text
stringlengths 2.85k
2.55M
| label
class label 11
classes |
|---|---|
arXiv:1304.0305v2 [math.AC] 3 Apr 2013
ON THE NUMBER OF GENERATORS OF IDEALS DEFINING
GORENSTEIN ARTIN ALGEBRAS
WITH
HILBERT FUNCTION
n+1
n+1
(1, n + 1, 1 + 2 , . . . , 2 + 1, n + 1, 1)
SABINE EL KHOURY, A. V. JAYANTHAN, AND HEMA SRINIVASAN
Abstract. Let R = k[w, x1 , . . . , xn ]/I be a graded Gorenstein Artin algebra . Then
I = ann F for some F in the divided power algebra kDP [W, X1 , . . . , Xn ]. If RI2 is
a height one ideal generated by n quadrics, then I2 ⊂ (w) after a possible change of
variables. Let J = I ∩ k[x1 , . . . , xn ]. Then µ(I) ≤ µ(J) + n + 1 and I is said to be generic
if µ(I) = µ(J) + n + 1. In this article we prove necessary conditions, in terms of F , for
an ideal to be generic. With some extra assumptions on the exponents of terms of F ,
we obtain a characterization for I = ann F to be generic in codimension four.
Introduction
Let R = k[x1 , . . . , xr ] and I be a homogeneous ideal of R. Let A = R/I be an Artinian
Gorenstein quotient of R. If r = 2, then it is known that I is a complete intersection. If r =
3, then by Buchsbaum and Eisenbud structure theorem [2], I is the (2n)th order pfaffians of
a (2n+1)th order skew symmetric matrix. When r = 4, Kustin and Miller gave a structure
theorem for Gorenstein Artinian ideals of the form (f, g, h, x4 J), where J is a Gorenstein
ideal of height three. Let I = ⊕n≥1 In be the direct sum decomposition of a Gorenstein
Artinian ideal in k[x, y, z, w]. In [4], Iarrobino and Srinivasan studied several properties
of the Gorenstein ideals I such that I2 ∼
= hwx, wy, w 2i. They gave
= hwx, wy, wzi or I2 ∼
a structure theorem for ideals I with I2 = hwx, wy, wzi and Hilbert function HR/I =
(1, 4, 7, . . . , 1). They did this by establishing a connection between some properties of the
ideal I and J = I ∩ R′ , where R′ = k[x, y, z], which is a height three Gorenstein ideal.
When I2 = hwx, wy, w 2i, they named these algebras mysterious Gorenstein algebras and
studied their various properties. They showed that such an ideal can be obtained as an
annihilator of a homogeneous form F = G(X, Y, Z) + W Z [j−1] ∈ kDP [X, Y, Z, W ], where
kDP [X, Y, Z, W ] denotes the divided power algebra. They studied the Hilbert function
properties of R/I connecting it with those of R′ /J.
In [3], El Khoury and Srinivasan studied certain properties of Gorenstein Artinian
algebras of the form R/I, where I2 = hwx, wy, w 2i. They gave a structure theorem for
1
2
SABINE EL KHOURY, A. V. JAYANTHAN, AND HEMA SRINIVASAN
such ideals. They showed that I is generated by I2 , J = I ∩ R′ and an element of the
form wz β − g, where g ∈ R′ ∈
/ I.
It can easily be seen that gx, gy ∈ J. El Khoury and Srinivasan proved that
wx, wy, w 2, α1 , . . . , αn−1 , z j , wz β − g is a minimal generating set for I unless gx or gy is
already a minimal generator of J = (α1 , . . . , αn−1 , z j ). In that case, dropping each of gx
and gy that is a minimal generator for J, the remaining n + 3 or n + 2 elements minimally
generate I. They obtained a minimal free resolution of R/I in all three cases where neither
gx nor gy is a part of a minimal generating set for J, either one of them is a minimal
generator and both are minimal generators for J. They also studied a special case of these
ideals namely, F = X [a] H(Y, Z) + W Z [j−1] . In that case, they described completely the
minimal free resolution of R/I. They concluded their article with an interesting question
on classification of ideals I with n + 2, n + 3 and n + 4 number of generators where n
denotes the minimum number of generators of J.
In this article, we consider the problem in higher embedding dimensions and partially
answer this question.
Let R = k[w = x0 , x1 , . . . xn ] be a standard graded ring of dimension n + 1 and I a ho
mogeneous Gorenstein ideal of height n+1 with H(R/I) = (1, n+1, 1+ n+1
, . . . , n+1
+
2
2
1, n + 1, 1). Suppose the ideal RI2 has height one. Then after a possible change of vari[j−1]
ables, I equals ann F , where F = G(X1 , . . . , Xn ) + W [j] or F = G(X1 , . . . Xn ) + W Xn
under Macaulay equivalence.
Let J = I ∩ k[x1 , . . . xn ] = I ∩ R′ so that R = R′ [w]. If F = G(X1 , . . . Xn ) + W [j], then
J is Gorenstein of height n and I is minimally generated by J, wxi , 1 ≤ i ≤ n − 1, w β − g
[j−1]
for some g ∈ R′ and not in I. If F = G(X1 , . . . Xn ) + W Xn , then I is generated by
J, wxi , 1 ≤ i ≤ n − 1, w 2, wxβn − g for some β ≤ j − 1 and g ∈ R′ \J. In the generic
case, all of these are minimal generators, that is gxi is not a minimal generator of J
for any i. Thus, if G is sufficiently general, µ(I) = µ(J) + n + 1 and we say G or
I = ann(G + W Z [j−1]) is generic. Since there is a one-to-one correspondence between
the height four Gorenstein ideals I with the property that I2 ∼
= (wxi , 1 ≤ i ≤ n − 1, w 2 )
and homogeneous forms F = G(X1 , . . . Xn ) + W Z [j−1] ∈ kDP [W, X1 , . . . Xn−1 , Xn = Z],
classifying such ideals is equivalent to classifying these homogeneous forms in the divided
power algebra. Therefore, we try to classify the property of I being generic in terms of
certain properties of the homogeneous form F such that I = ann F .
∂G
+ W Z [j−1] )
In Section 1, we begin by comparing I = ann(G + W Z [j] ) and Ixi = ann( ∂X
i
in general and show that when Xi divides G, I can be generic only if Ixi is generic. When
ON THE NUMBER OF GENERATORS OF IDEALS DEFINING GORENSTEIN ARTIN ALGEBRAS 3
n = 3, for Gorenstein Artinian algebras, we have more specific results. We prove that if
I = ann(F ) is generic, then there are some relations among the X, Y and Z-degrees of
F.
In Section 2, we study the case when F = G(Y, Z) + X [a] G1 + W Z [j−1] . When a > j/2
we prove that I is generic if and only if max degY (G) = max degY (G1 ). In the next section,
we discuss the general case, where F = G(Y, Z) + X [a1 ] G1 (Y, Z) + · · · + X [an ] Gn (Y, Z) +
W Z [j−1] . Assuming that j − an < a1 < · · · < an , we obtain necessary and sufficient
conditions for I to be generic.
We conclude our article by comparing the Hilbert functions of R/I and R′ /J. As
a consequence, we show that certain classes of Cohen-Macaulay height three ideals in
k[x, y, z] are unimodal, even though they are not Gorenstein.
1. Genericity
Let R be a standard graded k algebra of dimension n + 1 and I be a graded Gorenstein
ideal of height n+1 such that H(R/I) = (1, n+1, 1+ n+1
, . . . , n+1, 1). Then the ideal I
2
has n quadrics amongst its minimal generators. Suppose these quadrics generate an ideal
I2 of height one, then there exists a one form w such that I2 ⊂ (w) and R = k[w, x1 , . . . xn ]
for some suitable one form xi . Let R′ = k[x1 , . . . , xn ] and J = I ∩ R′ .
We will make use of the Macaulay equivalence between graded Gorenstein ideals of
socle degree j in R = k[w, x1 , . . . xn ] and j-forms in the divided power algebra R∗ =
kDP [X1 , . . . , Xn ] by the action of R on R∗ by differentiation. If F ∈ R∗ , then ann F =
{f ∈ R | (∂/∂f )(F ) = 0}.
The multiplication in the divided power algebra is different from the usual polynomial
X [a+b] .
algebra : X [a] · X [b] = (a+b)!
a!b!
[j−1]
In our situation, I = ann F , where F = G + W [j] or G + W Xn
depending on
whether w 2 ∈
/ I or w 2 ∈ I. When G is generic among R[X1 , . . . Xn ], µ(I) = µ(J) + n + 1.
We say that F is generic mod W and the corresponding ideal I = ann F is generic if
µ(I) = µ(J) + n + 1.
If w 2 ∈
/ I, then we can see that I2 = (wxi , 1 ≤ i ≤ n) and R′ /J is a Gorenstein Artin
algebra of embedding dimension n. The minimal number of generators for I, µ(I) is
precisely n+1 more than that of J for I = (J, wx1 , . . . , wxn , w j −g) for some g ∈ R′ , g ∈
/ J.
Thus µ(I) = µ(J) + n + 1 and in this case, I is always generic in our sense.
From now on, we consider the case where w 2 ∈ I. Then w 2 is one of the quadrics
minimally generating I and for a suitable choice of one forms x1 , . . . xn−1 , xn = z, I2 ∼
=
4
SABINE EL KHOURY, A. V. JAYANTHAN, AND HEMA SRINIVASAN
(w 2 , wx1 , . . . wxn−1 ). There exists a unique F = G(X1 , . . . , Xn−1 , Z) + W Z [j−1] of degree
j in the divided power algebra such that I = ann F . It has been proved in [3] that I is
generated by (I2 , J, wz β − g) with J = I ∩ k[x1 , . . . , xn = z], g ∈ k[x1 , . . . , xn−1 , z] \ J
and β ≤ j − 1. Suppose J is minimally generated by α1 , . . . , αn = z j . Since wxi are in
I, gxi , 1 ≤ i ≤ n − 1 belong to J. Therefore, if gxi ∈ nJ, where n = (x1 , . . . xn−1 , z) ⊂
k[x1 , . . . , xn−1 , z], then they are not part of any minimal generating set for J and hence
µ(I) = µ(J) + n + 1. Depending on whether gxi ’s are part of a minimal generating set
for J, the number of minimal generators for I will be µ(J) + t for 2 ≤ t ≤ n + 1. We
summarize this as:
Remark 1.1. If A = R/I is a Gorenstein Artin algebra with Hilbert function (1, n +
1, 1 + n+1
, . . . , n + 1, 1), then after a linear change of variables, R = k[x1 , . . . , xn , w]
2
and there exists a unique minimal generator for I of the form wxtn − g(x1 , . . . xn ). If
J = I ∩ k[x1 , . . . , xn ], then µ(I) = µ(J) + n + 1 − r, where r is the cardinality of {i | gxi ∈
J/mJ}.
Thus, I is generic if and only if gxi ∈ (x1 , . . . xn )J for all 1 ≤ i ≤ n − 1. The purpose of
this paper is to classify these polynomials F that give rise to generic ideals I.
We begin by proving a result which helps us restrict our study to a simpler class of
polynomials. Let F = G(X1 , . . . , Xn−1 , Z) + W Z [j−1] and I = ann F . Without loss of
generality, we may also assume that G does not have a term in LZ j−1 where L is a linear
form for in that case we can replace W by (W +L). In what follows we will take I = ann F
with F as above. For any form F ∈ kDP [X1 , . . . , Xn−1 , Z, W ], FX denotes the partial with
respect to X.
Theorem 1.2. Suppose F ′ = Xt G + W Z [j] for some t, 1 ≤ t ≤ n − 1. If ann F is generic,
then so is ann F ′ .
Proof. Let Xt = X, I ′ = ann F ′ and I = ann F . Let R′ = k[x1 , . . . , xn−1 , z] and J ′ =
′
I ′ ∩k[x1 , . . . , xn−1 , z]. Then I ′ = (wxi , 1 ≤ i ≤ n−1, w 2, J ′ , wz β −g ′ ) for some g ′ ∈ R′ \ J ′
′
and β ′ ≤ j. We first show that wz β −g ′ can be replaced with wz β+1 −h for some h ∈ R′ \ J ′
and hx ∈ nJ ′ .
Claim: wz β − g ∈ I for some g if and only if wz β+1 − gx ∈ I ′ .
ON THE NUMBER OF GENERATORS OF IDEALS DEFINING GORENSTEIN ARTIN ALGEBRAS 5
Proof of the claim: If wz β − g ∈ I, then
0 =
∂G
∂W Z [j−1] ∂G ∂W Z [j−1]
∂F
=
+
−
−
∂wz β − g
∂wz β
∂wz β
∂g
∂g
= Z [j−β−1] −
This implies
∂G
∂Z [j−1]
−W
.
∂g
∂g
∂Z [j−1]
∂G
= 0 and
= Z [j−β−1] . Now, consider
∂G
∂g
∂G′
∂W Z [j]
∂G′ ∂W Z [j]
∂(G′ + W Z [j] )
=
+
−
−
∂wz β+1 − gx
∂wz β+1 ∂wz β+1 ∂gx
∂gx
∂G
= Z [j−β−1] −
∂g
= 0.
Therefore wz β+1 − gx ∈ ann(G′ + W Z [j]) = I ′ .
Suppose wz t − h′ ∈ I ′ for some t < β + 1 and h′ ∈
/ J ′ . Then
∂G′
∂W Z [j] ∂G′ ∂W Z [j]
∂(G′ + W Z [j])
=
+
−
−
wz t − h′
∂wz t
∂wz t
∂h′
∂h′
∂Z [j]
∂G′
−
W
.
= Z [j−t] −
∂h′
∂h′
0 =
∂Z [j]
∂G′
=
0
and
= Z [j−t] . This implies that x divides h′ . Let h′ = xh for some
∂h′
∂h′
h. Consider wz t−1 − h,
Hence
∂G
∂W Z [j−1] ∂G ∂W Z [j−1]
∂G + W Z [j−1]
=
+
−
−
∂wz t−1 − h
∂wz t−1
∂wz t−1
∂h
∂h
[j−1]
∂G
∂Z
= Z [j−t] −
−W
.
∂h
∂h
∂Z [j−1]
is non-zero, then h contains a pure power of z, which must be z t . But
∂h
∂G′
= Z [j−t] would then mean that G′ has a term XZ j which is not possible by our
∂h′
assumption on F .
∂Z [j−1]
∂G′
∂G + W Z [j−1]
Therefore
= 0. Since G =
and h′ = xh, we get
= 0. There∂h
∂x
∂wz t−1 − h
fore wz t−1 − h ∈ I. This contradicts the minimality of β. This completes the proof of the
claim.
If
6
SABINE EL KHOURY, A. V. JAYANTHAN, AND HEMA SRINIVASAN
Suppose I = (I2 , J, wz β − g) is generic and let I ′ = (I2 , J ′ , wz β+1 − gx). First note that
if h ∈ nJ, then hxt ∈ nJ ′ for all t. Since gxt ∈ nJ, gxs xt ∈ nJ ′ for all 1 ≤ s, t ≤ n.
Therefore they are not part of minimal generating set for J ′ . Hence I ′ is generic.
We may also restate the theorem as a necessary condition as follows:
Let I = ann F = G+W Z [j]. Suppose G = Xt GXt for some 1 ≤ t ≤ n−1. Then I = ann F
is generic if ann GXt + W Z [j−1] is generic.
As a result of the above theorem, we concentrate on polynomials F = G + W Z [j−1] in
the divided power algebra such that none of the Xi , 1 ≤ i ≤ n − 1 divides G and of course
that G has no term containing Z [j−1].
The converse of the above theorem is not true in general as we can see in the Example
4.2.
We remark that it is not always the case that we can achieve ’genericity’ by multiplying
by an Xi even if one of gxj is a non-minimal generator for the corresponding ideal J.
See the examples in the last section. However, it is an interesting question, whether if
[j−1]
F = G + W Xn
∈ kDP [X1 , . . . Xn , W ] is not generic, does there exist a suitable power
of Xi , 1 ≤ i ≤ n − 1 multiplying G by which will result in a generic ideal, better yet, does
[j]
there exist a suitable one form L(X1 , . . . Xn−1 ) so that I ′ = ann(LG + W Xn ) is such that
µ(I ′ ) = n + 1 + µ(I ′ ) ∩ k[x1 , . . . , xn ]?
2. Embedding Dimension Four
In embedding dimension 4, we can get a stronger characterization. We now let n = 3
so that R = k[w, x, y, z].
Notation 2.1. For the rest of the paper, we set
F = G0 + X [a1 ] G1 + · · · + X [al ] Gl + W Z [j−1]
!
m
X
X
cr1k Y [r1k ] Z [s1k ]
=
cp−t Y [p−t] Z [q+t] + X [a1 ]
t=0
+ · · · + X [an ]
r1k +s1k =j−a1
X
rnk +snk =j−an
crnk Y [rnk ] Z [snk ]
!
+ W Z [j−1],
where cp 6= 0, a1 < · · · < an and one of the G′i s contain a pure power of Z.
For a polynomial h(x1 , . . . , xl ) ∈ k[x1 , . . . , xl ], let degxi h denote the highest power of
xi in h. Thus, in F as in 2.1, degX F = an ,.
ON THE NUMBER OF GENERATORS OF IDEALS DEFINING GORENSTEIN ARTIN ALGEBRAS 7
We first obtain a necessary condition for an ideal to be generic in terms of F . Before
we prove the result, we prove a technical, but very important lemma that is needed in
the proof of this theorem. This lemma will play a crucial role in the proofs of all the
forthcoming characterization results as well.
Lemma 2.2. With the notation as in 2.1, if degY G0 = p then ann G0 doesn’t have
elements of degree less than or equal to j − p other than y p+1 when p < j − p.
Proof. Clearly y p+1 ∈ ann G0 . Suppose ann G0 contains a generator of degree less than q,
P
say g = ki=0 αi y r−i z i for some 0 ≤ k ≤ r, where αk 6= 0. Then we have
0=
∂cp Y [p] Z [q] + cp−1 Y [p−1] Z [q+1] + . . . + cp−m Y [p−m] Z [q+m]
∂G0
=
∂g
∂αr y r + αr−1 y r−1z + · · · + αk y r−k z k
= αk cp Y [p−r+k]Z [q−k] + terms in Y of degree less than p − r + k.
Since cp 6= 0, αk = 0, which is a contradiction. Hence the assertion follows.
Theorem 2.3. With the notation as in 2.1, if I is generic, then either a1 ≤ degY G0 or
degY G0 ≤ max{degY Gi : i = 1, . . . , n}.
Proof. We first show that if f ∈ ann F with deg f < a1 , then either f ∈ ann G0 ∩
P
P
ann W Z [j−1] ∩ ann( ni=1 X [ai ] Gi ) or f ∈ ann(G0 + W Z [j−1]) ∩ ann( ni=1 X [ai ] Gi ). For if
f ∈ ann F , then
n
∂G0 X ∂X [ai ] Gi ∂W Z [j−1]
∂F
=
+
+
.
0=
∂f
∂f
∂f
∂f
i=1
Therefore
n
X ∂X [ai ] Gi
∂G0 ∂W Z [j−1]
+
=−
.
∂f
∂f
∂f
i=1
Since G0 is a polynomial in Y and Z, the term on the left hand side of the above equality
P
[a ]
Z [j−1] )
does not involve X. Since deg f < a1 we get that ni=1 ∂X∂fi Gi = 0 and ∂(G0 +W
= 0.
∂f
We now proceed to the proof of the theorem. We show that if ai > p > rik for every i
and k, where p = degY G0 , then I is not generic. We do this by considering the following
two cases:
Case 1. Suppose p ≤ j − p. By Lemma 2.2, y p+1 is a minimal generator for ann G0 since
all other generators have degree bigger than j − p. Since p > rik for all i, k, it can be seen
that y p+1 ∈ ann G0 ∩ ann W Z [j−1] ∩ ann X [ai ] Gi for all i. We then conclude that y p+1 is a
minimal generator for ann F .
On the other hand, note that f = wz p−1 − y p ∈ ann F . Since the degree of f = p < a1
8
SABINE EL KHOURY, A. V. JAYANTHAN, AND HEMA SRINIVASAN
P
and contains a term in w, we get f ∈ ann(G0 + W Z [j−1]) ∩ ann( li=1 X [ai ] Gi ). We show
that f is a minimal generator. Suppose wz β − g ∈ I for some β < p − 1. It implies
that y p − z p−1−β g ∈ I which is impossible since y p+1 is a minimal generator. Therefore,
wz β − g ∈
/ I for any β < p − 1 so that f is a minimal generator for I. Since y p+1 is a
minimal generator, I is not generic.
Case 2. Suppose p > j − p. In that case wz p−1 − y p ∈ ann F , but might not be a
minimal generator for I. Let f = wz β−1 − g(x, y, z) be a minimal generator for I, with
β ≤ p. We first show that we can replace g(x, y, z) by g(y, z). We write g(x, y, z) =
g1 (y, z) + xg2 (x, y, z). Then,
n
X
∂X [ai ] Gi
∂G0
∂W Z [j−1]
∂F
=−
−
+
= 0.
∂f
∂g1 (y, z) i=1 ∂g1 (y, z) + xg2 (x, y, z)
∂wz β−1
n
X
∂X [ai ] Gi
Since ai > p ≥ β, we get
= 0. Therefore, either g1 = y p
=Z
and
∂g
+
xg
1
2
i=1
or the degree of g1 is strictly less than p. But in both cases and by Lemma 2.2, the degree
of g1 is greater than j − p. On the other hand, the degree of Gi is at most j − p − 1. It
[a ]
follows that ∂X∂gi1 Gi = 0 for all i, and hence xg2 ∈ J. So g(x, y, z) can be replaced by
g(y, z).
∂F
We know that ∂g(y,z)
= Z [j−β] and yg(y, z) ∈ ann F . In fact, yg(y, z) ∈ ann G0 ∩
P
ann(W Z [j−1] )∩ann( ni=1 X [ai ] Gi ). As in the case of the proof of Lemma 4.2 in [3], it can be
seen that ann G0 is minimally generated in k[y, z] by a regular sequence (yδ(y, z), θ(y, z))
with θ(y, z) = z t + θ1 (y, z). By Lemma 2.2, the degrees of yδ(y, z) and θ(y, z) are
at least j − p + 1 and the degrees of the G′i s are at most j − p − 1. It follows that
P
/ ann W Z j−1 . Therefore
yδ(y, z) and θ(y, z) ∈ ann G0 ∩ ann( ni=1 X [ai ] Gi ). But θ(y, z) ∈
θ(y, z) ∈
/ ann F , but yθ(y, z) ∈ ann F . Hence, yδ(y, z) and yθ(y, z) are minimal generators
for ann F .
We now show that yg(y, z) is a minimal generator for I and can be chosen to be yδ(y, z).
Suppose yg(y, z) is not minimal then yg(y, z) = f1 (y, z)yδ(y, z) + f2 (y, z)yθ(y, z). It
implies that g(y, z) = f1 (y, z)δ(y, z) + f2 (y, z)θ(y, z). Consider
∂G0
∂g1
[j−β]
∂G0
∂G0
∂G0
=
=
= Z [j−β].
∂g(y, z)
∂f1 (y, z)δ(y, z) + f2 (y, z)θ(y, z)
∂f1 (y, z)δ(y, z)
We can choose g(y, z) = f1 (y, z)δ(y, z). If f1 (y, z) is a constant then we are done. Otherwise f1 (y, z) = cz u since yδ(y, z) is in J. On the other hand, δ(y, z) cannot have a pure
power of z, otherwise g(y, z) will have a pure power of z and will not belong to I. But
ON THE NUMBER OF GENERATORS OF IDEALS DEFINING GORENSTEIN ARTIN ALGEBRAS 9
∂G0
= Y ∂yδ(y,z)
+ αZ [γ] = αZ [γ] with α 6= 0 and the multiplication is the usual polynomial multiplication. It implies that f ′ = wz β−1−u − δ(y, z) is a minimal generator for I
which contradicts the minimality of f . Hence, g(y, z) = cδ(y, z) and yg(y, z) is minimal.
Hence I is not generic.
∂G0
∂δ(y,z)
In [3], it was proved that if F = X [a1 ] G1 + W Z [j−1] with a1 ≥ 1, then I is generic.
Example 4.8 shows that this property does not hold if F = G0 + X [a] G1 + W Z [j−1]. Let
F = G0 + X [a] G1 + W Z [j−1]
m
X
=
cp−r Y [p−r] Z [q+r] + X [a]
r=0
X
αri Y [ri ] Z [si ]
ri +si =j−a
!
+ W Z [j−1].
If we suppose that a > j − a ≥ p, then we get an improved version of Theorem 2.3 in
this case.
Theorem 2.4. Let F be as above and suppose degX F > deg G1 ≥ degY G0 . Then I is
generic if and only if degY G1 = p.
[p] [a+s]
Proof. Suppose degY G1 = p. We write F = c−1
Z
+ · · · + c2 Y [2] Z [a+s+p−2] +
p Y
X [a] (αp Y [p] Z [s] + · · · + α0 Z [p+s] ) + W Z [j−1] with cp , αp and α0 6= 0.
p s+1
We first show that wz p+s − c−1
is a minimal generator for I. Clearly, wz p+s −
p y z
p s+1
c−1
∈ I. Suppose there exists wz β − g ∈ I for some β < p + s. As in the proof of
p y z
Theorem 2.3, we can assume that g is a function of y and z. Then we have
∂F
0 =
∂wz β − g
=
Therefore
∂G0 + X [a] G1 + W Z [j−1]
.
∂wz β − g
∂G0 + X [a] G1
= Z [j−1−β] = Z [a] .
∂g
0
Since G0 does not involve X and g has degree p + s < a, we get ∂G
= Z [a] and
∂g
∂G1
= 0. Therefore, there exists 0 ≤ i ≤ p − 2 < a with cp−i 6= 0 such that g =
∂g
−1 p−i s+i
[i] [a−i]
0
= cp .c−1
+ · · · + Z [a] + · · · and g1
cp−i y z + g1 (y, z). If i 6= 0, then ∂G
p−i Y Z
∂g
does not annihilate G0 − cp−i Y [p−i]Z [a+s+i] by Lemma 2.2, which is a contradiction. Hence
[a]
∂G0
[a]
[a]
p s
and ∂X∂g G1 = c−1
i = 0 and g = c−1
p αp X . This
p y z + g1 (y, z). Therefore, ∂g = Z
implies that there exists a term g2 (y, z) in g1 (y, z) in y and z of total degree p + s such
[a]
∂G0 −cp Y [p] Z [a+s]
[a]
that ∂X∂g2G1 = −c−1
6= 0 by Lemma 2.2. Continuing in this
p αp X . Again
∂g2
β
manner, we see that wz − g does not annihilate F , which is a contradiction. Therefore
10
SABINE EL KHOURY, A. V. JAYANTHAN, AND HEMA SRINIVASAN
p s+1
p + s is the minimum exponent β of z such that wz β − g ∈ I and hence wz p+s − c−1
p y z
is a minimal generator in I.
Since y p+1 ∈ I, y ·y pz s+1 is not a minimal generator of I. Note also that if j is the smallest
integer so that αp−j 6= 0, then x(αp−j y pz s − αp y p−j z s+j ) ∈ I. Therefore, x · y p z s+1 is also
not a minimal generator of I. Hence I is generic.
We now prove the converse. We know by Theorem 2.3 that if a > p and I is generic,
then αp+i 6= 0 for some i ≥ 0. With the extra assumption that F has only two terms and
a > j − a, we show that αp 6= 0 and αp+i = 0 for i > 0. So it suffices to show that if there
exists an i > 0 such that αp+i 6= 0, then I is not generic. Suppose αp+i 6= 0 for some i.
Let
!
p+r
X
αm Y [m] Z [p+s−m] + W Z [j−1] ,
F = cp Y [p]Z [a+s] + · · · + c2 Y [2] Z [a+s+p−2] + X [a]
m=0
where αp+i 6= 0 for some i > 0. It is clear that wz p+s − cp −1 y p z s+1 ∈ I. If it is not a
minimal generator for I, then there exists β < p + s such that wz β − g(y, z) is minimal.
Then we have
∂F
∂G0
∂G1
= Z [j−1−β] −
− X [a]
= 0.
β
∂wz − g(y, z)
∂g
∂g
1
= 0 and ∂G
= Z [j−1−β]. We first claim that g(y, z) =
Since a > p + s, we get ∂G
∂g
∂g
cp −1 y p z q + g1 (y, z), where p + q = β + 1 and the exponents of y in g1 (y, z) are greater
p−i q
than p + 1. Let g(y, z) = c−1
z + g1 (y, z), with i ≥ 0 and the exponents of y
p−i y
∂(G −c
Y [p−i] Z [a+s+i])
0
p−i
in g1 strictly bigger than p − i. Then,
= 0. But by Lemma 2.2,
∂g(y,z)
[p−i] [a+s+i]
ann(G0 − cp−i Y
Z
) doesn’t contain generators of degree less than a + s. This
implies that i = 0. Hence g(y, z) is of the required form.
p q
Let q be the minimal power of z such that wz β −c−1
p y z −g1 (y, z) is a minimal generator
p q
−1 p q
for I with g1 ∈ ann G and c−1
p y z + g1 ∈ ann G1 . We show that f = y(cp y z +
g1 (y, z)) is a minimal generator for I. Suppose not, then f can be obtained from a
combination of a generator of the form y p+1z q−k + g2(y, z) and other generators of I. Note
that y p+1z q−k + g2 (y, z) ∈ ann G0 and hence belongs to ann G1 . We have also seen that
p q
p+1 q−k
c−1
z
+ g2 (y, z)
p y z + g1 (y, z)) ∈ ann G1 . It is clear that neither of the generators y
−1 p q
and cp y z + g1 (y, z) can be obtained from each other. We then study the minimal
generators of ann G1 . We know that ann G1 is minimally generated by a regular sequence
of the form (yδ(y, z), θ(y, z)) with θ(y, z) = z t + θ1 (y, z). For these generators to be in
I, no pure power of z of degree less than j and no terms with degree of y less than p
can appear in any of them. In that case, yδ(y, z) and θ(y, z) will be multiplied by the
ON THE NUMBER OF GENERATORS OF IDEALS DEFINING GORENSTEIN ARTIN ALGEBRAS 11
appropriate power of y to get rid of the power of z and have all powers of y in their terms
p q
greater or equal than p+1. Since both generators y p+1z q−k +g2 (y, z) and c−1
p y z +g1 (y, z)
p q
are independent in ann G1 , y p+1z q−k + g2 (y, z) and y(c−1
p y z + g1 (y, z)) are independent
in I. Hence f is minimal and I is not generic.
Example 4.9 shows that statement of Theorem 2.4 does not hold if p < a < p + s.
Now, we will consider the general case in codimension four. Let F be as in Notation
2.1. We now obtain a generalization of Theorem 2.4. Along with Notation 2.1 we further
assume that Gn has a pure power of Z (in other words, Y does not divide Gn ). We first
prove a necessary condition for an ideal to be generic.
Theorem 2.5. Let F be as in Notation 2.1 with j − an < a1 , Y not dividing Gn and
degY Gi < degY G0 for all i = 1, . . . , n − 1. If I is generic, then degY Gn = degY G0 .
Proof. By Theorem 2.3, it suffices to show that if there exists an k such that rik > p in
p s+1
Gn , then I is not generic. It is clear that wz p+s − c−1
∈ I. Suppose it is not a
p y z
minimal generator for I, then there exists β < p + s such that wz β − g(y, z) is minimal.
Since
∂G X ∂X [ai ] Y Gi
∂Gn
∂F
[j−1−β]
=
Z
−
−
− X [an ]
0=
β
∂wz − g(y, z)
∂g
∂g
∂g
i
we get ∂G
= Z [j−1−β] and the rest is zero.
∂g
p q
As in the case of the proof of Theorem 2.4, we can see that g(y, z) = c−1
p y z + g1 (y, z)
with p + q = β + 1 and g1 (y, z) having all powers of y greater than p + 1. We note that
p q
g1 ∈ ann G0 and c−1
p y z + g1 ∈ ann Gn ∩ ann Gi for all i = 1, . . . , n − 1 in that case. It
p q
can be seen that f = y(c−1
p y z + g1 (y, z)) is a minimal generator for I. For, if not, then
f can be expressed as a linear combination of other generators of I, one of which is of
the form y p+1z q−k + g2 (y, z). Note that y p+1z q−k + g2 (y, z) ∈ ann G0 ∩ (∩i ann Gi ) and
p q
p+1 q−k
c−1
z + g2(y, z)
p y z + g1 (y, z)) ∈ ∩i ann Gi . It is clear that neither of the generators y
−1 p q
and cp y z + g1 (y, z) can be obtained from each other.
Again, looking at the minimal generators of ann Gn , we know they are a regular sequence
of the following form (yδ(y, z), θ(y, z)) with θ(y, z) = z t + θ1 (y, z). For these generators to
be in I, they must be multiplied by the appropriate power of y to get rid of the power of z
and have all powers of y in their terms greater or equal than p + 1. Since both generators
p q
p+1 q−k
y p+1z q−k +g2 (y, z) and c−1
z +g2 (y, z) and
p y z +g1 (y, z) are independent in ann Gn , y
−1 p q
y(cp y z + g1 (y, z)) are independent in I. Hence f is minimal and I is not generic.
12
SABINE EL KHOURY, A. V. JAYANTHAN, AND HEMA SRINIVASAN
Examples 4.10 and 4.11 show that the converse of Theorem 2.5 is not true in general.
These examples suggest that we have to add more conditions on the exponents to obtain
a converse of Theorem 2.5. We prove the converse with some extra assumptions.
Theorem 2.6. Suppose degY Gn = p, degZ Gn = p + s, deg Gn < a1 and degY Gi <
degY G0 for all i = 1, . . . , n − 1. If F satisfies one of the two conditions below, then I is
generic.
(1) degY Gi < degY (Gn − Y p Z s );
(2) degZ Gi < degZ Gn for all i < n.
p s+1
Proof. We first show that wz p+s − c−1
is a minimal generator for I. Suppose not,
p y z
β
let wz − g ∈ I, where β < p + s. Then
∂F
∂wz β − g
∂[G0 (Y, Z) + X [a1 ] G1 + · · · + X [an ] Gn + W Z [j−1] ]
=
.
∂wz β − g
0 =
Therefore
∂[G0 (Y, Z) + X [a1 ] G1 + · · · + X [an ] Gn ]
= Z [j−1−β]
∂g
Since ai > j − an = p + s, g may be assumed to be a function of y and z. Hence we get
∂G0 (Y, Z)
= Z [j−1−β] and
∂g(y, z)
∂X [ai ] Y Gi (Y, Z)
= 0.
∂g(y, z)
By Lemma 2.2, ann G0 (Y, Z) does not contain generators of degree less than an + s =
p s
j − p. Therefore g = c−1
p y z + g1 for some polynomial g1 . But in that case, we have
∂X [an ] Gn (Y,Z)
p s
∂c−1
p y z
[an ]
= αp c−1
and hence g1 must contain a term g2 with the power of y less
p X
[an ]
than p such that ∂X ∂gG2n (Y,Z) = −X [an ] . Continuing in this manner, we see that wz β − g
can not annihilate F , which is a contradiction. Therefore wz β − g 6∈ I for β < p + s and
p s+1
hence wz p+s − c−1
is a minimal generator.
p y z
Note that since y p+1 ∈ I, y · y pz s+1 is not a minimal generator. Also we have in case (1)
that x(αp y p−k z s+k − αp−k y p z s ) ∈ I and in case (2), x(αp z s+p − α0 y p z s ) ∈ I, which implies
that xy p z s is not a minimal generator of I. Hence I is generic.
3. Hilbert Functions of I and J
In this section we compare the Hilbert functions of I and J = I ∩ k[x, y, z]. It can be
noted that these results are independent of whether the ideal is generic or not.
ON THE NUMBER OF GENERATORS OF IDEALS DEFINING GORENSTEIN ARTIN ALGEBRAS 13
Proposition 3.1. Let R = k[x, y, z, w], F = G + W Z [j−1] , I = ann(F ) and J = I ∩
k[x, y, z]. Then HR/I − HR′ /J = [0, 1, . . . , 1, 0, . . .], where the last 1 occurs at the degree β.
Proof. Let I = (I2 , J, wz β − g), where g ∈ k[x, y, z] \ J. Clearly HR/I (0) − HR′ /J (0) = 0
and HR/I (1) − HR′ /J (1) = 1. Note that I2 = hwx, wy, w 2i and J has generators of degree
at least 3. Therefore HR/I (2) = 10 − 3 = 7 and HR′ /J (2) = 6. For 3 ≤ n ≤ β, In ,
as a k-vector space, is generated by I2 Rn−2 and Jn . If a monomial m1 ∈ [R/I]n is a
k-basis element, then either m = wz n−1 or w does not divide m and m is a k-basis
element of [R′ /J]n . Therefore, dimk [R/I]n = dimk [R′ /J]n + 1 for all 3 ≤ n ≤ β. For
n = β + 1, one can see that wz β − g ∈ I and g ∈
/ I and hence g ∈
/ J. Moreover, for
n
n−β−1
n−β−1
n−β−1
n ≥ β + 1, wz = z
g(mod I) and z
g ∈ I if and only if z
g ∈ J. Therefore,
HR/I (n) − HR′ /J (n) = 0 for n ≥ β + 1. Hence the assertion is proved.
It is known that height three Gorenstein ideal in k[x, y, z] is unimodal, see [7]. As a
corollary of the previous result, we obtain a class of Artinian level algebras of embedding
dimension three, namely type two and at least one of the socle elements is a pure power
of a one form, having unimodal Hilbert function.
Corollary 3.2. If J ⊂ R′ = k[x, y, z] is an ideal such that J = ann(G, Z [deg G−1] ) for
some polynomial G in the divided power algebra kDP [X, Y, Z], then the Hilbert function
of J is unimodal.
Proof. Let J be an ideal of the given form. Let I = G + W Z [j−1]. Then it can be seen
that J = I ∩ R′ . From Proposition 3.1, it follows that HR′ /J = HR/I − [0, 1, . . . , 1, 0..],
where the last 1 occurs at the degree β. Since I is a Gorenstein ideal with initial degree
2, HR/I is unimodal, [6, Theorem 3.1]. Therefore HR′ /J is unimodal.
4. Examples
In this section we provide some examples to illustrate our results. We follow the notation that was set earlier, i.e., given a homogeneous form F in the divided power algebra
kDP [W, X1 , . . . , Xn ], I = ann(F ) ⊂ k[w, x1 , . . . , xn ] and J = I ∩ k[x1 , . . . , xn ]. The following is an example of Theorem 1.2.
Example 4.1. Let F ′ = G + W Z [9] = X [8] Y [2] + X [3] Y [3] Z [4] + X [2] Y Z [7] + W Z [9] and
∂2G
[7]
= X [6] Y [2] +XY [3] ]Z [4] +Y Z [7] +W Z [7]. If I ′ = ann(F ′ ) and I = ann(F ),
F = ∂X
2 +W Z
then it can be seen that µ(I ′ ) = 13 and µ(J ′ ) = 9 and µ(I) = 9 and µ(J) = 5.
The following example shows that the converse of Theorem 1.2 is not true in general.
14
SABINE EL KHOURY, A. V. JAYANTHAN, AND HEMA SRINIVASAN
Example 4.2. Let I = ann(Y [5] Z [8] +Y [4] Z [9] +X [5] (Y [6] Z [2] +Y [7] Z +Y [8] +Z [8] )+W Z [12] )
and I ′ = ann(XY [5] Z [8] + XY [4] Z [9] + X [6] (Y [6] Z [2] + Y [7] Z + Y [8] + Z [8] ) + W Z [13] )). Then
µ(I) = 11, µ(J) = 8, µ(I ′ ) = 13 and µ(J ′ ) = 9. Therefore I ′ is generic, but I is not.
The following is an example where ann(G + W Z [j−1]) is not generic and ann(X t G +
W Z [t+j−1] ) is also not generic for some t.
Example 4.3. Let F = G+W Z [12] = Y [3] Z [10] +Y [4] Z [9] +Y [5] Z [8] +Y [2] Z [11] +X [6] (Y [4] Z [3] +
Y [2] Z [5] + Z [7] ) + X [7] (Y [3] Z [3] + Y Z [5] + Z [6] ) + W Z [12] . It can be seen that I = ann(F ) is
not generic. Moreover, ann(X t G + W Z [12+t] ) is not generic for t = 1, 2, 3 and that I =
ann(X 4 G+W Z [16] ) is generic. In a similar manner, one can see that ann(Y t G+W Z [12+t] )
is also not generic for t = 1, 2, 3, 4 and that ann(Y 5 G + W Z [17] ) is generic.
The following example, in codimension five, shows that I = ann(G + W Z [j−1] ) ⊂
k[w, t, x, y, z] need not be generic but I = ann(T G + W Z [j]) is generic.
Example 4.4. Let I = ann(Y [3] Z [9] + Y [4] Z [8] + T [3] Z [9] + X [3] (Y [4] Z [4] T + Z [9] + +W Z [11] )
and I ′ = ann(XY [3] Z [9] + XY [4] Z [8 + XT [3] Z [9] + X [4] (Y [4] Z [4] T + Z [9] + W Z [12] )). Then
µ(I) = 14, µ(J) = 10, µ(I ′) = 16 and µ(J ′ ) = 11. Here, I is not generic as t4 is a
minimal generator of J. But I ′ is generic for t4 is no longer the obstruction, for wz 3 −xt3
is the minimal generator of I ′ .
The following is an examples of an ideal which is not generic when a > p and p > rik .
Example 4.5. If F = Y [10] Z [5] + Y [9] Z [6] + Y [6] Z [9] + X [11] (Y [2] Z [2] + Z [4] ) + W Z [14] ,
then µ(I) = 9 and µ(J) = 6. Here wz 7 − y 8 + y 7z − y 6z 2 + y 5z 3 − yz 7 is a minimal
generator for I. It can also be seen that ann G is minimally generated in k[y, z] by (θ, δ1 ) =
(y 8 +y 7 z −y 5 z 3 −2y 4 z 4 −y 3 z 5 +y 2 z 6 +3yz 7 +3z 8 , y 7z 2 −y 4 z 5 −y 3 z 6 +yz 8 +2z 9 ). Note that
we may replace δ1 by 3δ1 + (2z + y)θ = yg(y, z) which makes yg(y, z) a minimal generator
(see the proof of Theorem 2.3).
We have noticed that if p ≤ rik for some i and k, then the classification is more
complicated. It does not only depend on the powers alone, but also on the coefficients as
we see in the example below:
Example 4.6. If I = ann(Y [3] Z [8] + Y [4] Z [7] + X [5] (Y [4] Z [2] + Y [3] Z [3] + Z [6] ) + W Z [10] )
then µ(I) = 9 and µ(J) = 5 whereas if I = ann(Y [3] Z [8] + Y [4] Z [7] + X [5] (2.Y [4] Z [2] +
Y [3] Z [3] + Z [6] ) + W Z [10] ) then µ(I) = 11 and µ(J) = 8.
Examples 4.7 and 4.8 are examples of Theorem 2.4.
ON THE NUMBER OF GENERATORS OF IDEALS DEFINING GORENSTEIN ARTIN ALGEBRAS 15
Example 4.7. Let F = Y [4] Z [10] + Y [3] Z [11] + X [8] (Y [4] Z [2] + Y [3] Z [3] + Y [2] Z [4] + Z [6] ) +
W Z [13] . In this case, degY G1 = p and one can see that the ideal I = ann(F ) is generic
with µ(I) = 11 and µ(J) = 7.
Example 4.8. Let F = Y [4] Z [10] + Y [3] Z [11] + X [8] (Y [5] Z + Y [4] Z [2] + Y [3] Z [3] + Y [2] Z [4] +
Z [6] ) + W Z [13] . In this case degY G1 > p and we have µ(I) = 11 and µ(J) = 8.
Example below shows that Theorem 2.4 doesn’t work if p ≤ a:
Example 4.9. Let F = Y [4] Z [7] + Y [3] Z [8] + X [4] (Y [4] Z [3] + Z [7] ) + W Z [10] . Here, a <
degX F = 4 < deg G1 = 7. One can see that µ(I) = 9 and µ(J) = 6. Whereas if
F = Y [4] Z [7] + X [4] (Y [4] Z [3] + Z [7] ) + W Z [10] , then µ(I) = 9 and µ(J) = 5. In the above
examples, we have degX F < deg G1 . While the first one is generic, the second one is not.
This shows that our characterization is not valid without the given hypotheses.
The example below shows that the converse of Theorem 2.5 is not true.
Example 4.10. It can be seen that ann(Y [5] Z [12] + Y [4] Z [13] + X [8] (Y [4] Z [5] + Y [3] Z [6] +
Y [2] Z [7] +Y Z [8] )+X [9] (Y [4] Z [4] +Y [3] Z [5 )+X [10] (Y [5] Z [2] +Y [4] Z [3] +Y [3] Z [4] +Z [7] )+W Z [16] )
is generic whereas ann(Y [5] Z [12] + Y [4] Z [13] + X [7] (Y [4] Z [6] + Y [3] Z [7] + Y [2] Z [8] + Y Z [9] ) +
X [9] (Y [4] Z [4] + Y [3] Z [5] ) + X [11] (Y [5] Z + Y [4] Z [2] + Y [3] Z [3] + Z [6] ) + W Z [16] ) is not. Note
that in both cases j − a1 < a1 , degY Gi < degY G0 and degY G3 = degY G0 = 5.
The example below shows that we cannot remove the conditions degY = Gi < degY =
G0 , 1 ≤ i ≤ n − 1 in Theorem 2.5. In F , all conditions except degY G1 = 6 > degY G0 = 5
are satisfied and ann F is generic and the conclusion fails where as in G all conditions
except degY = G0 6= degY = Gn . Hence by the theorem 2.5, ann G cannot be generic.
Example 4.11. Let F = Y [5] Z [12] + Y [4] Z [13] + X [9] (Y [4] Z [4] ) + X [10] (Y [6] Z + Y [2] Z [5] +
Y [3] Z [4] + Z [7] ) + X [12] (Y Z [4] + Z [5] ) + W Z [16] . In this case, 6 = r21 > p = 5, µ(I) = 15
and µ(J) = 11, which implies that I is generic. On the other hand, if G = Y [5] Z [12] +
Y [4] Z [13] + X [9](Y [4] Z [4] ) + X [10] (Y [6] Z + Y [2] Z [5] + Y [3] Z [4] + Z [7] ) + W Z [16] , then µ(I) = 15
and µ(J) = 12 as predicted by Theorem 2.5.
References
[1] D. Eisenbud, Commutative algebra. With a view toward algebraic geometry, Graduate Texts in Mathematics, 150. Springer-Verlag, New York, 1995.
[2] D. A. Buchsbaum and D. Eisenbud. Algebra structures for finite free resolutions, and some structure
theorems for ideals of codimension 3, American Journal of Mathematics. 99 (1977), no.3, 447-485.
16
SABINE EL KHOURY, A. V. JAYANTHAN, AND HEMA SRINIVASAN
[3] S.El Khoury and H.Srinivasan, A Class of Gorenstein Artinian Algebras of Embedding Codimension
Four, Comm. in Algebra, Vol 37, (2009), Issue 9, 3259 – 3277.
[4] A. Iarrobino and H. Srinivasan, Artin Gorenstein algebras of embedding dimension four: components
of PGor(H) for H = (1, 4, 7, . . . , 1) J. Pure Appl. Algebra 201 (2005) 62–96.
[5] F.H.S. Macaulay, The Algebraic Theory of Modular Systems, Cambridge University Press, Cambridge, UK, (1916) reprinted with a foreword by P.Roberts, Cambridge University Press, London,
New York, 1994.
[6] J. Migliore, U. Nagel and F. Zanello, A Characterization of Gorenstein Hilbert functions in codimension four with small initial degree, Math. Res. Lett. 15 (2008) no. 2, 331 – 349.
[7] R. P. Stanley, Hilbert functions of graded algebras, Math. Res. Lett. 28(1978) no. 1 57 – 83.
Department of Mathematics, American University of Beirut, Beirut, Lebanon
E-mail address: [email protected]
Department of Mathematics, I.I.T. Madras, Chennai, Indian - 600036.
E-mail address: [email protected]
Department of Mathematics, University of Missouri, Columbia, Missouri, USA
E-mail address: [email protected]
| 0math.AC
|
Path Planning With Kinematic Constraints For Robot Groups
arXiv:1704.07538v1 [cs.AI] 25 Apr 2017
Wolfgang Hönig, T. K. Satish Kumar, Liron Cohen, Hang Ma, Sven Koenig, Nora Ayanian
Abstract— Path planning for multiple robots is well studied
in the AI and robotics communities. For a given discretized
environment, robots need to find collision-free paths to a set
of specified goal locations. Robots can be fully anonymous,
non-anonymous, or organized in groups. Although powerful
solvers for this abstract problem exist, they make simplifying assumptions by ignoring kinematic constraints, making it
difficult to use the resulting plans on actual robots. In this
paper, we present a solution which takes kinematic constraints,
such as maximum velocities, into account, while guaranteeing
a user-specified minimum safety distance between robots. We
demonstrate our approach in simulation and on real robots in
2D and 3D environments.
I. INTRODUCTION
Path planning for multiple robots has many applications, including improving traffic at intersections, search and
rescue, formation control, warehouse management, airport
scheduling, and assembly planning. There are two existing
major approaches: The first one works in continuous environments and can take kinematic constraints into account
but does not perform well in highly cluttered, puzzle-like
scenes, and the second one works in discrete environments
with artificial agents without motion constraints.
Hence, it is desirable to combine the two approaches by
providing a planner which can deal with highly cluttered,
puzzle-like scenes even under kinematic constraints. We
tackle this challenge by introducing a postprocessing step
that works on the output of a discrete solver. While the solver
itself is allowed to make simplifying assumptions in order to
run faster, our postprocessing step reinstates the adherence
to real-world kinematic constraints.
Solvers from the AI community for Multi-Agent PathFinding (MAPF) problems and Target-Allocation and PathFinding (TAPF) problems are able to solve instances with
hundreds of agents [1]. For the MAPF problem, each agent
has an assigned start- and goal location. The objective is to
find a set of synchronized paths, one path per agent, such
that each agent reaches its goal location without colliding
with other agents while minimizing the number of actions
required. Each agent can either move to an adjacent location
in one timestep or wait. The TAPF problem is a generalization in which the agents are partitioned into K groups.
A set of goal locations is assinged to each group. A solver
The
Science
authors
are
with
at the University
the
of
Department
of
Computer
Southern California, USA.
[email protected], [email protected],
{hangma,lcohen,skoenig,ayanian}@usc.edu.
Our
research was supported by ARL under grant number W911NF-14-D-0005,
ONR under grant numbers N00014-14-1-0734 and N00014-09-1-1031,
NASA via Stinger Ghaffarian Technologies, and NSF under grant numbers
1409987 and 1319966.
s11
s21
A
B
C
g12
g11
D
E
F
(a) Example environment with two
robots. The lighter circles on the
right mark the goal locations.
Fig. 1.
(b) Graph representation of the example environment.
Running TAPF example with 2 groups.
allocates a specific goal location to each agent and reports a
set of synchronized paths for all agents. If K = 1, all agents
are anonymous, and, if K equals the number of agents, TAPF
is the same as MAPF. Therefore, we will focus on TAPF in
the remainder of this paper.
Using a TAPF solution on real robots has several limitations: (a) robots have kinematic constraints, such as maximum velocities and accelerations; and (b) the generated
solution’s timing is inflexible, necessitating costly replanning
when robots execute the solution imperfectly. A framework
that does not explicitly address faulty execution can cause
undesirable robot-robot collisions or repeated replanning. To
overcome these limitations while still being able to make
use of powerful solvers developed in the AI community,
we propose the use of a postprocessing step based on
the algorithmic framework of Simple Temporal Networks
(STNs).
II. TAPF
We are given an undirected graph G1 = (V1 , E1 )
and K multi-agent groups {group1 , group2 , . . . , groupK },
where groupi consists of Ki interchangeable robots
{ai1 , ai2 , . . . , aiKi } for all i ∈ {1, 2, . . . , K}. Each robot aij
has a unique start location sij ∈ V1 , and the i-th group has a
i
}. A solution
set of unique target locations {g1i , g2i , . . . , gK
i
to the TAPF problem finds K permutations, one for each
group, to uniquely assign a target location to each robot
and a collision-free path for each robot to navigate from its
start location to the assigned goal location. A more rigorous
mathematical description is given in [1].
The makespan is the total time until the last robot reached
its goal location. A solution is optimal if the makespan is
minimal. For the case of a single group, the problem can be
solved in polynomial time. However, in general, it is NPhard to approximate an optimal solution within any constant
factor less than 4/3 [2].
The Conflict-Based Min-Cost-Flow (CBM) algorithm uses
a hierarchical approach, where the lower level uses a maxflow algorithm and the higher level uses a best-first search,
which tries to resolve conflicts as they occur [1]. It has been
A10
B11
C21
D31
E41
B02
C12
F22
C32
D42
Fig. 2.
Temporal plan graph with safety markers.
shown empirically that this method works well in warehouse
domains with dozens of teams and hundreds of agents.
An example of a TAPF instance is shown in Fig. 1. Here,
there are two groups, that each contain a single robot. The
robots are holonomic and move in a discretized environment. One optimal solution of the given TAPF problem is
hA, B, C, D, Ei for one robot and hB, C, F, C, Di for the
other robot.
III. STN
The output of the TAPF solver is a set of synchronized
paths, assuming that robots can traverse the unit-length edges
in unit time. We call the change of location of an agent event.
In order to include kinematic constraints, such as maximum
velocities of the different robots, we use a Temporal Plan
Graph (TPG). The TPG captures the partial order of the
events determined by the TAPF solver. More formally, the
TPG is an acyclic graph G2 = (V2 , E2 ), where each vertex
v ∈ V2 represents an event and each edge hu, vi ∈ E2
indicates that u should be scheduled before v.
In the TAPF case, there are two kinds of temporal precedences. First, the precedence which captures the location visit
order for each robot individually. Second, the TAPF solution
synchronizes the paths between robots to avoid conflicts.
This happens if robot 1 visits a location at timestep t1 and
robot 2 visits the same location at timestep t2 > t1 .
Moreover, we can add safety markers as additional events
to ensure a minimum safety distance between robots at any
time. There are several methods to do so. An LP-solver can
be used to optimize for maximum throughput or minimum
makespan [3]. If the desired safety distance divides the edge
length, the problem can be solved with a user-specified edgelength in strongly polynomial time by using a shortest path
algorithm [4]. One example of a TPG with safety markers is
shown in Fig. 2.
The TPG captures only the necessary partial order on
events. In order to include kinematic constraints such as
maximum velocities for robots or certain parts of the environment, we can extend it to a simple temporal network (STN).
An STN can be encoded as directed graph G3 = (V3 , E3 ),
where V3 = {X0 , X1 , . . . , XN } and E3 are the sets of events
and edges, respectively. Each edge e = hXi , Xj i ∈ E3 has
lower and upper bounds [LB(e), U B(e)], indicating that Xj
has to be scheduled between LB(e) and U B(e) time units
after Xi .
IV. EXPERIMENTS
We implement TAPF and MAPF solvers as well as various
variants of the STN framework in C++. For performance
evaluation, we randomly generate 10 × 10 × 5 maps with
Fig. 3. Examples of the experiments on robots. There are two groups
of four robots each trying to change formation in an environment with
obstacles.
varying numbers of robots, groups, and obstacles and measure how long the TAPF solver takes to find an optimal
solution. For example, a scenario with 150 obstacles and 100
robots in 5 groups can be solved in about 5 s on commodity
hardware. The time the STN requires to compute a solution
varies by method and desired safety distance. Interestingly,
its runtime is smaller in the common case of a large safety
distance, because fewer markers are required in that case.
In a warehouse-like domain with 100 robots, it takes 4 s to
compute a solution.
We verified our approach in simulation using the VREP robotics simulator for differential drive robots, 6-legged
robots, and quadcopters. Furthermore, we performed experiments with up to 8 iRobot Create2 robots, verifying
that actual robots are actually able to follow the computed
trajectories. An example is shown in Fig. 3.
V. CONCLUSION
We presented an approach for using powerful solvers
from the AI community for multi-agent path finding on
actual robots, which obey kinematic constraints. Although
not shown here, our approach is optimal with respect to
the makespan under some conditions. We demonstrate the
applicability of our approach through simulation and on
actual robots. In the future, we would like to test our
framework on physical quadcopters and add an execution
monitoring framework which makes use of the “slack” of
the STN to avoid costly replanning in case of inaccurate
execution.
R EFERENCES
[1] H. Ma and S. Koenig, “Optimal target assignment and path finding for
teams of agents,” in International Conference on Autonomous Agents
and Multiagent Systems, 2016.
[2] H. Ma, C. Tovey, G. Sharon, T. K. S. Kumar, and S. Koenig, “Multiagent path finding with payload transfers and the package-exchange
robot-routing problem,” in AAAI Conference on Artificial Intelligence,
2016.
[3] W. Hönig, T. K. S. Kumar, L. Cohen, H. Ma, H. Xu, N. Ayanian,
and S. Koenig, “Multi-agent path finding with kinematic constraints,”
in International Conference on Automated Planning and Scheduling,
2016.
[4] W. Hönig, T. K. S. Kumar, H. Ma, S. Koenig, and N. Ayanian,
“Formation change for robot groups in occluded environments,” in
IEEE/RSJ International Conference on Intelligent Robots and Systems,
2016.
| 2cs.AI
|
Temporal and Volumetric Denoising via
Quantile Sparse Image (QuaSI) Prior in Optical
Coherence Tomography and Beyond
Franziska Schirrmacher1,∗ , Thomas Köhler1,4,∗ , Tobias Lindenberger1 ,
Lennart Husvogt1 , Jürgen Endres1,3 , James G. Fujimoto2 , Joachim Hornegger1 ,
Arnd Dörfler3 , Philip Hoelter3 , and Andreas K. Maier1
arXiv:1802.03943v1 [cs.CV] 12 Feb 2018
1
2
Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg,
Germany
{franziska.schirrmacher, thomas.koehler}@fau.de
Department of Electrical Engineering & Computer Science and Research Laboratory
of Electronics, Massachusetts Institute of Technology, USA 3 Department of
Neuroradiology, Universitätsklinikum Erlangen, Germany 4 e.solutions GmbH,
Erlangen, Germany
∗
These authors contributed equally to this work.
Abstract. This paper introduces an universal and structure-preserving
regularization term, called quantile sparse image (QuaSI) prior. The prior
is suitable for denoising images from various medical image modalities. We
demonstrate its effectivness on volumetric optical coherence tomography
(OCT) and computed tomography (CT) data, which show differnt noise
and image characteristics. OCT offers high-resolution scans of the human
retina but is inherently impaired by speckle noise. CT on the other
hand has a lower resolution and shows high-frequency noise. For purpose
of denoising, we propose a variational framework based on the QuaSI
prior and a Huber data fidelity model that can handle 3-D and 3-D+t
data. Efficient optimization is fascilitated through use of an alternating
direction method of multipliers (ADMM) scheme and the linearization
of the quantile filter. Experiments on multiple dataset emphasize the
excellent performance of the proposed method.
1
Introduction
Optical coherence tomography (OCT) is a standard non-invasive imaging modality
within clinical workflows in ophthalmology. A fine spatial resolution and the
possibility of 3-D imaging aid the diagnosis and monitoring of eye diseases. In
OCT, low-coherent light passes through an interferometer-like setting and is
reflected back by the object. The amount of back-reflections in a certain axial
depth determines the amplitude of an interference signal at one position. From
the interference signal one can obtain depth-layers of the object. There are two
different types of systems: time domain OCT (TD-OCT) and Fourier domain
OCT (FD-OCT). In TD-OCT, light is decomposed into two parts and propagated
2
Schirrmacher, Köhler et al.
registered B-scans
denoised B-scan
B-scan
denoising
registered volumes
denoised volume
volumetric
denoising
Fig. 1: We propose two pipelines of our spatio-temporal denoising algorithm for
OCT data. In the first pipeline (top), hereinafter called B-scan denoising, single
B-scans or a sequence of registered B-scans are processed. The second pipeline
(bottom) processes volumes as well as a sequence of registered volumes and is
called volumetric denoising.
through a sample and a reference arm. The light is reflected back by a mirror to
assign the reflected light in the sample arm to an axial depth. In FD-OCT, the
Fourier transform enables a spectral resolution of the interference signal, without
scanning the reference arm. Predefined patterns are used to scan an object. At
each position a 1-D A-Scan is obtained encoding the back scatter intensities in
axial depth. A 2-D B-Scan comprises multiple of these A-Scans acquired with
linear or circular scan patterns. Current OCT systems also provide volumetric
data called C-scans. To this end, several parallel B-scans are conducted with
high speed to obtain 3-D volumes.
Photon interference during the acquisition causes speckle noise, which turns
out to be a major drawback of OCT. As a consequence, the quality of the scans
might be drastically reduced. This can limit the reliability of OCT in diagnostic
workflows [12] or affects image analysis for computer-aided diagnostics. Denoising
of OCT to mitigate these limitations emerged as an active research area within
the field of retinal image processing. In general, it can be approached using
customized hardware [21,7], spatial filtering of single scans [19,8], or temporal
merging of multiple consecutive scans [17]. However, denoising is still a challenging
task due to two reasons: 1) OCT measures fine morphological structures and it
would be unacceptable regarding clinical applications to impair these structures.
2) The inferring speckle noise follows a multiplicative model, which violates
the assumption on the noise statistics of most popular denoising methods that
emerged in the image processing community.
In this paper, we propose denoising by postprocessing noisy OCT scans within
a variational framework. As the key contribution, we introduce the class of
quantile sparse image (QuaSI) priors to model the appearance of noise-free OCT
data. Specifically, we propose a median filter based regularizer that is based
Temporal and Volumetric Denoising via QuaSI prior
3
on the QuaSI prior using the 0.5 quantile. This follows the idea that a noisefree scan should be a fixed point of the median filter and we show that this
approach facilitates structure-preserving denoising. To approach the resulting
non-linear and non-convex optimization derived in our framework, we present an
alternating direction method of multipliers (ADMM) scheme. Our algorithm can
handle spatio-temporal denoising by processing either single scans or sequences
of consecutive scans. Furthermore, it enables denoising on a B-scan level or on
a volume level. Thus, it can be adjusted to the clinical needs within a target
application. The proposed method can also be integrated to commercial OCT
systems without modifications on the existing hardware or scanning protocols.
This paper is an extension of our prior work in [25] and makes the following
additional contributions:
– The algorithm as well as the QuaSI prior are extended to process volumetric
OCT data.
– An investigation of the convergence and parameter sensitivity of the algorithm
is conducted.
– An extension of our algorithm is presented to process volumetric data in
C-arm CT imaging.
The remainder of this paper is organized as follows. In Section 2, we review
related work on OCT denoising. Section 3 comprises the objective function of
the energy minimization problem. In Section 4 the QuaSI prior is introduced.
The numerical optimization of our denoising framework is derived in Section 5.
In Section 6, an experimental evaluation of our method on publicly available
benchmark data and clinical OCT scans as well as CT data is reported. Finally,
Section 7 contains our conclusion.
2
Related work
The image-based techniques can be divided into two groups.
2.1
Spatial Denoising Methods
Spatial or single-image denoising have been extensively studied in the image
processing community and various approaches emerged over the past decades.
Local image filters performs smoothing of noisy images possibly in an adaptive
way to preserve image structures [27]. Non-local filtering also exploits the statistics
of similar and repeating patches within images. One representative from this
class is the successful BM3D method [5]. However, these methods have been
mainly designed for for natural images under simplified assumptions like additive
white Gaussian noise, which is inappropriate to describe speckle noise that is
multiplicative in nature. Learning-based denoising, e. g. based on multilayer
neural networks [3], hold the potential to handle speckle noise by learning noise
distributions from training data. However, large-scale training data required for
such methods is barley available for OCT.
4
Schirrmacher, Köhler et al.
ft
Edata
EQuaSI
ET V
Minimize
f t+1
Fig. 2: Method overview: The proposed spatio-temporal denoising algorithm is
based on an energy mininmization formulation with three terms.
Some spatial filters that have been adopted for OCT denoising are the hybrid
median filter, Lee filter, Wiener filter, or wavelet thresholding as investigated
by Ozcan et al. [19]. Global denoising methods for OCT have been introduced
by Salinas et al. [24] using non-linear diffusion and later by Duan et al. [8]
using second order total generalized variation. Wong et al. [30] have proposed
structure-adaptive Bayesian estimation to handle speckle noise. One interesting
approach has been proposed Fang et al. [9], where dictionary learning based on
B-scans with high signal-to-noise ratio (SNR) is used to denoise low SNR B-sans.
Temporal and Volumetric Denoising via QuaSI prior
5
Single-image denoising offers great flexibility in clinical applications of OCT
as few assumption on the the scanning protocol are made. However, the noise
reduction is limited as such methods can utilize single B-scans only.
2.2
Temporal Denoising Methods
Temporal or multi-image denoising methods considers coherence of consecutive
images to improve noise reduction over single-image denoising. Such methods have
been widely investigated for OCT and exploit sets of B-scans that are acquired
sequentially from the same location or nearby positions. A popular approach in
commercial systems is to register multiple of these B-scans and to average the
registered scans to cancel out random noise. Averaging is computationally efficient
but requires many repetitive acquisitions to effectively reduce speckle noise. Mayer
et al. [17] enhances simple averaging based on wavelet decompositions of B-scans
to estimate local image structures and noise. Denoising is conducted in the
wavelet domain by weighted averaging of wavelet coefficients according to the
local image structure. Cheng et al. [4] formulates OCT denoising from multiple
scans as a low-rank matrix completion problem. In [26], Thapa et al. follows a
similar notion and exploits the low-rank property on patch-based level of multiple
B-scans using weighted nuclear norm minimization. Bian et al. [2] have proposed
inter-frame and intra-frame priors for denoising using convex optimization.
All of these multi-image methods have in common that they require multiple
input scans. This increases the overall acquisition time and therefore they might
lead to a higher patient discomfort. Also, they perform denoising on a B-scan level
but ignore coherence of nearby B-scans within volumetric OCT data. If denoising
of entire volumes is desired, simple consecutive processing of individual B-scans
can lead to suboptimal results. In this paper, we mitigate both limitations by
proposing a unified approach to handle denoising on B-scan or volume level based
on single or multiple scans.
3
Background
This section presents the variational framework for denoising volumetric OCT
data. Fig. 1 illustrates the two pipelines of our proposed algorithm, namely
B-scan denoising and volumetric denoising. We denote a volume as a tensor
G̃ ∈ RL×Nx ×Ny composed of L B-scans G̃l , l = 1, . . . , L of size Nx × Ny pixels.
For the sake of convenience, we reshape volumes to vector notation according to
g̃ = (g̃1 , . . . , g̃L )> ∈ RLN , where g̃l ∈ RN , N = Nx Ny denotes a single B-scan
using row-wise scanning of G̃l .
3.1
Noise Model
In OCT, each captured B-scan g̃l of a volume g̃ is related to a respective noisefree scan f̃l in a volume f̃ = (f̃1 , . . . , f̃L )> ∈ RLN according to the multiplicative
model:
g̃l = f̃l ñl ,
(1)
6
Schirrmacher, Köhler et al.
where is the Hadamard (element-wise) product and ñl ∈ RN denotes multiplicative speckle noise. Following prior work on OCT denoising [30,8], we formulate
the multiplicative noise model in (1) in a logarithmic measurement domain. This
leads to the additive noise model:
gl = fl + nl ,
(2)
where gl = log(g̃l ), fl = log(f̃l ) and nl = log(ñl ).
3.2
Energy Minimization Formulation
Given a sequence of K ≥ 1 volumes g(k) with k = 1, . . . , K that are either
captured from the same position or from nearby positions and registered to each
other, we aim at finding a noise-free volume f̂ . We formulate denoising as the
minimization of the objective function:
f̂ = argmin
f
K
X
ρ f − g(k) + λRQuaSI (f ) + µk∇f k1 .
(3)
k=1
The first term in (3) denotes the data fidelity of f w.r.t. input volumes g1 , . . . , gL
followed by two regularization terms. The first regularization term is the proposed
quantile sparse image (QuaSI) prior weighted by λ ≥ 0. The second regularization
term denotes µ ≥ 0 weighted anisotropic total variation (TV), which regularizes
the image gradient ∇f = (∇x f , ∇y f , ∇z f )> . It is worth noting that the general
denoising framework in (3) can handle both noise reduction for entire volumes in
3-D as well as for individual B-scans in 2-D by constraining the domain of both
regularization terms.
In (3), the data fidelity term is customized to the image formation in OCT
using the loss function ρ : RN → R+
0 . In general, the image formation for clinical
OCT needs to consider a mixture of speckle noise, potential misalignments
between the input volumes, or motion artifacts. Following prior work on mixed
noise models in image restoration [11], we propose to use the Huber loss [18]:
ρ(z) =
N
X
φ(zi ),
(4)
i=1
where:
(
φ(z) =
1 2
2z
|z| −
1
2
if z ≤
otherwise,
(5)
and > 0 denotes the threshold of the Huber loss. This leads to an outlierinsensitive model while the underlying data fidelity is a convex term.
4
Quantile Sparse Image (QuaSI) Prior
A robust and efficient regularization term is of importance to achieve results with
a high signal-to-noise ratio (SNR). The better the regularization term is able to
Temporal and Volumetric Denoising via QuaSI prior
2
(a) Noisy B-scan fnoisy
(b) Gold standard B-scan fgold
(c) fnoisy − Q(fnoisy )
(d) fgold − Q(fgold )
·105
2
1.5
1.5
1
1
0.5
0.5
0
0
5 · 10−2
0.1
0.15
0.2
(e) Histogram of fnoisy − Q(fnoisy )
0
7
·105
0
5 · 10−2
0.1
0.15
0.2
(f) Histogram of fgold − Q(fgold )
Fig. 3: Analysis of our proposed QuaSI prior using median filtering Q(·) to model
the appearance of OCT B-scans. (a) and (b) depict a noisy B-scan along the
respective gold standard taken from the pig eye dataset [17]. (c) and (d) show
the residual f − Q(f ) of the QuaSI regularization term, where brighter pixels
express higher residuals (contrast enhanced for visualization). (e) and (f) depict
the corresponding histograms of the both residuals, where the histogram for the
gold standard is sparse. Our QuaSI prior exploits the sparsity of f − Q(f ) for
regularization in our variational denoising framework.
model natural or medical images, the better the result of the optimization. In
OCT denoising structure preservation is a sensitive issue. The B-scans contain
small morphological structures that need to be preserved for the purpose of
diagnosis. In order to tackle the problems referred to above, the so called quantile
sparse image (QuaSI) prior is introduced.
4.1
Definition of the Prior
The QuaSI prior is based on quantile filtering, where the quantile filter is denoted as f̃ = Q(f ). The p-quantile with p ∈ [0, 1] is determined within a local
neighborhood N (i), where d denotes the width of the cubic filter kernel. For
8
Schirrmacher, Köhler et al.
ft
1
2
3
j
1
i
1
current position
position of the
quantile
Move
kernel
Qi,j = 1
Fig. 4: Construction of the binary matrix to approximate the quantile filter
Q(f ) = Qf t .
the i-th pixel in f we filter according to f˜i = quantileN (i) (fi , p). Inspired by the
regularization by denoising priors by [22], the denoised volume is a fixed point
under the quantile filter. In this way:
RQuaSI (f ) = ||f − Q(f )||1 .
(6)
Specifically, regularization according to (6) enforces sparsity of residual f − Q(f ).
This offers a general model for regularization and – depending on the application
– various types of statistics can be chosen for Q(f ). In this paper, we propose the
median filter, where f˜i = medianN (i) (fi ). This follows the rationale that median
filtering facilitates structure-preserving denoising under non-Gaussian noise.
To validate the QuaSI prior using median filter regularization for OCT
denoising, we study its behavior under real measurement noise. For this purpose,
we use the publicly available pig eye dataset [17], which provides a gold standard
B-scan obtained from the average of 455 registered noisy B-scans. We compare
a noisy B-scan fnoisy with the gold standard fgold in Fig. 3a and Fig. 3b. The
residuals f − Q(f ) of the QuaSI regularization term are illustrated in Fig. 3c for
the noisy B-scan and in Fig. 3d for the gold standard. Compared to the gold
standard, the noisy B-scan yields a less sparse signal as shown in the histograms
of both residuals in Fig. 3e and Fig. 3f. Notice that the QuaSI regularization
does not penalize image discontinuities. The histogram using the noisy B-scan
contains less zero elements, while the histogram for the gold standard is sparse.
Our proposed QuaSI prior exploits these observations for structure-preserving
regularization in our variational denoising framework.
Temporal and Volumetric Denoising via QuaSI prior
4.2
9
Linearization
In order to deal with the non-linearity of the quantile operator Q(f ) the linearization Q(f ) = Qf , similar to the work of [20], is performed. The binary matrix Q
is assembled element-wise according to:
(
Qij =
1
0
if j = z,
otherwise,
(7)
where z = arg quantiler∈N (i) fr . This operation filters the i-th pixel according to
the p-quantile in its local neighborhood N (i). For f 0 = f the linearization fullfills
Q(f 0 ) = Qf 0 , while otherwise Q serves as an approximation of the quantile filter.
Fig. 4 illustrates the construction of the binary matrix Q in 2-D. Each pixel
is replaced by the quantile within its local neighborhood. The position of the
quantile is stored in the binary matrix. In this example, the quantile is at position
j. Thus, the i row of the matrix contains a one in the j-th column and zeros
otherwise. The multiplication Qf yields the quantile filtered result.
5
Deploying QuaSI for OCT Denoising
The non-smooth and non-linear QuaSI prior as well as the non-smooth TV prior
impede the optimization of (3) due to the quantile filter and the L1 -Norm. Using
the linearization of the quantile filter and an optimization scheme based on
ADMM [10] enable an efficient minimization. In order to tackle the L1 -Norm,
the auxiliary variables u and v are introduced. This leads to the constraint
optimization problem:
f̂ = argmin
f
K
X
ρ f − g(k) + λkuk1 + µkvk1
k=1
(8)
such that u = f − Q(f ), v = ∇f .
An unconstrained optimization problem is obtained using quadratic penalty
functions according to:
f̂ = argmin
f
K
X
k=1
α
ρ f − g(k) + ku − f + Q(f )k22 + λkuk1
2
(9)
β
+ kv − ∇f k22 + µkvk1 .
2
The Lagrangian multipliers α > 0 and β > 0 enforce the constraints u =
f − Q(f ) and v = ∇fl . If α, β → ∞, we end up at the original problem (3). In
order to strictly enforce the constraint, the Bregman variables bu and bv are
10
Schirrmacher, Köhler et al.
Algorithm 1 Denoising with the QuaSI prior using ADMM optimization
(1)
(K)
Initialize u1 = v1 = 0, b1u = b1v = 0 and fl1 = mean(gl , . . . , gl )
for t = 1, . . . , Touter do
Assemble QuaSI linearization Q from the intermediate estimate f t according to
(7)
for i = 1, . . . , Tinner do
Update weight matrices W(k) , k = 1, . . . , K for IRLS (14)
Update the intermediate estimate f t+1 using CG iterations for (11)
Update the auxiliary variables ut+1 and vt+1 using (15) – (16)
Update the Bregman variables bt+1
and bt+1
using (17) – (18)
v
u
end for
end for
introduced. Then, we minimize the augmented Lagrangian:
LAL (f , u, v, bu , bv ) =
K
X
ρ f − g(k)
k=1
α
+ ku − f + Q(f ) − bu k22 + λkuk1
2
β
+ kv − ∇f − bv k22 + µkvk1 .
2
(10)
We iteratively optimize (10) by alternating minimization w.r.t. the individual
parameters. Hence, three subproblems emerge, where the L1 -Norm is decoupled
from the L2 -Norm. We solve these problems separately in alternating minimization
scheme.
The minimization of the augmented Lagrangian (10) w.r.t. f can be solved
in a least square sense. Therefore, the binary matrix Q is constructed using the
result f t from the previous iteration, where t denotes the iteration index. In
order to cope with the Huber loss, iteratively re-weighted least squares (IRLS) is
applied. Solving the resulting least squares problem leads to the linear system:
Af t+1 = b
A=2
K
X
(W(k) )> + β∇> ∇ + αM> M
(11)
(12)
k=1
b=2
K
X
(W(k) )> g(k) + β∇> (v − bv ) + α(M)> (u − bu ),
(13)
k=1
where M = I − Q and W(k) are diagonal weight matrices constructed from f t .
Using the intermediate result f t , we can computed the weights for IRLS according
to:
(k)
φ0 (fit − gi )
(k)
Wii =
,
(14)
(k)
|fit − gi |
Temporal and Volumetric Denoising via QuaSI prior
11
where φ0 (z) is the derivative of the Huber loss. The threshold of the Huber
loss is set = 1.345σ to achieve 95-percent efficiency of the estimator under
Gaussian noise [18]. We use the mean absolute deviation (MAD) to estimate the
(k)
standard deviation σ = 1.4826mad(fit − gi ) [23]. To solve the linear system
(11), conjugate gradient (CG) iterations are used.
The minimization of the augmented Lagrangian (10) w.r.t. the auxiliary
variables can be done by exploiting the separability of the problem. Given the
estimate for the intermediate result f t+1 , this leads to the element-wise updates:
ut+1
= shrink([f t+1 − Qf t+1 + btu ]i , λ/α),
i
(15)
vit+1 = shrink([∇f t+1 + btv ]i , µ/β).,
(16)
where shrink(z, γ) = sign(z) max(z − γ, 0) denotes the shrinkage operator [10].
Given an estimate for the intermediate result f t+1 as well as the auxiliary
variables ut+1 and vt+1 , the Bregman variables are updated according to:
bt+1
= btu + (f t+1 − Qf t+1 − ut+1 ),
u
(17)
bt+1
v
(18)
=
btv
+ (∇f
t+1
−v
t+1
).
Algorithm 1 summarizes the proposed ADMM based iteration scheme. Overall,
we use two nested optimization loops to solve (8). We use the mean of the input
images as an initial guess f 1 as well as u1 = v1 = 0, b1u = b1v = 0. The weight
matrices for IRLS are updated in every iteration step.
The linearization Q of the quantile filter is updated every Tinner iterations,
assuming the position of the quantile does not change within the next Tinner
iterations. This assumption speeds up the algorithm, as the construction of the
matrix is time-consuming. Note that Tinner should not be choosen too large in
order to avoid a bad approximation of the quantile filter. A proper evaluation of
the convergence of the algorithm is presented in Sect. 6.5.
6
Experiments and Results
In this section, we evaluate our spatio-temporal framework for denoising of retinal
OCT data. Specifically, we quantitatively benchmark our method on different
datasets including comparisons to the state-of-the-art. In addition to OCT, we
also present an application of our QuaSI prior for removing structured noise in
C-arm CT imaging.
Throughout all experiments, we adopted our framework in two versions. For
denoising on a B-scan level, the parameters were set to µ = 0.075 · K, λ = 5.0 · K,
α = 100.0 · K, β = 1.5 · K, Touter = 20 and Tinner = 2 for K B-scans and 3 × 3
median filtering to setup the QuaSI prior. For volumetric denoising based on
6 adjacent B-scans, the parameters were set to µ = 0.0007 · K, λ = 1.0 · K,
α = 120.0 · K, β = 0.05 · K, Touter = 20 and Tinner = 2 for K volumes and
3 × 3 × 3 median filtering.
12
Schirrmacher, Köhler et al.
AVG
BM3D
BED
WMF
Ours
SSIM value
PSNR value
0.8
25
0.6
0.4
20
0.2
5
10
Number of input images
5
10
Number of input images
Fig. 5: Quantification of noise reduction in terms of mean PSNR and SSIM for
different denoising methods on the pig eye dataset for different numbers of input
images.
(a) Noisy input image
(b) AVG
(c) BED [30]
(d) BM3D [5]
(e) WMF [17]
(f) Ours (TV + QuaSI)
Fig. 6: Denoising on the pig eye dataset 9 using K = 5 B-scans. (a) Noisy image,
(b) – (f) AVG, BED [30], BM3D [5], WMF [17], and the proposed method.
6.1
Datasets
To evaluate the performance of the proposed OCT denoising algorithm, we conducted experiments on two different datasets. This comprises ex-vivo benchmark
data as well as real clinical data.
Pig Eye Data For an evaluation of OCT denoising on B-scan level, we used
the publicly available pig eye dataset provided by Mayer et al. [17]. The dataset
comprises 455 B-scans corresponding to 35 eye positions with 13 B-scans per
Temporal and Volumetric Denoising via QuaSI prior
AVG
BM3D
BED
4
2
Ours
6
CNR value
MSR value
6
WMF
13
1
2
3
4
Number of input images
5
4
2
1
2
3
4
Number of input images
5
Fig. 7: Quantification of noise reduction in terms of mean MSR and CNR measures
for denoising on a B-scan level on our clinical dataset for different numbers of
input images.
position and was captured ex-vivo with a Spectralis HRA & OCT. The published
B-scans were registered to each other to compensate for geometric shifts between
them. We apply denoising to sets of K registered B-scans with K ∈ [1, 13] to
demonstrate the influence of different numbers of input B-scans on the denoising
result.
The pig eye dataset provides a gold standard B-scans that was obtained by
averaging all 455 registered scans. The quality of the denoising algorithm was
evaluated by assessing the fidelity of a denoised B-scan w.r.t. to the gold standard
using the peak-signal-to-noise ratio (PSNR) as well as the structural similarity
index (SSIM).
Clinical Data In order to evaluate and compare B-scan with volumetric denoising, we use clinical data. A prototype ultrahigh-speed swept-source OCT system
with 1050 nm wavelength and a sampling rate of 400,000 A-scans per second
[29] was used to acquire volumetric data of 14 human subjects. Proliferative and
non-proliferative diabetic retinopathy, early age-related macular degeneration
and one healthy subject are covered on two volumes per subject, where each
B-scan was acquired five times in immediate succession. We use 500 A-scans by
500 B-scans for a field size of 3×3 mm.
For denoising on a B-scan level the central B-scan of each volume is used,
while volumetric denoising is performed on adjacent B-scans including the central
one. As the clinical data does not provide a gold standard, we follow prior work
[9,19,30] and the measure noise reduction using the mean-to-standard-deviation
14
Schirrmacher, Köhler et al.
(a) Noisy image (MSR: 2.68, CNR: 2.47)
(b) AVG (MSR: 3.17, CNR: 3.17)
(c) BM3D [5] (MSR: 4.61, CNR: 4.85) (d) BED [30] (MSR: 4.67, CNR: 4.85)
(e) WMF [17] (MSR: 3.67, CNR: 3.55)
(f) Ours (MSR: 5.02, CNR: 5.36)
Fig. 8: Visual comparison of denoising results using our clinical dataset with
K = 5 B-scans from a 46 years old male diabetic retinopathy patient. (a) Noisy
image with manually selected background (red) and foreground regions (green)
to determine MSR and CNR. (b) – (f) AVG, BM3D [5], BED [30], WMF [17],
and the proposed method.
ratio (MSR) and the contrast-to-noise ratio (CNR) according to:
µf
σf
|µf − µb |
CNR = q
,
1
(σf2 + σb2 )
2
MSR =
(19)
(20)
where µf and µb as well as σf and σb denote the means and standard deviations
of the intensities in a foreground and a background region, respectively. The
regions to determine MSR and CNR were manually selected for the central B-scan
of each volume, see Fig. 8a.
Temporal and Volumetric Denoising via QuaSI prior
6.2
15
Comparison to the State-of-the-art
We compared our proposed spatio-temporal denoising against four competing
noise reduction approaches. As a representative of general-purpose methods, we
evaluated the popular BM3D [5] that is state-of-the-art in the field of natural
image denoising. In terms of spatial filters customized for OCT, we used Bayesian
estimation denoising (BED) [30]. In the field of temporal methods using multiple
registered B-scans, we evaluate simple averaging (AVG) as a baseline as well as
wavelet multi-frame denoising (WMF) [17]. To ensure fair comparisons between
spatial and temporal methods, we provide the average of all B-scans as input for
single-image denoising (BM3D and BED). In contrast, AVG and WMF are pure
temporal approaches that process multiple registered B-scans. Notice that all of
these methods can only operate on individual 2-D B-scans to denoise volumetric
data and are therefore compared to our proposed method on a B-scan level.
First, we conducted experiments for denoising on a B-scan level on the pig eye
dataset described in Section 6.1. Figure 5 depicts the mean PSRN and SSIM of the
competing denoising methods w.r.t. to the gold standard for different numbers of
input B-scans. We observed quantitatively that our proposed method consistently
outperforms the competing BM3D, BED, and WMF denoising regardless of
the number of input frames. Moreover, using only 2 input B-scans, our spatiotemporal method achieved comparable results to averaging 5 B-scans. This reveals
that our method is more economic regarding the number of required input scans.
This property is essential for clinical applications, where acquiring more scans
might lead to unacceptable long acquisition times. Figure 6 depicts qualitative
denoising results for K = 5 B-scans. Here, the proposed spatio-temporal denoising
algorithm using the QuaSI prior achieved superior performance in terms of
noise reduction, while anatomical structures like retinal layers are preserved.
Second, denoising on a B-scan level was studied on our clinical datasets using the
non-reference MSR and CNR measures for a quantitative evaluation. Figure 7
depicts the averaged MSR and CNR measures for different numbers of input
images. Overall, we observed that BM3D and our proposed method achieved
the best noise reduction expressed by both measures. Figure 8 qualitatively
compares the denoising performance on one example dataset. We found that
AVG, WMF, BED facilitate structure-preserving denoising but were prone to
noise breakthroughs in homogeneous areas, which lowers their MSR and CNR.
In contrast, BM3D achieved superior noise reduction but suffered from streak
artifacts. Similar observations were made in related work on OCT denoising [9]
and can be explained by assumption of additive white Gaussian noise used for
BM3D. The proposed method achieved a decent tradeoff between noise reduction
and structure preservation.
6.3
Impact of the QuaSI Prior
We used the pig eye dataset as well as the clinical data to evaluate the performance
of our spatio-temporal denoising algorithm with and without the QuaSI prior.
Fig. 9 illustrates the impact of the QuaSI prior on the denoising result for the
16
Schirrmacher, Köhler et al.
(a) K = 1
(b) K = 5
(c) K = 13
Fig. 9: This comparison aims at demonstrating the improvement of the proposed
spatio-temporal denoising with TV + QuaSI regularization (third row) compared
to simple averaging of consecutive B-scans (top row) and the proposed spatiotemporal denoising with TV regularization only (second row) for different numbers
of input images.
pig eye data compared to simple averaging and pure TV regularization. In terms
of noise reduction, the proposed variational framework outperformed simple
averaging. Especially in the enlarged region a noticeable difference between
averaging and the proposed denoising algorithm is shown. In homogeneous
areas, the algorithm considerably suppressed speckle noise, while preserving
important structures. The noise reduction was superior when using a combination
of the QuaSI prior and the TV prior for regularization as shown for the retinal
structures in the enlarged region. In addition, the QuaSI prior contributed to
structure-preservation and avoided staircasing artifacts that typically appear in
TV denoising.
Fig. 10 illustrates the impact of the QuaSI prior using PSNR and SSIM (for
the pig eye data) as well as MSR and CNR (for clinical data) for different numbers
of input scans. Here, our denoising framework with QuaSI prior outperformed
TV denoising in terms of all measures.
Temporal and Volumetric Denoising via QuaSI prior
TV
Mean PSNR
TV + QuaSI
Mean SSIM
Mean MSR
Mean CNR
30
1
6
6
15
0.5
3
3
1
5
13
1
17
5
13
1
5
1
5
Fig. 10: Mean PSNR, SSIM, MSR and CNR measures to quantify noise reduction
with and without the QuaSI prior for 1, 5 and 13 input images.
(a) B-scan denoising K = 1
(MSR: 5.68, CNR: 5.25)
(b) volumetric denoising K = 1
(MSR: 6.43, CNR: 5.99)
(c) B-scan denoising K = 5
(MSR: 7.78, CNR: 7.15)
(d) volumetric denoising K = 5
(MSR: 6.85, CNR: 6.51)
Fig. 11: Denoising on the clincial dataset using 5 registered volumes from a 67
years old male mild non-proliferative diabetic retinopathy patient. Denoising is
performed on a B-scan level as well as on a volume level.
6.4
B-scan vs. Volumetric Denoising
So far, we evaluated denoising of volumetric OCT data by simply processing
individual B-scans. In order to evaluate the impact of true volumetric denoising
to simple B-scan wise B-scan denoising in our proposed framework, we used the
clinical dataset presented in Section 6.1. Our volumetric denoising processes 6
18
Schirrmacher, Köhler et al.
K = 1 scan
B-scan
Volumetric
denoising
denoising
K = 5 scans
B-scan
Volumetric
denoising
denoising
MSR
5.35
5.77
6.50
6.31
CNR
5.27
5.60
6.38
6.18
Table 1: Mean MSR and CNR measures for 1 and 5 registered input volumes
on the clinical data. For B-scan denoising, the central B-scan is used and for
volumetric denoising 6 adjacent B-scans including the central one are used.
consecutive B-scans including the central one. That way, CNR and MSR measures
from the previous experiments can be used for comparison. Table 1 shows the
mean MSR and CNR using one and five registered input volumes. Here, we found
that our volumetric denoising achieved better results in terms of noise reduction
for 1 input volume, as adjacent B-scans affect denoising positively. For 5 input
volumes, we found that our B-scan denoising achieved slightly better results in
terms of noise reduction. However, as opposed to noise reduction, volumetric
denoising achieved superior performance in structure preservation by exploiting
coherence between adjacent B-scans. This is depicted in Fig. 11, where the retinal
layers in the magnified region can be better distinguished.
6.5
Convergence and Parameter Sensitivity
The convergence of the proposed algorithm is shown experimentally on a B-scan
level. By our definition, the algorithm converges if a stationary point of the
objective function (3) is reached. The value of the objective, hereinafter referred
to as energy, is computed after every update of the intermediate image f t+1 .
In addition, PSNR and SSIM of the intermediate image are computed. Based
on the optimal parameter setting µ = 0.075 · K, λ = 5.0 · K, α = 100.0 · K,
β = 1.5 · K, Touter = 20, Tinner = 10 and Tcg = 3 for B-scan denoising, we denoise
the pig eye dataset 9 with K = 8 B-scans. Varying one parameter in an otherwise
fixed parameter setting shows the influence of the individual parameter on the
convergence.
Fig. 12 illustrates the influence of Touter , Tinner and Tcg on the convergence,
where Touter · Tinner = 300 for fair comparison. Three different settings are
compared, the optimal parameter setting, a special case where Tinner = 1 and
large Tcg . The approximation of the QuaSI prior is updated every Tinner iterations.
With increasing number of inner iterations, peaks appear in the energy as well
as in the PSNR value and impair the convergence properties of the algorithm.
This is mainly attributable to the rare assembly of Q. Hence, the special case
shows nice convergence properties, as no approximation is necessary. Due to
the high computational costs of the special case, the proposed optimal setting
if preferable. It requires half the computational cost with similar convergence
properties. Furthermore, the PSNR value increases faster. Increasing Tcg also
Temporal and Volumetric Denoising via QuaSI prior
Touter = 30 Tinner = 10 Tcg = 3
Touter = 300 Tinner = 1 Tcg = 1
19
Touter = 30 Tinner = 10 Tcg = 30
0.8
28
PSNR
SSIM
0.7
0.6
26
0.5
0.4
0
100
200
300
Number of inner and outer iterations
0
100
200
300
Number of inner and outer iterations
·105
Energy
1.5
1
0.5
0
100
200
300
Number of inner and outer iterations
Fig. 12: The division of number of iterations on Touter , Tinner and Tcg influences
the convergence behavior of the proposed algorithm. The intermediate energy
and PSNR measures of four different settings are displayed.
λ = 0.05
λ = 0.5
λ =5
λ = 50
0.8
0.6
0.5
0.4
0
100
200
300
Number of inner and outer iterations
28
PSNR
SSIM
0.7
26
0
100
200
300
Number of inner and outer iterations
Fig. 13: Intermediate SSIM and PSNR measures of four different λ values to
illustrate the impact of the QuaSI prior on the convergence behavior of the
algorithm.
induces peaks in the energy and PSNR value. Additionally, the PSNR value
decreases, which is not preferable.
The impact of the QuaSI prior on the denoising process is controlled by λ,
which is illustrated in Fig. 13. With decreasing λ the PSNR and SSIM value
increase slower, due to the low impact of the QuaSI prior. This underlines the
20
Schirrmacher, Köhler et al.
α =1
α = 10
α = 100
α = 1000
0.8
0.6
0.5
0.4
0
100
200
300
Number of inner and outer iterations
28
PSNR
SSIM
0.7
26
0
100
200
300
Number of inner and outer iterations
Fig. 14: Comparison of four different values for the Lagrangian multiplier α and
their influence on the convergence behavior of the proposed algorithm. The
intermediate energy and PSNR measures are displayed.
effectiveness of the proposed QuaSI prior. Increasing λ does not change the
convergence properties of the algorithm.
The Lagrangian multiplier α enforces the constraint u = f − Q(f ). For
α → ∞, the augmented Lagrangian (10) results in the objective function (3).
Hence, decreasing α impairs the convergence properties of the algorithm by
introducing peaks in the SSIM as well as in the PSNR value. PSNR and SSIM
value decrease, which is not preferable. Choosing a large α = 1000 reduces the
peaks but the PSNR value increases slower compared the the proposed parameter
setting. For a large α, the constraint must be satisfied more strictly which results
in higher computational cost due to the slow convergence.
6.6
Quantile Sparse Image Denoising for Volumetric C-Arm Data
Besides the OCT data presented in previous sections, we further applied the QuaSI
framework for denoising cerebral perfusion data acquired using a C-arm computed
tomography (CT) system. This data describes a sequence of volumetric CT images
and is commonly used for stroke assessment by measuring the distribution of
injected contrast agent within the brain over time. Photon effects as well as
patient movement and angular undersampling usually deteriorates the image
quality by introducing both structured and unstructured noise (see Figure 15
a,c).
Current approaches towards noise reduction in CT imaging are, for example,
based on anisotropic filtering or rely on a heuristic detection of streaks and vessel
structures [14,15,16].
Compared to OCT, where the proposed method takes multiple input images
and results in a single output image, denoising of C-arm perfusion data transforms
multiple input images to multiple output images, hereinafter called QuaSI 3D+t. The constraint optimization problem, as given in (8), thus slightly changes
by enforcing the data fidelity for each input volume and by incorporating an
L1 -Norm regularizer in temporal direction for subsequent volumes, resulting in
Temporal and Volumetric Denoising via QuaSI prior
f̂ = argmin ρ(f − g) + λkuk1 + µkvk1 + ωkdt k1 ,
f
21
(21)
where f and g ∈ RN with N = Nx Ny Nz Nt describe a sequence of volumetric
image data. In the last term of equation 21, ω ≥ 0 describes a weighting factor
and dt = ∇t f the gradient in temporal direction, respectively. According to (9),
the constraint optimization problem can thus be formulated as
α
ku − f + Q(f )k22 + λkuk1 +
2
f
β
γ
kv − ∇f k22 + µkvk1 + kdt − ∇t f k22 + ωkdt k1 ,
2
2
f̂ = argmin ρ(f − g) +
(22)
and is solved using the optimization approach given in Section 5.
We applied the QuaSI denoising both on simulated C-arm CT data and on
acquired, real patient data. The simulated data has been created using a digital
brain CT phantom [1] and a realistic scanning protocol (133 projections over an
angular range of 200◦ per volume). We added Poisson noise and introduced minor
patient movement during the generation of the simulated data by moving the head
up to 5◦ around z-axis between the individual scans. After reconstruction, the
individual scans are co-registered again to assert pixel correspondence between
the volumes.
Input BM4D QuaSI 3-D+t
PSNR 32.105 32.485
34.788
SSIM 0.883 0.914
0.943
Table 2: PSNR and SSIM for the input data, BM4D [13] and the QuaSI method.
We compared the results from the proposed method to the results from BM4D
[13], which is an extension of BM3D [6] to volumetric data. For a numerical
comparison of the algorithms we calculate the peak signal to noise ratio (PSNR)
and the structured similarity index measure (SSIM) [28]. We set the parameters
of our method to α = 0.1, λ = 0.0006, β = 0.09, µ = 0.004, γ = 90 and ω = 0.8.
The median filter regularization is computed on a 3 × 3 × 3 kernel.
The algorithms are applied to and evaluated on a subset of the brain volume
consisting of 30 consecutive slices. The slices show the complete head and contain
all structures of interest such as bones, white matter, gray matter and (contrastenhanced) vessels. The results from the evaluation of the realistic brain phantom
show that the proposed denoising algorithm outperforms BM4D with regards to
22
Schirrmacher, Köhler et al.
(a) Noisy input, syn-(b) Denoised output
thetic
(c) Noisy input, clini-(d) Denoised output
cal data
Fig. 15: Input to and result from the QuaSI method. (a) and (c) denote noisy
input to the algorithm, (b) and (d) denote the denoised result. The upper and
lower row shows synthetic and real data, respectively. Note that the input to the
algorithm is not just the slice as shown in the figure, but consists of a sequence
of volumetric data.
PSNR and SSIM, see Table 2. Vessel structures are well-preserved within both
volumes and boundaries between gray and white matter are perceivable.
Fig. 15 shows the input data and results from the QuaSI method for a single
slice out of one of the volume sequence.
7
Conclusion
In this paper, we have presented the quantile sparse image (QuaSI) prior and a
corresponding spatio-temporal denoising algorithm suitable for volumetric OCT
or CT data. For OCT denoising, we proposed two pipelines to either process
B-scans or volumetric OCT data. The numerical optimization is derived using
a linearization of the quantile filter and an alternating direction method of
multipliers scheme for efficient minimization. We can show that a combination of
QuaSI and Total Variation regularization outperformes state-of-the-art methods
in terms of quantitative measures. Interestingly, our method can be applied to
Temporal and Volumetric Denoising via QuaSI prior
23
both CT and OCT data through minor modifications of the denoising pipeline.
This suggestes that it may be worthwhile to evaluate the potential of the QuaSI
prior for inverse problems of other imaging modalities in future work.
References
References
1. Aichert, A., Manhart, M.T., Navalpakkam, B.K., Grimm, R., Hutter, J., Maier,
A.K., Hornegger, J., Doerfler, A.: A realistic digital phantom for perfusion c-arm
CT based on MRI data. In: Nuclear Science Symposium and Medical Imaging
Conference (NSS/MIC), 2013 IEEE. pp. 1–2. IEEE (2013)
2. Bian, L., Suo, J., Chen, F., Dai, Q.: Multiframe denoising of high-speed optical
coherence tomography data using interframe and intraframe priors. Journal of
Biomedical Optics 20, 20 – 20 – 11 (2015)
3. Burger, H.C., Schuler, C.J., Harmeling, S.: Image denoising: Can plain neural
networks compete with bm3d? In: IEEE Conference on Computer Vision and
Pattern Recognition. pp. 2392–2399 (June 2012)
4. Cheng, J., Duan, L., Wong, D.W.K., Tao, D., Akiba, M., Liu, J.: Speckle Reduction
in Optical Coherence Tomography by Image Registration and Matrix Completion.
In: MICCAI 2014, Proceedings, Part I. pp. 162–169 (2014)
5. Dabov, K., Foi, A., Kathovnik, V., Egiazarian, K.: Image denoising by sparse
3-d transform-domain collaborative filtering. IEEE Trans. Image Process. 16(8),
2080–2095 (2007)
6. Dabov, K., Foi, A., Katkovnik, V., Egiazarian, K.: Image denoising by sparse
3-d transform-domain collaborative filtering. IEEE Trans. Image Process. 16(8),
2080–2095 (2007)
7. Desjardins, A.E., Vakoc, B.J., Oh, W.Y., Motaghiannezam, S.M., Tearney, G.J.,
Bouma, B.E.: Angle-resolved optical coherence tomography with sequential angular
selectivity for speckle reduction. Opt. Express 15(10), 6200 (2007)
8. Duan, J., Lu, W., Tench, C., Gottlob, I., Proudlock, F., Samani, N.N., Bai, L.: Denoising optical coherence tomography using second order total generalized variation
decomposition. Biomed Signal Process Control 24, 120–127 (2016)
9. Fang, L., Li, S., Nie, Q., Izatt, J.A., Toth, C.A., Farsiu, S.: Sparsity based denoising
of spectral domain optical coherence tomography images. Biomed. Opt. Express
3(5), 927–942 (2012)
10. Goldstein, T., Osher, S.: The split bregman method for l1-regularized problems.
SIAM J. Imaging Sci. 2(2), 323–343 (2009)
11. Köhler, T., Huang, X., Schebesch, F., Aichert, A., Maier, A., Hornegger, J.: Robust
Multiframe Super-Resolution Employing Iteratively Re-Weighted Minimization.
IEEE Transactions on Computational Imaging 2(1), 42 – 58 (Mar 2016)
12. Kraus, M.: Motion Correction and Signal Enhancement in Optical Coherence
Tomography. Ph.D. thesis, Friedrich-Alexander-Universität Erlangen-Nürnberg
(FAU) (2016)
13. Maggioni, M., Katkovnik, V., Egiazarian, K., Foi, A.: Nonlocal transform-domain
filter for volumetric data denoising and reconstruction. IEEE Trans. Image Process.
22(1), 119–133 (2013)
24
Schirrmacher, Köhler et al.
14. Maier, A., Fahrig, R.: GPU Denoising for Computed Tomography, vol. 1. CRC
Press, Boca Raton, Florida, USA, 1 edn. (2015), http://www.crcnetbase.com/
doi/10.1201/b18968-9
15. Maier, A., Wigström, L., Hofmann, H., Hornegger, J., Zhu, L., Strobel, N.,
Fahrig, R.: Three-dimensional anisotropic adaptive filtering of projection data
for noise reduction in cone beam CT. Medical Physics 38(11), 5896–5909
(2011), http://www5.informatik.uni-erlangen.de/Forschung/Publikationen/
2011/Maier11-TAA.pdf
16. Manhart, M., Fahrig, R., Hornegger, J., Dörfler, A., Maier, A.: Guided
Noise Reduction for Spectral CT with Energy-Selective Photon Counting Detectors. In: Noo, F. (ed.) Proceedings of the Third CT Meeting. pp. 91–94 (2014), https://www5.informatik.uni-erlangen.de/Forschung/
Publikationen/2014/Manhart14-GNR.pdf
17. Mayer, M.A., Borsdorf, A., Wagner, M., Mardin, C.Y., Tornow, R.P.: Wavelet
denoising of multiframe optical coherence tomography. Biomed Opt Express 3(3),
2156–7085 (2012)
18. Ochs, P., Dosovitskiy, A., Brox, T., Pock, T.: On iteratively reweighted algorithms
for nonsmooth nonconvex optimization in computer vision. SIAM J. Imaging Sci.
8(1), 331–372 (2015)
19. Ozcan, A., Bilenca, A., Desjardins, A.E., Bouma, B.E., Tearney, G.J.: Speckle
reduction in optical coherence tomography images using digital filtering. Journal of
the Optical Society of America A 24(7), 1901 (2007)
20. Pan, J., Sun, D., Pfister, H., Yang, M.H.: Blind image deblurring using dark channel
prior. In: CVPR, Proceedings. pp. 1628–1636 (2016)
21. Pircher, M., Gotzinger, E., Leitgeb, R., Fercher, A.F., Hitzenberger, C.K.: Speckle reduction in optical coherence tomography by frequency compounding. J. Biomed. Opt.
8(3), 565 (2003)
22. Romano, Y., Elad, M., Milanfar, P.: The Little Engine that Could: Regularization
by Denoising (RED). arXiv preprint arXiv:1611.02862 (2016)
23. Rousseeuw, P.J., Leroy, A.M.: Robust Regression and Outlier Detection. John
Wiley & Sons (1987)
24. Salinas, H.M., Fernandez, D.C.: Comparison of PDE-Based Nonlinear Diffusion Approaches for Image Enhancement and Denoising in Optical Coherence Tomography.
IEEE Trans. Med. Imag. 26(6), 761–771 (2007)
25. Schirrmacher, F., Köhler, T., Husvogt, L., Fujimoto, J.G., Hornegger, J., Maier,
A.K.: Quasi: Quantile sparse image prior for spatio-temporal denoising of retinal
OCT data. In: MICCAI 2017, Proceedings, Part II. pp. 83–91 (2017)
26. Thapa, D., Raahemifar, K., Lakshminarayanan, V.: Reduction of speckle noise from
optical coherence tomography images using multi-frame weighted nuclear norm
minimization method. Journal of Modern Optics 62(21), 1856–1864 (2015)
27. Tomasi, C., Manduchi, R.: Bilateral Filtering for Gray and Color Images. In:
International Conference on Computer Vision (ICCV). pp. 839–846. IEEE, Bombay,
India (Jan 1998)
28. Wang, Z., Bovik, A.C., Sheikh, H.R., Simoncelli, E.P.: Image quality assessment:
from error visibility to structural similarity. IEEE Trans. Image Process. 13(4),
600–612 (2004)
29. W.Choi, Potsaid, B., Jayaraman, V., Baumann, B., Grulkowski, I., Liu, J.J., Lu,
C.D., Cable, A.E., Huang, D., Duker, J.S., Fujimoto, J.G.: Phase-sensitive sweptsource optical coherence tomography imaging of the human retina with a vertical
cavity surface-emitting laser light source. Opt Lett 38(3), 338 (2013)
Temporal and Volumetric Denoising via QuaSI prior
25
30. Wong, A., Mishra, A., Bizheva, K., Clausi, D.A.: General bayesian estimation for
speckle noise reduction in optical coherence tomography reinal imagery. Opt. Express
18(8), 8338–8352 (2010)
| 1cs.CV
|
This paper was presented at the 2017 51st Asilomar Conference
on Signals, Systems and Computers, Pacific Grove, CA, 2017.
Copyright (c) 2017 IEEE.
milliProxy: a TCP Proxy Architecture
for 5G mmWave Cellular Systems
Michele Polese∗ , Marco Mezzavilla† , Menglei Zhang† ,
Jing Zhu , Sundeep Rangan† , Shivendra Panwar† , Michele Zorzi∗
∗
arXiv:1712.02700v1 [cs.NI] 7 Dec 2017
†
Department of Information Engineering, University of Padova, Italy - e-mail: {polesemi, zorzi}@dei.unipd.it
NYU Wireless, New York University, NY, USA - e-mail: {menglei, mezzavilla, srangan}@nyu.edu, [email protected]
Intel Corporation - e-mail: [email protected]
Abstract—TCP is the most widely used transport protocol in
the internet. However, it offers suboptimal performance when
operating over high bandwidth mmWave links. The main issues
introduced by communications at such high frequencies are (i) the
sensitivity to blockage and (ii) the high bandwidth fluctuations due
to Line of Sight (LOS) to Non Line of Sight (NLOS) transitions
and vice versa. In particular, TCP has an abstract view of the endto-end connection, which does not properly capture the dynamics
of the wireless mmWave link. The consequence is a suboptimal
utilization of the available resources. In this paper we propose a
TCP proxy architecture that improves the performance of TCP
flows without any modification at the remote sender side. The
proxy is installed in the Radio Access Network, and exploits
information available at the Next Generation Node Base (gNB)
in order to maximize throughput and minimize latency.
I. I NTRODUCTION
Communication at mmWave frequencies represents the new
frontier for wireless networks [1]. Nowadays, indeed, most
wireless standards are constrained in a very small bandwidth
which can be allocated in the spectrum below 6 GHz. The
mmWave spectrum (and in general frequencies above 6 GHz)
would allow wireless communications to benefit from wide
chunks of untapped spectrum that could be used to reach very
high data rates.
Millimeter waves are being considered for the next generation of cellular networks (i.e., 5G), which is currently being
standardized by 3GPP as New Radio (NR) [2], as well as for
public safety [3] and vehicular communications [4], [5]. Several
applications in these scenarios require high data rate, and with
low latency: consider for example real-time video streaming
for remote interaction with autonomous robots, virtual reality
applications or raw sensor data (e.g., from RADAR or LIDAR)
exchanges between vehicles. However, the mmWave technology offers both a great potential and a number of challenges,
which are related to the harsh propagation conditions that
must be addressed before mmWaves can be reliably deployed.
The very high path loss can be overcome with beamforming
techniques, which in turn require the design of protocols
that account for directionality [6]. Moreover, mmWaves are
affected by blockage from obstacles (e.g., buildings, trees, the
human body itself) [7], and this has an impact on the service
availability and the achievable data rate.
These limitations have an impact not only on the protocol
stack of the wireless link but also on higher layers protocols,
such as TCP, which is the most widely used transport protocol
in the internet, for a variety of different applications that
require flow control and reliable transmissions. TCP however
suffers in terms of both reduced throughput and high latency
when deployed over an end-to-end connection whose last link
is based on mmWave radios [8], and it is not capable of
delivering the high data rate and low latency performance that
the aforementioned applications require [9].
In this paper we propose the design of milliProxy, a novel
TCP proxy for mmWave mobile networks aimed at fully reaping the benefits of mmWave links to achieve high throughput
with low latency. It is transparent to the end hosts of the
connection, and respects the end-to-end connection semantics.
The main rationale is to split the TCP control loop in the
mmWave RAN to optimize the flow control over the wireless
link. It is based on a cross-layer, data driven approach and
enables a number of optimizations for the operation of TCP in
mmWave networks.
The rest of the paper is organized as follows. In Sec. II, we
extensively discuss the main limitations of TCP over mmWave
links, and provide an overview of the literature related to
TCP proxies for traditional wireless networks. The architecture
of milliProxy is described in Sec. III, and the results of a
performance evaluation campaign are reported in Sec. IV.
Finally, conclusions and future extensions of this work are
provided in Sec. V.
II. R ELATED W ORK
A. 5G mmWave networks and challenges for TCP
In this section, we provide an overview of the main challenges related to the usage of TCP in mmWave cellular networks. They can potentially enable gigabit-per-second cell data
rates thanks to the very large bandwidth available [1], but the
deployment of a reliable 5G mmWave network is challenging,
mainly because of the harsh propagation environment at these
frequencies. The high propagation loss can be compensated
using beamforming techniques, with a large number of antennas
in the transceivers, but this requires the design of Physical
(PHY) and Medium Access Control (MAC) layer protocols able
to cope with directionality [6]. Moreover, mmWave frequencies
are sensitive to blockage from a wide range of materials [7],
and, while it is possible to transmit and receive data also in
NLOS exploiting reflections, there is a massive difference in
the achivable data rate and reliability when transitioning from
LOS to NLOS (with differences in the Signal to Interference
plus Noise Ratio (SINR) in the order of 30 dB) and vice versa.
Therefore, it is challenging to completely benefit from the
vast amount of resources, especially when taking into account
the end-to-end performance and the complex interaction between higher-layer protocols and the mmWave stack, as shown
in previous works [8], [10]–[14]. In particular, the main issues
related to the most widely used transport protocol in end-to-end
networks (i.e., TCP) are related to the slow reactiveness of TCP
with respect to changes in the channel quality of the mmWave
link [8], [10], [12]. This translates into a sub-optimal utilization
of the available resources, since the ramp up of the congestion
window of the most used congestion control algorithms (e.g.,
TCP CUBIC [15] and TCP NewReno [16]) is too slow and
limits the achievable goodput. This issue is more marked in
mmWave links than in traditional sub-6 GHz deployments,
because of an orders-of-magnitude difference in bandwidth
(and thus achievable rate), and it becomes more relevant as
the Round Trip Time (RTT) of the connection increases, as
shown in [12]. Another consequence is the emergence of
the bufferbloat phenomenon [8]. When the channel condition
changes from LOS to NLOS, there is a drop in the data rate
offered by the physical layer. This is a problem because the
packets in excess are buffered at the Radio Link Control (RLC)
layer, and the TCP sender is not aware of the update until
an Active Queue Management (AQM) mechanism or a buffer
overflow drops one or more packets. This causes an increase
in the buffer occupancy and, consequently, in the end-to-end
latency. Finally, there is the possibility of extended outages,
due to blockage and the lack of a fall-back link. TCP reacts
to these events with retransmission timeouts, and halves the
slow start threshold at each of these events. Then, when the
connection resumes, the duration of the slow start phase (with
the exponential increase of the congestion window) is limited,
and the TCP sender stays in congestion avoidance (with a linear
increase of the congestion window) for most of the time, thus
exacerbating the issue related to the slow congestion window
ramp up.
B. TCP Performance Enhancing Proxies
The performance of TCP on wireless networks has been
under the spotlight since the 1990s, when the first cellular
networks capable of data transmission were commercially deployed. Even though TCP faces more challenging conditions
when running on top of mmWave cellular networks, it is
worth describing the main approaches that can be found in
the literature related to the enhancement of TCP performance
on wireless links.
A first comprehensive review on the topic can be found in
a paper by Balakrishnan et al. [17]. The authors claim that
the poor performance of TCP in mobile networks is due to
packet losses over an unreliable channel. However, as shown
in [11], the channel losses can be masked by retransmission
mechanisms. Moreover, the considered links have very low data
rate and small buffers are used in the network. The settings in
a mmWave networks are very different, since large buffers and
retransmissions are already implemented in the wireless link
to make up for packet loss at the price of increased latency
and exposing more the network to the bufferbloat phenomenon.
However, the authors of [17] provide a comparison of different
strategies that can possibly be adapted to mmWave networks,
using TCP Reno as a baseline, and including also TCP split
approaches.
In a more recent paper [18], Liu et al. introduce a TCP proxy
middlebox for the optimization of TCP performance without
the need for any modification to the protocol stack of servers,
clients and base stations. They observe that the adoption of
a new end-to-end TCP congestion control mechanism may be
useless in the presence of HTTP proxies, which are frequently
used in mobile networks. Moreover, they design their solution
for modern LTE networks, characterized by large buffers (in
the order of 5 MB) and bandwidth fluctuations (even if not
as wide as those in mmWave networks [1], [9]), and a fixed
network which does not act as a bottleneck. Their solution is a
middlebox that can be placed anywhere in the mobile operator
core network, and breaks the TCP connection in two segments,
i.e., it does not respect end-to-end connection semantics1 . This
box performs some optimizations on the fly, such as (i) not
using the information of the receiver congestion window, which
may be too small with respect to the actual rate available on
the link, given that experimental evaluations on the receiver
buffer in real devices have highlighted that it is never filled;
(ii) changing retransmission patterns by intercepting duplicate
ACKs; (iii) tuning the congestion window with a rate estimation
algorithm. In this design, the TCP connection from the sender
to the receiver is terminated at the middlebox, which buffers
the packets for the final receiver until it can forward them.
A third approach is described in [21], where Ren et al.
introduce a TCP proxy in the mobile network base station. This
study, however, is focused on the UMTS architecture. Their
approach is based on a queue control mechanism: by using
the sliding mode variable structure (SMVS) control theory
the buffer queue length at the base station is kept at the
same size. This proxy does not respect the end-to-end TCP
semantics, because it terminates the connection at the proxy.
The advertised window at the proxy is used to limit the sending
rate of the server and to avoid buffering delays. At the proxy,
a control mechanism is used to keep the queue length at a
reference value, by inferring the bandwidth available at the base
station.
Some other interesting approaches that respect the end-toend TCP semantics are (i) Mobile TCP (M-TCP) [22], which
freezes the TCP sender when it senses imminent congestion,
in order to avoid packet loss and connection timeouts; and (ii)
Snoop [23], also from Balakrishnan et al., which performs local
retransmissions when TCP packet losses are sensed, in order
to improve the connection reactiveness. I-TCP [24], instead, is
a TCP split approach, not compliant with the end-to-end TCP
semantics, that uses a traditional TCP congestion control also
on the wireless link and does not yield a great performance
improvement.
In [25], a performance enhancement proxy for mmWave
1 According to [19], [20], the end-to-end principle, which states that certain
functions in the internet are designed to work at the end hosts, is a founding
paradigm of the internet. A proxy that splits the TCP connection into two
independent segments does not respect the end-to-end TCP semantics, meaning
that ACKs may be sent to the sender before the packet is actually received by
the other end host.
cellular networks is proposed. It is installed in the base stations,
and breaks the end-to-end TCP semantics by sending early
ACKs to the server. Moreover it performs batch retransmissions, i.e., it retransmits the packets that were detected as lost
as well as the segments with a sequence number which is close
to that of the lost packets. The detection of the lost packets
however assumes that on the link only Hybrid Automatic
Repeat reQuest (HARQ) retransmissions are performed, while
in general with TCP the RLC Acknowledged Mode (AM) is
also used. In the performance evaluation, moreover, the authors
of [25] limit the application data rate to 100 Mbit/s, which can
be usually sustained also in NLOS. Therefore, the performance
analysis does not account for the very high data rates that can
be achieved with mmWave and for the wide rate variations
of the LOS to NLOS (or vice versa) transitions. Finally, it
focuses only on the throughput and delivery ratio for a single
user, without considering the latency and thus the bufferbloat
problem.
III. E ND - TO - END P ROXY A RCHITECTURE FOR MM WAVE
In this section we describe our TCP proxy architecture for
mmWaves, called milliProxy, and highlight the main innovations with respect to the solutions reported in Sec. II. Importantly, by being transparent to both the end points of the TCP
flow, milliProxy respects the end-to-end semantics of the TCP
connection, as opposed to most of the proposed approaches
cited in Sec. II. The key functionalities of milliProxy are (1) the
ability to split the control loop of the connection with a different
and tunable Flow Window (FW) policy at the source server
and at the proxy, as well as (2) the capability of controlling
the Maximum Segment Size (MSS) of the connection in the
portion between the proxy and the User Equipment (UE).
A. Proxy Architecture
MilliProxy is a TCP proxy which can be implemented and
deployed as a network function, composed of several modules
that can be updated or changed. It can be placed in the
gNB, fully benefiting from the interaction with the mmWave
protocol stack, or in a node in the core network, sharing outof-band information with the gNB to which the TCP receiver
is connected. According to the position of the proxy, there
may be the need to design a mechanism to cope with the
user mobility. For example, if the proxy is in the gNB, when
the UE performs a handover the network has to transfer the
milliProxy’s state from the source to the target gNB. If instead
the proxy is in an edge node of the core network, then it can
manage multiple cells without the need to forward the state for
each UE handover. Additional considerations on this issue are
left for future work.
The basic structure of the proxy is shown in Fig. 1. An
instance of the proxy is created for each TCP flow which
goes through the node in which it is installed, so that different
policies can be enabled for different users, or different flows of
the same user. Each instance has its own customizable buffer
(set by default to 10 MB), flow window module and ACK
management unit. The buffer is used to store the payload of the
TCP packets before they can be delivered to the TCP receiver,
and the ACK management unit checks for incoming ACKs to
milliProxy instance
milliProxy instance
Flow window
management
Flow module
window
Flow
management module
BufferFlow
Buffer ACK management
module
ACK
management
Server
module
UE
End-to-end flows
UE
Server
Figure 1: Architecture of milliProxy
clear the contents of the buffer. The flow window policy is the
equivalent at the proxy of the congestion window mechanism
at the TCP sender, i.e., it controls the amount of data that
can be forwarded by the proxy. The policy is not hard-coded
into the proxy, but is loaded as a module, according to the
implementation of the TCP congestion control mechanisms in
the Linux kernel [26]. The ACK management unit of the proxy
modifies the advertised window in each ACK that is relayed to
the server in order to enforce the proxy flow window value also
at the TCP sender. According to [27], the TCP sender selects
the minimum between its congestion window and the received
value of the advertised window as the maximum number of
bytes it can send. Similarly to [10], the advertised window
in the modified ACKs is set to be equal to the flow window
determined at the proxy. This makes it possible to capture
both components of the network, and adapt accordingly: the
wired part is regulated by the classical TCP congestion control
selected, while the wireless channel is used in a cross-layer
fashion by the proxy, which selects the proper value of the
advertised window.
The presence of the buffer makes it possible to tune the
MSS of the connection between the proxy and the UE differently from that of the other part of the connection, enabling
further optimizations. If the MSS of the overall connection is
limited by the Maximum Transmission Unit (MTU) of some
intermediate networks using ethernet as link layer technology
(i.e., the MSS is at most 1460 bytes), then the proxy buffers
the 1460-byte payloads, and can send a larger segment which
aggregates multiple payloads of the end-to-end connection. For
example, fourteen 1460-bytes payloads received back-to-back
in a small time interval can be combined into a single 20440bytes segment which is sent from the proxy to the UE. This
increases the efficiency of the transmission process in the last
mile of the connection, i.e., in the mmWave wireless link,
because of the smaller overhead of the TCP/IP headers (in
the previous example, just one TCP/IP header is used instead
of fourteen), and because fewer uplink resources have to be
scheduled for the transmission of ACKs from the UE [28].
Notice that aggregation is generally performed also at the RLC
and MAC layers of very high-bandwidth connections in order to
improve the transmission efficiency [29], [30], and the larger
MSS helps also this process, since fewer concatenation and
Update FW
Data packet
IN
Payload (MSS 1)
Header
Flow
Window
Option
Processing
Header
Flow
Buffer
RTT Estimate
Window scaling
Data packet
OUT
Payload (MSS 2)
max TX SN
max acked SN
ACKs OUT
ACK
management
module
ACK IN
Figure 2: Packets processing in a milliProxy instance.
segmentation operations are required at the transmitter and the
receiver.
Fig. 2 depicts how a packet is processed by milliProxy. By
design, it is completely transparent to the UE, i.e., the TCP
receiver. It intercepts all the packets belonging to the flows it
is handling, and the payload of data packets is stored in the
proxy buffer. Any options in the packet header are processed,
for example to estimate the RTT, as will be described in
the following sections, or to handle the advertised window
scaling. The payload will then be sent as part of a larger
segment as soon as the flow window allows it. When an
ACK is received, the proxy checks its sequence number, and
marks the corresponding bytes in the buffer as received, which
will then be discarded, allowing the flow window to advance.
Consequently, a number of ACKs corresponding to the number
of original packets received (approximatively equal to the ratio
between the MSS of the proxy-UE connection and that of the
server-proxy connection) is sent to the TCP sender. In each
ACK the advertised window value is overwritten with the value
of the flow window in the proxy.
B. RTT estimation
The estimation of the RTT can be performed using the TCP
timestamp option [31]. This option is symmetric, i.e., it is added
both to data packets at the TCP sender and to ACKs at the
receiver. It has a total length of 10 bytes, and contains two
timestamps. The first (T Sval ) is that of the clock of the end
host that transmits the packet, the second (T Secho ) is the T Sval
of a recently received packet from the other end host. Its usage
is advised in [31] in order to improve the TCP performance,
in terms of both throughput and security.
If both the end hosts share the same clock, the estimation of
the RTT is composed by two phases as follows. In the first one,
which is shown in Fig. 3, the milliProxy instance estimates the
latency on the path from the UE to the server. The timestamp
T Secho of the data packet sent from the server to the UE
corresponds to the time t−1 at which the UE sent an ACK.
Similarly, the timestamp T Sval in the same packet corresponds
to the time instant t0 at which the server transmitted the
data packet. Given the very high packet rate that is sustained
in mmWave networks, it is unlikely to observe a significant
time interval between receiving the ACK corresponding to
T Secho and sending the data packet corresponding to T Sval .
!/*0
!+ = %&,-.
Data
milliProxy
ACK
!"# = %&'()*
Figure 3: RTT computation at the proxy.
Therefore, the latency of the uplink path can be estimated as
TUE→server = t0 − t−1 . In a similar fashion, it is possible to
use the timestamp values carried by ACK packets to estimate
the latency on the downlink path Tserver→UE . Finally, the RTT
is estimated as RT Te = Tserver→UE + TUE→server .
If instead the two end hosts do not have the same clock, or
if the TCP timestamp option is not supported, other methods
can be used to estimate the RTT as reported in [32].
C. Integration with the 5G protocol stack
The proxy is configured to collect some statistics from the
connected 5G gNB. According to the location of the proxy,
this data collection can be performed with or without delay.
If the proxy is installed in the gNB, the information can be
retrieved instantaneously, whereas if it resides in a node in the
core or edge network some signaling is necessary, which would
introduce some incremental latency. Thanks to this information
it is possible to enable a cross-layer approach, which is useful
for the design of flow window management algorithms driven
by the performance and the statistics of the mmWave link.
More information associated with each user can be retrieved
from the protocol stack of the gNB. The first is the RLC buffer
occupancy B, which can be seen as a signal of a congestion
event and a consequent increase in latency. The second is an
estimate of the PHY layer data rate between the UE and the
gNB. In [18] this is done by measuring the number of bytes
transmitted in the previous slots, dividing it by the duration of
the slots. This approach, however, is sensitive to the actual rate
that is injected in the network by the TCP source, and can lead
to an underestimation of the available rate if the source rate
does not saturate the connection. This limitation is particularly
relevant in mmWave networks, where it takes a long time for
the TCP source to reach a full utilization of the available
D. Window Management
The management of the flow window is an essential component of milliProxy. In this paper we propose a scheme based
on the computation of the Bandwidth-Delay Product (BDP).
The implementation and testing of alternative FW management
policies is left for future works.
In the BDP-based scheme, the FW management module
uses three different kinds of cross-layer data: the RLC buffer
occupancy B, the estimated data rate Re and the estimated
RTT RT Te . The latter is filtered, and the minimum value
RT Tmin is selected following the approach in [10], [33]–[35],
so that the queuing latency at the RLC layer or at intermediate
buffers is not taken into account. The flow window is then
computed as w = bRT Tmin Re c. When the RTT estimate is
not yet available (i.e., for the first ACK after the reception
of the SYN packet), the flow window is arbitrarily initialized
to a high value of 400 MB. Moreover, it is possible to make
the policy more conservative when the RLC buffer occupancy
exceeds a predefined value (e.g., 2 MB). In this case, the flow
window is set to w = max{bRT Tmin Re c − 2B, 0}.
IV. P ERFORMANCE E VALUATION
A. ns-3 mmWave module
In order to evaluate the performance of milliProxy in an endto-end scenario, we implemented the proxy in ns-3 [36], an
open source network simulator which also features a mmWave
cellular protocol stack. A complete description of the mmWave
module is provided in [37], [38]: it features a complete 3GPPlike protocol stack in the gNBs and UEs, with a custom
PHY/MAC layers implementation, based on a dynamic Time
Division Duplexing (TDD) scheme designed for low latency
communications [39], RLC, Packet Data Convergence Protocol
(PDCP) and Radio Resource Control (RRC) layers. The channel model is based on the 3GPP channel model for frequencies
above 6 GHz [40], and it models the time correlation among
the channel impulse responses computed for moving users in
order to account for spatial consistency.
B. Scenario and parameters
The main simulation parameters are reported in Table I. In
this paper we focus on testing the performance of milliProxy in
a single user scenario, in order to evaluate the responsiveness
of the proxy architecture to channel variations, from LOS to
NLOS and viceversa. In order to model them, some obstacles
are randomly deployed in the simulation scenario between the
100
mmWave
gNB
80
Y [m]
resources. In our previous works [10], [33], instead, we rely
on the information provided by the Adaptive Modulation and
Coding (AMC) module at the MAC layer. By knowing the
channel quality of a UE it is possible to compute the modulation
and coding scheme, predict how many bytes the scheduler could
allocate to the user (with full buffer assumption) in the next
time slot, and divide by its duration to obtain an achievable
data rate Re that is not influenced by the source rate. Another
useful metric that can be acquired in a cross-layer setup is the
SINR of the UE, which could give an indication on the link
status: for example, if it is below a certain threshold, then the
proxy will know that the UE is in outage.
60
40
20
0
UE path at
speed v
UE
0
10
20
30
X [m]
40
50
Figure 4: Randomly generated simulation scenario. The three grey rectangles
represent obstacles such as buildings, cars, trees.
Parameter
Value
mmWave carrier frequency
mmWave bandwidth
3GPP Channel Scenario
Max PHY layer rate
S1 link latency DS1
Latency from PGW to server DRS
RLC AM buffer size BRLC
RLC AM Reordering Timer
RLC AM Report Buffer Status timer
UE speed v
TCP MSS1 (server - proxy)
TCP MSS2 (proxy - UE)
28 GHz
1 GHz
Urban Micro
3.2 Gbit/s
1 ms
[1, 5, 10, 20] ms
[10, 20] MB
1 ms
2 ms
5 m/s
1400 byte
20000 byte
Table I: Simulation parameters
gNB (which is at coordinates (25, 100) m) and the UE (moving
from (0, 0) m to (50, 0) at speed v). As the user moves, it will
experience multiple transitions, with a random duration of each
LOS or NLOS phase in each different run of the simulation.
An example of scenario is shown in Fig. 4. All the results are
averaged over 50 independent simulation runs.
C. Results
Fig. 5 shows a comparison of both goodput and RAN latency
when milliProxy is deployed in the gNB or not, for different
RLC buffer sizes BRLC and fixed-network latencies. It can be
seen that milliProxy performs better in terms of both goodput
and latency, with a goodput gain of up to 2.24 times (combined
with a latency reduction of 1.98 times) with the highest DRS ,
or a latency reduction of 43 times with a similar goodput in the
edge server scenario (i.e., DRS = 1 ms). MilliProxy is therefore
effective at reducing the impact of the bufferbloat issue: when
the channel switches from a LOS to a NLOS state, milliProxy
can reduce the TCP sending rate faster, and thereby avoid
extra queuing latency. On the other hand, when the channel
quality improves, milliProxy is able to (i) track the available
data rate at the physical layer and (ii) promptly inform the
TCP sender of the increased resource availability, which indeed
results in higher goodput. The performance of milliProxy is
independent on the buffer size, since it manages to keep the
buffer occupancy and consequently the RLC queuing delay to a
minimum. As shown in Fig. 5 and extensively discussed in [28],
TCP NewReno, B = 10 MB
TCP NewReno, B = 20 MB
Goodput [Mbit/s]
1,500
Dinfo = 0, Tinfo = 10 ms
Dinfo = 3 ms, Tinfo = 10 ms
1,600
Goodput [Mbit/s]
milliProxy, B = 10 MB
milliProxy, B = 20 MB
UDP
1,400
1,200
1,000
1,000
2
500
6
11
21
One-way end-to-end latency DS1 + DRS [ms]
2
6
11
(a) TCP goodput
21
One-way end-to-end latency DS1 + DRS [ms]
30
milliProxy, B = 10 MB
milliProxy, B = 20 MB
TCP NewReno, B = 10 MB
TCP NewReno, B = 20 MB
RAN latency [ms]
150
RAN latency [ms]
(a) TCP goodput
20
10
100
Dinfo = 0, Tinfo = 10 ms
Dinfo = 3 ms, Tinfo = 10 ms
0
50
2
6
11
21
One-way end-to-end latency DS1 + DRS [ms]
(b) Latency in the RAN (from the PDCP at the eNB to that at the UE)
0
2
6
11
21
One-way end-to-end latency DS1 + DRS [ms]
Figure 6: Comparison of goodput and RAN latency with different milliProxy
configurations. Dinfo represents the latency needed to forward the cross-layer
information from the gNB to milliProxy, Tinfo is the periodicity at which this
information is collected.
(b) Latency in the RAN (from the PDCP at the eNB to that at the UE)
Figure 5: Comparison of goodput and Radio Access Network (RAN) latency
with and without milliProxy, for different buffer sizes B.
DS1 + DRS [ms]
2
6
11
21
BRLC = 10 MB
BRLC = 20 MB
11.8008
43.3299
4.7547
11.5578
2.5574
5.8104
1.9888
3.6988
(a) RAN latency reduction when using milliProxy, i.e., ratio between the latency
with TCP NewReno and that with milliProxy.
DS1 + DRS [ms]
2
6
11
21
BRLC = 10 MB
BRLC = 20 MB
1.1941
1.0135
1.6875
1.1448
1.7202
1.0765
2.2430
1.9901
(b) TCP goodput gain when using milliProxy, i.e., ratio between the goodput
with milliProxy and with TCP NewReno.
Table II: Goodput and latency performance gains with milliProxy.
traditional approaches without proxy result in higher goodput at
the price of increased RAN latency when using larger buffers.
A comparison between different configuration options for
milliProxy is given in Fig. 6. In particular, we are interested in
studying the sensitivity of goodput and latency with respect to
the delay Dinfo in the acquisition of the cross-layer information
from the gNB: it is equal to 0 when milliProxy is deployed in
the gNB, and greater than 0 when installed in a node in the core
or edge network. We consider Dinfo = 3 ms, i.e., we assume
that the latency between the proxy deployed in the core/edge
network and the gNB will be smaller than 3 ms. As shown in
Fig. 6, the two tested configurations have a similar behavior
in terms of both goodput and latency, showing that milliProxy
is robust with respect to different possible deployments in the
edge network or in the gNBs.
V. C ONCLUSIONS
In this paper we introduced milliProxy, a novel proxy designed to enhance the performance of TCP in mmWave cellular
networks. We described the main challenges related to the
usage of TCP on top of mmWave links, and the main proxy
designs from the literature. MilliProxy splits the TCP control
loop in two segments, while keeping the end-to-end semantics
of TCP. It has a modular design, which enables the use of
different MSS values and flow window management algorithms
in the two portions of the connection (i.e., wired and wireless).
The window control policy can benefit from the interaction of
milliProxy with the protocol stack of the mmWave networks,
which enables cross-layer approaches. We showed how a FW
policy based on the BDP of the end-to-end connection allows a
reduction in latency of up to 10 times or an increase in goodput
of up to 2 times with respect to traditional TCP NewReno, as
well as a robustness with respect to where milliProxy is placed
in the network.
MilliProxy is an option to reach high goodput with low
latency with TCP in mmWave networks. As part of our future
work we will test the proxy performance in a wide variety
of scenarios in ns-3, analyzing the performance with multiple
users, and with different flow window policies, and will consider the implementation in a real setup.
R EFERENCES
[1] S. Rangan, T. S. Rappaport, and E. Erkip, “Millimeter-wave cellular
wireless networks: Potentials and challenges,” Proc. IEEE, vol. 102, no. 3,
pp. 366–385, Mar 2014.
[2] 3GPP, “Study on New Radio (NR) Access Technology - Physical Layer
Aspects - Release 14,” TR 38.802, 2017.
[3] M. Mezzavilla, M. Polese, A. Zanella, A. Dhananjay, S. Rangan,
C. Kessler, T. Rappaport, and M. Zorzi, “Public Safety Communications
above 6 GHz: Challenges and Opportunities,” IEEE ACCESS Special
Section on Mission Critical Public-Safety Communications: Architectures,
Enabling Technologies, and Future Applications, 2017.
[4] J. Choi, V. Va, N. Gonzalez-Prelcic, R. Daniels, C. R. Bhat, and R. W.
Heath, “Millimeter-wave vehicular communication to support massive
automotive sensing,” IEEE Communications Magazine, vol. 54, no. 12,
pp. 160–167, Dec 2016.
[5] M. Giordani, A. Zanella, and M. Zorzi, “Millimeter wave communication
in vehicular networks: Challenges and opportunities,” in 2017 6th IEEE
International Conference on Modern Circuits and Systems Technologies
(MOCAST), May 2017.
[6] H. Shokri-Ghadikolaei, C. Fischione, G. Fodor, P. Popovski, and
M. Zorzi, “Millimeter wave cellular networks: A MAC layer perspective,”
IEEE Transactions on Communications, vol. 63, no. 10, pp. 3437–3458,
Oct 2015.
[7] S. Singh, F. Ziliotto, U. Madhow, E. M. Belding, and M. J. W.
Rodwell, “Millimeter Wave WPAN: Cross-Layer Modeling and MultiHop Architecture,” in 26th IEEE International Conference on Computer
Communications (INFOCOM), May 2007, pp. 2336–2340.
[8] M. Zhang, M. Mezzavilla, R. Ford, S. Rangan, S. S. Panwar, E. Mellios,
D. Kong, A. R. Nix, and M. Zorzi, “Transport Layer Performance
in 5G mmWave Cellular,” in 2016 IEEE Conference on Computer
Communications Workshops (INFOCOM WKSHPS), Apr 2016, pp. 730–
735.
[9] M. Polese, R. Jana, and M. Zorzi, “TCP and MP-TCP in 5G mmWave
Networks,” IEEE Internet Computing, vol. 21, no. 5, pp. 12–19, Sept
2017.
[10] M. Zhang, M. Mezzavilla, J. Zhu, S. Rangan, and S. Panwar, “The
Bufferbloat Problem over Intermittent Multi-Gbps mmWave Links,”
Nov. 2016. [Online]. Available: https://arxiv.org/abs/1611.02117
[11] M. Polese, R. Jana, and M. Zorzi, “TCP in 5G mmWave Networks:
Link Level Retransmissions and MP-TCP,” in 2017 IEEE Conference
on Computer Communications Workshops (INFOCOM WKSHPS), May
2017.
[12] M. Polese, M. Mezzavilla, S. Rangan, and M. Zorzi, “Mobility Management for TCP in mmWave Networks,” in Proceedings of the 1st ACM
Workshop on Millimeter-Wave Networks and Sensing Systems, co-located
with ACM MobiCom’17, ser. mmNets ’17, 2017, pp. 11–16.
[13] S. K. Saha, A. Garg, and D. Koutsonikolas, “A first look at TCP performance in indoor IEEE 802.11ad WLANs,” in 2015 IEEE Conference
on Computer Communications Workshops (INFOCOM WKSHPS), Apr
2015, pp. 63–64.
[14] S. Sur, I. Pefkianakis, X. Zhang, and K.-H. Kim, “WiFi-Assisted 60 GHz
Wireless Networks,” in Proceedings of the 23rd Annual International
Conference on Mobile Computing and Networking, ser. MobiCom ’17.
ACM, 2017, pp. 28–41.
[15] S. Ha, I. Rhee, and L. Xu, “CUBIC: a new TCP-friendly high-speed
TCP variant,” ACM SIGOPS Operating Systems Review, vol. 42, no. 5,
pp. 64–74, July 2008.
[16] T. Henderson, S. Floyd, A. Gurtov, and Y. Nishida, “The NewReno
Modification to TCP’s Fast Recovery Algorithm,” RFC 6582, 2012.
[17] H. Balakrishnan, V. N. Padmanabhan, S. Seshan, and R. H. Katz, “A
comparison of mechanisms for improving TCP performance over wireless
links,” IEEE/ACM Transactions on Networking, vol. 5, no. 6, pp. 756–
769, Dec 1997.
[18] K. Liu and J. Y. Lee, “On Improving TCP Performance over Mobile Data
Networks,” IEEE Transactions on Mobile Computing, vol. 15, no. 10, pp.
2522–2536, Oct 2016.
[19] J. Border, M. Kojo, J. Griner, G. Montenegro, and Z. Shelby, “Performance Enhancing Proxies Intended to Mitigate Link-Related Degradations,” RFC 3135, June 2001.
[20] J. H. Saltzer, D. P. Reed, and D. D. Clark, “End-to-end arguments in
system design,” ACM Trans. Comput. Syst., vol. 2, no. 4, pp. 277–288,
Nov 1984.
[21] F. Ren and C. Lin, “Modeling and improving TCP performance over
cellular link with variable bandwidth,” IEEE Transactions on Mobile
Computing, vol. 10, no. 8, pp. 1057–1070, Aug 2011.
[22] K. Brown and S. Singh, “M-TCP: TCP for mobile cellular networks,”
ACM SIGCOMM Computer Communication Review, vol. 27, no. 5, pp.
19–43, Oct 1997.
[23] H. Balakrishnan, S. Seshan, E. Amir, and R. H. Katz, “Improving TCP/IP
performance over wireless networks,” in Proceedings of the 1st annual
international conference on Mobile computing and networking. ACM,
1995, pp. 2–11.
[24] A. Bakre and B. Badrinath, “I-TCP: Indirect TCP for mobile hosts,”
in Proceedings of the 15th International Conference on Distributed
Computing Systems. IEEE, 1995, pp. 136–143.
[25] M. Kim, S.-W. Ko, and S.-L. Kim, “Enhancing TCP End-to-End
Performance in Millimeter-Wave Communications,” arXiv preprint
arXiv:1709.00717, 2017.
[26] M. Casoni, C. A. Grazia, M. Klapez, and N. Patriciello, “Implementation
and validation of TCP options and congestion control algorithms for ns3,” in Proceedings of the 2015 Workshop on ns-3. ACM, 2015, pp.
112–119.
[27] M. Allman, V. Paxson, and E. Blanton, “TCP Congestion Control,” RFC
5681, September 2009.
[28] M. Zhang, M. Polese, M. Mezzavilla, J. Zhu, S. Rangan, S. Panwar,
and M. Zorzi, “Will TCP Work in mmWave 5G Cellular Networks?”
submitted to IEEE Communication Magazine, 2017.
[29] 3GPP, “NR - Radio Link Control (RLC) protocol specification - Release
15,” TS 38.322, 2017.
[30] D. Skordoulis, Q. Ni, H. H. Chen, A. P. Stephens, C. Liu, and A. Jamalipour, “IEEE 802.11n MAC frame aggregation mechanisms for nextgeneration high-throughput WLANs,” IEEE Wireless Communications,
vol. 15, no. 1, pp. 40–47, Feb 2008.
[31] D. Borman, B. Braden, V. Jacobson, and R. Scheffenegger, “TCP Extensions for High Performance,” RFC 7323, Sep 2014.
[32] B. Veal, K. Li, and D. Lowenthal, “New Methods for Passive Estimation
of TCP Round-trip Times,” in Proceedings of the 6th International
Conference on Passive and Active Network Measurement, ser. PAM’05,
2005, pp. 121–134.
[33] T. Azzino, M. Drago, M. Polese, A. Zanella, and M. Zorzi, “X-TCP: A
Cross Layer Approach for TCP Uplink Flows in mmWave Networks,”
in 16th Annual Mediterranean Ad Hoc Networking Workshop (Med-HocNet’17), June 2017.
[34] H. Jiang, Y. Wang, K. Lee, and I. Rhee, “Tackling bufferbloat in
3G/4G networks,” Proceedings of the 2012 ACM conference on Internet
measurement, pp. 329–342, 2012.
[35] N. Cardwell, Y. Cheng, C. S. Gunn, S. H. Yeganeh, and V. Jacobson,
“BBR: Congestion-based congestion control,” Queue, vol. 14, no. 5, p. 50,
2016.
[36] T. R. Henderson, M. Lacage, G. F. Riley, C. Dowell, and J. Kopena,
“Network simulations with the ns-3 simulator,” ACM SIGCOMM demonstration, vol. 14, no. 14, p. 527, 2008.
[37] M. Mezzavilla, M. Zhang, M. Polese, R. Ford, S. Dutta, S. Rangan,
and M. Zorzi, “End-to-End Simulation of 5G mmWave Networks,”
submitted to IEEE Communication Surveys & Tutorials, 2017. [Online].
Available: https://arxiv.org/abs/1705.02882
[38] R. Ford, M. Zhang, S. Dutta, M. Mezzavilla, S. Rangan, and M. Zorzi, “A
Framework for End-to-End Evaluation of 5G mmWave Cellular Networks
in ns-3,” in Proceedings of the Workshop on Ns-3. ACM, June 2016,
pp. 85–92.
[39] S. Dutta, M. Mezzavilla, R. Ford, M. Zhang, S. Rangan, and M. Zorzi,
“Frame Structure Design and Analysis for Millimeter Wave Cellular
Systems,” IEEE Transactions on Wireless Communications, vol. 16, no. 3,
pp. 1508–1522, Mar 2017.
[40] 3GPP, “Study on channel model for frequency spectrum above 6 GHz Release 14.2.0,” TR 38.900, 2017.
| 7cs.IT
|
arXiv:1709.06990v1 [cs.NE] 20 Sep 2017
Text Compression for Sentiment Analysis via
Evolutionary Algorithms
Emmanuel Dufourq
Bruce A. Bassett
University of Cape Town
African Institute for Mathematical Sciences
Email: [email protected]
University of Cape Town
African Institute for Mathematical Sciences
South African Astronomical Observatory
Email: [email protected]
Abstract—Can textual data be compressed intelligently without
losing accuracy in evaluating sentiment? In this study, we propose
a novel evolutionary compression algorithm, PARSEC (PARts-ofSpeech for sEntiment Compression), which makes use of Partsof-Speech tags to compress text in a way that sacrifices minimal
classification accuracy when used in conjunction with sentiment
analysis algorithms. An analysis of PARSEC with eight commercial and non-commercial sentiment analysis algorithms on twelve
English sentiment data sets reveals that accurate compression is
possible with (0%, 1.3%, 3.3%) loss in sentiment classification
accuracy for (20%, 50%, 75%) data compression with PARSEC
using LingPipe, the most accurate of the sentiment algorithms.
Other sentiment analysis algorithms are more severely affected
by compression. We conclude that significant compression of text
data is possible for sentiment analysis depending on the accuracy
demands of the specific application and the specific sentiment
analysis algorithm used.
Index Terms—sentiment analysis, evolutionary algorithm, compression
I. I NTRODUCTION AND R ATIONALE
Sentiment analysis is the automated detection of emotions
and attitudes towards a particular subject, event or entity,
and has received an increasing amount of interest since the
year 2000 [1], [2]. Sentiment analysis has been applied to
many problem domains; for instance, determining sentiments
of consumers towards products, or mining social media to gain
an understanding of the public’s opinion on matters such as
corruption [1]–[3].
In this study, we propose a novel evolutionary algorithm
(EA) which is tasked with compressing text, whereby the
objective is to retain the correct sentiment classification of the
compressed text. To achieve this, the EA makes use of POS
(Parts of Speech) to determine which POS can be removed
from the text without hindering the classification performance.
EAs [4], [5] – inspired by nature – make use of a population of
individuals which are evolved over several iterations to solve
optimisation problems.
To illustrate the idea, consider the sentence “I went home
yesterday, opened the door and was immediately faced with a
terrible shock”. The key sentiment-encoding phrase is “terrible
shock”. We thus want to propose an algorithm that can learn
to evolve rules that can effectively do this compression. The
remaining words in the sentence do not contain any useful
information in determining the sentiment expressed, and thus
the rules can discard those words.
Feng et al. [6] proposed a Chinese text compression method
with the objective of preserving the opinions expressed in
the text to retain the sentiment, and to ensure that the
compressed sentence is grammatically correct. The authors
propose a score function which takes into consideration three
primary functions; namely a word significance, a linguistic and
an opinion scoring function. The word significance function
makes use of the word frequency properties to allocate a
score to each word. The scores were obtained by processing a
corpus containing 30,000 documents. The linguistic score was
obtained using a n-gram probabilistic function which made
use of Google search results. The opinion score made use of
an opinion lexicon to allocate a value to each word based
on the frequency of the word. The authors created the lexicon
from three different sources and the resulting lexicon contained
32,051 words for which the sentiment (positive, neutral or
negative) was known. A dynamic programming algorithm was
used in order to determine how many words should be present
in a compressed sentence by evaluating several candidate
compressed sentences. The score function was applied to each
candidate sentence and the result was divided by the number of
words in the compressed sentence. The compressed sentence
with the highest value after this operator was deemed to
be the final compressed sentence for a given uncompressed
sentence. One Chinese data set of compressed sentences was
manually created to evaluate their proposed approach. The
study revealed that their proposed approach was able to
outperform a traditional sentence compression method when
the sentences were evaluated - in terms of the opinion and
grammar - by a human.
Che et al. [7] also proposed a text compression method
for sentiment analysis. In their study, they make use of 10
features (which are determined for every word) to perform the
compression. The features are grouped into basic, sentiment
related, semantic and syntactic features. The basic features
include the words before and after the word for which the
features are being constructed for; they also include the POS
tags for those words. There are two sentiment features, namely
a binary field denoting if the word is a perception word and
if it is a polarity word. These words were obtained from
an existing lexicon. The semantic features include prefix and
suffix characters and also Brown word features which helps
identify that words which are written differently represent the
same thing. They also include word embedding which makes
use of Word2Vec [8] to determine if two words have a similar
meaning. The syntactic feature makes use of a single feature
which indicates the relationship between the words in the
sentence. The sentence compression is conducted prior to the
sentiment analysis.
With the limited amount of existing work in this field, we
propose and investigate PARSEC (PARts-of-Speech for sEntiment Compression), which makes use of an EA to achieve
text compression with minimal loss in sentiment accuracy.
II. PARTS - OF -S PEECH
POS, in languages, are defined as the primary groups of
words that are grammatically similar. Common POS include,
but are not limited to: adjective, verb and noun. For example,
the words “dog”, “house” and “student” are examples of
nouns.
In this study, the Stanford Log-linear Part-Of-Speech Tagger
[9] was used to convert the original data sets into their
corresponding POS tags. Thus, each word in a dataset was
converted into a POS tag, and the length of each sentence in
the original dataset was identical to that of the POS tagged
dataset. For example, if the following sentence was found
in the original dataset: “this is a great product”, it would be
converted into “DT VBZ DT JJ NN” (which corresponds to
the POS equivalent) in the POS tagged dataset.
We used the Penn Treebank tag set [10], [11]. This set
contains 36 tags and an additional 12 tags to denote certain
characters such as comma, semi-colon and other characters.
III. P ROPOSED C OMPRESSION M ETHOD
A. Compressor
This study proposes a new type of individual for EAs –
referred to as a compressor – which reduces the length of
sentences when applied to a dataset. The compressor is made
up of several rules, and in turn, each rule is made up of a
sequence of POS tags and a decision. The decision is applied
whenever it’s sequence of POS tags matches the POS tags
in the data. The decisions represent the indices of the words
which will be removed in the matched sequence. Figure 1
illustrates an example of a compressor.
the equivalent POS dataset. Furthermore, let Xi,j denote
sentence i and word j in X, and Yi,j denote sentence i and
word j in Y .
The compressor in the figure has 10 rules denoted as R1 ,
R2 , R3 , ..., R10 . The first rule, R1 , has 2 tags which are
represented by the sequence “JJ, NN”. This rule has one
decision, and that is, to delete the word at index 0. This implies
that when this particular compressor is applied, rule R1 will
find POS tags in Y corresponding to the sequence “JJ NN”.
Let Yl,m and Yl,m+1 denote two consecutive words in Y which
have the sequence “JJ NN”. The decision in R1 is to delete
index 0, thus R1 will delete Yl,m and consequently it will
delete Xl,m . Thus, the size of X and Y has been reduced by
one. Each time R1 finds the exact sequence “JJ NN”, the POS
tag and word corresponding to “JJ” (index 0) in Y and X are
deleted.
Similarly for R10 , Let Yl,m , Yl,m+1 and Yl,m+2 denote three
consecutive words in Y which have the sequence “NN VB
RB”. The decision in R10 is to delete index 0 and 2, thus
R10 will delete Yl,m , Yl,m+2 , Xl,m and Xl,m+2 from Y and
X respectively for each occurrence of the POS tags “NN VB
RB”.
When a rule is being processed, if a punctuation mark
occurs, then the rule evaluation immediately stops. This mechanism was incorporated so that the sequence of tags – which
is encapsulated by a rule – is applied to a single sentence and
is not applied across two separate sentences. For example,
consider the two following sentences “This book was really
funny. Great, tomorrow is Friday.” which corresponds to: “DT
NN VBD RB JJ. JJ, NN VBZ NNP.” Assume that some
rule has the following tags “JJ JJ” and the decision is to
delete both indices. If the punctuation marks are not taken
into consideration, then the rule will match with the words
“funny” and “great”. This will result in “This book was really
. , tomorrow is Friday.” Consequently, the sentiments in these
two sentences are no longer positive. However, if punctuation
is taken into consideration, then the rule will not delete any
words since the full stop separates the two sentences.
In this study, we define the term ‘wildcard’, which represents the notion of ‘any POS’ tag and is denoted with a ‘*’
symbol. Thus, when a rule is being evaluated, the wildcard
POS can match with any POS in the POS dataset. For example,
if a rule has the following tags: “JJ * NNP”, then any POS
sequence that has ‘JJ’, followed by any POS, followed by
‘NNP’ will cause the rule to match with that POS sequence.
Algorithm 1 presents the generalized pseudocode to match
the tags in a rule with the POS tags in the POS dataset. The
algorithm also presents how to delete words from the original
and POS dataset once a rule matches with a sequence.
B. PARSEC initial population generation
Fig. 1. An example of a PARSEC compressor with 10 rules.
In the following explanation, let X denote some original
dataset for which PARSEC is being applied to, and let Y denote
The pseudocode for creating an initial population of compressors is presented in algorithm 2. Several individuals are
created based on a predefined user parameter, namely the
population size. Several rules have to be created for each
compressor. There must be at least one rule, and a user-defined
Algorithm 1: Pseudocode to apply a compressor to a
sentence and compress the text.
input: compressor the compressor to evaluate
input: original sentence the original sentence
input: POS sentence the POS sentence
input: length sentence the length of the sentence to
evaluate
1 begin
2
for each rule in compressor do
3
tag in rule ← the tags inside rule
4
rule decisions ← the decisions inside rule
5
for i ← 0 to length sentence do
6
match ← true
7
for j ← 0 to length(tag in rule) do
8
if tag in rule [j] not equal
POS sentence [i + j] then
9
match ← false
10
break loop and go to line 5
Algorithm 2: Pseudocode to create the initial population
of compressors.
input: population size the number of compressors to
create
1 begin
2
for i ← 0 to population size do
3
CreateCompressor()
4
Evaluate compressor.
5
Add compressor to the initial population.
6
7
8
9
10
11
12
11
12
13
if match equal true then
for k ← 0 to length(rule decisions) do
flag original sentence
[i + rule decisions[k]]
13
14
15
14
15
compressed sentence ← delete all words in
original sentence that have been flagged.
return compressed sentence
16
17
18
parameter, namely maxrules , is implemented to restrict the total
number of rules. When creating the rules, the algorithm iterates
until it has reached maxrules , and a probability value of 0.5
was set to determine if a rule was to be created or not, as
can be seen on line 9 in algorithm 2. Each rule consists
of two primary parts, namely the tags and the decisions,
the pseudocode for the creation of those respective parts are
presented in algorithms 3 and 4.
C. PARSEC fitness function
The pseudocode for the fitness evaluation of a compressor
is presented in algorithm 6. Each instance in the dataset
is considered in turn. The sentiment for each instance in
the original dataset is computed using algorithm 5. When
a compressor is evaluated, it is applied to each instance
and a compressed sentence is created. The sentiment of
each compressed sentence is computed using algorithm 5.
The compressed sentiment is then compared to the original
sentiment. If these are the same then the compressed sentence
has had no effect on the sentiment when compared to the
original. Conversely, if the values are different, then there are
two possibilities. Either the compressed sentiment is equal to
the correct sentiment as labelled in the training data (in which
case the compressor is rewarded) or it is not (in which case
the compressor is penalised). Thus, the fitness function takes
Function CreateCompressor ()
new compressor ← create a blank compressor
for j ← 0 to max tags do
if rand() < 0.5 then
CreateRule() and add rule to
new compressor
if Compressor has no rule then
CreateRule() and add rule to
new compressor
return new compressor
Function CreateRule ()
new rule ← create a blank rule
Create the tags using algorithm 3 and add the
rules to new rule.
Create the decisions using algorithm 4 and add
the rules to new rule.
return The initial population.
Algorithm 3: Pseudocode to create POS tags.
input: minimum tags the minimum number of allowed
pos tags
input: maximum tags the maximum number of allowed
pos tags
1 begin
2
pos list ← { }
3
number tags ← random[minimum tags,
maximum tags ]
4
for i ← 0 to number tags do
5
if i = 0 then
6
Create a tag (excluding wildcard) and add it
to pos list
7
else if i = number tags-1 then
8
Create a tag (excluding wildcard) and add it
to pos list
9
else
10
Create a tag (including wildcard) and add it
to pos list
11
return pos list
Algorithm 4: Pseudocode to create decisions.
input: number tags the number of tags
1 begin
2
for i ← 0 to number tags do
3
if random[0,1] < 0.5 then
4
Add i to decision list
5
return decision list.
Algorithm 5: Pseudocode for baseline evaluation.
input : sentence the sentence to be evaluated
input : dictionary the dictionary of sentiment words
input : negation words the list of negation words
output: The sentiment for the evaluated sentence
1 begin
2
sentiment score ← 0
3
negation ← false
4
for each word in the sentence do
5
if word is in dictionary and negation is true
then
6
sentiment score ← sentiment score + (-1
× word’s sentiment value)
7
if word is in dictionary and negation is false
then
8
sentiment score ← sentiment score +
word’s sentiment value
9
if word is in negation words then
10
negation ← swap polarity
12
else
negation ← false
13
return sentiment score
11
Algorithm 6: Pseudocode to compute fitness of a compressor.
input : compressor the compressor to evaluated
input : sentence a list of sentences for which
compressor will be evaluated on
input : original lengths a list of lengths for each
sentence in the original data set
input : original sentiments a list of sententiment
values determined by applying algorithm 5 on
the original data set
input : correct sentiments a list of correct
sententiment values for the original data set
input : number rules the number of rules that
compressor has
output: The sentiment for the evaluated sentence
1 begin
2
for i ← 1 to number of sentences do
3
compressed sentence ← apply compression to
sentences [i].
4
compressed sentiment ← apply algorithm 5.
5
if compressed sentiment is not equal to
original sentiments [i] then
6
if compressed sentiment is equal to
correct sentiments [i] then
7
raw fitness ← raw fitness + 1
8
9
10
11
12
into consideration the number of new instances for which the
compressor produces the correct result.
A multi-objective fitness function was created to combine
the fitness, the average reduction in terms of the length of
the sentences before and after compression, and the size of
the compressor in terms of the number of rules. Thus, the
objective was to maximize the fitness and average reduction
in sentences, and to minimize the size of the compressor.
IV. E XPERIMENTAL S ETUP
A. Datasets and Setup
Twelve data sets were created by randomly selecting reviews
from corresponding Amazon review datasets [12]. For each
created data set, 1000 positive and 1000 negative reviews were
randomly selected from the larger corresponding Amazon data
set. The larger Amazon data sets were obtained from [12],
[13]. The data sets were: Amazon Instant Video (AIV), Apps
for Android, Automotive, Baby, Beauty, Digital Music, Health
and Personal Care, Musical Instruments (MI), Patio Lawn and
else
raw fitness ← raw fitness- 1
average change ← average change +
(Length(sentences [i]) Length(compressed sentence))
raw fitness ← raw fitness + [0.5 ×
(average change/ number of sentences)] - (0.1
× number rules)
return raw fitness
Garden, Pet Supplies, Tools and Home Improvement, and Toys
and Games.
We propose a set of experiments whereby the compression
rate is a user parameter and PARSEC must generate compressors that can compress the original datasets to reach the
specified compression rate.
To achieve this, a lower and upper user-defined compression
bound were defined. The lower compression bound (LCB) was
implemented to ensure that the compressors did not result in a
compression ratio less than the specified value. Similarly, the
upper compression bound (UCB) ensured that the compressors
did not result in a compression ratio greater than the specified
value. Thus, every compressor in the EA population had to
have a compression rate between the LCB and UCB.
Furthermore, two additional user-defined parameters were
implemented to ensure that the total number of rules within
each compressor remained within a certain bound. The
two parameters were named compressionRulesmin and
TABLE I
T HE LOWER (LCB) AND UPPER (UCB) COMPRESSION BOUND CONSTRAIN
THE ALGORITHM BY ENSURING THAT THE COMPRESSION RATE FOR EACH
COMPRESSOR IS BETWEEN THE TWO VALUES RESPECTIVELY. T HE TABLE
PRESENTS THE LCB AND UCB VALUES (%) ALONG WITH THE MINIMUM
AND MAXIMUM NUMBER OF RULES USED IN THE EXPERIMENTS . T HESE
SET COMPRESSION RATES WERE ENFORCED AND THE PERFORMANCE OF
PARSEC WAS MEASURED ON SEVERAL DATASETS . T HE VALUES FOR THE
MINIMUM AND MAXIMUM NUMBER OF RULES WERE DETERMINED BY
ADDITIONAL TRIAL RUNS . I T WAS OBSERVED THAT A GREATER NUMBER
OF RULES WERE NEEDED TO ACHIEVE HIGHER COMPRESSION RATES .
LCB
UCB
10
15
20
25
30
50
13
18
13
28
33
53
compression
Rulesmin
5
10
20
40
90
350
compression
Rulesmax
50
70
90
120
150
500
compression rates of 10% and 15% respectively. When experimenting with PARSEC we posed the following question. For
a predetermined drop in accuracy, how much compression is
achievable by PARSEC? For a threshold of -1% compression,
Vivek and LingPipe were able to achieve up to 30% compression. For this same threshold five methods could achieve up
to 20% compression rate. Regardless of the compression rate
used, both Stanford and Sentiment140 did not obtain any result
less than -1%. These two algorithms produced the weakest
performance with PARSEC.
0
−1
−2
B. Sentiment analysis algorithms
Several sentiment analysis algorithms were applied to the
original and compressed datasets to determine the difference
in accuracy and to investigate the performance of PARSEC.
The following algorithms were used: SentiStrength [14], [15],
MeaningCloud [16], Vivek [17], Stanford [18], uClassify [19],
Sentiment140 [20], Intellexer [21] and LingPipe [22].
C. Execution of experiments
We first applied the sentiment analysis algorithms described
in section IV-B to the original datasets and record the test
accuracy. Then, we ran PARSEC using the baseline evaluation to create compressor rules. Once the compressors were
created, we applied the compressors to the original datasets
to generate the compressed data. We applied the sentiment
analysis algorithms on the compressed data and recorded
the test accuracy. In the following section we report on the
change in accuracy between the compressed and original data.
PARSEC evolved a population of 250 compressors over 100
generations, the crossover rate was 60% and mutation rate was
40%.
V. R ESULTS AND C ONCLUSION
Figure 2 shows the average change in test accuracy across
the different methods based on the compression rates. The
general trend is a decrease in performance with the largest
decrease in performance observed by Sentiment140. LingPipe
and SentiStrength were the only two methods that achieved
a positive change in performance (on certain datasets) on
−3
−4
Change in Test Accuracy (%)
compressionRulesmax . The compressionRulesmax parameter prevents compressors from having an extremely large
number of rules. Both parameters aid in restricting the search
space in terms of the number of rules to create and maintain for
each compressor during the evolutionary process. The initial
population generation had to respect the LCB and UCB, and
respect the minimum and maximum number of rules. The
mutation operator also had to respect these constraints.
−5
Method
IN
LP
MC
S140
SS
ST
uC
V
−6
−7
−8
−9
−10
−11
−12
−13
−14
−15
20
40
60
Compression Rate (%)
Fig. 2. Change in average test accuracy across the different sentiment analysis
algorithms for different text compression rates enforced by PARSEC, averaged
across all the data sets. A higher positive value indicates a better result.
The various sentiment analysis technologies are denoted as follows: IN, LP,
MC, S140, SS, ST, uC and V refer to Intellexer, LingPipe, MeaningCloud,
Sentiment140, SentiStrength, Stanford, uClassify and Vivek respectively.
The accuracy obtained by LingPipe outperformed all the other methods.
Furthermore, LingPipe obtained an average positive change in accuracy of
0.1% for 10% compression. At a compression rate of 50%, LingPipe obtained
an change in accuracy of -1.2% averaged over all the datasets, illustrating that
it works well with PARSEC.
For a threshold of -2%, five algorithms could achieve up to
30% compression rate and LingPipe was able to achieve up
to 50% compression. When the threshold is set to -3% four
methods could achieve up to 50% compression.
We additionally wanted to determine the change in test
accuracy when a much larger compression rate was used. Table
II presents the results when a compression rate of 75% was
used. Only certain algorithms were tested due to our limited
access to the sentiment analysis algorithms. These results
reveal that a large compression rate is achievable without a
great loss in test accuracy: LingPipe obtained a loss of -3.3%
accuracy.
An example of PARSEC compression is as follows. The
original sentence (which contains 56 words and is already
stemmed): “this be a wonderful movie involve fascinating
people who be such great actor and actress. the event be
interesting and I never get tire of watch the episode. I be so sad
to see that they be not go to keep go with the life of all these
inhabitant of Lark rise and candleford. Mrs. Hamlin”. The
compressed sentence (22 words): “this be wonderful movie
involve fascinating people who actress. get tire. so sad to see
not keep go rise candleford. Mrs. Hamlin”.
TABLE II
AVERAGE CHANGE IN TEST ACCURACY (%) ACROSS ALL THE DATA SETS
WHEN PARSEC MODELS WERE CREATED FOR 75% COMPRESSION RATE .
Algorithm
LingPipe
MeaningCloud
SentiStrength
uClassify
Change in
Test Accuracy
-3.2
-4.2
-6.9
-6.4
Original
Accuracy
79.7
62.5
67.8
80.7
Compressed
Accuracy
76.4
58.2
60.8
74.3
Below are four rules which were randomly selected from a
PARSEC model to illustrate some of the evolved rules. This
model had a total of 38 rules. The POS tags are presented in
order of appearance between square brackets. The indices start
from zero. Rule 1 denotes that if the POS “NNP” is followed
by “JJ”, then delete both words. Rule 4 denotes that if “IN” is
followed by any POS and then followed by “VB”, then delete
the word at index 1.
• Rule 1: [NNP, JJ]. Number of decisions: 2.
Decisions: [0, 1]
• Rule 2: [VBG, NNP]. Number of decisions: 2.
Decisions: [0, 1]
• Rule 3: [IN, CC]. Number of decisions: 2.
Decisions: [0, 1]
• Rule 4: [IN, *, VB]. Number of decisions: 1.
Decisions: [1]
This study empirically demonstrates that machine learning
can reduce the amount of data needed to determine the
sentiments within sentences with little loss in accuracy. To
achieve this, we proposed an evolutionary algorithm, PARSEC,
which evolves a population of chromosomes which encode
rules for compression based on the removal of POS.
We studied the test accuracies achieved by the eight algorithms when set compression rate thresholds were enforced.
Two algorithms showed marginal improvements in accuracy
on several data sets for a compression rate of 15%. Additionally, there were 9 data sets whereby some of the algorithms were able to yield an improvement in accuracy for a
compression rate of 30%. At a threshold of -1% sentiment
accuracy loss, two methods were able to achieve up to 30%
compression, five methods up to 20% compression, and six
methods up to 10% compression. The best performing algorithm, LingPipe, showed only modest losses in accuracy even
at high compression: 3.3% for a four-fold data compression.
It would be of interest to incorporate the PARSEC algorithm directly into a sentiment analysis algorithm instead of
via a simple dictionary approach. One could achieve this
by replacing the fitness function used in the evolutionary
phase with a more sophisticated sentiment analysis algorithm.
This would allow the emergence of an algorithm where the
compression and sentiment analysis are mutually optimised
for best performance.
ACKNOWLEDGMENT
The financial assistance of the National Research Foundation (NRF) towards this research is hereby acknowledged.
Opinions expressed and conclusions arrived at, are those of the
authors and are not necessarily to be attributed to the NRF.
R EFERENCES
[1] B. Pang and L. Lee, “Opinion mining and sentiment analysis,” Found.
Trends Inf. Retr., vol. 2, no. 1-2, pp. 1–135, Jan. 2008.
[2] B. Liu, Sentiment Analysis: Mining Opinions, Sentiments, and Emotions.
Cambridge University Press, 2015.
[3] K. Ravi and V. Ravi, “A survey on opinion mining and sentiment
analysis,” Know.-Based Syst., vol. 89, no. C, pp. 14–46, Nov. 2015.
[4] A. E. Eiben and J. E. Smith, Introduction to Evolutionary Computing.
SpringerVerlag, 2003.
[5] X. Yu and M. Gen, Introduction to Evolutionary Algorithms. Springer
Publishing Company, Incorporated, 2012.
[6] S. Feng, D. Wang, G. Yu, B. Li, and K.-F. Wong, A Chinese Sentence
Compression Method for Opinion Mining. Berlin, Heidelberg: Springer
Berlin Heidelberg, 2010, pp. 320–329.
[7] W. Che, Y. Zhao, H. Guo, Z. Su, and T. Liu, “Sentence compression for
aspect-based sentiment analysis,” IEEE/ACM Trans. Audio, Speech and
Lang. Proc., vol. 23, no. 12, pp. 2111–2124, Dec. 2015.
[8] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of
word representations in vector space,” arXiv preprint arXiv:1301.3781,
2013.
[9] K. Toutanova, D. Klein, C. D. Manning, and Y. Singer, “Featurerich part-of-speech tagging with a cyclic dependency network,” in
Proceedings of the 2003 Conference of the North American Chapter
of the Association for Computational Linguistics on Human Language
Technology - Volume 1, ser. NAACL ’03.
Stroudsburg, PA, USA:
Association for Computational Linguistics, 2003, pp. 173–180.
[10] M. P. Marcus, M. A. Marcinkiewicz, and B. Santorini, “Building a large
annotated corpus of english: The penn treebank,” Comput. Linguist.,
vol. 19, no. 2, pp. 313–330, Jun. 1993.
[11] A. Taylor, M. Marcus, and B. Santorini, Treebanks: Building and Using
Parsed Corpora. Dordrecht: Springer Netherlands, 2003, ch. The Penn
Treebank: An Overview, pp. 5–22.
[12] J. Leskovec and A. Krevl, “SNAP Datasets: Stanford large network
dataset collection,” http://snap.stanford.edu/data, Jun. 2014.
[13] J. McAuley, R. Pandey, and J. Leskovec, “Inferring networks of substitutable and complementary products,” in Proceedings of the 21th ACM
SIGKDD International Conference on Knowledge Discovery and Data
Mining, ser. KDD ’15. New York, NY, USA: ACM, 2015, pp. 785–794.
[14] M. Thelwall, K. Buckley, G. Paltoglou, D. Cai, and A. Kappas, “Sentiment in short strength detection informal text,” J. Am. Soc. Inf. Sci.
Technol., vol. 61, no. 12, pp. 2544–2558, Dec. 2010.
[15] M. Thelwall, K. Buckley, and G. Paltoglou, “Sentiment strength detection for the social web,” J. Am. Soc. Inf. Sci. Technol., vol. 63, no. 1,
pp. 163–173, Jan. 2012.
[16] MeaningCloud, “Meaningcloud is a brand by meaningcloud llc, a
s—ngular company.” https://www.meaningcloud.com/, 2015.
[17] V. Narayanan, I. Arora, and A. Bhatia, “Fast and accurate sentiment
classification using an enhanced naive bayes model,” in Proceedings
of the 14th International Conference on Intelligent Data Engineering
and Automated Learning — IDEAL 2013 - Volume 8206, ser. IDEAL
2013. New York, NY, USA: Springer-Verlag New York, Inc., 2013,
pp. 194–201.
[18] R. Socher, A. Perelygin, J. Y. Wu, J. Chuang, C. D. Manning, A. Y. Ng,
and C. Potts, “Recursive deep models for semantic compositionality over
a sentiment treebank,” in Proceedings of the conference on empirical
methods in natural language processing (EMNLP), 2013, pp. 1631–
1642.
[19] uClassify, “uclassify,” https://www.uclassify.com/, 2008.
[20] A. Go, R. Bhayani, and L. Huang, “Twitter sentiment classification using
distant supervision,” Technical report, Stanford, 2009.
[21] EffectiveSoft, “Intellexer,” http://www.intellexer.com/, 2000.
[22] Alias-i., “Lingpipe 4.1.0,” http://alias-i.com/lingpipe, 2008.
A PPENDIX
Tables III to VIII present the difference in test accuracy results for the various compression rates. The various sentiment
analysis algorithms are abbreviated as follows: IN, LP, MC,
S140, SS, ST, uC and V refer to Intellexer, LingPipe, MeainingCloud, Sentiment140, SentiStrength, Stanford, uClassify and
Vivek respectively. For a compression rate of 10%, the best
and worst change in accuracy were obtained by LingPipe and
Sentiment140 respectively. Based on all the findings for 10 per
cent compression, there were 21 cases across all the methods
and data sets whereby the change in accuracy was greater
than zero. For a compression rate of 15% the number of cases
was 17 whereby LingPipe obtained the best average change
in accuracy with a value of 0.0%.
The best change in accuracy for 20% compression rate was
obtained by uClassify with a value of 0.7 on two data sets.
There were six data sets for which uClassify reduced the data
up to 20% and for which the test accuracy on the compressed
data was better than on the original data. Additionally, there
were 14 cases in total across all the algorithms whereby the
change in accuracy was greater than zero for 20% compression.
The findings revealed that a for the compression rate of 25%
there were 14 cases for which the change was greater than
zero, and the best change in average accuracy was obtained
once again by LingPipe with a value of -0.3%. There was a
reduction performance in terms of the number of cases having
a change greater than zero when a compression of 30% was
used; notably 9 cases. uClassify however, on 5 datasets, had
a test accuracy greater than zero. Finally, only two algorithms
were able to obtain improvements in terms of the changes for
a compression rate of 50%, namely uClassify and LingPipe,
with changes in 3 and 2 data sets respectively.
AIV
MI
Digital
Music
Baby
Patio &
Garden
Automotive
Pet
Supplies
Apps for
Android
Beauty
Tools &
Home
Toys &
Games
Health &
Personal
Average
IN
-0.9
0.1
-1.0
LP
-0.0
0.3
0.2
MC
0.4
-0.7
-0.3
S140
-2.1
-1.9
-2.5
SS
-0.3
-0.4
-0.1
ST
-3.0
0.5
-1.9
uC
-1.9
-0.4
-0.4
V
-0.9
-1.4
-2.0
0.0
-0.1
0.4
0.4
-0.2
-0.2
-2.7
-2.1
0.0
-0.2
-1.2
-0.7
0.3
0.9
0.2
-0.9
-1.1
-1.2
0.4
0.3
-0.5
-0.4
-2.1
-2.9
0.0
-0.4
-1.1
0.3
0.7
0.0
-0.9
-1.6
-1.4
0.0
-0.5
-2.5
-0.3
-3.5
0.0
-1.6
-0.5
0.8
0.0
-0.1
0.2
-1.0
-1.2
-2.3
0.0
-0.6
-0.5
-1.7
-0.7
0.5
-0.4
0.0
-0.5
0.0
-1.5
-1.8
0.2
-3.1
-0.1
-1.4
0.4
0.1
-0.4
-1.6
-0.3
-1.4
0.5
-0.4
-0.4
0.1
-0.4
-2.1
TABLE III
-0.2
-1.4
0.0
-0.9
D IFFERENCE BETWEEN THE COMPRESSED AND ORIGINAL TEST
ACCURACY WHEN A COMPRESSION RATE BETWEEN 10% AND 13% WAS
IMPOSED .
AIV
MI
Digital
Music
Baby
Patio &
Garden
Automotive
Pet
Supplies
Apps for
Android
Beauty
Tools &
Home
Toys &
Games
Health &
Personal
Average
IN
-1.1
1.1
-1.6
LP
-1.1
0.5
0.4
MC
-0.2
-0.8
-0.7
S140
-2.7
-2.9
-3.6
SS
0.1
-0.8
-0.5
ST
-3.2
0.0
-2.3
uC
-2.0
-1.3
-1.2
V
-0.6
-0.8
-2.1
0.0
-0.3
-0.9
-0.4
-0.5
-0.3
-2.4
-3.5
0.2
-0.6
-2.0
-1.2
0.6
0.5
-0.1
-1.0
-1.4
-1.1
0.6
0.6
-1.0
-0.2
-2.8
-3.8
-0.3
-0.6
-2.2
-1.3
0.9
0.2
-0.8
-2.9
-1.7
-0.3
-0.9
-4.2
-0.5
-3.4
-1.0
-3.2
-0.9
-0.2
-0.3
0.0
-0.7
0.0
-3.5
-3.0
-0.2
-0.3
-1.3
-1.7
-1.8
1.2
-1.2
0.7
-1.3
-0.2
-1.2
-1.6
-0.1
-1.1
-0.8
-1.8
0.1
0.0
-0.6
-2.8
-0.5
-2.6
0.2
-0.4
-0.7
0.0
-0.6
-3.0
TABLE IV
-0.3
-1.8
-0.3
-1.2
D IFFERENCE BETWEEN THE COMPRESSED AND ORIGINAL TEST
ACCURACY WHEN A COMPRESSION RATE BETWEEN 15% AND 18% WAS
IMPOSED .
AIV
MI
Digital Music
Baby
Patio & Garden
Automotive
Pet Supplies
Apps for Android
Beauty
Tools & Home
Toys & Games
Health & Personal
Average
IN
-1.7
0.2
-2.2
-0.7
-0.9
-1.6
-1.0
-1.7
-1.1
0.6
-1.5
0.0
-0.9
LP
-1.2
1.1
-0.4
-0.4
-1.1
0.2
0.4
-0.1
0.2
-0.1
0.0
-0.4
-0.1
MC
S140
0.3
-4.0
-1.1
-2.5
-0.2
-3.5
-0.4
-2.7
-1.2
-3.7
-0.3
-3.5
-1.3
-3.5
-1.6
-4.8
-0.6
-4.0
-1.2
-4.4
-0.4
-3.3
-1.0
-4.2
-0.7
-3.7
TABLE V
SS
-0.4
-0.6
-1.2
-0.8
-0.5
-0.9
-0.4
-0.4
-1.0
-0.7
-0.2
-1.2
-0.7
ST
-3.1
0.4
-3.1
-2.2
-2.2
-2.8
-2.1
-4.5
-1.1
-1.9
-2.8
-1.7
-2.2
uC
-1.8
-1.2
-1.7
0.5
0.5
0.3
0.7
-1.7
-1.2
0.0
-1.5
0.7
-0.5
V
-1.8
-1.4
-2.0
-1.1
-0.9
-1.7
-2.2
-3.3
-1.2
-1.5
-1.9
-0.9
-1.6
D IFFERENCE BETWEEN THE COMPRESSED AND ORIGINAL TEST ACCURACY WHEN A COMPRESSION RATE BETWEEN 20% AND 23% WAS IMPOSED .
AIV
MI
Digital Music
Baby
Patio & Garden
Automotive
Pet Supplies
Apps for Android
Beauty
Tools & Home
Toys & Games
Health & Personal
Average
IN
-2.4
-0.6
-1.9
0.1
-0.9
-2.0
-0.8
-2.4
-1.1
-0.8
-1.2
0.3
-1.1
LP
-1.6
1.0
-0.1
-0.9
-1.4
0.1
0.0
-0.4
0.2
-1.0
-0.5
0.9
-0.3
MC
S140
-0.6
-3.7
-0.1
-4.8
-0.5
25.1
0.5
-4.4
0.1
-4.7
0.1
-3.3
-0.7
-5.3
-1.4
-6.7
-0.7
-5.5
-1.6
-5.2
-1.1
-5.7
0.0
-3.9
-0.5
-2.3
TABLE VI
SS
-0.2
-1.0
-0.5
-0.5
-0.9
-1.5
-0.6
-0.9
-0.6
-1.1
-0.2
-0.3
-0.7
ST
-6.0
-1.6
-2.4
-2.6
-1.5
-4.2
-1.7
-3.4
-2.1
-3.6
-2.7
-2.8
-2.9
uC
-2.5
-0.1
-2.9
-0.4
0.6
0.5
0.1
-0.9
-1.8
0.0
-1.0
-0.9
-0.7
V
-2.0
-1.8
-3.5
-0.6
-1.6
-0.4
-2.9
-3.4
-2.4
-0.3
-3.6
-1.1
-1.9
D IFFERENCE BETWEEN THE COMPRESSED AND ORIGINAL TEST ACCURACY WHEN A COMPRESSION RATE BETWEEN 25% AND 28% WAS IMPOSED .
AIV
MI
Digital Music
Baby
Patio & Garden
Automotive
Pet Supplies
Apps for Android
Beauty
Tools & Home
Toys & Games
Health & Personal
Average
IN
-2.3
-0.5
-3.3
-0.8
-1.8
-1.2
-2.2
-3.0
-2.4
-0.9
-2.2
-0.6
-1.8
LP
-2.1
0.6
-1.1
-0.7
-1.2
-0.4
-0.9
0.0
-0.2
0.0
0.0
-0.3
-0.5
MC
-0.1
-1.1
-1.0
0.2
-1.7
-0.9
-1.1
-2.1
-1.1
-1.3
-1.6
-0.60
-1.0
TABLE
S140
-8.2
-6.9
-6.2
-6.1
-6.6
-4.8
-5.8
-7.7
-7.7
-5.9
-6.0
-6.10
-6.5
VII
SS
-1.3
-1.6
-1.7
-1.1
-1.5
-1.5
-1.0
-1.1
-1.2
-1.5
-0.2
-0.70
-1.2
ST
-7.0
-2.4
-3.6
-5.3
-4.3
-4.7
-4.4
-6.4
-4.6
-5.0
-5.0
-5.30
-4.8
uC
-2.3
-1.9
-2.8
0.7
-0.3
0.8
1.6
-2.1
-1.9
0.4
-0.7
0.70
-0.6
V
-2.5
-2.8
-4.9
-1.3
-2.2
-2.6
-2.8
-5.3
-3.2
-1.1
-3.0
-2.00
-2.8
D IFFERENCE BETWEEN THE COMPRESSED AND ORIGINAL TEST ACCURACY WHEN A COMPRESSION RATE BETWEEN 30% AND 33% WAS IMPOSED .
AIV
MI
Digital Music
Baby
Patio & Garden
Automotive
Pet Supplies
Apps for Android
Beauty
Tools & Home
Toys & Games
Health & Personal
Average
IN
-5.8
-1.9
-5.5
-1.4
-3.8
-3.7
-1.8
-6.2
-3.0
-2.2
-4.8
-2.2
-3.5
LP
-4.2
0.8
-1.3
-1.4
-1.8
-2.4
-1.6
-1.6
-0.1
0.1
-1.2
-0.2
-1.2
MC
-1.8
-0.9
-1.9
-1.7
-2.1
-1.9
-2.2
-4.8
-1.8
-2.2
-2.0
-2.0
-2.1
TABLE
S140
-15.6
-15.1
-15.3
-12.4
-13.9
-12.6
-14.9
-17.1
-15.3
-14.3
-12.7
-12.1
-14.3
VIII
SS
-3.4
-2.6
-3.5
-2.4
-2.5
-2.8
-2.5
-4.4
-3.0
-3.6
-1.8
-2.2
-2.9
ST
-13.2
-6.9
-9.9
-11.2
-10.1
-7.9
-8.0
-12.4
-8.7
-9.5
-10.1
-10.1
-9.8
uC
-7.8
-3.9
-5.5
-1.5
-0.5
0.1
0.6
-4.3
-4.1
1.0
-4.3
-1.1
-2.6
V
-8.2
-7.6
-9.0
-5.6
-4.7
-4.6
-6.2
-11.0
-7.8
-4.9
-7.2
-5.0
-6.8
D IFFERENCE BETWEEN THE COMPRESSED AND ORIGINAL TEST ACCURACY WHEN A COMPRESSION RATE BETWEEN 50% AND 53% WAS IMPOSED .
| 9cs.NE
|
Resilient Distributed State Estimation for LTI Systems ⋆
Aritra Mitra and Shreyas Sundaram a
arXiv:1802.09651v1 [cs.SY] 27 Feb 2018
a
School of Electrical and Computer Engineering at Purdue University, West Lafayette, IN 47907, USA
Abstract
Consider a linear time-invariant (LTI) dynamical system monitored by a network of sensors, modeled as nodes of an underlying
directed communication graph. We study the problem of collaboratively estimating the state of the system when certain nodes
are compromised by adversarial attacks. Specifically, we consider a Byzantine adversary model, where a compromised node
possesses complete knowledge of the system dynamics and the network, and can deviate arbitrarily from the rules of any
prescribed algorithm. We first characterize certain fundamental limitations of any distributed state estimation algorithm in
terms of the measurement and communication structure of the nodes. We then develop an attack-resilient, provably correct
state estimation algorithm that admits a fully distributed implementation. To characterize feasible network topologies that
guarantee success of our proposed technique, we introduce a notion of ‘strong-robustness’ that captures both measurement
and communication redundancy. Finally, by drawing connections to bootstrap percolation theory, we argue that given an LTI
system and an associated sensor network, the ‘strong-robustness’ property can be checked in polynomial time.
Key words: Resilient state estimation; distributed estimation; Byzantine attacks.
1
Introduction
The control of large-scale complex networked systems
such as power grids, transportation networks, water and
gas distribution networks, and multi-agent robotic systems requires precise estimation of the state of the underlying dynamical process. Typically, in these applications, sensors (nodes) collecting information about the
process are scattered over a geographical network. As
the diameter of such networks increase, routing information from all the sensors to a central computational resource induces large delays and creates communication
bottlenecks. To bypass these difficulties, it thus becomes
important to consider distributed algorithms where individual sensors communicate only with sensors within
a physical proximity. However, the potential merits [1]
of such a distributed approach are matched by various
challenges. In particular, a key challenge is to design networks and distributed algorithms that guarantee reliable
operation of the system in the face of faults or sophisticated adversarial attacks on certain sensors. This leads
⋆ This paper was not presented at any IFAC meeting. This
research was supported in part by NSF CAREER award
1653648. Corresponding author: Shreyas Sundaram. Telephone of the corresponding author: 1-(765)-496-0406.
Email addresses: [email protected],
[email protected] (Aritra Mitra and Shreyas
Sundaram).
Preprint submitted to Automatica
to the motivation behind our present work.
In a classical distributed state estimation setup, each
node receives partial measurements of the state of an
LTI process, and seeks to asymptotically estimate the
entire state by exchanging information with its neighbors in the network. Given that our primary focus will be
on security related issues associated with this problem,
we direct the interested reader to recent work on singletime-scale distributed observers in [2–6] that make minimal system-theoretic and graph-theoretic assumptions.
For literature on distributed Kalman filtering, see [7–9].
However, none of these papers address the challenges
associated with tolerating unreliable components in the
network. Accordingly, we now provide a survey of the
cyber-security literature that is most relevant to our
present cause.
Related Work: Over the last decade, a significant
amount of research has focused on security in networked
control systems. In particular, for noiseless dynamical
systems, it has been established that zero-dynamics
plays a key role in characterizing the stealth of an attack [10, 11]. For networked control systems affected
by noise, the authors in [12] recently introduced an
information-theoretic metric that quantifies the detectability of an attack. An unifying feature of these
works, and the ones in [13–15], is that they involve
systems where all the sensor measurements are avail-
28 February 2018
conditions are met. A thorough analysis of the proposed
filtering scheme is then presented in Section 5.
able at a single location. In the sequel, we shall refer
to such systems as centralized control systems. Our
problem formulation and subsequent analysis differs
from the above literature by constraining each sensor
to exchange information with only its neighbors in the
communication graph.
As our third contribution, in Section 7, we introduce a
topological property called ‘strong-robustness’ to characterize feasible systems and networks for which our proposed approach is guaranteed to work. By drawing connections to bootstrap percolation theory, we show that
the ‘strong-robustness’ property can be checked in polynomial time (in the size of the system and the network).
While the study of security in centralized control systems is now mature, there lacks a comprehensive theoretical understanding of analogous questions in a distributed setting. Preliminary attempts to counter adversarial behavior in a distributed state estimation context
are reported in [16], [17]. However, unlike our results,
these papers neither provide any theoretical guarantees
of success, nor allude to graph-theoretic conditions that
are necessary for their respective algorithms to work.
Recently, in [18], the authors employ an H∞ based approach for detecting biasing attacks in distributed estimation networks. Our present work deviates from [18]
in several aspects, namely (i) while the analysis in [18]
is limited to a certain class of attack inputs, our attack
model allows compromised nodes to behave arbitrarily,
i.e., no restrictions are placed on the inputs that can be
injected by an adversary, (ii) unlike [18], we develop a filtering algorithm that allows each uncompromised node
to asymptotically recover the state of the plant without
explicitly detecting which nodes are under attack, and
(iii) the existence of the attack detection filter proposed
in [18] relies on solving an LMI; however, the authors
neither provide graph-theoretic insights regarding the
solvability of such an LMI nor discuss whether the LMI
can be solved in a distributed manner. In contrast, we
detail graph-theoretic conditions that allow each step of
our approach to have a resilient, distributed implementation. Summarily, this paper attempts to bridge the gap
between centralized and distributed resilient state estimation. In this context, our contributions are as follows.
We reported certain preliminary developments in [21].
In this paper, we significantly expand upon our prior results in the following ways. (i) While the analysis in [21]
was limited to system matrices with real, distinct eigenvalues, our present framework allows the system matrix
to have arbitrary spectrum. (ii) Section 3 is a new addition entirely, and discusses fundamental limitations of
any distributed state estimation algorithm in the face
of arbitrary adversarial attacks. (iii) Section 7 contains
additional details about properties of feasible network
topologies and discusses the applicability of our proposed approach.
Notation: A directed graph is denoted by G = (V, E),
where V = {1, · · · , N } is the set of nodes and E ⊆ V × V
represents the edges. An edge from node j to node i, denoted by (j, i), implies that node j can transmit information to node i. The neighborhood of the i-th node is
defined as Ni , {j | (j, i) ∈ E}. A node j is said to be an
out-going neighbor of node i if (i, j) ∈ E. By an induced
subgraph of G obtained by removing certain nodes C ⊂
V, we refer to the subgraph that has V \ C as its node set
and contains only those edges of E with both end points
in V \C. The notation |V| is used to denote the cardinality
of a set V. The set of all eigenvalues (or modes) of a matrix A is denoted by sp(A) = {λ ∈ C | det(A− λI) = 0},
and the set of all unstable eigenvalues by ΛU (A) = {λ ∈
sp(A) | |λ| ≥ 1}. We use aA (λ) and gA (λ) to denote
the algebraic and geometric multiplicities, respectively,
of an eigenvalue λ ∈ sp(A). An eigenvalue λ is said to
be simple if aA (λ) = gA (λ) = 1. Given a set of matrices {A1 , · · · , An }, we use diag(A1 , · · · , An ) to refer
to a block diagonal matrix with Ai as its i-th block
entry. For a set J = {m1 , · · · , m|J | } ⊆ {1, · · · , N },
h
iT
and a matrix C = CT1 · · · CTN , we define CJ ,
h
iT
CTm1 · · · CTm|J | . The identity matrix of dimension r
is denoted Ir , and N+ is used to refer to the set of all
positive integers. The terms ‘communication graph’ and
‘network’ are used interchangeably. Throughout this paper, we shall use the term ‘resilient’ in the same context
as that used traditionally in the computer science literature to deal with worst-case adversarial attack models [22] (a formal definition of such a model is provided
in the next section).
Contributions: Our contributions are threefold. First,
in Section 3, we characterize certain necessary conditions that need to be satisfied by the sensor measurements and the communication graph for the distributed
state estimation problem to be solvable in the presence
of arbitrary adversarial behavior. Our results hold for
any algorithm and hence identify fundamental limitations that are of both theoretical and practical importance in the design of attack-resilient robust networks.
We also argue that our impossibility results in the distributed setting generalize those existing for centralized
control systems subject to sensor attacks [19, 20].
For the problem under consideration, it is imperative to
understand which (potentially adversarial) neighbors a
given node should listen to, and subsequently, how it
should process the information received from neighbors
it chooses to listen to. Consequently, our second contribution is to develop a distributed filtering algorithm in
Section 4 that enables each uncompromised node to recover the entire state dynamics, provided certain graph
2
2
Definition 2 (f -local set) A set C ⊂ V is f -local if it
contains at most f nodes in the neighborhood of the other
nodes, i.e., |Ni ∩ C| ≤ f, ∀i ∈ V \ C.
System and Attack Model
System Model: Consider the LTI dynamical system
x[k + 1] = Ax[k],
(1)
Definition 3 (f -local and f -total adversarial models) A set A of adversarial nodes is f -locally bounded
(resp., f -totally bounded) if A is an f -local (resp., f total) set.
where k ∈ N is the discrete-time index, x[k] ∈ Rn is the
state vector and A ∈ Rn×n is the system matrix. The
system is monitored by a network G = (V, E) consisting
of N nodes. The i-th node receives a measurement of the
state, given by
yi [k] = Ci x[k],
(2)
ri ×n
ri
. We use C =
where yi [k] ∈ R and Ci ∈ R
h
iT
T
T
to represent the collection of the
C1 · · · CN
individual node observation matrices; accordingly,
h
iT
T
y[k] = y1T [k] · · · yN
represents the cumulative
[k]
In the literature dealing with distributed fault-tolerant
algorithms, it is a common assumption to consider an
f -total adversarial model. However, to allow for a large
number of adversaries in large scale networks, we will allow the adversarial set to be f -local. Summarily, the adversary model considered throughout this paper will be
referred to as an f -locally bounded Byzantine adversary
model. The non-adversarial nodes will be referred to as
regular nodes and be represented by the set R = V \ A.
Finally, note that the actual number and identities of the
adversarial nodes are not known to the regular nodes;
only the upper bound f (on the total number of adversarial nodes in the neighborhood) is known.
measurement vector, i.e., y[k] = Cx[k].
Each node is tasked with estimating the entire system
state x[k] based on information received from its neighbors and its local measurements (if any). As such, we
assume that the pair (A, C) is detectable (this is a necessary condition for solving the distributed state estimation problem even in the absence of adversaries); however, we do not assume that the pair (A, Ci ) is detectable
for any i ∈ V. Two immediate challenges are as follows:
(i) As the pair (A, Ci ) may not be detectable for some
(or all) i ∈ {1, · · · , N }, information exchange is necessary; and (ii) information exchange is restricted by the
underlying communication graph G. In addition to the
above challenges, in this paper, we allow for the possibility that certain nodes in the network are compromised
by an adversary, and do not follow their prescribed state
estimate update rule. We will use the following adversary model in this paper.
Throughout this paper, we shall consider causal, deterministic algorithms; note however that the notions of
causal and deterministic behavior apply only to the regular nodes. With x̂i [k] representing the estimate of x[k]
maintained by node i, the problem studied in this paper
can be formally stated as follows.
Problem 1 (Resilient Distributed State Estimation Problem) Given an LTI system (1), a linear
measurement model (2), and a time-invariant directed
communication graph G, design a set of state estimate update and information exchange rules such that
limk→∞ ||x̂i [k] − x[k]|| = 0, ∀i ∈ R, regardless of the
actions of any f -locally bounded set of Byzantine adversaries.
Adversary Model: We consider a subset A ⊂ V of the
nodes in the network to be adversarial. We assume that
the adversarial nodes are completely aware of the network topology, the system dynamics and the algorithm
employed by the non-adversarial nodes. In terms of capabilities, an adversarial node can leverage the aforementioned information to arbitrarily deviate from the
rules of any prescribed algorithm, while colluding with
other adversaries in the process. Furthermore, following
the Byzantine fault model [22], adversaries are allowed
to send differing state estimates to different neighbors at
the same instant of time. This assumption of omniscient
adversarial behavior is standard in the literature on resilient and secure distributed algorithms [23–26], and allows us to provide guarantees against “worst-case” adversarial behavior. To characterize the threat model in
terms of the number of adversaries in the network, we
will use the following definitions from [27], [28].
The interplay between the measurement structure of the
nodes and the underlying communication graph results
in certain conditions being necessary for solving Problem
1, irrespective of the choice of algorithm. We provide
such conditions in the following section.
3
Fundamental Limitations of any Distributed
State Estimation Algorithm
Intuitively, the network must possess a certain degree
of measurement redundancy as well as redundancy in
its communication structure so as to counter the effect
of adversarial behavior. More specifically, the measurements of the regular nodes must ensure global detectability of the state, and the network structure should prevent the malicious nodes from acting as bottlenecks between correctly functioning nodes. To identify necessary
conditions for resilient distributed state estimation that
Definition 1 (f -total set) A set C ⊂ V is f -total if it
contains at most f nodes in the network, i.e., |C| ≤ f .
3
Execution σ: The initial condition is x[0] = 0. The
nodes in H1 are regular while the nodes in H2 are adversarial. The nodes in H2 pretend that their state estimates
are x̂H2 [k] and that their measurements are CH2 Ak η,
where x̂H2 [k] represents the collection of the state estimates maintained by the nodes in H2 during the execution σ ′ of T . Additionally, at each time-step, the nodes
in H2 perform the exact same actions that they perform
during the execution σ ′ .
capture the above notions of redundancy, we first introduce some terminology.
Definition 4 (Critical Set) A set of nodes F ⊂ V is
said to be a critical set if the pair (A, CV\F ) is not detectable.
Note that detectability of (A, C) implies that a critical
set must necessarily be non-empty.
Execution σ ′ : The initial condition is x[0] = η. The
nodes in H2 are regular while the nodes in H1 are adversarial. The nodes in H1 pretend that their state estimates are x̂H1 [k] and that their measurements are zero,
where x̂H1 [k] represents the collection of the state estimates maintained by the nodes in H1 during the execution σ of T . Additionally, at each time-step, the nodes
in H1 perform the exact same actions that they perform
during the execution σ.
Definition 5 (Minimal Critical Set) A set F ⊂ V is
said to be a minimal critical set if F is a critical set and
no subset of F is a critical set.
Let M = {F1 , · · · , F|M| } denote the set of all minimal
critical sets. With each set Fi ∈ M, we associate a virtual node si as follows. Directed edges are added from si
to each node in Fi and the resulting network is denoted
′
by Gi = (V ∪ si , E ∪ Ei ), where Ei represents the set of
edges from si to Fi .
Since the actions of the adversaries in the two executions
described above are coupled, it becomes important to
establish that such actions are in fact well-defined. To
do so, we argue as follows. Consider the actions of the
adversarial set H2 at time k = 0 of execution σ. Due
to their omniscient nature, these adversaries can anticipate the information that a regular set H2 is supposed
to transmit at time k = 0 of execution σ ′ based on algorithm T . Thus, their actions are well-defined at time
k = 0. An identical argument defines the actions of the
adversarial set H1 at time k = 0 of execution σ ′ . Since
the actions of both the sets H1 and H2 at time k = 0
are well-defined in each of the executions σ and σ ′ , the
response of the regular nodes to such actions (in the respective executions) at time k = 1 can be anticipated
by any adversarial set. Specifically, to generate their actions at time k = 1 of execution σ (resp., execution σ ′ ),
the adversarial set H2 (resp., H1 ) simply simulates execution σ ′ (resp., execution σ) for time k = 0 to figure
out how a regular set H2 (resp., H1 ) would act at time
k = 1 of execution σ ′ (resp., execution σ). Repeating the
above argument reveals that the actions of the respective adversarial sets in each of the executions σ and σ ′
are well-defined at every time step.
Definition 6 (f -local pair and f -total pair cuts
w.r.t. si ) Consider a minimal critical set Fi ∈ M. A
set H ⊂ V is called a cut w.r.t. si if removal of H from
′
′
Gi results in an induced subgraph of Gi whose node set
can be partitioned into two non-empty sets X and Y with
si ∈ X , and no directed paths from X to Y in the induced
subgraph. A cut H w.r.t. si is called an f -local pair cut
(resp., f -total pair cut) w.r.t. si if it can be partitioned
as H = H1 ∪ H2 such that both H1 and H2 are f -local
(resp., f -total) in G.
The following result identifies a fundamental limitation
for f -local adversarial models.
Theorem 1 If there exists an f -local pair cut w.r.t. si
′
in Gi for some minimal critical set Fi ∈ M, then it is
impossible for any algorithm to solve Problem 1.
PROOF. Suppose there exists an f -local pair cut H =
H1 ∪ H2 w.r.t. si for some minimal critical set Fi ∈ M.
For the sake of contradiction, suppose there exists an algorithm T that solves Problem 1 for the given network
G. From the definition of H, we see that Y contains no
elements of Fi . Since Fi is a critical set, it then follows
that the pair (A, CY ) is not detectable. Thus, there exists an initial condition x[0] = η that causes the measurement set yY [k] corresponding to Y to be identically
zero for all time, while the state x[k] remains bounded
away from zero. The idea of the proof will be to demonstrate that the nodes in Y cannot distinguish between
the zero initial condition and the initial condition η under an appropriately constructed attack. To this end,
noting that each of the sets H1 and H2 are f -local and
can hence act as valid adversarial sets, we consider the
following executions σ and σ ′ of T .
Based on the attack described above, it is clear that the
nodes in Y receive the same state estimate and measurement information from the nodes in H in each of the two
executions. Further, their own measurements are identically zero for all time in each of the two executions.
Hence, based on such identical information, it is impossible for the nodes in Y to resolve the difference in the underlying initial conditions via algorithm T . This leads to
the desired contradiction and completes the proof.
Remark 1 Interestingly, the necessary condition presented in the above theorem bears close resemblance to
the necessary condition in [28] for resilient broadcasting
subject to the same f -local Byzantine adversary model
4
of adversaries that can be tolerated when one seeks to
solve Problem 1. Leveraging Theorem 2, we can provide
an upper bound on this number, as follows.
that we consider here. This similarity can be attributed to
the following analogy: viewing the virtual nodes as originators of messages in a broadcasting context, Problem
1 can be interpreted as a version of the resilient broadcasting problem where the regular nodes are required to
agree (asymptotically) on a time-varying message that
captures the state evolution of the system.
Corollary 1 Let k denote the smallest positive integer
such that there exists a k-total pair cut w.r.t. si for some
Fi ∈ M. Then, the total number of adversaries f must
satisfy the inequality f < k for Problem 1 to have a
solution. 1
Our next result provides a necessary condition for an f total (and hence, also an f -local) adversarial model.
Corollary 2 The condition |Fi | ≥ (2f + 1) ∀Fi ∈ M is
necessary for resilient distributed state estimation subject
to the f -local or f -total adversarial model.
Theorem 2 Suppose there exists an algorithm that
solves the variant of Problem 1 corresponding to an f total Byzantine adversary model. Then, the following
equivalent statements are true.
The proof of the above result is straightforward and is
hence omitted here. With the above corollary in hand,
one can gain insights regarding the distribution of certain specific critical sets in the network. To do so, given
(j)
(j)
an eigenvalue λj ∈ ΛU (A), let {ρ1 , · · · , ρgA (λj ) } represent a basis for the null space of (A − λj In ), and let
(j)
(j)
φi = span{ρi }, i ∈ {1, · · · , gA (λj )}. We say that
(j)
(j)
node i can detect the subspace φi if Ci ρi 6= 0. 2 Let
(j)
Wi ⊆ V denote the set of all nodes that can detect
(j)
φi . The next result then readily follows from Corollary
2 and the PBH test [31].
(i) Consider any minimal critical set Fi ∈ M. There exists no f -total pair cut w.r.t. si .
(ii) Consider a node i ∈ V such that (A, Ci ) is not detectable. Let Xi denote the set of all nodes in G that have
directed paths to node i, and consider a set Di ⊂ Xi
such that |Di | ≤ 2f . Let Pi ⊂ Xi represent the set
of nodes that have directed paths to node i in the induced subgraph obtained by removing Di from G. Then,
(A, Ci∪Pi ) is detectable.
The proof of necessity follows similar arguments as in
Theorem 1, while the equivalence between the two conditions stated in Theorem 2 is established in Appendix
A. We now compare our results with certain relevant existing conditions.
(j)
Proposition 1 For each λj ∈ ΛU (A), if Wi ⊂ V,
(j)
where 1 ≤ i ≤ gA (λj ), then |Wi | ≥ (2f + 1) is a necessary condition for resilient distributed state estimation
subject to the f-local or f-total adversarial model.
Remark 2 In [19, 20], the authors showed that for centralized systems subject to f sensor attacks, a necessary condition for estimating the state asymptotically is
that the system should remain detectable after the removal of any 2f sensors. In our present distributed setting, the maximum information about the state that any
given node i can hope to obtain is from the set {i ∪ Xi },
where Xi is defined as in Theorem 2. Thus, the second
part of Theorem 2 generalizes the necessary conditions
in [19, 20]. In [11, 29], the authors established that the
graph-connectivity metric plays a pivotal role in the analysis of fault-tolerant and resilient distributed consensus
algorithms for settings where there are no underlying
state dynamics that need to be estimated. The results
stated in Theorems 1 and 2 differ from those in [11, 29]
since they blend both graph-theoretic and system-theoretic
requirements. Finally, it can be easily shown that when
there are no adversaries, i.e., when f = 0, the conditions
identified in Theorem 2 reduce to the necessary and sufficient condition for distributed state estimation, namely
every source component (strong components with no incoming edges) of the graph should be detectable [2–6].
For systems with distinct eigenvalues, a direct consequence of the above result is the requirement of at least
(2f +1) nodes that can detect each unstable stable eigenvalue of the system. The preceding analysis builds up to
the distributed estimation strategy adopted in this paper. In particular, our approach shall involve identifying
the locally detectable and undetectable eigenvalues associated with a given node, and subsequently devising
separate estimation strategies for the subspaces associated with such eigenvalues. We formalize this idea in the
rest of the paper.
Remark 3 Two important directions of future investigation are (i) finding an efficient algorithm (if one exists) for computing k in Corollary 1 either exactly or approximately, and (ii) determining whether the conditions
stated in Theorem 1 (resp., Theorem 2) are sufficient for
achieving resilient distributed state estimation subject to
1
Similar bounds for static power system models subject to
attacks were obtained in [30].
2
Throughout the paper, for the sake of conciseness, we use
the terminology
“node
i i can detect eigenvalue λj ” to imply
h
A−λj In
= n. Each stable eigenvalue is considthat rank
Ci
ered detectable w.r.t. to the measurements of every node.
We now discuss certain implications of Theorem 2. Given
an LTI system (1), a measurement model specified by
(2), and a communication graph G, it is of both theoretical and practical interest to know the maximum number
5
Definition 7 (Source nodes) For each λj ∈ ΛU (A),
the set of nodes that can detect λj is denoted by Sj , and
called the set of source nodes for λj .
an f -local (resp., f -total) Byzantine adversary model.
We point out that the main source of computational complexity associated with the first point lies in finding all the
minimal critical sets associated with the given system.
4
4.1
We now proceed to develop an estimation scheme that
enables each regular node to estimate z[k] (from which
(j)
they can obtain x[k] = Tz[k]). Accordingly, let ẑi [k]
denote the estimate of z(j) [k] (the portion of z[k] corresponding to the eigenvalue λj 3 ) maintained by node
i ∈ R. For each λj ∈ ΛU (A), our estimation scheme relies on separate strategies for nodes in Sj and V \ Sj .
In particular, each node in Sj employs a Luenberger observer for estimating z(j) [k]. The nodes in V \ Sj , on
the other hand, cannot detect the eigenvalue λj , and
thus rely on a resilient consensus algorithm to estimate
z(j) [k]. In what follows, we discuss these ideas in detail.
Resilient Distributed State Estimation
Preliminaries
For an eigenvalue λ ∈ sp(A), let V(λ) represent a
block diagonal matrix with the Jordan blocks corresponding to λ (in the standard Jordan canonical representation of A) along the main block diagonal. We
begin by recalling certain properties of the real Jordan
canonical form of a square matrix that will be useful for our subsequent development [32]. We first note
that if λ represents a non-real eigenvalue of A and
λ̄ represents its complex-conjugate, then [32, Lemma
3.1.18] ensures that λ and λ̄ have the same Jordan
structure. Next, let λ = a + ib" where
# a, b ∈ R, and
a b
D(a, b) be defined as D(a, b) ,
. Then, the ma−b a
trix diag(V(λ), V(λ̄)) is similar to a real block upper
triangular matrix W(λ) ∈ R2aA (λ)×2aA (λ) which has
aA (λ) 2-by-2 blocks D(a, b) on the main block diagonal
and (aA (λ) − 1) blocks I2 on the block superdiagonal. Henceforth, for a non-real eigenvalue λ ∈ sp(A),
W(λ) will have the meaning discussed above. Let
sp(A) = {{λ1 , λ̄1 }, · · · , {λp , λ̄p }, λp+1 , · · · , λγ } with
the first p pairs representing the non-real eigenvalues,
and λp+1 to λγ representing the real eigenvalues of A.
Then, the real Jordan canonical form theorem [32, Theorem 3.4.1.5] can be stated as follows.
The first step in the estimation process involves the
above common coordinate transformation given by
z[k] = T−1 x[k], to be performed by each regular node
of the graph. As this only requires knowledge of the
system matrix A (which is assumed to be known by all
the nodes), all of the nodes can do this in a distributed
manner (e.g., by using an agreed-upon convention for
ordering the eigenvalues and corresponding eigenvectors). Building on the general theme in [3], we first
present a method that allows a regular node i ∈ R to
estimate the locally detectable portion of the state z[k]
without communicating with neighbors. To this end,
consider the following result.
Lemma 1 Let λj ∈ Oi be a non-real eigenvalue. Let
(j)
C̄i denote the columns of C̄i corresponding to W(λj )
(j)
in (3). Then, the pair (W(λj ), C̄i ) is detectable.
The proof of the above lemma follows readily from
the PBH test and is hence omitted here. Let Oi =
{{λn1 , λ̄n1 }, · · · , {λnpi , λ̄npi }, λnpi +1 , · · · , λnγi }, where
the first pi pairs represent the non-real eigenvalues,
and λnpi +1 to λnγi represent the real eigenvalues of A
that are detectable w.r.t. the measurements of node
i. Let MOi = diag(W(λn1 ), · · · , W(λnpi ), V(λnpi +1 ),
· · · , V(λnγi )). Let COi represent the columns of C̄i
corresponding to the matrix MOi and zOi [k] denote the
portion of the state z[k] corresponding to the detectable
eigenvalues of node i, i.e., corresponding to Oi . Based
on Lemma 1, it is easy to see that the pair (MOi , COi )
is detectable. Thus, a standard Luenberger observer can
be locally constructed by node i for estimating zOi [k].
The specifics of such a construction are straightforward,
and are similar to those in Section VI-A of [4]. We thus
Theorem 3 There exists a real similarity transformation matrix T that transforms the state transition matrix A in (1) to a real block diagonal matrix M given by
M = diag(W(λ1 ), · · · , W(λp ), V(λp+1 ), · · · , V(λγ )).
With T as in the above theorem, and z[k] = T−1 x[k],
the dynamics (1) are transformed into the form
z[k + 1] = Mz[k]
yi [k] = C̄i z[k],
∀i ∈ {1, · · · , N }
(3)
where M = T−1 AT and C̄i = Ci T. For a non-real
eigenvalue pair {λj , λ̄j } ∈ sp(A), let z(j) [k] ∈ R2aA (λj )
represent the portion of the state z[k] associated with
the matrix W(λj ). Similarly, for a real eigenvalue λj ∈
sp(A), z(j) [k] ∈ RaA (λj ) is the portion of the state z[k]
associated with the matrix V(λj ). For each node i, we
denote the detectable and undetectable eigenvalues by
the sets Oi and UO i , respectively. Next, we introduce
the notion of source nodes.
3
Throughout the rest of the paper, the terminology “z(j) [k]
corresponds to the eigenvalue λj ” should be interpreted as
z(j) [k] corresponds to the eigenvalue pair {λj , λ̄j } for a nonreal eigenvalue λj ∈ sp(A).
6
(j)
skip minor details and state the following result which
shall be useful later on.
in hand, node i updates ẑi [k] as follows:
(
(j)
V(λj )z̄i [k], if λj ∈ UO i is real
(j)
ẑi [k+1] =
(j)
W(λj )z̄i [k], if λj ∈ UO i is not real,
(5)
iT
h
(j)
(jσj )
(j1)
where z̄i [k] = z̄i [k], · · · , z̄i
[k] , σj =
aA (λj ) if λj ∈ UO i is real and σj = 2aA (λj ) if
λj ∈ UO i is not real.
Lemma 2 For each regular node i ∈ R and each λj ∈
Oi , a Luenberger observer can be locally constructed by
(j)
node i such that limk→∞ ||ẑi [k] − z(j) [k]|| = 0.
Based on the previous result, we see that a regular node
i can estimate certain portions of the state space without having to exchange information with neighbors. The
challenge, however, lies in estimating the locally undetectable portion of the state in the presence of adversaries. The following section presents a resilient consensus based strategy to address this issue.
4.2
We refer to the above algorithm as the Local-Filtering
based Resilient Estimation (LFRE) algorithm. For implementing this algorithm, a regular node i needs to con(j)
struct the set Ni , ∀λj ∈ UOi , based on the relative positions of its neighbors (with respect to its own position)
(j)
in G. We will provide the exact definition of Ni , and a
distributed algorithm for constructing such a set in the
following sections where we analyze the convergence of
the LFRE algorithm. We conclude this section by commenting on certain features of the LFRE algorithm.
Local-Filtering Based Resilient Estimation
For any λj ∈ sp(A), let z (jm) [k] denote the m-th com(jm)
[k] denote the
ponent of the vector z(j) [k], and let ẑi
estimate of that component maintained by node i ∈ V.
Consider an unstable eigenvalue λj ∈ UO i . For such an
eigenvalue, node i has to rely on the information received
from its neighbors, some of whom might be adversarial,
in order to estimate z(j) [k]. To this end, we propose a
resilient consensus algorithm that requires each regular
node i ∈ V \ Sj to update its estimate of z(j) [k] using
the following two stage filtering strategy:
Remark 4 The rationale behind performing a real Jordan canonical decomposition at every node (as opposed
to a standard Jordan transformation) is to ensure that
the state estimates featuring in equations (4) and (5) are
real at every time-step, thereby making the sorting operation performed in Step 1 of the algorithm meaningful. At
any time-step, if a regular node i either receives a non(j)
real estimate of z (jm) [k] from some node l ∈ Ni or does
not receive an estimate at all, it would immediately identify node l as an adversarial node, and simply assign a
0 value to node l’s estimate of z (jm) [k]. Note that every
(j)
regular node in Ni will always transmit a real estimate
to node i at every time-step.
1) At each time-step k, each regular node i collects the
state estimates of z(j) [k] received from only those
(j)
neighbors that belong to a certain subset Ni ⊆ Ni
(to be defined later). For every component m of
z(j) [k], the estimates of z (jm) [k] received from nodes
(j)
in Ni are sorted from largest to smallest.
2) For each component m of z(j) [k], node i removes the
largest and smallest f estimates (i.e., removes 2f esti(j)
mates in all) of z (jm) [k] received from nodes in Ni ,
and computes the quantity:
(jm)
z̄i
[k] =
X
(jm)
wil
(jm)
[k]ẑl
[k],
Remark 5 The strategy of disregarding the most extreme values in one’s neighborhood, and using a convex
combination of the rest for performing linear scalar updates, has been used for designing resilient distributed
algorithms for consensus [23, 24, 33] and optimization
[25, 26] problems. In this paper, we show that such algorithms can also be used for resilient distributed state estimation, with certain substantial differences arising from
the fact that the nodes are trying to track the state of an
external dynamical system.
(4)
(jm)
l∈Mi
[k]
(jm)
5
(j)
where Mi [k] ⊂ Ni (⊆ Ni ) is the set of nodes from
which node i chooses to accept estimates of z (jm) [k]
at time-step k, after removing the f largest and f
(j)
smallest estimates of z (jm) [k] from Ni . Node i as(jm)
signs the weight wil [k] to the l-th node at the kth time-step for estimating the m-th component of
z(j) [k]. The weights are nonnegative and chosen to satP
(jm)
isfy l∈M(jm) [k] wil [k] = 1, ∀λj ∈ UOi and for each
i
(jm)
component m of z(j) [k]. With the quantities z̄i
Analysis of the Resilient Distributed Estimation Strategy
In this section, we provide our main result concerning the
convergence of the LFRE algorithm. Let ΩU (A) , {λj ∈
ΛU (A)|V \Sj is non-empty}. By this definition, all nodes
are source nodes for each eigenvalue in ΛU (A) \ ΩU (A),
and are hence capable of recovering the corresponding
portions of the state based on locally constructed Luenberger observers (as discussed in Section 4.1). Consequently, the LFRE algorithm specifically applies to only
[k]
7
x[k + 1] = ax[k]
x[k + 1] = ax[k]
1
2
3
1
2
3
4
5
6
4
5
6
7
7
(a)
(b)
property indicates that all its regular neighbors in such
a subgraph belong to levels strictly preceding its own
level. In essence, the edges of the MEDAG Gj represent a
medium for transmitting information securely from the
source nodes Sj to the non-source nodes, by preventing the adversaries from forming a bottleneck between
such nodes. Intuitively, this requires redundant nodes
and edges, and such a requirement is met by the first
property of the MEDAG. In particular, as regards measurement redundancy, it follows from the definition that
for each λj ∈ ΩU (A), a MEDAG Gj contains at least
(2f + 1) source nodes that can detect λj . 4 The LFRE
algorithm described in the previous section relies on a
special information flow pattern that requires a node i to
(j)
listen to only its neighbors in Ni for estimating z(j) [k].
The second property of a MEDAG then indicates that
nodes in level m only use the estimates of regular nodes
in levels 0 to m − 1 for recovering z(j) [k]. The implications of the above properties will become apparent in the
proof of the following result which provides a sufficient
condition for solving Problem 1 based on our approach.
Fig. 1. A scalar unstable plant is monitored by a network
of 7 nodes as depicted by the figure on the left. Nodes 1, 2
and 3 are the source nodes for this system. The figure on the
right represents a subgraph of the original graph satisfying
the properties of a MEDAG in Definition 8 for all 1-local
sets (i.e., with f = 1). For example, when A = {1} (as
shown in the right figure), every non-source node has at least
2f + 1 = 3 neighbors. The levels that partition R = V \ A
are level 0 with nodes 2 and 3, level 1 with nodes 4, 5 and 6,
and level 2 with node 7. Note that each regular node has all
its regular neighbors in levels that are numbered lower than
its own.
Theorem 4 Suppose that G contains a MEDAG Gj for
each λj ∈ ΩU (A). Then, based on the LFRE dynamics described by equations (4) and (5), each regular node
i ∈ R can asymptotically estimate the state of the plant,
despite the actions of any f -locally bounded set of Byzantine adversaries.
those eigenvalues that belong to ΩU (A). Consider the
following definition.
Definition 8 (Mode Estimation Directed Acyclic
Graph (MEDAG)) Consider an eigenvalue λj ∈
ΩU (A). Suppose there exists a spanning subgraph
Gj = (V, Ej ) of G with the following properties for all
f -local sets A and R = V \ A.
(j)
(i) If i ∈ {V \ Sj } ∩ R, then |Ni | ≥ 2f + 1, where
(j)
Ni = {l|(l, i) ∈ Ej } represents the neighborhood of
node i in Gj .
(j)
(j)
(ii) There exists a partition of R into the sets {L0 , · · · , LTj },
(j)
(j)
where Tj ∈ N+ , L0 = Sj ∩ R, and if i ∈ Lq (where
Sq−1 (j)
(j)
1 ≤ q ≤ Tj ), then Ni ∩ R ⊆ r=0 Lr . Further(j)
(j)
more, Ni = ∅, ∀i ∈ L0 .
Then, we call Gj a Mode Estimation Directed Acyclic
Graph (MEDAG) for λj ∈ ΩU (A).
An example of a MEDAG is shown in Figure 1. The “for
all A” in the definition accounts for the fact that the
set of adversarial nodes during the process of state estimation is unknown, and hence can be any f -local set
(j)
(j)
of V. Note that Tj and the levels L0 to LTj can vary
across different f -local sets. For a given f -local set A,
(j)
we say a regular node i ∈ Lm “belongs to level m”,
where the levels are indicative of the distances of the
regular nodes from the source set Sj . The first property
indicates that every regular node i ∈ V \ Sj has at least
(2f + 1) neighbors in the subgraph Gj , while the second
The proof of the above theorem is given in Appendix
B. Notice that Theorem 4 hinges on the existence of
a MEDAG Gj , for each λj ∈ ΩU (A); in the following
section we describe an approach for checking whether a
given graph G contains such MEDAGs.
6
Checking the Existence of a MEDAG
From the foregoing discussion, it is apparent that the
MEDAGs described in Definition 8 play a key role in
solving Problem 1 based on our proposed technique. In
particular, recall that for each λj ∈ ΩU (A), the LFRE
algorithm described in Section 4.2 requires a regular
node i ∈ V \Sj to accept estimates from only its neighbor
(j)
set Ni in the MEDAG Gj for estimating z(j) [k]. With
these points in mind, our immediate goal in this section will be to develop a distributed algorithm, namely
Algorithm 1, that constructs a MEDAG Gj for each
λj ∈ ΩU (A), and in the process enables each regular
(j)
node i to determine the set Ni for each λj ∈ UO i . The
construction of these MEDAGs constitutes the initialization phase of our design, which can then be followed
up by the LFRE algorithm described earlier. We briefly
describe the implementation of Algorithm 1 as follows.
4
Recall from the discussion immediately following Proposition 1 that such a condition is in fact necessary for systems
with distinct eigenvalues.
8
Algorithm 1 MEDAG Construction Algorithm
the termination of the MEDAG construction algorithm
under arbitrary adversarial behavior. For the following
discussion, we characterize the properties of the output
of Algorithm 1 if it terminates. To this end, consider
the spanning subgraph Gj = (V, Ej ) induced by the sets
(j)
{Ni } returned by Algorithm 1. Keeping in mind that
′
R ⊇ R represents the set of nodes that behave regularly during the execution of Algorithm 1, we have the
following results (the proofs of these results are omitted
due to space constraints, and can be found in [21]).
For each eigenvalue λj ∈ ΩU (A) do:
(j)
Initialization: Initialize ci (j) = 0, Ni = ∅, ∀i ∈ V.
Each node determines whether it belongs to Sj .
Actions of the source nodes: Each node in Sj updates its counter value ci (j) = 1, and transmits the
message “1” to its out-going neighbors. Following this
(j)
step, it does not listen to any other node, i.e., Ni = ∅
and ci (j) = 1, ∀i ∈ Sj for the remainder of the algorithm.
Actions of the non-source nodes: Each node i ∈
V \ Sj does the following:
• If ci (j) = 0 and node i has received “1” from at least
(2f +1) distinct neighbors (not necessarily all in the
same time-step), it updates ci (j) to 1, appends the
labels of the neighbors from which it received “1” to
(j)
Ni , and transmits “1” to its out-going neighbors.
• If ci (j) = 1, it discards all messages received from
(j)
its neighbors, i.e., it does not update ci (j) or Ni .
(j)
Return : A set of sets {Ni }, λj ∈ ΩU (A), i ∈ V.
Proposition 2 Suppose Algorithm 1 terminates for
(j)
some λj ∈ ΩU (A), and returns the sets {Ni }. Then,
the spanning subgraph Gj induced by these sets contains
no directed cycles where every node belongs to R′ .
(j)
Let Lm−1 denote the set of all nodes in R′ that update
their counter value from 0 to 1 in round m of Algorithm
(j)
1, i.e., L0 = Sj ∩ R′ , and so on. 5
Proposition 3 Suppose Algorithm 1 terminates for
some λj ∈ ΩU (A). Let Tj denote the smallest integer
such that in round Tj , ci (j) = 1 ∀i ∈ R′ . Then, the sets
(j)
(j)
{L0 , · · · , LTj } form a partition of the set R′ in Gj .
Algorithm 1 requires each node i to maintain a counter
(j)
ci (j) and a list of indices Ni for each λj ∈ ΩU (A). The
(j)
nodes in Ni ⊆ Ni will be the parents of node i in the
DAG constructed for the estimation of z(j) [k]. Algorithm
(j)
1 is initialized with ci (j) = 0 and Ni = ∅, for each
i ∈ V. Subsequently, the algorithm proceeds in rounds
where in the first round each node in Sj broadcasts the
message “1” to its out-going neighbors, sets ci (j) = 1,
(j)
maintains Ni = ∅ for all future rounds, and goes to
sleep. Each node i ∈ V \ Sj waits until it has received
“1” from at least (2f + 1) distinct neighbors, at which
point it sets ci (j) = 1, appends the labels of each of the
(j)
neighbors from which it received “1” to Ni , broadcasts
the message “1” to its out-going neighbors, and goes to
sleep. Let R′ ⊆ V denote the set of nodes that behave
regularly during the execution of Algorithm 1. We say
that the MEDAG construction algorithm “terminates
for λj ” if there exists Tj ∈ N+ such that ci (j) = 1 ∀i ∈
R′ , for all rounds following round Tj . The objective of
(j)
the algorithm is to return a set of sets {Ni }, where
λj ∈ ΩU (A), and i ∈ V.
Theorem 5 Suppose the MEDAG construction algorithm terminates for λj ∈ ΩU (A). Then, there exists a
subgraph Gj satisfying the properties of a MEDAG for
all f -local sets A that contain V \ R′ as a subset.
PROOF. The result follows from Propositions 2 and
3.
Remark 6 Based on the above theorem, we make the following observations. If Algorithm 1 terminates for each
λj ∈ ΩU (A), and V \ R′ = ∅, then the Gj subgraphs satisfy all the properties of a MEDAG and we can directly
invoke Theorem 4. If Algorithm 1 terminates for each
λj ∈ ΩU (A) and V \ R′ 6= ∅, (i.e., there is some adversarial activity during the MEDAG construction phase),
then we do not need to provide any guarantees of state
estimation for the set of misbehaving nodes V \ R′ , since
V \ R′ ⊆ A. In this case too, the subgraphs returned
by Algorithm 1 have enough redundancy to ensure that
Problem 1 can be solved based on our proposed approach;
this fact can be established using arguments identical to
those used for proving Theorem 4. In what follows, we
summarize our overall approach.
We emphasize that in addition to misbehavior during the
state estimation phase (run-time), an adversarial node
is allowed to misbehave during the implementation of
Algorithm 1 (design-time) as well. For example, it can
transmit the message out of turn, i.e., before receiving
“1” from at least (2f + 1) neighbors. It can also choose
not to transmit the message at all. Note however that
we must have V \ R′ ⊆ A, i.e., the f -local set of adversaries during the estimation phase must contain the
set of adversaries during the design phase. In the next
section, we shall detail graph conditions that guarantee
6.1
Summary of the Resilient Distributed State Estimation Scheme
1) Each regular node i ∈ R performs the coordinate
transformation z[k] = T−1 x[k] described in Section
5
Note that the method developed in this paper allows even
some of the source nodes in Sj to be adversarial.
9
the set of source nodes Sj . Suppose that the MEDAG
construction algorithm for λj does not terminate. Let
C ⊆ V \ Sj be the set of regular nodes that never update
their counter ci (j) from 0 to 1, where i ∈ C. As G is
strongly (3f + 1)-robust w.r.t. Sj , it follows that C is
(3f + 1)-reachable, i.e., there exists a node i ∈ C which
has at least (3f + 1) neighbors outside C. Under the
f -local adversarial model, at least (2f + 1) of them are
regular nodes with ci (j) = 1. Thus, at least (2f + 1)
regular nodes must have transmitted the message “1”
to node i. Thus, based on the rules of Algorithm 1, node
i must have updated ci (j) from 0 to 1 at some point of
time, leading to a contradiction.
4.1; accordingly, it identifies its detectable and undetectable eigenvalues (Oi and UO i respectively).
2) The MEDAG construction algorithm described by Algorithm 1 is implemented for each λj ∈ ΩU (A); graph
conditions for termination of this algorithm are provided in the next section. At the end of this algorithm,
(j)
each regular node i knows the subset Ni of neighbors it should use in the LFRE algorithm.
3) Each regular node i employs a locally constructed
Luenberger observer (refer to Lemma 2 and the discussion preceding it) for estimating zOi [k], namely
the portion of the state z[k] corresponding to its detectable eigenvalues.
4) Each regular node i employs the LFRE algorithm governed by equations (4) and (5) for estimating zU Oi [k],
namely the portion of the state z[k] corresponding to
its undetectable eigenvalues.
Remark 8 Whereas the (2f + 1) term appears in various contexts when dealing with security problems on networks (such as distributed consensus, broadcasting and
optimization), the (3f + 1) term featuring in our analysis accounts for misbehavior that involves transmission
of no messages by the adversarial nodes during execution of the MEDAG construction algorithm described in
Section 6.
Remark 7 Whereas steps 1 and 2 correspond to the initial design phase of our scheme, steps 3 and 4 constitute the estimation phase. A key benefit of the proposed
method is that if certain graph-theoretic conditions (to be
discussed in the following section) are met, then our overall scheme provably admits a fully distributed implementation even under worst-case adversarial behavior.
7
We now present the main result of this paper which ties
together the previous results presented in this paper,
and in turn provides a connection between feasible graph
topologies and the solution to Problem 1 based on our
proposed approach.
Feasible Graph Topologies
In this section, we characterize feasible graph topologies that guarantee the termination of the MEDAG construction algorithm described in the previous section. In
other words, based on Remark 6, feasible graph topologies guarantee that Problem 1 can be solved based on
our proposed approach (summarized in Section 6.1). We
first recall the following definition from [24, 33].
Theorem 6 Consider an LTI system (1) and a measurement model (2). Let the communication graph G be
strongly (3f + 1)-robust w.r.t. Sj , ∀λj ∈ ΩU (A). Then,
the proposed algorithm summarized in Section 6.1 provides a solution to Problem 1.
PROOF. From Lemma 3, it follows that if G is strongly
(3f + 1)-robust w.r.t. Sj for every λj ∈ ΩU (A), then
the MEDAG construction algorithm terminates for each
such eigenvalue. Combining Theorem 5, Remark 6 and
Theorem 4 then leads to the desired result.
Definition 9 (r-reachable set) For a graph G =
(V, E), a set S ⊂ V, and an integer r ∈ N+ , S is
an r-reachable set if there exists an i ∈ S such that
|Ni \ S| ≥ r,
Thus, if a set S is r-reachable, then it contains a node
which has at least r neighbors outside S. We modify the
notion of a strongly-r robust graph from [33] as follows.
If the adversarial attacks are restricted to the estimation
phase only, i.e., if there are no attacks during the initial
MEDAG construction phase, then the following result
provides a tight graph condition for our algorithm.
Definition 10 (strongly r-robust graph w.r.t. Sj )
For r ∈ N+ and λj ∈ ΩU (A), a graph G = (V, E) is
strongly r-robust w.r.t. to the set of source nodes Sj , if
for any non-empty subset C ⊆ V \ Sj , C is r-reachable.
Lemma 3 The MEDAG construction algorithm terminates for λj ∈ ΩU (A) if G is strongly (3f + 1)-robust
w.r.t. Sj .
Theorem 7 Consider an LTI system (1) and a measurement model (2). Suppose adversarial behavior is restricted to the estimation phase (steps 3 and 4) of the proposed algorithm summarized in Section 6.1. Then, this
algorithm solves Problem 1 if and only if G is strongly
(2f + 1)-robust w.r.t. Sj , ∀λj ∈ ΩU (A).
PROOF. We prove by contradiction. Consider any
λj ∈ ΩU (A) and let G be strongly (3f + 1)-robust w.r.t.
The proof of the above theorem is presented in Appendix
C. Essentially, Theorem 7 alludes to the fact that G contains a MEDAG Gj for each λj ∈ ΩU (A) if and only if
10
G is strongly (2f + 1)-robust w.r.t. Sj , ∀λj ∈ ΩU (A). It
should be noted that Theorem 7 provides a graph condition that is necessary and sufficient for the algorithm
developed in this paper. However, such a condition may
not be necessary for solving Problem 1 in general. Investigation of an algorithm-independent tight graph condition that solves Problem 1 is an interesting avenue of
future research.
non-source node owing to the r-feasibility of the original
network. As r > k, such sets are not k-local. Next, pick
any λj ∈ ΩU (A) and let C be a non-empty subset of
′
′
′
′
V \ Sj , where V and Sj represent the vertex set and
′
source node set for λj , respectively, in G . Since C was rreachable in G, it contained some node v with r neighbors
′
outside C. While constructing G , node v can lose at most
k of such neighbors, and hence C is (r − k)-reachable in
′
G . The rest of the proof follows trivially.
Theorems 6 and 7 reveal that ‘strong r-robustness w.r.t.
Sj , ∀λj ∈ ΩU (A)’ is the key topological property required for guaranteeing success of our proposed algorithm. Accordingly, given a system model (1) and measurement model (2), a network that is strongly r-robust
w.r.t. Sj , ∀λj ∈ ΩU (A), will be called an ‘r-feasible network ’ for simplicity. We summarize certain features of
an r-feasible network in the following result.
We remark on certain implications of the above result.
The first property provides a procedure for constructing
r-feasible networks with N nodes (where N > r) starting
from r-feasible networks with fewer than N nodes. The
second property shows that the measurement structure
of the nodes provides an upper bound on the robustness
of the overall network. The third property places a lower
bound on the minimum in-degree of any node that cannot estimate the entire state on its own in an r-feasible
network. Finally, a direct implication of the fourth property is that a loss of k source nodes (where k < r) for any
unstable eigenvalue of the system (possibly due to sensor failures) leaves the resulting network at least (r − k)feasible if the original network is r-feasible to begin with.
Proposition 4 An r-feasible network G has the following properties.
′
(i) The graph G = (V ∪ vnew , E ∪ Enew ), where vnew is a
new vertex added to G and Enew is the edge set associated with vnew , is an r-feasible network if |Nvnew | ≥ r.
(ii) r ≤ minλT
|Sj |.
j ∈ΩU (A)
(iii) Let S = λj ∈ΩU (A) Sj . Then |Ni | ≥ r, ∀i ∈ V \ S.
(iv) Removal of a k-local set from G, where 0 < k < r,
′
results in a network G that is (r − k)-feasible.
Applicability of the Proposed Approach: Building
on the insights developed in this section, we make a case
for the applicability of the approach developed in this paper by addressing the following question: how efficiently
can one verify whether a given network is r-feasible? To
answer the above question we will exploit a connection
between the ‘strong r-robustness property w.r.t. a certain set of nodes’ and the dynamic process of ‘bootstrap
percolation’ on networks [34]. Given a graph G and a
threshold r ≥ 2, bootstrap percolation can be viewed
as a process of spread of activation where one starts off
with a set I ⊆ V of initially active nodes. Subsequently,
the process evolves over the network based on the rule
that an inactive node becomes active if and only if it has
at least r active neighbors, with active nodes remaining active forever. The process terminates when no more
nodes become active; an initial set I is said to percolate
if upon termination the final active set equals the entire node set V. Consider the following simple, yet key
observation.
PROOF. (i) Consider any λj ∈ ΩU (A). If vnew is
′
a source node for λj , then it is easily seen that G is
strongly r-robust w.r.t. Sj . For the case when vnew is
not a source node for λj , consider any non-empty set
C ⊆ {V ∪ vnew } \ Sj . If C = {vnew }, then r-reachability
of C follows from the fact that |Nvnew | ≥ r. In every
other case, C contains some nodes of the original graph
′
G and is hence r-reachable as G is r-feasible. Thus G is
strongly r-robust w.r.t. Sj . A similar analysis holds for
each λj ∈ ΩU (A), leading to the desired result.
(ii) Suppose r > |Sρ | where ρ = arg minλj ∈ΩU (A) |Sj |.
Since the set C = V \ Sρ can be at most |Sρ |-reachable
and |Sρ | < r, it follows that G is not r-feasible.
(iii) Suppose i ∈ V \ S with |Ni | < r. As i ∈ V \ S, there
exists some λj ∈ ΩU (A) such that i ∈ V \ Sj . Consider
the set C = {i}. As |Ni | < r, the set C is not r-reachable.
Thus, G is not strongly r-robust w.r.t. Sj , implying that
G is not r-feasible.
Lemma 4 Given a graph G and a threshold r ≥ 2, an
initial set I percolates via the process of bootstrap percolation if and only if G is strongly r-robust w.r.t. I.
The proof of the above result follows similar arguments
as Lemma 3, and is hence omitted. Leveraging the above
equivalence, the next result establishes that one can verify whether a given network is r-feasible in polynomial
time (in the size of the system and the network).
(iv) First, observe that as G is r-feasible and k < r,
removal of a k-local set from G cannot cause the removal
of an entire source node net Sj for any λj ∈ ΩU (A).
This follows from noting that any source set Sj where
λj ∈ ΩU (A) (or any set containing Sj ) will have an
overlap of at least r nodes with the neighborhood of some
Proposition 5 Given a system matrix A ∈ Rn×n (1),
11
a measurement model (2), a communication graph G =
(V, E) with |V| = N , and an integer r ≥ 2, one can verify
whether the network is r-feasible in O(nN |E|) time.
graph conditions can be checked in polynomial time as
discussed in the previous section.
There are various interesting directions of future research, some of which were pointed out in Remark 3.
Finding an algorithm-independent necessary and sufficient condition for the problem posed in this paper will
likely be a challenging proposition. Whereas the focus
of this paper has been on obtaining sufficient graphtheoretic conditions that account for worst-case adversarial behavior, it might be of interest to see if such conditions can be relaxed when confronted with less sophisticated adversarial attacks. Extensions of our framework
to account for network-induced issues such as packetdrops, delays and asynchronicity also merit attention.
Preliminary investigations along this line have been conducted in [35].
PROOF. Notice that |ΩU (A)| ≤ n, i.e., there are
at most n source sets Sj for which we need to verify the strong r-robustness property in Definition 10.
Based on Lemma 4, for each Sj corresponding to some
λj ∈ ΩU (A), verifying whether G is strongly r-robust
w.r.t. Sj is equivalent to verifying whether Sj percolates
via the process of bootstrap percolation with threshold r. Thus, we analyze the complexity of simulating
a bootstrap percolation process on a given network. 6
First, notice that it takes at most N iterations/rounds
for a bootstrap percolation process to terminate on a
network of N nodes. In each round, every inactive node
checks whether it has at least r active neighbors; the enP
tire process of checking is thus completed in O ( N
i=1 di )
= O (|E|) time, where di represents the in-degree of node
i. Thus, for a given initial set, it takes O (N |E|) time to
simulate the bootstrap percolation process. The result
then follows readily.
References
[1] Deborah Estrin, Ramesh Govindan, John Heidemann,
and Satish Kumar.
Next century challenges: Scalable
coordination in sensor networks. In Proceedings of the
5th annual ACM/IEEE international conference on Mobile
computing and networking, pages 263–270. ACM, 1999.
Remark 9 Based on the above result, one can check
whether the approach developed in this paper is applicable
for a given system and network in polynomial time. Interestingly, leveraging the equivalence described in Lemma
4, it is possible to show that the strong r-robustness property described in Definition 10 is exhibited by various
large-scale complex network models such as the BarabásiAlbert (BA) preferential attachment model, the ErdősRényi random graph model, and the 2-dimensional random geometric graph model. A detailed discussion of this
topic can be found in Appendix D.
8
[2] Shinkyu Park and Nuno C Martins. Design of distributed
lti observers for state omniscience. IEEE Transactions on
Automatic Control, 62(2):561–576, 2017.
[3] Aritra Mitra and Shreyas Sundaram.
An approach
for distributed state estimation of LTI systems.
In
Proceedings of the 2016 54th Annual Allerton Conference on
Communication, Control, and Computing, pages 1088–1093,
2016.
[4] A. Mitra and S. Sundaram. Distributed observers for LTI
systems. IEEE Transactions on Automatic Control, 2018.
[5] L Wang and AS Morse. A distributed observer for a timeinvariant linear system. IEEE Transactions on Automatic
Control, 2017.
[6] Weixin Han, Harry L Trentelman, Zhenhua Wang, and
Yi Shen. A simple approach to distributed observer design
for linear systems. arXiv preprint arXiv:1708.01459, 2017.
Conclusions
We studied the problem of collaboratively estimating the
state of an LTI system subject to worst-case adversarial
behavior. For the attack models under consideration, we
identified certain necessary conditions that need to be
satisfied by any system and network for the problem
posed in this paper to have a feasible solution. We then
developed a local-filtering algorithm to enable each noncompromised node to recover the entire state. Finally,
using a topological property called strong r-robustness,
we characterized networks that guarantee success of our
proposed strategy. Two notable features of our approach
are as follows: (i) each step of our approach admits an
attack-resilient, completely distributed implementation
provided certain graph conditions are met; and (ii) these
[7] Usman Khan and José MF Moura. Distributing the Kalman
filter for large-scale systems. IEEE Transactions on Signal
Processing, 56(10):4919–4935, 2008.
[8] Reza Olfati-Saber. Distributed Kalman filter with embedded
consensus filters. In Proceedings of the 44th IEEE Conference
on Decision and Control and European Control Conference,
pages 8179–8184, 2005.
[9] Ion Matei and John S Baras. Consensus-based linear
distributed filtering. Automatica, 48(8):1776–1782, 2012.
[10] Fabio Pasqualetti, Florian Dörfler, and Francesco Bullo.
Attack detection and identification in cyber-physical systems.
IEEE Transactions on Automatic Control, 58(11):2715–2729,
2013.
[11] Shreyas Sundaram and Christoforos N Hadjicostis.
Distributed function calculation via linear iterative strategies
in the presence of malicious agents. IEEE Transactions on
Automatic Control, 56(7):1495–1508, 2011.
6
Algorithm 1 essentially simulates the evolution of a bootstrap percolation process with threshold r = (2f + 1), provided there is no adversarial activity during the distributed
implementation of such an algorithm.
[12] Cheng-Zong Bai, Fabio Pasqualetti, and Vijay Gupta. Datainjection attacks in stochastic control systems: Detectability
and performance tradeoffs. Automatica, 82:251–260, 2017.
12
[29] Fabio Pasqualetti, Antonio Bicchi, and Francesco Bullo.
Consensus computation in unreliable networks: A system
theoretic approach.
IEEE Transactions on Automatic
Control, 57(1):90–104, 2012.
[13] André Teixeira, Iman Shames, Henrik Sandberg, and
Karl Henrik Johansson. A secure control framework for
resource-limited adversaries. Automatica, 51:135–148, 2015.
[14] Miroslav Pajic, Insup Lee, and George J Pappas. Attackresilient state estimation for noisy dynamical systems. IEEE
Transactions on Control of Network Systems, 4(1):82–92,
2017.
[30] Oliver Kosut, Liyan Jia, Robert J Thomas, and Lang Tong.
Malicious data attacks on the smart grid. IEEE Transactions
on Smart Grid, 2(4):645–658, 2011.
[31] Joao P Hespanha.
Linear systems theory.
university press, 2009.
[15] Shaunak Mishra, Yasser Shoukry, Nikhil Karamchandani,
Suhas N Diggavi, and Paulo Tabuada.
Secure state
estimation against sensor attacks in the presence of noise.
IEEE Transactions on Control of Network Systems, 4(1):49–
59, 2017.
Princeton
[32] Roger A Horn and Charles R Johnson. Matrix analysis.
Cambridge university press, 2012.
[33] Haotian Zhang and Shreyas Sundaram.
Robustness
of information diffusion algorithms to locally bounded
adversaries.
In Proceedings of the American Control
Conference, pages 5855–5861, 2012.
[16] Ion Matei, John S Baras, and Vijay Srinivasan. Trust-based
multi-agent filtering for increased smart grid security. In
Proceedings of the Mediterranean Conference on Control &
Automation, pages 716–721, 2012.
[34] Svante Janson, Tomasz Luczak, Tatyana Turova, Thomas
Vallier, et al. Bootstrap percolation on the random graph
GN,P . The Annals of Applied Probability, 22(5):1989–2047,
2012.
[17] Usman Khan and Aleksandar M Stankovic.
Secure
distributed estimation in cyber-physical systems.
In
Proceedings of the IEEE International Conference on
Acoustics, Speech and Signal Processing, pages 5209–5213,
2013.
[35] Aritra Mitra and Shreyas Sundaram. Secure distributed state
estimation of an LTI system over time-varying networks and
analog erasure channels. In Proceedings of the American
Control Conference, to appear.
[18] Mohammad Deghat, Valery Ugrinovskii, Iman Shames, and
Cédric Langbort. Detection of biasing attacks on distributed
estimation networks. In Proceedings of the IEEE Conference
on Decision and Control, pages 2134–2139, 2016.
[36] Réka Albert and Albert-László Barabási.
Statistical
mechanics of complex networks. Reviews of modern physics,
74(1):47, 2002.
[19] Hamza Fawzi, Paulo Tabuada, and Suhas Diggavi. Secure
estimation and control for cyber-physical systems under
adversarial attacks.
IEEE Transactions on Automatic
Control, 59(6):1454–1467, 2014.
[37] Paul Erdős and Alfréd Rényi.
On the strength of
connectedness of a random graph.
Acta Mathematica
Academiae Scientiarum Hungarica, 12(1-2):261–267, 1964.
[20] Michelle S Chong, Masashi Wakaiki, and Joao P Hespanha.
Observability of linear systems under adversarial attacks.
In Proceedings of the American Control Conference, pages
2439–2444, 2015.
[38] Milan Bradonjić and Iraj Saniee. Bootstrap percolation on
random geometric graphs. Probability in the Engineering and
Informational Sciences, 28(2):169–181, 2014.
[39] Gregory J Pottie and William J Kaiser. Wireless integrated
network sensors. Communications of the ACM, 43(5):51–58,
2000.
[21] Aritra Mitra and Shreyas Sundaram. Secure distributed
observers for a class of linear time invariant systems in the
presence of byzantine adversaries. In Proceedings of the 55th
IEEE Conference on Decision and Control, pages 2709–2714,
2016.
A
[22] Danny Dolev, Nancy A Lynch, Shlomit S Pinter, Eugene W
Stark, and William E Weihl.
Reaching approximate
agreement in the presence of faults. Journal of the ACM
(JACM), 33(3):499–516, 1986.
Proof of Theorem 2
PROOF. “(i)=⇒(ii)” We prove by contraposition.
Suppose statement (ii) is violated for some node i ∈ V,
i.e., there exists a set Di such that its removal from
G causes the pair (A, Ci∪Pi ) to become undetectable
(where Di and Pi have the same meaning as in the statement of Theorem 2). It then follows that F = V \{i∪Pi }
is a critical set. Suppose it is also a minimal critical set.
′
We construct G by adding directed edges from a virtual
node s to each node in F . 7 Observe that H = Di satisfies all the properties of an f -total pair cut w.r.t. s. In
particular, Y = {i ∪ Pi } and X = {V \ {Di ∪ Y}} ∪ {s}.
Thus, statement (i) is violated. A similar argument
holds when F contains a minimal critical set.
[23] Nitin H Vaidya, Lewis Tseng, and Guanfeng Liang. Iterative
approximate byzantine consensus in arbitrary directed
graphs. In Proceedings of the ACM symposium on Principles
of distributed computing, pages 365–374, 2012.
[24] Heath J LeBlanc, Haotian Zhang, Xenofon Koutsoukos,
and Suresh Sundaram.
Resilient asymptotic consensus
in robust networks. IEEE Journal on Selected Areas in
Communications, 31(4):766–781, 2013.
[25] Shreyas Sundaram and Bahman Gharesifard. Consensusbased distributed optimization with malicious nodes.
In Proceedings of the Annual Allerton Conference on
Communication, Control and Computing, 2015.
[26] Lili Su and Nitin Vaidya.
Byzantine multi-agent
optimization: Part i. arXiv preprint arXiv:1506.04681, 2015.
“(i)⇐=(ii)” We again prove by contraposition. Suppose
statement (i) is violated, i.e., there exists an f -total pair
[27] Chiu-Yuen Koo. Broadcast in radio networks tolerating
byzantine adversarial behavior. In Proceedings of the twentythird annual ACM symposium on Principles of distributed
computing, pages 275–282. ACM, 2004.
7
′
Throughout this proof, we drop the subscript on G , F and
s, unlike the notation in Section 3. This is done since the
subscript i is used to denote sets defined w.r.t. a node i in
this proof.
[28] Andrzej Pelc and David Peleg. Broadcasting with locally
bounded byzantine faults. Information Processing Letters,
93(3):109–115, 2005.
13
cut H w.r.t. a virtual node s corresponding to some minimal critical set F . Consider a node i in Y (recall that Y
is non-empty based on Definition 6). First consider the
case when node i is not reachable from any node in H
in the graph G. It then follows that in the graph G, directed paths to node i can only exist from the set Y. But
since i ∈ Y and (A, CY ) is not detectable, it is trivially
impossible for node i to estimate the state. We thus focus on the case where node i is reachable from a certain
set of nodes, say Di , within the set H. Since |H| ≤ 2f
and Di ⊆ H, we have that |Di | ≤ 2f . It can be easily argued that the removal of Di from G results in an induced
subgraph where node i can only be reached from the set
Y. In other words, the set Pi , as defined in the statement of Theorem 2, is a subset of Y. As (A, CY ) is not
detectable, it follows that (A, Ci∪Pi ) is not detectable
either, and thus statement (ii) is violated.
B
using (4), we obtain
P
(j)
ei [k + 1] = W(λj )
P
|
(jσj )
l∈Mi
[k]
(jσj )
[k]
(j1)
wil [k]el
..
.
wil
{z
[k]
(jσj )
[k]el
[k]
,
}
(j)
ēi [k]
(B.1)
where σj = 2aA (λj ) (since λj is non-real). For arriving
P
(jm)
at (B.1), we used the fact that l∈M(jm) [k] wil [k] = 1
i
for every component m of z(j) [k]. We now analyze the
error dynamics (B.1). To this end, for each component
(j)
m of the vector z(j) [k], we partition the set Ni into
(jm)
(jm)
(jm)
[k], and Mi [k], such that the
[k], Ji
the sets Ui
(jm)
(jm)
[k] contain f nodes each, with
[k] and Ji
sets Ui
the highest and lowest estimate values for z (jm) [k] respectively, transmitted to node i at time-step k, and
(j)
(jm)
Mi [k] contains the rest of the nodes in Ni . According to the update rule (4), node i only uses esti(jm)
mates from the set Mi [k] (which is non-empty at
all time-steps based on the properties of a MEDAG)
(jm)
[k]. Now, for any compoto compute the quantity z̄i
(j)
nent m of z [k], consider the following two cases. (i)
(jm)
Mi [k] ∩ A = ∅, i.e., there are no adversarial nodes
(jm)
in the set Mi [k]: in this case, all the nodes in the set
(j)
(j)
(jm)
Mi [k] are regular and belong to L0 (as Ni ∩ R ⊆
(j)
(jm)
L0 ). (ii) Mi [k] ∩ A is non-empty, i.e., there are
(jm)
some adversarial nodes in the set Mi [k]: based on
the f -local adversarial model, it is apparent that each
(jm)
(jm)
[k] contain at least one
[k] and Ji
of the sets Ui
(j)
regular node belonging to L0 . Let u and v be two such
(jm)
(jm)
[k], re[k] and Ji
regular nodes belonging to Ui
(jm)
[k],
spectively. Based on the definitions of the sets Ui
(jm)
(jm)
(jm)
(jm)
[k] ≤
[k], and Mi [k], we have ẑv [k] ≤ ẑl
Ji
(jm)
(jm)
(jm)
(jm)
ẑu [k], and hence ev [k] ≤ el [k] ≤ eu [k], for
(j)
(jm)
every node l ∈ Mi [k]. In particular, since u, v ∈ L0 ,
(jm)
(jm)
it follows that for any node l ∈ Mi [k], emin [k] ≤
(jm)
(jm)
(jm)
(jm)
el [k] ≤ emax [k], where emin [k] = minu∈L(j) eu [k]
Proof of Theorem 4
PROOF. Let A be the (unknown) set of f -local adversaries, and consider R = V \ A. Given a node i ∈ R,
the state vector z[k] can be partitioned into the components zOi [k] and zU Oi [k] that correspond to the detectable and undetectable eigenvalues, respectively, of
node i. Based on Lemma 2, we know that node i can
estimate zOi [k] asymptotically via a locally constructed
Luenberger observer. It remains to show that node i can
recover zU Oi [k], or in other words, for each λj ∈ UO i ,
(j)
we need to prove that limk→∞ ||ẑi [k] − z(j) [k]|| = 0.
To this end, consider a non-real eigenvalue λj ∈ ΩU (A).
As G contains a MEDAG for each λj ∈ ΩU (A), the sets
(j)
(j)
(j)
(j)
{L0 , L1 , · · · , Lq , · · · LTj } form a partition of the set
R. We prove that each node in R can asymptotically
estimate z(j) [k] by inducting on the level number q.
(j)
(j1)
(j1)
l∈Mi
(j)
For q = 0, by definition of the set L0 , all nodes in L0
are regular and belong to the set Sj , i.e., λj ∈ Oi for each
(j)
node i in L0 . Thus, by Lemma 2, each node in level 0
can estimate z(j) [k] asymptotically. Notice that for any
node i belonging to a level q, where 1 ≤ q ≤ Tj , we have
(j)
λj ∈ UOi . Consider a node i in L1 and let its error
(jm)
in estimation of the component z
[k] be denoted by
(jm)
(jm)
(jm)
[k] − z
[k]. The estimation errors of
ei [k] , ẑi
the individual components are aggregated in the vector
(j)
(j)
ei [k] = ẑi [k] − z(j) [k]. Subtracting z(j) [k + 1] from
both sides of equation (5), noting that z(j) [k + 1] =
W(λj )z(j) [k] (based on the dynamics given by (3)), and
(jm)
(jm)
and emax [k] = maxu∈L(j) eu
0
0
[k]. This property holds
for every component m of z(j) [k]. Analyzing each of the
two cases, we infer that at every time-step k, each com(j)
ponent of the vector ēi [k] in (B.1) lies in the convex
hull of the corresponding components of the error vec(j)
(j)
tors eu [k], u ∈ L0 = Sj ∩ R. Based on Lemma 2, we
(j)
have that limk→∞ eu [k] = 0, ∀u ∈ Sj ∩ R, and hence
(j)
it follows that ẑi [k] converges asymptotically to z(j) [k]
14
(j)
violating the first property of a MEDAG in Definition
8. We thus arrive at a contradiction, and the proof is
complete.
for every regular node i in L1 .
Suppose the result holds for all levels from 0 to q (where
1 ≤ q ≤ Tj − 1). It is easy to see that the result holds for
(j)
all regular nodes in Lq+1 as well, by noting the following.
Sq
(j)
(j)
(j)
(i) A regular node i ∈ Lq+1 has Ni ∩ R ⊆ r=0 Lr .
D
(j)
(ii) For each i ∈ Lq+1 , a similar analysis reveals that at
every every time-step k, each component of the vector
(j)
ēi [k] lies in the convex hull of the corresponding comSq
(j)
(j)
ponents of the error vectors eu [k], u ∈ r=0 Lr . The
desired result then follows from the induction hypothesis. An identical argument can be sketched for a real
eigenvalue λj ∈ ΩU (A), and thus the result holds for
any λj ∈ ΩU (A). We arrive at the conclusion that every
node i ∈ R can asymptotically estimate z(j) [k] for every
eigenvalue λj ∈ UO i . Thus, each node i ∈ R can asymptotically estimate z[k], and hence x[k] = Tz[k].
C
The Strong r-robustness Property of Random Graphs
The focus of this section is to address the following
question. Given a dynamical system and an associated
large-scale complex sensor network monitoring the system, under what conditions is the system and network
pair r-feasible? To provide an answer to this question,
we study the ‘strong r-robustness’ property in three relatively common random graph models for large-scale
complex networks, namely the Barabási-Albert (BA)
preferential attachment model, the Erdős-Rényi random
graph model, and the 2-dimensional random geometric
graph model.
Consider a scenario where we are given a dynamical system and an associated wireless sensor network such that
the system and network pair is r-feasible, and hence,
resilient to adversarial attacks. We wish to expand the
network via addition of more sensors without disrupting
the r-feasibility property, i.e., we intend to tolerate the
same number of adversaries as earlier. As the first property in Proposition 4 suggests, this can be achieved by
continually adding new nodes with incoming edges from
at least r nodes in the existing network. The specific construction where the neighbors of a new node are selected
with a probability proportional to the number of edges
they already have leads to the BA preferential attachment model. Such a model is thought of as a plausible
mechanism for the formation of many real-world complex networks [36]. Based on our foregoing discussion, it
then follows that such real-world networks would facilitate the LFRE dynamics introduced in Section 4.2, and
would hence be resilient to the worst-case attack model
considered in this paper.
Proof of Theorem 7
PROOF. For sufficiency, it is easily noted that the
conditions stated in the theorem guarantee termination
of the MEDAG construction algorithm for every λj ∈
ΩU (A). The rest of the proof for sufficiency follows identical arguments as the proof of Theorem 6.
For proving necessity, we first note that the proposed algorithm summarized in Section 6.1 is applicable only if
the MEDAG construction algorithm (Algorithm 1) terminates for each λj ∈ ΩU (A) and returns a subgraph Gj
satisfying the properties of a MEDAG for all f -local sets
containing V \ R′ . Here R′ denotes the set of nodes that
behave regularly during the execution of Algorithm 1.
Based on the hypothesis of the theorem, since R′ = V,
the existence of a MEDAG Gj ∀λj ∈ ΩU (A) is necessary
in this case for running the LFRE algorithm. The rest of
the proof proceeds via contradiction. Suppose G is not
strongly (2f + 1)-robust w.r.t. Sj for some λj ∈ ΩU (A)
and yet there exists a MEDAG Gj for λj . Since G is not
strongly (2f + 1)-robust w.r.t. λj , there exists a nonempty set C ⊆ V \ Sj that is not (2f + 1)-reachable.
Consider the trivial f -local set A = ∅. The subgraph Gj
must contain a partition of R = V \ A = V into levels
that satisfy the second property of a MEDAG in Definition 8. With this point in mind, let C be partitioned as
S
(j)
C = qr=1 Fr , where Fr = C ∩ Lnr for some set of integers {n1 , · · · , nq |1 ≤ ni ≤ Tj ∀i ∈ {1, · · · , q}}. Here,
(j)
{Lnr }qr=1 represents a subset of the levels that partition
R in the MEDAG Gj (that exists based on the hypothesis). Without loss of generality, let n1 < n2 < · · · < nq .
Then, from the definition of a MEDAG, it follows that
(j)
for any i ∈ Fn1 , Ni contains elements from only V \ C.
(j)
As C is not (2f +1)-reachable, |Ni | < (2f +1), thereby
Next, we turn our attention to one of the most common
mathematical models for large-scale networks, namely
Erdős-Rényi random graphs [37]. We denote an ErdősRényi random graph on N nodes by GN,p , where all possible edges between pairs of different nodes are present
independently and with the same probability p. We further note that p is in general a function of the network
size N . From the perspective of a network designer, we
will be interested in answering the following questions.
(i) How should the size of the source sets Sj (for each
λj ∈ ΩU (A)) scale with the size of the network to maintain r-feasibility in an Erdős-Rényi random graph? (ii)
Which nodes should be chosen as the source nodes? Prior
to answering these questions, we briefly remark on the
notation to be used for the remainder of this section.
The term w.h.p. (with high probability) will be used for
events with probability tending to 1 as N → ∞. Given
15
two non-negative sequences aN and bN , the notation
aN ≪ bN will convey the same meaning as aN = o(bN ).
To make use of Lemma 4, we first recall a few definitions
from [34]. Given an integer r ≥ 2, define
Tc (N, p) ,
(r − 1)!
N pr
1
(r−1)
1
Tc (N, p).
, Ac (N ) , 1 −
r
(D.1)
We have the following result for an Erdős-Rényi random
graph model.
Proposition 6 Given an LTI system (1), a measurement model (2), and a network modeled by an
Erdős-Rényi random graph GN,p , suppose that for each
λj ∈ ΩU (A), the source set Sj (N ) is chosen randomly. 8
Then, the following are true.
Among the three random graph models mentioned
earlier, the one most relevant to our cause is the twodimensional random geometric graph (RGG) model [38].
RGGs are typically used to model networks where a
notion of spatial proximity governs the interaction between the nodes. A wireless sensor network where randomly deployed nodes communicate with nodes only in
a geographical vicinity, constitutes an ideal setup for an
RGG model [39]. We will consider a two-dimensional
RGG model generated by first placing N nodes randomly within the unit square [0, 1]2 . Undirected edges
are placed between two nodes if and only if the Euclidean distance between such nodes is at most d(N ),
where d(N ) is a positive number that may depend on the
network size N . We will denote such a RGG by GN,d(N ) .
Like the Erdős-Rényi case, our focus will be on understanding how the source sets should be chosen to ensure
r-feasibility of GN,d(N ) with high probability. To provide
such a characterization, we first recall a few functions
from [38]. Let H(x) , x lnx − x + 1 be defined on [0, ∞)
and J(x) , lnx − 1 + x−1 be defined on (0, ∞). Further−1
more, let JR
: [0, ∞) → [1, ∞) denote the inverse of
J(x) when the domain of J(x) is [1, ∞). We then have
the following result.
1
(i) Let p = p(N ) be such that N −1 ≪ p ≪ N − r . If for
|S (N )|
each λj ∈ ΩU (A), Ajc (N ) ≥ 1 + δ, for some δ > 0,
and |Sj (N )| ≤ N2 , then GN,p is r-feasible w.h.p. if
and only if N p − (lnN + (r − 1)ln lnN ) → ∞ as
N → ∞.
1
(ii) Let p = p(N ) be such that p ≫ N − r . If for each
λj ∈ ΩU (A), |Sj (N )| ≥ r, then GN,p is r-feasible
w.h.p.
Proposition 7 Given an LTI system (1) and a measurement model (2), let the communication
q graph be modeled
lnN
by the RGG GN,d(N ), where d(N ) = aπN
and a > 1.
For each λj ∈ ΩU (A), let a node be chosen as a source
node for λj with a probability p independently of the other
1
).
nodes in the network. Let r = γa lnN , where γ ∈ (0, 5π
5π
Suppose a ≥ H(5πγ) and
PROOF. (i) If the conditions in part (i) are met, then
for each λj ∈ ΩU (A), Sj percolates via bootstrap percolation with threshold r on GN,p w.h.p. based on [34, Theorem 3.2]. Lemma 4 then implies that GN,p is strongly
r-robust w.r.t. each such source set Sj w.h.p., i.e., GN,p
is r-feasible w.h.p. . The proof for part (ii) follows similarly by leveraging [34, Theorem 5.8] and Lemma 4.
(
5πγ
p ≥ min γ, −1 1
JR ( aγ )
Remark 10 We glean the following insights from the
above result. First, we observe that if either condition (i)
or condition (ii) is met, then our proposed algorithm will
enable each regular node to asymptotically estimate the
state of the system w.h.p. in the presence of any ⌊ r−1
3 ⌋
locally-bounded set of Byzantine adversaries. This is a
direct consequence of Theorem 6. The first part of Proposition 2 indicates that although the source sets can be
chosen randomly, their size needs to scale appropriately
with the size of the network to maintain r-feasibility. The
second part states that if the probability of edge formation is large enough, then it suffices to pick source sets
of constant size equal to the bare minimum required for
achieving r-feasibility (which equals r based on part (ii)
of Proposition 4).
)
.
(D.2)
Then, GN,d(N ) is r-feasible. 9
PROOF. For each λj ∈ ΩU (A), if the source set Sj
is chosen as described above, then it percolates GN,d(N )
w.h.p. if the conditions of the proposition are met [38,
Theorem 4]. The result then follows from Lemma 4.
Remark 11 In an attack-prone wireless sensor network, one might be interested in tolerating f -local adversarial sets where the paramater f scales with the size of
the network. Such a possibility is captured by Proposition
−1
9, based on which, ⌊ γa lnN
⌋-local Byzantine adversar3
ial sets can be accounted for by our proposed algorithm.
8
By choosing Sj (N ) randomly, we imply that the measurement set needed to detect λj is allocated to |Sj (N )| nodes
picked uniformly at random. The notation Sj (N ) is used to
explicitly point out that the size of the source sets scales
with the size of the network.
q
lnN
.
A RGG GN,d(N) is connected w.h.p. for d(N ) >
πN
The choice of a > 1 thus allows one to deal with an asymptotically connected GN,d(N) .
9
16
| 3cs.SY
|
Nonparametric Inference for Auto-Encoding
Variational Bayes
arXiv:1712.06536v1 [stat.ML] 18 Dec 2017
Erik Bodin *
Iman Malik *
*
Carl Henrik Ek *
University of Bristol
†
Neill D. F. Campbell †
University of Bath
Variational approximations are an attractive approach for inference of latent variables in unsupervised
learning. However, they are often computationally intractable when faced with large datasets.
Recently, Variational Autoencoders (VAEs) Kingma and Welling [2014] have been proposed as a
method to tackle this limitation. Their methodology is based on formulating the approximating
posterior distributions in terms of a deterministic relationship to observed data consequently the
title “Auto-Encoding Variational Bayes”. Importantly, this is a decision regarding an approximate
inference scheme that should not be confused with an auto-encoder as a model.
Unsupervised learning is a ill-conditioned problem that requires prior knowledge to reach a solution.
We would like to learn latent representations that are low-dimensional and highly interpretable.
A model that has these characteristics is the Gaussian Process Latent Variable Model (GP-LVM)
Lawrence [2005]. The benefits and negative of the GP-LVM are complementary to the VAE, the
former provides useful low-dimensional latent representations while the latter is able to handle large
amounts of data and can use non-Gaussian likelihoods. Our inspiration for this paper is to marry these
two approaches and reap the benefits of both. In order to do so we will introduce a novel approximate
inference scheme inspired by the GP-LVM to the VAE.
The standard VAE formulation Kingma and Welling [2014] adopts a unit Gaussian prior, creating a
trade-off between the embedded data residing at the same location in the latent space and the ability
to reconstruct the data in the observed space. This encourages a “tight packing” of the data around a
shared origin, with hope that similar data in the observed space have overlapping probability mass in
the latent space i.e. mutual information. It has been shown that a simple prior over-regularizes the
latent space leading to poor reconstructions Hoffman and Johnson [2016]. Other more flexible priors
can be used to change the dynamic between reconstruction and mutual information, such as in recent
work of using a mixture Tomczak and Welling [2017] or let the prior be autoregressive Chen et al.
[2016]. The quality of the output from a VAE can be improved by more expressive generative models,
but this has been shown to lead to a tendency of ignoring the latent space, defeating the purpose of
unsupervised learning Zhao et al. [2017]. In summary, reconstruction quality and mutual information
in the latent space is traded against each other.
In this paper we address this limiting trade-off by escaping it; we let the space where we encourage
sharing be separated from the space where the generative capacity is set. We do this by an approximation of a model where the observations can be generated from either space. We show experimentally
that the approximation allows the capacity of the generative bottle-neck (Z) to be arbitrarily large
without losing sharing and the beneficial properties of the sharing space (X), allowing reconstruction
quality to be unlimited by Z at the same time as a low-dimensional space can be used to perform
ancestral sampling from as well as a means to reason about the embedded data.
1
Method
The VAE Kingma and Welling [2014] inference scheme optimises a traditional evidence lower bound
where the latent space posterior is approximated as a deterministic relationship from the observed
QN
data as q(Z|Y) = i=1 q(zi |yi ) where each latent variable is conditionally independent given the
31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.
Y
z1
Z
z2
→
(a) Model
z1
y2
z2
zN
→
...
y1
Y
zN
...
yN
y1
(b) VAE Approximation
y2
yN
(c) VAE Inference Scheme
Y
X
X
Distribution
z1
z2
zN
z1
z2
zN
→
Point Estimate
...
y1
Z̃
...
y2
yN
y1
y2
yN
Observed
(d) Our Nonparametric VAE
(e) Our Inference Scheme
Figure 1: Contrasting variational approximation schemes for unsupervised learning. (a) We specify an
unsupervised generative model from latent Z to observed Y . (b) The VAE proposes a fully factored
variational approximation. (c) Inference proceeds by conditioning the variational latent parameters on
observed data through an explicit deterministic function (e.g. MLP network). (d) Our model proposes
an additional latent space X that ties together the factored Z space. (e) Inference then proceeds
with X also conditioned on observed data through an additional deterministic function. For tractable
inference we match moments between Z and Z̃.
observed data. In this paper we introduce an additional latent variable X that model the interaction
between the latent variables Z.
Our approach means that Z are no longer independent but conditionally independent given X. This
leads to the following updated evidence lower bound with an additional divergence term as,
Ls = Lg − KL(qg (Z|Y)||ps (Z|X)),
(1)
where Lg and qg (zi |yi ) is the standard VAE lower bound and approximative posterior respectively.
To facilitate the use of batch processing we will rather than matching the joint distribution of the latent
space match the predictive posteriors. As each Z is conditionally independent given the observed
data Y this leads to the following updated objective function,
Y
Y
L0s = Lg − KL( qg (zi |yi )||
ps (zi |X, z¬i )).
(2)
i
i
The predictive posterior of the Gaussian process is,
p(zi |X, z¬i ) = N (k(xi , X¬i )k(X¬i , X¬i )−1 z¬i ,
k(xi , xi ) − k(xi , X¬i )k(X¬i , X¬i )k(X¬i , xi ))
(3)
where k(·, ·) is the covariance function. Evaluating this posterior for N datapoints is computationally
expensive due to the inverse of the covariance function. To proceed we will introduce two additional
approximations, first we will approximate the mean of the predictive posterior of the GP by directly
parametrising it as,
µ(zi ) = k(xi , X¬i )k(X¬i , X¬i )−1 z¬i ≈ Wi z¬i .
(4)
1×N
In specific we will parametrise the weight matrix Wi ∈ RP
such that the approximative predictive
mean for zi is a convex combination of z¬i , i.e. that j Wij = 1, ∀i and Wij ≥ 0, ∀i, j. The
2
intuition behind this is that we want to encourage sharing in the latent space X such that the space Z
is represented in a distributed fashion. Secondly rather than minimising the KL-divergence we will
match the first mode of the two distributions. This leads to our final objective function,
X
L̃s = Lg −
(E[zi ] − E[z̃i ])2 ,
(5)
i
where z̃ is the prediction of the latent space from X while z is the prediction using the approximative
posterior from the VAE. In effect we have separated the two models while retaining a connection
by matching their first modes when predicting the latent space Z. We will now proceed to show
the experimental evaluation of the model showing that we are capable of using the additional lowdimensional latent space X as a proxy for the VAE latent space Z.
2
Experiments
In this paper we choose to model Wi,j as a squared exponential covariance kernel between input
locations, xi and xj ,
kxi − xj k2
(6)
Wi,j = k(xi , xj ) = σ 2 exp −
i 6= j
2l2
and infer its parameters. We ensure Wi,i = 0 by subtracting the diagonal from the computed W and
normalise by rows to create a convex combination. The latent locations X are represented implicitly
as a function of the observed data as in Lawrence and Quiñonero-Candela [2006].
We trained models to illustrate that our extension can be used to obtain a low-dimensional X space
that is highly interpretable while permitting the use of a high-dimensional Z space to provide high
quality data generation. We validate our approach, with comparison to a standard VAE, by showing
data embeddings in the X space and generation of new data.
All experiments were performed with the decoder and both encoders as Multilayer Perceptrons (MLP)
with the same architecture as in the original VAE Kingma and Welling [2014]; we used two hidden
layers of 500 units each, mini-batch sizes of 128 and a drop-out probability of 0.9 throughout training.
The decoder used was the Bernoulli MLP variant. Furthermore, the ADAM Kingma and Ba [2014]
optimiser was used with a learning rate of 10−3 . We varied the dimensionality of the inner most layer
of the autoencoder (the Z space) for the experiments. We used the MNIST data set from LeCun et al.
[1998] comprised of 55 000 training examples and 10 000 test examples of 28 × 28 pixel greyscale
images, corresponding to 784 data dimensions.
In Fig. 2 we show a 2-dimensional X space corresponding to a 500-dimensional Z space with both
the training and test data embedded as well as examples of data generation. Despite using a high
capacity Z space, the nonparametric VAE can still sample from a low dimension (in our example
2D) using ancestral sampling from the X space. This ensures that test samples maintain a high
fidelity from a space that is highly interpretable, easy to visualise and easy to sample from. In Fig. 3
we show corresponding X space embeddings for different dimensionalities of Z; this demonstrates
that the X space maintains its virtues independent of the Z dimension. Finally, in Fig. 4, we show
sample interpolations in the latent space for the standard VAE (directly in Z) and with our extension
(in X) illustrating that not only is the reconstruction quality preserved but interpolations are more
meaningful.
3
Conclusions
We have presented a hierarchical model for unsupervised learning and an associated efficient approximative inference scheme. The inference takes inspiration from amortised inference and use
a recognition model to parameterise the approximate posterior using a deterministic relationship
from the observed data. Rather than using a traditional mean-field approximation which forces
the latent representation to be independent we introduce an additional latent representation that
models their dependence. Our model results in a significantly lower dimensional latent representation
allowing us to visualise and generate data in a intuitive manner without sacrificing the quality of the
reconstruction. We have shown experimental results on how we can retain the representative power
of a 500 dimensional model with just a 2 dimensional latent space.
3
(a) Training X
(b) Test X
(c) Samples
Figure 2: Learned X space embeddings from the nonparametric VAE. Inferred X locations for (a) the
training data and (b) the test data with colors encoding the MNIST digit classes. (c) Generated
samples from the corresponding locations in (b) using a Z space with 500 dimensions.
VAE
np-VAE
zi ∈ R2
zi ∈ R10
zi ∈ R100
zi ∈ R1000
Figure 3: Latent space visualisation. Upper row: the Z space embedding is visualised for the standard
VAE where possible (we are unable to do this for high-dimensional Z). Bottom row: the same Z
space dimensionalities are used but the nonparametric VAE allows the X space to be visualised and
sampled (set to be 2-dimensional). Z spaces of higher dimension become impractical to visualise
and interpret whereas the X space provides an embedding for easy display and interpretation.
VAE
zi ∈ R 2
zi ∈ R500
np-VAE
zi ∈ R 2
zi ∈ R500
Figure 4: Latent space interpolation. The upper two rows show interpolants between two MNIST
training examples for a standard VAE with a Z latent dimensionality of 2 and 500. The bottom two
rows show interpolants between the same training examples for our nonparametric VAE with the same
respective dimensionalities for Z but where the interpolation is performed in the inferred latent space
X of dimension 2. We observe that a similar reconstruction quality is obtained by corresponding
Z-dimensionalities, however, the interpolants from the X space of the nonparametric VAE are more
meaningful with credible intermediate states between digits. Thus we can obtain a low dimensional
latent space that provides interpretability without sacrificing reconstruction quality.
4
References
Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. In International Conference
on Learning Representations (ICLR), 2014.
Neil D Lawrence. Probabilistic non-linear principal component analysis with Gaussian process latent
variable models. Journal of Machine Learning Research, 6:1783–1816, 2005.
Matthew D Hoffman and Matthew J Johnson. ELBO Surgery: Yet Another Way to Carve up the
Variational Evidence Lower Bound. In Workshop in Advances in Approximate Bayesian Inference,
NIPS, 2016.
Jakub M Tomczak and Max Welling. VAE with a VampPrior. arXiv preprint arXiv:1705.07120,
2017.
Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya
Sutskever, and Pieter Abbeel. Variational Lossy Autoencoder. arXiv preprint arXiv:1611.02731,
2016.
Shengjia Zhao, Jiaming Song, and Stefano Ermon. Towards a Deeper Understanding of Variational
Autoencoding Models. arXiv preprint arXiv:1702.08658, 2017.
Neil D Lawrence and J Quiñonero-Candela. Local distance preservation in the GP-LVM through
back constraints. Proceedings of the 23rd international conference on Machine learning, pages
513–520, 2006.
Diederik Kingma and Jimmy Ba. ADAM: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-Based Learning Applied
to Document Recognition. Proceedings of the IEEE, 86(11):2278–2323, 1998.
5
| 2cs.AI
|
Robustly Learning a Gaussian: Getting Optimal Error, Efficiently
Ilias Diakonikolas∗
CS, USC
[email protected]
arXiv:1704.03866v2 [cs.DS] 5 Nov 2017
Jerry Li §
EECS & CSAIL, MIT
[email protected]
Gautam Kamath†
EECS & CSAIL, MIT
[email protected]
Ankur Moitra¶
Math & CSAIL, MIT
[email protected]
Daniel M. Kane‡
CSE & Math, UCSD
[email protected]
Alistair Stewartk
CS, USC
[email protected]
November 7, 2017
Abstract
We study the fundamental problem of learning the parameters of a high-dimensional
Gaussian in the presence of noise — where an ε-fraction of our samples were chosen by
an adversary. We give robust estimators that achieve estimation error O(ε) in the total
variation distance, which is optimal up to a universal constant that is independent of
the dimension.
In the case√where just the mean is unknown, our robustness guarantee is optimal up
to a factor of 2 and the running time is polynomial in d and 1/ε. When both the mean
and covariance are unknown, the running time is polynomial in d and quasipolynomial
in 1/ε. Moreover all of our algorithms require only a polynomial number of samples.
Our work shows that the same sorts of error guarantees that were established over fifty
years ago in the one-dimensional setting can also be achieved by efficient algorithms in
high-dimensional settings.
∗
Supported by NSF CAREER Award CCF-1652862, a Sloan Research Fellowship, and a Google Faculty Research
Award.
†
Supported by NSF CCF-1551875, CCF-1617730, CCF-1650733, and ONR N00014-12-1-0999.
‡
Supported by NSF CAREER Award CCF-1553288 and a Sloan Research Fellowship.
§
Supported by NSF CAREER Award CCF-1453261, CCF-1565235, a Google Faculty Research Award, and an
NSF Graduate Research Fellowship.
¶
Supported by NSF CAREER Award CCF-1453261, CCF-1565235, a Packard Fellowship, a Sloan Research Fellowship, a grant from the MIT NEC Corporation, and a Google Faculty Research Award.
k
Supported by a USC startup grant.
1
Introduction
1.1
Background
The most popular and widely used modeling assumption is that data is approximately Gaussian. This is a convenient simplification to make when modeling velocities of particles in an ideal
gas [Goo15], measuring physical characteristics across a population (after controlling for gender),
and even modeling fluctuations in a stock price on a logarithmic scale. However, real data is not
actually Gaussian and is at best crudely approximated by a Gaussian (e.g., with heavier tails).
What’s worse is that estimators designed under this assumption can perform poorly in practice
and be heavily biased by just a few errant samples that do not fit the model.
For over fifty years, the field of robust statistics [HR09, HRRS86, RL05] has studied exactly
this phenomenon — the sensitivity or insensitivity of estimators to small deviations in the model.
Unsurprisingly, one of the central questions that shaped its development was the problem of learning
the parameters of a one-dimensional Gaussian distribution when a small fraction of the samples
are arbitrarily corrupted. More precisely, in 1964, Huber [Hub64] introduced the following model:
Definition 1. In Huber’s contamination model, we are given samples from a distribution
D = (1 − ε)N (µ, σ 2 ) + εZ ,
where N (µ, σ 2 ) is a Gaussian of mean µ and variance σ 2 , and Z is an arbitrary distribution chosen
by an adversary.
Intuitively, among our samples, about a (1−ε) fraction will have been generated from a Gaussian
and are called inliers, and the rest are called outliers or gross corruptions. We will work with an
even more challenging1 model — called the strong contamination model (Definition 2) — where the
adversary is allowed to look at the inliers and then decide on the outliers. The literature on robust
statistics has given numerous explanations and empirical investigations [GCSR14, Ham01] into how
such outliers might arise as the result of equipment failure, data being entered incorrectly, or even
from a subpopulation that was not accounted for in a medical study. These types of errors are
erratic and difficult to model, so instead our goal is to design a procedure that accurately estimates
µ and σ 2 without making any assumptions about them.
In one dimension, the median and median absolute deviation are well-known robust estimators
for the mean and variance respectively. In particular, given samples X1 , X2 , . . . , Xn , we can compute
µ
b = median(X1 , X2 , . . . , Xn ) and σ
b=
median(|Xi − µ
b|)
,
−1
Φ (3/4)
where Φ is the cumulative distribution of the standard Gaussian. (This scaling constant is needed
to ensure that σ
b is an unbiased estimator when there is no noise.) If n ≥ C logε21/δ , then with
probability at least 1 − δ we have that dT V (N (µ, σ 2 ), N (b
µ, σ
b2 )) ≤ Cε. In Huber’s contamination
model, this is the strongest type of error guarantee we could hope for2 and captures both the task
of learning the underlying parameters µ and σ 2 , and finding the approximately best fit to the
observed distribution within the family of one-dimensional Gaussians. In fact there are plentifully
many other estimators — such as the trimmed mean, winsorized mean, Tukey’s biweight function,
1
None of the results in our paper were previously known in Huber’s contamination model either. The reason we
work with this stronger model is because we can — nothing in our analysis relies on the inliers and outliers being
independent.
2
See Lemma 17.
1
and the interquartile range — that achieve the same sorts of error guarantees, up to constant factors.
The design of robust estimators for location (e.g., estimating µ) and scale (e.g., estimating σ 2 ) is
guided by certain overarching principles, such as the notion of the influence curve [HRRS86] or the
notion of breakdown point [RL05]. In some cases, it is even possible to design robust estimators
that are minimax optimal [Hub64].
These days, much of modern data analysis revolves around high-dimensional data — for example, when we model documents [BNJ03], images [OF96], and genomes [NJB+ 08] as vectors in a very
high-dimensional space. The need for robust estimators is even more pressing in these applications,
since it is infeasible to remove obvious outliers by inspection. However, adapting robust statistics
to high-dimensional settings is fraught with challenges. The principles that guided the design of
robust estimators in one dimension seem to inherently lead to high-dimensional estimators that are
hard to compute [Ber06, HM13].
In this paper, we focus on the central problem of learning the parameters of a multivariate
Gaussian N (µ, Σ) in the strong contamination model. The textbook estimators for the mean and
covariance – such as the Tukey median [Tuk75] and minimum volume enclosing ellipsoid [Rou85]
– essentially search for directions where the projection of D is suitably non-Gaussian. However,
trying to find a direction where the projection is non-Gaussian can be like looking for a needle
in an exponentially-large haystack – these statistics are not efficiently computable, in general.
Furthermore, a random projection will look Gaussian with high probability [Kla07].
In this paper, our main result is an efficiently computable estimator for a high-dimensional
Gaussian that achieves error
b ≤ Cε
dT V (N (µ, Σ), N (b
µ, Σ))
in the strong contamination model, for a universal constant C that is independent of the dimension.
For a Gaussian distribution, we consider estimation in terms of total variation distance, which is
equivalent to estimating the parameters under the natural measures. Our main idea is to use various
regularity conditions satisfied by the inliers to make the problem of searching for non-Gaussian
projections easier. When just the mean µ is unknown, our algorithm runs in time polynomial in
the dimension d and 1/ε. When both the mean and covariance are unknown, our algorithm runs
in time polynomial in d and quasi-polynomial in 1/ε. All of our algorithms achieve polynomial
sample complexity.
Prior to our work, the best known algorithm of Diakonikolas et al. [DKK+ 16] achieved estimation error O(ε log 1/ε) for this problem3 , again with respect to total variation distance. Concurrently, Lai, Rao and Vempala [LRV16] gave an algorithm which achieves estimation error roughly
O(ε1/2 log1/2 d). In fact, the algorithm of Diakonikolas et al. [DKK+ 16] works in a stronger model
than what we consider here, where an adversary gets to look at the samples and then decides on
an ε-fraction to move arbitrarily. Such errors are both additive and subtractive (because inliers
are removed). Interestingly, Diakonikolas, Kane and Stewart [DKS17] proved that any Statistical
Query learning algorithm that works in such an additive and subtractive model and achieves an
error guarantee asymptotically better than O(ε log1/2 1/ε) must make a super-polynomial number
of statistical queries. Our work shows a natural conclusion that in an additive only model it is
possible to algorithmically achieve the same error guarantees as are possible in the one-dimensional
case, up to a universal constant.
3
We note that, as stated, the results in [DKK+ 16] give estimation error O(ε log3/2 1/ε). However, combining the
techniques in [DKK+ 16] with the arguments in Section 7 of this paper gives the stated bound. This argument will
be included in the full version of [DKK+ 16].
2
1.2
Our Results and Techniques
In what follows, we will explain both our work as well as prior work through the following lens:
At the core of any robust estimator is some procedure to certify that the estimates have
not been moved too far away from the true parameters by a small number of corruptions.
First, we consider the subproblem where the covariance Σ = I is known and only the mean µ is
unknown. In the terminology of robust statistics, this is called robust estimation of location. If we
could compute the Tukey median, we would have an estimate that satisfies dT V (N (µ, I), N (b
µ, I)) ≤
Cε. The way that the Tukey median guarantees that it is close to the true mean is that along every
direction u it is close to the median of the projection of the samples. More precisely, at least a
1−ε
T
T b, and at least a 1−ε fraction of the samples satisfy
2 fraction of the samples satisfy u Xi ≥ u µ
2
T
T
u µ
b ≥ u Xi . However, if we have a candidate µ
b, finding a direction u that violates this condition
is again like searching for a needle in an exponentially large haystack.
The approach of Diakonikolas et al. [DKK+ 16] was essentially a data-dependent way to search
for appropriate directions u, by looking for directions where the empirical variance is larger than
it should be (if there were no corruptions). However, because their approach considers only a
single direction at a time, it naturally gets stuck at error Θ(ε log1/2 1/ε). This is because along the
direction u, only when a point is Ω(log1/2 1/ε) away from most of the rest of the samples can we be
relatively confident that it is an outlier. Thus, an adversary could safely place all the corruptions in
the tails and move the mean by as much as Θ(ε log1/2 1/ε). This would not affect the Tukey median
by as much, but would affect an estimate based on the empirical mean (because the algorithm could
find no other outliers to remove) by considerably more.
Our approach is to consider logarithmically many directions at once. Even though an inlier
can be logarithmically many standard deviations away from the mean along a single direction u
with reasonable probability, it is unlikely to be that many standard deviations away simultaneously
across many orthogonal directions. Essentially, this allows us to remove the influence of outliers on
all but a logarithmic dimensional subspace. Combining this with an algorithm for robustly learning
the mean in time exponential in the dimension (but polynomial in the number of samples), we obtain
our first main result:
Theorem 1. Suppose we are given a set of n = poly(d, 1/ε) samples from the strong contamination
model, where the underlying d-dimensional Gaussian is N (µ, I). Let ε ≤ ε0 , where ε0 is a positive
universal constant. For any β > 0, there is an algorithm to learn an estimate N (b
µ, I) that with
high probability satisfies
1
1
2
ε.
dT V (N (µ, I), N (b
µ, I)) ≤ √ + O √ + ε
β
2
Moreover, the algorithm runs in time poly(n, (1/ε)β ).
We prove an almost matching lower bound of 2ε + Ω(ε2 ) on the estimation error. Thus, our
√
robustness guarantee is optimal up to a factor√of 2, even among computationally inefficient robust
estimators. Interestingly, our extra factor of 2 comes from the following geometric fact which we
make crucial
√ use of: Any convex body of diameter D in any dimension can be covered by a ball of
radius D/ 2, and moreover such a ball can be (approximately) found in time exponential in the
dimension. Suppose that along some direction u we have an estimate p that is guaranteed to be
within ε/2 of the projection of the true mean µ. We can now confine µ to a slab of width ε, and by
taking the intersection of all such slabs we get a convex body that contains µ and has diameter of
3
√
at most ε. By covering
√ the body with a ball of radius ε/ 2, we are guaranteed that the center of
the ball is within ε/ 2 of the true mean. This gives us a general way to combine one-dimensional
robust estimates along a net of directions.
We note that, for general isotropic sub-Gaussian distributions, the bound of O(ε log1/2 1/ε)
of [DKK+ 17] is optimal for robust mean estimation, even in one dimension. See Section A for
a proof of this fact. However, our results can be seen to hold more generally than stated above
– indeed, the same arguments work for a class of symmetric isotropic sub-Gaussian distributions
which are sufficiently smooth near their mean. More precisely, we require that along any univariate
projection, the mean is robustly estimated by the median.
We next consider the subproblem where the mean µ = 0 is known and only the covariance Σ is
unknown. In the terminology of robust statistics, this is called robust estimation of scale. In this
b that satisfies4 kΣ − Σk
b F ≤ Cε. When Σ
b does not satisfy
case, we want to compute an estimate Σ
this condition, it can be shown (in Section 6.2.3) that there is a degree-two polynomial p(X), where
E
[p(X)] = 1 and
X∼N (0,Σ)
E
[p(X)] = 1 + C ′ ε .
b
X∼N (0,Σ)
It turns out that, even given the polynomial p(X), deciding whether or not the above conditions
b we can certainly compute E
approximately hold is challenging. Given p(X) and Σ,
b [p(X)].
X∼N (0,Σ)
But given only contaminated samples from N (0, Σ) and without knowing what Σ is, can we estimate
EX∼N (0,Σ) [p(X)]?
Often, univariate robust estimation problems are considered easy, with a simple recipe: Construct an unbiased estimator for the statistic for which each sample point has low influence. However, in our setting, it is highly non-trivial to construct such an estimator. The naive attempt in
this case would be the median – this immediately fails since the distribution of p(X) is asymmetric.
Even if there were no noise, that would not necessarily be an unbiased estimator. So how can we
dampen the influence of outliers, if there is no natural symmetry in the distribution? We construct
a robust estimator crucially using the fact that p(X) is the weighted sum of chi-squared random
variables when there is no noise. The key structural fact we exploit is the following: Given two
sums of chi-squared random variables, if the random variables are far in total variation distance,
most of their difference must lie close to their means. We use this fact to show how, given a weak
estimate of the mean (i.e., one which is only accurate up ω(ε)), one can improve the estimate by a
constant factor. Our result follows by an iterative application of this technique.
However, there is still a major complication in utilizing our low-dimensional estimator to obtain
a high-dimensional estimator. In the unknown mean case, we knew the higher-order moments (since
we assumed that the covariance is the identity). Here, we do not have control over the higher-order
moments of the unknown Gaussian. Overcoming this difficulty requires several new techniques,
which are quite complicated, and we defer the full details to Section 6. Our second main result is:
Theorem 2. Suppose we are given a set of n = poly(d, 1/ε) samples from the strong contamination
model, where the underlying d-dimensional Gaussian is N (0, Σ). There is an algorithm to learn an
b that runs in time poly(n, (1/ε)O(log 4 1/ε) ) and with high probability satisfies
estimate N (0, Σ)
b ≤ Cε ,
dT V (N (0, Σ), N (0, Σ))
for a universal constant C that is independent of the dimension.
4
More precisely, to obtain O(ε) error guarantee with respect to the total variation distance, we need to robustly
approximate Σ within O(ε) in Mahalanobis distance, which is a stronger metric than the Frobenius norm. As part
of our approach, we are able to efficiently reduce to the case that Σ is close to the identity matrix, in which case the
Frobenius error suffices.
4
A key technical problem arises when we attempt to combine estimates for the covariance restricted to a subspace and its orthogonal complement. We refer to this as a stitching problem,
where if we write Σ as
ΣV AT
,
Σ=
A ΣV ⊥
and have accurate estimates for ΣV and ΣV ⊥ , we still need to accurately estimate A. Our algorithm
utilizes an unexpected connection to the unknown mean case: We show that, under a carefully
chosen projection scheme, we can simulate noisy samples from a Gaussian with identity covariance,
where the mean of this distribution encodes the information needed to recover A. We defer the full
details to Section 6.4.
It turns out that we can solve the general case when both µ and Σ are unknown, by directly
reducing to the previous subproblems, exactly as was done in [DKK+ 16] (with some caveats, addressed in Section 4.4). Since all of our error guarantees are optimal up to constant factors, there
is only a constant factor loss in this reduction. Finally, we obtain the following corollary:
Corollary 1. Suppose we are given a set of n = poly(d, 1/ε) samples from the strong contamination
model, where the underlying d-dimensional Gaussian is N (µ, Σ). There is an algorithm to learn an
b that runs in time poly(n, (1/ε)O(log 4 1/ε) ) and with high probability satisfies
estimate N (b
µ, Σ)
b ≤ Cε ,
dT V (N (µ, Σ), N (b
µ, Σ))
for a universal constant C that is independent of the dimension.
This essentially settles the complexity of robustly learning a high-dimensional Gaussian. The
sample complexity of our algorithm depends polynomially on d and 1/ε, and the running time
depends polynomially on d and quasi-polynomially on 1/ε. Up to a constant factor, ours is the first
high-dimensional algorithm that achieves the same error guarantees as in the one-dimensional case,
where results were known for more than fifty years! It is an interesting open problem to reduce the
running time to polynomial in 1/ε (while still being polynomial in d). As we explain in Section 6.6,
this seems to require fundamentally new ideas.
More Related Work
In addition to the works mentioned above, there has been an exciting flurry of recent work on
robust high-dimensional estimation. This includes studying graphical models in the presence of
noise [DKS16], tolerating much more noise by allowing the algorithm to output a list of candidate hypotheses [CSV17], formulating general conditions under which robust estimation is possible [SCV18], developing robust algorithms under sparsity assumptions [Li17, DBS17, BDLS17]
where the number of samples is sublinear in the dimension, and leveraging theoretical insights to
give practical algorithms that can be applied to genomic data [DKK+ 17]. We note that, in comparison to all these other works, ours is the only to efficiently achieve the information-theoretically
optimal error guarantee (up to constant factors). Despite all of this rapid progress, there are still
many interesting theoretical and practical questions left to explore.
1.3
Organization
In Section 2, we go over preliminaries and notation that we will use throughout the paper. In Section 3, we describe an algorithm for robustly estimating the mean of a Gaussian in low-dimensional
settings, and crucially apply it in the design of an algorithm for mean-estimation in high dimensions, described in Section 4. Similarly, in Section 5, we give an algorithm for robustly estimating
5
the mean of degree-two polynomials in certain settings, which is applied in the context of our
covariance-estimation algorithm in Section 6. Finally, we put these tools together and describe our
general algorithm for robustly estimating a Gaussian in Section 7.
2
Preliminaries
In this section, we give various definitions and lemmata we will require throughout the paper. First,
given a distribution F , we let EF [f (X)] = EX∼F [f (X)] denote the expectation of f (X) under F .
If S is a finite set, we let ES [f (X)] = EX∼unif(S) [f (X)] denote the expectation of f (X) under the
uniform distribution over points in S (i.e., the empirical mean of f under S). Given any subspace
V ⊆ Rd , we let ΠV : Rd → Rd be the projection operator onto V . If V = span(v) is 1-dimensional,
we will denote this projection as Πv .
2.1
The Strong Contamination Model
Here we formally define the strong contamination model.
Definition 2. Fix ε > 0. We say a set of samples X1 , . . . , Xn was generated from the strong
contamination model on a distribution F , if it was generated via the following process:
1. We produce (1 − ε)n i.i.d. samples G from F .
2. An adversary is allowed to observe these samples and add εn points E arbitrarily.
We are then given the set of samples G ∪ E in random order. Also, we will say that the samples
X1 , . . . , Xn are ε-corrupted. Moreover given an ε-corrupted set of samples S, we will write S =
(G, E) where G is the set of uncorrupted points and E is the set of corrupted points. Moreover,
given a subset S ′ ⊂ S, we will also write S ′ = (G′ , E ′ ), where G′ = S ′ ∩ G and E ′ = S ′ ∩ E denote
the set of uncorrupted points and corrupted points remaining in S ′ . L will denote G \ G′ , which is
the set of “lost” uncorrupted points.
Given a contaminated set S ′ = (G′ , E ′ ) and a set G so that G′ ⊆ G, define the following
quantities
φ(S ′ , G) =
|G \ G′ |
,
|S ′ |
ψ(S ′ , G) =
|E ′ |
,
|S ′ |
∆(S ′ , G) = ψ(S ′ , G) + φ(S ′ , G) log
1
.
φ(S ′ , G)
(1)
In particular, observe that if ∆(S ′ , G) < O(ε), then a simple calculation implies that φ(S ′ , G) ≤
O(ε/ log 1/ε). Equivalently, we have removed at most an O(ε/ log 1/ε) fraction of good points from
G. This is crucial, as if we throw out an ε-fraction of good points then we essentially put ourselves
in the subtractive model, and there our guarantees no longer hold.
There are two differences between the strong contamination model and Huber’s contamination
model. First, the number of corrupted points is fixed to be εn instead of being a random variable.
However, this difference is negligible. It follows from basic Chernoff bounds that n samples from
Huber’s contamination model with parameter ε (for n sufficiently large) can be simulated by a
(1 + o(1))ε-corrupted set of samples, except with negligible failure probability. Hence, we lose
only an additive o(ε) term when translating from Huber’s contamination model to the strong
contamination model, which will not change any of the guarantees in our paper. The second
difference is that the adversary is allowed to inspect the uncorrupted points before deciding on the
corrupted points. This makes the model genuinely stronger since the samples we are given are no
longer completely independent of each other.
6
2.2
Deterministic Regularity Conditions
In analyzing our algorithms, we only need certain deterministic regularity conditions to hold on the
uncorrupted points. In this subsection, we formally state what these conditions are. It follows from
known concentration bounds that these conditions all hold with high probability given a polynomial
number of samples. Now with these regularity conditions defined once and for all, we will be able to
streamline our proofs in the sense that each step in the analysis will only ever use one of these fixed
set of conditions and will not use the randomness in the sampling procedure. We remark that some
subroutines in our algorithm only need a subset of these conditions to hold, so we could improve the
sample complexity by changing the regularity conditions we need at each step. However, since we
will not be concerned with optimizing the sample complexity beyond showing that it is polynomial,
we choose not to complicate our proofs in this manner.
2.2.1
Regularity Conditions for Unknown Mean
In the unknown mean case, we will require the following condition:
Definition 3. Let G be a multiset of points in Rd and η, δ > 0. We say that G is (η, δ)-good with
respect to N (µ, I) if the following hold:
p
• For all x ∈ G we have kx − µk2 ≤ O( d log(|G|/δ)).
• For every affine function L : Rd → R we have | PrG (L(X) ≥ 0) − PrN (µ,I) (L(X) ≥ 0)| ≤
η/(d log(d/ηδ)) .
• We have that k EG [X] − EN (µ,I) [X]k2 ≤ η.
• We have that kCovG [X] − Ik2 ≤ η/d.
• For any even degree-2 polynomial p : Rd → R we have that
E[p(X)] −
G
E[p2 (X)] −
G
E [p(X)] ≤ η
N (µ,I)
E [p2 (X)] ≤ η
N (µ,I)
E [p2 (X)]1/2 ,
N (µ,I)
E [p2 (X)], and
N (µ,I)
Pr[p(X) ≥ 0] ≤ Pr [p(X) ≥ 0] +
G
N (µ,I)
η
.
d log(|G|/δ)
It is easy to show (see Lemma 6) that given enough samples from N (µ, I), the empirical data set
will satisfy these conditions with high probability.
2.2.2
Regularity Conditions for Unknown Covariance
In the unknown covariance case, we will require the following condition:
Definition 4. Let G be a set of n points of Rd , and η, δ > 0. We say that G is (η, δ)-good with
respect to N (0, Σ) if the following hold:
• For all x ∈ G we have that xT Σ−1 x = O(d log(|G|/δ)).
7
• For any even degree-2 polynomial p : Rd → R we have
E[p(X)] −
G
E[p2 (X)] −
G
E [p(X)] ≤ η
N (0,Σ)
E [p2 (X)] ≤ η
N (0,Σ)
E [p2 (X)]1/2 ,
N (0,Σ)
E [p2 (X)] and
N (0,Σ)
Pr [p(X) ≥ 0] ≤ Pr [p(X) ≥ 0] +
X∼G
N (0,Σ)
η2
.
d log(|G|/δ)
• For any even degree-4 polynomial p : Rd → R we have
E[p(X)] −
G
E [p(X)] ≤ η Var [p(X)]1/2 ,
N (0,Σ)
N (0,Σ)
Pr[p(X) ≥ 0] ≤ Pr [p(X) ≥ 0] +
G
N (µ,I)
η2
.
2 log(1/ε)(d log(|G|/δ))2
As before, it is easy to show (see Lemma 14) that given enough samples from N (0, Σ), the empirical
data set will satisfy these conditions with high probability.
2.3
Bounds on the Total Variation Distance
We will require some simple bounds on the total variation distance between two Gaussians. These
bounds are well-known. Roughly speaking, they say that the total variation distance between
two Gaussians with identity covariance is governed by the ℓ2 norm between their means, and the
total variation distance between two Gaussians with mean zero is governed by the Frobenius norm
between their covariance matrices, provided that the matrices are close to the identity.
Lemma 1. Let µ1 , µ2 ∈ Rd be such that kµ1 − µ2 k2 = ε for ε < 1. Then
1
dTV (N (µ1 , I), N (µ2 , I)) = √ + o(1) ε .
2π
For clarity of exposition we defer this calculation to the Appendix.
We also need to bound the total variation distance between two Gaussians with zero mean and
different covariance matrices. The natural norm to use is the Mahalanobis distance. But in our
setting, we will be able to use the more convenient Frobenius norm instead (because we effectively
reduce to the case that the covariance matrices will be close to the identity):
b be such that kΣ−IkF ≤ O(ε log 1/ε), and kΣ− Σk
b F ≤
Lemma 2 (Cor. 2.14 in [DKK+ 16]). Let Σ, Σ
b ≤ O(ε).
Cε. Then dTV (N (0, Σ), N (0, Σ))
These lemmata show that parameter estimation and approximation in total variation distance
are essentially equivalent. Indeed, in this paper, we achieve both guarantees, but state our results
in terms of total variation estimation.
3
Robustly Learning the Mean in Low Dimensions
This section is dedicated to the proof of the following theorem:
8
Theorem 3. Fix µ ∈ Rd , and let ε, γ, δ > 0. Let S0 = (G0 , E0 ) be such that G0 is a (γε, δ)-good
set with respect to N (µ, I), and |E0 |/|S0 | ≤ ε. Let S = (G, E) be another set with ∆(S, S0 ) < ε.
Let V ⊆ Rd be a subspace. For all 0 < ρ < 1, the algorithm LearnMeanLowD(V, γ, ε, δ, S, ρ)
runs in time poly(d, |S|, (1/ρ)O(dim(V )) , log(ρε/(1 − ρ)), log(1/ρ)) and returns a µ
e so that
kΠV (µ − µ
e)k2 =
γ
1 + 2ρ √
ε.
π+O
1−ρ
d
√
In particular, as we let ρ, γ → 0, the √
parameter estimation error approaches πε (corresponding to
a total variation approximation of ε/ 2). In Lemma 17
p in the Appendix we show√that no algorithm
can achieve parameter estimation error better than π2 ε. Thus, we achieve a 2 approximation
to the optimal error.
For simplicity, in the rest of this section, we will let V = Rd , that is, we assume there is no
projection. It should be clear that this can be done without loss of generality. Our algorithm
proceeds as follows: First, we show that in one dimension, the median produces an estimate which
is optimal, up to lower order terms, if the sample set is (γε, δ)-good with respect to the underlying
d
Gaussian. Then, we show that
p π by using a net argument, we can produce a convex body in R with
diameter at most 2R = 2( 2 + o(1))ε which must contain the true mean. Finally, we use an old
√
result of Jung [Jun01] that such a set can be circumscribed by a ball of radius 2R (see [BW41]
for an English language version of the result). We use the center of the ball as our estimate µ
e.
3.1
Robustness of the Median
First we show that if we projectponto one dimension, then the median of the corrupted data differs
from the true mean by at most π2 ε + o(ε). Our proof will rely only on the notion of a (γε, δ)-good
set with respect to N (µ, I) and thus it works even in the strong contamination model. Formally,
we show:
Lemma 3. Fix any v ∈ Rd . Fix µ ∈ Rd , and let δ > 0. Let S0 = (G0 , E0 ) be so that G0 be a
set
with
(γε, δ)-good set with respect to N (µ, I), and |E0 |/|S0 | ≤ ε. Let S = (G, E) be another
pπ
γ
ε.
∆(S, S0 ) < ε. Let b be the median of S when projected onto v. Then, |b − Πv µ| ≤
2 +O d
Proof. For any a ∈ R, we have
Pr [hv, Xi > a] =
X∼S
Observe that we have
|E|
|S|
|G|
|E|
Pr [hv, Xi > a] +
Pr [hv, Xi > a] .
|S| X∼G
|S| X∼E
PrX∼E [hv, Xi > a] ≤ ψ(S, G). Moreover, by simple calculation we have
Pr [hv, Xi > a] − Pr [hv, Xi > a] ≤ 2φ(S, G) .
X∼G
X∼G0
Hence, we have
Pr [hv, Xi > a] −
X∼S
|G|
Pr [hv, Xi > a] ≤ ψ(S, G) + o(ε) ,
|S| X∼G0
since by assumption ∆(S, G) ≤ ε. Similarly, we have that for all a ∈ R,
Pr [hv, Xi < a] −
X∼S
|G|
Pr [hv, Xi < a] ≤ ψ(S, G) + o(ε) .
|S| X∼G0
9
For |a| = O(ε) we have that PrN (0,I) [X > a] = 12 − √12π a + O(ε3 ). Thus, by (γε, δ)-goodness of G0 ,
this implies that for |a| = O(ε), we have
γε
|G| 1
1
Pr [hv, Xi > Πy µ + a] −
.
(2)
− √ a ≤ ψ(S, G) + O
X∼S
|S| 2
d
2π
Thus, for |a| = O(ε) we have
r
γε
|G| 1
1
|E|
π
+ o(ε)
a] ≤
−√ a +
+O
Pr [hv, Xi > Πy µ +
X∼S
2
|S| 2
|S|
d
2π
γε
1 |G| 1
|E|
√ a+
≤ −
+ o(ε) .
+O
2
|S| 2π
2|S|
d
pπ
p
In particular, we see that if a > π2 ε + O γε
d + o(ε), then PrX∼S
p π hv, Xi > Πy µ + 2 ε < 1/2.
By symmetric logic, we also have that Pr
hv, Xi > Πy µ − 2 ε > 1/2. Thus, the median in
pX∼S
π
direction v differs from Πy µ by at most 2 ε + O γε
d + o(ε).
3.2
Finding a Minimum Radius Circumscribing Ball
For any x ∈ Rd and r > 0, let B(x, r) = {y ∈ Rd : kx − yk2 ≤ r} denote the closed ball of
radius r centered at x. The following classical result of Gale gives a bound on the radius of the
circumscribing ball of any convex set in terms of its diameter:
Theorem 4 (see [Jun01, BW41]). Fix R > 0. Let C ⊆ Rd be a convex √
body so that for all x, y ∈ C,
we have kx − yk2 ≤ 2R. Then C is contained within a ball of radius R 2.
The bound is asymptotically achieved for the standard simplex as we increase its dimension.
The goal of this subsection is to show that the (approximately) minimum radius circumscribing
ball can be found efficiently. We will assume we are given an approximate projection oracle for the
convex body that given a point y ∈ Rd , outputs a point which is almost the closest point in C to x:
Definition 5. A ρ-projection oracle for a convex body C is a function O : Rd → Rd , which, given
a point y ∈ Rd , outputs a point x ∈ C so that kx − yk2 ≤ inf x′ ∈C kx′ − yk2 + ρ.
Our first step is to use such an oracle to construct a net for C. First, we need the following
well-known bound on the size of the net.
Claim 1. Fix r > 0. Then, for any β > 0, there is a β-net F for the sphere of radius r around 0
in Rd of size (r/β)O(d) . Moreover, this net can be constructed in time poly(d, |F|).
With this, we can show:
Lemma 4. Fix R, C as in Theorem 4, and let 1 > ρ > 0. Let x ∈ C be arbitrary. Let O be a
(ρR/3)-projection oracle for C. Suppose a call to O runs in time T . Then, there is an algorithm
CircumscribeNet(R, ρ, O, x) which runs in time poly((R/ρ)O(d) , T ) and outputs a set X ⊆ Rd
so that X is a (ρR)-net for C, and moreover, |X | ≤ (R/ρ)O(d) .
The algorithm is fairly straightforward. First, we observe that C is contained within B(x, 2R).
We then form a (ρR)/3-net of B(x, 2R) using Claim 1. We then iterate over every element v of this
net, and use our projection oracle to (approximately) find the closest point in C to v. If this point
is too far away, we throw it out, otherwise, we add this projected point into the net. The formal
pseudocode for CircumscribeNet is given in Algorithm 1.
10
Algorithm 1 Generating a net of C
1: function CircumscribeNet(R, ρ, O, x)
2:
Form an ρ/3-net F ′ of the sphere of radius 2 of size (1/ρ)O(d) as in Claim 1.
3:
Let F = R · F ′ + x.
4:
Let X ← ∅
5:
for each v ∈ F do
6:
Let uv ← O(v)
7:
if kv − uv k2 ≤ 2ρR/3 then
8:
Add uv to X
9:
end if
10:
end for
11:
return X
12: end function
Proof. The runtime bound follows from Claim 1. We now turn our attention to correctness. By
Claim 1, and rescaling and shifting, the set F is clearly a (ρR)/3-net for a ball B of radius 2R
containing C. We now claim that the set X is indeed a (ρR)/3-net for C. Fix y ∈ C. Since C ⊆ B,
this implies there is some v ∈ F so that ky − vk2 ≤ ρR/3. Thus, in Line 7, when processing v,
we must find some uv ∈ C so that kuv − vk2 ≤ 2ρR/3. The claim then follows from the triangle
inequality.
With this, we obtain:
Corollary 2. Fix R, C, ρ, O, x as in Lemma 4. Suppose a call to O runs in time T . Then, there is
O(d) , T ) and returns a point
an algorithm Circumscribe(R, ρ, O, x) which runs
√ in time poly((R/ρ)
yb so that C is contained within a ball of radius 2(1 + 2ρ)R.
The algorithm at this point is very simple. Using the output of CircumscribeNet, we iterate
over all points √
in a net over B(x, 2R), find an x in this net so that the distance to all points in the
net is at most 2(1 + ρ)R, and output any such point. The formal pseudocode for Circumscribe
is given in Algorithm 2.
Algorithm 2 Finding a circumscribing ball of small radius
1: function Circumscribe(R, ρ, O, x)
2:
Form an ρ/3-net F ′ of B(0, 2) of size (1/ρ)O(d) as in Claim 1.
3:
Let F = R · F ′ + x.
4:
Let X ← CircumscribeNet(R, ρ, O, x).
5:
for each v ∈ F do
√
6:
if for all u ∈ X , we have ku − vk2 ≤ 2(1 + ρ)R then
7:
return u
8:
end if
9:
end for
10: end function
Proof. The√runtime bound is immediate. By Theorem 4, there is some y ∈ B(x, √
2R) so that
′
C ⊆ B(y, R 2). Thus, by the triangle inequality, there is some y ∈ F so that C ⊆ B(y, 2(1+ρ)R).
Thus, the algorithm will output some point y ′′ ∈ F. By an additional
application of the triangle
√
′′
inequality, since X is a ρR-net for C, this implies that C ⊆ B(y , 2(1 + 2ρ)R), as claimed.
11
3.3
The Full Low-Dimensional Algorithm
We now have all the tools to describe the full algorithm in low-dimensions. Let S be our corrupted
dataset as in Theorem 3. Fix ρ > 0. We first produce a ρ-net F over the unit sphere in Rd .
By (a slight modification of) Claim 1, this net has size (1/ρ)O(d) and can be constructed in time
poly(d, |F|). For each v ∈ F, we project all points in S onto v, and take the median of these points
to produce bv . We then construct the following set:
\
C=
{y ∈ Rd : hv, yi ∈ [bv − β, bv + β]} ,
(3)
v∈F
pπ
γε
where β = 2 ε + O d + o(ε) is as in Lemma 3. We now show two properties of this set, which
in conjunction with the machinery above, allows us to prove Theorem 3. The first shows that C
has small diameter:
Claim 2. For all x, y ∈ C, we have kx − yk2 ≤ 2β/(1 − ρ).
Proof. Fix any x, y ∈ C. By definition of C, it follows that for all v ∈ F, we have |hx − y, vi| ≤ 2β.
For any u with kuk2 = 1, there is some v ∈ F with ku − vk2 ≤ ε, and so we have
|hx − y, ui| ≤ |hx − y, vi| + |hx − y, u − vi|
≤ 2β + ρkx − yk2 .
Taking the supremum over all unit vectors u and simplifying yields that kx − yk2 ≤ 2β/(1 − ρ), as
claimed.
The second property shows that we may find an α-projection oracle for C efficiently.
Claim 3. Fix ρ′ > 0. There is a ρ′ -projection oracle ProjOracle(y, ρ′ , C) for C which runs in
time poly((1/ρ)O(d) , log(γε/(1 − ρ)), log(1/ρ′ )).
Proof. The projection problem may be stated as
min kx − yk2 s.t.hv, yi ∈ [bv − β, bv + β], ∀v ∈ F .
This is a convex minimization problem with linear constraints. By the classical theory of optimization [GLS88], finding a ρ-approximate y can be done in poly(d, log(vol(C)/ρ′ )) queries to a
separation oracle for C. Since the separation oracle must only consider the constraints induced by
F, this can be done in time (1/ρ)O(d) . Since by Claim 2 we have vol(C) ≤ (2β/(1 − ρ))O(d) , the
desired runtime follows immediately.
We now finally describe LearnMeanLowD. Using convex optimization, we first find an arbitrary x ∈ C. By Lemma 3 we know µ ∈ C and so this step succeeds. After constructing C,
we run Circumscribe with appropriate parameters, and return the outputted point. The formal
pseudocode for LearnMeanLowD is given in Algorithm 3.
Proof. The runtime claim follows from the runtime claims for Circumscribe and ProjOracle.
Thus, it suffices to prove correctness of this algorithm.
C. By
√ 1+2ρBy Lemma 3, we know that µ√∈ 1+2ρ
Claim 2 and Corollary 2, the output y satisfies B(y, 2 1−ρ β). Thus, we have kµ−yk2 ≤ 2 1−ρ β,
as claimed.
12
Algorithm 3 Finding a circumscribing ball of small radius
1: function LearnMeanLowD(ε, δ, S, ρ)
2:
Form a ρ-net F of B(0, 1) of size (1/ρ)O(d) as in Claim 1.
3:
for each v ∈ F do
4:
Let bv be the median of S projected onto v.
5:
end for
6:
Form C as in Equation (3).
7:
Find an p
x ∈ C using convex
optimization.
+
o(ε)
8:
Let β = π2 ε + O γε
d
9:
Let R = β/(1 − ρ)
10:
Let O(·) = ProjOracle(·, (ρR)/3, C)
11:
return the output of Circumscribe(R, ρ, O, x)
12: end function
4
Robustly Learning the Mean in High Dimensions
In this section, we prove the following theorem, which is our first main result:
Theorem 5. Fix ε, γ, δ > 0, and let X1 , . . . , Xn be an ε-corrupted set of points from N (µ, I), where
kµk2 ≤ O(ε log 1/ε), and where
(d log(d/γεδ))6
.
n=Ω
γ 2 ε2
Then, for every α, β > 0, there is an algorithm RecoverMean(X1 , . . . , Xn , ε, δ, γ, α, β) which
runs in timepoly(d,
1/γ, 1/εβ , 1/α, log 1/δ) and outputs a µ
b so that with probability 1 − δ, we have
√
π+O(γ)
1
kb
µ − µk2 ≤
+ √β ε.
1−α
In particular, observe that Theorem 5, in conjunction with Lemma 1, gives us Theorem 1, if we set
γ = o(1). With this, we may state our primary algorithmic contribution:
Theorem 6. Fix ε, γ, α, δ, β > 0, and let S0 = (G0 , E0 ) be an ε-corrupted set of samples of size
n from N (µ, I), where kµk2 ≤ O(ε log 1/ε), and where n = poly(d, 1/(γε), log 1/δ). Suppose that
G0 is (γε, δ)-good with respect to N (µ, I). Let S ⊆ S0 be a set so that ∆(S, G0 ) ≤ ε. Then, there
exists an algorithm FilterMeanOpt that given S, ε, γ, α, β outputs one of two possible outcomes:
√
π+O(γ)
√1
+
ε.
(i) A µ
b, so that kb
µ − µk2 ≤
1−α
β
(ii) A set S ′ ⊂ S so that ∆(S ′ , G0 ) < ∆(S, G0 ).
Moreover, FilterMeanOpt runs in time poly(d, 1/γ, 1/εβ , 1/α, log 1/δ).
+
By
p first running the algorithm of [DKK 16] to obtain an estimate of the mean to error
O(ε log 1/ε), then running FilterMeanOpt at most polynomially many times, we clearly recover the guarantee in Theorem 5. Thus, the rest of the section is dedicated to the proof of Theorem
6.
At a high level, the structure of the argument is as follows: We first show that if there is a
subspace of eigenvectors of dimension at least O(log 1/ε) of the empirical covariance matrix with
large associated eigenvalues, then we can produce a filter using a degree-2 polynomial (Section 4.1).
Otherwise, we know that there are at most O(log 1/ε) eigenvectors of the empirical covariance
13
with a large eigenvalue. We can learn the mean in this small dimensional subspace using our
learning algorithm from the previous section, and then we can argue that the empirical mean on
the remaining subspace is close to the true mean (Section 4.2).
This outline largely follows the structure of the filter arguments given in [DKK+ 16], however,
the filtering algorithm we use here requires a couple of crucial new ideas. First, to produce the
filter, instead of using a generic degree-2 polynomial over this subspace, we construct an explicit,
structured, degree-2 polynomial which produces such a filter. Crucially, we can exploit the structure
of this polynomial to obtain very tight tail bounds, e.g., via the Hanson-Wright inequality. This is
critical to avoid a quasi-polynomial runtime. If instead we used arbitrary degree-2 polynomials in
this subspace, it would need to be of dimension O(log2 1/ε) and the low-dimensional algorithm in
the second step would take quasi-polynomial time.
Second, we must be careful to throw out far fewer good points than corrupted points. In particular, by our definition of ∆ (which gives an additional logarithmic penalty to discarding good
points) and our guarantee that ∆ decreases, our filter can only afford to throw out an ε/ log(1/ε)
fraction of good points in total, since ∆ is initially ε. This is critical, as if we threw away an
ε-fraction of good points, then proving that the problem remains efficiently solvable becomes problematic. In particular, if these points were thrown away arbitrarily, then this becomes the full
additive and subtractive model, for which a statistical query lower bound prevents us from getting
an O(ε)-approximate answer in polynomial time [DKS17]. To avoid discarding too many good
points, we exploit tight exponential tail bounds of Gaussians, and observe that by slightly increasing the threshold at which we filter away points, we decrease the fraction of good points thrown
away dramatically.
4.1
Making Progress with Many Large Eigenvalues
We now give an algorithm for the case when there are many eigenvalues which are somewhat large.
Formally, we show:
Theorem 7. Fix ε, γ, δ, α, β > 0, and let S0 = (G0 , E0 ) be an ε-corrupted set of samples of size
n from N (µ, I), where kµk2 ≤ O(ε log 1/ε), and where n = poly(d, 1/(γε), log 1/δ). Suppose that
b be the
G0 is (γε, δ)-good with respect to N (µ, I). Let S ⊆ S0 be a set so that ∆(S, G0 ) ≤ ε. Let Σ
sample covariance of S, let µ
b be the sample mean of S, and let V be the subspace of all eigenvectors
b − I with eigenvalue more than 1 ε. Then, there exists an algorithm FilterMeanManyEig
of Σ
β
that given S, ε, γ, δ, α, β outputs one of two possible outcomes:
1. If dim(V ) ≥ C1 β log(1/ε), then it outputs an S ′ so that ∆(S ′ , G0 ) < ∆(S, G0 ).
2. Otherwise, the algorithm outputs “OK”, and outputs an orthonormal basis for V .
b − I, and if there are too
Our algorithm works as follows: It finds all large eigenvalues of Σ
many, produces an explicit degree-2 polynomial which, as we will argue, produces a valid filter.
The formal pseudocode for our algorithm is in Algorithm 4.
For clarity of exposition, we defer the proof of Theorem 7 to Appendix C.
4.2
Returning an Estimate When There are Few Large Eigenvalues
At this point, we have run the filter of Algorithm 4 until there are few large eigenvalues. In the
subspace with large eigenvalues, we again run the low dimensional algorithm to obtain an estimate
for the mean in this subspace. Recall that Lemma 3 guarantees the accuracy of this estimator
within this subspace. In the complement of this subspace, where the empirical covariance is very
14
Algorithm 4 Filter if there are many large eigenvalues of the covariance
1: function FilterMeanManyEig(S, ε, γ, δ, α, β)
2:
Let C1 , C2 , C3 > 0 be sufficiently large constants.
b be the empirical mean and covariance of S, respectively.
3:
Let µ
b and Σ
b − I with eigenvalue more than
4:
Let V be the subspace of Rd spanned by eigenvectors of Σ
1
β ε.
5:
if dim(V ) ≥ C1 β log(1/ε) then
6:
Let V ′ be a subspace of V of dimension C1 β log(1/ε).
√
7:
Let µ̃ be an approximation to ΠV ′ (µ) with ℓ2 -error π+O(γ)
1−α ε, computed using
LearnMeanLowD(V, γ, ε, δ, S, γ).
8:
Let p(x) be the quadratic polynomial p(x) = kΠV ′ (x) − µ̃k22 − dim(V ′ ).
9:
Find a value T > 0 so that either:
(a) T > C2 d log(|S|/δ) and p(x) > T for at least one x ∈ S, or
(b) T > 2C3 log(1/ε)/c0 and PrS (p(x) > T ) > exp(−c0 T /(2C3 )) + γε/(d log(|S|/δ)).
10:
11:
12:
13:
14:
return S ′ = {x ∈ S : p(x) ≤ T }
else
return an orthonormal basis for V .
end if
end function
close to the identity, Lemma 5 (stated below) shows that the empirical mean is close to the true
mean. This leads to a simple algorithm which outputs an estimate for the mean, described in
Algorithm 5.
Algorithm 5 Return a mean if there are few large eigenvalues of the covariance
1: function FilterMeanFewEig(S, ε, γ, δ, α, β, V )
√
π+O(γ)
2:
Let µ̃V be an approximation to ΠV (µ) with ℓ2 -error
1−α ε, computed using
LearnMeanLowD(V, γ, ε, δ, S, γ).
3:
Let µ̃V ⊥ be the empirical mean on V ⊥ , ΠV ⊥ µ̂.
4:
return µ̃V + µ̃V ⊥ .
5: end function
Lemma 5. Let µ, η, G0 , S be as in Theorem 7. Let µ
b be the sample mean of S, and let v be a unit
ε
vector. Suppose that hv, µ − µ
bi > β 1/2
. Then VarS [hv, Xi] > 1 + βε .
For clarity of exposition, we defer the proof of Lemma 5 to Appendix C.
4.3
The Full High-Dimensional Algorithm
We now have almost all the pieces needed to prove the full result. The last ingredient is the
fact that, given enough samples, the good set condition is satisfied by the samples from the true
distribution. Formally,
Lemma 6. Fix η, δ > 0. Let X1 , . . . , Xn be independent samples from N (µ, I), where n =
Ω((d log(d/ηδ))6 /η 2 ). Then, S = {X1 , . . . , Xn } is (η, δ)-good with respect to N (µ, I) with probability at least 1 − δ.
15
Proof. This follows from Lemmas 8.3 and 8.16 of [DKK+ 16].
At this point, we conclude with the proof of Theorem
5. Within the subspace V , Lemma
√
π+O(γ)
3 guarantees that the mean is accurate up to ℓ2 -error 1−α ε. Within the subspace V ⊥ , the
ε
contrapositive of the statement of Lemma 5 guarantees the mean is accurate up to ℓ2 -error β 1/2
.
The desired result follows from the Pythagorean theorem.
4.4
An Extension, with Small Spectral Noise
For learning of arbitrary Gaussians, we will need a simple extension that allows us to learn the mean
even in the presence of some spectral norm error in the covariance matrix. Since the algorithms
and proofs are almost identical to the techniques above, we omit them for conciseness. Formally,
we require:
Theorem 8. Fix χ, ε, δ > 0, and let X1 , . . . , Xn be an ε-corrupted set of points from N (µ, Σ),
where kΣ − Ik2 ≤ O(χ), kµk2 ≤ O(ε log 1/ε), and where n = poly(d, 1/χ, 1/ε, log 1/δ). For
any γ > 0, there is an algorithm RecoverMeanNoisy(X1 , . . . , Xn , ε, δ, γ, χ) which runs in time
poly(d, 1/χ, 1/ε, log 1/δ) and outputs a µ
b so that with probability 1 − δ, we have kb
µ − µk2 ≤
(C + γ)ε + O(χ).
This extension follows from two elementary observations:
1. For the learning in low dimensions, observe that the median is naturally robust to error in the
covariance, and in general, by the same calculation we did, the error of the median becomes
O(ε + α).
2. For the filter, observe that we only need concentration of squares of linear functions, and
whatever error we have in this concentration goes directly into our error guarantee. Thus, by
the same calculations that we had above, if we filtered for eigenvalues above 1 + O(ε + α), we
would immediately get the desired bound.
5
Robustly Estimating the Mean of Degree Two Polynomials
In this section, we give robust estimates of E[p2 (X)] for degree-2 polynomials p in subspaces of
small dimension, which is an important prerequisite to learning the covariance in high-dimensions.
A crucial ingredient in our algorithm is the following improvement theorem (stated and proved in
the next section) which shows how to take any weak high-dimensional estimate for the covariance
and use it to get an even better robust estimate for E[p2 (X)].
5.1
Additional Preliminaries
Here we give some additional preliminaries we require for the low-dimensional learning algorithm
we present here. We will need the following well-known tail bound for degree-2 polynomials:
Lemma 7 (Hanson-Wright Inequality [LM00, Ver10]). Let X ∼ N (0, I) ∈ Rd and A be a d × d
matrix. Then for some absolute constant c0 , for every t ≥ 0,
2
t
t
T
T
,
.
Pr X AX − E[X AX] > t ≤ 2 exp −c0 · min
kAk2F kAk2
We will also require the following lemmata:
16
Lemma 8 (Hölder’s inequality for Schatten norms). Let A, B be matrices. Then, for all p, q so
that 1p + 1q = 1, we have kABkS 1 ≤ kAkS p kBkS q .
This implies the following corollary:
b M be so that kΣ − Σk
b F ≤ O(δ), and so that kM kF = 1. Then, we have
Corollary 3. Let Σ, Σ,
1/2
1/2
1/2
1/2
b
b
kΣ M Σ − Σ M Σ kS 1 ≤ 5δ.
Proof. We have
b 1/2 M Σ
b 1/2 kS 1 ≤ kΣ1/2 M Σ1/2 − M kS 1 + kM − Σ
b 1/2 M Σ
b 1/2 kS 1 .
kΣ1/2 M Σ1/2 − Σ
We will bound the first term on the RHS by 5δ/2; the second term is bounded symmetrically. We
have
kΣ1/2 M Σ1/2 − M kS 1 ≤ kΣ1/2 M Σ1/2 − M Σ1/2 kS 1 + kM Σ1/2 − M kS 1
kΣ1/2 − IkS 2 kM Σ1/2 kS 2 + kΣ1/2 − IkS 2 kM kS 2
≤ 5δ/2 ,
where the last line follows from Hölder’s inequality for Schatten norms.
5.2
An Improvement Theorem
Here we state and prove one of the main technical ingredients in our algorithm for robustly learning
the covariance.
Theorem 9. Fix ε, δ, τ > 0. Let Σ be so that kΣ − IkF ≤ O(ε log 1/ε), and fix a p ∈ P2 , where P2
denotes the set of even degree-2 polynomials in d variables. Let G0 be an (ε, δ)-good set of samples
from N (0, Σ), and let S = {X1 , . . . , Xn } be so that ∆(S, G0 ) ≤ ε. Then, for any C > 0 there is an
algorithm LearnMeanChiSquared which, given p, X1 , . . . , Xn , and ε, outputs a µ
b so that with
probability 1 − τ over the randomness of the algorithm,
µ
b−
E
X∼N (0,Σ)
[p(X)] ≤ kΣ − IkF /C + O(log(C)ε) .
Moreover, the algorithm runs in time O(|S| + log(1/τ )/ε2 ).
The way to think about how this result fits into the overall strategy is that robustly estimating
the covariance is equivalent to robustly estimating the mean of every (normalized) degree-two
polynomial p. The above theorem shows how a weak estimate in high-dimensions can be used
to obtain stronger estimates in one dimension, which ultimately we will use to improve the highdimensional estimate as well. The above theorem is the workhorse in our proof.
Our algorithm itself is simple, however, its correctness is quite non-trivial. We define some
threshold T . Given our corrupted set of samples from N (0, Σ), we use our corrupted data set to
estimate the mean of p(X) conditioned on the event that |p(X)| ≤ T . Then, to estimate the contribution of the mean from points X so that |p(X)| > T , we estimate this by EX∼N (0,I) [p(X)1|p(X)|>T ].
In other words, we are replacing the contribution of the true tail by an estimate of the contribution
of p(X) when X ∼ N (0, I) on this tail. The formal pseudocode is given in Algorithm 6.
Intuitively, this algorithm works because of two reasons. First, it is not hard to show that the
influence of points p(X) within the threshold T on the estimator are bounded by at most T . Hence,
the adversary cannot add corrupted points within this threshold and cause our estimator to deviate
17
too much. Secondly, because we know that kΣ − IkF is small, by carefully utilizing smoothness
properties of sums of chi-squared random variables, we are able to show that our estimate for the
contribution of the tail is not too large. At a high level, this is because “most” of the distance
between two chi-squared random variables must remain close to the means, so the difference in the
tails is much smaller. Proving that this holds in a formal sense is the majority of the technical
work of this section.
Proof. We know the distribution of p(X ′ ) for X ′ ∼ N (0, I) explicitly and wish to use this to get
a better estimate for the mean of p(X) for X ∼ N (0, Σ) than might be given by the mean of the
ε-corrupted set of samples.
Algorithm 6 Approximating E[p(X)] for X ∼ N (0, Σ) with corrupted samples.
1: function LearnMeanChiSquared(X1 , . . . , Xn , p(x), ε, τ )
2:
Let T = O(log
C).
x − T, for x ≥ T
3:
Let f (x) =
0,
for |x| ≤ T .
x + T, for x ≤ −T
P
4:
Compute α = ni=1 (p(Xi ) − f (p(Xi )))/n.
′ from X ′ ∼ N (0, I).
2
5:
Simulate m = O((ln
X1′ , . . . , Xm
Pmτ )/ε ) samples
′
6:
Return µ
b = α + i=1 f (p(Xi ))/n.
7: end function
It follows from (ε, δ)-goodness that | PrX ′ ∼N (0,Σ) [p(X ′ ) > t] − #{Xi : p(Xi ) > t}/n| ≤ 2ε for all
t. We need to express the expectation in terms that we can use this to bound. For Z = p(X ′ ), we
have that
Z T
Z T
Pr[Z < −t]dt .
Pr[Z > t]dt −
E[Z − f (Z)] = E[max{T, min{−T, Z}}] =
0
0
Similarly, the samples have
α=
n
X
i=1
(p(Xi ) − f (p(Xi ))/n =
Z
T
0
#{Xi : p(Xi ) > t}
dt −
n
Z
t
0
#{Xi : p(Xi ) < −t}
dt .
n
Thus, we have | E[Z − f (Z)] − α| ≤ 2T ε.
Since p ∈ P2 , we have E[p(X ′ )] = 1 for X ′ ∼ N (0, I). Thus, we have Var[f (p(X ′ ))] ≤
E[f (p(X ′ ))2 ] ≤ E[p(X ′ )2 ] = 1. It follows by
results that the empirical after
Pstandard concentration
′ )/n − E[p(X ′ )]| ≤ ε with probability 1 − τ .
taking m = O(ln(1 − τ )/ε2 ) samples has | m
f
(p(X
)
i
i=1
When this holds, we have
µ
b−
E
X∼N (0,Σ)
[p(X)] ≤ (2T + 1)ε +
E
[f (p(X))] −
X∼N (0,I)
E
[f (p(X))] .
X∼N (0,Σ)
To prove the correctness of the algorithm it remains to show that:
Lemma 9. For any constant C > 0, for T = O(log C), we obtain
E
[f (p(X))] −
X∼N (0,I)
E
[f (p(X ′ ))] ≤
X ′ ∼N (0,Σ)
18
kΣ − IkF
.
C
Proof. Let M be the symmetric matrix with kM kF = 1 such that p(x) = xT M x for x ∈ Rd . We
can write p(X) with X ∼ N (0, I) as
p(X) + tr(M ) = X T M X = X T O T DOX = Y T DY =
d
X
ai Yi2 ,
i=1
where O is orthogonal, D is diagonal, ai are the eigenvalues
of M , and Y ∼ N (0, I) hence Yi are
P 2
i.i.d. from N (0, 1). Since kM kF = 1, here we have
ai = 1. If instead we express p(X ′ ) in terms
of X ′ ∼ N (0, σ 2 ), we obtain:
′
p(X ) + tr(M ) = Y
′T
Σ
1/2
MΣ
1/2
′
Y =Y
′T
′T
′
′
′
T
′
O DOY =Y DY =
d
X
bi Yi2 ,
i=1
where O′ is orthogonal, D ′ is diagonal, bi are the eigenvalues of Σ1/2 M Σ1/2 , and Y ′ , Y ∼ N (0, I)
hence Yi are i.i.d. from N (0, 1). P
By Corollary 3, we have that i |bi − ai | ≤ (5/2)kΣ − IkF . Now consider the random variables
Zi,λ = −tr(M ) +
where
d
X
j=1
cj (Yi2 − 1) ,
bj ,
cj = (1 − λ)ai + λbi
aj
for j < i
for j = i ,
for j > i
for 1 ≤ i ≤ d and 0 ≤ λ ≤ 1. Note that Zi,1 = Zi+1,0 , for 1 ≤ i ≤ d − 1. Note that, to prove the
lemma, it suffices to show that
X
2|bi − ai |/5C .
| E[f (Z1,0 )] − E[f (Zd,1 )]| ≤
i
To this end,
Pconsider how
P E[f (Zi,λ )] varies with λ. We can write Zi,λ = Z−i + Zi , where Z−i =
−tr(M ) + j<i bj Yj2 + j>i aj Yj2 ; and Zi = ci Yi2 , where ci = ((1 − λ)ai + λbi ). We assume for
now that ci 6= 0. Since only Zi depends on λ, we have
d E[f (Zi,λ )]
d
=
E[f (Z−i + Zi )]
dλ
dλ
d
E [f (Z−i + Zi )]
= E
Z−i dλ Zi
Z ∞
d
= E
f (Z−i + x)Pi (x)dx
Z−i dλ −∞
Z ∞
f (Z−i + x)dPi (x)/dλdx ,
= E
Z−i
−∞
where Pi (x) is the probability density function of the random variable Zi . Standard results about
the χ2 distribution give that:
Fact 1. Let Y1 , Y2 , Y3 ∼ N (0, 1). Then the probability density function of Y12 is
Y12
+
Y22
is
1 −x/2
,
2e
and of
Y12
+
Y22
+
Y32
is
√
√ x e−x/2 .
2π
19
√ 1 e−x/2 ,
2πx
of
This gives that Pi (x) =
√ 1
e−x/2ci .
2πxci
Now consider the derivative:
1
1
/dλ · e−x/2ci + √
· de−x/2ci /dλ
2πxci
2πxci
1
dci
x
1
dci −x/2ci
=− √
· 2 · e−x/2ci ·
·e
+√
·
3/2 dλ
dλ
2πxci 2ci
2 2πxci
√
(bi − ai ) x −x/2ci
bi − a i
e−x/2ci + √
e
=− √
3/2
5/2
2 2πxci
2πci
dPi (x)/dλ = d √
(3)
= ((bi − ai )/2ci )(Pi (x) − Pi (x)) ,
(3)
where Pi (x) is the distribution of Zi + Zi′ + Zi′′ , where Zi′ and Zi′′ are i.i.d. copies of Zi . We thus
have
Z ∞
d E[f (Zi,λ )]
f (Z−i + x)dPi (x)/dλdx
= E
Z−i
dλ
−∞
= ((bi − ai )/2ci ) E [E[f (Z−i + Zi )] − E[f (Z−i + Zi + Zi′ + Zi′′ )]]
Z−i
= ((bi − ai )/2ci ) E[f (Zi,λ ) − f (Zi,λ + Zi′ + Zi′′ )]
= ((bi − ai )/2ci ) E [f (Zi,λ ) − ′E ′′ [f (Zi,λ + Zi′ + Zi′′ )]] .
Zi,λ
Zi ,Zi
Since f has Lipschitz constant 1, |f (Zi,λ ) − EZi′ ,Zi′′ [f (Zi,λ + Zi′ + Zi′′ ]| ≤ E[Zi′ + Zi′′ ] = 2ci whatever
value Zi,λ takes.
Using the probability distribution of Zi′ + Zi′′ , in the case ci > 0, we have for all |z| ≤ T ,
Z ∞
′′
′
e−x/2ci /2ci · (z + x − T )dx
E ′′ [f (z + Zi + Zi )] =
′
Zi ,Zi
T −z
−(T −z)/2ci
=e
(z − T + 2ci + (T − z))
= 2ci e−(T −z)/2ci .
Since ci ≤ max ai , bi ≤ 1 + O(ε log(1/ε)) ≤ 2, for large enough T = O(log C), we have e−(T /2)/2ci ≤
1/C 2 T . We assume that this holds.
For −T ≤ z ≤ T /2, we have 2ci e−(T −z)/2ci ≤ 2ci /C 2 T . A similar argument when ci < 0 gives
that for −T /2 ≤ z ≤ T , we have 2|ci |e−(T −z)/2ci ≤ 2|ci |/C 2 T . Now we have enough to show that
d E[f (Zi,λ )]
≤ |(bi − ai )/2ci || E [f (Zi,λ ) − ′E ′′ [f (Zi,λ ) + Zi′ + Zi′′ )]|
Zi,λ
dλ
Zi ,Zi
≤ Pr [|Zi,λ − tr(M )| ≥ T /2]|(bi − ai )| + |(bi − ai )|/C 2 .
Zi,λ
By the Hanson-Wright inequality, we have that for any x,
√
Pr [|Zi,λ − E[Zi,λ ]| ≥ 2kck2 x + kck∞ x] ≤ 2e−x ,
Zi,λ
where c = (b1 , . . . , bi − 1, ci , ai+1 , . . . , ad ). Note that kck2 ≤ 1 + O(kΣ − IkF ) ≤ 2. Also, E[Zi,λ ] +
P
tr(M ) is the sum of coordinates of c, and so | E[Zi,λ ]| ≤ ij=1 |bj − aj | ≤ O(kΣ − IkF ) ≤ 1. Putting
this together, we obtain
Pr [|Zi,λ − tr(M )| ≥ T /2] ≤ 2 exp(−((T − 1)/8)2 ) ≤ 1/C 2 .
Zi,λ
20
Finally, we have that, assuming ci 6= 0,
d E[f (Zi,λ )]
≤ 2|(bi − ai )|/C 2 .
dλ
Now we need to deal with the special case ci = 0. Note that since f is Lipschitz, for any β ∈ R,
we have | E[f (Z−i + βYi2 ) − f (Z−i )]|/|β| ≤ | E[Yi2 ]| ≤ 1. By considering the limit as β tends to
d E[f (Zi,λ )]
≤ 1.
zero from above or below, we get that when ci = 0, the derivative still exists and
dλ
Since the limit 2|(bi − ai )|/C 2 only does not apply at one point where E[f (Zi,λ )] is still continuous
as a function of λ, this is not an issue. We still obtain that
| E[f (Zi,0 )] − E[f (Zi,1 )]| ≤ 2|(bi − ai )|/C 2 .
Recalling that Zi,1 = Zi+1,0 with Z1,0 = p(X ′ ) and Zd,1 = p(X) for X ′ ∼ N (0, 1) and X ∼ N (0, Σ),
we have by the Mean Value Theorem, that
X
| E[f (p(X ′ ))] − E[f (p(X)]| ≤
| E[f (Zi,0 )] − E[f (Zi,1 )]|
i
≤ 2|(bi − ai )|/C 2 ≤ O(kΣ − IkF /C 2 )
≤ kΣ − IkF /C ,
as required.
This completes the proof of the theorem.
5.3
Working in a Low-Dimensional Space of Degree-Two Polynomials
We now show that via similar techniques as before, we can patch our estimates together to find
a matrix which agrees with the ground truth on all degree-two polynomials in a fixed subspace of
low dimension. Formally, we show:
Theorem 10. Fix ε, τ > 0. Let Σ be so that kΣ−IkF ≤ O(ε log 1/ε). Let G0 be an (ε, δ)-good set of
samples from N (0, Σ), and let S = {X1 , . . . , Xn } be so that ∆(S, G0 ) ≤ ε. Let W1 be a subspace of
degree-2 polynomials, and let W2 be an orthogonal subspace of degree-2 polynomials, so that we have
b so that EX∼N (0,Σ) [p(X)] − E
aΣ
b [p(X)] ≤ ξ for all p ∈ W2 . Then there is an algorithm
X∼N (0,Σ)
b runs in time poly(d, |S|, 2O(dim(W1 )) , log 1/τ ),
LearnMeanPolyLowD which given ε, S, W1 , W2 , Σ
and returns a Σ′ so that
E
N (0,Σ)
[p(X)] −
E [p(X)] ≤ 4 (kΣ − IkF /C + O(log(C)ε) + ξ ,
N (0,Σ′ )
for all p ∈ span(W1 ∪ W2 ) ∩ P2 , with probability 1 − τ .
In particular, this implies:
Corollary 4. Fix ε, τ > 0. Let Σ be so that kΣ − IkF ≤ O(ε log 1/ε). Let G0 be an (ε, δ)good set of samples from N (0, Σ). Let S = {X1 , . . . , Xn } be so that ∆(S, G0 ) ≤ ε. Let V be a
subspace of Rd . Then there is an algorithm LearnCovLowDim which given S, ε, ξ, τ, V runs in
2
time poly(|S|, 2O(dim(V ) ) , log 1/τ ) and returns a Σ′ so that
kΠV Σ − Σ′ ΠV k2 ≤ 4 (kΣ − IkF /C + O(log(C)ε) ,
with probability 1 − τ .
21
Proof. Observe that the dimension of the space of degree-2 polynomials W in V is O(dim(V )2 ).
Run the algorithm in Theorem 10 with the same parameters as before, with W1 = W and W2 = ∅
(so that we may take ξ = 0), and then the guarantee of that algorithm, along with Lemma 10,
gives our desired guarantee.
We now describe the algorithm for Theorem 10. Essentially, we do the same thing as we did
for low-dimensional learning in the unknown mean case: we take a constant net over V ∩ P2 , l earn
the mean over every polynomial in the net, and then find a Σ′ which is close in each direction to
the learned mean. Since we will not attempt to optimize the constant factor here, will will use a
naive LP-based approach to find a point which is close to optimal. The formal pseudocode is given
in Algorithm 7.
Algorithm 7 Filter if there are many large eigenvalues of the covariance
1: function LowDimCovLearning(S, ε, ξ, τ, W1 , W2 )
2:
Generate a 1/2-cover C for W1 ∩ P2 .
3:
Let τ ′ = 2−|C| τ
4:
for p ∈ C do
5:
Compute mp = LearnMeanChiSquared(S, p, ε, τ ′ ).
6:
Generate a linear constraint cp (Σ′ ): EN (0,Σ′ ) [p(X)] − mp ≤ kΣ − IkF /C + O(log(C))ε.
7:
end for
8:
Generate the convex constraint that EN (0,Σ′ ) [p(X)] − EN (0,Σ′ ) [p(X)] ≤ ξ for all p ∈ W2 .
9:
Using a convex program, return any matrix Σ′ which obeys cp (Σ′ ) for all p ∈ C.
10: end function
Observe that every constraint for each polynomial in W1 is indeed linear in Σ′ , by Lemma 10.
Moreover, the constraint for W2 has an explicit separation oracle, since it induces a norm, and
for any p ∈ W2 , we may explicitly compute EN (0,Σ′ ) [p(X)] − EN (0,Σ′ ) [p(X)]. Thus, we may use
separating hyperplane techniques to solve this convex program in the claimed running time.
Proof of Theorem 10. Let us condition on the event that LearnMeanChSquared succeeds for
each p ∈ C. By a union bound, this occurs with probability at least 1 − τ . Thus, in each p ∈ C, we
have that |mp − EX∼N (0,Σ) [p(X)]| ≤ β, where β = kΣ − IkF /C + O(log(C))ε. Let Σ′ be the matrix
we find. By the triangle inequality, we then have that for every p ∈ C, that | EN (0,Σ′ ) [p(X)] −
EN (0,Σ) [p(X)]| ≤ 2β. Hence, by the usual net arguments, we know that for every p ∈ V ∩ P2 ,
|
E
[p(X)] −
N (0,Σ′ )
E [p(X)]| ≤ 4β .
N (0,Σ)
Moreover, by triangle inequality, for every p ∈ W2 , we have EN (0,Σ′ ) [p(X)] − EN (0,Σ′ ) [p(X)] ≤ 2ξ.
The result then follows from the Pythagorean theorem.
6
Robustly Learning the Covariance in High-Dimensions
In this section, we show how to robustly estimate the covariance of a mean-zero Gaussian in highdimensions up to error O(ε). We use our low-dimensional learning algorithm from the previous
section as a crucial subroutine in what follows.
Our main algorithmic contribution is as follows:
Theorem 11. Fix ε, δ > 0, and let S0 = (G0 , E0 ) be an ε-corrupted set of samples of size n from
N (0, Σ), where kΣ − IkF ≤ ξ where ξ = O(ε log 1/ε), and where n = poly(d, 1/ε, log 1/δ). Suppose
22
that G0 is (ε, δ)-good with respect to N (0, Σ). Let S ⊆ S0 be a set so that ∆(S, G0 ) ≤ ε. Then, there
exists an algorithm ImproveCov that given S, ξ, ε, fails with probability at most poly(ε, 1/d, δ),
and otherwise outputs one of two possible outcomes:
b so that kΣ
b − ΣkF ≤ kΣ − IkF /2.
(i) A matrix Σ,
(ii) A set S ′ ⊂ S so that ∆(S ′ , G0 ) < ∆(S, G0 ).
Moreover, ImproveCov runs in time poly(d, (1/ε)O(log
4
1/ε) , log 1/δ).
By first applying the algorithm in [DKK+ 16] to produce an initial estimate for Σ, and then
iterating the above algorithm polynomially many times, this immediately yields:
Corollary 5. Fix ε, δ > 0, and let G0 be a set of i.i.d. samples from N (0, Σ), where n =
poly(d, 1/ε, log 1/δ). Let S be so that ∆(S, G0 ) ≤ ε. There is an universal constant C and an
b so that with probability 1 − δ, we have kΣ
b −1/2 ΣΣ
b −1/2 − IkF ≤ Cε. In
algorithm which outputs a Σ
particular, this implies that
b ≤ 2Cε .
dTV N (0, Σ), N (0, Σ)
6.1
Technical Overview
Our strategy for obtaining a high-dimensional estimate for the covariance based on solving lowdimensional subproblems will be substantially more challenging than it was for the unknown mean
case. The natural approach is to take the poly log(1/ε)-dimensional subspace of degree-2 polynomials of largest empirical variance and construct a filter. However, this fails because, unlike in the
mean case, we do not know the variance of these degree-2 polynomials to small error. For the unknown mean case, because we assumed that we knew the covariance was the identity (or spectrally
close to the identity), this was not an issue. Now, the variance of our polynomials depends on the
(unknown) covariance of the true Gaussian, which may be more than O(ε)-far from our current
estimate. Indeed, it is not difficult to come up with counterexamples where there are many large
eigenvalues of the empirical covariance matrix, but no filter can make progress.
We overcome this hurdle in several steps. First, in Section 6.3, we show how to find a filter
if there are many medium-sized eigenvalues of the empirical covariance matrix. This will proceed
roughly in the same way that the filter for the unknown mean does. If no filter is created, then
we know there are at most logarithmically large eigenvalues of the empirical covariance. In the
subspace V ⊆ Rd spanned by their eigenvectors, we can then learn the covariance to high accuracy
using our low-dimensional estimator.
Then, in Section 6.3, we show that if we restrict to the orthogonal subspace, i.e., the subspace where the empirical covariance matrix does not have large eigenvalues, we can indeed either
produce a filter or improve our estimate of the covariance restricted to this subspace using our
low-dimensional estimator. While the blueprint is similar to the filter for the unknown mean, the
techniques are much more involved and subtle.
Supposing we have not yet created a filter, we have now estimated the covariance on a polylogarithmic dimensional subspace V , and on V ⊥ . This does not in general imply that we have
learned the covariance in Frobenius norm. In block form, if we write
ΣV AT
,
Σ=
A ΣV ⊥
23
where here Rd is written as V ⊕ V ⊥ , this implies we have learned ΣV and ΣV ⊥ to high accuracy.
Thus, it remains to estimate the cross term A.
In Section 6.4, we show, given a polylogarithmically sized subspace V , and a good estimate of
the covariance matrix on V and V ⊥ , how to fill in the entire covariance matrix. Roughly, we do
this by randomly fixing directions in V , and performing rejection sampling based on the correlation
in the direction in V , and showing that the problem reduces to one of robustly learning the mean
of a Gaussian, which (conveniently) we have already solved. These steps together yield our overall
algorithm ImproveCov. Finally, in Section 6.6 we explain why there is a natural barrier that
makes reducing the running time from quasi-polynomial to polynomial (in 1/ε) difficult.
6.2
Additional Preliminaries
Here we give some additional preliminaries we will require in this Section.
6.2.1
The Agnostic Tournament
We also require the following classical result, which allows us to do agnostic hypothesis selection
with corrupted samples (see e.g., [DL01, DDS12, DK14, DDS15]).
Theorem 12. Fix ε, δ > 0. Let D1 , . . . , Dk , D be a set of distributions where mini dTV (Di , D) = γ.
1/δ
. There is an algorithm Tournament which given oracles for evaluating
Set n = Ω log k+log
ε2
the pdfs of D1 , . . . , Dk along with n independent samples X1 , . . . Xn from D, outputs a Di so that
dTV (Di , D) ≤ 3γ + ε with probability 1 − δ. Moreover, the running time and number of oracle calls
needed is at most O(n2 /ε2 ).
Remark 1. As a simple corollary of the agnostic tournament, observe that this allows us to do
agnostic learning without knowing the precise error rate ε. Throughout the paper, we assume the
algorithm knows ε. However, if the algorithm is not given this information, and instead given an
η and asked to return something with error at most O(ε + η), we may simply grid over {η, (1 +
γ)η, (1 + γ)2 η, . . . , 1} (here γ is some arbitrary constant that governs a tradeoff between runtime
and accuracy), run our algorithm with ε set to each element in this set, and perform hypothesis
selection via Tournament. Then it is not hard to see that we are guaranteed to output something
which has error at most O(ε + (1 + γ)η).
6.2.2
The Fourth Moment Tensor of a Gaussian
As in [DKK+ 16], it will be crucial for us to understand the behavior of the fourth moment tensor
of a Gaussian. Let ⊗ denote the Kronecker product on matrices. We will make crucial use of the
following definition:
2
Definition 6. For any matrix M ∈ Rd×d , let M ♭ ∈ Rd denote its canonical flattening into a vector
2
2
in Rd , and for any vector v ∈ Rd , let v ♯ denote the unique matrix M ∈ Rd×d so that M ♭ = v.
We will also require the following definition:
2
Definition 7. Let Ssym = {M ♭ ∈ Rd : M is symmetric}.
The following result was proven in [DKK+ 16]:
Theorem 13 (Theorem 4.15 in [DKK+ 16]). Let X ∼ N (0, Σ). Let M be the d2 × d2 matrix given
by M = E[(X ⊗ X)(X ⊗ X)T ]. Then, as an operator on Ssym , we have
T
.
M = 2Σ⊗2 + Σ♭ Σ♭
24
6.2.3
Polynomials in Gaussian Space
Here we review some basic facts about polynomials under Gaussian measure, which will be crucial
for our algorithm for learning Gaussians with unknown covariance. We equip the set of polynomials
over Rd with the Gaussian inner product, defined by hf, gi = EX∼N (0,I) [f (X)g(X)], and we let
kf k22 = hf, f i.
For any symmetric M with kM kF = 1, define the degree-2 polynomial p(x) = √12 (xT M x −
tr(M )). We call p the polynomial associated to M . Observe that p is even (i.e., has no degree-1
terms). We will use the following properties of such polynomials:
Lemma 10. Let M be symmetric, so that kM kF = 1. Let p be its associated polynomial. Then,
we have:
(i) EX∼N (0,I) [p(X)] = 0.
(ii) More generally, for any positive definite matrix Σ, we have EX∼N (0,Σ) [p(X)] = hM, Σ − Ii.
(iii) VarX∼N (0,I) [p(X)] = EX∼N (0,I) [p2 (X)] = hp, pi = 1.
(iv) More generally, for any positive definite matrix Σ, we have
2
T
1
E
[p2 (X)] = M ♭ Σ⊗2 M ♭ +
hΣ − I, M ♭ i .
2
X∼N (0,Σ)
Proof. The first three properties are a straightforward calculation. We show the last one here. By
definition, we have
E
X∼N (0,Σ)
1
E
[(X T M X − tr(M ))2 ]
2 X∼N (0,Σ)
1
E
[(X T M X)2 − 2(X T M X)tr(M ) + tr(M )2 ]
=
2 X∼N (0,Σ)
1
T
♭
2
♭T
=
E
[M (X ⊗ X)(X ⊗ X) M ] − 2hM, Σitr(M ) + tr(M )
2 X∼N (0,Σ)
T
T
(a) 1
M ♭ 2Σ⊗2 + Σ♭ Σ♭ M ♭ − 2hM, Σitr(M ) + tr(M )2
=
2
T
1 ♭
= M ♭ Σ⊗2 M ♭ +
hM , Σi2 − 2hM, Σitr(M ) + tr(M )2
2
2
1
♭
♭ T ⊗2
=M Σ M +
hΣ − I, M ♭ i ,
2
[p2 (X)] =
as claimed, where (a) follows from Theorem 13.
Observe that Lemma 10(iv) implies that if we take the top eigenvector of the d2 × d2 matrix
Σ⊗2 +
on the linear subspace
1 ♭ ♭ ⊤
M
M
2
V = {M ♭ : M is a symmetric d × d matrix} ,
then the associated polynomial maximizes EX∼N (0,Σ) [p2 (X)], and so we can find these polynomials
efficiently. More generally, if we take any linear subspace of degree two polynomials with associated
25
matrix subspace V ′ , so that V ′ ⊆ V , then the top eigenvector of the same matrix restricted to V ′
allows us to find the polynomial in this subspace which maximizes EX∼N (0,Σ) [p2 (X)] efficiently.
We have the following tail bound for degree-2 polynomials in Gaussian space: We will use ΠV (x)
and ΠV (S) to denote projection to a subspace V , of a point x and a set of points S, respectively.
We will also need the following hypercontractivity theorem for low-degree polynomials in Gaussian
space:
Theorem 14 (Hypercontractivity in Gaussian space, see e.g. [O’D14]). Let p : Rd → R be a degree
√
m polynomial, and let q ≥ 2 be even. Then EX∼N (0,I) [p(X)q ]1/q ≤ ( q − 1)m kpk2 .
We need the following definition:
Definition 8. Let Pk denote the set of even degree-k polynomials over d variables satisfying
VarX∼N (0,1) [p(X)] = 1. Moreover, for any subspace W ⊆ Rd , let Pk (W ) denote the set of even
polynomials over d variables which only depend on the coordinates in W .
b
Then by the arguments above, we have that for any two matrices Σ, Σ,
!
b F = sup
kΣ − Σk
E
[p(X)] −
E
[p(X)] .
p∈P2
X∼N (0,Σ)
b
X∼N (0,Σ)
In particular, by Lemma 2, this implies that when kΣ − Ik2 is small, then learning a Gaussian with
unknown covariance in total variation distance is equivalent to learning the expectation of every
even degree-2 polynomial.
Theorem 14 implies the following concentration for degree-4 (more generally, low-degree) polynomials of Gaussians:
Corollary 6. Let p be a degree-4 polynomial. Then there is some A, C ≥ 0 so that for all t ≥ C,
we have
Pr [|p(X) − E [p(X)]| ≥ tkpk2 ] ≤ exp −At1/2 .
N (0,I)
N (0,I)
Proof. Hypercontractivity in particular implies the following moment bound: for all q ≥ 2 even,
we have
E [(p(X) − E [p(X)])q ] ≤ (q − 1)qm/2 kp(X) − E [p(X)]kq2 .
N (0,I)
N (0,I)
N (0,I)
By a typical moment argument, and optimizing the choice of q, this gives the desired bound.
Hermite polynomials Hermite polynomials are what arise by Gram-Schmidt orthogonalization
applied with respect to this inner product. For a vector of non-negative integers a = (a1 , . . . , ad ),
we let Ha (x) : Rd → R be the Hermite
Pd polynomial associated with multi-index a. It is well-known
that the degree of Ha is |a| =
i=1 ai , and moreover, hHa , Hb i = δa,b . In particular, for any
r ≥ 1, the Hermite polynomials of degree at most r form an orthonormal basis with respect to the
Gaussian inner product for all polynomials with degree at most r.
Therefore, given any polynomial p : Rd → R with degree r, we may write it uniquely as
X
p(x) =
ca (p)Ha (x), where ca (p) = hp, Ha i .
|a|≤r
We define the kth harmonic component of p to be
X
p[k](x) =
ca (p)Ha (x) ,
|a|=k
and we say p is harmonic of degree k if it equals its kth part.
26
6.3
Working with Many Large Eigenvalues of the Second and Fourth Moment
As in the unknown mean case, we will need a filter to detect if there are many directions of the
empirical covariance which have too large an eigenvalue. Formally, we need:
Theorem 15. Fix ε, δ > 0. Assume kΣ − IkF ≤ ξ, where ξ = O(ε log 1/ε). Suppose that G0
b = ES [XX T ].
is (ε, δ)-good with respect to N (0, Σ). Let S be a set so that ∆(S, G0 ) ≤ ε. Let Σ
Then there is an algorithm FilterCovManyDeg2Eig and a universal constant C such that the
following guarantee holds:
b − I has more than O(log 1/ε) eigenvalues larger than Cξ, then the algorithm outputs a
1. If Σ
′
S so that ∆(S ′ , G0 ) < ∆(S, G0 ).
2. Otherwise, the algorithm outputs “OK”, and outputs an orthonormal basis v1 , . . . , vk for the
b − I with eigenvalue larger than Cξ.
subspace V of vectors spanned by all eigenvectors of Σ
The filter developed here is almost identical to the one developed for unknown mean. Thus, for
conciseness we describe and prove the theorem in Appendix D.1.
We will also need a subroutine to enforce the condition that not only does the fourth moment
tensor have spectral norm which is at most O(ε log2 1/ε) (restricted to a certain subspace of polynomials), but there can only be at most O(poly log 1/ε) directions in which the eigenvalue is large.
However, the techniques here are a bit more complicated, for a number of reasons. Intuitively,
the main complication comes from the fact that we do not know what the fourth moment tensor
looks like, whereas in the unknown mean case, we knew that the covariance was the identity by
assumption. Our main result in this subsection is the following subroutine:
Theorem 16 (Filtering when there are many large eigenvalues). Fix ε, δ > 0. Assume kΣ − IkF ≤
ξ, where ξ = O(ε log 1/ε). Let C be the universal constant in FilterCovManyDeg2Eig. Let
W ⊆ Rd be a subspace, so that for all v ∈ W with kvk2 = 1, we have v T ES [XX T ]v ≤ 1 + Cξ.
Suppose that G0 is (ε, δ)-good with respect to N (0, Σ). Let S be a set so that ∆(S, G0 ) ≤ ε. Let
k = O(log4 1/ε). Then there is an algorithm FilterCovManyDeg4Eig and universal constants
C1 , C2 such that the following guarantee holds:
1. If there exist p1 , . . . , pk ∈ P2 (W ) so that hpj , pℓ i = δjℓ for all j, ℓ, and so that ES [p2j (Y )] − 1 ≥
C1 ε for all j, then the algorithm outputs an S ′ so that ∆(S ′ , G0 ) < ∆(S, G0 ).
2. Otherwise, the algorithm outputs “OK”, and outputs an orthonormal basis p1 , . . . , pk′ for a
subspace V of degree-2 polynomials in P2 (W ) with k′ ≤ k so that for all p ∈ V ⊥ ∩ P2 , we
have ES [p2 (X)] − 1 ≤ C2 ε.
Moreover, FilterCovManyEig runs in time poly(d, 1/ε, log 1/δ).
Roughly, we will show that if there are many polynomials with large empirical variance, this
implies that there is a degree-four polynomial whose value is much larger than it could be if w
were the set of uniform weights over the uncorrupted points. Moreover, we can explicitly construct
this polynomial, and it has a certain low-rank structure which allows us to use the concentration
bounds we have previously derived.
6.4
Stitching Together Two Subspaces
This section is dedicated to giving an algorithm which allows us to fully reconstruct the covariance
matrix given that we know it up to small error on a low-dimensional subspace V and on W = V ⊥ .
27
Algorithm 8 Filter if there are many large eigenvalues of the fourth moment tensor
1: function FilterCovManyEig(S, ε, ξ, δ, W )
b = ES [XX T ]
2:
Let Σ
3:
Let C1 , C2 , C3 be some universal constants sufficiently large
4:
Let A be the constant in Corollary 6
5:
Let B be the constant in Claim 20
6:
Let m = 0
7:
Let k = O(log4 1/ε)
8:
while there exists p ∈ P2 (W ) so that p ∈ V ⊥ and ES [p2 (X)] − 1 > C1 ξ do
9:
Let Vm+1 = span(Vm ∪ p)
10:
Let m ← m + 1
11:
end while
12:
Let p1 , . . . , pm be an orthonormal basis for Vm
13:
if m ≥ k then
[4]
14:
Let qi = (p2i ) be the 4th harmonic component of p2i
15:
Let ri = p2i − qi be the degree-2 component of p2i
P
16:
Let Q(x) = ki=1 qi
17:
Find a T so that either:
√
• T > C3 d2 k log(|S|) and p(X) > T for at least one x ∈ S ′ , OR
√
• T > 4A2 C2 B k log2 (1/ε) and
√
Pr [Q(X) > T ] > exp(−A(T /4B k)1/2 ) + ε2 /(d log(|S|/δ))2 .
X∈u S
18:
19:
20:
21:
22:
return the set S ′ = {X ∈ S : Q(X) ≤ T }
else
return “OK”, and output p1 , . . . , pm
end if
end function
Theorem 17. Let 1 > ξ > η > ε > 0, and let τ > 0. Let Σ so that kΣ − IkF ≤ ξ. Suppose that
Rd is written as V ⊕ W for orthogonal subspaces V and W with dim(V ) = O(log(1/ε)). Suppose
furthermore that
ΣV AT
,
Σ=
A ΣW
with kΣV − IV kF , kΣW − IW kF = O(η). Let S0 = (G0 , E0 ) be an ε-corrupted set of samples from
N (0, Σ), and let S ⊆ S0 with ∆(S, G) ≤ O(ε) of size poly(d, 1/η, log 1/δ).
Then, there exists a universal constant C5 and an algorithm Stitching that given V, W, ξ, η, ε, τ
and S runs in polynomial time and with probability at least 1 − τ returns a matrix Σ0 with kΣ0 −
ΣkF = C5 η + O(ξ 2 ).
In the latter, we will show the algorithm works when τ = 2/3. As usual the probability of
success can be boosted by repeating it independently.5 The basic idea of the proof is as follows.
Since we already know good approximations to ΣV and ΣW , it suffices to find an approximation to
5
Observe the only randomness at this point is in the random choices made by the algorithm. Thus, one can just
(j)
(ℓ)
(1)
run this algorithm O(log 1/δ) times to obtain Σ0 , . . . , Σ0 and find any Σ0 which is O(η + ξ 2 ) close to at least a
2/3 + o(1) fraction of the other outputs.
28
A. In order to do this, we note that if we take a sample x from G conditioned on its projection to V
being some vector v, we find that the distribution over W is a Gaussian with mean approximately
Av. Running our algorithm for approximating the mean of a noisy-Gaussian, we can then compute
the mapping v → Av, which will allow us to compute A.
There are three main technical obstacles to this approach. The first is that we cannot condition
on xV taking a particular value, as we will likely see no samples from X with exactly that projection.
Instead, what we will do is given samples from X we will reject them with probabilities depending
on their projections to V in such a way to approximate the conditioning we require. The second
obstacle is that the errors in X may well be concentrated around some particular projection to
V . Therefore, some of these conditional distributions may have a much larger percentage of errors
than ε. To circumvent this, we will show that by carefully choosing how we do our conditioning
and by carefully picking the correct distribution over vectors v, that on average these errors are
only O(ε). Finally, we need to be able to reconstruct A from a collection of noisy approximations
to Av. We show that this can be done by computing these approximations at a suitably large
random set of v’s, and finding the matrix A that minimizes the average ℓ2 error between Av and
its approximation.
Our algorithm is given in Algorithm 9:
Algorithm 9 Stitching the two subspaces together
1: function Stitching(V, W, δ, ε, τ, S)
2:
Given a vector x, let xV and xW be the projections onto V and W , respectively.
3:
Let C be a sufficiently large constant (where C may depend on the constants in the big-O
terms in the guarantee that dim(V ) = O(log(1/ε))).
4:
Generate a set V = {v1 , . . . , vm } of (n/ε)C independent random samples from N (0, 2IV ).
5:
for v ∈ V do
6:
For each sample x ∈ S, add xW to a new set T independently with probability
exp(−kxV − vk2 /2) .
7:
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
Treat T as a collection of independent samples from a noisy Gaussian with covariance
matrix IW + O(η).
Set av equal to 0 if T did not contain enough samples for our algorithm or if kµ̃k2 >
C log(1/ε).
for ε ∈ {1, 1/2, 1/4, 1/8, . . . , η} do
Let µ
e be the output of RecoverMeanNoisy(T, ε, (ε/n)2C , o(1), O(η)).
end for
Run Tournament with the output hypotheses.
Set av = µ̃, where µ
e is the winning hypothesis.
end for
Use linear programming to find the dim(W ) × dim(V )-matrix B that minimizes the convex
function Ev∈u S [|av − Bv|].
return
IV 2B T
.
Σ0 =
2B IW
end function
Before we prove Theorem 17, we will need the following definition.
29
Definition
9. A function f : Rd → R is a positive measure if f ≥ 0 and
R
|f |1 = f , and we say f1 ≤ f2 if f1 (x) ≤ f2 (x) pointwise.
R
f ≤ 1. We write
Proof of Theorem 17. Throughout this proof, let G = N (0, Σ). It is clear that this algorithm has
polynomial runtime and sample complexity. We have yet to show correctness. The first thing
that we need to understand is the procedure of rejection sampling, where we reject x except with
probability exp(−kxV − vk2 /2). Therefore, given a distribution D, we let the positive measure
Dv be what is obtained by sampling from D and accepting a sample x only with probability
exp(−kxV − vk2 /2). We need to understand the distributions Gv and Xv .
Note that the pdf of G is given by
p
(2π)−n/2 det(Σ−1 ) exp(−xΣ−1 x/2)dx.
Therefore, the density of Gv is
p
(2π)−n/2 det(Σ−1 ) exp(−xΣ−1 x/2 − kxV − vk22 /2)dx
p
=(2π)−n/2 det(Σ−1 ) exp(−x(Σ−1 + IV )x/2 + x · v − kvk22 /2)dx.
Note that letting µv := (Σ−1 + IV )−1 v = v/2 + O(δkvk2 ) that this equals
p
=(2π)−n/2 det(Σ−1 ) exp(−(x − µv )(Σ−1 + IV )(x − µv )/2 − kvk22 /2 + µv · v/2)dx
p
=(2π)−n/2 det(Σ−1 ) exp(−(x − µv )(Σ−1 + IV )(x − µv )/2 − kvk22 (1 + O(δ))/4)dx.
Note that this is a Gaussian with mean µv weighted by
s
det(Σ−1 )
exp(−kvk22 (1 + O(δ))/4) = Θ(2− dim(V )/2 exp(−kvk22 /4)) ,
det(Σ−1 + IV )
so long as kvk2 ≪ δ−1/2 . Therefore, if this condition holds, a random sample from G is accepted
by this procedure with probability Θ(2− dim(V ) exp(−kvk22 /4).
We also need to understand the fraction of samples from Xv that are erroneous. In a slight
abuse of notation, let E also denote the distribution which is uniform over the points in E, and let
L be the distribution which is uniform over the points in L. Therefore, we define
εv := 2n/2 exp(kvk22 /4)(|Ev |1 + |Lv |1 log(1/ε)) ,
that is approximately the fraction of samples from Xv that are errors (where subtractive errors are
weighted more heavily).
p
Note that so long as kvk2 < C log(1/ε) that so long as |Lv |1 = o(1) that T in Step 5 will
have sufficiently many samples with high probability. Furthermore, if this is the case, we have
(assuming our mean estimation algorithm succeeds) that kµ̃ − πW (µv )k2 ≤ O(εv + η). Therefore,
unless kµv k2 = Ω(C log(1/ε)) (which can only happen in kvk2 = Ω(C log(1/ε))), we have with high
probability that kav − πW (µv )k2 = O(εv + η). In fact, we assume that this holds for all v ∈ S with
kvk2 ≪ (C log(1/ε)).
In order to show that this is generally a good approximation, we need to know that εv is not
too large on average. In particular, we show:
Lemma 11. If v ∼ N (0, 2IV ), then Ev [εv ] = O(ε).
30
Proof. Letting F = E + log(1/ε)L, we have that
Z
− dim(V )/2
E[εv ] = (2π)
exp(kvk22 /4) exp(−kvk22 /4)dFv dv
v
Z
Z
− dim(V )/2
2
= (2π)
exp(−kx − vk /2)dF (x)dv = dF = O(ε).
Note that πW (µv ) = M v, where M = πW (Σ−1 + IV )−1 πVT . Let B be any dim(W ) × dim(V )
matrix. Note that
E
[kBv − M vk22 ] = 2kB − M k2F .
v∼N (0,2IV )
Therefore, Ev∼N (0,2IV ) [kBv − M vk2 ] = O(kB − M kF ). Since kBv − M vk22 is a degree-2 polynomial
in v, by Corollary 4 in [Kan12],
#
"
r
E
[kBv − M vk42 ]/2 > 1/81.
Pr
[kBv − M vk2 > kB − M kF /2] >
Pr
kBv − M vk22 >
v∼N (0,2IV )
v∼N (0,2IV )
v∼N (0,2IV )
Combining these we find that
E
[kBv − M vk2 ] = Θ(kB − M kF ) .
v∼N (0,2IV )
Next we show that our choice of S derandomizes this result.
Lemma 12. With high probability over the choice of S, we have that
E [kBv − M vk2 ] = Θ(kB − M kF ).
v∈u S
for all dim(W ) × dim(V ) matrices B.
Proof. This is equivalent to showing that for all matrices U = B − M it holds
E [kU vk] = Θ(kU kF ).
v∈u S
By the standard scaling laws, it suffices to show this only for U with kU kF = 1.
We also note that it suffices to show this only for U in an ε-net for all such matrices. This is
because if kU − U ′ kF < ε and if U ′ satisfies the desired condition, then
E [kU vk2 ] = E [kU ′ vk2 ] + O( E [k(U − U ′ )vk2 ]) = Θ(1) + O(ε) E [kvk2 ] ,
v∈u S
v∈u S
v∈u S
v∈u S
and with high probability Ev∈u S [kvk2 ] = O(log(1/ε)).
Note that such nets exist with size exp(poly(n/ε)). Therefore, it suffices to show that this
condition holds for each such U with probability exp(−(n/ε)Ω(C) ).
As noted above, it suffices that
E [kU vk2 ] = O( E [kU vk2 ]) = O(1)
v∈u S
v∈u S
and that
Pr [kU vk2 > 1/4] ≥ Pr [kU vk2 ≥ 1/4] − 1/100.
v∈u S
v∈u S
The first follows because if S = {v1 , . . . , vm } then Ev∈u S [kU vk2 ] is a degree-2 polynomial in the
√
vi with mean O(1) and variance O(1/ m), so by standard concentration results, is O(1) with 1 −
exp(−Ω(|S|)) probability. The latter follows from standard concentration bounds. This completes
the proof.
31
We also note that the random choice of S has another nice property:
Lemma 13. With S and av as above, with high probability we have that
E [kav − M vk2 ] = O(η) .
v∈u S
Proof. Let S = {v1 , . . . , vm }. Then kf (vi ) − M vi k2 are independent random variables with mean
E[O(εv + η)] = O(η) and variance at most
E[kav k22 + kM vk22 ] ≤ O(log2 (1/ε)) + E[kvk22 ] = O(log2 (1/ε)).
The result follows by Chernoff bounds.
Now by Lemmas 12 and 13 with high probability over the choice of S in Step 4, and the av in
Step 5 we have that for all dim(W ) × dim(V )-matrices B that
E [kBv − M vk2 ] = Θ(kB − M kF ),
v∈u S
and have that
E [kav − M vk2 ] = O(η).
v∈u S
Combining these statements, we have that for all dim(W ) × dim(V )-matrices B that
E [kav − Bk2 ] = Θ(kB − M kF ) + O(η).
v∈u S
Note that by taking B = M , this quantity is O(η). Therefore, the B found in Step 15 satisfies
O(ε) = E [kav − Bk2 ] = Θ(kB − M kF ) + O(η) ,
v∈u S
and therefore, kB − M kF = O(η).
The rest of the proof is a simple computation of the matrices involved. In particular, recall that
IV AT
ΣV AT
+ O(η) ,
=
Σ=
A IW
A ΣW
where the O(η) denotes a matrix with Frobenius norm O(η) and where kAkF = O(δ). It is easy to
see that
I
−AT
+ O(η + δ2 ).
Σ−1 = V
−A IW
Therefore,
Σ
−1
Hence,
(Σ
−1
2IV −AT
+ O(η + δ2 ).
+ IV =
−A IW
−1
+ IV )
IV /2 AT /2
=
+ O(η + δ2 ).
A/2
IW
Therefore, M = A/2 + O(η + δ2 ). Therefore, A = 2M + O(η + δ2 ) = 2B + O(η + δ2 ). And finally,
we conclude that
IV 2B T
+ O(η + δ2 ) = Σ0 + O(η + δ2 ).
Σ=
2B IW
32
6.5
The Full High-Dimensional Algorithm
We now show how to prove Theorem 11, given the pieces we have. We first show that given enough
samples from N (0, Σ), the empirical data set without corruptions satisfies the regularity conditions
in Section 2.2.2 with high probability. For clarity of exposition, the proof of this lemma is deferred
to Appendix D.3.
Lemma 14. Fix η, δ > 0. Let X1 , . . . , Xn be independent samples from N (µ, I), where n =
poly(d, 1/η, log 1/δ). Then, S = {X1 , . . . , Xn } is (η, δ)-good with respect to N (µ, I) with probability
at least 1 − δ.
Finally, we require the following guarantee, which states that if there is a degree-2 polynomial
whose expectation under S and the truth differs by a lot (equivalently, if the empirical covariance
differs from the true covariance in Frobenius norm substantially), then it must also have very large
variance under S.
Lemma 15. Fix ε, δ > 0. Assume kΣ − IkF ≤ ξ, where ξ = O(ε log 1/ε). Suppose that G0 is
(ε, δ)-good with respect to N (0, Σ), and let S ⊆ S0 be a set so that ∆(S, G0 ) ≤ ε. There is some
√
absolute constant C5 so that if p ∈ P2 is a polynomial so that ES [p(X)] − EN (0,Σ) [p(X)] > C5 ξε,
then ES [p2 (X)] − 1 > C1 ξ.
We defer the proof of this lemma to the Appendix.
We are now ready to present the full algorithm as Algorithm 10.
Algorithm 10 Filter if there are many large eigenvalues of the covariance
1: function ImproveCov(S, ξ, ε, δ)
2:
Let C be the universal constant in FilterCovManyDeg2Eig
3:
Let τ = poly(ε, 1/d, δ).
4:
Run FilterCovManyDeg2Eig(S, ε, ξ)
5:
if FilterCovManyDeg2Eig outputs S ′ then
6:
return S ′
7:
else
8:
Let V be the subspace returned by FilterCovManyDeg2Eig
9:
Let W = V ⊥ .
10:
Run FilterCovManyDeg4Eig(S, ε, ξ, δ, W )
11:
if FilterCovManyDeg4Eig outputs S ′ then
12:
return S ′
13:
else
14:
Let U1 be the subspace of degree 2 polynomials over W it returns
15:
Let U2 be the perpendicular subspace of degree 2 polynomials over W
b = ES [XX T ]
16:
Let Σ
17:
Let ΣV = LearnCovLowDim(S, ε, ξ, τ, V )
b
18:
Let ΣW = LearnMeanPolyLowD(S, ε, ξ, τ, U1 , U2 , Σ)
′
19:
Take poly(n, 1/ε) fresh ε-corrupted samples S
20:
return Stitching(V, W, ΣV , ΣW , ξ, ε, S ′ ).
21:
end if
22:
end if
23: end function
33
Proof of Theorem 11. Condition on the events that neither LearnCovLowDim nor Stitching
fail. This happens with probability at least poly(ε, 1/d, δ). Observe that if we pass the “if”
statement in Line 5, then by the guarantee of FilterCovManyDeg2Eig this is indeed an S ′
satisfying the desired properties. Otherwise, by the guarantees of FilterCovManyDeg2Eig, we
have that W satisfies the conditions needed by FilterCovManyDeg4Eig. Hence, if we pass
the “if” statement in Line 11, then the guarantee of FilterCovManyDeg4Eig this is indeed a
S ′ satisfying the desired properties. Otherwise, by Lemma 15, we know
√ that for all polynomials
p ∈ P2 over W orthogonal to U1 , we have | EN (0,Σ) − ES [XX T ]| ≤ C5 ξε. Thus, ΣW satisfies the
conditions needed by Stitching.
By Corollary 4, we know that ΣV satisfies the conditions for Stitching, and so the correctness
of the algorithm follows from Theorem 17.
6.6
The Barrier at Quasi-Polynomial
Here we explain why improving the running time from quasi-polynomial to polynomial in 1/ε will
likely be rather difficult. Recall that our strategy is to project the problem onto lower dimensional
subproblems and stitch together the answer. We need the dimension of the subspace to be large
enough that we can find a polynomial Q that is itself the sum of squares of k orthogonal degree
two polynomials pi so that the value of Q on the corrupted points is considerably larger than the
value on the uncorrupted points. More precisely, if we let S = (G, E) denote our corrupted set of
samples then we want EE [Q(X)] to be larger than Q(X) for all but a poly(ε) fraction of X ∈ G.
We then remove all points X ∈ S with large Q(X) and by the properties of Q we are guaranteed
that we throw out mostly corrupted points. It
√ turns out that the most aggressive we could be
is removing points where Q(X) is more than k standard deviations away from its expectation
under the true Gaussian. But since Q is a degree-four polynomial and we want Q(X)
√ to be smaller
than our cutoff for all but a poly(ε) fraction of X ∈ G, we are forced to choose k = Ω(log 1/ε),
which means that we need to reduce to k = Ω(log2 1/ε) dimensional subproblems. Thus, if we
solve low-dimensional subproblems in time exponential in the dimension, we naturally arrive at
a quasi-polynomial running time. It seems that any approach for reducing the running time to
polynomial would require fundamentally new ideas.
7
The General Algorithm
We now have all the tools to robustly learn the mean and covariance of an arbitrary high-dimensional
Gaussian. We first show how to reduce the problem of robustly learning the covariance of N (µ, Σ) to
learning the covariance of N (0, Σ), by at most doubling error, a trick previously used in [DKK+ 16]
and [LRV16]. Given an
√ ε-corrupted set of samples X1 , . . . , X2n of size 2n from N (µ, Σ), we may
let Yi = (Xi − Xn+i )/ 2. Then we see that if Xi and Xn+i are uncorrupted, then Yi ∼ N (0, Σ).
Moreover, at most 2εn of the Yi can be corrupted, since there are at most 2εn corrupted Xi .
Therefore, by doubling the error rate, we may assume that µ = 0. We may then apply the algorithm
b so that with high probability, we have kΣ
b −1/2 ΣΣ
b −1/2 − IkF ≤ O(ε)
in Corollary 5 to obtain a Σ
4
O(log
1/ε)
with polynomially many samples, and in poly(d, (1/ε)
) time.
b −1/2 X ′ .
We may then take an additional set of ε-corrupted samples {Xi′ , . . . , Xn′ }, and let Yi′ = Σ
i
′
′
b we have that if X is uncorrupted, then Y ∼ N (0, Σ)
e where
Then, by our guarantee on Σ,
i
i
e − IkF ≤ O(ε). We then run RecoverMeanNoisy with the Y ′ to obtain a µ
kΣ
b so that
i
−1/2
−1/2
b
e
b
b
kb
µ−Σ
µk2 ≤ O(ε). This guarantees that dTV (N (b
µ, Σ), N (Σ
µ, Σ)) ≤ O(ε), which in turn
b N (µ, Σ)) ≤ O(ε), as claimed.
implies that dTV (N (b
µ, Σ),
34
Therefore, we have shown:
Theorem 18. Fix ε, δ > 0. Given an ε-corrupted set of samples S from N (µ, Σ), where n =
poly(d, 1/ε, log 1/δ), there is an algorithm RecoverGaussian which takes as input S, ε, δ, and
b so that
outputs a µ
b, Σ
b ≤ O(ε) .
dTV (N (µ, Σ), N (b
µ, Σ))
Moreover, the algorithm runs in time poly(d, (1/ε)O(log
4
1/ε) , log 1/δ).
References
[BDLS17] S. Balakrishnan, S. S. Du, J. Li, and A. Singh. Computationally efficient robust sparse
estimation in high dimensions. In Proceedings of the 30th Annual Conference on Learning Theory, COLT ’17, 2017.
[Ber06]
T. Bernholt. Robust estimators are hard to compute. Technical report, University of
Dortmund, Germany, 2006.
[BNJ03]
D. M Blei, A. Y. Ng, and M. I. Jordan. Latent dirichlet allocation. Journal of machine
Learning research, 3(Jan):993–1022, 2003.
[BW41]
L. M. Blumenthal and G. E. Wahlin. On the spherical surface of smallest radius enclosing a bounded subset of n-dimensional euclidean space. Bull. Amer. Math. Soc.,
47:771–777, 1941.
[CSV17]
M. Charikar, J. Steinhardt, and G. Valiant. Learning from untrusted data. In Proc.
49th Annual ACM Symposium on Theory of Computing (STOC), pages 47–60. ACM
Press, 2017.
[CW01]
A. Carbery and J. Wright. Distributional and Lq norm inequalities for polynomials over
convex bodies in Rn . Mathematical Research Letters, 8(3):233–248, 2001.
[DBS17]
S. S. Du, S. Balakrishnan, and A. Singh. Computationally efficient robust estimation
of sparse functionals. CoRR, abs/1702.07709, 2017.
[DDS12]
C. Daskalakis, I. Diakonikolas, and R.A. Servedio. Learning Poisson Binomial Distributions. In Proceedings of the 44th Symposium on Theory of Computing, pages 709–728,
2012.
[DDS15]
A. De, I. Diakonikolas, and R. Servedio. Learning from satisfying assignments. In
Proceedings of the 26th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA
2015, pages 478–497, 2015.
[DK14]
C. Daskalakis and G. Kamath. Faster and sample near-optimal algorithms for proper
learning mixtures of gaussians. In Proceedings of The 27th Conference on Learning
Theory, COLT 2014, pages 1183–1213, 2014.
[DKK+ 16] I. Diakonikolas, G. Kamath, D. M. Kane, J. Li, A. Moitra, and A. Stewart. Robust
estimators in high dimensions without the computational intractability. In Proceedings
of FOCS’16, 2016.
35
[DKK+ 17] I. Diakonikolas, G. Kamath, D. M. Kane, J. Li, A. Moitra, and A. Stewart. Being
robust (in high dimensions) can be practical. In Proceedings of the 34th International
Conference on Machine Learning, ICML ’17, pages 999–1008. JMLR, Inc., 2017.
[DKS16]
I. Diakonikolas, D. M. Kane, and A. Stewart. Robust learning of fixed-structure bayesian
networks. CoRR, abs/1606.07384, 2016.
[DKS17]
I. Diakonikolas, D. M. Kane, and A. Stewart. Statistical query lower bounds for robust
estimation of high-dimensional gaussians and gaussian mixtures. In Proceedings of
the 58th Annual IEEE Symposium on Foundations of Computer Science, FOCS ’17,
Washington, DC, USA, 2017. IEEE Computer Society.
[DL01]
L. Devroye and G. Lugosi. Combinatorial methods in density estimation. Springer Series
in Statistics, Springer, 2001.
[GCSR14] A. Gelman, J. B. Carlin, H. S. Stern, and D. B. Rubin. Bayesian data analysis, volume 2.
Chapman & Hall/CRC Boca Raton, FL, USA, 2014.
[GLS88]
M. Grötschel, L. Lovász, and A. Schrijver. Geometric Algorithms and Combinatorial
Optimization, volume 2. Springer, 1988.
[Goo15]
R. Goodell. Personal communication, 2015.
[Ham01]
F. Hampel. Robust statistics: A brief introduction and overview. In First International
Symposium on Robust Statistics and Fuzzy Techniques in Geodesy and GIS. A. Carosio,
H. Kutterer (editors), Swiss Federal Institute of Technology Zurich (ETH), Institute of
Geodesy and Photogrammetry, IGP-Bericht, number 295, pages 13–17, 2001.
[HM13]
M. Hardt and A. Moitra. Algorithms and hardness for robust subspace recovery. In
COLT 2013, pages 354–375, 2013.
[HR09]
P.J. Huber and E. M. Ronchetti. Robust statistics. Wiley New York, 2009.
[HRRS86] F. R. Hampel, E. M. Ronchetti, P. J. Rousseeuw, and W. A. Stahel. Robust statistics.
The approach based on influence functions. Wiley New York, 1986.
[Hub64]
P. J. Huber. Robust estimation of a location parameter. The Annals of Mathematical
Statistics, 35(1):73–101, 1964.
[Jun01]
H. W. E. Jung. Über die kleinste kugel, die eine räumliche figur einschliesst. J. Reine
Angew. Math, 123:241–257, 1901.
[Kan12]
D. M. Kane. A structure theorem for poorly anticoncentrated gaussian chaoses and
applications to the study of polynomial threshold functions. In FOCS, pages 91–100,
2012.
[Kla07]
B. Klartag. A central limit theorem for convex sets.
168(1):91–131, 2007.
[Li17]
J. Li. Robust sparse estimation tasks in high dimensions. CoRR, abs/1702.05860, 2017.
[LM00]
B. Laurent and P. Massart. Adaptive estimation of a quadratic functional by model
selection. Ann. Statist., 28(5):1302–1338, 2000.
36
Inventiones mathematicae,
[LRV16]
K. A. Lai, A. B. Rao, and S. Vempala. Agnostic estimation of mean and covariance. In
Proceedings of FOCS’16, 2016.
[NJB+ 08] J. Novembre, T. Johnson, K. Bryc, Z. Kutalik, A. R. Boyko, A. Auton, A. Indap, K. S.
King, S. Bergmann, M. R. Nelson, et al. Genes mirror geography within europe. Nature,
456(7218):98–101, 2008.
[O’D14]
R. O’Donnell. Analysis of Boolean Functions. Cambridge University Press, 2014.
[OF96]
B. A. Olshausen and D. J. Field. Emergence of simple-cell receptive field properties by
learning a sparse code for natural images. Nature, 381(6583):607, 1996.
[RL05]
P. J. Rousseeuw and A. M. Leroy. Robust regression and outlier detection, volume 589.
John Wiley & Sons, 2005.
[Rou85]
P. Rousseeuw. Multivariate estimation with high breakdown point. Mathematical Statistics and Applications, pages 283–297, 1985.
[SCV18]
J. Steinhardt, M. Charikar, and G. Valiant. Resilience: A criterion for learning in
the presence of arbitrary outliers. In Proc. of the 9th Conference on Innovations in
Theoretical Computer Science, 2018. to appear.
[Tuk75]
J. W. Tukey. Mathematics and picturing of data. In Proceedings of ICM, volume 6,
pages 523–531, 1975.
[Ver10]
R. Vershynin. Introduction to the non-asymptotic analysis of random matrices, 2010.
A
Lower Bounds on Agnostic Learning
In this section, we prove information theoretic lower bounds for robust distribution learning. In
particular, we will prove that no algorithm (efficient or inefficient) can learn a general distribution
ε
. We note that our lower
from an ε-corrupted set of samples to total variation distance less than 1−ε
bound applies in more general settings than we consider in this paper in couple of ways:
1. Our construction is univariate;
2. Our construction works for any pair of distributions which are
Gaussian distributions; and
ε
1−ε -close,
and not just for
3. Our construction holds for Huber’s ε-contamination model, which is weaker than the noise
model studied in this paper.
In Huber’s ε-contamination model, data is drawn from a mixture distribution (1 − ε)P + εQ,
where P is some distribution that we wish to estimate and Q is arbitrary (in particular, it might
ε
-close can be made indistindepend on P ). We will show that any two distributions which are 1−ε
guishable under this contamination model.
ε
Lemma 16. Let p1 and p2 be two distributions such that dTV (p1 , p2 ) = 1−ε
. Then there exist
Huber ε-contaminations p1′ and p2′ of p1 and p2 respectively, such that p1′ = p2′ . Therefore, no
ε
in Huber’s ε-contamination model.
algorithm can learn a distribution p up to accuracy < 1−ε
37
Proof. Since dTV (p1 , p2 ) =
ε
1−ε ,
we can write
ε
ε
p1 = 1 −
q1 ,
pc +
1−ε
1−ε
ε
ε
p2 = 1 −
q2 ,
pc +
1−ε
1−ε
where pc , q1 , q2 are distributions. Let p1′ be the Huber ε-contamination of p1 with q2 , p2′ be the
Huber ε-contamination of p2 with q1 . In other words,
p1′ = (1 − 2ε) pc + εq1 + εq2 ,
p2′ = (1 − 2ε) pc + εq2 + εq1 ,
and thus p1′ = p2′ as desired.
We note that a similar lower bound holds when p1 and p2 are required to be Gaussians, but
weakened by a factor of 2. This is due to the geometry of the space of Gaussian distributions, and
we note that in the case where the variance is known, this is achieved by the median.
Lemma 17.pNo algorithm
can output an estimate for the mean of a unit-variance Gaussian at
π
accuracy <
2 − o(1) ε with probability > 1/2 in Huber’s ε-contamination model. Consequently,
by Lemma 1, no algorithm can learn a Gaussian to total variation distance < 12 − o(1) ε in Huber’s
ε-contamination model.
Proof. We will consider the distributions p1 = N (−α, 1) and p2 = N (α, 1), where α is to be
specified later. We note that if p1 and p2 can be ε-corrupted into the same distribution, then the
best estimate for the mean is to output 0 (by symmetry), and the result holds.
2 (x)}
and
We will show that p1 can be ε-corrupted into a distribution f , where f (x) = max{p1 (x),p
η
η is a normalizing constant (the case of p2 follows similarly). In other words, f can be written as
(1 − ε)p1 + εq1 for some distribution q1 . Since q1 is a distribution, it is non-negative, and thus we
2 (x)}
require that (1 − ε)p1 (x) ≤ f (x) for all x ∈ R. Note that f (x) = max{p1 (x),p
≥ p1η(x) .
η
We can compute η as follows:
Z ∞
Z ∞
p2 (x)dx
max {p1 (x), p2 (x)} dx = 2
η=
0
−∞
r
α
α
1 1
2
+ erf √
α − O(α3 )
= 1 + erf √
=1+
=2
2 2
π
2
2
Thus,
p (x)
p1 (x)
q 1
=
f (x) ≥
≥
η
1 + π2 α − O(α3 )
1−
r
!
2
α + O(α2 ) p1 (x).
π
q
Again, since we require that (1 − ε)p1 (x) ≤ f (x), it suffices that 1 − ε ≤ 1 − π2 α + O(α2 ) and thus
q
pπ
that π2 α − O(α2 ) ≤ ε. This corresponds to α ≤
2 − o(1) ε, and the proof is complete.
Observe that this lower bound can be strengthened by a factor of two in the subtractive adversary model, where the adversary is able to both remove εN samples and add εN samples. In this
38
infinite sample regime, this means that a distribution can be corrupted to any distribution which is
ε-far in total
√ variation
distance. By noting that N (−α, 1) and N (α, 1) are both ε-far from N (0, 1)
for α =
2π + o(1) ε, we obtain the following lemma. Note that this lower bound is also achieved
by the median in this model.
Lemma 18.√No algorithm
can output an estimate for the mean of a unit-variance Gaussian at
2π + o(1) ε with probability > 1/2 in the subtractive adversary model. Consequently,
accuracy <
by Lemma 1, no algorithm can learn a Gaussian to total variation distance < ε in the subtractive
adversary model.
Finally, we conclude by sketching a lower bound for mean estimation of sub-Gaussian distributions.
Lemma 19. No algorithm can output an estimate for the mean of a sub-Gaussian distribution at
accuracy o(ε log1/2 (1/ε)) with probability > 1/2 in the Huber’s ε-contamination model.
Proof. We start with the distribution q = N (0, 1). We construct p1 by truncating the right tail of
q at the point xr = c1 log1/2 (1/ε) for some constant c1 , and rescaling the rest of the distribution
appropriately. Observe that, for an appropriate choice of c1 :
1
;
• p1 is sub-Gaussian with a constant of O 1−ε
• The mean of p1 is −c2 ε log1/2 (1/ε) for some constant c2 ;
• p1 can be corrupted in Huber’s ε-contamination model to be q.
We can similarly consider p2 , which is constructed by truncating the left tail of q at the point
xl = −c1 log1/2 (1/ε). Since p1 and p2 are indistinguishable when they are both ε-corrupted to q,
and the mean of all three distributions are separated by ≥ c2 ε log1/2 (1/ε), the lemma follows.
B
B.1
Omitted Proofs from Section 2
Proof of Lemma 1
Proof. Observe that by rotational and translational invariance, it suffices to consider the problem
when µ1 = −εe1 /2 and µ2 = εe1 /2, where e1 is the first standard basis vector. By the decomposability of TV distance, we have that the TV distance can in fact be written as a 1 dimensional
integral:
Z ∞
1
1
2
2
dTV (N (µ1 , I), N (µ2 , I)) = · √
e−(x−ε/2) /2 − e−(x+ε/2) /2 dx .
2
2π −∞
2 /2
The value of the function f (x) = e−(x−ε/2)
when x > 0, hence this integral becomes
2 /2
− e−(x+ε/2)
is negative when x < 0 and positive
Z ∞
1
2
2
e−(x−ε/2) /2 − e−(x+ε/2) /2 dx
dTV (N (µ1 , I), N (µ2 , I)) = √
2π 0
= F (ε/2) − F (−ε/2) ,
39
Rx
2
where F (x) = √12π −∞ e−t /2 dt is the CDF of the standard normal Gaussian. By Taylor’s theorem,
and since F ′′ (x) is bounded when x ∈ [−1, 1], we have
F (ε/2) − F (−ε/2) = F ′ (−ε/2)ε + O(ε3 )
1
2
= √ e−(ε/2) /2 ε + O(ε3 )
2π
1
= √ + o(1) ε ,
2π
which proves the claim.
C
Omitted Proofs from Section 4
C.1
Proof of Theorem 7
First, we require the following several claims about p(x) under various distributions, including the
true Gaussian and from choosing points uniformly at random from the sets of interest. Note that
throughout this section we will use Lemma 3, which implies that kΠV ′ µ − µ̃k2 < 2ε when the
parameters α, γ are chosen appropriately.
Claim 4. EN (µ,I) [p(X)] = kΠV ′ µ − µ̃k22 .
Proof. Letting v1 , . . . , vC1 β log(1/ε) be an orthonormal basis of V ′ , we have
C1 β log(1/ε)
E [p(X)] =
N (µ,I)
X
i=1
C1 β log(1/ε)
E
N (viT µ,1)
2
[hvi , X − µ̃i − 1] =
X
i=1
(viT (µ − µ̃))2 = kΠV ′ µ − µ̃k22 .
Claim 5. For some absolute constant c0 and all t ≥ kΠV ′ µ − µ̃k22 ,
(t − kΠV ′ µ − µ̃k22 )2
2
Pr (p(X) > t) ≤ 2 exp −c0 · min
, t − kΠV ′ µ − µ̃k2
.
C1 β log(1/ε)
N (µ,I)
Proof. This follows from Lemma 7, after re-centering the polynomial using Claim 4 and noting that
the spectral norm and squared Frobenius norm of the corresponding A matrix are at most 1 and
C1 β log(1/ε), respectively.
Claim 6. ES [p(X)] ≥ C1 ε log(1/ε).
Proof. Recall that µ̂ is the empirical mean of the point set.
E[p(X)] = E[kΠV ′ X − µ̃k22 ] − dim(V ′ )
S
S
≥ E[kΠV ′ X − ΠV ′ µ̂k22 ] − dim(V ′ )
S
1
≥ 1 + ε dim(V ′ ) − dim(V ′ ) = C1 ε log(1/ε) .
β
Claim 7. EG0 [p(X)] ≤ kµ − µ̃k22 + O(γ)ε.
40
Proof.
E [p(X)] ≤
G0
≤
≤
Z
0
Z
∞
Pr[p(X) ≥ t]dt
G0
O(d log(d/εδ))
Pr[p(X) ≥ t]dt
G0
0
Z
O(d log(d/εδ))
Pr [p(X) ≥ t]dt + O(γ)ε
N (µ,I)
0
≤ kµ − µ̃k22 + O(γ)ε
The inequalities follow from (γε, δ)-goodness and Claim 4.
Claim 8. φ(S, G0 ) EL [p(X)] = (C1 β + O(γ) + o(1))ε.
Proof.
φ(S, G0 ) E[p(X)] =
L
Z
∞
φ(S, G0 ) Pr[p(X) > t]dt
L
0
=
Z
O(d log(|G|/δ))
φ(S, G0 ) Pr[p(X) > t]dt
=
Z
(4)
L
0
C1 β log(1/ε)+4ε2
φ(S, G0 ) Pr[p(X) > t]dt
0
+
Z
L
O(d log(|G|/δ))
C1 β log(1/ε)+4ε2
φ(S, G0 ) Pr[p(X) > t]dt
L
2
≤ (C1 β log(1/ε) + 4ε )φ(S, G0 ) + 2
Z
O(d log(|G|/δ))
Pr[p(X) > t]dt
C1 β log(1/ε)+4ε2 G0
≤ (C1 β log(1/ε) + 4ε2 )φ(S, G0 )
Z O(d log(|G|/δ))
Pr [p(X) > t] +
+2
C1 β log(1/ε)+4ε2 N (µ,I)
2
γε
dt
d log(|G|/δ)
≤ (C1 β log(1/ε) + 4ε )φ(S, G0 ) + O(γε) + 8εc0 C1 β
≤ (C1 β + O(γ) + o(1))ε.
(5)
(6)
(7)
(4) and (5) follow from G0 being (γε, δ)-good, (6) is from Claim 5, and (7) is because
φ(S, G0 ) log(1/φ(S, G0 )) ≤ (1 + o(1))ε .
This gives:
Claim 9. ψ EE [p(X)] ≥ C1 ε log 1/ε − (C1 β + 1 + O(γ) + o(1))ε.
Proof. This immediately follows from Claims 6, 7, and 8.
We now show that in the case that there are many large eigenvalues, there will be a T satisfying
the conditions of the filter.
Claim 10. Suppose dim(V ) ≥ C1 β log(1/ε). Then there is a T satisfying the conditions in the
algorithm.
41
Proof. Suppose not. Then, we have
ψ E[p(X)] ≤
E
Z
2C3 log(1/ε)/c0
0
ψ Pr[p(X) ≥ t]dt +
E
≤ 2(C3 /c0 )ψ log 1/ε +
≤ 2(C3 /c0 )ψ log 1/ε +
≤ (2C3 /c0 )ε log 1/ε +
Z
Z
Z
C2 d log |S|/δ
Z
2(C3 /c0 ) log(1/ε)
C2 d log |S|/δ
∞
2C3 log(1/ε)/c0
ψ Pr[p(X) ≥ t]dt
E
ψ Pr[p(X) ≥ t]dt
E
Pr[p(X) ≥ t]dt
2(C3 /c0 ) log(1/ε) S
C2 d log |S|/δ
Pr[p(X) ≥ t]dt .
2(C3 /c0 ) log(1/ε) S
By applying the contradiction assumption, we have
c0 T
γε
Pr[p(X) ≥ t]dt ≤
exp −
dt
+
2C3
d log |S|/δ
2(C3 /c0 ) log(1/ε) S
2(C3 /c0 ) log(1/ε)
Z C2 d log |S|/δ
c0 T
exp −
≤
dt + C2 γε
2C3
2(C3 /c0 ) log(1/ε)
Z ∞
c0 T
dt + C2 γε
exp −
≤
2C3
2(C3 /c0 ) log(1/ε)
Z ∞
exp (−t) dt + C2 γε
≤ (C3 /c0 )
Z
Z
C2 d log |S|/δ
C2 d log |S|/δ
2 log(1/ε)
2
≤ (C3 /c0 ) · O(ε ) + C2 γε ,
and thus, we have
ψ E[p(X)] ≪ C1 ε log 1/ε − (C1 β + 1 + O(γ) + o(1))ε ,
E
which contradicts Claim 9.
It now suffices to prove that if we construct a filter, then the invariant that ∆ decreases is
preserved. Formally, we show:
Claim 11. Suppose dim(V ) ≥ C1 β log(1/ε). Let S ′ be the set of points we return. Then ∆(S ′ , G) <
∆(S, G).
Proof. Let T be the threshold we pick. If T > C2 d log(|S|/δ) then the invariant is satisfied since we
remove no good points, by (γε, δ)-goodness. It suffices to show that in the other case, we remove
log(1/ε) times many more bad points than good points. By definition we remove at least
γε
c0 T
+
.
|S| · exp −
2C3
d log |S|/δ
points. On the other hand, by (γε, δ)-goodness, Claim 5, we know that we throw away at most
|G| · Pr[p(X) > T ] ≤ |G0 | · Pr[p(X) > T ]
G
G0
γε
(T − kµ − µ̃k22 )2
2
, (T − kµ − µ̃k2 )
+
≤ |G0 | exp −c0 · min
.
C1 β log(1/ε)
2 log(1/ε)(d log(|S|/δ))
42
By our choice of T , and since kµ − µ
ek2 ≤ O(δ) ≪ 1, the first term is upper bounded by
(C3 − 1)2
(T − kµ − µ̃k22 )2
2
, (T − kµ − µ̃k2 )
≤ exp −c0 · min
T, (C3 − 1)T )
exp −c0 · min
C1 β log(1/ε)
C1 β
and so
exp −c0 · min
(C3 − 1)2
c0 T
T, (C3 − 1)T )
≤ exp −
C1 β
C3
c0 T 2
= exp −
2C3
c0 T
·ε
≤ exp −
2C3
1
c0 T
≪
.
exp −
2C3
log2 1/ε
Since we have maintained this invariant so far, in particular, we have thrown away more bad points
than good points, and so |G0 | ≤ (1 + ε)|S|.
γε
c0 T
Therefore, we have log(1/ε) · |G| · PrG [p(X) > T ] ≤ |S| · exp − 2C
+
d log |S|/δ , and hence,
3
the invariant is satisfied.
Claims 10 and 11 together imply the correctness of Theorem 7.
C.2
Proof of Lemma 5
We require the following basic fact of good sets (see, e.g., Fact 8.6 in [DKK+ 16]):
Fact 2. Let v ∈ Rd be any unit vector. Let G be (η, δ)-good with respect to N (µ, I). Then for any
T > 0,
Pr(|v · (X − µ)| > T ) ≤ 2 exp(−T 2 /2) + η/(d log(d/ηδ)) ,
G
and
Pr (|v · (X − µ)| > T ) ≤ 2 exp(−T 2 /2).
N (µ,I)
ε
Proof. Let hv, µ − µ
bi = R > β 1/2
. Observe that by direct calculation, we have ES [hv, X −
2
2
2
µi ] − ES [hv,
bi ] ≤ 3R . Hence, it suffices to show that ES [hv, X − µi2 ] ≥ 1 + Ω(R2 /ε) −
X − µ
ε.
γ + O β1
We write S = (G0 \ L, E) where |E| = ψ|S| and |L| = φ|S|. First we consider the expectation
of v · X over X in S. This is
µ0 − φµL + ψµE
.
1−φ+ψ
where µ0 is the mean over G0 , µL the mean over L and µE the mean over E. By (γε, δ)-goodness,
we have that |hv, µ0 − µi| ≤ γε. We also have that
Z ∞
Pr [|hv, X − µi| ≥ t] dt
|hv, µL − µi| ≤
L
0
Z √
O(
d log(d/εδ))
Pr [|hv, X − µi| ≥ t] dt
=
L
0
43
Since we have
|G0 |
Pr [|hv, X − µi| > t] ,
Pr [|hv, (X − µ)i| > t] ≤ min 1,
L
|L| G0
(8)
by Fact 2 and the calculations done in the proof of Corollary 8.8 in [DKK+ 16], we have that
ε
,
|hv, µL − µi| ≤ O(log |S|/|L|) ≤ O(log 1/φ). Since by assumption we have hv, µ − µ
bi = R > β 1/2
this implies that hv, µE − µi = Ω(R/ε). In particular, this implies that
E hv, X − µi2 ≥ E [hv, X − µi]2
E
E
2
R
≥Ω
.
ε2
Next we consider the expectation of hv, X −µi2 . By (γε, δ)-goodness, we have EG0 [hv, X −µi2 ] =
1 + O(ε). We also have that
Z ∞
2
φ Pr |hv, X − µi|2 ≥ t dt
φ E hv, X − µi ≤
L
0
L
5
≤ log(1/ε)φ +
β
≤
5
log(1/ε)φ +
β
Z
O(d log d/εδ)
(5/β) log(1/ε)
Z
O(d log d/εδ)
φ Pr |hv, X − µi|2 ≥ t dt
L
Pr |hv, X − µi|2 ≥ t dt
(5/β) log(1/ε) G0
Z
O(d log d/εδ)
5
Pr |hv, X − µi|2 ≥ t dt + γε
log(1/ε)φ +
β
(5/β) log(1/ε) N (µ,I)
1
= O
+ γ + o(1) ε.
β
=
Since
EG0 [hv, (X − µ)i2 ] − φ EL [hv, (X − µ)i2 ] + ψ EE [hv, (X − µ)i2 ]
,
S
1−φ+ψ
ε.
this is at least 1 + Ω(R2 /ε) − γ + O β1
E[hv, (X − µ)i2 ] =
D
D.1
Omitted Proofs from Section 6
Proof of Theorem 15
b − I, and
Our algorithm works as follows, just as for Algorithm 4. It finds all large eigenvalues of Σ
if there are too many, produces an explicit degree-2 polynomial which, as we will argue, produces
a valid filter. The formal pseudocode for our algorithm is in Algorithm 11.
The proofs of the following claims are identical to the proofs of Claim 6-9, by applying the
corresponding property of (γε, δ)-goodness for this setting, and so we omit them.
Claim 12. ES [p(x)] ≥ Cξ.
Claim 13. EG0 [p(X)] ≤
ε
.
(d log 1/ε)2
Claim 14. φ(S, G0 ) EL [p(X)] = (C + O(γ) + o(1))ε.
44
Algorithm 11 Filter if there are many large eigenvalues of the covariance
1: function FilterCovManyDeg2Eig(S, ε, ξ, δ)
b be the empirical second moment of S, respectively.
2:
Let Σ
b − I with eigenvalue more than
3:
Let V be the subspace of Rd spanned by eigenvectors of Σ
Cξ.
4:
if dim(V ) ≥ C1 log(1/ε) then
5:
Let V ′ be a subspace of V of dimension C1 log(1/ε).
6:
Let p(x) be the quadratic polynomial
p(x) = kΠV ′ (x)k22 − dim(V ′ ).
7:
Find a value T > 0 so that either:
• T > C2 d log(|S|/δ) and p(x) > T for at least one x ∈ S, or
• T > 2C3 log(1/ε)/c0 and PrS (p(x) > T ) > exp(−c0 T /(2C3 )) + ε/(d log(|S|/δ)).
8:
return
S ′ = {x ∈ S : p(x) ≤ T } .
else
return an orthonormal basis for V .
11:
end if
12: end function
9:
10:
Claim 15. ψ EE [p(X)] ≥ Cξ − (C1 + 1 + O(γ) + o(1))ε.
These claims imply just as before that there is a T with the desired properties. Again, the proof
is identical. Formally:
Claim 16. Suppose dim(V ) ≥ C log(1/ε). Then there is a T satisfying the conditions in the
algorithm.
Finally, we show the invariant that ∆ decreases. This is almost identical to the proof of Claim
11, however, we need to slightly change our application of the Hanson-Wright inequality. Formally,
we show:
Claim 17. Suppose dim(V ) ≥ C1 log 1/ε. Let S ′ be the set of points we return. Then ∆(S ′ , G) <
∆(S, G).
Proof. Let T be the threshold we pick. If T > C2 d log(|S|/δ) then the invariant is satisfied since we
remove no good points, by (γε, δ)-goodness. It suffices to show that in the other case, we remove
log 1/ε times many more bad points than good points. By definition we remove at least
c0 T
ε
|S| · exp −
+
2C3
d log |S|/δ
points. On the other hand, observe that if v1 , . . . , vk is an orthonormal basis for V , we have
X
kΠV xk22 =
hvi , Xi2
X
vi viT Σ1/2 X ,
= yΣ1/2
45
so that ifPX ∼ N (0, Σ), we have Y ∼ N (0, I). Let M = Σ1/2
kM kF ≤ ni=1 kΣk2 ≤ (1 + ξ)k, and kM k2 ≤ kΣk2 ≤ (1 + ξ). Since
E [kΠV xk22 ] − k =
N (0,Σ)
≤
k
X
i=1
k
X
i=1
P
vi viT Σ1/2 . We have that
viT (Σ − I)vi
kΣ − Ik2
≤ (1 + ξ)k ,
we have by Hanson-Wright that
Pr
N (0,Σ)
(T − (1 + ξ)k)2
, (T − kµ − (1 + ξ)kk22 )
,
kΠV xk22 − dim(V ) > T ≤ exp −c0 · min
C1 log(1/ε)
so by our choice of T , we have
(T − kµ − µ̃k22 )2
(C3 − 1)2
2
exp −c0 · min
, (T − kµ − µ̃k2 )
≤ exp −c0 · min
T, (C3 − 1)T )
.
C1 log(1/ε)
C1
The remaining proof now proceeds identically to the proof of Claim 17.
D.2
Proof of Theorem 16
Clearly, the only non-trivial condition to certify for Theorem 16 is that if we are in Case (1), the
returned set satisfies the desired properties.
Our proof will roughly follow the same structure as the proof of Theorem 7. We will first show
that the empirical average of the polynomial Q can only be large because of the contribution of the
points in E (Claim 21). We will then show that this implies that there exists a threshold T which
the algorithm will find in this case (Claim 22). Finally, we will show that for any such T we find,
the returned set of points will indeed satisfy ∆(S ′ , G0 ) < ∆(S, G0 ), which implies the correctness
of the algorithm (Claim 23).
We first show the following claim:
Claim 18. ES [ri ] − EN (0,Σ) [ri ] ≤ (4C + 1)ξ
√
Proof. Let us suppose that pi corresponds to the matrix Ai , given by (Ai )a,b = ∇a ∇b pi / 2!, so
that the Ai are orthonormal with respect to the Frobenius norm. Then the constant harmonic
2
2
2
part
√ of2 pi corresponds to kAi kF ≤ 1. The degree-2 harmonic part of p2i corresponds to the matrix
2 2Ai . This is because if we let Bi be the matrix corresponding to pi , we have
1
∇a ∇b p2i (X)
E
(Bi )a,b = √
2 X∼N (0,I)
2
=√
E
[∇b (pi (X)∇a pi (X))]
2 X∼N (0,I)
2
E
[∇a pi (X)∇b pi (X) + pi (X)∇a ∇b pi (X)]
=√
2 X∼N (0,I)
2
=√
E
[∇a pi (X)∇b pi (X)]
2 X∼N (0,I)
√
= 2 2h(Ai )a , (Ai )b i ,
46
where the second to last line follows since ∇a ∇b pi (X) is a constant, and the last line follows from
explicit computation. In particular,√this implies that the non-constant component of
Pri corresponds
to matrix with trace norm at most
2
2
≤
4.
Therefore,
r
can
be
written
as
r
(x)
=
αi hvi , xi2 +C0
i
i
P
for some constant C0 , where
|αi | ≤ 4. Thus, by our assumption, we have ES [ri ] ≤ 1 + Cξ. The
claim then follows since Σ and I are differ in Frobenius norm by at most ξ.
Claim 19. ES [Q(X)] − EX∼N (0,Σ) [Q(X)] ≥ (C1 − 6)ξk
Proof. Observe that since kΣ − IkF ≤ ξ, in particular we have kΣ⊗2 − I ⊗2 k2 ≤ ξ, and hence
by Lemma 10 we have that EN (0,Σ) [p2 (X)] − EX∼N (0,I) [p2 (X)] ≤ 2ξ for all p ∈ P2 . Hence, in
particular, we have
E[p2i (X)] − E [p2i (X)] ≥ (C1 − 2)ξ .
S
By Claim 18, we have
E
S
"
In particular, this implies that
N (0,Σ)
k
X
i=1
#
ri −
E [Q(X)] −
S
E
N (0,Σ)
"
k
X
i=1
#
ri ≤ (4C + 1)k .
E [Q(X)] ≥ (C1 − 4C − 5)ξk ,
N (0,I)
as claimed.
We now show:
Claim 20. There is some universal constant B so that
E
√
[Q2 (X)]1/2 ≤ B k .
X∼N (0,Σ)
Before we prove this, we need the following lemma:
Lemma 20. For any degree-4 polynomial p, and any Σ, if we let F (y, p, Σ) denote the yth percentile
of p under Σ, then we have F (1/4, p2 , Σ), F (3/4, p2 , Σ) = Θ(EX∼N (0,Σ) [p2 (X)]).
Proof. Let µ′ = EX∼N (0,Σ) [p2 (X)]. First, we note that Pr(p2 (X) > 4µ′ ) ≤ 1/4, so F (3/4, p2 , Σ) ≤
4µ′ . On the other hand,
anti-concentration bounds [CW01], we have that Pr(p2 (X) ≤
√ ′ by known
1/8
′
εµ ) = Pr(|p(X)| ≤ εµ ) = O(ε ). So, for ε a sufficiently small constant, Pr(p2 (X) ≤ εµ′ ) < 1/4,
and therefore, F (1/4, p2 , Σ) ≥ εµ′ . Since εµ′ ≤ F (1/4, p2 , Σ) ≤ F (3/4, p2 , Σ) ≤ 4µ′ , this completes
our proof.
Proof of Claim 20. Since dTV (N (0, Σ), N (0, I)) = O(ε log 1/ε) = o(1), we have that F (3/4, Q2 , Σ) ≤
F (1/4, Q2 , Σ), so by the above lemma, we have that
E
[Q2 (X)]1/2 ≤ O(
X∼N (0,Σ)
E
[Q2 (X)]1/2 ) .
X∼N (0,I)
Hence, it suffices to bound kQk22 .
47
√
Let us again suppose that pi corresponds to the matrix Ai , given by (Ai )a,b = ∇a ∇b pi / 2!.
Note that the Ai are symmetric matrices that form an orthonormal set. We note that qi is the
harmonic degree-4 polynomial corresponding to the rank-4 tensor
√
Ta,b,c,d = ∇a ∇b ∇c ∇d qi / 24
√
= ∇a ∇b ∇c ∇d (p2i )/ 24
√
= (2(∇a ∇b pi )(∇c ∇d pi ) + 2(∇a ∇c pi )(∇b ∇d pi ) + 2(∇a ∇d pi )(∇b ∇c pi ))/ 24
√
= (4(Ai )a,b (Ai )c,d + 4(Ai )a,c (Ai )b,d + 4(Ai )a,d (Ai )b,c )/ 24
√
= 6Sym(Ai ⊗ Ai ).
√ P
By linearity, Q corresponds to the rank-4 tensor T = 6 ki=1 Sym(Ai ⊗ Ai ). It thus suffices to
show that kT k2 = O(k), where here k · k2 denotes the square root of the sum of the squares of the
entries of the tensor. In order to show this, we note that kT k2 = supkV k2 =1 hV, T i. Therefore, it
√
suffices to show that for all 4-tensors V with kV k2 ≤ 1, that hV, T i = O( k). We note that
+
+
*
*
k
k
X
X
√
√
hV, T i = 6 V,
Sym(Ai ) ⊗ Ai = 6 Sym(V ),
Ai ⊗ Ai .
i=1
i=1
Note that Sym(V ) is a symmetric 4-tensor of ℓ2 norm at most 1. Thinking of Sym(V ) as a symmetric
Pn 2
matrix over 2-Tensors, we can write it as Sym(V ) =
j=1 λj Bj ⊗ Bj , where the Bj ’s are an
Pn 2 2
orthonormal basis for the set of 2-tensors and j=1 λj = kSym(V )k22 ≤ 1. Then, if ci,j := hAi , Bj i,
P
and cj = ki=1 c2i,j , we have that
2
2
k
n
n
X
√ X
√ X
c2i,j = 6
λj
λj cj .
hV, T i = 6
i=1
j=1
j=1
Since the Ai are orthonormal, cj ≤ 1 for all j. Furthermore,
2
2
n
X
cj =
c2i,j =
k
X
i=1
i=1 j=1
j=1
Therefore,
n
k X
X
kAi k2F = k.
v
v
u n2
u n2
n2
uX
√
√
√ X
√ uX
hV, T i = 6
λj cj ≤ 6t
λ2j t
c2j ≤ 6 · 1 · k = O( k) .
j=1
j=1
j=1
Let µ′ = EX∼N (0,Σ) [Q(X)]. We now show that since G0 is ε-good, then almost all of the
difference in Claim 19 must be because of the points in E.
Claim 21. EE [Q(Y )] − µ′ ≥
C1 −6
2
·
ξk
φ
Proof. It suffices to show that
E[Q(X)] − µ′ ≤
G
48
C1 − 6
ξk ,
2
(9)
that is, the good points do not contribute
√ much to the difference. By (ε, δ)-goodness of G0 and
′
Claim 20 we have E0 [Q(Y )] − µ ≤ O ε k . Moreover, we have
φ(E[Q(X)] −
L
E
[Q(X)]) ≤
X∼N (0,Σ)
(a)
=
Z
∞
0
Z
φ Pr Q(X) − µ′ ≥ t dt
L
√ 2
O( kd log |S|/δ)
φ Pr Q(X) − µ′ ≥ t dt
L
0
Z
√
2
≤ φ · O( k log 1/ξ) +
√
≤ φ · O( k log2 1/ξ) +
√
≤ φ · O( k log2 1/ξ) +
(b)
√
O( kd2 log |S|/δ)
φ Pr
√
L
O( k log2 1/ξ)
Z O(√kd2 log |S|/δ)
Pr
√
G0
O( k log2 1/ξ)
Z O(√kd2 log |S|/δ)
Q(X) − µ′ ≥ t dt
Q(X) − µ′ ≥ t dt
Pr [Q(X) − µ′ ≥ t]dt + O(ξ)
√
O( k log2 1/ξ)
N (0,Σ)
√
≤ O(ξ k) + exp(−Ω(log 1/ξ))
(c)
≪ 5ξk ,
where (a) follows from the boundedness condition of (ε, δ)-goodness, (b) follows from the last
condition of goodness, and (c) follows from hypercontractivity. This shows (9), which completes
the proof.
We now show that this implies that there must be a T satisfying the conditions in Algorithm 8.
Claim 22. If dim(Vm ) ≥ k, then Algorithm 8 returns a T satisfying the conditions in the algorithm.
Proof. Suppose not. By the assumption that kI − ΣkF ≤ ξ, we have |µ′ | = | EX∼N (0,Σ) [Q(X)]| ≤
P
2ξkQk2 ≤ ki=1 kpi k2 + kri k2 ≤ 5ξk. Thus, we have
C1 − 6
C1 − 6
− 5 ξk ≫
E[Q(X)] ≥
ξk .
E
2ψ
3ψ
But since we assume there is no T satisfying the conditions in Algorithm 8, we have
Z ∞
Pr[p(X) > t]dt
ψ E [p(X)] ≤ ψ
X∈E
0
E
√
2
≤ ψ · 4B k log (1/ξ) +
≤ 4Bψ ·
√
k log2 (1/ξ) +
≤ 4C2 Bψ ·
√
Z
√
O(d2 k log(|S|/δ))
ψ Pr[p(X)
√
E
4B k log2 1/ξ
Z O(d2 √k log(|S|/δ))
Pr[p(X)
√
S
4B k log2 1/ξ
Z O(d2 √k log(|S|/δ))
k log2 (1/ξ) +
√
4A2 C2 B k log2 1/ξ
√ Z ∞
√
≤ 4A2 C2 Bψ k log2 (1/ξ) + 4B k
√
e 2 ) ≪ 10ξk ,
≤ O( kξ) + O(ε
A2 log2 1/ξ
> t]dt
> t]dt
t
√
exp −A
4B k
1/2 !
ε2
+ 2
d log |G|/δ
e 2)
exp(−Ω(t1/2 ))dt + O(ε
which is a contradiction, for our choice of C1 , C2 , C3 , and since we chose k = O(log4 1/ε).
49
!
dt
The final thing we must verify is that the number of good points we remove is much smaller
than the number of bad points we remove. Formally, we show:
Claim 23. If dim(Vm ) ≥ k, then Algorithm 8 returns a S ′ satisfying ∆(S ′ , G0 ) < ∆(S, G0 ).
√
Proof. Let T be the threshold we pick. If T > C3 d2 k log(|S|) then the invariant is satisfied since
we remove no good points, by (ε, δ)-goodness. It suffices to show that we remove log 1/ε times
many more bad points than good points. Otherwise, by definition we remove a total of
!
1/2 !
ε2
T
√
+ 2
|S| · exp −A
d log2 |S|/δ
2B k
points. On the other hand, by (ε, δ)-goodness, hypercontractivity, and Claim 20, we know that the
total number of points we throw away is at most
|G| · Pr[Q(X) > T ] ≤ |G0 | · Pr[Q(X) > T ]
G
G0
−1/2 !
!
ε2
+
≤ |G0 | exp −A
2 log(1/ε)(d log(|G|/δ))2
!
1/2 !
T
ε2
√
≤ |G0 | exp −A
.
+
2 log(1/ε)(d log(|G|/δ))2
B k
T
kQk2
Since we have maintained this invariant so far, in particular, we have√
thrown away more bad points
than good points, and so |G0 | ≤ (1 + ε)|S|. Moreover, since T ≥ 4B k log2 1/ε, we have
exp −A
T
√
B k
1/2 !
= exp −A
≤ exp −A
≤ exp −A
≤ exp −A
T
√
4B k
T
√
4B k
T
√
4B k
T
√
4B k
1/2 !2
1/2 !
1/2 !
1/2 !
· exp −A
· ε−1
·
!1/2
√
2
4B k log 1/ε
√
4B k
1
.
log 1/ε
1/2
Therefore, we have log(1/ε) · |G| · PrG [Q(X) > T ] ≤ |S| · exp −A 4BT√k
+
hence, the invariant is satisfied.
ε2
d2 log |G|/δ
, and
Claims 22 and 23 together prove the theorem.
D.3
Proof of Lemma 6
By Lemma 8.16 in [DKK+ 16], the first two items hold together with probability 1 − O(δ) after
taking O(poly(d, 1/η, log 1/δ)) samples. Thus, it suffices to show that the last property holds with
probability 1 − O(δ) given O(poly(d, 1/η, log 1/δ)) samples. We may clearly WLOG take Σ = I.
Moreover, these properties are clearly invariant under scaling, and hence, it suffices to prove them
50
for degree-four polynomials with VarN (0,I) [p(X)] = 1. For any fixed degree-4 polynomial p, by
P
hypercontractivity, since P (X) = n1 ni=1 p(Xi ) has variance Var[P (X)] = n1 , we have that
"
#
n
1X
Pr
p(Xi ) − E [p(X)] ≥ t ≤ exp −At1/2 n1/4 .
n
N (0,I)
i=1
In particular, if we take
n=Ω
then
Pr
"
log4 1/δ
η2
,
#
n
1X
p(Xi ) − E [p(X)] ≥ η ≤ O(δ) .
n
N (0,I)
i=1
4
Since there is a 1/3-net over all degree-4polynomials
with unit variance of size (1/3)O(d ) , by union
d4 log1/δ
samples, then the second to last property holds
bounding, we obtain that if we take Ω
η2
with probability 1 − δ.
Finally, the same net technique used to prove Lemma 8.16, along with hypercontractivity, may
be used to show the last property holds when given poly(d, log 1/η, log 1/δ) samples.
D.4
Proof of Lemma 15
Proof. Let µ′ = EN (0,Σ) [p(X)], s = EN (0,Σ) [p2 (X)], and let ES [p(X)] − EN (0,Σ) [p(X)] = K. By
(ε, δ)-goodness and Lemma 10, we have
E [p(X)] − µ′ ≤ ε
E [p2 (X)]1/2 ≤ 2ε .
G0
N (0,Σ)
(10)
Moreover, for some appropriate choice of β1 , we have
Z ∞
′
φ Pr [|p(X) − µ| ≥ t] dt
φ E[p(X)] − φµ ≤
L
0
L
≤ β1 φ log 1/ε +
(a)
Z
≤ β1 φ log 1/ε +
(b)
≤ β1 φ log 1/ε +
(c)
O(d log |S|/δ)
β1 log 1/ε
|G0 |
|S|
|G0 |
|S|
Z
Z
φ Pr [|p(X) − µ| ≥ t] dt
L
O(d log |S|/δ)
β1 log 1/ε
O(d log |S|/δ)
β1 log 1/ε
Pr [|p(X) − µ| ≥ t] dt
G0
Pr [|p(X) − µ| ≥ t] dt + O(ε)
N (0,Σ)
≤ β1 φ log 1/ε + εO(β1 ) + O(ε)
= O(ε) ,
(11)
where (a), (b), follows from the definition of goodness, and (c) follows from standard Guassian
concentration bounds. In particular, (10) and (11) together imply that
2K
K
≥ E[p(X) − µ′ ] ≥
,
E
ψ
2ψ
51
so this implies
K2
.
4ψ 2
E[(p(X) − µ′ )2 ] ≥
E
Expanding this yields that
E[(p(X) − µ′ )2 ] = E[p2 (X)] − 2µ′ E[p(X)] + (µ′ )2 ,
E
E
E
so since |µ′ | ≤ kΣ − IkF ≤ ξ, we have
E[p2 (X)] ≥
E
2K
K2
−
+ O(ξ 2 ) .
4ψ 2
ψ
(12)
On the other hand, by Lemma 10, we have
1
|s − 1| ≤ kΣ⊗2 − I ⊗2 k2 + kΣ − Ik2F
2
≤ ξ + O(ξ 2 ) .
(13)
Also, by (ε, δ)-goodness and Lemma 10, we have
E [p2 (X)] − s ≤ ε(1 + O(ξ)) ,
(14)
G0
and, for A as in Corollary 6, and β appropriately chosen,
Z ∞
2
φ(E[p (X)] − s) ≤
φ Pr[|p2 (X) − s| ≥ t]dt
L
0
L
2
≤ βAφ log 1/ε +
Z
O((d log |S|/δ)2 )
|G0 |
≤ βAφ log 1/ε +
|S|
2
(a)
≤ βAφ log2 1/ε +
(b)
≤ βAφ log2 1/ε +
φ Pr[|p2 (X) − s| ≥ t]dt
L
2A log2 1/ε
|G0 |
|S|
Z ∞
Z
O((d log |S|/δ)2 )
βA log 1/ε
Z
Pr[|p2 (X) − s| ≥ t]dt
G0
2
O((d log |S|/δ)2 )
2
βA log 1/ε
1/2 /2
e−At
2
Pr [|p2 (X) − s| ≥ t]dt + O(ε2 )
N (0,Σ)
dt + O(ε2 )
βA log 1/ε
2
≪ 2βAξ + O(ε) + O(ε ) ,
(15)
where (a) follows from goodness, and (b) follows from Corollary 6 and since kΣ − IkF = o(1). Thus,
(12), (13), (14), and (15) together imply that
E[p2 (X)] − 1 ≥ 1 − (ξ + O(ξ 2 )) − O(ε) − 2βAξO(ε) +
S
and so since K ≥ Ω
√
εξ , this gives the desired bound.
52
K2
− 2K + O(ξ 2 ) ,
4ψ
| 10math.ST
|
ON THE ALGEBRAIC AND ARITHMETIC STRUCTURE
OF THE MONOID OF PRODUCT-ONE SEQUENCES
arXiv:1802.00991v1 [math.AC] 3 Feb 2018
JUN SEOK OH
Abstract. Let G be a finite group. A finite unordered sequence S = g1 · . . . · gℓ of terms from G, where
repetition is allowed, is a product-one sequence if its terms can be ordered such that their product equals
1G , the identity element of the group. As usual, we consider sequences as elements of the free abelian
monoid F (G) with basis G, and we study the submonoid B(G) ⊂ F (G) of all product-one sequences.
This is a finitely generated C-monoid, which is a Krull monoid if and only if G is abelian. In case of
abelian groups, B(G) is a well-studied object. In the present paper we focus on non-abelian groups, and
we study the class semigroup and the arithmetic of B(G).
1. Introduction
Let G be a finite group. By a sequence over G, we mean a finite unordered sequence of terms from
G, where the repetition of elements is allowed (the terminology stems from Arithmetic Combinatorics).
A sequence S is a product-one sequence if its terms can be ordered such that their product equals the
identity element of the group. Clearly, juxtaposition of sequences is a commutative operation on the set
of sequences. As usual we consider sequences as elements of the free abelian monoid F (G) with basis G,
and clearly the subset B(G) ⊂ F (G) of all product-one sequences is a submonoid. The small Davenport
constant d(G) is the maximal integer ℓ for which there is a sequence of length ℓ which has no productone subsequence. The large Davenport constant D(G) is the maximal length of a minimal product-one
sequence (by a minimal product-one sequence we mean an irreducible element in the monoid B(G)).
Suppose that G is abelian, and let us use additive notation in this case. Then product-one sequences
are zero-sum sequences and their study is the main objective of Zero-Sum Theory ([12, 25]). The monoid
B(G) is a Krull monoid with class group G (apart from the exceptional case where |G| = 2), and because
it has intimate relationship with general Krull monoids having class group G, the study of B(G) is a
central topic in factorization theory ([18, 15]).
Although the abelian setting has been the dominant one, many of the combinatorial problems on
sequences over abelian groups have also been studied in the nonabelian setting. For example, an upper
bound on the small Davenport constant was given already in the 1970s [29], and in recent years Gao’s
Theorem E(G) = |G| + d(G) ([25, Chapter 16]) has been generalized to a variety of nonabelian groups
([3, 14, 13, 26, 27]). Fresh impetus came from invariant theory. If G is abelian, then d(G) + 1 = β(G) =
D(G), where β(G) is the Noether number. If G is nonabelian and has a cyclic subgroup of index two,
then the Davenport constants and the Noether number have been explicitely determined ([8, 17]), and it
turned out that d(G) + 1 ≤ β(G) ≤ D(G). For a survey on the interplay with invariant theory we refer to
[9] and for recent progress to [24, 6, 7, 10].
Let F be a factorial monoid. A submonoid H ⊂ F is a C-monoid if H × = H ∩ F × and the reduced
class semigroup is finite. A commutative ring is called a C-ring if its monoid of regular elements is a
b is
C-monoid. To give an example of a C-ring, consider a Mori domain R. If its conductor f = (R : R)
2010 Mathematics Subject Classification. 13A50, 20D60, 20M13.
Key words and phrases. product-one sequences, Davenport constant, C-monoids, sets of lengths.
This work was supported by the Austrian Science Fund FWF, W1230 Doctoral Program Discrete Mathematics.
1
2
JUN SEOK OH
b is finite, then R is a C-ring. For more on C-rings we refer to
nonzero and the residue class ring R/f
[30, 21]. The finiteness of the reduced class semigroup is used to derive arithmetical finiteness result for
C-monoids ([18]). However, the structure of the (reduced) class semigroup has never been studied before.
The monoid B(G) of product-one sequences is a finitely generated C-monoid, which is Krull if and only
if G is abelian (in the Krull case, the class semigroup coincides with the class group which is isomorphic
to G, apart from the exceptional case where |G| = 2). After putting together the needed background in
Section 2, we study the structure of the class semigroup of B(G) in Section 3. Among others we show that
the unit group of the class semigroup is isomorphic to the center of G, and we reveal a subgroup of the
class semigroup which is isomorphic to G/G′ , which is the class group of the complete integral closure of
B(G) (see Theorems 3.8 and 3.10; G′ denotes the commutator subgroup of G). In Section 4 we provide a
complete description of the class semigroup for non-abelian groups of small order. In Section 5 we study
the arithmetic of B(G) and our focus is on sets of lengths. Among others we show that unions of sets of
lengths are finite intervals (Theorem 5.5).
Throughout this paper, let G be a multiplicatively written finite group with identity element 1G ∈ G.
2. Preliminaries
We denote by N the set of positive integers. For integers a, b ∈ Z, [a, b] = {x ∈ Z | a ≤ x ≤ b} means
the discrete interval. For every n ∈ N, Cn denotes a cyclic group of order n. For an element g ∈ G,
ord(g) ∈ N is its order, and for a subset G0 ⊂ G, hG0 i ⊂ G denotes the subgroup generated by G0 . We
will use the following standard notation of group theory:
• Z(G) = {g ∈ G | gx = xg for all x ∈ G} ⊳ G is the center of G,
• [x, y] = xyx−1 y −1 ∈ G is the commutator of the elements x, y ∈ G, and
• G′ = [G, G] = h[x, y] | x, y ∈ Gi ⊳ G is the commutator subgroup of G.
Semigroups. All our semigroups are commutative and have an identity element. Let S be a semigroup.
We denote by S × its group of invertible elements and by E(S) the set of all idempotents of S. E(S) is
endowed with the Rees order ≤, defined by e ≤ f if ef = e. Clearly, ef ≤ e and ef ≤ f for all e, f ∈ E(S).
If E ⊂ E(S) is a finite subsemigroup, then E has a smallest element.
By a monoid, we mean a semigroup which satisfies the cancellation laws. Let H be a monoid. Then
q(H) denotes the quotient group of H and A(H) the set of irreducibles (atoms) of H. The monoid H is
called atomic if every non-unit of H can be written as a finite product of atoms. We say that H is reduced
if H × = {1}, and we denote by Hred = H/H × = {aH × | a ∈ H} the associated reduced monoid of H. A
monoid F is called free abelian with basis P ⊂ F if every a ∈ F has a unique representation of the form
Y
a=
pvp (a) with vp (a) = 0 for almost all p ∈ P .
p∈P
If F is free abelian with basis P , then P is the set of primes of F , we set F = F (P ), and denote by
P
• |a| = p∈P vp (a) the length of a, and by
• supp(a) = {p ∈ P | vp (a) > 0} the support of a.
A monoid F is factorial if and only if Fred is free abelian if and only if F is atomic and every atom is a
prime. We denote by
• H ′ = {x ∈ q(H) | there is an N ∈ N such that xn ∈ H for all n ≥ N } the seminormalization of
H, by
e = {x ∈ q(H) | xN ∈ H for some N ∈ N} the root closure of H, and by
• H
b
• H = {x ∈ q(H) | there is a c ∈ H such that cxn ∈ H for all n ∈ N} the complete integral closure
of H,
e ⊂H
b ⊂ q(H). Then the monoid H is called
and observe that H ⊂ H ′ ⊂ H
PRODUCT-ONE SEQUENCES
3
• seminormal if H = H ′ (equivalently, if x ∈ q(H) and x2 , x3 ∈ H, then x ∈ H),
e
• root closed if H = H,
b
• completely integrally closed if H = H.
If D is a monoid and H ⊂ D a submonoid, then H is a divisor-closed submonoid if a ∈ H, b ∈ D, and
b | a implies that b ∈ H. A monoid homomorphism ϕ : H → D is said to be
• cofinal if for every α ∈ D there is an a ∈ H such that α | ϕ(a).
• a divisor homomorphism if a, b ∈ H and ϕ(a) | ϕ(b) implies that a | b.
• a divisor theory if D is free abelian, ϕ is a divisor homomorphism, and for all α ∈ D there are
a1 , . . . , am ∈ H such that α = gcd(ϕ(a1 ), . . . , ϕ(am )).
• a transfer homomorphism if it satisfies the following two properties :
(T 1) D = ϕ(H)D× and ϕ−1 (D× ) = H × .
(T 2) If u ∈ H, b, c ∈ D and ϕ(u) = bc, then there exist v, w ∈ H such that u = vw,
ϕ(v)D× = bD× , and ϕ(w)D× = cD× .
Transfer Krull monoids. A monoid H is said to be a Krull monoid if it satisfies one of the following
equivalent conditions (see [18, Theorem 2.4.8]):
(a) H is completely integrally closed and satisfies the ACC on divisorial ideals.
(b) There is a divisor homomorphism ϕ : H → F , where F is free abelian monoid.
(c) H has a divisor theory.
A commutative domain R is a Krull domain if and only if its multiplicative monoid of nonzero elements
is Krull. Further examples of Krull monoids can be found in [18, 16]. Let H be a Krull monoid. Then a
divisor theory ϕ : H → F is unique up to isomorphism and
C(H) = q(F )/q ϕ(H)
is called the class group of H.
A monoid H is said to be a transfer Krull monoid if it allows a transfer homomorphism to a Krull
monoid (since in our setting all monoids are commutative, [2, Lemma 2.3.3] shows that the present
definition coincides with the definition in [16]). The significance of transfer homomorphisms θ : H → B
stems from the fact that they allow to pull back arithmetical properties from the (simpler monoid) B to
the monoid H (of original interest). In particular, if H is a transfer Krull monoid, then sets of lengths in
H coincide with sets of lengths in a Krull monoid (see Equation (5.1)).
Since the identity map is a transfer homomorphism, every Krull monoid is a transfer Krull monoid. For
a list of transfer Krull monoids which are not Krull we refer to [16]. To give one such example, consider
a ring of integers O in an algebraic number field K, a central simple algebra A over K, and a classical
maximal O-order R of A. If every stably free left R-ideal is free, then the (non-commutative) semigroup
of cancellative elements of R is a transfer Krull monoid over a finite abelian group ([33]).
Class semigroups and C-monoids. (a detailed presentation can be found in [18, Chapter 2]). Let F
be a monoid and H ⊂ F a submonoid. Two elements y, y ′ ∈ F are called H-equivalent, denote y ∼H y ′ ,
−1
if y −1 H ∩ F = y ′ H ∩ F . H-equivalence is a congruence relation on F . For y ∈ F , let [y]F
H denote the
congruence class of y, and let
×
C(H, F ) = [y]F
and C ∗ (H, F ) = [y]F
H |y ∈ F
H | y ∈ (F \ F ) ∪ {1} .
Then C(H, F ) is a commutative semigroup with unit element [1]F
H (called the class semigroup of H in F )
and C ∗ (H, F ) ⊂ C(H, F ) is a subsemigroup (called the reduced class semigroup of H in F ).
As usual, class groups and class semigroups will both be written additively. The following lemma
describes the relationship between class groups and class semigroups. Its proof is elementary and can be
found in [18, Propositions 2.8.7 and 2.8.8]
Lemma 2.1. Let F be a monoid and H ⊂ F be a submonoid.
4
JUN SEOK OH
1. If H ⊂ F is cofinal, then the map θ : C(H, F ) → q(F )/q(H), [y]F
H 7→ yq(H) for all y ∈ F , is an
epimorphism. Moreover, θ is an isomorphism if and only if H ֒→ F is a divisor homomorphism.
2. If C(H, F ) is a group, then H ⊂ F is cofinal, and if C(H, F ) is a torsion group, then H ֒→ F is a
divisor homomorphism.
A monoid H is called a C-monoid if H is a submonoid of a factorial monoid F such that H ∩ F × = H ×
and C ∗ (H, F ) is finite. A commutative ring R is a C-ring if its monoid of regular elements is a C-monoid.
A Krull monoid is a C-monoid if and only if it has finite class group. We refer to [18, 30, 21, 28] for
b
more on C-monoids. To give an explicit example, consider a Mori ring R. If the conductor f = (R : R)
b is finite, then R is a C-ring. We will need the following lemma (for a proof see [18,
is nonzero and R/f
Theorem 2.9.11]).
Lemma 2.2. Let F = F × × F (P ) be factorial and H ⊂ F be a C-monoid. Suppose that vp (H) ⊂ N0
is a numerical monoid for all p ∈ P (this is a minimality condition on F which can always be assumed
without restriction).
b is a Krull monoid with nonempty conductor (H : H)
b and finite class group C(H).
b
1. H
b → F (P ), defined by
2. The map ∂ : H
∂(a) =
Y
pvp (a) ,
p∈P
b
is a divisor theory, and there is an epimorphism C ∗ (H, F ) → C(H).
3. Algebraic Properties of the monoid of product-one sequences
We introduce sequences over the finite group G. Our notation and terminology follows the articles
[17, 24, 9]. Let G0 ⊂ G be a subset. The elements of the free abelian monoid F (G0 ) will be called
sequences over G0 . This terminology goes back to Arithmetic Combinatorics. Indeed, a sequence over
G0 can be viewed as a finite unordered sequence of terms from G0 , where the repetition of elements is
allowed. In order to avoid confusion between the multiplication in G and multiplication in F (G0 ), we
denote multiplication in F (G0 ) by the boldsymbol · and we use brackets for all exponentiation in F (G0 ).
In particular, a sequence S ∈ F (G0 ) has the form
(3.1)
S = g1 · . . . · gℓ =
• gi =
i∈[1,ℓ]
• g [vg (S)] ∈ F (G0 ),
g∈G0
where g1 , . . . , gℓ ∈ G0 are the terms of S. Moreover, if S1 , S2 ∈ F (G0 ) and g1 , g2 ∈ G0 , then S1 ·S2 ∈ F (G0 )
has length |S1 |+ |S2 |, S1 ·g1 ∈ F (G0 ) has length |S1 |+ 1, g1 g2 ∈ G is an element of G, but g1 ·g2 ∈ F (G0 )
is a sequence of length 2. If g ∈ G0 , T ∈ F (G0 ), and k ∈ N0 , then
. . · T} ∈ F (G0 ) .
g [k] = g · . . . · g ∈ F (G0 ) and T [k] = |T · .{z
| {z }
k
k
Let S ∈ F (G0 ) be a sequence as in (3.1). Then we denote by
π(S) = {gτ (1) . . . gτ (ℓ) ∈ G | τ a permutation of [1, ℓ]} ⊂ G
and
Π(S) =
[
π(T ) ⊂ G ,
T |S
|T |≥1
the set of products and subsequence products of S, and it can easily be seen that π(S) is contained in
a G′ -coset. Note that |S| = 0 if and only if S = 1F (G0 ) , and in that case we use the convention that
π(S) = {1G }. The sequence S is called
• a product-one sequence if 1G ∈ π(S), and
• product-one free if 1G ∈
/ Π(S).
PRODUCT-ONE SEQUENCES
5
Definition 3.1. Let G0 ⊂ G be a subset.
1. The submonoid
B(G0 ) = {S ∈ F (G0 ) | 1G ∈ π(S)} ⊂ F (G0 )
is called the monoid of product-one sequences, and A(G0 ) = A B(G0 ) denotes its set of atoms.
2. We call
D(G0 ) = sup{|S| | S ∈ A(G0 )} ∈ N ∪ {∞}
the large Davenport constant of G0 and
d(G0 ) = sup{|S| | S ∈ F (G0 ) is product-one free} ∈ N0 ∪ {∞}
the small Davenport constant of G0 .
Note that obviously the set B(G0 ) is a multiplicatively closed subset of the commutative cancellative
semigroup F (G0 ) whence B(G0 ) is indeed a monoid. Our object of interest is the monoid B(G) but for
technical reasons we also need the submonoids B(G0 ) for subsets G0 ⊂ G. The next lemma gathers some
elementary properties. Its simple proof can be found in [9, Lemma 3.1].
Lemma 3.2. Let G0 ⊂ G be a subset.
1. B(G0 ) ⊂ F (G0 ) is a reduced finitely generated submonoid, A(G0 ) is finite, and D(G0 ) ≤ |G|.
2. Let S ∈ F (G) be product-one free.
(a) If g0 ∈ π(S), then g0−1 · S ∈ A(G). In particular, d(G) + 1 ≤ D(G).
(b) If |S| = d(G), then Π(S) = G \ {1G } and hence d(G) = max{|S| : S ∈ F (G) with Π(S) =
G \ {1G}}.
3. If G is cyclic, then d(G) + 1 = D(G) = |G|.
The Davenport constant of finite abelian groups is a frequently studied invariant in Zero-sum Theory
(for an overview see [18] and for recent progress [31, 5]). For the Davenport constant of nonabelian groups
we refer to [17, 24, 10].
Lemma 3.3. Let G0 ⊂ G be a subset. A submonoid H ⊂ B(G0 ) is divisor-closed if and only if H = B(G1 )
for a subset G1 ⊂ G0 .
Proof. If G1 ⊂ G0 , then clearly B(G1 ) ⊂ B(G0 ) is a divisor-closed submonoid. Conversely, let H ⊂ B(G0 )
be a divisor-closed submonoid. We set
[
G1 =
supp(B)
B∈H
and obtain that H ⊂ B(G1 ). We have to show that equality holds. Let S = g1 · . . . · gℓ ∈ B(G1 ). Then
for every i ∈ [1, ℓ] there is some Bi ∈ H such that gi ∈ supp(Bi ). Then B1 · . . . · Bℓ ∈ H and S divides
B1 · . . . · Bℓ , which implies that S ∈ H.
Proposition 3.4. The following statements are equivalent :
(a) G is abelian.
(b) B(G) is a Krull monoid.
(c) B(G) is a transfer Krull monoid.
Proof. The implications (a) ⇒ (b) and (b) ⇒ (c) are well-known and immediate. Indeed, if G is abelian,
then the embedding B(G) ֒→ F (G) is a divisor homomorphism whence B(G) is Krull. By definition, every
Krull monoid is a transfer Krull monoid.
(c) ⇒ (a) Assume to the contrary that G is not abelian but B(G) is a transfer Krull monoid. Thus
there is a transfer homomorphism θ1 : B(G) → H, where H is a Krull monoid. Since there is a transfer
homomorphism θ2 : H → B(G0 ), where G0 ⊂ C(H) is a subset of the class group, and since the composition
6
JUN SEOK OH
of transfer homomorphisms is a transfer homomorphism, we obtain a transfer homomorphism θ : B(G) →
B(G0 ), where G0 is a subset of an abelian group.
Since G is non-abelian, there exist g, h ∈ G such that gh 6= hg. Consider the sequence
U = g · h · g −1 · (gh−1 g −1 ) ∈ A(G).
From [18, Proposition 3.2.3], we have that θ(U ) ∈ A(G0 ), say θ(U ) = a1 · . . . · aℓ , where ai ∈ G0 for all
i ∈ [1, ℓ]. Let m = ord (hgh−1 g −1 ) ∈ N. Then
U [m] = (g · g −1 )[m] · (h · (gh−1 g −1 ))[m] ,
and hence it follows that
(a1 · . . . · aℓ )[m] = θ(U [m] )
[m]
= θ(g · g −1 )
· θ (h · (gh−1 g −1 ))[m] .
[m]
Since B(G0 ) ֒→ F (G0 ) is a divisor homomorphism (and so B(G0 ) is root closed), the fact that θ(g·g −1 )
divides (a1 ·. . .·aℓ )[m] in B(G0 ) implies that θ(g ·g −1 ) divides a1 ·. . .·aℓ in B(G0 ). Since θ(g ·g−1 ) and a1 ·
. . . · aℓ ∈ A(G0 ), it follows that θ(g · g −1 ) = a1 · . . . · aℓ = θ(U ). Thus θ (h · (gh−1 g −1 ))[m] = 1F (G0 ) , a
contradiction.
Clearly, if |G| ≤ 2, then B(G) is factorial whence it is both a Krull monoid (with trivial class group)
and a C-monoid (with trivial class semigroup). In order to avoid trivial case distinctions, we exclude
this case whenever convenient. By Proposition 3.4, B(G) is not Krull when G is not abelian. The next
proposition reveals that B(G) is a C-monoid in all cases and determines the class group of its complete
integral closure.
Proposition 3.5. Suppose that |G| ≥ 3.
1. If G0 ⊂ G is a subset, then B(G0 ) ⊂ F (G0 ) is cofinal and B(G0 ) is a C-monoid.
[ ֒→ F (G) is a divisor theory and the map
2. The embedding B(G)
[ = q F (G) /q B(G)
Φ : C(B(G))
−→ G/G′
gG′ ,
where S ∈ F (G) and g ∈ π(S), is an isomorphism. Clearly, vg B(G) = N0 for all g ∈ G.
3. There is an epimorphism C B(G), F (G) → G/G′ .
Sq(B(G))
7−→
Proof. 1. and 2. Let G0 ⊂ G. If S = g1 · . . . · gℓ ∈ F (G0 ) and g ∈ π(S), then S ′ = g −1 · S ∈ B(G0 ), S | S ′ ,
and hence B(G0 ) ⊂ F (G0 ) is cofinal. Let k ∈ N0 . If g ∈ G with ord(g) > 2, then U = g · g −1 ∈ B(G),
vg (U [k] ) = k. If g = 1G , then vg (g [k] ) = k. If ord(g) = 2, then there is an h ∈ G \ {1G , g} and
U = g · h · (gh)−1 ∈ B(G) and vg (U [k] ) = k. Thus in all cases we have vg (B(G)) = N0 .
[
It is easy to check that Φ is an isomorphism (details can be found in [9, Theorem 3.2]). Since C(B(G))
\
is finite, C(B(G
0 )) is finite and since B(G0 ) is finitely generated, it is a C-monoid by [19, Proposition 4.8].
3. This follows from 2. and from Lemma 2.2.2.
We start with the study of the class semigroup and recall that, by definition, for two sequences S, S ′ ∈
F (G) the following statements are equivalent:
• S ∼B(G) S ′ .
• For all T ∈ F (G), we have S · T ∈ B(G) if and only if S ′ · T ∈ B(G).
• For all T ∈ F (G), we have 1G ∈ π(S · T ) if and only if 1G ∈ π(S ′ · T ).
If S = g1 · . . . · gℓ ∈ B(G) such that 1G = g1 . . . gℓ , then 1G = gi . . . gℓ g1 . . . gi−1 for every i ∈ [1, ℓ]. We will
F (G)
use this simple fact without further mention. Moreover, ∼ means ∼B(G) and we write [S] = [S]B(G) .
PRODUCT-ONE SEQUENCES
7
Lemma 3.6. Let S ∈ F (G).
1. If S ′ ∈ F (G) such that S ∼ S ′ , then π(S) = π(S ′ ).
2. Let S ′ ∈ F (G). In the following cases, S ∼ S ′ if and only if π(S) = π(S ′ ) :
(a) S and S ′ are sequences over the center of G.
(b) There is g ∈ π(S) such that π(S) = gG′ .
3. If g, h ∈ G with g 6= h, then g ≁ h. In particular, |G| ≤ |C B(G), F (G) |.
4. If g ∈ Z(G) and h ∈ G, then g · h ∼ gh.
5. |π(S)| = 1 if and only if hsupp(S)i is abelian.
Proof. 1. Suppose that S ′ ∈ F (G) with S ∼ S ′ . Then for every g ∈ G we obtain that
g ∈ π(S)
⇔
g −1 · S ∈ B(G)
⇔
g −1 · S ′ ∈ B(G)
⇔
g ∈ π(S ′ ) ,
which implies that π(S) = π(S ′ ).
2. From 1., it remains to verify the reverse implication. Suppose that π(S) = π(S ′ ).
(a) Since S, S ′ ∈ F (Z(G)), we have |π(S)| = |π(S ′ )| = 1, say π(S) = π(S ′ ) = {g}. Then for every
T ∈ F (G) we have
π(S · T ) = gπ(T )
and
π(S ′ · T ) = gπ(T ) ,
which implies that S ∼ S ′ .
(b) Suppose that there is a g ∈ π(S) such that π(S) = gG′ . Let T ∈ F (G) such that T · S ∈ B(G).
Then we have
π(T )π(S) ⊂ π(T · S) ⊂ G′ .
′
Since π(S) = gG , there are t ∈ π(T ) and e ∈ G′ such that tg = e. Hence we obtain that
1G = ee−1 = tge−1 ∈ π(T )π(S) = π(T )π(S ′ ) ⊂ π(T · S ′ ).
It follows that T · S ′ ∈ B(G). By symmetry, we infer that S ∼ S ′ .
3. Let g, h ∈ G. Then g, h ∈ F (G) and if g ∼ h, then 1. implies that {g} = π(g) = π(h) = {h}.
Therefore, g 6= h implies that g ≁ h, and hence we obtain that |G| ≤ |C B(G), F (G) |.
4. This follows from the fact that g ∈ Z(G).
5. Obvious.
Lemma 3.7. Let S ∈ F (G).
1. If [S] is an idempotent of C B(G), F (G) , then π(S) ⊂ G′ is a subgroup.
2. [1F (G)] = B Z(G) . In particular, if g ∈ Z(G), then g [ord(g)] ∼ 1F (G) .
3. [S] is the smallest idempotent in E C B(G), F (G) (with respect to the Rees order) if and only if
π(S) = G′ .
Proof. 1. Suppose that [S] is an idempotent of C B(G), F (G) . Then [S] = S [2] whence S ∼ S [2] , and
Lemma 3.6.1 implies that
π(S)π(S) ⊂ π S [2] = π(S) .
Thus π(S) ⊂ G′ is a subgroup.
2. Suppose that S ∈ B Z(G) . Then π(S) = {1G }, and for all T ∈ F (G) we have
π(S · T ) = π(T ) = π(1F (G) · T ) ,
whence S ∼ 1F (G) .
Conversely, suppose that S ∼ 1F (G) . Then S ∈ B(G), and we set S = g1 ·. . .·gℓ such that g1 . . . gℓ = 1G .
Assume to the contrary that there is an i ∈ [1, ℓ] such that gi ∈
/ Z(G), say i = 1. Then there is an element
h ∈ G such that hg1 6= g1 h. Then T = (hg1 ) · (h−1 g1−1 ) ∈ F (G) \ B(G), but
8
JUN SEOK OH
1G = g1 (hg1 )(g2 . . . gℓ )(h−1 g1−1 ) ∈ π(T · S) .
Since S ∼ 1F (G), we infer that 1G ∈ π(T · 1F (G)) = π(T ), a contradiction.
In particular, if g ∈ Z(G), then g [ord(g)] ∈ B Z(G) and hence g [ord(g)] ∼ 1F (G) .
3. First, we suppose that π(S) = G′ . Then [S] is an idempotent
by Lemma 3.6.2, and it remains
to verify that [S] is the smallest idempotent of C B(G), F (G) . Let S ′ ∈ F (G) such that [S ′ ] is an
idempotent. We have to show that S ∼ S · S ′ . Since π(S ′ ) ⊂ G′ is a subgroup by 1., S, S ′ ∈ B(G), and
since π(S · S ′ ) is a G′ -coset, it follows that
G′ ⊂ π(S)π(S ′ ) ⊂ π(S · S ′ ) ⊂ G′
whence
G′ = π(S · S ′ ) .
Again Lemma 3.6.2 implies that S ∼ S · S ′ .
To show the converse, suppose that [S] is the smallest idempotent. We construct an element S ′ ∈ F (G)
such that π(S ′ ) = G′ . Then the proof above shows that [S ′ ] is the smallest idempotent whence S ∼ S ′
and π(S) = π(S ′ ) = G′ . We set G′ = {g1 , . . . , gn }, and for each i ∈ [1, n] we set
gi =
ki
Y
−1 −1
ai,ν bi,ν ai,ν
bi,ν ,
where ki ∈ N and all
ai,ν , bi,ν ∈ G ,
ν=1
and define
Si =
−1
(ai,ν · bi,ν · a−1
i,ν · bi,ν ) ∈ B(G) .
•
ν∈[1,ki ]
Then obviously π(S1 · . . . · Sn ) = G′ .
Theorem 3.8.
1. Suppose that G/G′ = {g0 G′ , . . . , gk G′ }, and for each i ∈ [0, k] let Si ∈ F (G) such that π(Si ) = gi G′ .
Then the map
G/G′ →
[Si ] | i ∈ [0, k] = C ⊂ C B(G), F (G)
gi G′
7→
[Si ]
is a group isomorphism. Thus C B(G), F (G) contains a subgroup isomorphic to the class group
[ Moreover, for any i ∈ [0, k] and for any S ∈ F (G), we have [S · Si ] ∈ C.
of B(G).
2. The map
Z(G)
→
g
7→
×
C B(G), F (G)
[g]
is a group isomorphism.
Proof. 1. We first verify
the
existence
of
such
sequences
S
,
.
.
.
,
S
.
Since
C
B(G),
F
(G)
is finite, the
0
k
set E C B(G), F (G) has the smallest element, say [S]. For each i ∈ [0, k], we set Si = gi · S. Since
π(S) = G′ by Lemma 3.7.3, it follows that for each i ∈ [0, k]
gi G′ ⊂ π(gi · S) ⊂ gi G′ , whence π(Si ) = π(gi · S) = gi G′ .
We now use Lemma 3.6.2. If i, j ∈ [0, k], then gi gj ∈ π(Si )π(Sj ) ⊂ π(Si · Sj ), and hence π(Si · Sj ) =
gi gj G′ = gℓ G′ = π(Sℓ ) for some ℓ ∈ [0, k]. Thus it follows that C is a subgroup of the class semigroup and
[
the given map is an isomorphism. By Proposition 3.5.2, G/G′ is isomorphic to the class group of B(G).
Moreover, let S ∈ F (G), g ∈ π(S), and i ∈ [0, k]. Then
ggi G′ = gπ(Si ) ⊂ π(S)π(Si ) ⊂ π(S · Si ) ,
whence π(S · Si ) = ggi G′ = gj G′ = π(Sj ) for some j ∈ [0, k]. Again by Lemma 3.6.2, we have S · Si ∼ Sj ,
and thus [S · Si ] ∈ C.
PRODUCT-ONE SEQUENCES
9
2. Let S ∈ F (G) such that [S] ∈ C F (G), B(G) is invertible. Then there is an S ′ ∈ F (G) such that
0C(F (G),B(G)) = [1F (G) ] = [S] + [S ′ ] = [S · S ′ ] ,
whence S · S ′ ∼ 1F (G) . Then Lemma 3.7.2 implies that S · S ′ ∈ B Z(G) whence S, S ′ ∈ F Z(G) . If
g ∈ π(S), then Lemma 3.6.2 implies that S ∼ g. This proves that the given map is well-defined and
surjective. Lemma 3.6 (items 3 and 4) shows that the map is a monomorphism.
Our next goal is a detailed investigation of the case where |G′ | = 2. We derive a couple of special
properties which do not hold in general (Theorem 3.10 and Remark 3.11).
Lemma 3.9. Suppose that |G′ | = 2, and let g ∈ G with ord(g) = n.
1. We have g ∼ g [n+1] , and hence g [n] ∈ C B(G), F (G) is idempotent.
2. If k ∈ [1, n] is odd and g k ∈
/ Z(G), then g [k] ∼ g k .
3. If g, h ∈ G \ Z(G) with gh = hg, then g · h ≁ gh provided that one of the following conditions holds :
(a) gh ∈ Z(G).
(b) There is g0 ∈ G such that g0 g 6= gg0 and g0 h 6= hg0 .
Proof. 1. Let T ∈ F (G). Since g n = 1G , 1G ∈ π(g · T ) implies that 1G ∈ π g [n+1] · T . Conversely,
suppose that 1G ∈ π g [n+1] · T . If every element in supp(T ) commutes with g, then
π g [n+1] · T = π(g · T ) and hence 1G ∈ π(g · T ) .
Now suppose that there is at least one h ∈ supp(T ) such that hg 6= gh. Then π(g · T ) has at least two
elements. Since |G′ | = 2 and
π(g · T ) ⊂ π g [n+1] · T ⊂ G′ ,
we obtain that π(g · T ) = G′ and hence 1G ∈ π(g · T ). Thus [g] = g [n+1] and thus
[n] [n] [2n] [n+1] [n−1] [n]
g
+ g
= g
= g
·g
= g
.
2. Let k ∈ [1, n] be odd, g k ∈
/ Z(G), and T ∈ F (G). If 1G ∈ π g k · T , then 1G ∈ π g [k] · T . Conversely,
suppose that 1G ∈ π T · g [k] . If hg = gh for all h ∈ supp(T ), then obviously 1G ∈ π g k · T .
Suppose there is an element h ∈ supp(T ) such that hg 6= gh. Then π T ·g [k] = G′ . We set T = h1 ·. . .·hℓ
with h1 = h and consider the elements
g 0 = h1 g k h2 . . . hℓ
and
(1)
g0 = gh1 g k−1 h2 . . . hℓ in G′ .
(1)
(1)
Then G′ = g0 , g0 . If g0 = 1, then we are done. Suppose that g0 = 1G . Then
(2)
(1)
g0 = ggh1 g k−2 h2 . . . hℓ 6= g0
(2)
whence g0 = g0 , and hg 2 = g 2 h. Thus we obtain
(1)
(3)
(k)
1G = g0 = g0 = . . . = g0
= g k h1 . . . hℓ ∈ π T · g k .
3. Let g, h ∈ G \ Z(G) with gh = hg.
(a) Suppose that gh ∈ Z(G) and assume to the contrary that g · h ∼ gh. By Lemma 3.7.2, we infer that
(g · h)[ord(gh)] ∼ gh[ord(gh)] ∼ 1F (G) .
a contradiction to [1F (G) ] = B Z(G) .
(b) Let g0 ∈ G such that g0 g 6= gg0 and g0 h 6= hg0 . If
T = (g −1 g0 ) · (h−1 g0−1 ) ∈ F (G) ,
then 1G ∈
/ π T · (gh) but 1G = g(g −1 g0 )h(h−1 g0−1 ) ∈ π(T · g · h). This shows that g · h ≁ gh.
10
JUN SEOK OH
It is well-known that a finitely generated monoid is Krull if and only if it is root-closed. Thus if G is
non-abelian, then B(G) is not root-closed by Proposition 3.4. The next result shows that it is seminormal
if |G′ | = 2. An element c of a semigroup is called a regular if c lies in a subgroup of the semigroup, and
the semigroup is called Clifford if every element is regular.
Theorem 3.10. Suppose that |G′ | = 2.
1. B(G) is seminormal.
2. C B(G), F (G) is a Clifford semigroup. In particular, if S ∈ F (G) with π(S) = {g}, then [S]
generates a cyclic subgroup of order ord(g).
Q
3. |C B(G), F (G) | ≤ |Z(G)| + g∈G\Z(G) ord(g).
Proof. 1. Let S ∈ q B(G) such that S [2] , S [3] ∈ B(G). Since S ∈ q F (G) with S [2] , S [3] ∈ F (G), we
have that S ∈ F (G). Let g ∈ π(S). Since S [2] , S [3] ∈ B(G), it follows that
g 2 ∈ π S [2] ⊂ G′ and g 3 ∈ π S [3] ⊂ G′ .
Since G′ is a subgroup of G, we obtain g ∈ G′ . If g = 1G , then we are done. Suppose that g 6= 1G . Then
G′ = {1G , g}. If each two elements in supp(S) would commute, then
π S [3] = {g 3 } ,
π(S) = {g}, π S [2] = {g 2 }, and
which implies g 2 = 1G = g 3 and g = 1G , a contradiction. Thus supp(S) contains two elements which do
not commute, say S = g1 · . . . · gℓ , g1 g2 6= g2 g1 , and g = g1 . . . gℓ . Then 1G = g2 g1 g3 . . . gℓ ∈ π(S).
2. Let S ∈ F (G). If π(S) has two elements, then [S] lies in the group given in Theorem 3.8.1. Suppose
that π(S) has onlynone element, say
π(S) = {g} and ord(g) = n. We assert that S ∼ S [n+1] . Suppose this
[n] o
is a cyclic subgroup of the class semigroup, and hence the assertion
holds true. Then [S], . . . , S
[m]
follows. Let m ∈ [1, n] and S
the identity element of the subgroup. Then it is an idempotent of the
class semigroup, and Lemma 3.7.1 shows that π S [m] = {g m } ⊂ G′ is a subgroup. This implies m = n.
Thus it remains to show that S ∼ S [n+1] . To do so, let T ∈ F (G) be given. Since S [n] ∈ B(G),
1G ∈ π(T · S) implies that 1G ∈ π T · S [n+1] .
Conversely, suppose that 1G ∈ π T · S [n+1] . If every element of supp(T ) commutes with every element of
supp(S), then π T · S [n+1] = π(T · S) and thus 1G ∈ π(T · S). If there are t ∈ supp(T ) and s ∈ supp(S)
such that ts 6= st, then |π(T · S)| ≥ 2 and hence 1G ∈ π(T · S).
3. We set G = {g1 , . . . , gn } and Z(G) = {g1 , . . . , gk } for some k ∈ [1, n]. Let S ∈ F (G). Then we can
write S of the form
[k ord(g1 )+r1 ]
S = g1 1
· . . . · gn[kn ord(gn )+rn ] ,
where k1 , . . . , kn ∈ N0 and ri ∈ [1, ord(gi )] for each i ∈ [1, n]. By Lemma 3.9.1, we have
[r ]
S ∼ g1 1 · . . . · gn[rn ] .
[r ]
[r ]
By Lemma 3.6.4., g1 1 · . . . · gk k ∼ g0 , where g0 = g1r1 . . . gkrk ∈ Z(G). Thus
[r
]
k+1
S ∼ g0 · gk+1
· . . . · gn[rn ] ,
and hence the assertion follows.
Remark 3.11.
1. Suppose that |G′ | = 2. Let g, h ∈ G be distinct with ord(g) = n and ord(h) = m. By Theorem
3.10.2, [g] ⊂ C B(G), F (G) generates a cyclic subgroup of order n and [h] generates a cyclic subgroup of
order m.
PRODUCT-ONE SEQUENCES
11
Suppose that h[g]i ∩ h[h]i =
6 ∅. Then
there are i ∈ [1, n] and j ∈ [1, m] such that g [i] ∼ h[j] . Let
m
j
[m]
mj = ord(h ) = gcd(j,m) . Since h
is an idempotent, we have
g [imj ] ∼ h[m] .
By Lemma 3.6.1, imj = kn for some k ∈ N. Since g [n] is also idempotent, we have
g [n] ∼ h[m] .
Again by Lemma 3.6.1, we obtain gh = hg because B(G)-equivalence is a congruence relation on F (G).
It follows that if gh 6= hg, then h[g]i ∩ h[h]i = ∅, and for each i ∈ [1, n] and j ∈ [1, m],
[i] [j] [i] [j]
g + h
= g ·h
∈C,
where C is the group given in Theorem 3.8.1.
2. Suppose that B(G) is seminormal and let G0 ⊂ G be a subset. Let S ∈ q B(G0 ) such that
S [2] , S [3] ∈ B(G0 ). Since B(G) is seminormal, it follows that S ∈ B(G) and hence S ∈ B(G) ∩ q B(G0 ) =
B(G0 ). Thus B(G0 ) is seminormal.
3. Let D2n = ha, bi = {1G , a, . . . , an−1 , b, ab, . . . , an−1 b} be the dihedral group with n ≡ 3 mod 4,
where ord(a) = n, ord(b) = 2, and ak baℓ b = ak−ℓ for all k, ℓ ∈ Z. Then
n−1
S = a[ 2 ] · b[2] ∈ q B(D2n ) \ B(D2n ), but S [2] , S [3] ∈ B(D2n ).
Thus if G contains D2n as a subgroup, then 2. shows that B(G) is not seminormal.
4. Examples of Class Semigroups for Non-abelian Groups of small order
In this section, we discuss the three smallest non-abelian groups and provide a complete description
of the class semigroup. Among others we will see that the monoid of product-one sequences over the
dihedral group with 6 elements is not seminormal and the associated class semigroup is not Clifford.
Example 4.1. Let G = Q8 = {E, I, J, K, −E, −I, −J, −K} be the quaternion group with the relations
IJ = −JI = K,
′
JK = −KJ = I,
KI = −IK = J,
and
IJK = −E .
′
Then Z(G) = G = {E, −E} and G/G ≃ C2 ⊕ C2 . Furthermore, d(G) = 4 and D(G) = 6 by [17,
Theorem 1.1].
Let S ∈ F (G). If |π(S)| = 2, then, by Theorem 3.8.1, S is B(G)-equivalent to an element in the group
C which is isomorphic to G/G′ . We only consider
the case |π(S)| = 1. By Lemma 3.6.5, hsupp(S)i is an
abelian subgroup of G. Suppose that S ∈ F hIi . Then S has of the form
S = E [n1 ] · I [n2 ] · (−E)[n3 ] · (−I)[n4 ] ,
where n1 , . . . , n4 ∈ N0 . By Lemma 3.6.4 and Lemma 3.9 (items 1 and 2), we have
S ∼ I [n] for some n ∈ [1, 4] .
By symmetry,
we
obtain
the
same
results
in
the
case
S
∈
F
hJi
and
S
∈
F
hKi
. Moreover, If
S ∈ F h−Ei , then, by Theorem
3.8.2, S is B(G)-equivalent to an element in the group of units of the
class semigroup C B(G), F (G) which is isomorphic to Z(G). It follows that
|C B(G), F (G) | = 18 .
Figure 1 presents the subgroup lattice of the class
semigroup. Note that the subgroup lattice of G is
not preserved in the class semigroup C B(G), F (G) . Furthermore, observe that the bottom elements in
the lattice are all idempotent elements of the class semigroup.
12
JUN SEOK OH
C B(G), F (G)
h[I · J], [I · K]i
h[I · K]i
h[J · K]i
h[I · J · K]i
h[I · J]i
h[K]i
h[J]i
h[I]i
h[−E]i
h K [2] i
h J [2] i
h I [2] i
h[1F (G) ]i
h K [4] i
h J [4] i
h I [4] i
Figure 1. Subgroup Lattice of the Class Semigroup over Q8
Example 4.2. Let G = D8 = ha, b | a4 = b2 = 1G and ba = a3 bi = {1G , a, a2 , a3 , b, ab, a2 b, a3 b} be
the dihedral group of order 8. Then Z(G) = G′ = ha2 i = {1G , a2 } and G/G′ ≃ C2 ⊕ C2 . Furthermore,
d(G) = 4 and D(G) = 6 by [17, Theorem 1.1].
All arguments run along the same lines as the ones given in the previous example. However, in this
case, there are two elements g, h ∈ G \ Z(G) such that gh = hg. Consider the sequence b · a2 b ∈ F (G)
having π(b · a2 b) = {a2 }, and suppose that b · a2 b ∼ S for some S ∈ F (G). By Lemma 3.6 (items 1 and
5), we have π(S) = {a2 } and hence hsupp(S)i ⊂ G is abelian subgroup containing a2 . It follows that
hsupp(S)i is one of hai, ha2 i, ha2 , bi and ha2 , abi.
CASE 1. S ∈ F ha2 i .
Then S = a2 is only possible choice, but it never happen by Lemma 3.9.3.
CASE 2. S ∈ F hai .
Then S = a[2] is only possible choice by Lemma 3.6.4 and Lemma 3.9 (items 1 and 2). Since B(G)equivalence is a congruence relation on F (G),
b · a2 b ∼ a[2] implies that b · a2 b · a ∼ a[3] .
But {a3 } = π(a[3] ) = π(b · a2 b · a) = {a, a3 }, a contradiction.
In the case that S ∈ F ha2 , abi , we can obtain a contradiction by the same argument of CASE 2. It
follows that S = b · a2 b, and by Lemma 3.6.4,
S [2] ∼ b[2] is an idempotent element of the class semigroup C B(G), F (G)
with the relation b[2] ∼ (a2 b)[2] ∼ S [2] .
For the sequence T = ab · a3 b ∈ F (G), we can obtain the same result by above argument. Thus T [2]
is an idempotent element of the class semigroup C B(G), F (G) with the relation (ab)[2] ∼ (a3 b)[2] ∼ T [2].
It follows that
|C B(G), F (G) | = 18 .
Figure 2 presents the subgroup lattice of the class semigroup. Note that, as in
the previous example,
the subgroup lattice of G is not preserved in the class semigroup C( B(G), F (G) . Furthermore, observe
that the bottom elements in the lattice are all idempotent elements of the class semigroup.
×
Remark 4.3. In general, the group of units C B(G), F (G) of the class semigroup is not a direct factor.
For example, let G be the group described in Example 4.1. Assume to the contrary that C B(G), F (G)
has a decomposition of the form
×
C B(G), F (G) = C0 × C B(G), F (G) ,
where C0 is a subsemigroup having 9 elements. Figure 1. shows that there are three elements x ∈
C B(G), F (G) such that x 6= x3 . It follows that there exist three elements satisfying such property in the
PRODUCT-ONE SEQUENCES
13
C B(G), F (G)
h[ab], a3 b i
h[b], a2 b i
h a[3] · b[2] , a[2] · b[3] i
h[a]i
h a[3] · b[2] i h a[2] · b[3] i h a[2] · b[2] · ab i h b · a2 b i h a2 b i h[b]i h a2 i h ab · a3 b i h a3 b i h[ab]i h a[2] i
h a[2] · b[2] i
h b[2] i
h (ab)[2] i
h[1F (G) ]i
h a[4] i
Figure 2. Subgroup Lattice of the Class Semigroup over D8
×
decomposition. Since |C B(G), F (G) | = 2, C0 has at least two such elements, whence |C0 | > 9. Indeed,
if [I] and [J] are in C0 , then [I · J] = [I] + [J] ∈ C0 and thus we obtain that C0 has at least 10 elements.
The similar argument works for the group described in Example 4.2.
Example 4.4. Let G = D6 = ha, b | a3 = b2 = 1G and ba = a2 bi = {1G , a, a2 , b, ab, a2 b} be the dihedral
group of order 6. Then Z(G) = {1G }, G′ = hai = {1G, a, a2 }, and G/G′ ≃ C2 . Furthermore, d(G) = 3
and D(G) = 6 by [17, Theorem 1.1].
Let S ∈ F (G). If |π(S)| = 3, then, by Theorem 3.8.1, we obtain
S ∼ a[2] · b[2]
S ∼ a · a2 · b ,
or
where π(a[2] · b[2] ) = {1G, a, a2 } and π(a · a2 · b) = {b, ab, a2 b}.
If |π(S)| = 2, then S is B(G)-equivalent with one of the following sequences :
(4.1)
a · b[n] ,
b[n] · ab,
a · (ab)[n] ,
b · (ab)[n] ,
a · (a2 b)[n] ,
b[n] · a2 b,
b · (a2 b)[n] ,
where n ∈ N.
We now start with the following claims.
CLAIM.A : b[2] ∼ b[4] , (ab)[2] ∼ (ab)[4] ,
(a2 )[2] ∼ a[4] ,
a2 · b[n] ,
a2 · (ab)[n] ,
a2 · (a2 b)[n] ,
(ab)[n] · a2 b,
ab · (a2 b)[n] ,
(a2 b)[2] ∼ (a2 b)[4] ,
(a2 )[3] ∼ a[3] ,
(a2 )[4] ∼ a[2] ,
a[2] ∼ a[5] ,
(a2 )[2] ∼ (a2 )[5] ,
a · a2 ∼ a[3] ,
a · b ∼ a2 · b ∼ a · b[3] ∼ b[2] · ab ∼ b[2] · a2 b ,
a · ab ∼ a2 · ab ∼ a · (ab)[3] ∼ ab[2] · a2 b ∼ ab[2] · b ,
a · a2 b ∼ a2 · a2 b ∼ a · (a2 b)[3] ∼ (a2 b)[2] · ab ∼ (a2 b)[2] · b .
CLAIM.B :
b ≁ b[3] ,
ab ≁ (ab)[3] ,
b · ab ≁ b · a2 b,
a2 b ≁ (a2 b)[3] ,
ab · a2 b,
a · b[2] ,
b · a2 b ≁ ab · a2 b,
a · b[2] ,
ab · a2 b ≁ a · b[2] ,
a · (ab)[2] ,
a ≁ a[4] ,
a · (ab)[2] ,
a · (ab)[2] ,
a[2] ≁ a2 ,
a · (a2 b)[2] ,
a · (a2 b)[2] ,
a · (a2 b)[2] .
Suppose the Claims hold true. Then we obtain that
|C B(G), F (G) | = 26
and
C B(G), F (G) = G1 ∪ G2 ∪ G3 ∪ G4 ,
where
n
o
• G1 = [b], b[2] , b[3] , [ab], (ab)[2] , (ab)[3] , [a2 b], (a2 b)[2] , (a2 b)[3] , [a], a[2] , a[3] , a[4] , a2 ,
14
JUN SEOK OH
n
o
• G2 = [a · b], [a · ab], a · a2 b , a · b[2] , a · (ab)[2] , a · (a2 b)[2] , [b · ab], b · a2 b , ab · a2 b ,
n
o
• G3 = a[2] · b[2] , a · a2 · b (that is isomprphic to G/G′ ), and
n
o
• G4 = [1F (G) ] that is isomorphic to Z(G) .
Figure 3 presents the subgroup lattice of the class semigroup. Note that, as in
the previous example,
the subgroup lattice of G is not preserved in the class semigroup C B(G), F (G) . Furthermore, observe
that the elements in the set
n
o
[b], [ab], a2 b , [a], a2 ∪ G2
are not regular, whence C B(G), F (G) is not Clifford semigroup, and the bottom elements in the lattice
are all idempotent elements of the class semigroup.
C B(G), F (G)
h a · a2 · b i
h b[3] i
h a[2] · b[2] i
h b[2] i
h (ab)[3] i
h (ab)[2] i
h (a2 b)[3] i
h[1F (G) ]i
h a[2] i
h (a2 b)[2] i
h a[3] i
Figure 3. Subgroup Lattice of Class Semigroup over D6
Proof of the Claim. We only show the first assertions, and others can be proved by the same way. Let
T ∈ F (G). Since b2 = 1G , it suffices to show that if b[4] · T ∈ B(G), then b[2] · T ∈ B(G). Suppose that
b[4] · T ∈ B(G).
CASE 1. |π(T )| = 3.
Then we have
T ∼ a[2] · b[2] or T ∼ a · a2 · b .
Since b[4] · T ∈ B(G), it follows that we have the only case T ∼ a[2] · b[2] . Then, since b[2] · a · b · a · b ∈ B(G),
b[2] · T ∼ b[2] · a[2] · b[2] implies that b[2] · T ∈ b[2] · a[2] · b[2] ⊂ B(G) .
CASE 2. |π(T )| = 2.
Then T is B(G)-equivalent with one of the sequence described in (4.1). Since b[4] · T ∈ B(G), the
possible choice of T under B(G)-equivalence is one the following sequences :
a · (ab)[even] ,
b · (ab)[odd≥3 ] ,
a · (a2 b)[even],
a2 · (ab)[even],
a2 · (a2 b)[even] ,
b · (a2 b)[odd≥3 ] ,
(ab)[odd] · a2 b,
ab · (a2 b)[odd] .
Then we obtain the following simple calculation, and it can cover all other cases :
ab · b · ab · b · a, a2 b · b · a2 b · a · b, ab · b · ab · a2 · b, a2 b · b · a2 b · b · a2 ,
ab · b · ab · b · ab · b, a2 b · b · a2 b · b · a2 b · b, ab · b · a2 b · b,
which are all product-one sequences, and thus it follows that b[2] · T ∈ B(G).
CASE 3. |π(T )| = 1.
Then hsupp(T )i is abelian subgroup by Lemma 3.6.5. If hsupp(T )i = h1G i, then we are done because
T ∈ B(G).
PRODUCT-ONE SEQUENCES
15
If hsupp(T )i = hbi, then T = b[n] for some n ∈ N by Lemma 3.6.4. Since b[4] · T ∈ B(G), it follows that
n should be even number, and hence we are done.
By symmetry, we can obtain the same result whenever hsupp(T )i = habi or ha2 bi.
Suppose now that hsupp(T )i = hai. Then T = a[n1 ] · (a2 )[n2 ] for some n1 , n2 ∈ N0 .
i) n1 = 0.
In this case, b[4] · (a2 )[n2 ] ∈ B(G) implies that n2 ≥ 2. It follows that b[2] · (a2 )[n2 ] ∈ B(G).
ii) n2 = 0.
In this case, b[4] · a[n1 ] ∈ B(G) implies that n1 ≥ 2. It follows that b[2] · a[n1 ] ∈ B(G).
iii) n1 ≥ 1 and n2 ≥ 1.
To avoid the trivial case, we may assume that n1 6= n2 .
If |n1 − n2 | = 1, then, since any choice of n1 , n2 can be reduced to the case
n1 = 1, n2 = 2
it suffices to verify only such two cases,
a · b · a2 · a2 · b[3] ∈ B(G)
a · a · b · a2 · b[3] ∈ B(G)
If |n1 − n2 | ≥ 2, then it can be reduced
or
n1 = 2, n2 = 1 ,
and we have the followings :
implies that a · b · a2 · a2 · b ∈ B(G),
implies that a · a · b · a2 · b ∈ B(G).
to the case n1 = 0 or n2 = 0.
Therefore, in any cases, we can obtain b[2] · T ∈ B(G), and it follows that b[2] ∼ b[4] . Moreover, b ≁ b[3]
because b[3] · (ab)[3] ∈ B(G), but b · (ab)[3] ∈
/ B(G). Hence
[2] [3]
b
[b],
b ,
are distinct B(G)-equivalence classes in the class semigroup C B(G), F (G) .
5. Arithmetic Properties of the monoid of product-one sequences
The goal of this section is to study the arithmetic of the monoid B(G) of product-one sequences. To
do so we briefly recall some arithmetical concepts (details can be found in [18]).
Let H be an atomic monoid and a, b ∈ H. If a = u1 ·. . .·uk , where k ∈ N and u1 , . . . , uk ∈ A(H), then k
is called the length of the factorization and LH (a) = L(a) = {k ∈ N | a has a factorization of length k} ⊂ N
is the set of lengths of a. As usual we set L(a) = {0} if a ∈ H × , and then
L(H) = {L(a) | a ∈ H}
denotes the system of sets of lengths of H. Let k ∈ N and suppose that H 6= H × . Then
[
L ⊂ N
Uk (H) =
k∈L,L∈L(H)
denotes the union of sets of lengths containing k, and we set
ρk (H) = sup Uk (H) .
If L = {m1 , . . . , mk } ⊂ Z is a finite subset of the integers, where k ∈ N and m1 < . . . < mk , then
∆(L) = {mi − mi−1 | i ∈ [2, k]} ⊂ N is the set of distances of L. If all sets of lengths are finite, then
[
∆(L)
∆(H) =
L∈L(H)
is the set of distances of H. It is easy to check that min ∆(H) = gcd ∆(H). If θ : H → B is a transfer
homomorphism between atomic monoids having finite sets of lengths, then ([18, Proposition 3.2.3])
(5.1)
L(H) = L(B) whence ∆(H) = ∆(B) and Uk (H) = Uk (B) for all k ∈ N.
16
JUN SEOK OH
It is well-known that if H is finitely generated, then then all unions Uk (H) of sets of lengths (in particular,
all sets of lengths) and the set of distances ∆(H) are finite ([18, Theorem 3.1.4]).
We will study the system of sets of lengths and all further mentioned invariants for the monoid B(G).
As usual, we set
L(G) = L B(G) , ∆(G) = ∆ B(G) , Uk (G) = Uk B(G)
and ρk (G) = ρk B(G) for all k ∈ N. Since B(G) is finitely generated by Lemma 3.2, all these invariants
are finite. The significance of the monoid B(G) for abelian groups stems from its relevance for general
Krull monoids. Indeed, if H is a commutative Krull monoid with class group G such that every class
contains a prime divisor, then there is a transfer homomorphism θ : H → B(G) and hence L(H) = L(G).
This transfer process from monoids of zero-sum sequences to more general monoids carries over to transfer
Krull monoids ([16]).
We study the structure of Uk (G) and of ∆(G) for general finite groups G and we will derive canonical
bounds for their size. Most results are known in the abelian case partly with different bounds (for a recent
survey on the abelian case we refer to [32]). However, if G is non-abelian, then B(G) is not a transfer
Krull monoid by Proposition 3.4. Thus the present results cannot be derived from the abelian setting. If
|G| ≤ 2, then B(G) is factorial, L(G) = {{k} | k ∈ N0 }, ∆(G) = ∅, and Uk (G) = {k} for all k ∈ N. To
avoid annoying case distinctions we exclude this trivial case.
Throughout this section, let G be a finite group with |G| ≥ 3.
Although the forthcoming proofs parallel the proofs given in the commutative setting there is a main
difference. It stems from the fact that, in the non-abelian case, the embedding B(G) ֒→ F (G) is not a
divisor homomorphism. Thus there exist U, V ∈ B(G) such that U divides V in F (G), but not in B(G).
Moreover, U and V can be atoms (e.g., if G is the quaternion group, as discussed in Examples 4.1, then
U = I [4] ∈ A(G) and V = I [4] · J [2] ∈ A(G) have this property).
Lemma 5.1. Let k, ℓ ∈ N with k < ℓ and U1 , . . . , Uk , V1 , . . . , Vℓ ∈ A(G) such that U1 ·. . .·Uk = V1 ·. . .·Vℓ .
There exist µ ∈ [1, k], λ, λ′ ∈ [1, ℓ] with λ 6= λ′ , and g1 , g2 ∈ G such that Uµ = g1 · g2 · . . . · gm with m ≥ 2,
1G = g1 . . . gm , g1 | Vλ , and g2 | Vλ′ in F (G).
Proof. Assume to the contrary that the assertion does not hold. Since 1G is a prime element of B(G), we
may suppose without restriction that all Ui and Vj have length at least two. We set U1 = g1 · g2 · . . . · gm
where m ≥ 2 and 1G = g1 . . . gm . Then g1 · g2 divides Vλ for some λ ∈ [1, ℓ], say λ = 1. Then we consider
g2 · g3 . Since the assertion does not hold, it follows that g1 · g2 · g3 divides V1 . Proceeding recursively we
obtain that U1 | V1 , say V1 = U1 · V1′ with V1′ ∈ F (G). Thus we obtain the equation
U2 · . . . · Uk = V1′ · V2 · . . . · Vℓ .
Proceeding in this way we end up with an equation of the form
1F (G) = S1 · . . . · Sη · Vη+1 · . . . · Vℓ ,
where η ≤ k and S1 , . . . , Sη ∈ F (G), a contradiction to k < ℓ.
We consider the following property :
P. If U = g1 · . . . · gℓ ∈ A(G) and g1 = h1 h2 with h1 , h2 ∈ G, then U ′ = h1 · h2 · g2 · . . . · gℓ is either an
atom or a product of two atoms at most.
Remark 5.2. Of course, every abelian group satisfies Property P and the same is true for some nonabelian groups such as the quaternion group. However, for every n ≥ 9, the dihedral group D2n does not
have Property P as the following example shows.
PRODUCT-ONE SEQUENCES
17
Example 5.3. 1. Let G = Q8 be the quaternion group as discussed in Example 4.1. Then D(G) = 6,
and clearly any atom having length at most 4 satisfies Property P. Any atom U having the length 6 has
the form
[4]
[2]
U = g1 · g2 , where g1 , g2 ∈ {I, J, K, −I, −J, −K} with g2 6= ±g1 ,
and any atom V having length 5 is one of the following three types :
[3]
• V = g1 · g2 · g3 ,
[2]
g1
where g1 , g2 , g3 ∈ {I, J, K, −I, −J, −K} with g2 6= g3 and g2 , g3 6= ±g1 .
[2]
g2
• V =
·
· (−E), where g1 , g2 ∈ {I, J, K, −I, −J, −K} with g2 6= ±g1 .
• V = g1 · (−g1 ) · g2 · (−g2 ) · (−E), where g1 , g2 ∈ {I, J, K} with g1 6= g2 .
It is easy to check that G satisfies Property P.
2. Let G = D2n = ha, b | an = b2 = 1G and ba = a−1 bi be the dihedral group, where n ≥ 9. Consider
the following sequences :
U = b[2] · (ba)[3] · a2 · ba5
and
U ′ = b[2] · (ba)[3] · ba2 · ba4 · ba5 .
Then U ′ is a product of b[2] , (ba)[2] , and ba2 · ba · ba4 · ba5 , which are atoms. Now we need to verify that
U is an atom. Assume to the contrary that U is not atom, say U = W1 · W2 with W1 , W2 ∈ B(G), and
let W1 be an atom with a2 | W1 . Write W1 = g1 · g2 · . . . · gℓ with g1 = a2 and g1 g2 . . . gℓ = 1G . Then
ℓ = 3 or ℓ = 5. If ℓ = 3, then g2 g3 = a−2 and g2 , g3 ∈ {b, ba, ba5 }. Since n ≥ 9, this is impossible. If
ℓ = 5, then W2 = b[2] or W2 = (ba)[2] which implies that W1 = (ba)[3] · a2 · ba5 or W1 = b[2] · ba · a2 · ba5 .
It is easy to check that W1 is not an atom. Thus G does not have Property P.
Lemma 5.4. Suppose that G satisfies Property P. Then for every S ∈ B(G) with max ∆ L(S) ≥ 2,
there exists T ∈ B(G) such that |T | < |S| and max ∆ L(T ) ≥ max ∆ L(S) − 1.
Proof. Let S ∈ B(G) and d = max ∆ L(S) ≥ 2. Then there are k, ℓ ∈ N and U1 , . . . , Uk , V1 , . . . , Vℓ ∈
A(G) such that
S = U1 · . . . · Uk = V1 · . . . · Vℓ
with ℓ − k = d and U1 · . . . · Uk has no factorization of length in [k + 1, ℓ − 1]. Since 1G ∈ B(G) is a prime
element, we may assume that 1G does not divide S, and thus all Ui and Vj have length at least 2.
By Lemma 5.1, we may assume that there are g1 , g2 ∈ G such that
g1 · g2 | U1 ,
Let U1 = g1 · g2 ·
and consider
U1′ ,
V1 = g1 ·
V1′
g1 | V1 and g2 | V2 in F (G).
and V2 = g2 · V2′ with U1′ , V1′ , V2′ ∈ F (G). Then we set g0 = g1 g2 ∈ G,
U1′′ = g0 · U1′ and V1′′ = g0 · V1′ · V2′ .
Clearly, U1′′ ∈ A(G). Since V1 gives rise to a product-one equation with g1 being the finial element and
V2 gives rise to a product-one equation with g2 being the first element, it follows that V1′′ ∈ B(G). We
obtain that
U1′′ · U2 · . . . · Uk = V1′′ · V3 · . . . · Vℓ .
We set T = U1′′ · U2 · . . . · Uk and observe that |T | < |S|. If T = W1 · . . . · Wt with W1 , . . . , Wt ∈ A(G)
and g0 | W1 in F (G), then W1 = g0 · W1′ and W1′′ = g1 · g2 · W1′ is either atom or product of precisely two
atoms. Thus S = U1 · . . . · Uk has a factorization of length t or t + 1. It follows that T has no factorization
of length in [k + 1, ℓ − 2], and thus we obtain
max ∆ L(T ) ≥ ℓ − 1 − k = d − 1 = max ∆ L(S) − 1 .
The next result shows that the set of distances and all unions of sets of lengths of B(G) are finite
intervals. This is far from being true in general. Indeed, for every finite set ∆ ⊂ N with min ∆ = gcd ∆
there is a finitely generated Krull monoid H such that ∆(H) = ∆ ([22]).
18
JUN SEOK OH
Theorem 5.5.
1. Uk (G) is a finite interval for all k ∈ N.
2. If G satisfies Property P, then ∆(G) is a finite interval with min ∆(G) = 1.
Proof. 1. Let k ∈ N. We set λk (G) = min Uk (G). Then Uk (G) ⊂ [λk (G), ρk (G)], and we have to show
equality. Note that it is sufficient to prove that [k, ρk (G)] ⊂ Uk (G). Indeed, suppose that this is done,
and let ℓ ∈ [λk (G), k]. Then ℓ ≤ k ≤ ρℓ (G), hence k ∈ Uℓ (G) and consequently ℓ ∈ Uk (G). It follows that
[λk (G), ρk (G)] = Uk (G).
To prove the assertion, let ℓ ∈ [k, ρk (G)] be minimal such that [ℓ, ρk (G)] ⊂ Uk (G). Assume to the
contrary that ℓ > k. We set Ω = {A ∈ B(G) | {k, j} ⊂ L(A) for some j ≥ ℓ} and choose B ∈ Ω such that
|B| is minimal. Then
B = U1 · . . . · Uk = V1 · . . . · Vj ,
where U1 , . . . , Uk , V1 , . . . , Vj ∈ A(G) ,
and by Lemma 5.1, we may assume that
Uk = g1 · g2 · U with g1 | Vj−1 and g2 | Vj ,
where U = g3 · . . . · gt and g1 . . . gt = 1G . We set g0 = g1 g2 ∈ G and obtain that Uk′ = g0 · U ∈ A(G).
Write
Vj−1 = h1 · . . . · hn · g1 and Vj = g2 · s1 · . . . · sm
′
with h1 . . . hn g1 = 1G and g2 s1 . . . sm = 1G . Let Vj−1
= g0 · V ∈ B(G), where V = h1 · . . .· hn · s1 · . . . · sm ∈
F (G), and consider
B ′ = U1 · . . . · Uk−1 · Uk′ .
Then |B ′ | < |B|, and
B ′ = U1 · . . . · Uk−1 · Uk′ = V1 · . . . · Vj−2 · W1 · . . . · Wt ,
′
= W1 · . . . · Wt with W1 , . . . , Wt ∈ A(G). By the minimality of |B|, we have (j − 2) + t < ℓ ≤ j.
where Vj−1
Hence t = 1 and j = ℓ. Thus ℓ − 1 ∈ Uk (G), a contradiction to the minimality of ℓ. Therefore ℓ = k and
hence [k, ρk (G)] ⊂ Uk (G).
2. Since ∆(G)
is finite as mentioned before, we can take S0 ∈ B(G) with minimal length such that
max ∆ L(S0 ) = max ∆(G). By Lemma 5.4, we can find S1 ∈ B(G) with minimal length such that
|S1 | < |S0 | and max ∆ L(S1 ) ≥ max ∆(G) − 1. By the minimality of S0 ,
|S1 | < |S0 |
implies
max ∆ L(S1 ) < max ∆(G) .
It follows that max ∆(G) − 1 = max ∆ L(S1 ) ∈ ∆(G). Again Lemma 5.4 implies that we can find
S2 ∈ B(G) with minimal length such that |S2 | < |S1 | and max ∆ L(S2 ) ≥ max ∆(G) − 2. By the
minimality of S1 ,
|S2 | < |S1 |
implies
max ∆ L(S2 ) < max ∆(G) − 1 .
It follows that max ∆(G) − 2 = max ∆ L(S2 ) ∈ ∆(G). Continuing this process, we can obtain Sn ∈ B(G)
such that
1 = max ∆(G) − n = max ∆ L(Sn ) ∈ ∆(G) .
Thus ∆(G) = [1, max ∆(G)] is an interval.
Next we study the maxima of the sets Uk (G). Recall that we have defined ρk (G) = max Uk (G) for all
k ∈ N. Even in case of abelian groups, precise values of ρ2k+1 (G) are unknown in general ([11]).
Proposition 5.6. Let k ∈ N.
for all k ∈ N.
1. ρk (G) ≤ kD(G)
2
2. ρ2k (G) = kD(G) for all k ∈ N.
In particular, 1 + kD(G) ≤ ρ2k+1 (G) ≤ kD(G) +
D(G)
2
for all k ∈ N.
PRODUCT-ONE SEQUENCES
19
Proof. 1. Let k ∈ N. Let A ∈ B(G) with k ∈ L(A). We have to show that max L(A) ≤ kD(G)
2 . There are
U1 , . . . , Uk ∈ A(G) such that
A = U1 · . . . · Uk .
Since 1G ∈ B(G) is a prime element, we may assume that 1G does not divide A, and hence any atom
dividing A has length at least 2. If A = V1 · . . . · Vℓ for V1 , . . . , Vℓ ∈ A(G), then we obtain
2ℓ ≤
ℓ
X
i=1
|Vi | = |A| =
k
X
|Uj | ≤ kD(G) .
j=1
kD(G)
It follows that ℓ ≤ kD(G)
2 , and thus max L(A) ≤
2 .
2. Let k ∈ N. Then, by 1., we have ρ2k (G) ≤ kD(G). Now let U = g1 · . . . · gℓ ∈ A(G) with
ℓ = |U | = D(G), where g1 , . . . , gℓ ∈ G. Consider the sequence
V = g1−1 · . . . · gℓ−1 ∈ A(G) .
Then kD(G) ∈ L (U · V )[k] . Since 2k ∈ L (U · V )[k] , we obtain ρ2k (G) ≥ kD(G), and hence ρ2k (G) =
kD(G).
It remains to prove the “In particular” statement. For all i, j ∈ N, we have Ui (G) + Uj (G) ⊂ Ui+j (G)
whence ρi (G) + ρj (G) ≤ ρi+j (G). Thus the “In particular” statement follows from 1. and 2.
Our next goal is to study the maximum of the set of distances ∆(G). If G is abelian, then precise
values for max ∆(G) are known only for very special classes of groups including cyclic groups ([23]). For
general groups G we study max ∆(G) via a further invariant (introduced in Definition 5.7). The main
result is Proposition 5.12.
Definition 5.7. Let H be an atomic monoid.
1. For b ∈ H, let ω(H, b) denote the smallest N ∈ N0 ∪ {∞} with the following property :
For all n ∈ N and a1 , . . . , an ∈ H, if b | a1 . . . an , then
Q there exists a subset Ω ⊂ [1, n] such that
|Ω| ≤ N and b | ν∈Ω aν .
2. We set
ω(H) = sup {ω(H, u) | u ∈ A(H)} ∈ N0 ∪ {∞}.
If H is an atomic monoid with ∆(H) 6= ∅, then 2 +max ∆(H) ≤ ω(H) (see [18, Theorem 1.6.3] and
[20, Proposition 3.6]). As usual we set ω(G) = ω B(G) and observe that 2 + max ∆(G) ≤ ω(G). If G is
a cyclic group or an elementary 2-group, then 2 + max ∆(G) = ω(G) = D(G).
Lemma 5.8. We have D(G) ≤ ω(G), and equality holds if G is abelian.
Proof. Let U = g1 · . . . · gℓ ∈ A(G) with ℓ = |U | = D(G). Put Vi = gi · gi−1 ∈ A(G) for all i ∈ [1, ℓ]. Then
we obtain
U | V1 · . . . · Vℓ in B(G) .
Assume to the contrary that U divides a proper subproduct in B(G), say U | V1 ·. . .·Vk for some k ∈ [1, ℓ−1].
Then 2k ≥ ℓ. If 2k = ℓ, then U = V1 · . . . · Vk is not atom in B(G). If 2k > ℓ, then ℓ − k < k, and, for
each j = [1, ℓ − k], we may assume that gk+j = gj−1 . Hence we have
U = (g1 · gk+1 ) · . . . · (gℓ−k · gℓ ) · (gℓ−k+1 · . . . · gk ) ,
and each terms is in B(G). It follows that U is not atom in B(G), a contradiction. Thus U cannot divide
a proper subproduct, and hence D(G) ≤ ω(G).
Suppose that G is abelian. It remains to show that ω(G) ≤ D(G). Let U ∈ A(G). If U | V1 · . . . · Vk for
V1 , . . . , Vk ∈ B(G), then there is a subproduct of at most |U | factors such that U divides this subproduct in
F (G). Since B(G) ֒→ F (G) is a divisor homomorphism, this divisibility holds in B(G). Hence ω(G, U ) ≤
|U | ≤ D(G), and it follows that ω(G) ≤ D(G).
20
JUN SEOK OH
Next we define Davenport constants of additive commutative semigroups (as studied in [35, 1, 34]) and
will apply these concepts to the class semigroup of B(G). Recall that all our semigroups have an identity
element (a zero-element in the additive case) and we use the convention that an empty sum equals the
zero-element.
Definition 5.9. Let C be an additive commutative semigroup.
1. We denote by d(C) the smallest d ∈ N0 ∪ {∞} with the following property :
P
P
For all n ∈ N and c1 , . . . , cn ∈ C, there exists Ω ⊂ [1, n] such that |Ω| ≤ d and nν=1 cν = ν∈Ω cν .
2. We denote by D(C) the smallest ℓ ∈ N ∪ {∞} with the following property :
P
P
For all n ∈ N with n ≥ ℓ and c1 , . . . , cn ∈ C, there exists Ω ( [1, n] such that nν=1 cν = ν∈Ω cν .
If G is a finite abelian group, then of course the definitions for D(G) and d(G) given in Definition 3.1
and in Definition 5.9 coincide.
Lemma 5.10. If C is a finite commutative semigroup, then D(C) = d(C) + 1 ≤ |C|.
Proof. The equality D(C) = d(C) + 1 is verified in [1, Proposition 1.2]. Since we could not find a reference
that |C| is an upper bound, we provide the short argument. To show that D(C) ≤ |C|, it suffices to verify
that |C| has the property given in Definition 5.9. So let n ∈ N with n ≥ |C| and c1 , . . . , cn ∈ C. If all sums
X
cν , c1 , c1 + c2 , . . . , c1 + . . . + cn−1
0=
ν∈∅
are pairwise distinct, then one of the sums coincides with c1 + . . . + cn as required. Suppose there are
k, ℓ ∈ [0, n − 1] with k < ℓ and c1 + . . .+ ck = c1 + . . .+ cℓ . Then c1 + . . .+ cn = c1 + . . .+ ck + cℓ+1 + . . .+ cn
as required.
Lemma 5.11. Let F be a monoid, H ⊂ F a submonoid with H × = H ∩ F × and C = C ∗ (H, F ).
If n ∈ N and f, a1 , . . . , an ∈ F with f a1 . . . an ∈ H, then there exists a subset Ω ⊂ [1, n] such that
Y
|Ω| ≤ d(C) and f
aν ∈ H .
ν∈Ω
Proof. By the definition of d(C), there exists a subset Ω ⊂ [1, n] such that |Ω| ≤ d(C) and
n
X
[aν ]F
H =
ν=1
X
[aν ]F
H.
ν∈Ω
Since f a1 . . . an ∈ H, we have
n
n
iF
h Y iF
h Y
Y
X
X
F
F
F
F
f
aν ∈ f
aν
= [f ]H +
[aν ]H = [f ]H +
[aν ]H = f
aν
⊂H.
ν∈Ω
Proposition 5.12.
ν∈Ω
H
ν∈Ω
ν=1
ν=1
H
1. D(G) ≤ ω(G) ≤ D(G) + d C B(G), F (G) .
2. D(G/G′ ) ≤ D C B(G), F (G) , and equality holds if G is abelian.
Proof. We set C = C(B(G), F (G) .
1. The first inequality follows from Lemma 5.8. For the second inequality, let U ∈ A(G), and
A1 , . . . , An ∈ B(G) such that
U | A1 · . . . · An in B(G) .
PRODUCT-ONE SEQUENCES
21
After renumbering if necessary, we may assume that U | A1 · . . . · A|U| in F (G), say A1 · . . . · A|U| = U · W
with W ∈
F (G). Then
W · A|U|+1 · . . . · An ∈ B(G). By Lemma 5.11, there is a subset Ω ⊂ |U | + 1, n ,
say Ω = |U | + 1, m , such that |Ω| ≤ d(C) and W · A|U|+1 · . . . · Am ∈ B(G). Thus we obtain
A1 · . . . · Am = U · (W · A|U|+1 · . . . · Am )
and m ≤ |U | + d(C) ≤ D(G) + d(C).
2. By Theorem 3.8.1, G/G′ is an isomorphic to a subsemigroup of C. Clearly, this implies that
D(G/G′ ) ≤ D(C). Suppose that G is abelian. Then |G′ | = 1 and G/G′ is isomorphic to G. Furthermore,
B(G) ֒→ F (G) is a divisor theory, the class semigroup C is isomorphic to the class group of B(G) by
Lemma 2.1, and the class group of B(G) is isomorphic to G. Thus equality holds.
References
[1] S.D. Adhikari, W. Gao, and Guoqing Wang, Erdős-Ginzburg-Ziv theorem for finite commutative semigroups, Semigroup
Forum 88 (2014), 555 – 568.
[2] N.R. Baeth and D. Smertnig, Factorization theory: From commutative to noncommutative settings, J. Algebra 441
(2015), 475 – 551.
[3] J. Bass, Improving the Erdős-Ginzburg-Ziv theorem for some non-abelian groups, J. Number Theory 126 (2007), 217 –
236.
[4] F.E. Brochero Martı́nez and S. Ribas, Extremal product-one free sequences in Dihedral and Dicyclic Groups, Discrete
Mathematics(2017), https://doi.org/10.1016/j.disc.2017.09.024.
[5] F. Chen and S. Savchev, Long minimal zero-sum sequences in the groups C2r−1 ⊕ C2k , Integers 14 (2014), Paper A23.
[6] K. Cziszter and M. Domokos, On the generalized Davenport constant and the Noether number, Central European J.
Math. 11 (2013), 1605 – 1615.
[7]
, Groups with large Noether bound, Ann. Inst. Fourier (Grenoble) 64 (2014), 909 – 944.
, The Noether number for the groups with a cyclic subgroup of index two, J. Algebra 399 (2014), 546 – 560.
[8]
[9] K. Cziszter, M. Domokos, and A. Geroldinger, The interplay of invariant theory with multiplicative ideal theory and
with arithmetic combinatorics, Multiplicative Ideal Theory and Factorization Theory, Springer, 2016, pp. 43 – 95.
[10] K. Cziszter, M. Domokos, and I. Szöllősi, The Noether number and the Davenport constants of the groups of order less
than 32, arXiv:1702.02997.
[11] Y. Fan and Q. Zhong, Products of k atoms in Krull monoids, Monatsh. Math. 181 (2016), 779 – 795.
[12] W. Gao and A. Geroldinger, Zero-sum problems in finite abelian groups : a survey, Expo. Math. 24 (2006), 337 – 369.
[13] W. Gao and Yuanlin Li, The Erdős-Ginzburg-Ziv theorem for finite solvable groups, J. Pure Appl. Algebra 214 (2010),
898 – 909.
[14] W. Gao and Zaiping Lu, The Erdős-Ginzburg-Ziv theorem for dihedral groups, J. Pure Appl. Algebra 212 (2008), 311
– 319.
[15] A. Geroldinger, Additive group theory and non-unique factorizations, Combinatorial Number Theory and Additive
Group Theory (A. Geroldinger and I. Ruzsa, eds.), Advanced Courses in Mathematics CRM Barcelona, Birkhäuser,
2009, pp. 1 – 86.
[16]
, Sets of lengths, Amer. Math. Monthly 123 (2016), 960 – 988.
[17] A. Geroldinger and D.J. Grynkiewicz, The large Davenport constant I: Groups with a cyclic index 2 subgroup, J. Pure
Appl. Algebra 217 (2013), 863 – 885.
[18] A. Geroldinger and F. Halter-Koch, Non-Unique Factorizations. Algebraic, Combinatorial and Analytic Theory, Pure
and Applied Mathematics, vol. 278, Chapman & Hall/CRC, 2006.
[19] A. Geroldinger and W. Hassler, Arithmetic of Mori domains and monoids, J. Algebra 319 (2008), 3419 – 3463.
[20] A. Geroldinger and F. Kainrath, On the arithmetic of tame monoids with applications to Krull monoids and Mori
domains, J. Pure Appl. Algebra 214 (2010), 2199 – 2218.
[21] A. Geroldinger, S. Ramacher, and A. Reinhart, On v-Marot Mori rings and C-rings, J. Korean Math. Soc. 52 (2015),
1 – 21.
[22] A. Geroldinger and W. A. Schmid, A realization theorem for sets of distances, J. Algebra, 481 (2017), 188 – 198.
[23] A. Geroldinger and Q. Zhong, The catenary degree of Krull monoids II, J. Australian Math. Soc. 98 (2015), 324 – 354.
[24] D.J. Grynkiewicz, The large Davenport constant II: General upper bounds, J. Pure Appl. Algebra 217 (2013), 2221 –
2246.
, Structural Additive Theory, Developments in Mathematics, vol. 30, Springer, 2013.
[25]
[26] Dongchun Han, The Erdős-Ginzburg-Ziv Theorem for finite nilpotent groups, Archiv Math. 104 (2015), 325 – 332.
[27] Dongchun Han and Hanbin Zhang, Erdős-Ginzburg-Ziv Theorem and Noether number for Cm ⋉ϕ Cmn , arXiv:1707.03639.
[28] F. Kainrath, Arithmetic of Mori domains and monoids : The Global Case, Multiplicative Ideal Theory and Factorization
Theory, Springer, 2016, pp. 183 – 218.
22
JUN SEOK OH
[29] J.E. Olson and E.T. White, Sums from a sequence of group elements, Number Theory and Algebra (H. Zassenhaus,
ed.), Academic Press, 1977, pp. 215 – 222.
[30] A. Reinhart, On integral domains that are C-monoids, Houston J. Math. 39 (2013), 1095 – 1116.
[31] W.A. Schmid, The inverse problem associated to the Davenport constant for C2 ⊕ C2 ⊕ C2n , and applications to the
arithmetical characterization of class groups, Electron. J. Comb. 18(1) (2011), Research Paper 33.
, Some recent results and open problems on sets of lengths of Krull monoids with finite class group, Multiplicative
[32]
Ideal Theory and Factorization Theory (S.T. Chapman, M. Fontana, A. Geroldinger, and B. Olberding, eds.), Springer,
2016, pp. 323 – 352.
[33] D. Smertnig, Sets of lengths in maximal orders in central simple algebras, J. Algebra 390 (2013), 1 – 43.
[34] Guoquing Wang, Davenport constants for semigroups II, J. Number Theory 153 (2015), 124 – 134.
[35] Guoquing Wang and Weidong Gao, Davenport constants for semigroups, Semigroup Forum 76 (2008), 234 – 238.
Institute for Mathematics and Scientific Computing, University of Graz, NAWI Graz, Heinrichstraße 36,
8010 Graz, Austria
E-mail address: [email protected]
| 0math.AC
|
.
TOP LOCAL COHOMOLOGY, MATLIS DUALITY AND TENSOR PRODUCTS
MAJID EGHBALI
A BSTRACT. Let a be an ideal of a local ring ( R, m) with d = dim R. For i > d the local cohomology
module Hai ( R) vanishes. In the case that Had ( R) 6 = 0, at first we give a bound for depth D ( Had ( R)),
arXiv:1302.1274v1 [math.AC] 6 Feb 2013
where d > 2 and ( R, m) is complete. More precisely we show
grade(b, D ( Had ( R))) + dim R/b ≥ depth( D ( Had ( R))) ≥ 2,
where b is an ideal of R of dimension ≤ 2 and D(-) denotes the Matlis duality functor. Later we
give conditions that Cohen-Macaulayness of D ( Had ( R)) fails, where d ≥ 3. Finally, Had ( R) ⊗ R Had ( R),
D ( Had ( R)) ⊗ R D ( Had ( R)) and Had ( R) ⊗ R D ( Had ( R)) are examined. In particular, we give the necessary
and sufficient condition for Cohen-Macaulayness of D ( Had ( R)) ⊗ R/Q D ( Had ( R)), where Q is a certain
ideal of R.
1. I NTRODUCTION
Let a ⊂ R denote an ideal of a local ring ( R, m) with d = dim R. Let M be a finitely generated
R-module. For an integer i ∈ Z, let Hai ( M ) denote the ith local cohomology module of M with
respect to a as introduced by Grothendieck (see [G] and [Br-Sh]). Let E := ER ( R/m) be the
injective hull of the R-module R/m. By D (−) we denote the Matlis dual functor HomR (−, E).
In algebraic geometry a (set-theoretic) complete intersection is a variety Y that can be cut out
by codim(Y ) many equations. By a result of Hellus [H, Corollary 1.1.4] Matlis dual of local cohomology can be useful to see whether we have a set-theoretic complete intersection ideal or not.
Of a particular interest is the last non-vanishing degree of the local cohomology module Hai ( R)
known as the cohomological dimension cd(a, R). In general cd(a, R) ≤ d and in the case that
a = m, the equality occurs. Hartshorne-Lichtenbaum Vanishing Theorem gives the necessary and
sufficient condition for non-vanishing of Had ( R) (cf. [Br-Sh, Theorem 8.2.1]). Here we assume that
Had ( R) 6= 0 and give some information on D ( Had ( R)) and the tensor products of these modules.
First, for a complete local ring ( R, m) we show that depth D ( Had ( R)) ≥ min{2, d} (see Theorem
3.1 ). Then, we give an upper bound for depth( D ( Had ( R))):
grade(b, D ( Had ( R))) + dim R/b ≥ depth( D ( Had ( R))) ≥ 2,
where b is an ideal of R of dimension at most 2 and d ≥ 3. Hence, it provides conditions in which
Cohen-Macaulayness of D ( Had ( R)) fails (d ≥ 3), see Remark 3.5.
Recently, there were some researches on the endomorphism ring of local cohomology modules
and their importance, see for instance [H-St], [H-Sch], [Sch3] and [E-Sch]. However not so much
is known on its dual, that is on the tensor product of local cohomology modules. Section four
is devoted to examining the modules Had ( R) ⊗ R Had ( R), D ( Had ( R)) ⊗ R D ( Had ( R)) and Had ( R) ⊗ R
2000 Mathematics Subject Classification. 13D45, 13C14.
Key words and phrases. Top local cohomology, Depth, Tensor products, Matlis duality, Cohen-Macaulayness.
The author is supported in part by a grant from IPM (No. 91130026).
1
2
EGHBALI
D ( Had ( R)). In particular, in the case R is complete and pdR/Q (K R/Q ) < ∞ for a certain ideal Q
of R we give the necessary and sufficient conditions for Cohen-Macaulayness of D ( Had ( R)) ⊗ R/Q
D ( Had ( R)) (see Theorem 4.2). We denote by pdR ( M ) resp. K M the projective dimension resp. the
canonical module of a finite R-module M (if it exists).
2. P RELIMINARIES
Let ( R, m) be a local ring of dimension d. We denote by ER ( R/m) the injective hull of the
residue field R/m = k. The Matlis duality functor Hom R (−, ER ( R/m)) is denoted by D (−). We
state the following lemmas which we need in our proofs.
Lemma 2.1. Let ( R, m) be a complete local ring. Suppose that M and N are either Artinian or finitely
generated R-modules. Then we have the following isomorphism
Hom( M, N ) ∼
= Hom( D ( N ), D ( M )).
Proof. Since N ∼
= D ( D ( N )) so by [R, Theorem 2.11] we get the claim as follows:
Hom R ( M, N )
∼
= Hom R ( M, D ( D ( N )))
∼
= Hom R ( D ( N ), D ( M )).
Lemma 2.2. Let ( R, m) be a complete local ring. Let M and N be Artinian R-modules. Then we have the
following isomorphism
D( N ) ⊗R M ∼
= D (HomR ( M, N )).
Proof. It should be noted that, by Matlis duality D ( N ) is a finitely generated R-modules. By
Hom −⊗-adjointness we can prove the claim as follows:
D( N ) ⊗R M
∼
= D ( N ) ⊗ R D ( D ( M ))
∼
= D ( N ) ⊗ R Hom R ( D ( M ), E( R/m))
∼
= Hom R (HomR ( D ( N ), D ( M )), E( R/m))
∼
= Hom R (HomR ( M, N ), E( R/m)).
(cf. [Br-Sh, 10.2.16])
The last isomorphism follows by 2.1.
We need the definition of the canonical module of a finitely generated R-module M. To this
end let ( R, m) be the epimorphic image of a local Gorenstein ring (S, n) and n = dim S.
Definition 2.3. (A) For a finitely generated R-module M we consider
t
K ( M ) = HomR ( Hm
( M ), ER ( R/m)), t = dim M.
t ( M ) is an Artinian R-module it turns out that K ( M ) is a finitely generated R̂-module.
Because Hm
(B) This is closely related to the notion of the canonical module K M of a finitely generated Rmodule M. To this end we have to assume that ( R, m) is the epimorphic image of a local Gorenstein ring (S, n) with n = dim S. Then K M = ExtnS−t ( M, S) is called the canonical module of M.
TOP LOCAL COHOMOLOGY
3
By virtue of the Cohen’s Structure Theorem and (A) there is an isomorphism K ( M ) ≃ K M̂ . By the
Local Duality Theorem (see [G]) there are the following isomorphisms
t
Hom R ( Hm
( M ), ER ( R/m)) ≃ K ( M ) ≃ K M̂ .
Here R̂ resp. M̂ denote the m-adic completion of R resp. of M.
In the following, we summarize a few results about the last non-vanishing local cohomology
module Had ( R). To this end note that a finite module M over a Noetherian ring R satisfies Serre’s
condition Sn if depth Mp ≥ min{n, dim Mp } for all p ∈ Supp M.
Lemma 2.4. Let ( R, m) be a complete local ring of dimension d and Had ( R) 6= 0.
(a) Had ( R) is an Artinian R-module.
(b) H d ( R) ∼
= H d ( R/Qa ( R)), where
a
m
Qa ( R) =
T
{q : q primary component of R, dim R/q = d and Rad(a + q) = m}.
(c) D ( Had ( R)) ∼
= K R/Qa ( R) is a finitely generated R-module.
(d) Ann R Had ( R) = Qa ( R).
(e) dim D ( Had ( R)) = d.
(f) The module D ( Had ( R)) satisfies Serre’s condition S2 .
(g) If R/Qa ( R) is a Cohen-Macaulay ring, then the module D ( Had ( R)) is Cohen-Macaulay.
Proof. For the proof of (a), see [Br-Sh, Theorem 7.1.6]. For the proof of (b), (c), (d) and (g) we
refer to [E-Sch, 3.3, 4.1(c), 4.2(a) and 4.3]. Part (e) follows by (b) and (d). The prove of (f) is a
consequence of (c) and [Sch1, Lemma 1.9].
As a technical tool, in Section three, we use the so-called formal local cohomology modules.
i ( M/an M ) the ith formal local cohomology of M with respect to
We denote by Fai ( M ) := limn Hm
←−
a. see [Sch2] for more information and also [E].
Lemma 2.5. (see [Sch2, Remark 3.6 and Theorem 4.5]) Let ( R, m) be a local ring.
(a) If M is a finitely generated R-module, then
dimR M/aM = sup{i ∈ Z : Fai ( M ) 6= 0}.
(b) If the local ring ( R, m) is a quotient of a local Gorenstein ring (S, n), Then
n−i
Fai ( M ) ∼
( M, S), E), i ∈ Z,
= HomR ( HaS
where M is considered as an S-module, n = dim S and E denotes the injective hull of the residue
field R/m.
It should be noted that Hai ( M, N ) = limn ExtiR ( M/an M, N ) is known as the ith generalized
−→
local cohomology module provided that M and N are finitely generated R-modules. In case
R = M, the module Hai ( M, N ) is the same as the ordinary local cohomology module Hai ( N ).
4
EGHBALI
3.
ON THE DEPTH
Let ( R, m) be a complete local ring of dimension d and a be an ideal of R. It is a well-known
result that D ( Had ( R)) is a Cohen-Macaulay module of dimension d for d ≤ 2 (see for example
2.4(f)). In this section we examine the depth of Matlis dual of top local cohomology modules.
Theorem 3.1. Let ( R, m) be a complete local ring of dimension d. Let Had ( R) be nonzero. Then
depth( D ( Had ( R)) ≥ min{d, 2}.
Proof. We prove the claim by induction on d. For d ≤ 2, the claim is clear. Assume that d ≥ 3.
∼ H d ( R/Q), where Q := Qa ( R). By
First note that by virtue of Lemma 2.4 we have Had ( R) =
m
the definition of Q we may deduce that m ∈
/ Ass R/Q and so depth R/Q > 0. Hence, there exists
an element x ∈ m which is nonzero divisor on R/Q. From the short exact sequence
x
0 → R/Q → R/Q → R/( x, Q) → 0
we get the following long exact sequence
x
d ( R/Q )) → D ( H d ( R/Q )) → D ( H d −1 ( R/( x, Q )) → · · · .
0 → D ( Hm
m
m
Since dim( R/( x, Q)) = d − 1, so by induction hypothesis one can see that
Hence, there are u, v ∈ m \
S
d −1
depth D ( Hm
( R/( x, Q)) ≥ 2.
d ( R/Q )) /xD ( H d ( R/Q ))) p.
p∈Ass( D ( Hm
m
Thus it follows that
d
depth( D ( Hm
( R/Q)) ≥ 2.
According to [H, Corollary 1.1.4] the grade(a, D ( Had ( R))) plays a significant role to see whether
a is a set-theoretic complete intersection. In this direction, we try to find an explicit relation
between depth D ( Had ( R)) and grade(a, D ( Had ( R))) (see Theorem 3.2). To this end we denote by
gradea ( M, N ) the least integer i such that Hai ( M, N ) is nonzero for finitely generated R-modules
M and N.
Theorem 3.2. Let a be an ideal of a local ring ( R, m). Let M and N be finitely generated R-modules. Then
gradea ( M, N ) ≥ max{0, depth N − dim M/aM }.
In order to prove Theorem 3.2 we need some preparations.
Remark 3.3. Let ( R, m) be a local ring which is a factor ring of a Gorenstein local ring (S, n). In the
light of Lemma 2.5 we have
dim M/aM
= sup{i : Fai ( M ) 6= 0}
dim S − i
= sup{i : HaS
( M, S) 6= 0}
j
= sup{dim S − j : HaS ( M, S) 6= 0}
i ( M, S) 6 = 0}
= dim S − inf{i : HaS
= dim S − gradea ( M, S),
TOP LOCAL COHOMOLOGY
5
where
dim M/aM = dim S − grade(a + AnnS M ).
See [B, Proposition 5.5].
Lemma 3.4. Let M and N be two finitely generated R-modules and a be an ideal of R. Then
gradea ( M, N ) ≥ gradea ( M, R) − pdR N,
provided that pdR N < ∞ .
Proof. We argue by induction on n := pdR N.
Put n = 0 so N is a finite free R-module of the form N ∼
= ⊕t R, where t is an integer. Hence, the
isomorphisms
H i ( M, N ) ∼
= H i ( M, ⊕t R) ∼
= ⊕t H i ( M, R)
a
a
a
hold. Clearly gradea ( M, N ) ≥ gradea ( M, R).
Next, suppose that n > 0 and the claim holds for n − 1. There is a finite free R-module F such
that
0→L→F→N→0
which induces pdR L = n − 1. By induction hypothesis gradea ( M, L) ≥ gradea ( M, R) − n + 1. If
i < gradea ( M, R) − n, then it follows from the long exact sequence
· · · → Hai ( M, F ) → Hai ( M, N ) → Hai+1 ( M, L) → · · ·
that Hai ( M, N ) is zero. Hence, gradea ( M, N ) ≥ gradea ( M, R) − pdR N.
Now is the time to give the proof of Theorem 3.2 as follows:
Proof of Theorem 3.2. First note that by passing to the completion we may assume that R is
a complete local ring. Then, by Cohen’s structure Theorem R is a factor ring of a regular local
ring ( T, n). In the light of Independence Theorem we may assume that Fi ( M ) ∼
= Fi ( M ), where
a
b
b := a ∩ T.
Using Remark 3.3, we have
dim M/aM = dim T − gradea ( M, T ).
So by Lemma 3.4 and Auslander-Buchsbaum Theorem we prove the claim as follows:
dim M/aM
= dim T − gradea ( M, T )
≥ dim T − gradea ( M, N ) − pdT N
= depth N − gradea ( M, N ).
Assume that b is an ideal of R of dimension at most 2 and d ≥ 3. By virtue of Theorem 3.2 and
3.1 we have
grade(b, D ( Had ( R))) + dim R/b ≥ depth( D ( Had ( R))) ≥ 2,
which is an upper bound for depth( D ( Had ( R))).
Remark 3.5. Let ( R, m) be a local ring of dimension d.
6
EGHBALI
(a) Let x be a parameter element of R. Consider 0 : R < x >=
S
n ≥1 0 : R
x n and following exact
sequence
0 → 0 : R < x >→ R → R/0 : R < x >→ 0.
One can see that dim(0 : R < x >) < d and dim R/(0 : R < x >) = d. Now apply Had (−) to
the above short exact sequence to get the following exact sequence
Had (0 : R < x >) → Had ( R) → Had ( R/0 : R < x >) → 0.
Hence, by Grothendieck’s vanishing Theorem, we have the isomorphism Had ( R) ∼
=
d
Ha ( R/0 : R < x >).
(b) Now we give conditions that Cohen-Macaulayness of D ( Had ( R)) fails. Let ( R, m) be a
complete local ring, d ≥ 3 and b be a one-dimensional ideal containing a regular element
x of R (part one guarantees the existence of a regular element of R). Then, it implies the
following short exact sequence
x
0 → R → R → R/xR → 0.
Thus we have the following long exact sequence
x
· · · → Had−1 ( R/xR) → Had ( R) → Had ( R) → 0.
By applying D (−) we get
x
0 → D ( Had ( R)) → D ( Had ( R)) → D ( Had−1( R/xR)) → · · · .
Now it yields that x is a D ( Had ( R))-regular element. So grade(b, D ( Had ( R))) > 0. Suppose that b does not contain a D ( Had ( R))/xD ( Had ( R))-regular element. Hence, [Bru-Her,
1.2.3 and 1.2.4] imply that Ext1R ( R/b, D ( Had ( R))) 6= 0 and then, grade(b, D ( Had ( R))) = 1.
Therefore, in the view of Theorem 3.1 and Theorem 3.2 we have
2 ≤ depth( D ( Had ( R)) ≤ grade(b, D ( Had ( R))) + dim R/b = 2.
It shows that with the above assumptions D ( Had ( R)) is not Cohen-Macaulay for d ≥ 3.
4.
TENSOR PRODUCTS
Throughout this section, ( R, m) is a local ring of dimension d. We assume that Had ( R) 6= 0
and Q := Qa ( R). In recent research there is an interest in endomorphism rings of certain local
cohomology modules Hai ( R). This was done in the case of i = d by the author and Schenzel
(see [E-Sch]) and some conditions for Cohen-Macaulayness of Hom R ( Had ( R), Had ( R)) have been
investigated. For the sake of completeness, we state the following result.
Remark 4.1. Let a be an ideal of a complete local ring ( R, m). Then,
depthR (Hom R ( Had ( R), Had ( R))) ≥ min{2, d}.
To this end note that Hom R ( Had ( R), Had ( R)) ∼
= HomR ( D ( Had ( R)), D ( Had ( R))) (see 2.1). Hence, the
claim is clear by virtue of [Bru-Her, Exer. 1.4.19] and Theorem 3.1.
The next part of our work is to investigate the tensor product of top local cohomology modules
and their Matlis duals.
TOP LOCAL COHOMOLOGY
7
Theorem 4.2. Let ( R, m) be a complete local ring of dimension d. Then, the following statements hold:
(a) Had ( R) ⊗ R Had ( R) is of finite length.
(b) Assume that R/Q is a Gorenstein ring. Then,
Had ( R) ⊗ R Had ( R) = 0.
(c) The ideal Q is contained in AnnR ( D ( Had ( R)) ⊗ R D ( Had ( R)) and the Krull dimension of
D ( Had ( R)) ⊗ R D ( Had ( R)) is equal to d.
(d) Suppose that pdR/Q (K R/Q ) < ∞ then, the following statements are equivalent :
(1) R/Q is a Cohen-Macaulay ring.
(2)
R/Q
Tori
( D ( Had ( R)), D ( Had ( R))) = 0 for all i ≥ 1
and
D ( Had ( R)) ⊗ R/Q D ( Had ( R)) is a Cohen-Macaulay R/Q-module.
(e) Assume that R/Q satisfies S2 situation. Then, there is an isomorphism
Had ( R) ⊗ R D ( Had ( R)) ∼
= ER/Q ( R/m).
Proof.
(a) As Had ( R) is an Artinian R-module, then it follows by [F-H, Proposition 6.1] that
Had ( R) ⊗ R Had ( R) is of finite length.
(b) Since Had ( R) ⊗ R Had ( R) is of finite length so by Hom −⊗-adjointness we have
Had ( R) ⊗ R Had ( R)
∼
= D ( D ( Had ( R) ⊗ R Had ( R)))
∼
= D (HomR ( Had ( R), D ( Had ( R)))
∼
= D (HomR ( Had ( R), K R/Q )).
As R/Q is a Gorenstein ring of positive dimension then, the later module is zero. To this
end note that the last isomorphism follows by Lemma 2.4.
(c) It is clear that
AnnR ( D ( Had ( R)) ⊗ R D ( Had ( R)) ⊇ AnnR ( D ( Had ( R))) = Q
(see 2.4) and
dim R ( D ( Had ( R)) ⊗ R D ( Had ( R)))
= dimR (Supp( D ( Had ( R)) ⊗ R D ( Had ( R))))
= dimR D ( Had ( R)) = d
(see 2.4).
(d) It should be noted that for a finite module M which has R-module and R/Q-module
structures then, it is clear that depthR M = depthR/Q M. Moreover, by virtue of 2.4 it
follows that D ( H d ( R)) ∼
= K R/Q .
a
In order to prove that 1 implies 2, take into account that if R/Q is a Cohen-Macaulay
ring, then D ( Had ( R)) is a Cohen-Macaulay module (see 2.4). Thus by Auslander-Buchsbaum
Theorem we deduce that K R/Q is a free R/Q-module. Hence,
ToriR/Q ( D ( Had ( R)), D ( Had ( R))) = 0, for all i ≥ 1.
8
EGHBALI
Now using [Hu-W, Lemma 2.2] we have
depth( D ( Had ( R)) ⊗ R/Q D ( Had ( R))) = d.
Thus D ( Had ( R)) ⊗ R/Q D ( Had ( R)) is a Cohen-Macaulay R/Q-module.
To show that (2) implies (1), we first use the hypothesis and [Hu-W, Lemma 2.2]. Then
we have
2 depth D ( Had ( R)) = depth R/Q + depth( D ( Had ( R)) ⊗ R/Q D ( Had ( R))).
(4.1)
As pdR/Q (K R/Q ) < ∞, then by Auslander-Buchsbaum Theorem it follows that
depth D ( Had ( R)) ≤ depth R/Q.
Since D ( Had ( R)) ⊗ R/Q D ( Had ( R)) is a Cohen-Macaulay R/Q-module, then in the light of
the equality (4.1) we prove the claim as follows:
depth R/Q
≥ depth( D ( Had ( R)) ⊗ R/Q D ( Had ( R)))
= dimR/Q ( D ( Had ( R)) ⊗ R/Q D ( Had ( R)))
= dim R/Q.
(e) As Had ( R) is an Artinian R-module so D ( Had ( R)) is finite. Hence,
Had ( R) ⊗ R D ( Had ( R))
∼
= D (HomR ( Had ( R), Had ( R)))
∼
= D ( R/Q)
∼
= ER/Q ( R/m).
To this end note that the first isomorphism follows by Lemma 2.2 and the second one
follows by [E-Sch, Theorem 4.2].
Let M and N be nonzero finitely generated modules over a local ring ( R, m). M and N satisfy
the depth formula provided that
depth M + depth N = depth R + depth( M ⊗ R N ).
A necessary condition for the depth formula to hold is that depth M + depth N ≥ depth R (see
[Hu-W, pp. 163]). As an application of 3.1 we deduce the following result:
Corollary 4.3. Let ( R, m) be a d-dimensional complete local ring. In case d = 3 or 4, then we have
depth( D ( Had ( R) ⊗ R D ( Had ( R)) ≥ 4 − depth R.
Corollary 4.4. Let ( R, m) be a complete local ring with d ≥ 4. Suppose that pdR/Q (K R/Q ) < ∞ and
ToriR/Q ( D ( Had ( R)), D ( Had ( R))) = 0 for all i > 0. Then,
(a) depth( D ( Had ( R)) ⊗ R/Q D ( Had ( R))) ≥ 4 − depth R/Q,
(b) pdR/Q ( D ( Had ( R)) ⊗ R/Q D ( Had ( R))) ≤ 2 depth R/Q − 4.
Proof. By Lemma 2.4 it follows that D ( Had ( R)) ∼
= K R/Q is a finitely generated module. Part (a) is
a consequence of 3.1 and [A, Theorem 1.2]. Part (b) follows by (a) and the Auslander-Buchsbaum
Theorem.
TOP LOCAL COHOMOLOGY
9
R EFERENCES
[A] M. Auslander, Modules over unramified regular local rings, Ill. J. Math. 5 (1961), 631–647.
[B] M. H. Bijan-Zadeh, A common generalization of local cohomology theories, Glasgow Math. J. 21, 173–181, (1980).
[Br-Sh] M. Brodmann and R.Y. Sharp, Local cohomology: an algebraic introduction with geometric applications, Cambridge
Univ. Press, 60, Cambridge, (1998).
[Bru-Her] W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge Univ. Press, (1993).
[E] M. Eghbali, On Artinianness of formal local cohomology, colocalization and coassociated primes. Math. Scand. , to appear.
arXiv:1011.4141 [math.AC].
[E-Sch] M. Eghbali and P. Schenzel, On an endomorphism ring of local cohomology , Comm. Alg. 40: 4295–4305, (2012).
[F-H] C. Faith and D. Herbera, Endomorphism rings and tensor products of linearly compact modules, Comm. Alg. 25(4), (1997),
1215–1255.
[G] A. Grothendieck, Local cohomology, Notes by R. Hartshorne, Lect. Notes in Math. 20, Springer, 1966.
[H] M. Hellus, Local Cohomology and Matlis duality, Habilitationsschrift, Leipzig (2006).
[H-Sch] M. Hellus and P. Schenzel, On cohomologically complete intersections, J. Algebra 320 (2008) 3733–3748.
[H-St] M. Hellus and J. Stückrad, On endomorphism rings of local cohomology modules, Proc. A.M.S Vol. 136(7), (2008), (2333–
2341).
[Hu-W] C. Huneke and R. Wiegand, Tensor products of modules, rigidity and local cohomology , Math. Scand. 81 (1997), 161–
183.
[R] J. Rotman, An Introduction to Homological Algebra, Academic Press, Orlando, FL, 1979.
[Sch1] P. Schenzel, On the use of local cohomology in algebra and geometry, In: J. Elias, J.M. Giral, R.M. Miró-Roig, S. Zarzuela
(Eds.), Six Lectures in Commutative Algebra, Proceed. Summer School on Commutative Algebra at Centre de Recerca
Matemtica, Progr. Math., Vol. 166, Birkhäuser, 1998, 241-292.
[Sch2] P. Schenzel, On formal local cohomology and connectedness, J. Algebra, 315(2), (2007), 894–923.
[Sch3] P. Schenzel, On endomorphism rings and dimensions of local cohomology modules , Proc. A.M.S Vol. 137(4), (2009), (1315–
1322).
M. E GHBALI , S CHOOL
OF
M ATHEMATICS , I NSTITUTE
B OX : 19395-5746, T EHRAN -I RAN .
E-mail address: [email protected]
FOR
R ESEARCH
IN
F UNDAMENTAL S CIENCES (IPM), P. O.
| 0math.AC
|
What Doubling Tricks Can and Can’t Do for Multi-Armed Bandits
Lilian Besson†
L ILIAN .B ESSON @C ENTRALE S UPELEC . FR
CentraleSupélec (campus of Rennes), IETR, SCEE Team,
Avenue de la Boulaie – CS 47601, F-35576 Cesson-Sévigné, France
arXiv:1803.06971v1 [stat.ML] 19 Mar 2018
Emilie Kaufmann
EMILIE . KAUFMANN @ UNIV - LILLE 1. FR
CNRS & Université de Lille, Inria SequeL team
UMR 9189 – CRIStAL, F-59000 Lille, France
Abstract
An online reinforcement learning algorithm is anytime if it does not need to know in advance the
horizon T of the experiment. A well-known technique to obtain an anytime algorithm from any nonanytime algorithm is the “Doubling Trick”. In the context of adversarial or stochastic multi-armed
bandits, the performance of an algorithm is measured by its regret, and we study two families of
sequences of growing horizons (geometric and exponential) to generalize previously known results
that certain doubling tricks can be used to conserve certain regret bounds. In a broad setting,√we
prove that a geometric doubling trick can be used to conserve (minimax) bounds in RT = O( T )
but cannot conserve (distribution-dependent) bounds in RT = O(log T ). We give insights as to
why exponential doubling tricks may be better,
as they conserve bounds in RT = O(log T ), and are
√
close to conserving bounds in RT = O( T ).
Keywords: Multi-Armed Bandits; Anytime Algorithms; Sequential Learning; Doubling Trick.
1. Introduction
Multi-Armed Bandit (MAB) problems are well-studied sequential decision making problems in
which an agent repeatedly chooses an action (the “arm” of a one-armed bandit) in order to maximize
some total reward (Robbins, 1952; Lai and Robbins, 1985). Initial motivation for their study came
from the modeling of clinical trials, as early as 1933 with the seminal work of Thompson (1933).
In this example, arms correspond to different treatments with unknown, random effect. Since then,
MAB models have been proved useful for many more applications, that range from cognitive radio
(Jouini et al., 2009) to online content optimization (e.g., news article recommendation (Li et al.,
2010), online advertising (Chapelle and Li, 2011), A/B Testing (Kaufmann et al., 2014; Yang et al.,
2017)), or portfolio optimization (Sani et al., 2012).
While the number of patients involved in a clinical study (and thus the number of treatments
to select) is often decided in advance, in other contexts the total number of decisions to make (the
horizon T ) is unknown. It may correspond to the total number of visitors of a website optimizing
its displays for a certain period of time, or to the number of attempted communications in a smart
radio device. In such cases, it is thus crucial to devise anytime algorithms, that is algorithms that do
no rely on the knowledge of this horizon T to sequentially select arms. A general way to turn any
base algorithm into an anytime algorithm is the use of the so-called Doubling Trick, first proposed
by Auer et al. (1995), that consists in repeatedly running the base algorithm with increasing horizons.
Motivated by the frequent use of this technique and the absence of a generic study of its effect
c 2018 L. Besson & E. Kaufmann.
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
on the algorithm’s efficiency, this paper investigates in details two families of doubling sequences
(geometric and exponential), and shows that the former should be avoided for stochastic problems.
More formally, a MAB model is a set of K arms, each arm k being associated to a (unknown)
reward stream (Yk,t )t∈N . Fix T a finite (possibly unknown) horizon. At each time step t ∈ {1, . . . , T }
an agent selects an arm A(t) ∈ {1, . . . , K} and receives as a reward the current value of the associated
reward stream, r(t) := YA(t),t . The agent’s decision strategy (or bandit algorithm) AT := (A(t), t ∈
{1, . . . , T }) is such that A(t) can only rely on the past observations A(1), r(1), . . . , A(t−1), r(t−1),
on external randomness and (possibly) on the knowledge of the horizon T . The objective of the agent
is to find an algorithm A that maximizes the expected cumulated rewards, where the expectation
is taken over the possible randomness used by the algorithm and the possible randomness in the
generation of the rewards stream. In the oblivious case, in which the reward streams are independent
of the algorithm’s choice, this is equivalent to minimizing the regret, defined as
" T
#
X
RT (AT ) := max E
Yk,t − YA(t),t .
(1)
k∈{1,...,K}
t=1
This quantity, referred to as pseudo-regret in Bubeck et al. (2012), quantifies the difference
between the expected cumulated reward of the best fixed action, and that of the strategy AT . For the
general adversarial bandit problem (Auer et al., 2002b), in which the rewards streams are arbitrary
(picked by an adversary), a worst-case lower bound has been given. It says that for
√ every algorithm,
there exists (stochastic) reward streams such that the regret is larger than (1/20)p KT (Auer et al.,
2002b). Besides, the EXP3 algorithm has been shown to have a regret of order KT log(K).
Much smaller regret may be obtained in stochastic MAB models, in which the reward stream
from each arm k is assumed to be i.i.d., from some (unknown) distribution νk , with mean µk . In
that case, various algorithms have been proposed with problem-dependent regret upper bounds
of the form C(ν) log(T ), where C(ν) is a constant that only depend on the arms distributions.
Different assumptions on the arms distributions lead to different problem-dependent constants. In
particular, under some parametric assumptions (e.g., Gaussian distributions, exponential families),
asymptotically optimal algorithms have been proposed and analyzed (e.g., kl-UCB (Cappé et al.,
2013) or Thompson sampling (Agrawal and Goyal, 2012; Kaufmann et al., 2012)), for which the
constant C(ν) obtained in the regret upper bound matches exactly that of the lower bound given
by Lai and Robbins (1985). Under the non-parametric assumption that the νk are bounded in
[0, 1],Pthe regret of the UCB1 algorithm (Auer et al., 2002a) is of the above form with C(ν) =
8 × k:µk >µ∗ (µ∗ − µk )−1 , where µ∗ = maxk µk is the mean of the best arm. Like in this last
example, all the available constants C(ν) become very large on “hard” instances, in which some
arms are very close
√ to the best arm. On such instances, C(ν) log(T ) may be much larger than the
worst-case (1/20) KT , and distribution-independent guarantees may actually be preferred.
The MOSS algorithm, proposed by Audibert and Bubeck (2009), is the first stochastic bandit
algorithm to enjoy a problem-dependent logarithmic
regret and to be optimal in a minimax sense, as
√
its regret is proved to be upper bounded by KT , for bandit models with rewards in [0, 1]. However
the corresponding constant C(ν) is proportional to K/∆min , where ∆min = mink (µ∗ − µk ) is the
minimal gap, which worsen the constant of UCB1. Another drawback of MOSS is that it is not
anytime. These two shortcoming have been overcame recently in two different works. On the one
hand, the MOSS-anytime algorithm (Degenne and Perchet, 2016) is minimax optimal and anytime,
but its problem-dependent regret does not improve that of MOSS. On the other hand, the kl-UCB++
2
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
algorithm (Ménard and Garivier, 2017) is simultaneously minimax optimal and asymptotically
optimal (i.e., it has the best problem-dependent constant C(ν)), but it is not anytime. A natural
question is thus to know whether a Doubling Trick could overcome this limitation.
This question is the starting point of our comprehensive study of the Doubling Trick: can a
single Doubling Trick be used to preserve both problem-dependent (logarithmic) regret and minimax
(square-root) regret? We answer this question in the negative, by showing that two different types of
Doubling Trick may actually be needed. In this paper, we investigate how algorithms enjoying regret
guarantees of the generic form
∀T ≥ 1, RT (AT ) ≤ c T γ (log(T ))δ + o(T γ log(T ))δ
(2)
may be turned into an anytime algorithm enjoying similar regret guarantees with an appropriate
Doubling Trick. This does not come for free, and we exhibit a “price of Doubling Trick”, that is a
constant factor larger than 1, referred to as a constant manipulative loss.
The rest of the paper is organized as follows. The Doubling Trick is formally defined in Section 2,
along with a generic tool for its analysis. In Section 3, we present upper and lower bounds on the
regret of algorithms to which a geometric Doubling Trick is applied. Section 4 investigates regret
guarantees that can be obtained for a “faster” exponential Doubling Trick. Experimental results are
then reported in Section 5. Complementary elements of proofs are deferred to the appendix.
2. Doubling Tricks
The Doubling Trick, denoted by DT , is a general procedure to convert a (possibly non-anytime)
algorithm into an anytime algorithm. It is formally stated below as Algorithm 1 and depends on a
non-decreasing diverging doubling sequence (Ti )i∈N (i.e., Ti → ∞ for i → ∞). DT fully restarts
the underlying algorithm A at the beginning of each new sequence (at t = Ti + 1), and run this
algorithm on a sequence of length (Ti − Ti−1 ).
Input: Bandit algorithm A, and doubling sequence (Ti )i∈N .
1
2
3
4
5
6
7
8
Let i = 0, and initialize algorithm A(0) = AT0 .
for t = 1, . . . , T − 1 do
if t > Ti then
// Next horizon Ti+1 from the sequence
Let i = i + 1.
Initialize algorithm A(i) = ATi −Ti−1 .
// Full restart
end
Play with A(i) : play arm A0 (t) := A(i) (t − Ti ), observe reward r(t) = YA0 (t),t .
end
Algorithm 1: The Generic Doubling Trick Algorithm, A0 = DT (A, (Ti )i∈N ).
Related work. The Doubling Trick is a well known idea in online learning, that can be traced
back to Auer et al. (1995). In the literature, the term Doubling Trick usually refers to the geometric
sequence Ti = 2i , in which the horizon is actually doubling, that was popularized by Cesa-Bianchi
and Lugosi (2006) in the context of adversarial bandits. Specific doubling tricks have also been used
for stochastic bandits, for example in the work of Auer and Ortner (2010), which uses the doubling
i
sequence Ti = 22 to turn the UCB-R algorithm into an anytime algorithm.
3
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
Elements of regret analysis. For a sequence (Ti )i∈N , with Ti ∈ N for all i, we denote T−1 = 0,
and T0 is always taken non-zero, T0 > 0 (i.e., T0 ∈ N∗ ). We only consider non-decreasing and
diverging sequences (that is, ∀i, Ti+1 ≥ Ti , and Ti → ∞ for i → ∞).
Definition 1 (Last Term LT )
For a non-decreasing diverging sequence (Ti )i∈N and T ∈ N, we can define LT (Ti )i∈N by
∀T ≥ 1, LT (Ti )i∈N := min {i ∈ N : Ti > T } .
(3)
It is simply denoted LT when there is no ambiguity (e.g., if the doubling sequence is chosen).
DT (A) reinitializes its underlying algorithm A at each time Ti , and in generality the total regret
is upper bounded by the regret on each sequence {Ti , . . . , Ti+1 − 1}. By considering the last partial
sequence {TLT −1 , . . . , T − 1}, this splitting can be used to get a generic upper bound (UB) by
taking into account a larger last sequence (up to TLT − 1). And for stochastic bandit models, the
i.i.d. hypothesis on the rewards streams makes the splitting on each sequence an equality, so we
can also get the lower bound (LB) by excluding the last partial sequence. Lemma 2 is proved in
Appendix A.1.
Lemma 2 (Regret Lower and Upper Bounds for DT )
For any bandit model and algorithm A and horizon T , one has the generic upper bound
LT
X
RT (DT (A, (Ti )i∈N )) ≤
RTi −Ti−1 (ATi −Ti−1 ).
(LB)
i=0
Under a stochastic bandit model, one has furthermore the lower bound
LT −1
RT (DT (A, (Ti )i∈N )) ≥
X
RTi −Ti−1 (ATi −Ti−1 ).
(UB)
i=0
As expected, the key to obtain regret guarantees for a Doubling Trick algorithm is to carefully
choose the doubling sequence (Ti )i∈N . Empirically, one can verify that sequences with slow growth
gives terrible results, and for example using an arithmetic progression typically gives a linear regret.
Building on this result, we will prove that if A satisfies a certain regret bound (RT = O(T γ ),
O((log T )δ ), or O(T γ (log T )δ )) then an appropriate anytime version of A with a certain doubling
trick can conserve the regret bound, with an explicit constant multiplicative loss ` > 1. In this paper,
we study in depth two families of sequences: first geometric and then exponential growths.
3. What the Geometric Doubling Trick Can and Can’t Do
We define geometric
doubling sequences, and prove that they can be used to conserve
bounds in
O T γ (log T )δ with γ > 0 but cannot be used to conserve bounds in O (log T )δ .
Definition 3 (Geometric Growth) For b ∈ R, b > 1 and T0 ∈ N∗ , the sequence defined by
Ti = bT0 bi c is non-decreasing and diverging, and satisfies
T
∀T < T0 , LT = 0, and ∀T ≥ T0 , LT = logb
,
(4)
T0
∀i > 0, T0 (b − 1)bi−1 − 1 ≤ Ti − Ti−1 ≤ 1 + T0 (b − 1)bi−1 .
Asymptotically for i and T → ∞, Ti = O bi and LT ∼ logb (T ) = O(log T ).
4
(5)
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
3.1. Conserving a Regret Upper Bound with Geometric Horizons
√
A geometric doubling sequence allows to conserve a minimax bound (i.e., RT = O( T )). It was
suggested, for instance, in (Cesa-Bianchi and Lugosi, 2006, Ex.2.9).
√ We generalize this result in the
following theorem, proved in Appendix A.2, by extending it from T bounds to bounds of the form
T γ (log T )δ for any 0 < γ < 1 and δ ≥ 0. Note that no distinction is done on the case δ = 0 neither
in the expression of the constant loss, nor in the proof.
δ
Theorem 4 If an algorithm A satisfies RT (AT ) ≤ c T γ (log
T ) + f (T ), for 0 < γ < 1, δ ≥ 0
γ
δ
and for c > 0, and an increasing function f (t) = o t (log t) (at t → ∞), then the anytime version
A0 := DT (A, (Ti )i∈N ) with the geometric sequence (Ti )i∈N of parameters T0 ∈ N∗ , b > 1 (i.e.,
Ti = bT0 bi c) with the condition T0 (b − 1) > 1 if δ > 0, satisfies,
RT (A0 ) ≤ `(γ, δ, T0 , b) c T γ (log T )δ + g(T ),
with a increasing function g(t) = o tγ (log t)δ , and a constant loss `(γ, δ, T0 , b) > 1,
`(γ, δ, T0 , b) :=
log(T0 (b − 1) + 1)
log(T0 (b − 1))
δ
×
bγ (b − 1)γ
.
bγ − 1
(6)
(7)
For a fixed γ and δ, minimizing `(γ, δ, T0 , b) does not always give a unique solution. On the one
γ
γ
hand, if γ & 0.384, there is a unique solution b∗ (γ) > 1 minimizing the b b(b−1)
term, solution of
γ −1
γ+1
b
− 2b + 1 = 0, but without a closed form if γ is unknown. On the other hand, for any γ, the
term depending on δ tends quickly to 1 when T0 increases.
Practical considerations. Empirically, when γ and δ are fixed and known, there is no need to
γ
γ
minimize ` jointly. It can be minimized separately by first minimizing b b(b−1)
γ −1 , that is by solving
bγ+1 − 2b + 1 = 0 numerically (e.g., with Newton’s method), and then by taking T0 large enough so
that the other term is close enough to 1.
√
√
For the usual case of γ = 1/2 and δ = 0 (i.e., bounds in T ), the optimal choice of b is 3+2 5
leading to ` ' 3.33, and the usual choice of b = 2 gives ` ' 3.41 (see Corollary 10 in appendix).
Any large enough T0 gives similar performance, and empirically T0 K is preferred, as most
algorithms explore each arm once in their first steps (e.g., T0 = 200 for K = 9 for our experiments).
3.2. A Regret Lower Bound with Geometric Horizons
We observe that the constant loss in Eq. (7) from the previous Theorem 4 blows up when γ goes to
zero, giving the intuition that no geometric doubling trick could be used to preserve a logarithmic
bound (i.e., with γ = 0, δ > 0). This is confirmed by the lower bound given below.
Theorem 5 For stochastic models, if A satisfies RT (AT ) ≥ c (log T )δ , for c > 0 and δ > 0,
then the anytime version A0 := DT (A, (Ti )i∈N ) with the geometric sequence (Ti )i∈N of parameters
T0 ∈ N∗ , b > 1 (i.e., Ti = bT0 bi c) satisfies this lower bound for a certain constant c0 > 0,
∀T ≥ 1, LT ≥ 2 =⇒ RT (A0 ) ≥ c0 (log T )δ+1 .
(8)
This implies that RT (A0 ) = Ω((log T )δ+1 ), which proves that a geometric sequence cannot be used
to conserve a logarithmic regret bound.
5
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
Theorem 5 implies that a geometric sequence cannot be used to conserve a finite-horizon
lower bound like RT (AT ) ≥ c log(T ). A complementary lower bound, stated as Theorem
11 in
√
Appendix B, shows that if the regret is lower bounded at finite horizon by RT (AT ) ≥ c T , then a
comparable lower bound for the Doubling Trick algorithm DT (A), possibly with a larger constant.
This special case (δ = 1) is indeed the most interesting, as in the stochastic case the regret of any
uniformly efficient algorithm is at least logarithmic (Lai and Robbins (1985)), and efficient algorithm
with logarithmic regret have been exhibited. If RT (AT )/log T is bounded, then using a geometric
sequence in the doubling trick algorithm is a bad idea as it guarantees a blow up in the regret, that
is RT (DT (A, (Ti )i∈N )) = Ω((log T )2 ). This result is the reason we need to consider successive
horizons growing faster than a geometric sequence (i.e., such that log(Ti ) i), like the exponential
sequence, which is studied in Section 4.
3.3. Proof of Theorem 5
Let A0 := DT (A, (Ti )i∈N ) and consider a fixed stochastic bandit problem. The lower bound (LB)
from Lemma 2 gives
0
RT (A ) ≥
LX
T −1
RTi −Ti−1 (AT =Ti −Ti−1 )
i=0
We bound Ti − Ti−1 ≥ T0 (b − 1)bi−1 − 1 for any i > 0, thanks to Definition 3, and we can use the
hypothesis on A for each regret term.
≥
LX
T −1
δ
c(log(Ti − Ti−1 )) ≥ c
i=0
LX
T −2
=c
LX
T −1
δ
log T0 (b − 1)bi−1 − 1
i=1
log T0 (b − 1)bi − 1
δ
(with i := i − 1)
i=0
Let xi := T0 (b − 1)bi > 0. If we have T0 (b − 1) > 1 (see below (♣) in Page 7 for a discussion on
0 (b−1)−1)
the other case), then Lemma 15 (Eq. (26)) gives log(xi − 1) ≥ log(T
log(T0 (b−1)) log(xi ) as xi > 1. For
lower bounds, there is no need to handle the constants tightly, and we have xi ≥ bi by hypothesis, so
δ
0 (b−1)−1)
let call this constant c0 = c log(T
> 0, and thus it simplifies to
log(T0 (b−1))
≥ c0
LX
T −2
δ
log(bi )
i=0
δ
P T −2
A sum-integral minoration for the increasing function t 7→ tδ (as δ > 0) gives L
log(bi ) =
i=0
R
δ
P T −2 δ
δ LT −2 tδ dt = (log b) (L − 2)δ+1 (if L ≥ 2), and so
(log b)δ L
T
T
i=1 i ≥ (log b) 0
δ+1
RT (A0 ) ≥ c0
(log b)δ
(LT − 2)δ+1
δ+1
6
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
For the geometric sequence, we know that LT ≥ logb
T
T0
≥ logb (T ), and logb (T ) − 2 ∼ logb (T )
c00
at T → ∞ so there exists a constant 0 < < 1 such that LT − 2 ≥ c00 logb (T ) for T large enough
2
(≥ b−1
), and such that LT ≥ 2. And thus we just proved that there is a constant c000 > 0 such that
RT (A0 ) ≥ c000 (log T )δ+1 =: g(T ).
So this proves that for T large enough, RT (A0 ) ≥ g(T ) with g(T ) = O (log T )δ+1 , and so
RT (A0 ) = Ω((log T )δ+1 ), which also implies that RT (A0 ) cannot be a O (log T )δ .
(♣) If we do not have the hypothesis T0 (b − 1) > 1, the same proof could be done, by observing
1
that from i ≥ i0 large enough, we have xi ≥ bi−i0 (as soon as bi0 ≥ T0 (b−1)
> 0, i.e., i0 ≥
d− logb (T0 (b − 1))e ≥ 1), and so the same arguments can be used, to obtain a sum from i = i0 + 1
instead of from i = 1. For a fixed i0 , we also have LT − 2 − i0 ≥ c00 log(T ) for a (small enough)
constant c00 , and thus we obtain the same result.
4. What Can the Exponential Doubling Trick Do?
We define exponential
doubling sequences, and prove that they can be used to conserve bounds in
δ
O (log T ) , unlike the previously studied geometric sequences. Furthermore,
we provide elements
showing that they may also conserve bounds in O(T γ ) or O T γ (log T )δ .
Definition 6 (Exponential Growth) For a, b ∈ R, a, b > 1 and T0 ∈ N∗ , if τ := Ta0 ∈ R, ≥ 0,
i
then the sequence defined by Ti := bτ ab c is non-decreasing and diverging, and satisfies
T
.
(9)
∀T < T0 , LT = 0, and ∀T ≥ T0 , LT = logb loga
τ
i
Asymptotically for i and T → ∞, Ti = O ab and LT ∼ logb (loga ( Tτ )) = O(log log T ).
4.1. Conserving a Regret Upper Bound with Exponential Horizons
An exponential doubling sequence allows to conserve a problem-dependent bound on regret (i.e.,
RT = O(log T )). This was already used in particular cases by Auer and Ortner (2010) and more
recently by Liau et al. (2018). We generalize this result in the following theorem.
Theorem 7 If an algorithm A satisfies RT (AT ) ≤ c T γ (log
T )δ + f (T ), for 0 ≤ γ < 1, δ ≥ 0,
and for c > 0, and an increasing function f (t) = o tγ (log t)δ (at t → ∞), then the anytime version
A0 := DT (A, (Ti )i∈N ) with the exponential sequence (Ti )i∈N of parameters T0 ∈ N∗ , a, b > 1 (i.e.,
i
Ti = b Ta0 ab c), satisfies the following inequality,
γ
RT (A0 ) ≤ `(γ, δ, T0 , a, b) c T b (log T )δ + g(T ).
(10)
with an increasing function g(t) = o (tb )γ (log t)δ , and a constant loss `(γ, δ, T0 , a, b) > 0,
a (b−1)γ b2δ > 0
if δ > 0
T0
bδ −1
`(γ, δ, T0 , a, b) :=
(11)
1
1 +
>
1
if
δ
=
0
(log(a))(log(bγ ))
7
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
This result first shows that an exponential doubling trick can preserve a logarithmic regret bound
(O(log(T )), which corresponds to γ = 0 and δ = 1), with a multiplicative
constant loss ` ≥ 4. It can
γ
δ
further be applied to bounds of the generic form RT = O T (log T ) , but with a significant loss as
T γ becomes T bγ , additionally to the constant multiplicative loss ` > 0. However, it is important to
notice that for γ > 0, the constant ` can be made arbitrarily small (with a large enough first step T0 ).
This observation is encouraging, and let the authors think that a tighter upper bound could be proved.
Remark 8 An interesting particular case of Theorem 7 is the following (γ = 0, δ = 1 and f (t) = 0).
i
RT (AT ) ≤ c log(T ) =⇒ RT (DT (A, (bT0b c)i∈N )) ≤
b2
c log(T ).
b−1
(12)
In this upper bound, the optimal choice of b is b = 2, which yields a constant multiplicative loss of
`(γ = 0, b) = 4. It can be observed that this loss is twice smaller as the loss of 8.0625 obtained
by (Auer and Ortner, 2010, Sec.4)1 .
4.2. A Regret Lower Bound with Exponential Horizons
Assuming the upper bound of Theorem 7 obtained for γ > 0 are tight would lead√to think that
exponential doubling tricks cannot preserve minimax regret bounds of the form O( T )). If true,
such a conjecture would need to be supported by a lower bound (a counterpart of Theorem 5).
Theorem 9 provides such a lower bound, but as discussed below, combined with Theorem 7, this
result rather advocates the use of exponential doubling tricks. Theorem 9 is proved in Appendix A.3.
Theorem 9 For stochastic models, if A satisfies RT (AT ) ≥ c T γ , for c > 0 and 0 < γ ≤ 1, then
the anytime version A0 := DT (A, (Ti )i∈N ) with the exponential sequence (Ti )i∈N of parameters
i
T0 ∈ N∗ , a > 1, b > 1 (i.e., Ti = b Ta0 ab c), satisfies this lower bound for a certain constant c0 > 0,
∀T ≥ 1,
RT (A0 ) ≥ c0
1
Tb
γ
.
(13)
We already saw that any exponential doubling trick can conserve logarithmic problem-dependent
regret bounds. If we could take b → 1 in the two previous Theorems 7 and 9, both results would
match and prove that there exists an exponential doubling trick that can also be used to conserve
minimax regret bounds. This argument is not so easy to formulate, as b cannot depend on T , but
it supports our belief that exponential doubling tricks are good candidates for (asymptotically)
preserving both problem-dependent and minimax regret bounds.
4.3. Proof of Theorem 7
Let A0 := DT (A, (Ti )i∈N ), and consider a fixed bandit problem. We first consider the harder case
of δ > 0, see below in Page 10 in (♠) for the other case. The lower bound (LB) from Lemma 2 gives
RT (A0 ) ≤
LT
X
RTi −Ti−1 (AT =Ti −Ti−1 )
i=0
1. In Auer and Ortner (2010), the authors obtained a loss of 258/32 = 8.0625 ≥ 8, as the ratio between the constants
for the log(T ) terms, respectively 258 in Th.4.1 and 32 in Th.3.1.
8
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
i
We bound naively2 Ti − Ti−1 ≤ Ti ≤ Ta0 ab , and we can use the hypothesis on A for each regret
term, as f and t 7→ ctγ (log t)δ are non-decreasing for t ≥ 1 (by hypothesis for f and by Lemma 16).
≤
LT
X
f (Ti ) + c
LT
X
i=0
γ
δ
(Ti ) (log (Ti )) ≤ g1 (T ) + c
i=0
LT
X
T0
i=0
γ
δ
T0 bi
log
a
a
a
a
bi
PLT
The
first part is denoted g1 (T ) := i=0 f (Ti ) and is dealt with Lemma 17: the sum of f (Ti ) is
P
γ
LT
δ
γ
δ by hypothesis, and this sum of T γ (log(T ))δ is
ao
i
i=0 Ti (log(Ti )) , as f (t) = o t (log t)
i
bγ
δ
bγ
δ
proved below to be bounded by T (log(T )) . So g1 (T ) = o T (log T ) . The second part is
γ PLT bi γ T bi δ
c Ta0
. Define log+ (x) := max(log(x), 0) ≥ 0, so whether Ta0 ≤ 1
log a0 a
i=0 a
i
i
or > 1, we always have log Ta0 ab ≤ log+ Ta0 +log ab . Then we can use Lemma 14 (Eq. (23))
δ
δ
δ
i
+ (log(a))δ bi
to distribute the power on δ (as it is < 1). So log Ta0 ab
≤ log+ Ta0
with the convention that 0δ = 0 (even if δ = 0), and so this gives
#
γ "
δ X
LT
LT
X
T0
δ
+ T0
bi γ
bi γ
i δ
≤ g1 (T ) + c
log
(a ) + (log(a))
(a ) b
a
a
i=0
i=0
If γ = 0 then
the first sum is just LT + 1 = O(log(log(T ))) which can be included in g1 (T ) =
o (log T )δ (still increasing), and so only the second sum has to be bounded, and a geometric
P T i δ
δ
sum gives L
≤ bδb−1 (bLT )δ . But if γ > 0, we can naively bound the first sum by
i=0 b
PLT bi γ
bLT )γ Observe that abLT = (abLT −1 )b ≤ (a T )b . So abLT = O T b
i=0 (a ) ≤ (LT + 1)(a
T0
and LT + 1 = O(log(log(T ))), thus the first sum is a O T bγ log(log(T )) = o T bγ (logT )δ (as
δ > 0). In both cases, the first sum can be included in g2 (T ) which is still a o T bγ (log T )δ Another
P T b i γ i δ
P T i δ
δ
L
geometric sum bounds the second sum by L
b ≤ (ab T )γ L
≤ bδb−1 (bLT )δ .
i=0 (a )
i=0 b
≤ g1 (T ) + c1 (ab
LT
)γ (bLT )δ
T0 γ bδ
a
bδ −1
We identify a constant multiplicative loss c1 := c
depends on LT is (ab
LT
(log a)δ > 0. The only term left which
)γ (bLT )δ , and it can be bounded by using bLT = bbLT −1 ≤ b loga (a TT0 ) =
L
T
b T ≤ (a T )b . The constant part of
b + b log+
a ( T0 ) ≤ b + b loga (T ) (as T ≥ 1), and again with a
T0
bLT also gives a O T bγ term, that can be included in g(T ) := g2 (T ) + (a TT0 )bγ which is still a
o T bγ (log T )δ , and is still increasing as sum of increasing functions. So we can focus on the last
term, and we obtain
#γ
δ " b
b
a
b
≤ g(T ) + c1
T
(log T )δ
log(a)
T0
=⇒ RT (A0 ) ≤ g(T ) + `(γ, δ, T0 , a, b) cT bγ (log T )δ with an increasing g(t) = o tbγ (log t)δ .
i−1
i−1
2. Here, using the more subtle bound Ti − Ti ≤ Ta0 ab (αb ) + 1, with α = ab−1 , from Definition 6, does not seem
to help as it becomes too complicated to handle clearly in the log terms.
9
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
So the constant multiplicative loss ` depends on γ and δ as well as on T0 , a and b and is
(b−1)γ
b2δ
a
× δ
`(γ, δ, T0 , a, b) :=
> 0, if δ > 0.
T0
b −1
(14)
If T0 = a, the loss `(γ, δ, T0 , a, b) is minimal at b∗ (δ) = 21/δ > 1 and for a minimal value of
minb>1 `(γ, δ, T0 , a, b) = 4 (for any δ and γ). Finally, the a/T0 part tends to 0 if T0 → ∞ so the
loss can be made as small as we want, simply by choosing a T0 large enough (but constant w.r.t. T ).
(♠) Now for the other case of δ = 0, we can start similarly, but instead of bounding naively
PLT bi γ
PLT bi γ
bLT )γ we use Lemma 13 to get a more subtle bound:
i=0 (a ) by (LT + 1)(a
i=0 (a ) ≤
LT γ
1
γ
b
a + (1 + (log(a))(log(bγ )) )(a ) . The constant term gets included in g(T ), and for the non-constant
LT
part, (ab
)γ is handled similarly. Finally we obtain the loss
`(γ, 0, T0 , a, b) := 1 +
1
> 1.
(log(a))(log(bγ ))
(15)
5. Numerical Experiments
We illustrate here the practical cost of using Doubling Trick, for two interesting non-anytime
algorithms that have recently been proposed in the literature: Approximated Finite-Horizon Gittins
indexes, that we refer to as AFHG, by Lattimore (2016) (for Gaussian bandits with known variance)
and kl-UCB++ by Ménard and Garivier (2017) (for Bernoulli bandits).
We first provide some details on these two algorithms, and then illustrate the behavior of Doubling
Tricks applied to these algorithms with different doubling sequences.
5.1. Two Index-Based Algorithms
P
We denotePby Xk (t) :=
s<t YA(s),s 1(A(s) = k) the accumulated rewards from arm k, and
Nk (t) := s<t 1(A(s) = k) the number of times arm k was sampled. Both algorithms A assume
to know the horizon T . They compute an index IkA (t) ∈ R for each arm k ∈ {1, . . . , K} at
each time step t ∈ {1, . . . , T }, and use the indexes to choose the arm with highest index, i.e.,
A(t) := arg maxk∈{1,...,K} Ik (t) (ties are broken uniformly at random).
• The algorithm AFHG can be applied for Gaussian bandits with variance V (= 1 for our
experiments). Let m(T, t) = T − t + 1 ≥ 1, and let
v
u
u
Xk (t) u V
m(T, t)
.
IkAFHG (t) :=
+t
log
(16)
Nk (t)
Nk (t)
N (t) log1/2 m(T,t)
k
Nk (t)
• The algorithm kl-UCB++ can be applied for bounded rewards in [0, 1], and in particular for
Bernoulli bandits. The binary Kullback-Leibler divergence is kl(x, y) := x log(x/y) + (1 −
+
x) log((1 − x)/(1 − y)) (for
0 < x, y < 1), andletlog (x) := max(0, log(x)) ≥ 0. Let the
2
T
T
function g(n, T ) := log+ Kn
(1 + log+ ( Kn
) ) for n ≤ T , and finally let
Xk (t)
g(Nk (t), T )
kl-UCB++
Ik
(t) := sup q : kl
,q ≤
.
(17)
Nk (t)
Nk (t)
q∈[0,1]
10
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
5.2. Experiments
We present some results from numerical experiments on Bernoulli and Gaussian bandits. More results
are presented in Appendix E. We present in pages 12 and 13 results for K = 9 arms and horizon
T = 45678 (to ensure that no choice of sequence were lucky and had TLT −1 = T or too close to it).
We ran n = 1000 repetitions of the random experiment, either on the same “easy” bandit problem
µ (with evenly spaced means), or on n different random instances µ sampled uniformly in [0, 1]K ,
and we plot theP
average regret on n simulations. The black line without markers is the (asymptotic)
lower bound in k6=k∗ (kl(µk , µ∗ ))−1 log T , from Lai and Robbins (1985). We consider kl-UCB++
for Bernoulli bandits (Figures 2, 3) or AFHG for Gaussian bandits (Figures 4, 5),
Each doubling trick algorithm uses the same T0 = 200 as a first guess for the horizon. We include
both the non-anytime version that knows the horizon T , and different anytime versions to compare
the choice of doubling trick. To compare against an algorithm that does not need the horizon, we also
include kl-UCB (Cappé et al., 2013) as a baseline for Bernoulli bandits and for Gaussian bandits (in
the Gaussian version, the divergence used is kl(x, y) = (x − y)2 /2, and the algorithm is referred to
as UCB). We consider are a geometric doubling sequence with b = 2, and two different exponential
doubling sequences: the “classical” b = 2 and a “slower” one with b = 1.1. Both use a = T0 = 200,
and the last one is using a = 2, b = 2. Despite what was proved theoretically in Theorem 7, using
a = T0 and a large enough T0 improves regarding to using a = 2 and a leading (T0 /a) factor.
Another version of the Doubling Trick with “no restart”, denoted DT no-restart , is presented in
Appendix C, but it is only an heuristic and cannot be applied to any algorithm A. Algorithm 2 can be
applied to kl-UCB++ or AFHG for instance, as they use T just as a numerical parameter (see Eqs. 16
and 17), but its first limitation is that it cannot be applied to DMED+ (Honda and Takemura, 2010)
or EXP3++ (Seldin and Lugosi, 2017), or any algorithms based on arm eliminations, for example.
A second limitation is the difficulty to analyze this “no restart” variant, due to the unpredictable
effect on regret of giving non-uniform prior information to the underlying algorithm A on each
successive sequence. An interesting future work would be to analyze it, either in general or for a
specific algorithm like kl-UCB++ . Despite its limitations, this heuristic exhibits as expected better
empirical performance than DT , as can be observed in Appendix E.
6. Conclusion
We formalized and studied the well-known “Doubling Trick” for generic multi-armed bandit problems, that is used to automatically obtain an anytime algorithm from any non-anytime algorithm.
Our results are summarized in
√ Table 1. We show that a geometric doubling can be used to conserve
minimax regret bounds (in T ), with a constant loss (typically ≥ 3.33), but cannot be used to
conserve problem-dependent bounds (in log T ), for which a faster doubling sequence is needed. An
exponential doubling sequence can conserve logarithmic regret bounds also with a constant loss, but it
is still an open question to know if minimax bounds can be obtained for this faster growing sequence.
Partial results of both a lower and an upper bound, for bounds of the generic form T γ (log T )δ , let
use believe in a positive answer.
It is still an open problem to know if an anytime algorithm can be both asymptotically optimal
for the problem-dependent
regret (i.e., with the exact constant) and optimal in a minimax regret
√
(i.e., have a KT regret), but we believe that using a doubling trick on non-anytime algorithms
like kl-UCB++ cannot be the solution. We showed that it cannot work with a geometric doubling
sequence, and conjecture that exponential doubling trick would never bring the right constant either.
11
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
i
Bound \ Doubling
Geometric, Ti = bT0 bi c
Exponential, Ti = b Ta0 ab c
(log T )δ
× Known to fail
RT (DT ) ≥ c0 (log T )1+δ if
RT (AT ) ≥ c(log T )δ .
(Theorem 5)
X Known to work, with loss
2δ
`(δ, b) = bbδ −1 > 1.
(Theorem 7)
Tγ
X Known toγ work,γ with loss
`(γ, b) = b b(b−1)
> 1.
γ −1
(Theorem 4)
T γ (log T )δ
for both
γ > 0, δ > 0
X Known to work, with loss
`(γ, δ, T0 , b) =
δ γ
log(T0 (b−1)+1)
b (b−1)γ
bγ −1 > 1.
log(T0 (b−1))
(Theorem 4 )
? Partial, best known bound is
1
c00 (T b )γ ≤ RT (DT ≤ `c(T b )γ with a
loss ` > 1, if c0 T γ ≤ RT (AT ) ≤ cT γ .
(Theorems 7, 9)
? Partial, best known bound is
RT (DT ) ≤ `c(T b )γ if
RT (AT ) ≤ cT γ , with a loss ` → 0 for
T0 → ∞.
(Theorem 7)
Figure 1: Summary of known positive X and negative × and partial results ?.
Cumulated regrets for different bandit algorithms, averaged 1000 times
9 arms: Bayesian MAB, Bernoulli with means on [0, 1]
400
Cumulated regret Rt = tµ ∗ −
s=0 k=1
9
t−1 X
X
µk
1000 [Tk (s)]
500
KLUCB
KLUCB+ + (T = 45678)
DT(Ti = 200 × 2i )[KLUCB+ + ]
DT(Ti = 2002i )[KLUCB+ + ]
DT(Ti = 2001.1i )[KLUCB+ + ]
DT(Ti = (200/2)22i )[KLUCB+ + ]
Lai & Robbins lower bound = 3.74 log(t)
300
200
100
0
0
10000
20000
30000
Time steps t = 0. . . T − 1, horizon T = 45678
40000
Figure 2: Regret for DT , for K = 9 Bernoulli arms, horizon T = 45678, n = 1000 repetitions and
µ taken uniformly in [0, 1]K . Geometric doubling (b = 2) and slow exponential doubling (b = 1.1)
are too slow, and short first sequences make the regret blow up in the beginning of the experiment.
At t = 40000 we see clearly the effect of a new sequence for the best doubling trick (Ti = 200 × 2i ).
As expected, kl-UCB++ outperforms kl-UCB, and if the doubling sequence is growing fast enough
then DT (kl-UCB++ ) can perform as well as kl-UCB++ (see for t < 40000).
12
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
Cumulated regrets for different bandit algorithms, averaged 1000 times
9 arms: [B(0.1), B(0.2), B(0.3), B(0.4), B(0.5), B(0.6), B(0.7), B(0.8), B(0.9) ∗ ]
400
300
Cumulated regret Rt = tµ ∗ −
s=0 k=1
9
t−1 X
X
µk
1000 [Tk (s)]
350
KLUCB
KLUCB+ + (T = 45678)
DT(Ti = 200 × 2i )[KLUCB+ + ]
DT(Ti = 2002i )[KLUCB+ + ]
DT(Ti = 2001.1i )[KLUCB+ + ]
DT(Ti = (200/2)22i )[KLUCB+ + ]
Lai & Robbins lower bound = 7.52 log(t)
250
200
150
100
50
0
0
10000
20000
30000
Time steps t = 0. . . T − 1, horizon T = 45678
40000
Figure 3: Similarly but for µ evenly spaced in [0, 1]K ({0.1, . . . , 0.9}). Both kl-UCB and kl-UCB++
are very efficient on “easy” problems like this one, and we can check visually that they match the
lower bound from Lai and Robbins (1985). As before we check that slow doubling are too slow to
give reasonable performance.
Cumulated regrets for different bandit algorithms, averaged 1000 times
9 arms: Bayesian MAB, Gaussian with uniform means on [−5, 5]
2000
Cumulated regret Rt = tµ ∗ −
s=0 k=1
9
t−1 X
X
µk
1000 [Tk (s)]
2500
UCB
ApprFHG(T = 45678)
DT(Ti = 200 × 2i )[ApprFHG]
DT(Ti = 2002i )[ApprFHG]
DT(Ti = 2001.1i )[ApprFHG]
DT(Ti = (200/2)22i )[ApprFHG]
Lai & Robbins lower bound = 4.06 log(t)
1500
1000
500
0
0
10000
20000
30000
Time steps t = 0. . . T − 1, horizon T = 45678
40000
Figure 4: Regret for K = 9 Gaussian arms N (µ, 1), horizon T = 45678, n = 1000 repetitions and
µ taken uniformly in [−5, 5]K and variance V = 1. On “hard” problems like this one, both UCB
and AFHG perform similarly and poorly w.r.t. to the lower bound from Lai and Robbins (1985). As
before we check that geometric doubling (b = 2) and slow exponential doubling (b = 1.1) are too
slow, but a fast enough doubling sequence does give reasonable performance for the anytime AFHG
obtained by Doubling Trick.
13
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
Acknowledgments
This work is supported by the French National Research Agency (ANR), under the project BADASS
(grant coded: N ANR-16-CE40-0002), by the French Ministry of Higher Education and Research
(MENESR) and ENS Paris-Saclay.
References
S. Agrawal and N. Goyal. Analysis of Thompson sampling for the Multi-Armed Bandit problem. In
Conference On Learning Theory. PMLR, 2012.
J-Y. Audibert and S. Bubeck. Minimax policies for adversarial and stochastic bandits. In Conference
on Learning Theory, pages 217–226. PMLR, 2009.
P. Auer and R. Ortner. UCB Revisited: Improved Regret Bounds For The Stochastic Multi-Armed
Bandit Problem. Periodica Mathematica Hungarica, 61(1-2):55–65, 2010.
P. Auer, N. Cesa-Bianchi, Y. Freund, and R. Schapire. Gambling in a Rigged Casino: The Adversarial
Multi-Armed Bandit Problem. In Annual Symposium on Foundations of Computer Science, pages
322–331. IEEE, 1995.
P. Auer, N. Cesa-Bianchi, and P. Fischer. Finite-time Analysis of the Multi-armed Bandit Problem.
Machine Learning, 47(2):235–256, 2002a.
P. Auer, N. Cesa-Bianchi, Y. Freund, and R. Schapire. The Nonstochastic Multiarmed Bandit
Problem. SIAM journal on computing, 32(1):48–77, 2002b.
S. Bubeck, N. Cesa-Bianchi, et al. Regret Analysis of Stochastic and Non-Stochastic Multi-Armed
Bandit Problems. Foundations and Trends R in Machine Learning, 5(1), 2012.
O. Cappé, A. Garivier, O-A. Maillard, R. Munos, and G. Stoltz. Kullback-Leibler upper confidence
bounds for optimal sequential allocation. Annals of Statistics, 41(3):1516–1541, 2013.
N. Cesa-Bianchi and G. Lugosi. Prediction, Learning, and Games. Cambridge University Press,
2006.
O. Chapelle and L. Li. An Empirical Evaluation of Thompson Sampling. In Advances in Neural
Information Processing Systems, pages 2249–2257. Curran Associates, Inc., 2011.
R. Degenne and V. Perchet. Anytime Optimal Algorithms In Stochastic Multi Armed Bandits. In
International Conference on Machine Learning, pages 1587–1595, 2016.
A. Garivier, E. Kaufmann, and T. Lattimore. On Explore-Then-Commit Strategies. volume 29 of
Advances in Neural Information Processing Systems (NIPS), Barcelona, Spain, 2016.
J. Honda and A. Takemura. An Asymptotically Optimal Bandit Algorithm for Bounded Support
Models. In Conference on Learning Theory, pages 67–79. PMLR, 2010.
W. Jouini, D. Ernst, C. Moy, and J. Palicot. Multi-Armed Bandit Based Policies for Cognitive Radio’s
Decision Making Issues. In International Conference Signals, Circuits and Systems. IEEE, 2009.
14
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
E. Kaufmann, N. Korda, and R. Munos. Thompson Sampling: an Asymptotically Optimal Finite-Time
Analysis, pages 199–213. PMLR, 2012.
E. Kaufmann, O. Cappé, and A. Garivier. On the Complexity of A/B Testing. In Conference on
Learning Theory, pages 461–481. PMLR, 2014.
T. L. Lai and H. Robbins. Asymptotically Efficient Adaptive Allocation Rules. Advances in Applied
Mathematics, 6(1):4–22, 1985.
T. Lattimore. Regret Analysis Of The Finite Horizon Gittins Index Strategy For Multi Armed Bandits.
In Conference on Learning Theory, pages 1214–1245. PMLR, 2016.
L. Li, W. Chu, J. Langford, and R. E. Schapire. A Contextual-Bandit Approach to Personalized News
Article Recommendation. In International Conference on World Wide Web, pages 661–670. ACM,
2010.
D. Liau, E. Price, Z. Song, and G. Yang. Stochastic Multi-Armed Bandits in Constant Space. In
International Conference on Artificial Intelligence and Statistics, 2018.
P. Ménard and A. Garivier. A Minimax and Asymptotically Optimal Algorithm for Stochastic
Bandits. In Algorithmic Learning Theory, volume 76, pages 223–237. PMLR, 2017.
H. Robbins. Some Aspects of the Sequential Design of Experiments. Bulletin of the American
Mathematical Society, 58(5):527–535, 1952.
A. Sani, A. Lazaric, and R. Munos. Risk-Aversion In Multi-Armed Bandits. In Advances in Neural
Information Processing Systems, pages 3275–3283, 2012.
Y. Seldin and G. Lugosi. An Improved Parametrization and Analysis of the EXP3++ Algorithm for
Stochastic and Adversarial Bandits. In Conference on Learning Theory, volume 65, pages 1–17.
PMLR, 2017.
W. R. Thompson. On the Likelihood that One Unknown Probability Exceeds Another in View of the
Evidence of Two Samples. Biometrika, 25, 1933.
F. Yang, A. Ramdas, K. Jamieson, and M. Wainwright. A framework for Multi-A(rmed)/B(andit)
Testing with Online FDR Control. In Advances in Neural Information Processing Systems, pages
5957–5966. Curran Associates, Inc., 2017.
Note: the simulation code used for the experiments is using Python 3. It is open-sourced at
https://GitHub.com/SMPyBandits/SMPyBandits and fully documented at
https://SMPyBandits.GitHub.io.
15
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
Appendix A. Omitted Proofs
We include here the proofs omitted in the main document.
A.1. Proof of Lemma 2, “Regret Lower and Upper Bounds for DT ”
Let A0 denote DT (A, (Ti )i∈N ). For every k ∈ {1, . . . , K},
" T
# L −1 T
T
T
i
X
X
X
X
E
(Xk,t − XA(t),t ) =
E
(Xk,t − XA(t),t ) + E
(Xk,t − XA(t),t )
t=1
≤
i=0
LX
T −1
i=0
≤
LX
T −1
max
k∈{1,...,K}
t=Ti−1
Ti
X
E
t=TLT −1
(Xk,t − XA(t),t ) +
t=Ti−1
max
k∈{1,...,K}
E
T
X
(Xk,t − XA(t),t )
t=TLT −1
RTi −Ti−1 A0 + RT −TLT −1 A0 .
i=0
Thus, by definition of the regret
0
RT (A ) ≤
=
LX
T −1
i=0
LX
T −1
i=0
RTi −Ti−1 (A0 ) + RT −TLT −1 (A0 )
RTi −Ti−1 ATi −Ti−1 + R T − TL −1 ATLT −TLT −1
| {z T }
≤TL
≤
LT
X
T
−TL −1
T
RTi −Ti−1 ATi −Ti−1 .
i=0
In the stochastic case, it is well known that the regret can be rewritten in the following way,
introducing µk the mean of arm k and µ∗ the mean of the best arm:
" T
#
X
0
∗
RT (A ) = E
(µ − µA(t) )
t=1
LT −1
=
X
i=0
=
LX
T −1
i=0
E
Ti
X
(µ∗ − µA(t) ) + E
t=Ti−1
T
X
(µ∗ − µA(t) )
t=TLT −1
RTi −Ti−1 ATi −Ti−1 ) + RT −TLT −1 (A0 ) .
|
{z
}
≥0
and the lower bound follows.
16
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
A.2. Proof of Theorem 4, “Conserving a Regret Upper Bound with Geometric Horizons”
It is interesting to note that the proof is valid for both the easiest case when δ = 0 (as it was known in
Cesa-Bianchi and Lugosi (2006) for γ = 1/2) and the generic case when δ ≥ 0, with no distinction.
As far as the authors know, this result in its generality with δ ≥ 0 is new.
Proof Let A0 := DT (A, (Ti )i∈N ), and consider a fixed bandit problem. The upper bound (UB) from
Lemma 2 gives
0
RT (A ) ≤
LT
X
RTi −Ti−1 (AT =Ti −Ti−1 )
i=0
We can use the hypothesis on A for each regret term, as f and t 7→ ctγ (log t)δ are non-decreasing
for t ≥ 1 (by hypothesis for f and by Lemma 16).
≤
LT
X
f (Ti − Ti−1 ) +
cT0γ (log T0 )δ
i=0
+c
LT
X
(Ti − Ti−1 )γ (log (Ti − Ti−1 ))δ
i=1
The first part is denoted g1 (T ) := i=0 f (Ti − Ti−1 ) + cT0γ (log T0 )δ , it is an increasing function
as a sum
Pof increasing functions, and it is dealt
with by using Lemma17: the sum of f (Ti − Ti−1 )
LT
γ
δ
γ
δ by hypothesis, and this sum
is a o
i=0 (Ti − Ti−1 ) (log(Ti − Ti−1 )) , as f (t) = o t (log t)
PLT
of (Ti − Ti−1 )γ (log(Ti − Ti−1 ))δ is proved below to
be bounded by c0 T γ (log(T ))δ for a certain
constant c0 > 0, which gives g1 (T ) = o T γ (log T )δ . For the second part, we bound Ti − Ti−1 ≤
T0 (b − 1)bi−1 + 1 thanks to Definition 3. Moreover, as γ < 1 we can use Lemma 14 (Eq. (23)) to
distribute the power on γ, so (T0 (b − 1)bi−1 + 1)γ ≤ (T0 (b − 1)bi−1 )γ + 1(γ 6= 0) (indeed if γ = 0
both sides are equal to 1). This gives
≤ g1 (T ) + c(T0 (b − 1))γ
LT
LT
X
X
(bi−1 )γ (log(Ti − Ti−1 ))δ + c1(γ 6= 0)
(log(Ti − Ti−1 ))δ
i=1
i=1
P T
PLT δ
δ
δ
δ
If γ =
6 0, thelast sum is bounded by L
i (log Ti ) ≤ (log T0 ) (LT + 1) + (log b)
i i which
δ+1
δ+1
γ
δ
is a O LT
= O (log T )
= o T (log T ) (as γ > 0, thanks to a geometric sum), and
so it can be included in g2 (T ) = o T γ (log T )δ . If γ = 0, there is only the first sum. We bound
again Ti − Ti−1 ≤ T0 (b − 1)bi−1 + 1 and use Lemma 15 to bound log(T0 (b − 1)bi−1 + 1) by
log(T0 (b−1)+1)
i−1 ) term (as T (b − 1) > 1 by hypothesis).
0
log(T0 (b−1)) log(T0 (b − 1)b
δ
LT
X
log(T0 (b − 1) + 1)
i−1
i−1 γ
≤ g2 (T ) + c(T0 (b − 1))
(b )
log(T0 (b − 1)b )
log(T0 (b − 1))
γ
i=1
i−1
We split the
log(T0 (b − 1)b ) term in two, and once again, the termγwith log(T
0 (b − 1)) gives a
L
−1
O b T
(by a geometric sum), which gets included in g3 (T ) = o T (log T )δ . We focus on the
fastest term, and we can now rewrite the sum from i = 0 to LT − 1,
LT −1
log(T0 (b − 1) + 1) δ X
≤ g2 (T ) + c(T0 (b − 1)) log(b)
(bi )γ iδ
log(T0 (b − 1))
γ
i=0
17
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
We naively bound iδ by (LT − 1)δ , and use a geometric sum to have
γ
≤ g2 (T ) + c(T0 (b − 1))
log(T0 (b − 1) + 1)
log(b)
log(T0 (b − 1))
δ
(LT − 1)δ
bγ
(bLT −1 )γ
bγ − 1
Finally, observe that LT − 1 ≤ logb ( TT0 ) ≤ logb (T ), so the (log b)δ term simplifies, and observe
that bLT −1 ≤ TT0 so the T0γ term also simplifies. Thus we get
≤ g2 (T ) + c
log(T0 (b − 1) + 1)
log(T0 (b − 1))
δ
bγ (b − 1)γ γ
T (log T )δ .
bγ − 1
The constant multiplicative loss ` depends on γ and δ as well as on T0 and b, and is `(γ, δ, T0 , b) :=
δ γ
log(T0 (b−1)+1)
b (b−1)γ
bγ −1 > 1.
log(T0 (b−1))
Minimizing the constant loss? This constant loss has two distinct part, `(γ, δ, T0 , b) = `1 (δ, T0 , b)
`2 (γ, b), with `1 depending on δ, T0 and b (equal to 1 if δ = 0), and `2 depending on γ and b.
δ
0 (b−1)+1)
• Minimizing this constant loss `1 (δ, T0 , b) := log(T
≥ 1 is independent of δ (even
log(T0 (b−1))
if it is 0). If we assume b to be fixed, `1 (δ, T0 , b) → 1 when T0 → ∞. Moreover, for any
δ and b > 1, `1 (δ, T0 , b)√goes to 1 very quickly when T0 is large enough. For instance, for
γ = δ = 12 and b = 3+2 5 (see Corollary 10), then `1 (δ, T0 , b) ' 1.109 for T0 = 2, ' 1.01
for T0 = 10 and ' 1.0004 for T0 = 100.
γ
γ
• To minimize this constant loss `2 (γ, b) := b b(b−1)
> 1, we fix γ and study h : b 7→
γ −1
`2 (γ, b). The function h is of class C 1 on (1, ∞) and h(b) → +∞ for b → 1+ and b →
∞, so h has a (possibly non-unique) global minimum and attains it. Moreover h0 (b) =
γbγ−1 (b−1)γ−1
γ+1 − 2b + 1 has the sign of bγ+1 − 2b + 1, which does not have a constant
b
γ
2
(b −1)
sign and does not have explicit root(s) for a generic γ. However, it is easy to minimizing
`2 (γ, b) for b numerically when γ is known and fixed (with, e.g., Newton’s method).
The result from Theorem 4 of course implies the result from (Cesa-Bianchi and Lugosi, 2006,
Ex.2.9), in the special case of δ = 0 and γ = 12 (for minimax bounds), as stated numerically in the
following Corollary 10.
1
1
, 0, T0 , b) does not
2 and δ = 0, the multiplicative loss `( 2 q
√
√
5
b∗ ( 21 ) = 3+2 5 ' 2.62 and its minimum is 11+5
' 3.33.
2
√
2
loss of √2−1 ' 3.41, close to the optimal value.
Corollary 10 If γ =
depend on T0 . It
is then minimal for
Usually b = 2 is
used, which gives a
In particular, order-optimal and optimal algorithms for the minimax bound have γ =
f (t) = 0, for which Theorem 7 gives a simpler bound
√
√
2 √
i
RT (AT ) ≤ c T =⇒ RT (DT (A, (T0 2 )i∈N )) ≤ √
c T.
2−1
18
1
2
and
(18)
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
A.3. Proof of Theorem 9, “Minimax Regret Lower Bound with Exponential Horizons”
Proof Let A0 := DT (A, (Ti )i∈N ), and consider a fixed stochastic bandit problem. The lower bound
(LB) from Lemma 2 gives
RT (A0 ) ≥
LX
T −1
RTi −Ti−1 (AT =Ti −Ti−1 )
i=0
We can use the hypothesis on A for each regret term, and as 0 < γ ≤ 1, we can use Lemma 14
(Eq. (24)) to distribute the power on γ to ease the proof and obtain
≥
cT0γ
+c
≥ cT0γ + c
LX
T −1
(Ti − Ti−1 )γ
i=1
LX
T −1
γ
Tiγ − Ti−1
(it is a telescopic sum and simplifies)
i=1
≥ cTLγT −1
1
Observe that TLT −1 ≥ (TLT ) b by definition of the exponential sequence (Def. 6), and TLT ≥ T
(Def. 1). For the log(TLT −1 ) term, we simply have log(TLT −1 ) ≥ 1b log(T ) so if c0 = c/b, then we
obtain what we want
γ
RT (A0 ) ≥ c0 T b .
γ
b , and it looks very similar to
This lower bound goes from RT (AT ) = ΩT γ to RT (DT (A)) = ΩT
bγ
the upper bound from Theorem 7 where RT (DT (A)) = O T
was obtained from RT (AT ) =
O(T γ ).
1
Remark It does seem sub-optimal to lower bound TLT −1 like this (TLT −1 ≥ (TLT ) b ), but we
remind that T can be located anywhere in the discrete interval {TLT −1 , . . . , TLT − 1}, so in the
worst case when T is very close to TLT (and for large enough T ), we indeed have TLb T −1 ∼ TLT and
1
TLT ∼ T , so with this approach, the lower bound TLT −1 ≥ T b cannot be improved.
Appendix B. Minimax Regret Lower Bound with Geometric Horizons
We include here a last result that partly replies to Theorem 4. It is more subtle that the lower bound in
Theorem 5 but still provides an interesting insight: if b is not chosen carefully (i.e., if `0 (γ, b) > 1),
then the anytime version of AT using a geometric Doubling Trick suffers a non-improvable constant
multiplicative loss compared to AT .
Theorem 11 For stochastic models, if A satisfies RT (AT ) ≥ c T γ , for 0 < γ < 1 and c > 0,
then the anytime version A0 := DT (A, (Ti )i∈N ) with the geometric sequence (Ti )i∈N of parameters
T0 ∈ N∗ , b > 1 (i.e., Ti = bT0 bi c) satisfies
LT ≥ 2 =⇒ RT (A0 ) ≥ `0 (γ, b) c T γ + g0 (T ).
19
(19)
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
with g0 (t) = O(log t) = o(tγ ), and a constant loss `0 (γ, b) depending only on γ and b,
(b − 1)γ
> 0.
− 1)
`0 (γ, b) =
(20)
bγ (bγ
`0 (γ, b) is always > 0 and tends to 0 for b → ∞, and some choice of b gives `0 (γ, b) > 1.
Proof Let A0 := DT (A, (Ti )i∈N ), and consider a fixed stochastic bandit problem. Assume LT ≥ 2.
The lower bound (LB) from Lemma 2 gives
0
RT (A ) ≥
LX
T −1
RTi −Ti−1 (AT =Ti −Ti−1 )
i=0
We bound Ti − Ti−1 ≥ T0 (b − 1)bi−1 − 1 for i > 0, thanks to Definition 3, and we can use the
hypothesis on A for each regret term. Additionally, we have (Ti − Ti−1 )γ ≥ (T0 (b − 1)bi−1 − 1)γ ≥
(T0 (b − 1))γ (bi−1 )γ − 1 by Lemma 14 (Eq. (24), as b > 1 and 0 < γ < 1), thus
≥
LX
T −1
c(Ti − Ti−1 )γ ≥ cT0γ + cT0γ (b − 1)γ
LX
T −1
LX
T −1
i=1
i=1
i=0
≥
cT0γ
+
cT0γ (b
γ
− 1)
(bi−1 )γ − c
1
LX
T −2
(bi )γ − c(LT − 1)
i=0
We have
PLT −1
i=0
(bi )γ =
(bLT −1 )γ −1
bγ −1
thanks to a geometric sum (with γ > 0) and thus
≥ cT0γ + cT0γ (b − 1)γ
(bLT −1 )γ − 1
+ c(1 − LT )
bγ − 1
γ
Thanks to Definition 3, bLT −1 satisfies bLT −1 ≥ 1b TT0 . Let g0 (T ) := cT0γ (b−1)
−
1
+ c(LT −
bγ −1
1) = O(1) + O logb ( TT0 ) = O(log T ) = o(T γ ) and g0 (T ) > 0, then we have
(b − 1)γ γ
(b − 1)γ
γ
≥c γ γ
T − cT0
− 1 + c(LT − 1) .
b (b − 1)
bγ − 1
We obtain as announced, RT (A0 ) ≥ `0 (b) cT γ + g0 (T ) with g0 (T ) = O(log T ) = o(T γ ).
γ
Maximizing the constant loss? To maximize3 `0 (γ, b) := bγ(b−1)
(bγ −1) > 0, we fix γ and study
1
the function h : b 7→ `0 (γ, b). The function h is of class C on (1, ∞) and h(b) → +∞ for
(b−1)γ−1
b → 1+ and h(b) → 0 for b → ∞. Moreover h0 (b) = −γ bγ+1
−2bγ + bγ+1 + 1 has the
γ −1)2
(b
same sign as f (b) := − −2bγ + bγ+1 + 1 . The function f is of class C 1 , with f (1) = 0 and
2γ
2γ
f 0 (b) = −(γ + 1)(b − γ+1
)bγ−1 , and as 0 < γ < 1, γ+1
< 1 so f 0 (b) < 0 for all b > 1. Thus f is
decreasing, and ∀b > 1, f (b) < f (1) = 0. So h0 has a negative sign, and this allows to conclude that
h is decreasing, and so b 7→ `0 (γ, b) has no global maximum at fixed γ, and `0 → ∞ if b → 1+ .
3. For the largest possible lower bound, we try to maximize the constant loss in the lower bound.
20
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
Relationship with the upper bound. For any b > 1, we compare `0 (γ, b) with `(γ, b) and we see
that, interestingly, `0 (γ, b) = `2 (γ, b)/b2γ , with `2 (γ, b) from Theorem 4. For the particular case
of γ = 12 , this lower bound also leads to another interesting remark: if b is chosen to minimize the
loss in the upper bound (Theorem 4, b∗ ( 12 ) =
√
√
3+ 5
2 ),
then this lower bound gives `0 ( 12 , b∗ ( 12 )) =
1 + 22 ' 1.71 > 1, which proves that this choice of geometric doubling trick cannot be used to
conserve an optimal algorithm, i.e., the constant loss cannot be made as close to 1 as we want.
Appendix C. An Efficient Heuristic, the Doubling Trick with “No Restart”
Let A0 := DT no-restart (A, (Ti )i∈N ) denotes the following Algorithm 2. The only difference with
DT (Algorithm 1) is that the history from all the steps from t = 1 to t = Ti is used to reinitialize
the new algorithm A(i) . To be more precise, this means that a fresh algorithm A(i) is created, and
then fed with successive observations (A0 (s), YA0 (s),s ) for all 1 ≤ s < t, like if it was playing
from the beginning. Note that A(i) could have chosen a different path of actions, but we give it the
observations from the previous plays of A0 .
This obviously cannot be applied to any kind of algorithm A, and for instance any algorithm
based on arm elimination (e.g., Explore-Then-Commit approaches like in Garivier et al. (2016)) will
most surely fail with this approach. This second doubling trick algorithm DT no-restart can be applied
in practice if A is index-based and uses the horizon T as a simple numerical parameter in its indexes,
like it is the case for instance for AFHG (cf. Eq. (16)). or kl-UCB++ (cf. Eq. (17)).
Input: Bandit algorithm A, and sequence (Ti )i∈N
Let i = 0, and initialize algorithm A(0) = AT0 .
2 for t = 1, . . . , T do
3
if t > Ti then
// Next horizon Ti+1 from the sequence
4
Let i = i + 1.
5
Initialize algorithm A(i) = ATi −Ti−1 .
6
Update internal memory of A(i) with the history of plays and rewards from Ai−1 .
7
end
8
Play with A(i) : play arm A0 (t) := A(i) (t − Ti ), observe reward r(t) = YA0 (t),t .
9 end
Algorithm 2: The Non-Restarting Doubling Trick Algorithm, A0 = DT no-restart (A, (Ti )i∈N ).
1
However, it is much harder to state any theoretical result on this heuristic DT no-restart . We
conjecture that a regret upper bound similar to (UB) from Lemma 2 could still be obtained, but
it is still an open problem that the authors do not know how to tackle for a generic algorithm.
The intuition is that starting A(i) with some previous observations from the (same) i.i.d. process
(Yk,s )k∈{1,...,K},s∈N can only improve the performance of A(i) , and lead to a smaller regret on the
interval {Ti , . . . , Ti+1 − 1}.
21
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
Appendix D. Basic but Useful Results
All the logarithm log are taken in base e = exp(1) (i.e., natural logarithm ln, but log is preferred for
x
readability). Logarithms in a basis b > 1 are denoted logb (x) := log
log b , for any x ∈ R, x > 0.
We remind that bxc denotes the integer part of x ∈ R, and for x > 0, that is the unique integer i
such that i ≤ x < i + 1. The only property we use is its definition and the fact that bxc ≤ x. We
also define dxe := 1 + bxc for x ≥ 0, which is the unique integer j such that j − 1 ≤ x < j.
D.1. Weighted Geometric Inequality
Lemma 12 (Weighted Geometric Inequality) For any n ∈ N∗ , b > 1 and δ > 0, and if f is a
function of class C 1 , non-decreasing and non-negative on [0, ∞), we have
n
X
bδ
f (n)(bn )δ .
bδ − 1
f (i)(bi )δ ≤
i=0
(21)
Proof By hypothesis, f is non-decreasing, so ∀i ∈ {0, . . . , n}, f (i) ≤ f (n), and so by using the
sum of a geometric sequence, we have
!
n
n
X
X
bδ
1
(bn+1 )δ = δ
f (n)(bn )δ .
f (i)(bi )δ ≤ f (n)
(bi )δ ≤ f (n) δ
b −1
b −1
i=0
i=0
f (i) = 1 gives the geometric inequality. Note that if we make δ → 0, the left sum converges to
P
n−1
i=0 f (i) and the right term diverges to +∞, making this inequality completely uninformative.
D.2. Another Sum Inequality
i
This second result is similar to the previous one but for a “doubly exponential” sequence, i.e., ab , as
it also bounds a sum of increasing terms by a constant times its last term.
Lemma 13 For any n ∈ N∗ , a > 1, b > 1 and γ > 0, we have
n
n
X
1
i
bn γ
(ab )γ ≤ aγ + 1 +
(a
)
=
O
(ab )γ .
(log(a))(log(bγ ))
(22)
i=0
Proof We first isolate both the first and last term in the sum and focus on the from i = 1 sum up to
t
i = n − 1. As the function t 7→ (ab )γ is increasing for t ≥ 1, we use a sum-integral inequality, and
then the change of variable u := γbt , of Jacobian dt = log1 b du
u , gives
n−1
X
bi γ
Z
(a ) ≤
i=1
n
a
1
γbt
1
dt ≤
log(bγ )
Z
γbn
γb
au
du
u
u
Now for u ≥ 1, observe that au ≤ au , and as γb > 1, we have
Z γbn
1
1
1
1
n
n γ
≤
au du ≤
aγb =
(ab ) .
γ
γ
γ
log(b ) γb
log(b ) log(a)
(log(a))(log(b ))
Finally, we obtain as desired,
n
P
i
n
(ab )γ ≤ aγ + (ab )γ +
i=0
22
1
bn γ
(log(a))(log(bγ )) (a ) .
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
D.3. Basic Functional Inequalities
These functional inequalities are used in the proof of the main theorems.
Lemma 14 (Generalized Square Root Inequalities) For any x, y ≥ 0 and 0 < δ < 1,
(x + y)δ ≤ xδ + y δ .
(23)
And conversely for any 0 < δ < 1, and x, y ≥ 0, if x ≥ y then
(x − y)δ ≥ xδ − y δ .
(24)
Proof Fix y ≥ 0. Let f (x) := (x+y)δ −(xδ +y δ ) for x ≥ 0. First, f (0) = y δ −y δ = 0, and as δ > 0,
f is differentiable on [0, ∞), with f 0 (x) = (log δ)(x + y)δ − (log δ)xδ = (log δ)((x + y)δ − xδ ), and
as δ < 1, log δ < 0, and (x + y)δ ≥ xδ , so f 0 (x) ≤ 0 for any x ≥ 0. Therefore, f is non-increasing,
and so ∀x ≥ 0, f (x) ≤ f (0) = 0, so f is non-positive, giving the desired inequality (for any y ≥ 0
and any x ≥ 0).
The second inequality is a direct application of the first one. Assume x ≥ y, and let x0 = x − y ≥
0, then (x0 + y)δ ≤ (x0 )δ + y δ . This gives (x − y)δ = (x0 )δ ≤ (x0 + y)δ − y δ = xδ − y δ .
Lemma 15 (Bounding log(x − ∆)) Let x0 > 1 and 0 < ∆ < x0 (e.g., ∆ ≤ 1), then
∀x ≥ x0 ,
log(x0 − ∆)
log(x) ≤ log(x − ∆) ≤ log(x).
log(x0 )
(25)
With ∆ = 1, it implies that if T0 > 1, b > 1 satisfy T0 (b − 1) > 1, then for any i ∈ N, we have
log(T0 (b − 1)bi − 1) ≥
log(T0 (b − 1) − 1)
log(T0 (b − 1)bi ).
log(T0 (b − 1))
(26)
log(x−∆)
1
log(x) , defined for x ≥ x0 . It is of class C , and by differentiating, we have
f 0 (x) = log(x)−log(x−∆)
> 0 as log is increasing. So f is increasing, and its minimum is attained at
x(log x)2
0 −∆)
x = x0 , i.e., ∀x ≥ x0 , f (x) ≥ f (x0 ) = log(x
log(x0 ) > 0, which gives Eq. (25).
Proof Let f (x) :=
The corollary is immediate but stated explicitly for clarity when used in page 6.
Lemma 16 For any γ > 0 and δ > 0, the function f : x 7→ xγ (log x)δ is increasing on [1, ∞).
Proof f is of class C 1 on [1, ∞). First, if γ > 0, we have f 0 (x) = xγ−1 (log x)δ−1 (δ + γ log(x)).
So f 0 (x) > 0 if and only if δ + γ log(x) ≥ 0, that it x ≥ exp(− γδ ). But x ≥ 1 and < 0 so f 0 (x) is
always positive, and thus f is increasing. Then, if γ = 0, we have f 0 (x) = δ x1 (log x)δ−1 > 0 as
x > 1 gives log x > 0 and so (log x)δ−1 > 0.
It is also true if γ ≥ 0, δ ≥ 0 if not both are zero simultaneously.
23
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
D.4. Controlling an Unbounded Sum of Dominated Terms
This Lemma is used in the proofs of our upper bounds (Theorems 4 and 7), to handle the sum of
f (Ti ) terms. In particular, it can be applied to (Ti ) the geometric sequence and g(t) = h(t) = tγ or
g(t) = h(t) = tγ (log t)δ (for Theorem 4) for γ > 0 and δ ≥ 0; or (Ti ) the exponential sequence,
g(t) = tγ (log t)δ and h(t) = tbγ (log t)δ (for Theorem 7) for γ ≥ 0 and δ ≥ 0. Note that it would be
obvious if LT was bounded for T → ∞, but a more careful analysis has to be given as LT → ∞.
Lemma 17 Let f , g and h be three positive, diverging and non-decreasing functions on [1, ∞),
such that f (t) = o(g(t)) for t → ∞. Let a non-decreasing diverging sequence (Ti )i∈N , and a
diverging sequence (LT )T ∈N (i.e., Ti → ∞Pfor i → ∞ and LT → ∞ if T → ∞), such that there
T
exists a constant c ≥ 0 satisfying ∀T ≥ 1, L
i=0 g(Ti ) ≤ c × h(T ). Then the (unbounded) sum of
dominated terms f (Ti ) is still dominated by h(T ), i.e.,
f (t) = o(g(t)) and ∃c ≥ 0,
T →∞
LT
X
g(Ti ) ≤ c × h(T ) =⇒
∀T ≥1
i=0
LT
X
f (Ti ) = o(h(T )) .
i=0
T →∞
(27)
Proof By hypothesis, if f is dominated by g, formally f (t) = o(g(t)), then there exists a positive
function ε(t) such that f (t) = g(t)ε(t) and ε(t) → 0 for t → 0. Fix η > 0, as small as we want,
then there exists Tη ≥ 1 such that ∀t ≥ Tη , ε(t) ≤ η. Let iη such that Tiη −1 < Tη ≤ Tiη . Now for
any T ≥ Tη and large enough so that LT ≥ Tη , we can start to split the sum
LT
X
i=0
iη −1
f (Ti ) =
X
i=0
f (Ti ) +
LT
X
f (Ti )
i=iη
The first sum is naively bounded by iη × f (Tiη −1 ) as f is increasing, and for the second sum, for
any i ≥ iη , Ti ≥ Tη and so f (Ti ) = ε(Ti )g(Ti ) ≤ ηg(Ti ), thus
LT
X
≤ iη f (Tiη −1 ) + η ×
g(Ti )
i=iη
The sum is smaller than a sum on a larger interval, as g(Ti ) ≥ 0 for any i, and f is increasing so
!
LT
X
≤ iη f (Tη ) + η
g(Ti )
i=0
But now, f (Tη ) ≤ ηg(Tη ) by hypothesis, and this sum is smaller than c × h(T ) also by hypothesis
≤ iη η × g(Tη ) + ηc × h(T ) = η (iη g(Tη ) + c × h(T ))
Finally, we use the hypothesis that h(T ) is diverging and as η and Tη are both fixed, there exists a
f
fη . And so we have finally proved that
Tη ≥ Tη large enough so that iη g(Tη ) ≤ h(T ) for all T ≥ T
fη ≥ 1, ∀T ≥ T
fη ,
∀η > 0, ∃T
LT
X
f (Ti ) ≤ η(c + 1) × h(T ).
i=0
This concludes the proof and shows that
L
T
P
f (Ti ) = o(h(T )) as wanted.
i=0
24
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
Appendix E. Additional Experiments
We presents additional experiments, for Gaussian bandits and for the heuristic DT no-restart .
E.1. Experiments with Gaussian Bandits (with Known Variance)
We include here another figure for experiments on Gaussian bandits, see Fig. 5.
Cumulated regrets for different bandit algorithms, averaged 1000 times
9 arms: [N(−4, 1), N(−3, 1), N(−2, 1), N(−1, 1), N(0, 1), N(1, 1), N(2, 1), N(3, 1), N(4, 1) ∗ ]
350
300
Cumulated regret Rt = tµ ∗ −
s=0 k=1
9
t−1 X
X
µk
1000 [Tk (s)]
400
UCB
ApprFHG(T = 45678)
DT(Ti = 200 × 2i )[ApprFHG]
DT(Ti = 2002i )[ApprFHG]
DT(Ti = 2001.1i )[ApprFHG]
DT(Ti = (200/2)22i )[ApprFHG]
Lai & Robbins lower bound = 5.44 log(t)
250
200
150
100
50
0
0
10000
20000
30000
Time steps t = 0. . . T − 1, horizon T = 45678
40000
Figure 5: Regret for DT , for K = 9 Gaussian arms N (µ, 1), horizon T = 45678, n = 1000
repetitions and µ uniformly spaced in [−5, 5]K . On “easy” problems like this one, both UCB and
AFHG perform similarly and attain near constant regret (identifying the best Gaussian arm is very
easy here as they are sufficiently distinct). Each doubling trick also appear to attain near constant
regret, but geometric doubling (b = 2) and slow exponential doubling (b = 1.1) are slower to
converge and thus less efficient.
E.2. Experiments with DT no-restart
As mentioned previously, the DT no-restart algorithm (Algorithm 2) is only an heuristic so far, as no
theoretical guarantee was proved for it. For the sake of completeness, we also include results from
numerical experiments with it, to compare its performance against the “with restart” version DT .
As expected, DT no-restart enjoys much better empirical performance, and in Figs. 6 and 7 we see
that a geometric or a slow exponential doubling trick with no restart with kl-UCB++ can outperform
kl-UCB and perform similarly to the non-anytime kl-UCB++ . But as observed before, the regret
blows up after the beginning of each new sequence if the doubling sequence increase too fast (e.g.,
exponential doubling). Despite what is proved theoretically in Theorem 5, here we observe that the
geometric doubling is the only one to be slow enough to be efficient.
25
W HAT D OUBLING T RICKS C AN AND C AN ’ T D O FOR M ULTI -A RMED BANDITS
Cumulated regrets for different bandit algorithms, averaged 1000 times
9 arms: Bayesian MAB, Bernoulli with means on [0, 1]
400
Cumulated regret Rt = tµ ∗ −
s=0 k=1
9
t−1 X
X
µk
1000 [Tk (s)]
500
KLUCB
KLUCB+ + (T = 45678)
DTnr(Ti = 200 × 2i )[KLUCB+ + ]
DTnr(Ti = 2002i )[KLUCB+ + ]
DTnr(Ti = 2001.1i )[KLUCB+ + ]
DTnr(Ti = (200/2)22i )[KLUCB+ + ]
Lai & Robbins lower bound = 6.81 log(t)
300
200
100
0
0
10000
20000
30000
Time steps t = 0. . . T − 1, horizon T = 45678
40000
Figure 6: Regret for K = 9 Bernoulli arms, horizon T = 45678, n = 1000 repetitions and µ taken
uniformly in [0, 1]K , for DT no-restart . Geometric doubling (e.g., b = 2) and slow exponential doubling
(e.g., b = 1.1) are too slow, and short first sequences make the regret blow up in the beginning of the
experiment. At t = 40000 we see clearly the effect of a new sequence for the best doubling trick
(Ti = 200 × 2i ). As expected, kl-UCB++ outperforms kl-UCB, and if the doubling sequence is
growing fast enough then DT no-restart (kl-UCB++ ) can perform as well as kl-UCB++ .
Cumulated regrets for different bandit algorithms, averaged 1000 times
9 arms: [B(0.1), B(0.2), B(0.3), B(0.4), B(0.5), B(0.6), B(0.7), B(0.8), B(0.9) ∗ ]
140
100
Cumulated regret Rt = tµ ∗ −
s=0 k=1
9
t−1 X
X
µk
1000 [Tk (s)]
120
KLUCB
KLUCB+ + (T = 45678)
DTnr(Ti = 200 × 2i )[KLUCB+ + ]
DTnr(Ti = 2002i )[KLUCB+ + ]
DTnr(Ti = 2001.1i )[KLUCB+ + ]
DTnr(Ti = (200/2)22i )[KLUCB+ + ]
Lai & Robbins lower bound = 7.52 log(t)
80
60
40
20
0
0
10000
20000
30000
Time steps t = 0. . . T − 1, horizon T = 45678
40000
Figure 7: K = 9 Bernoulli arms with µ evenly spaced in [0, 1]K . On easy problems like this
one, both kl-UCB and kl-UCB++ are very efficient, and here the geometric allows the DT no-restart
anytime version of kl-UCB++ to outperform both kl-UCB and kl-UCB++ .
26
| 10math.ST
|
Algorithmic Linearly Constrained Gaussian Processes
Markus Lange-Hegermann 1
arXiv:1801.09197v1 [stat.ML] 28 Jan 2018
Abstract
We algorithmically construct multi-output Gaussian process priors which satisfy linear differential
equations. Our approach attempts to parametrize
all solutions of the equations using Gröbner bases.
If successful, a push forward Gaussian process
along the paramerization is the desired prior. We
consider several examples, among them the full
inhomogeneous system of Maxwell’s equations.
By bringing together stochastic learning and computeralgebra in a novel way, we combine noisy
observations with precise algebraic computations.
1. Introduction
In recent years, Gaussian process regression has become a
prime regression technique (Rasmussen & Williams, 2006).
Roughly, a Gaussian process can be viewed as a suitable2
probability distribution on a set of functions, which we can
condition on observations using Bayes’ rule. The resulting
mean function is used for regression. Additionally, one can
also easily sample random functions, so-called realizations,
from this distribution. The strength of Gaussian process
regression lies in avoiding overfitting while still finding
functions complex enough to describe any behavior present
in given observations, even in noisy or unstructured data.
Gaussian processes are usually applied when observations
are rare or expensive to produce. Applications range, among
many others, from robotics (Deisenroth et al., 2015), biology
(Honkela et al., 2015), global optimization (Osborne et al.,
2009), astrophysics (Garnett et al., 2015) to engineering
(Thewes et al., 2015).
Incorporating justified assumptions into the prior helps these
applications: the full information content of the scarce observations can be utilized to create a more precise regression
model. Examples of such assumptions are smooth or rough
behavior, trends, homogeneous or heterogeneous noise, local or global behavior, and periodicity (cf. §4 in (Rasmussen
1
Correspondence to:
Markus Lange-Hegermann
<[email protected]>.
2
Gaussian processes are the maximum entropy prior for finite
mean and variance in the unknown behavior (Jaynes, 1968; Jaynes
& Bretthorst, 2003).
& Williams, 2006),(Duvenaud, 2014)). Such assumptions
are usually incorporated in the covariance structure of the
Gaussian process.
Even certain physical laws, given by certain linear differential equations, could be incorporated into the covariance
structures of Gaussian process priors. Thereby, despite their
random nature, all realizations and the mean function of
the posterior strictly adhere to these physical laws3 . For
example, (Macêdo & Castro, 2008) constructed covariance
structures for divergence-free and curl-free vector fields,
which (Wahlström et al., 2013; Solin et al., 2015) used to
model electromagnetic phenomena.
A first step towards systematizing this construction was
achieved in (Jidling et al., 2017). In certain cases, a
parametrization of all solutions for physical laws could be
found by a computation that does not necessarily terminate.
Having found a parametrization, one could assume a Gaussian process prior for the parametrizing functions and push
it forward. This results in a Gaussian process prior for the
solutions of the physical laws.
In Section 3, this paper combines these ideas from (Jidling
et al., 2017) with an algorithm which computes this
parametrization if it exists, or reports failure if it does not
exist. This algorithm is a homological result in algebraic
system theory (cf. §7.(25) in (Oberst, 1990)).
This paper adds information to Gaussian processes in two
ways:
(i) restricting to solutions of linear operator matrices by
constructing a suitable prior and
(ii) conditioning on observations using Bayes’ rule.
Since these two constructions are compatible, we can combine strict, global information from equations with noisy,
local information from observations. The author views
this combination of techniques from homological algebra
and machine learning as the main result of this paper, and
3
For notational simplicity, we refrain from using the phrases
“almost surely” and “up to equivalence” in this paper. We say that
all realizations of a Gaussian process adhere to such laws if there
is an element in its equivalence class w.r.t. almost sure equality
that has this property, and similarly for continuity and smoothness.
Algorithmic Linearly Constrained Gaussian Processes
the construction of covariance functions satisfying physical
laws as a proof of concept.
Example 4.2 shows a typical application. It constructs a
Gaussian process such that all of its realizations satisfy the
inhomogeneous Maxwell equations of electromagnetism.
Conditioning this Gaussian process on a single observation of electric current yields, as expected, a magnetic field
circling around this electric current.
Even though Gaussian processes are a highly precise interpolation tool, they lack in two regards: missing extrapolation
capabilities and high computation time, cubically in the
amount of observations. These problems have, to a certain degree, been addressed: several fast approximations to
Gaussian process regression (Titsias, 2009; Hensman et al.,
2013; Wilson et al., 2015a; Izmailov et al., 2017; Dong
et al., 2017) and more powerfull covariance structures (Lee
et al., 2017; Wilson & Adams, 2013; Wilson et al., 2015b;
Calandra et al., 2016) have been proposed. This paper addresses these two problems from a complementary angle.
The linear differential equations allow to extrapolate and
reduce the needed amount of observations, which improves
computation time.
2. Differential Equations and Gaussian
Processes
This section is mostly expository and summarizes Gaussian processes and how differential operators act on them.
Subsection 2.1 summarizes Gaussian process regression.
We then introduce differential (Subsection 2.2) and other
operators (Subsection 2.4), and sketch their connection to
constructing priors (Subsection 2.3).
2.1. Gaussian processes
A Gaussian process g = GP(µ, k) is a distribution on the
set of functions Rd → R` such that the function values
g(x1 ), . . . , g(xn ) at x1 , . . . , xn ∈ Rd have a joint Gaussian
distribution. It is specified by a mean function
µ : Rd → R` : x 7→ E(g(x))
and a positive semidefinite covariance function
`×`
k : Rd ⊕ Rd → R0
:
(x, x0 ) 7→ E((g(x) − µ(x))(g(x0 ) − µ(x0 ))T ) .
Assume the regression model yi = g(xi ) and condition on
n observations
{(xi , yi ) ∈ R1×d ⊕ R1×` | i = 1, . . . , n} .
Denote by k(x, X) ∈ R`×`n resp. k(X, X) ∈ R`n×`n
the
0
(covariance) matrices obtained by concatenating the matrices k(x, xj ) resp. the positive semidefinite block partitioned
matrix with blocks k(xi , xj ). Write µ(X) ∈ R`×n for the
matrix obtained by concatenating the vectors µ(xi ) and
y ∈ R1×`n for the row vector obtained by concatenating the
rows yi . The posterior
x 7→ µ(X) + (y − µ(X))k(X, X)−1 k(x, X)T ,
(x, x0 ) 7→ k(x, x0 ) − k(x, X)k(X, X)−1 k(x0 , X)T ,
GP
is a Gaussian process and its mean function x →
7 µ(X) +
(y − µ(X))k(X, X)−1 k(x, X)T is used as regression
model.
2.2. Differential equations
From now on, let R = R[∂x1 , . . . , ∂xd ] be the polynomial
ring in the partial differential operators. This ring models linear (partial) differential equations with constant coefficients,
as it acts on the vector space F = C ∞ (Rd , R) of smooth
functions, where ∂xi acts by partial derivative w.r.t. xi . The
set of realizations of a Gaussian process with squared exponential covariance function is dense in F (cf. Thm. 12,
Prop. 42 in (Simon-Gabriel & Schölkopf, 2016)).
Roughly speaking, Gaussian processes are the linear objects
among stochastic processes. Hence, it is not surprising
to find a rich interplay of Gaussian processes and linear
differential equations.
The class of Gaussian processes is closed under matrices
00
B ∈ R`×` of linear differential operators with constant
coefficients. Let g = GP(µ, k) be a Gaussian process with
00
realizations in the space F ` of vectors with functions in F
as entries. Define the Gaussian process B∗ g as the Gaussian
process induces by the pushforward measure under B of the
Gaussian measure induced by g. It holds that
B∗ g = GP(Bµ(x), Bk(x, x0 )(B 0 )T ) ,
(1)
where B 0 denotes the operation of the operators in B on
functions with argument x0 ∈ Rd (cf. Thm. 9 in (Bertinet &
Agnan, 2004)).
The covariance matrices for such Gaussian processes as in
(1) are often singular. This is to be expected, as B∗ g is rarely
dense in F ` . For numerical stability, we tacitely assume the
model yi = g(xi ) + ε for small Gaussian white noise term ε
and adopt k by adding var(ε) to k(xi , xi ) for observations
xi .
Example 2.1. Let g = GP(0, k(x, x0 )) be a scalar univariate Gaussian process with differentiable realizations. Then,
the Gaussian process of derivatives of functions is given by
∂
∂2
0
g = GP 0,
k(x, x ) .
∂x ∗
∂x∂x0
∂
One can interpret this Gaussian process ∂x
g as taking
∗
derivatives as measurement data and producing a regression
model of derivatives.
Algorithmic Linearly Constrained Gaussian Processes
2.3. Parametrizations
2.4. Further operator rings
We say that a Gaussian process is in a function space, if
0
its realizations are contained in said space. For A ∈ R` ×`
define the solution set
The theory presented for differential equations with constant
coefficients also holds for other rings R of linear operators
and solution spaces F. The following three operator rings
are prominent examples.
solF (A) := {f ∈ F ` | Af = 0} .
Such solutions sets and Gaussian processes are connected.
Lemma 2.2. Let g = GP(µ, k) be a Gaussian process in
F `×1 . Then g is also a Gaussian process in solF (A) for
0
A ∈ R` ×` if and only if µ ∈ solF (A) and A∗ (g − µ) is the
constant zero process.
Proof. Assume that g is a Gaussian process in solF (A).
Then, as the mean function is a realization, µ ∈ solF (A).
Furthermore, for g̃ := (g − µ) = GP(0, k) we have that
all realizations are annihilated by A, and hence A∗ g̃ is the
constant zero process.
Conversely, assume that µ ∈ solF (A) and A∗ (g − µ) is
the constant zero process. This implies 0 = A∗ (g − µ) =
A∗ g − A∗ µ = A∗ g, i.e. all realizations of g become zero
after a pushforward by A. In particular, all realizations of g
are contained in solF (A).
Our goal is to construct Gaussian processes with realizations
dense in the solution set solF (A) of an operator matrix
0
A ∈ R` ×` . The following remark, implicit in (Jidling et al.,
2017), is a first step towards an answer.
0
00
Remark 2.3. Let A ∈ R` ×` and B ∈ R`×` with AB = 0.
00
Let g = GP(0, k) be a Gaussian process in F ` . Then, the
set of realizations of B∗ g is contained in solF (A).
This follows easily from Lemma 2.2, as A∗ (B∗ g) =
(AB)∗ g = 0∗ g = 0.
00
We call B ∈ R`×` a parametrization of solF (A) if
00
solF (A) = BF ` . Parametrizations yield the denseness
of the realizations of a Gaussian process B∗ g in solF (A).
The polynomial ring R = R[x1 , . . . , xd ] models polynomial
equations when it acts on the set F of smooth functions
defined on a (Zariski-)open set in Rd .
To model ordinary linear differential equations with rational4 coefficients consider the Weyl algebra R = R(t)h∂t i,
with the non-commutative relation ∂t t = t∂t + 1 representing the product rule of differentiation. Here, we consider
solutions in the set F of smooth functions defined on a
co-finite set.
The polynomial ring R = R[σx1 , . . . , σxd ] models linear
shift equations with constant coefficients when it acts on the
d
set F = RZ≥0 of d-dimensional sequences by translation of
the arguments.
3. Computing parametrizations
By the last section, constructing a parametrization B of
solF (A) yields a Gaussian process dense in the solution set
0
solF (A) of an operator matrix A ∈ R` ×` . Subsection 3.1
gives necessary and sufficient conditions for a parametrization to exist and Subsection 3.2 describes their computation.
It turns out that these computations are purely algebraic over
R.
3.1. Existence of parametrizations
It turns out that we can decide whether a parametrization
exists purely algebraically, with operations over R that do
not involve F.
0
Proposition 2.4. Let B ∈ R`×` be a parametrization of
0
solF (A) for A ∈ R` ×` . Let g = GP(0, k) be a Gaussian
00
process dense in F ` . Then, the set of realizations of B∗ g
is dense in solF (A).
By r-ker(A) we denote the right kernel of A ∈ R` ×` , i.e.
r-ker(A) = {m ∈ R`×1 | Am = 0}. By l-ker(A) we
0
denote the left kernel of A, i.e. l-ker(A) = {m ∈ R1×` |
mA = 0}. Abusing notation, denote any matrix as left resp.
right kernel if its rows resp. columns generate the kernel as
an R-module.
This proposition is a consequence of partial derivatives being bounded, and hence continuous, when F is equipped
with the Fréchet topology generated by the family of seminorms
Theorem 3.1. Let A ∈ R` ×` . Define matrices B =
r-ker(A) and A0 = l-ker(B). Then solF (A0 ) is the
largest subset of solF (A) that is parametrizable and B
parametrizes solF (A0 ).
00
0
kf ka,b := sup
sup
z∈[−b,b]d
i∈Zd
≥0
|i|≤a
|
∂
f (z)|
∂z i
for a, b ∈ Z≥0 (cf. §10 in (Treves, 1967)). Now, the continuous surjective map induced by B maps a dense set to a
dense set.
In constrast to vector spaces, the left kernel of the right
kernel of A is not necessarily A (up to an equivalence)
in case of operator rings. For example, the solution set
solF (A0 ) is the subset of controllable behaviors in solF (A).
4
No major changes for polynomial, holonomic, or meromorphic coefficients.
Algorithmic Linearly Constrained Gaussian Processes
Corollary 3.2. In Theorem 3.1, solF (A) is parametrizable
if and only if the rows of A and A0 generate the same rowmodule. Since AB = 0, this is the case if all rows of A0 are
contained in the row module generated by the rows of A.
(Decker et al., 2016) and Macaulay2 (Grayson & Stillman)
are two well-known open source implementations.
In this case, solF (A) is parametrized by B. Furthermore, a
00
Gaussian process g with realizations dense in F ` leads to
a Gaussian process B∗ g with realizations dense in solF (A).
Many covariance functions5 incorporate hyperparameters
and advanced methods specifically add more hyperparameters to Gaussian processes, see e.g. (Snelson et al., 2003;
Calandra et al., 2016; Wilson & Adams, 2013), for additional flexibility.
For a proof of this theorem and its corollary see Thm. 2 in
(Zerz et al., 2010), (cf. also Thm. 3, Alg. 1, and Lemma 1.2.3
in (Zerz, 2000) or §7.(24) in (Oberst, 1990)) and for additional characterizations, generalizations, and proofs using
more homological machinery see (Quadrat, 2013; 2010b;
Barakat, 2010; Seiler & Zerz, 2010; Chyzak et al., 2005;
Robertz, 2015) and references therein.
The approach assigns a prior to the parametrising functions
and pushes this prior forward to a prior of the solution set
solF (A). The paramerization is not canonical, and hence
different parametrizations might lead to different priors.
This did not lead to practical problems, yet.
3.2. Algorithms
Summarizing Theorem 3.1 and Corollary 3.2 algorithmically, we need to compute right kernels (of A), compute
left kernels (of B), and decide whether rows (of A0 ) are
contained in a row module (generated by the rows of A).
All these computations are an application of Gröbner basis
algorithms.
In the recent decades, Gröbner bases algorithms have become one of the core algorithms of computeralgebra, with
manifold applications among others in in geometry, system
theory, natural sciences, and automatic theorem proving.
Generalizing the Gaussian algorithm, reduced Gröbner
bases generalize the reduced echolon form to systems of
linear operator equations. In particular, using them, one can
compute all solutions in R (not in F) of the homogeneous
system and compute, if it exists, a particular solution in R
(not in F) for an inhomogeneous system. Solving homogeneous systems is the same as computing its right resp.
left kernel or finding all relations (the generalization of linear dependencies) between columns resp. rows of a matrix.
Solving inhomogeneous equations decides whether an element is contained in a module. Alternatively, the uniqueness
of reduced Gröbner bases also decides submodule equality.
A formal description of Gröbner bases would exceed the
scope of this note. Instead, we refer to the excellent literature (Sturmfels, 2005; Eisenbud, 1995; Adams & Loustaunau, 1994; Greuel & Pfister, 2002; Gerdt, 2005; Buchberger,
2006) and explicitely show in the next section how to use
Gröbner bases in computer algebra systems. They are implemented in various computer algebra systems, e.g., Singular
3.3. Hyperparameters
The approach in this paper is the opposite. We restrict the
Gaussian process prior, specifically to solutions of an operator matrix, guarding against overfitting. The prior of the
parametrizing functions can, and usually does, still contain
hyperparameters. These can be determined by maximizing
the likelihood, as usual.
Many important applications contain unknown parameters
in the equations. Such parameters can also be estimated by
the likelihood, when conditioning on data.
For ordinary differential equations, the solution set of an
operator matrix is a direct sum of parametrizable functions
and a finite dimensional set of functions, both with constant resp. variable coefficients, due to the Smith form resp.
Jacobson form. In many cases, in particular the case of constant coefficients, the solution set of the finite dimensional
summand can easily be computed. This paper also allows to
compute with the parametrizable summand of the solution
set and estimate parameters and hyperparameters of both
summands together.
4. Examples
Example 4.1. We reproduce the well-known fact that
divergence-free fields can be parametrized by the curl operator. This has been used in connection with Gaussian processes to model electric and magnetic phenomena (Macêdo
& Castro, 2008; Wahlström et al., 2013; Solin et al., 2015).
The same algebraic computation also constructs a prior for
tangent fields of a sphere.
The computer algebra system Macaulay2 (Grayson & Stillman) performs the Gröbner basis computations.
Let R be the polynomial ring in three indeterminates, which
we can both interpret as the polynomial ring Q[∂1 , ∂2 , ∂3 ]
in the differential operators resp. the polynomial ring
Q[x1 , x2 , x3 ] in the indeterminates.
i1 : R=QQ[d1,d2,d3]
o1 = R
o1 : PolynomialRing
5
Sometimes even the mean function contains hyperparameters.
These additional hyperparameters are usually not very expressive,
compared to the non-parametric Gaussian process model.
Algorithmic Linearly Constrained Gaussian Processes
and consider the matrix A = ∂1 ∂2 ∂3 representing
the divergence resp. the normals of circles centered around
the origin.
i2 : A=matrix{{d1,d2,d3}}
o2 = | d1 d2 d3 |
1
3
o2 : Matrix R <--- R
The right kernel of A is given by the operator B representing
the curl resp. tangent spaces of circles centered around the
origin.
i3 : B = generators kernel A
o3 = 1 | -d2 0
-d3 |
1 | d1 -d3 0
|
1 | 0
d2 d1 |
3
3
o3 : Matrix R <--- R
Since the right kernel A0 of B is again the A, the matrix B
is really parametrization matrix of the solutions of A.
i4 : A1 = transpose generators kernel
transpose B
o4 = | d1 d2 d3 |
1
3
o4 : Matrix R <--- R
(As kernel in Macaulay2 yields right kernels, we
compute the left kernel by transposition.)
Figure 1. Taking the squared exponential covariance function for
k in Example 4.1 yields the above smooth mean tangent field
on the sphere after conditioning at 4 evenly distributed points on
the equator with two opposite tangent vectors pointing north and
south each. The two visible of these four vectors are displayed
significantly bigger.
tor matrix
We construct a prior for tangent fields on the sphere by
assuming equal covariance functions k for 3 uncorrelated
parametrizing functions. A mean field is demonstrated in
Figure 1 and the covariance function for the tangent field is
k(x1 , y1 , z1 , x2 , y2 , z2 )·
−y1 y2 − z1 z2
y1 x2
x1 y2
−x1 x2 − z1 z2
x1 z2
y1 z2
z1 x2
.
z1 y2
−x1 x2 − y1 y2
Example 4.2. Maxwell’s equations of electromagnetism
uses curl and divergence operators as building blocks. It is a
well-known result that the solutions of the inhomogeneous
Maxwell equations are parametrized by the electric and magnetic potentials. We verify this and use the parametrization
to construct a Gaussian process, such that its realizations
adhere to Maxwell’s equations. We condition this prior on a
single observation of flowing electric current, which leads
to the magnetic field circling around the current. All computations have been performed in Maple using the Involutive
package (Blinkov et al., 2003).
The homogenous Mawell equations are given by the opera-
0
∂z
−∂y
0
Mh :=
−∂t
0
0
∂x
−∂z
0
∂x
0
0
−∂t
0
∂y
∂y
−∂x
0
0
0
0
−∂t
∂z
∂t
0
0
∂x
0
∂z
−∂y
0
0
∂t
0
∂y
−∂z
0
∂x
0
0
0
∂t
∂z
∂y
−∂x
0
0
applied to three components of the electric field and three
components of the magnetic (pseudo) field. We have set all
constants to 1. The right kernel of Mh is zero, in particular
no parametrization exists.
The inhomogeneous Maxwell equations with three additional components of the electric current and one additional
component of electric
flux aregiven by the 8 × 10 operator
04×4
matrix Mi := Mh
. Using Gröbner bases, one
−I4
computes the right kernel
∂x
∂t
0
0
∂y
∂z
0
0
0
−∂t ∂x
−∂t ∂y
−∂t ∂z
2
2
2
∂x
+ ∂y
+ ∂z
0
0
0
−∂z
∂y
2
2
∂y
+ ∂z
− ∂t2
−∂y ∂x
−∂z ∂x
∂t ∂x
∂t
0
∂z
0
−∂x
−∂y ∂x
2
2
∂x
+ ∂z
− ∂t2
−∂z ∂y
∂t ∂y
0
∂t
−∂y
∂x
0
−∂z ∂x
−∂z ∂y
2
2
∂x
+ ∂y
− ∂t2
∂t ∂z
of Mi and verifies that it is a parametrization of the set of
Algorithmic Linearly Constrained Gaussian Processes
solutions of the inhomogeneous Maxwell equations.
We assume squared exponential covariance functions
k := exp −
1
(x1 − x2 )2 + (y1 − y2 )2
2
+(z1 − z2 )2 + (t1 − t2 )2
and a zero mean function for four uncorrelated parametrising functions (electric and magnetic potential). The resulting covariance matrix can be found in Appendix A and for
a demonstration see Figure 2.
y
> T := E[3];
0
1
1
1
t3 Dt
> Mult( T1, [[ Dt, -tˆ3 ]], T, Alg );
1 0
In particular, after the base change with the matrix T , the
systemis free on one generator and parametrizable by the
0
matrix
.
1
The original system is thus parametrizable by
1
0
B=T·
= 1
.
1
t3 ∂ t
1
1
x
For a parametrizing functions with squared exponential covariance functions k(t1 , t2 ) = exp(− 12 (t1 − t2 )2 ) and a
zero mean function, the covariance function for (x, u) is
"
#
t1 −t2
1
1
t32
2
(t2 −t1 −1)(t1 −t2 −1) exp − (t1 − t2 )
t2 −t1
2
t3
t3 t3
1
2 1
There are no exceptional points in the domain of the transformed system, but T removes {0} from the domain of the
original system. We restrict ourselves to the set of smooth
functions C ∞ (R>0 , R) defined on positive real numbers.
For a demonstration of these priors see Figures 3 and 4.
Figure 2. We condition the prior in Example 4.2 on an electric
current in z-direction and zero electric flux at the origin x = y =
z = t = 0. The diagram shows the mean posterior magnetic
field in the (z, t) = (0, 0)-plane. As expected, it circles around
the point with electric current. This mean field
has closed form
−y
2
2
2
2
1
1
·
.
x + y − 4 exp − 2 x + y
10
x
Example 4.3. We consider the one-dimensional Weyl
algebra R = R(t)h∂t i. It allows a stronger notion of a basis
than a Gröbner basis: the Jacobson form (Jacobson, 1943).
It is similar to the Smith normal form for PID’s, in that
multiplying with invertible matrices on both sides yields
a diagonal matrix. We use the Maple packages Janet and
OreModules (Blinkov et al., 2003; Chyzak et al., 2007).
> with( Janet ): with( OreModules ):
> Alg := DefineOreAlgebra( diff=[ Dt, t
], polynom=[ t ] ):
We consider the time-varying control system
d
3
dt x(t) = t u(t) from Example
1.5.7
in (Quadrat,
2010a). The Jacobson form is just 1 0 .
> E := ElementaryDivisors( [
diff(x(t),t)-tˆ3*u(t) ], [ t ], [ x,
u ] ):
> T1 := E[1];
1
− t3
x(t)
1
u(t)
0
1
2
3
4
5
Figure 3. The state function x(t) of the system in Example 4.3 can
be influenced by assigning an input function u(t). For example
setting x(1) = 0 and u(t) = t41+1 for t ∈ {1, 11
, 12 , . . . , 5}
10 10
leads to the above posterior mean. This model yields x(5) ≈
R5 3
1.436537, close to 1 t4t+1 dt ≈ 1.436551.
A. Covariance matrices
The covariance matrix for the 3-dimensional electric field
E, the 3-dimensional magnetic (pseudo) field B, the 3dimensional electric current C and 1-dimensional electric
flux F in Maxwell’s equations as constructed in Example 4.2
is
KE,E KE,B KE,C KE,F
KB,E KB,B KB,C KB,F
K := 4k ·
KC,E KC,B KC,C KC,F
KF,E KF,B KF,C KF,F
Algorithmic Linearly Constrained Gaussian Processes
1
0
Symposium on Mathematical Theory of Networks and Systems - MTNS 2010, pp. 1657–1661, Budapest, Hungary,
2010.
x(t)
1
u(t)
2
3
Bertinet, A. and Agnan, Thomas C. Reproducing Kernel
Hilbert Spaces in Probability and Statistics. Kluwer
Academic Publishers, 2004.
-1
-2
Figure 4. We prescribe a desired behavior for the state x(t) in
Example 4.3 and let the Gaussian process construct a suitable input
u(t). Starting with x(1) = 1 we give u(t) one time step to control
20 21
x(t) to zero, e.g., by setting x(t) = 0 for t ∈ { 10
, 10 , . . . , 5}.
for matrices
Buchberger, Bruno. An algorithm for finding the basis
elements of the residue class ring of a zero dimensional
polynomial ideal. J. Symbolic Comput., 41(3-4):475–
511, 2006. Translated from the 1965 German original by
Michael P. Abramson.
KE,E = (2 − (t1 − t2 )2 )I3 − v T v
KB,B = (2 − vv T )I3 − v T v
KC,C = (4 + 3(t1 − t2 )2 − (vv T ))v T v
+((vv T − (t1 − t2 )2 )2 − 3(t1 − t2 )2 − 7vv T + 10)I3
KF,F = (t1 − t2 )2 (vv T − 3) + (vv T )2 − 11vv T + 18
KB,E = (t1 − t2 )(v ∧ v)
KC,E = (t1 − t2 )((2 + vv T − (t1 − t2 )2 )I3 − 2v T v)
KF,E = ((t1 − t2 )2 + vv T − 6)v
Calandra, Roberto, Peters, Jan, Rasmussen, Carl E., and
Deisenroth, Marc P. Manifold Gaussian processes for
regression. In International Joint Conference on Neural
Networks, pp. 3338–3345, 2016. doi: 10.1109/IJCNN.
2016.7727626.
Chyzak, Frédéric, Quadrat, Alban, and Robertz, Daniel.
Effective algorithms for parametrizing linear control systems over Ore algebras. Appl. Algebra Engrg. Comm.
Comput., 16(5):319–376, 2005.
KC,B = (vv T − (t1 − t2 )2 − 4)(v ∧ v)
KF,B = 0 0 0
KF,C = −(t1 − t2 )(vv T + (t1 − t2 )2 − 8)v
x1 − x2 y2 − y2 z1 − z2 ,
using v
:=
for
the 3 × 3 identity matrix, and
v ∧ v
0
−(z1 − z2 )
y1 − y2
z1 − z2
0
−(x1 − x2 ).
−(y1 − y2 )
x1 − x2
0
Blinkov, Yuri A., Cid, Carlos F., Gerdt, Vladimir P., Plesken,
Wilhelm, and Robertz, Daniel. The MAPLE Package
JANET: I. Polynomial Systems. II. Linear Partial Differential Equations. In Proceedings 6th International
Workshop on Computer Algebra in Scientific Computing,
pp. 31–40 and 41–54, 2003. (http://www.mathb.
rwth-aachen.de/Janet).
I3
:=
Acknowledgments
The authors thanks M. Barakat, S. Gutsche, C. Kaus, D.
Moser, S. Posur, and O. Wittich for discussions concerning this paper, W. Plesken, A. Quadrat, D. Robertz, and E.
Zerz for introducing him to the algebraic background of
this paper, S. Thewes for introducing him to the stochastic
background of this paper, and the authors of (Jidling et al.,
2017) for providing the starting point of this work.
References
Adams, William W. and Loustaunau, Philippe. An introduction to Gröbner bases. Graduate Studies in Mathematics.
American Mathematical Society, 1994.
Barakat, Mohamed. Purity filtration and the fine structure
of autonomy. In Proceedings of the 19th International
Chyzak, Frédéric, Quadrat, Alban, and Robertz, Daniel.
OreModules: a symbolic package for the study of multidimensional linear systems. In Applications of time
delay systems, volume 352 of Lecture Notes in Control and Inform. Sci., pp. 233–264. Springer, Berlin,
2007. (http://www.mathb.rwth-aachen.de/
OreModules).
Decker, Wolfram, Greuel, Gert-Martin, Pfister, Gerhard, and
Schönemann, Hans. S INGULAR 4-1-0 — A computer
algebra system for polynomial computations. http:
//www.singular.uni-kl.de, 2016.
Deisenroth, Marc Peter, Fox, Dieter, and Rasmussen,
Carl Edward. Gaussian processes for data-efficient learning in robotics and control. IEEE Trans. Pattern Anal.
Mach. Intell., 37(2):408–423, 2015.
Dong, Kun, Eriksson, David, Nickisch, Hannes, Bindel,
David, and Wilson, Andrew Gordon. Scalable log determinants for gaussian process kernel learning. 2017.
(arXiv:1711.03481).
Duvenaud, David. Automatic Model Construction with
Gaussian Processes. PhD thesis, University of Cambridge, 2014.
Algorithmic Linearly Constrained Gaussian Processes
Eisenbud, David. Commutative Algebra with a View Toward
Algebraic Geometry, volume 150 of Graduate Texts in
Mathematics. Springer-Verlag, 1995.
Garnett, Roman, Ho, Shirley, and Schneider, Jeff G. Finding galaxies in the shadows of quasars with Gaussian
processes. In Bach, Francis R. and Blei, David M. (eds.),
ICML, volume 37 of JMLR Workshop and Conference
Proceedings, pp. 1025–1033. JMLR.org, 2015.
Gerdt, Vladimir P. Involutive algorithms for computing
Gröbner bases. In Computational commutative and noncommutative algebraic geometry, volume 196 of NATO
Sci. Ser. III Comput. Syst. Sci., pp. 199–225. 2005.
Grayson, Daniel R. and Stillman, Michael E. Macaulay2,
a software system for research in algebraic geometry.
http://www.math.uiuc.edu/Macaulay2/.
Greuel, G. and Pfister, G. A Singular introduction to commutative algebra. Springer-Verlag, 2002. With contributions by Olaf Bachmann, Christoph Lossen and Hans
Schönemann.
Hensman, James, Fusi, Nicoló, and Lawrence, Neil D. Gaussian processes for big data. In Proceedings of the TwentyNinth Conference on Uncertainty in Artificial Intelligence,
2013.
Lee, Jaehoon, Bahri, Yasaman, Novak, Roman, Schoenholz, Samuel S., Pennington, Jeffrey, and Sohl-Dickstein,
Jascha. Deep neural networks as Gaussian processes,
2017. (arXiv:1711.00165).
Macêdo, Ives and Castro, Rener. Learning divergence-free
and curl-free vector fields with matrix-valued kernels.
Instituto Nacional de Matematica Pura e Aplicada, Brasil,
Tech. Rep, 2008.
Oberst, Ulrich. Multidimensional constant linear systems.
Acta Appl. Math., 20(1-2):1–175, 1990.
Osborne, Michael A., Garnett, Roman, and Roberts,
Stephen J. Gaussian processes for global optimization. In
3rd international conference on learning and intelligent
optimization (LION3), pp. 1–15, 2009.
Quadrat, Alban. An introduction to constructive algebraic analysis and its applications. In Journées
Nationales de Calcul Formel, volume 1 of Les cours
du CIRM, pp. 279–469. CIRM, Luminy, 2010a.
(http://ccirm.cedram.org/ccirm-bin/
fitem?id=CCIRM_2010__1_2_281_0).
Quadrat, Alban. Systèmes et Structures – Une approche
de la théorie mathématique des systèmes par l’analyse
algébrique constructive. April 2010b. Habilitation thesis.
Honkela, Antti, Peltonen, Jaakko, Topa, Hande, Charapitsa, Iryna, Matarese, Filomena, Grote, Korbinian, Stunnenberg, Hendrik G., Reid, George, Lawrence, Neil
D., and Rattray, Magnus. Genome-wide modeling of
transcription kinetics reveals patterns of rna production
delays. Proceedings of the National Academy of Sciences, 112(42):13115–13120, 2015. doi: 10.1073/pnas.
1420404112.
Quadrat, Alban. Grade filtration of linear functional
systems. Acta Appl. Math., 127:27–86, 2013. doi:
10.1007/s10440-012-9791-2.
Izmailov, Pavel, Novikov, Alexander, and Kropotov,
Dmitry. Scalable Gaussian processes with billions of
inducing inputs via tensor train decomposition, 2017.
(arXiv:math/1710.07324).
Robertz, Daniel. Recent progress in an algebraic analysis approach to linear systems. Multidimensional
Syst. Signal Process., 26(2):349–388, April 2015. doi:
10.1007/s11045-014-0280-9.
Jacobson, N. The Theory of Rings. Mathematical surveys
and monographs. American Mathematical Society, 1943.
Seiler, Werner M. and Zerz, Eva. The inverse syzygy problem in algebraic systems theory. PAMM, 10(1):633–634,
2010.
Jaynes, Edwin T. Prior probabilities. IEEE Transactions on
systems science and cybernetics, 4(3):227–241, 1968.
Jaynes, Edwin T. and Bretthorst, G. Larry. Probability
Theory: The Logic of Science. Cambridge University
Press, 2003.
Jidling, Carl, Wahlström, Niklas, Wills, Adrian, and Schön,
Thomas B. Linearly constrained Gaussian processes.
2017. (arXiv:1703.00787).
Rasmussen, Carl Edward and Williams, Christopher K. I.
Gaussian Processes for Machine Learning (Adaptive
Computation and Machine Learning). The MIT Press,
2006.
Simon-Gabriel, C.-J. and Schölkopf, B. Kernel distribution
embeddings: Universal kernels, characteristic kernels and
kernel metrics on distributions. Technical report, 2016.
(arXiv:1604.05251).
Snelson, Edward, Rasmussen, Carl Edward, and Ghahramani, Zoubin. Warped gaussian processes. In Thrun,
Sebastian, Saul, Lawrence K., and Schlkopf, Bernhard
(eds.), NIPS, pp. 337–344. MIT Press, 2003.
Algorithmic Linearly Constrained Gaussian Processes
Solin, Arno, Kok, Manon, Wahlström, Niklas, Schön,
Thomas B., and Särkkä, Simo. Modeling and interpolation of the ambient magnetic field by Gaussian processes.
2015. (arXiv:1509.04634).
Sturmfels, Bernd. What is... a Gröbner basis? Notices of
the AMS, 52(10):2–3, 2005.
Thewes, Silja, Lange-Hegermann, Markus, Reuber,
Christoph, and Beck, Ralf. Advanced Gaussian Process
Modeling Techniques. In Design of Experiments (DoE)
in Powertrain Development. Expert, 2015.
Titsias, Michalis K. Variational learning of inducing variables in sparse Gaussian processes. In Artificial Intelligence and Statistics 12, pp. 567–574, 2009.
Treves, F. Topological Vector Spaces, Distributions and
Kernels. Dover books on mathematics. Academic Press,
1967.
Wahlström, Niklas, Kok, Manon, Schön, Thomas B., and
Gustafsson, Fredrik. Modeling magnetic fields using
Gaussian processes. In in Proceedings of the 38th International Conference on Acoustics, Speech, and Signal
Processing (ICASSP), 2013.
Wilson, Andrew G. and Adams, Ryan Prescott. Gaussian
process kernels for pattern discovery and extrapolation. In
ICML (3), volume 28 of JMLR Workshop and Conference
Proceedings, pp. 1067–1075. JMLR.org, 2013.
Wilson, Andrew G., Dann, Christoph, and Nickisch, Hannes.
Thoughts on massively scalable Gaussian processes.
2015a. (arXiv:1511.01870).
Wilson, Andrew G., Hu, Zhiting, Salakhutdinov, Ruslan, and Xing, Eric P. Deep kernel learning. 2015b.
arXiv:1511.02222).
Zerz, Eva. Topics in multidimensional linear systems theory,
volume 256 of Lecture Notes in Control and Information
Sciences. London, 2000.
Zerz, Eva, Seiler, Werner M, and Hausdorf, Marcus. On the
inverse syzygy problem. Communications in Algebra, 38
(6):2037–2047, 2010.
| 0math.AC
|
Decarbonized Demand Response for Residential
Plug-in Electric Vehicles in Smart Grids
Farshad Rassaei, Wee-Seng Soh and Kee-Chaing Chua
arXiv:1604.00812v1 [math.OC] 4 Apr 2016
Department of Electrical and Computer Engineering
National University of Singapore, Singapore
Email: [email protected], {weeseng,chuakc}@nus.edu.sg
Abstract—Recently, in Paris, the world has reached an agreement whereby many countries commit to bolster their efforts
about reducing adverse climate changes. Hence, we can expect
that decarbonization will even attract more attention in different
energy sectors in near future. In particular, both generation side
and consumption side are required to be run more congruently
and environmentally friendly. Thus, employing the renewables at
the generation side along with our proposed decarbonized demand
response (DDR) at the consumption side could significantly
reduce deleterious impacts on the climate. Such ambition, at the
consumption side, necessitates symbiosis and synergy between the
customers and the retailer, and among customers, respectively. In
other words, there should be some incentive-based collaboration
between customers and the retailer as well as coordination among
customers to make the objective be achieved successfully. In
this paper, we present such matching demand response (DR)
algorithm for residential users owning vehicle-to-grid (V2G)
enabled plug-in electric vehicles (PEVs) who obtain electricity
from a common retailer. The retailer itself is connected to the
wholesale electricity market to purchase and sell electricity.
Furthermore, we explain the details of the existing symbiosis
and synergy in our system. Our simulation results illustrate that
substantial cost savings can be achieved along with pollution
reduction by our proposed technique.
Index Terms—Climate change, demand response (DR), electricity retailer, plug-in electric vehicles (PEVs), power demand
elasticity, residential load, smart grids, vehicle-to-grid (V2G).
I. I NTRODUCTION
Climate change has become one of the major concerns
worldwide. Recently, in December 2015, many countries have
agreed to further enhance their efforts to confront adverse
climate changes which are mainly because of tremendous
green house gases (GHGs) emissions, e.g., CO2 and CH4 [1].
One of the significant reasons for GHG emissions is the
transportation sector. Thus, decarbonization this sector has
attracted much research, e.g, [2]–[6]. Meanwhile, plug-in
electric vehicles (PEVs) are good alternatives for traditional
cars to diminish carbon emissions provided their electricity
consumption is managed properly.
However, in the literature, PEVs’ charging and/or discharging management and scheduling are mainly investigated for
cost savings purposes, e.g., [7]–[10]. In these papers, the
emphasis is mostly on increasing the users’ utility, welfare,
the billing strategies, etc.
On the other hand, in near future, we are going to face a
new paradigm in power system, e.g., new ways of electricity
generation, market liberalization, storage capability, two-way
Fig. 1: The United States of America GHGs emissions by sector in 2014 [13].
electricity delivery, demand side management (DSM), demand
response (DR) and environmentally concious transportation
[11] and [12]. Let us add the salience of decarbonization to
the above list. Hence, a practical technique is incumbent to
consider this new paradigm in order to be competent enough
to be employed in a real-world smart grid.
The share of the electricity generation and transportation
sectors in GHGs emissions depend on many factors, e.g.,
type and age of generators, different regions, traffic congestion
management, etc. It varies from one region to another one in
the world. For instance, Fig. 1 illustrates the GHGs emissions
by sector in 2014 in the U.S. and we can observe that the quota
from electricity generation and transportation sectors stand for
more than half of the overall emissions.
However, we believe that minimizing and even nullifying
the share of GHGs emissions from the transportation and the
power generation sectors will be possible in the long run
by further utilizing renewables, electromobility and proper
DDR techniques. Besides, the emissions from power generators could be reduced in smart grids wherein there is high
penetration of renewables and distributed generation (DG).
Widespread penetration of PEVs could inherently reduce the
GHGs emissions from the transportation sector. Fig. 2 shows
that the share from light duty vehicles (LDV) accounts for
Independent System Operator
Energy
Generator 1
Energy
Retailer 1
The Electricity Market
Energy
Generator 2
Energy
Retailer N
Energy
Generator M
Fig. 2: The share of GHGs emissions in the transportation sector by mode in
the U.S. [13].
Annual standard deviation of the price ($)
120
100
Day−ahead market 2014
Spot market 2014
Day−ahead market 2015
Spot market 2015
80
60
40
20
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time of day (Hour)
Fig. 3: Annual standard deviation of the electricity price in 2014 and 2015
for each our of a day in PJM for DA and RT (spot) markets.
59%. Nonetheless, PEVs’ electricity demand adds a huge
burden on the power generation side.
We should note that striving to make decarbonized energy
supply alone is not adequate [14] nor is electrification of
transportation sector. In order to triumph in GHGs reduction,
congruous DR techniques are also needed. In other words,
taking into account the level of GHGs emissions for the
generation and consumption sides in DR can further diminish
the emissions even in an already electrificated sector.
In this paper, we consider Pennsylvania-Jersey-Maryland
(PJM) day-ahead (DA) and real-time (RT) electricity market.
We use its pricing data for the years 2014 and 2015 [15],
the average prices of electricity per MWh which has been
sold over those two years are very close for both DA and
RT markets in each year, i.e., $48.95 and $48.21, in 2014,
and $33.94 and $33.34 in 2015, respectively. The reason for
cheaper average price in 2015 compared to its antecedent year
could be the unprecedented falling down of the oil price.
Fig. 3 shows the annual standard deviation of electricity
price in 2014 and 2015 for each hour of a day in PJM for
both DA and RT (spot) markets. Although the prices are
Fig. 4: An archetype of a smart electricity system encompassing several
retailers, utility companies and multiple users sharing one electricity retailer
or an aggregator.
much cheaper in 2015, we observe that hourly pricing data
for PJM’s DA and RT markets can be significantly distinct
and unpredictable. Another point is that we see RT market
prices has more fluctuations than the DA market, as we
could expect it. Therefore, the high uncertainty, particularly
in the RT market, can remarkably affect the overall electricity
procurement cost for a retailer especially in the long term. This
fact is much more expected when the power system is relying
on a large number of intermittent energy resources with more
uncertainty.
The role of intelligence along with significant architectures
and concepts in future power systems are reviewed in [16]. A
good overview of DR and their different classifications in a
deregulated electricity market is discussed in [17]. In [18], we
present a statistical modelling and a closed-form expression
for PEVs’ uncoordinated charging power demand. Our paper
[9] proposes a decentralized demand shaping algorithm for an
a priori known demand profile for the next day or for flattening
the aggregated daily demand profile. In [10] we consider both
DA and RT markets of PJM in our proposed DR algorithm.
In this paper, for the transportation sector, by adding the
significance of reducing GHGs emissions, we discuss decarbonized DR (DDR) techniques for residential users owning
vehicle to grid (V2G) enabled PEVs by which we strive to
decrease the emissions from the electric power sector, see Fig.
1. Hence, we contemplate lessening both carbon emissions and
electricity procurement costs.
II. S YSTEM M ODEL
In this section, we provide the underlying model and assumptions of the power system in this paper which entails the
energy markets, the electricity retailers or the aggregators, and
the residential users. Similar models for future smart power
systems are advocated in [8] and [19]. We discuss this model
in the sequel.
Fig. 4 represents our model of a smart electricity system
where multiple users share one electricity retailer or an aggregator. The users’ overall load can be differentiated into two
distinct types; typical household load which normally needs
on-demand electricity supply, e.g. air conditioning, heating,
lighting, audio visual devices, cooking and refrigerator, and
PEV as a flexible or programmable load. Here, the dotted
lines show the underlying information system and the solid
lines represent the power transmission and distribution infrastructure.
We assume that an electricity retailer (which may own its
generation capacity) bids to the energy market, e.g., on a DA
basis. Then, based on its energy needs and the market situation,
it buys electricity from the market at market clearing prices
(MCPs). Then, we assume that the retailer is willing to handle
its customers’ PEVs’ electricity assignments (charging and
discharging) such that the shape of the resulting aggregated
power demand profile matches the electricity profile resulted
from the successful bids in the DA market.
This enables the retailer to minimize its demand from the
RT market -which has more price volatility according to Fig.
3- for balancing the load in the following day. Accordingly,
it can reduce the overall electricity procurement cost. This
cost reduction makes the energy retailer afford to offer more
attractive deals to the customers in the form of pricing,
rewarding, promotions, etc [9].
On the other hand, we assume that there are some incentives
or limits from a regulator or the government which make the
retailers interested in or have to reduce the GHGs emissions.
We note that the incentives and limits can be translated to payoffs and fines, in terms of fulfilment and violation, respectively.
As we indicated in [10], retailers adjust their electricity deals
(purchase and sell) in response to market prices. Nevertheless,
the regulator can put some limits on power consumption, e.g.,
the mean of daily electricity consumption. This could reduce
or cancel the need of turning on the traditional generators
which are accounted for 30% of the GHGs emissions Fig. 1.
In practice, the shaped aggregated profile does not exactly
suite the retailer’s purchased DA energy profile. Thus, the retailer needs to reciprocate the load imbalances in the following
day by referring to the RT market and buy electricity at RT
prices. Therefore, we assume that a retailer should consider
three directions when designing its DR technique: DA market,
RT market and GHGs emissions.
III. A NALYSIS
In this section, we provide the electricity procurement cost
for a retailer bidding to the electricity market and then present
our proposed DDR algorithm.
The overall electricity procurement cost for the next day
can be formulated as follows:
Cost =< lDA , pDA > + < li , pRT >
=
24
X
t=1
DA
pDA
+
t lt
24
X
t=1
i
pRT
t lt ,
(1)
where Cost is the overall electricity procurement cost over the
scheduling horizon. The 24-elements vectors lDA and pDA
represent the power demand and electricity price cleared in
the DA market for the following day. The units of liDA and
pDA
are MWh and $/MWh, respectively. Similarly, li and pRT
i
are load imbalance and the electricity price vectors in the RT
market for the next day. The values of the elements of these
two vectors will be known to the retailer only at each time
slot of the next day.
Then, the following sequential optimization programming
technique is used. The users individually contribute in this
program (see Algorithm 1):
(λ < l0PEV,n , l0A,n + l0−n − ld >
minimize
0
lPEV,n
t0
0t0
0t0
+ lA,n
+ lPEV,n
)),
+(1 − λ)(l−n
(2)
0
0
t0 −1
t0 −1
1
1
[lPEV,n
, · · · , lPEV,n
] = [lPEV,n
, · · · , lPEV,n
],
βn
X
0t
lPEV,n
= EPEV,n −
t=t0
N
X
tX
0 −1
t
lPEV,n
,
(4)
t=αn
P24 PN
t
(lPEV,n
(3)
+
t
lA,n
)
= 1.5 ×
t=1
n=1
t
n=1 (lPEV,n
t
+ lA,n
)
N
,
∀t,
(5)
0t
|lPEV,n
| ≤ pmax , ∀t ∈ TP
PEV,n ,
0t
P
lPEV,n = 0, ∀t ∈
/ TPEV,n ,
t
X
t=t0 −1
0k
SOCPEV,n
+
lPEV,n
≥ 0.2
k=t0
(6)
(7)
× CPEV,n , ∀t ∈ TP
PEV,n ,
(8)
Here, lPEV,n and lA,n show the energy assignment vectors for
user n’s PEV and the aggregated load from its household
appliances, respectively. Besides, l0PEV,n and l0A,n vectors show
the same things whenever load altering is needed in RT market,
see Algorithm 1. Furthermore, ld is the purchased load profile
from DA market. EPEV,n is the user n’s required energy to be
delivered to its PEV which is associated with the total required
charging time TPEV,n as follows:
EPEV,n = ω × TPEV,n ,
(9)
where ω is the charging power rate of the outlet to which it is
plugged in. Likewise, αn and βn represent the arrival time and
t
departure time of the PEV. Furthermore, |lPEV,n
| ≤ pmax limits
the maximum power that can be delivered to/from the PEV, we
may presume pmax = ω, and TP
PEV,n describes the permissible
charging time set or simply the set of time slots during the
PEV’s connection time to the power grid. This is simply the
set of time slots between αn and βn . Constraint (5) limits
excessive power consumption at each time slot. This prevents
the need to turn on or use traditional thermal generators.
Additionally, l−n is the aggregated power profile from other
N − 1 users in the system which is described as follows:
X
l−n =
(lPEV,i + lA,i ).
(10)
i∈N
i6=n
Algorithm 1 Decarbonized Demand Response (DDR)
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
Fig. 5: Distributions of (a) arrival time, (b) departure time, (c) charging time
and (d) initial SOC, for 1,000 electric vehicles.
In (8), CPEV,n is the total storage capacity of the user n’s PEV
and we assume that in case of employing V2G in the system,
PEV’s state of charge (SOC) should not fall below 20% of that
total capacity in order to make sure that the adverse impacts on
PEV’s battery lifetime due to complete depletion are avoided.
We know the fact that (li , pRT ) is unknown to the retailer
a priori, at each time slot t0 of the next day after getting this
information, the retailer may decide to alter the previously
shaped DA demand profile. It may want to minimize its RT
electricity purchase to balance the load if the RT prices rise
unexpectedly and even sell back some of its pre-purchased
electricity from DA market to the RT market by using the
PEVs’ available demand elasticity. RT prices may fluctuate
significantly due to the state of the RT market or contingencies.
We should notice that in the proposed programming method
λ is one for shaping the aggregated demand profile, lN , for
the DA market. Then, as described in Algorithm 1, λ could
take a desired value to alter the aggregated demand profile,
l0N , in response to price fluctuations in the RT market.
As the chances for the price to remain that high during
all the next remaining hours of the day is low [20], reshaping
the load profile by lowering the electricity consumption at that
time slot and purchasing electricity at the further time slots can
yield to a lower electricity procurement total cost in practice.
This is also true for purchasing electricity at those time slots
when price, unexpectedly, falls down significantly. The retailer
may buy extra electricity at those specific time slots (based on
the overall storage capacity coming from connected PEVs).
We should note that the retailer is assumed to be allowed
to employ the existing flexibility (offered by each user’s PEV)
and the diversity (resulting from the users’ different usage
patterns). We refer to these two as the system’s elasticity.
Nevertheless, the electricity consumption behaviours of the
users (their PEVs’ usage patterns) are not to be changed and
hence the algorithm preserves users’ comfort. Moreover, users’
privacy is not violated as the information about their individual
appliances, including PEV, is not revealed.
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
Each user initializes its respective load profile over the
assignment horizon based on its power demands, i.e., ln
for n = 1, . . . , N .
All N users send their initialized load profiles to the
retailer.
while not reaching convergence do
for n = 1 to N do
λ is set 1 in the proposed problem.
The retailer calculates the state information l−n
according to (10) for user n.
The retailer sends (l−n − ld ) to user n.
Each user n solves the proposed problem and
updates its load profile ln .
User n sends back the new demand profile to the
retailer.
The retailer updates ln .
end for
end while
for t = 1 to 24 do
The retailer receives information from RT market, i.e.,
pRT
t .
The retailer decides whether or not it proceeds for
demand altering.
λ is set to a desired value in the proposed problem.
while not reaching convergence do
for n = 1 to N do
The retailer sends demand altering signal at
time slot t to user n.
User n solves problem (2) and updates its
load profile l0n .
User n sends back the new demand profile
to the retailer.
The retailer updates l0n .
The retailer calculates the state information
l0−n according to (10) for user n.
end for
end while
26: end for
24:
25:
The convergence criterion in Algorithm 1 can be simply
assumed as a desired number of iterations of updating all
users’ demand profiles or it can be determined to be lower
than some pre-set mean square error (MSE) between two
subsequent iterations of achieving aggregated demand profiles.
As we have discussed in [9], the convergence is guaranteed to
be obtained. Furthermore, users’ contribution can be modelled
as a cooperative game with complete information wherein a
Nash equilibrium exists [9].
IV. S IMULATION R ESULTS
In this section, we evaluate and present the results of our
proposed model and programming technique articulated in
the previous sections through computer simulations. In the
(a)
800
400
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
(b)
2400
2000
1600
1200
800
400
0
Cleared bids from DA market (kW)
Power Consumption (kW)
1200
1500
1200
900
600
300
0
1 2 3 4 5 6 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time of day (Hour)
Fig. 7: The assumed electricity profile purchased by the retailer from the DA
market by the bids that could be cleared.
450
RT market
DA market
400
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time of Day (Hour)
Fig. 6: Electricity demand profile from (a) normal household appliances, i.e.,
without PEVs and (b) the overall electricity demand profile when users owns
PEVs with different usage patterns based on NHTS data.
simulations, we consider the number of residential users, N ,
to be 1,000 and the horizon for testing and evaluation is
considered to be 24 hours for a DA programming scenario
with a time granularity of one hour.
For the PEVs usage patterns, our data and distributions
are based on 2009 NHTS data [21]. Fig. 5 displays the
distributions for arrival time, departure time, charging time and
PEVs’ state of charge (SOC) at the arrival time. Furthermore,
we considered new standard outlets, NEMA 5-15, with 1.8
kW power transfer limit. We assumed that PEVs are needed
to be fully charged by their respective next departure time.
Additionally, we considered 24 kWh energy storage capacity
for all PEVs according to Nissan Leaf model [22]. Moreover,
we adopted the PJM interconnection electricity market pricing
data for both DA and RT markets in the year 2015 [15] and
assumed that PEVs are all V2G enabled.
Next, we examine the DDR scheme introduced in Algorithm
1. Fig. 6 shows the assumed daily aggregated electricity
demand profile of the users with and without the presence
of PEVs with different usage patterns based on NHTS data.
Fig. 7 shows an assumed electricity profile cleared for the
retailer in the DA market. In other words, it shows the bids
that could be cleared in the market at different hours of the
following day. The results of Algorithm 1 is depicted in Fig. 9.
For this, we assumed λ = 0.5 in (2). For a particular day as an
extreme example- Fig. 8 - it can be observed that at the eighth
hour of the day the highest RT price occurs. Online demand
altering can reduce the aggregated demand from 1258.7 kWh
to 878.2 kWh, i.e., we can obtain almost %30 reduction in
the overall demand at that hour. This is when the constraint
(5) in the proposed programming technique and algorithm is
350
Price ($/MWh)
1600
300
250
200
150
100
50
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time of day (Hour)
Fig. 8: DA and RT prices for 28 November 2015 in PJM Interconnection
electricity market (the number of days with such black swan behaviour in
2014 and 2015 is several).
not complied. In case of the absence and presence of that
constraint -Decarbonization constraint- DDR algorithm results
can be seen in Fig. 9. We see that the picks are curbed in Fig.
9(c). But, the power demand at eighth hour is now 985.6 kWh
which cause some extra cost. This additional money could
be paid back by the regulator to the retailer as subsidies for
instance.
In our simulations, convergence has been attained only after
one single iteration of updating all users’ electricity demand
profiles in Algorithm 1.
TABLE I: OVERALL ENERGY PROCUREMENT COSTS FOR THE RETAILER
Case
Energy procurement cost ($)
1
5,382.3
2
4,463.9
3
4,149.2
4
4,262.8
Power Consumption (kW)
2000
1750
1500
1250
1000
750
500
250
0
2000
1750
1500
1250
1000
750
500
250
0
2000
1750
1500
1250
1000
750
500
250
0
(a)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
(b)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
(c)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
Time of Day (Hour)
Fig. 9: Aggregated power demand profiles for (a) after using DR technique in
[9], (b) when the algorithm in [10] is used and (c) when the proposed DDR
algorithm is employed.
It should be emphasized that this could be achieved since
at that hour of the day we had almost 405 V2G enabled
PEVs available at users’ dwellings. Different results would be
obtained for the other hours of that day. Also, it is obvious that
the amount of cost savings would be dissimilar on weekdays
and in the weekend.
In Table I, we compare the overall electricity procurement
costs for the retailer for four cases: case (1) is purchasing
electricity without any DR, case (2) when DR technique in
[9] is used, i.e., only DA demand shaping is implemented,
case (3) when joint shaping and altering demand is applied as
in [10] and case (4) when DDR is being employed.
It can be seen that in the first case, when no DR method is
used and the power demand is directly purchased from the
RT market, total cost is the highest. For the second case,
around $920 is saved and in the third case the cost is further
reduced by $314.7. In the fourth case, however, when DDR
is employed there is some extra cost, $113.6, because of
complying with the GHGs emissions reduction in Algorithm
1.
V. C ONCLUSION AND F UTURE W ORK
In this paper, we proposed a fast converging and decentralized algorithm for managing V2G enabled PEVs’ electricity
assignments (charging and discharging) in order to simultaneously reduce the overall electricity procurement cost and
GHGs emission for electricity retailers. We illustrated that
with some incentives and/or regulations from the regulator, the
retailer or aggregator could help lessening GHGs emissions
by using our proposed decarbonized demand response (DDR)
technique.
In this work, we emphasised on the importance of considering decarbonization in DR algorithms for PEVs. However,
various other combinatorial optimization methods could be
investigated. Furthermore, regional GHGs emission factors can
be captured into the evaluations in practice.
R EFERENCES
[1] E. Burleson, “Paris agreement and consensus to address climate challenge,” ASIL INSIGHT, Forthcoming, 2016.
[2] H. Zhang, W. Chen, and W. Huang, “Times modelling of transport sector
in china and usa: Comparisons from a decarbonization perspective,”
Applied Energy, vol. 162, pp. 1505–1514, 2016.
[3] K. C. Hargroves, C. Desha, and E. von Weisaecker, “Introducing
carbon structural adjustment: energy productivity and decarbonization
of the global economy,” Wiley Interdisciplinary Reviews: Energy and
Environment, vol. 5, no. 1, pp. 57–67, 2016. [Online]. Available:
http://dx.doi.org/10.1002/wene.181
[4] J. Liu and G. Santos, “Decarbonizing the road transport sector: breakeven point and consequent potential consumers’ behavior for the us
case,” International Journal of Sustainable Transportation, vol. 9, no. 3,
pp. 159–175, 2015.
[5] K. B. Abdallah, M. Belloumi, and D. De Wolf, “International comparisons of energy and environmental efficiency in the road transport
sector,” Energy, vol. 93, pp. 2087–2101, 2015.
[6] S. Mittal, H. Dai, and P. Shukla, “Low carbon urban transport scenarios for china and india: A comparative assessment,” Transportation
Research Part D: Transport and Environment, 2015.
[7] B. G. Kim, S. Ren, M. van der Schaar, and J. W. Lee, “Bidirectional
energy trading and residential load scheduling with electric vehicles in
the smart grid,” IEEE Journal on Selected Areas in Communications,
vol. 31, no. 7, pp. 1219–1234, Jul. 2013.
[8] A. H. Mohsenian-Rad, V. Wong, J. Jatskevich, R. Schober, and
A. Leon-Garcia, “Autonomous demand-side management based on
game-theoretic energy consumption scheduling for the future smart
grid,” IEEE Transactions on Smart Grid, vol. 1, no. 3, pp. 320 –331,
Dec. 2010.
[9] F. Rassaei, W. S. Soh, and K. C. Chua, “Demand response for residential
electric vehicles with random usage patterns in smart grids,” Sustainable
Energy, IEEE Transactions on, vol. 6, no. 4, pp. 1367–1376, Oct 2015.
[10] ——, “Joint shaping and altering the demand profile by residential plugin electric vehicles for forward and spot markets in smart grids,” in
Smart Grid Technologies - Asia (ISGT ASIA), 2015 IEEE Innovative,
Nov 2015, pp. 1–6.
[11] C. Marnay, “Challenges of supply: An evolving energy paradigm [in my
view],” Power and Energy Magazine, IEEE, vol. 11, no. 5, pp. 104–100,
2013.
[12] M. Kutz, Environmentally conscious transportation.
Wiley Online
Library, 2008, vol. 1.
[13] Center for climate change and energy solutions). [Online]. Available:
http://www.c2es.org/facts-figures/us-emissions.
[14] J. H. Williams, A. DeBenedictis, R. Ghanadan, A. Mahone, J. Moore,
W. R. Morrow, S. Price, and M. S. Torn, “The technology path to deep
greenhouse gas emissions cuts by 2050: the pivotal role of electricity,”
science, vol. 335, no. 6064, pp. 53–59, 2012.
[15] Pennsylvania–jersey–maryland (PJM) interconnection. [Online]. Available:
http://www.pjm.com/markets-and-operations/etools/data-miner.
aspx.
[16] T. Strasser, F. Andren, J. Kathan, C. Cecati, C. Buccella, P. Siano,
P. Leitao, G. Zhabelova, V. Vyatkin, P. Vrba et al., “A review of
architectures and concepts for intelligence in future electric energy
systems,” Industrial Electronics, IEEE Transactions on, vol. 62, no. 4,
pp. 2424–2438, 2015.
[17] M. H. Albadi and E. El-Saadany, “A summary of demand response in
electricity markets,” Electric Power Systems Research, vol. 78, no. 11,
pp. 1989–1996, 2008.
[18] F. Rassaei, W. S. Soh, and K. C. Chua, “A statistical modelling and
analysis of residential electric vehicles’ charging demand in smart grids,”
in Innovative Smart Grid Technologies Conference (ISGT), 2015 IEEE
Power Energy Society, Feb 2015, pp. 1–5.
[19] Y. Liu, C. Yuen, S. Huang, N. Ul Hassan, X. Wang, and S. Xie, “Peakto-average ratio constrained demand-side management with consumer’s
preference in residential smart grid,” Selected Topics in Signal Processing, IEEE Journal of, vol. 8, no. 6, pp. 1084–1097, Dec 2014.
[20] J. M. Poterba and L. H. Summers, “Mean reversion in stock prices:
Evidence and implications,” Journal of financial economics, vol. 22,
no. 1, pp. 27–59, 1988.
[21] National household travel survey (2009). [Online]. Available: http:
//nhts.ornl.gov.
[22] Nissan Leaf. [Online]. Available: http://www.nissanusa.com/
electric-cars/leaf.
| 5cs.CE
|
Recursive Sampling for the Nyström Method
arXiv:1605.07583v5 [cs.LG] 3 Nov 2017
Cameron Musco
MIT
[email protected]
Christopher Musco
MIT
[email protected]
November 6, 2017
Abstract
We give the first algorithm for kernel Nyström approximation that runs in linear time in the
number of training points and is provably accurate for all kernel matrices, without dependence
on regularity or incoherence conditions. The algorithm projects the kernel onto a set of s
landmark points sampled by their ridge leverage scores, requiring just O(ns) kernel evaluations
and O(ns2 ) additional runtime. While leverage score sampling has long been known to give
strong theoretical guarantees for Nyström approximation, by employing a fast recursive sampling
scheme, our algorithm is the first to make the approach scalable. Empirically we show that it
finds more accurate, lower rank kernel approximations in less time than popular techniques such
as uniformly sampled Nyström approximation and the random Fourier features method.
1
Introduction
The kernel method is a powerful tool for applying linear learning algorithms (SVMs, linear regression, etc.) to nonlinear problems. The key idea is to map data to a higher dimensional kernel
feature space, where linear relationships correspond to nonlinear relationships in the original data.
Typically this mapping is implicit. A kernel function is used to compute inner products in the
high-dimensional kernel space, without ever actually mapping original data points to the space.
Given n data points x1 , . . . , xn , the n × n kernel matrix K is formed where Ki,j contains the highdimensional inner product between xi and xj , as computed by the kernel function. All computations
required by a linear learning method are performed using the inner product information in K.
Unfortunately, the transition from linear to nonlinear comes at a high cost. Just generating the
entries of K requires Θ(n2 ) time, which is prohibitive for large datasets.
1.1
Kernel approximation
A large body of work seeks to accelerate kernel methods by finding a compressed, often lowrank, approximation K̃ to the true kernel matrix K. Techniques include random sampling and
embedding [AMS01, BBV06, ANW14], random Fourier feature methods for shift invariant kernels
[RR07, RR09, LSS13], and incomplete Cholesky factorization [FS02, BJ02].
One of the most popular techniques is the Nyström method, which constructs K̃ using a subset
of “landmark” data points [WS01]. Once s data points are selected, K̃ (in factored form) takes
just O(ns) kernel evaluations and O(s3 ) additional time to compute, requires O(ns) space to store,
and can be manipulated quickly in downstream applications. E.g., inverting K̃ takes O(ns2 ) time.
The Nyström method performs well in practice [YLM+ 12, GM13, TRVR16], is widely implemented [HFH+ 09, PVG+ 11, IBM14], and is used in many applications under different names such as
“landmark isomap” [DST03] and “landmark MDS” [Pla05]. In the classic variant, landmark points
are selected uniformly at random. However, significant research seeks to improve performance via
data-dependent sampling that selects landmarks which more closely approximate the full kernel
matrix than uniformly sampled ones [SS00, DM05, ZTK08, BW09, KMT12, WZ13, GM13, LJS16].
Theoretical work has converged on leverage score based approaches, as they give the strongest
provable guarantees for both kernel approximation [DMM08, GM13] and statistical performance
in downstream applications [AM15, RCR15, Wan16]. Leverage scores capture how important an
individual data point is in composing the span of the kernel matrix.
Unfortunately, these scores are prohibitively expensive to compute. All known approximation
schemes require Ω(n2 ) time or assume strong conditions on K – e.g. good conditioning or data
“incoherence” [DMIMW12, GM13, AM15, CLV16]. Hence, leverage score-based approaches remain
largely in the domain of theory, with limited practical impact [KMT12, LBKL15, YPW15].
1.2
Our contributions
In this work, we close the gap between strong approximation bounds and efficiency: we present
a new Nyström algorithm based on recursive leverage score sampling which achieves the “best of
both worlds”: it produces kernel approximations provably matching the accuracy of leverage score
methods while only requiring O(ns) kernel evaluations and O(ns2 ) runtime for s landmark points.
Theoretically, this runtime is surprising. In the typical case when s ≪ n, the algorithm evaluates
just a small subset of K, ignoring most of the kernel space inner products. Yet its performance
1
guarantees hold for general kernels, requiring no assumptions on coherence or regularity.
Empirically, the runtime’s linear dependence on n means that our method is the first leverage
score algorithm that can compete with the most commonly implemented techniques, including the
classic uniform sampling Nyström method and random Fourier features sampling [RR07]. Since
our algorithm obtains higher quality samples, we show experimentally that it outperforms these
methods on benchmark datasets – it can obtain as accurate a kernel approximation in significantly
less time. As a bonus, our approximations have lower rank, so they can be stored in less space and
processed more quickly in downstream learning tasks.
1.3
Paper outline
Our recursive sampling algorithm is built on top of a Nyström scheme of Alaoui and Mahoney that
samples landmark points based on their ridge leverage scores [AM15]. After reviewing preliminaries
in Section 2, in Section 3 we analyze this scheme, which we refer to as RLS-Nyström. To simplify
prior work, which studies the statistical performance of RLS-Nyström for specific kernel learning
tasks [AM15, RCR15, Wan16], we prove a strong, application independent approximation guarantee: for any λ, if K̃ is constructed with s = Θ(dλeff log dλeff ) samples1 , where dλeff = tr(K(K + λI)−1 )
is the so-called “λ-effective dimensionality” of K, then with high probability, kK − K̃k2 ≤ λ.
In Appendix C, we show that this guarantee implies prior results on the statistical performance
of RLS-Nyström for kernel ridge regression and canonical correlation analysis. We also use it
to prove new results on the performance of RLS-Nyström for kernel rank-k PCA and k-means
clustering – in both cases just O(k log k) samples are required to give a solution with good accuracy.
After affirming the favorable theoretical properties of RLS-Nyström, in Section 4 we show that
its runtime can be significantly improved using a recursive sampling approach. Intuitively our
algorithm is simple. We show how to approximate the kernel ridge leverage scores using a uniform
sample of 21 of our input points. While the subsampled kernel matrix still has a prohibitive n2 /4
entries, we can recursively approximate it, using our same sampling algorithm. If our final Nyström
approximation will use s landmarks, the recursive approximation only needs rank O(s), which lets
us estimate the ridge leverage scores of the original kernel matrix
O(ns2 ) time.
Since n
in just
2
ns2
ns
2
is cut in half at each level of recursion, our total runtime is O ns + 2 + 4 + ... = O(ns2 ),
significantly improving upon the method of [AM15], which takes Θ(n3 ) time in the worst case.
Our approach builds on recent work on iterative sampling methods for approximate linear
algebra [CLM+ 15, CMM17]. While the analysis in the kernel setting is technical, our final algorithm
is simple and easy to implement. We present and test a parameter-free variation of Recursive RLSNyström in Section 5, confirming superior performance compared to existing methods.
2
Preliminaries
Consider an input space X and a positive semidefinite kernel function K : X × X → R. Let F be an
associated reproducing kernel Hilbert space and φ : X → F be a (typically nonlinear) feature map
such that for any x, y ∈ X , K(x, y) = hφ(x), φ(y)iF . Given a set of n input points x1 , . . . , xn ∈ X ,
define the kernel matrix K ∈ Rn×n by Ki,j = K(xi , xj ).
It will often be natural to consider the kernelized data matrix that generates K. Informally,
′
let Φ ∈ Rn×d be the matrix containing φ(x1 ), ..., φ(xn ) as its rows (note that d′ may be infinite).
1
O(dλeff log dλeff ) samples is within a log factor of the best possible for any low-rank approximation with error λ.
2
K = ΦΦT . While we use Φ for intuition, in our formal proofs we replace it with any matrix
B ∈ Rn×n satisfying BBT = K (e.g. a Cholesky factor).
We repeatedly use the singular value decomposition, which allows us to write any rank r matrix
M ∈ Rn×d as M = UΣVT , where U ∈ Rn×r and V ∈ Rd×r have orthogonal columns (the left and
right singular vectors of M), and Σ ∈ Rr×r is a positive diagonal matrix containing the singular
values: σ1 (M) ≥ σ2 (M) ≥ . . . ≥ σr (M) > 0. M’s pseudoinverse is given by M+ = VΣ−1 UT .
2.1
Nyström approximation
The Nyström method selects a subset of “landmark” points and uses them to construct a low-rank
approximation to K. Given a matrix S ∈ Rn×s that has a single entry in each column equal to 1
so that KS is a subset of s columns from K, the associated Nyström approximation is:
K̃ = KS(ST KS)+ ST K.
(1)
K̃ can be stored in O(ns) space by separately storing KS ∈ Rn×s and (ST KS)+ ∈ Rs×s . Furthermore, the factors can be computed using just O(ns) evaluations of the kernel inner product to form
KS and O(s3 ) time to compute (ST KS)+ . Typically s ≪ n so these costs are significantly lower
than the cost to form and store the full kernel matrix K.
We view Nyström approximation as a low-rank approximation to the dataset in feature space.
Recalling that K = ΦΦT , S selects s kernelized data points ST Φ and we approximate Φ using its
′
′
projection onto these points. Informally, let PS ∈ Rd ×d be the orthogonal projection onto the row
def
span of ST Φ. We approximate Φ by Φ̃ = ΦPS . We can write PS = ΦT S(ST ΦΦT S)+ ST Φ. Since
it is an orthogonal projection, PS PTS = P2S = PS , and so we can write:
K̃ = Φ̃Φ̃T = ΦP2S ΦT = Φ ΦT S(ST ΦΦT S)+ ST Φ ΦT = KS(ST KS)+ ST K.
This recovers the standard Nyström approximation (1). Note that we present the above for
intuition and do not rigorously handle possibly infinite dimensional feature spaces. To formalize
the argument, replace Φ with any B ∈ Rn×n satisfying BBT = K.
3
The RLS-Nyström method
We now introduce the RLS-Nyström method, which uses ridge leverage score sampling to select
landmark data points, and discuss its strong approximation guarantees for any kernel matrix K.
3.1
Ridge leverage scores
In classical Nyström approximation (1), S is formed by sampling data points uniformly at random.
Uniform sampling can work in practice, but it only gives theoretical guarantees under strong regularity or incoherence assumptions on K [Git11]. It will fail for many natural kernel matrices where
the relative “importance” of points is not uniform across the dataset
For example, imagine a dataset where points fall into several clusters, but one of the clusters
is much larger than the rest. Uniform sampling will tend to oversample landmarks from the large
cluster while undersampling or possibly missing smaller but still important clusters. Approximation
of K and learning performance (e.g. classification accuracy) will decline as a result.
3
(a) Uniform landmark sampling.
(b) Improved landmark sampling.
Figure 1: Uniform sampling for Nyström approximation can oversample from denser parts of the
dataset. A better Nyström scheme will select points that more equally cover the relevant data.
To combat this issue, alternative methods compute a measure of point importance that is used
to select landmarks. For example, one heuristic applies k-means clustering to the input and takes
the cluster centers as landmarks [ZTK08]. A large body of theoretical work measures importance
using variations on the statistical leverage scores. One natural variation is the ridge leverage score:
Definition 1 (Ridge leverage scores [AM15]). For any λ > 0, the λ-ridge leverage score of data
point xi with respect to the kernel matrix K is defined as
def
liλ (K) = K(K + λI)−1
i,i
,
(2)
where I is the n × n identity matrix. For any B ∈ Rn×n satisfying BBT = K, we can also write
liλ (K) = bTi (BT B + λI)−1 bi ,
(3)
where bTi ∈ R1×n is the ith row of B.
For conciseness we write liλ (K) as liλ and include the argument only when referring to the ridge
note that
leverage scores of a kernel matrix other than
K. To check that (2) and (3) are equivalent
T
T
−1
T
−1
T
T
bi (B B+λI) bi = B(B B + λI) B i,i . Using the SVD to write B = UΣV and accordingly
−1 T
K = UΣ2 UT confirms that K(K + λI)−1 = B(BT B + λI)−1 BT = UΣ2 Σ2 + λI
U .
+
It’s not hard to check (see [CLM 15]) that the ridge scores can be defined alternatively as:
liλ = minn
y∈R
1 T
kb − yT Bk22 + kyk22 .
λ i
(4)
This formulation provides better insight into the meaning of these scores. Since BBT = K, any
kernel learning algorithm effectively performs linear learning with B’s rows as data points. So the
ridge scores should reflect the relative importance or uniqueness of these rows. From (4) it’s clear
that liλ ≤ 1 since we can set y to the ith standard basis vector. A row bTi will have ridge score ≪ 1
(i.e. is less important) when it’s possible to find a more “spread out” y that uses other rows in B
to approximately reconstruct bTi – in other words when the row is less unique.
4
3.2
Sum of ridge leverage scores
As is standard in leverage score methods, we don’t directly select landmarks to be the points with
the highest scores. Instead, we sample each point with probability proportional to liλ . I.e. if a
point has the highest possible ridge leverage score of 1, we will select it with probability 1 to be a
landmark. If a point has leverage score 1/100, we select it with probability 1/100.2
Accordingly, the number of landmarks selected, which controls K̃’s rank, is a random variable with expectation equal to the sum of the λ-ridge leverage scores. To ensure compact kernel
approximations, we want this sum to be small. Immediately from Definition 1, we have:
Fact 2 (Ridge leverage scores sum to the effective dimension).
n
X
liλ (K) = tr(K(K + λI)−1 ).
(5)
i=1
tr(K(K + λI)−1 ) is a natural quantity, called the “effective dimension” or “degrees of freedom” for
a ridge regression problem on K with regularization λ [HTF02, Zha06]. We use the notation:
def
dλeff = tr(K(K + λI)−1 ).
(6)
dλeff increases monotonically as λ decreases. For any fixed λ it is essentially the smallest possible rank
achievable for K̃ satisfying the approximation guarantee given by RLS-Nyström: kK − K̃k2 < λ.
3.3
The basic sampling algorithm
We can now introduce the RLS-Nyström method of Alaoui and Mahoney as Algorithm 1. Our
pseudocode allows sampling each point by any probability greater than liλ . This is useful later
when we compute ridge leverage scores approximately. Naturally, oversampling landmarks can
only improve K̃’s accuracy. It could cause us to take more samples, but we will always ensure that
the sum of our approximate ridge leverage scores is not much higher than that of the exact scores.
Algorithm 1 RLS-Nyström Sampling
input: x1 , . . . , xn ∈ X , kernel matrix K, ridge parameter λ > 0, failure probability δ ∈ (0, 1/8)
output: kernel approximation K̃
˜lλ > lλ for the λ-ridge leverage score of each x1 , . . . , xn
1: Compute an over-approximation,
io
i
n
P
2: Set pi := min 1, ˜
lλ · 16 log( ˜
lλ /δ) .
i
i
Construct S ∈ Rn×s by sampling x1 , . . . , xn each independently with probability pi . In other
words, for each i add a column to S with a 1 in position i with probability pi .
4: return the Nyström factors KS ∈ Rn×s and (ST KS)+ ∈ Rs×s .
3:
Note that this implementation of RLS-Nyström Sampling does not form K̃ explicitly in Step 4,
as this would take space and time quadratic in n. It simply returns the factors KS ∈ Rn×s and
(ST KS)+ ∈ Rs×s . Any kernel learning method can then access K̃ implicitly. For example, the
kernel method can be implemented as a linear method run on the n × s matrix KS(ST KS)+/2
whose rows serves as a compression of the data points in kernel space
2
To ensure concentration in our sampling algorithm, we will actually take points with probability qliλ where q is
a small oversampling parameter.
5
3.4
Accuracy bounds
Like other leverage scores methods, RLS-Nyström sampling is appealing because it provably approximates any kernel matrix. In particular, we show that the algorithm produces a K̃ which spectrally
approximates K up to a small additive error. This is the strongest type of approximation offered by
any known Nyström method [GM13] and, importantly, it guarantees that K̃ will provide provable
accuracy when used in place of K in many downstream machine learning applications.
Theorem 3 (Spectral error approximation). For any
P λ > 0 and δ ∈ (0, 1/8), Algorithm 1 returns
an S ∈ Rn×s such that with probability 1 − δ, s ≤ 2 i pi and K̃ = KS(ST KS)+ ST K satisfies:
K̃ K K̃ + λI.
When ridge scores are computed exactly,
P
(7)
λ log(dλ /δ) .
p
=
O
d
i i
eff
eff
denotes the standard Loewner matrix ordering on positive semi-definite matrices3 . Note that
(7) immediately implies the well studied (see e.g [GM13]) spectral norm guarantee, kK − K̃k2 ≤ λ.
Intuitively, Theorem 3 guarantees that the K̃ produced by RLS-Nyström well approximates the
top of K’s spectrum (i.e. any eigenvalues > λ) while allowing it to lose information about smaller
eigenvalues, which are less important for many learning tasks.
Proof. It is clear from the view of Nyström approximation as a low-rank projection of the kernelized
data (see Section 2.1) that K̃ K. Formally, for any B ∈ Rn×n with BBT = K:
K̃ = KS(ST KS)+ ST K = BPS BT ,
where PS = BT S(ST BBT S)+ ST B is the orthogonal projection onto the row span of ST B. Since
PS is a projection kPS k2 ≤ 1. So, for any x ∈ Rn :
xT K̃x = xT BPS Bx = kPS Bxk22 ≤ kBxk22 = xT Kx,
which is equivalent to K̃ K. It remains to show that K K̃ + λI.
In Lemma 11, Appendix A, we apply a matrix Bernstein bound [Tro15] to prove that, when S’s
columns are reweighted by the inverse of their sampling probabilities, with probability 1 − δ/2:
1
3
BT B + λI BT SST B + λI
BT B + λI .
2
2
It is not hard to show (Corollary 13, Appendix A) that even if S is unweighted, as in Algorithm 1,
this bound implies the existence of some finite scaling factor C > 0 such that:
BT B C · BT SST B + λI.
(8)
Let P̄S = I − PS be the projection onto the complement of the row span of ST B. By (8):
P̄S BT BP̄S C · P̄S BT SST BP̄S + λP̄S IP̄S .
3
M N means that N − M is positive semidefinite.
6
(9)
Since P̄S projects to the complement of the row span of ST B, ST BP̄S = 0. So (9) gives:
P̄S BT BP̄S 0 + λP̄S IP̄S λI.
In other notation, kP̄S BT BP̄S k2 ≤ λ. This in turn implies kBP̄S BT k2 ≤ λ and hence:
BP̄S BT = B(I − PS )BT λI.
Rearranging and using K = BBT and K̃ = BPS BT gives theP
result. A Chernoff bound (see Lemma
11, Appendix A), gives that with probability 1 − δ/2, s ≤ 2 i pi , completing the theorem.
Often a regularization parameter λ is specified for a learning task, and for near optimal performance on this task, we set the approximation factor in Theorem 3 to ǫλ. In this case we have:
1
Corollary 4 (Tighter spectral error approximation). For any λ > 0 and δ ∈ (0, 1/8),Algorithm
run with ridge parameter ǫλ returns S ∈ Rn×s such that with probability 1 − δ, s = O
dλ
eff
ǫ
log
dλ
eff
δǫ
and K̃ = KS(ST KS)+ ST K satisfies K̃ K K̃ + ǫλI.
−1 1 (K + λI)−1 .
λ
Proof. This follows from Theorem 3 by noting dǫλ
eff ≤ deff /ǫ since (K + ǫλI)
ǫ
Corollary 4 is sufficient to prove that K̃ can be used in place of K without sacrificing performance
on kernel ridge regression and canonical correlation tasks (see [AM15] and [Wan16]). We also use it
to prove a projection-cost preservation guarantee (Theorem 14, Appendix B). Specifically, we show
that if O((k log k)/ǫ) landmarks are sampled with an appropriately chosen ridge parameter λ, then
for any rank-k projection matrix X, K̃ will satisfy, for some fixed c > 0:
tr(K − XKX) ≤ tr(K̃ − XK̃X) + c ≤ (1 + ǫ) tr(K − XKX).
(10)
(10) allows us to prove approximation guarantees for kernel PCA and k-means clustering. Projectioncost preservation has proven a powerful concept in the matrix sketching literature [FSS13, CEM+ 15,
CMM17, BWZ16, CW17]. We hope that an explicit guarantee for kernels will lead to applications
of RLS-Nyström beyond those considered in this work.
Our results on downstream learning bounds that can be derived from Theorem 3 are summarized
in Table 1. Details can be found in Appendices B and C.
4
Recursive sampling for efficient RLS-Nyström
Having established strong approximation guarantees for RLS-Nyström, it remains to provide an
efficient implementation. Specifically, Step 1 of Algorithm 1 naively requires Θ(n3 ) time. We show
that significant acceleration is possible using a recursive sampling approach, which is adapted from
techniques developed in [CLM+ 15] and [CMM17].
4.1
Ridge leverage score approximation via uniform sampling
The key idea is to approximate the ridge leverage scores of K using a uniform sample of the data
points. To ensure accuracy, the sample must be large – a constant fraction of the points. We later
show how to recursively approximate this large sample to achieve our final runtimes. We first prove:
7
Application
Kernel Ridge
Regression w/
Parameter λ
Kernel
k-means
Clustering
Rank k
Kernel PCA
Kernel CCA w/
Regularization
Params λx , λy
∗
Downstream
Guarantee
Relevant
Theorem
Space to
store K̃
Time to
compute K̃
(1 + ǫ) relative error
risk bound
Thm 15
Õ(
(1 + ǫ) relative error
Thm 16
Õ( nk
ǫ )
nk
Õ( nk
ǫ2 ) + Õ( ǫ ) kernel evals.
(1 + ǫ) relative
Frobenius norm error
Thm 17
Õ( nk
ǫ )
nk
Õ( nk
ǫ2 ) + Õ( ǫ ) kernel evals.
Õ(
ndλ
eff
ǫ )
Õ(
ndλ
eff
ǫ
2
n(dλ
eff )
)+
ǫ2
) kernel evals.
2
2
λ
ǫ additive error to
canonical correlation
λ
Thm 18
ndλx +nd y
Õ( eff ǫ eff
)
x )2 +n(d y )2
n(dλ
eff
eff
)+
ǫ2
λy
λx
ndeff +ndeff
Õ(
) kernel evals.
ǫ
Õ(
For conciseness, Õ(·) hides log factors in the failure probability, deff , and k.
Table 1: Downstream guarantees for K̃ obtained from RLS-Nyström (Algorithm 1). For all
problems, the runtime and space cost depends linearly on the number of training data points n.
Lemma 5. For any B ∈ Rn×n with BBT = K and S ∈ Rn×s chosen by sampling each data point
independently with probability 1/2, let
˜lλ = bT (BT SST B + λI)−1 bi
i
i
(11)
P
and pi = min{1, 16˜liλ log( i ˜liλ /δ)} for any δ ∈ (0, 1/8). Then with probability at least 1 − δ:
1. l˜iλ ≥ liλ for all i.
P
P λ
P λ
2.
i pi ≤ 64
i li log( i li /δ).
The first condition ensures that the approximate scores ˜liλ suffice for use in Algorithm 1. The
second ensures that the Nyström approximation obtained will have, up to constant factors, the
same size as if we used the true ridge leverage scores. Note that it is not obvious how to compute
˜lλ using the formula in (11) without explicitly forming B. We discuss how to do this in Section 4.2.
i
Proof. The first bound follows trivially since BT SST B BT B so:
˜lλ = bT (BT SST B + λI)−1 bi ≥ bT (BT B + λI)−1 bi = lλ .
i
i
i
i
The challenge is the second bound. The key observation is that there exists a diagonal reweighting
def
1
matrix W ∈ Rn×n , 0 W I such that for all i, liλ (WKW) ≤ α where α = 12 · 16 log(P
. This
lλ /δ)
i
ensures that uniformly sampling rows with probability 1/2 from the reweighted
is a
P λ kernel
P WKW
λ
valid ridge leverage score sampling. Additionally, |{i : Wi,i < 1}| ≤ 32 log( li /δ) · li . That is,
we do not need to reweight too many columns to achieve the ridge leverage score bound.
Although W is never actually computed, its existence can be proved algorithmically: we can
construct a valid W by iteratively considering any i with liλ (WKW) ≥ α. Since λ > 0, it is always
possible to decrease the ridge leverage score to exactly α by decreasing Wi,i sufficiently.
It is clear from the interpretation of Definition 1 given in (4) that decreasing Wi,i , which
corresponds to decreasing the weight of row i of B, only increases the ridge leverage scores of other
8
rows. So, any reweighted row will always maintain leverage score ≥ α as other rows are reweighted.
Theorem 2 of [CLM+ 15] demonstrates rigorously that the reweighted rows’ leverage scores in fact
converge to α. Further, since W I, it is simple to show (see Lemma 19, Appendix D.1):
X
i
liλ (WKW) ≤
X
def
liλ (K) =
i
X
liλ .
i
Thus, since each reweighted row has liλ (WKW) ≥ α, α · |{i : Wi,i < 1}| ≤
|{i : Wi,i < 1}| ≤
We can now bound
Since liλ (WKW) ≤ 12 ·
P
i pi . For any
1
P
for
16 log( liλ /δ)
P
λ
i li
and so:
X
X
1X λ
li = 32 log
liλ /δ ·
liλ .
α
i
i that is reweighted by W we just trivially bound pi ≤ 1.
all i, and since S samples each i with probability 1/2, by
the matrix Bernstein bound of Lemma 11, with probability 1 − δ/2:
1 T 2
3
(B W B + λI) (BT WSST WB + λI) (BT W2 B + λI).
2
2
Hence:
˜lλ = bT (BT SST B + λI)−1 bi ≤ bT (BT WSST WB + λI)−1 bi
i
i
i
≤ 2bTi (BT W2 B + λI)−1 bi
= 2liλ (WBBT W) = 2liλ (WKW).
Again using that W I and Lemma 19,
X
i
pi =
X
P
˜λ
{i:Wi,i =1} li
pi +
{i:Wi,i <1}
≤2
X
{i:Wi,i =1}
≤ |{i : Wi,i < 1}| + 32 log
= 64 log
4.2
X
liλ /δ
pi
P
X
X
·
liλ .
λ
i li .
Overall:
X
liλ /δ ·
liλ
i
i
Computing ridge leverage scores from a sample
In order to utilize Lemma 5 we must show how to efficiently compute ˜liλ via formula (11) without
explicitly forming either K or B. We prove the following:
Lemma 6. For any sampling matrix S ∈ Rn×s , and any λ > 0:
−1 T
1
T
T
T
T
−1
˜lλ def
S K
K
−
KS
S
KS
+
λI
.
=
b
(B
SS
B
+
λI)
b
=
i
i
i
λ
i,i
It follows that we can compute ˜liλ for all i in O(ns2 ) time using just O(ns) kernel evaluations, to
compute KS and the diagonal of K.
9
Proof. Using the SVD write ST B = ŪΣ̄V̄T . V̄ ∈ Rn×s forms an orthonormal basis for the row
span of ST B. Let V̄⊥ be span for the nullspace of ST B. Then we can rewrite ˜liλ as:
˜lλ = bT BT SST B + λI
i
i
−1
T
bi = bTi V̄, V̄⊥ (Σ̄2 + λI)−1 V̄, V̄⊥ bi .
Here we abuse notation a by letting Σ̄ represent an n × n diagonal matrix whose first s entries are
the singular values of ST B and whose remaining entries are all equal to 0. Now:
˜lλ = bT V̄, V̄⊥ (Σ̄2 + λI)−1 V̄, V̄⊥ T bi = 1 bT V̄T V̄⊥ bi + bT V̄(Σ̄2 + λI)−1 V̄T bT .
i
i
i
i
λ i ⊥
(12)
Focusing on the second term of (12),
1
I − Σ̄2 (Σ̄2 + λI)−1 V̄T bi
λ
1 T
1
= bi V̄V̄T bi − bTi V̄ Σ̄2 (Σ̄2 + λI)−1 V̄T bi .
λ
λ
bTi V̄(Σ̄2 + λI)−1 V̄T bi = bTi V̄
(13)
Focusing on the second term of (13),
bTi V̄ Σ̄2 (Σ̄2 + λI)−1 V̄T bi = bTi V̄Σ̄ŪT Ū(Σ̄2 + λI)−1 ŪT ŪΣ̄V̄T bTi
= bTi BT S(ST KS + λI)−1 ST Bbi .
Substituting back into (13) and then (12), we conclude that:
˜lλ = 1 bT V̄T V̄⊥ bi + 1 bT V̄V̄T bi − 1 bT BT S(ST KS + λI)−1 ST Bbi
i
λ i ⊥
λ i
λ i
1
1
= bTi bi − bTi BT S(ST KS + λI)−1 ST Bbi
λ
λ
−1 T
1
1
= Ki,i −
S K
KS ST KS + λI
.
λ
λ
i,i
We can compute (ST KS+λI)−1 in O(s3 ) ≤ O(ns2 ) time and O(s2 ) ≤ O(ns) kernel evaluations.
−1 T
S K requires just O(ns)
Given this inverse, computing the diagonal entries of KS ST KS + λI
2
kernel evaluations to form KS and O(ns ) time to perform the necessary multiplications. Finally,
computing the diagonal entries of K requires n additional kernel evaluations.
4.3
Recursive RLS-Nyström
We are finally ready to use Lemmas 5 and 6 to give an efficient recursive method for ridge leverage
score Nyström approximation. We show that the output of Algorithm 2, S, is sampled according
to approximate ridge leverage scores for K and so satisfies the approximation bound of Theorem 3.
Theorem 7 (Main Result). Let S ∈ Rn×s be computed by Algorithm 2. With probability 1 − 3δ,
s ≤ 384 · dλeff log(dλeff /δ), S is sampled by overestimates of the λ-ridge leverage scores of K, and thus
by Theorem 3, the Nyström approximation K̃ = KS(ST KS)+ ST K satisfies:
K̃ K K̃ + λI.
Algorithm 2 uses O(ns) kernel evaluations and O(ns2 ) computation time.
10
Algorithm 2 RecursiveRLS-Nyström.
input: x1 , . . . , xm ∈ X , kernel function K : X × X → R, ridge λ > 0, failure prob. δ ∈ (0, 1/32)
output: weighted sampling matrix S ∈ Rm×s
1: if m ≤ 192 log(1/δ) then
2:
return S := Im×m .
3: end if
4: Let S̄ be a random subset of {1, ..., m}, with each i included independently with prob. 12 .
⊲ Let X̄ = {xi1 , xi2 , ..., xi|S̄ | } for ij ∈ S̄ be the data sample corresponding to S̄.
⊲ Let S̄ = [ei1 , ei2 , ..., ei|S̄| ] be the sampling matrix corresponding to S̄.
S̃ := RecursiveRLS-Nyström(X̄, K, λ, δ/3).
Ŝ := S̄ · S̃.
−1
3
7: Set ˜
liλ := 2λ
K − KŜ ŜT KŜ + λI
ŜT K
5:
6:
i,i
for each i ∈ {1, . . . , m} .
⊲ By Lemma 6, equals 23 (B(BT ŜŜT B+λI)−1 BT )i,i . K denotes the kernel matrix for datapoints
{x1 , . . . , xm } and kernel function
K.
P
8: Set pi := min{1, ˜
liλ · 16 log( ˜liλ /δ)} for each i ∈ {1, . . . , m} .
9: Initially set weighted sampling matrix S to be empty. For each i ∈ {1, . . . , m}, with probability
pi , append the column √1pi ei onto S.
10: return S.
√
Note that in Algorithm 2 the columns of S are weighted by 1/ pi . The Nyström approximation
K̃ = KS(ST KS)+ ST K is not effected by column weights (see derivation in Section 2.1). However,
weighting is necessary when the output is used in recursive calls (i.e., when S̃ is used in Step 6).
We prove Theorem 7 via the following intermediate result:
Theorem 8. For any inputs x1 , . . . , xm , K, λ > 0 and δ ∈ (0, 1/32), let K be the kernel matrix for
x1 , . . . , xm and kernel function K and let dλeff (K) be the effective dimension of K with parameter
λ. With probability (1 − 3δ), RecursiveRLS-Nyström outputs S with s columns that satisfies:
3
1 T
(B B + λI) (BT SST B + λI) (BT B + λI)
2
2
for any B with BBT = K.
(14)
def
Additionally, s ≤ smax (dλeff (K), δ) where smax (w, z) = 384 · (w + 1) log ((w + 1)/z). The algorithm
uses ≤ c1 msmax (dλeff (K), δ) kernel evaluations and ≤ c2 msmax (dλeff (K), δ)2 additional computation
time where c1 and c2 are fixed universal constants.
Proof. RecursiveRLS-Nyström is a recursive algorithm and we prove Theorem 8 via induction
on the size of the input, m. In particular, we will show that, if Theorem 8 holds for any all m < n,
then it also holds for m = n. Our base case is m = 1.
Base case: Theorem 8 holds for any inputs as long as m = 1.
Suppose m = 1, so the input data set just consists of a single point x1 . Then the if statement
on Line 1 evaluates to true since 192 log(1/δ) > 1. So, S is set to a 1 × 1 identity matrix and
(14) of Theorem 8 holds trivially since (BT B + λI) = (BT SST B + λI). Furthermore, s = 1 ≤
smax (dλeff (K), δ) for any dλeff (K) and δ, as required. The algorithm runs in O(1) time and performs
no kernel evaluations, so the runtime requirements of Theorem 8 also hold as long as c2 set to a
large enough constant. This all holds with probability 1, and so for any input failure probability δ.
11
Inductive Step: Theorem 8 holds for m = n as long as it holds for all m < n.
Depending on the setting of δ, we split our analysis into 2 cases:
Case 1: The number of input data points n is < 192 log(1/δ).
In this case, as for the base case, the if statement on Line 1 evaluates to true. S is set to an
n × n identity matrix so (14) holds trivially. Furthermore, the number of samples s is equal to
n, and n < 192 log(1/δ) ≤ smax (dλeff (K), δ) as required. Again the algorithm doesn’t compute any
kernel dot products, the runtime bound required by Theorem 8 holds, and all statements hold with
probability 1, which is > 1 − 3δ for any input failure probability δ.
Case 2: The number of input data points n is ≥ 192 log(1/δ).
For this case we will use our inductive assumption since RecursiveRLS-Nyström will call itself
recursively at Step 5, for a smaller input size m < n.
We first note that the expected number of samples taken in Step 4 is n/2. I.e. E |S̄| = n/2. By
a standard multiplicative error Chernoff bound, with high probability the number of samples taken
is not much larger than this expectation. This is important because it tells us that our problem
size decreases substantially before we make the recursive call in Step 5. Following the simplified
Chernoff bounds in e.g. [MU17], when n ≥ 192 log(1/δ), and thus E |S̄| ≥ 96 log(1/δ), we have :
P 1 ≤ |S̄| ≤ .56n ≥ (1 − δ)
(15)
as long as δ < 1/32, as required by Theorem 8.
So, with probability (1 − δ), on Step (5), RecursiveRLS-Nyström is called recursively on
a data set X̄ of size ≥ 1 and ≤ .56n. Accordingly, we can apply our inductive assumption that
Theorem 8 holds for all m between 1 and n − 1 to conclude that, with probability (1 − 3 · δ/3)4 :
1. Let KS̄ denote the kernel matrix for the data points in X̄ (corresponding to the sample S̄
with kernel function K. Then BS̄ = S̄T B satisfies BS̄ BTS̄ = KS̄ . Thus:
1 T
3
(BS¯ BS̄ + λI) (BTS¯ S̃S̃T BS¯ + λI) (BTS¯ BS̄ + λI).
2
2
(16)
2. S̃ has ≤ smax (dλeff (KS̄ ), δ/3) columns.
¯ · smax (dλ (K ), δ/3)
3. The recursive call at Step 5 evaluates K, the kernel function, ≤ c1 · |S|
S̄
eff
2
λ
times and uses ≤ c2 · |S̄| · smax (deff (KS̄ ), δ/3) additional runtime steps.
We first use (16) to prove (14). We can write KS̄ = S̄T KS̄. For all i ∈ {1, . . . n} let
X
−1 T
¯lλ /δ)}.
B
ℓ̄λi = B BT S̄S̄T B + λI
and
p̄i = min{1, 16¯liλ log(
i
i,i
i
By Lemma 5, since S̄ is constructed by sampling with probability 12 , with probability 1 − δ, ∀ i:
ℓ̄λi ≥ ℓλi (K)
n
X
and
i=1
p̄i ≤ 64dλeff (K) log dλeff (K)/δ .
Here ℓλi (K) is the exact ith λ-ridge leverage score of K.
4
Note that in Step 5 we run RecursiveRLS-Nyström with failure probability δ/3
12
(17)
Now, since BS¯ = S̄T B, it follows from (16) and from the well known fact that M N =⇒
N−1 M−1 , that for any vector z,
−1
−1
−1
2 T
z.
z ≤ 2zT BT S̄S̄T B + λI
z ≤ zT BT S̄S̃S̃T S̄T B + λI
z BT S̄S̄T B + λI
3
Accordingly, since we set Ŝ := S̄ · S̃, for all i ∈ {1, . . . , n}
−1
3
λ
T
T
T
ℓ̄i ≤
B B ŜŜ WB + λI
B
≤ 3ℓ̄λi .
2
i,i
(18)
By Lemma 6, the middle term is exactly equal to ˜liλ as computed in Step 7 of RecursiveRLS-Nyström.
So combining (18) and (17) we have that:
ℓ̃λi ≥ ℓλi (K)
n
X
and
i=1
pi ≤ 192dλeff (K) log dλeff (K)/δ .
(19)
The second bound holds because, as computed on Step 8 of RecursiveRLS-Nyström,
X
X
˜lλ /δ)} ≤ 3 min{1, ¯lλ · 16 log(
¯lλ /δ)} = 3p̄i ≤ 192dλ (K) log dλ (K)/δ
pi = min{1, ˜liλ · 16 log(
i
i
i
eff
eff
by (18). Equation (19) guarantees that S is sampled by over-estimates of the ridge leverage scores
and we have a bound on the sum of the sampling probabilities. So, to establish (14), we just apply
the matrix Bernstein results of Lemma 11. We conclude that, with probability (1 − δ),
3
1 T
(B B + λI) (BT SST B + λI) (BT B + λI)
2
2
for any B with BBT = K.
The same lemma guarantees that S will have s columns where
X
1X
pi ≤ s ≤ 2
pi .
2
(20)
pi ≤ 384dλeff (K) log dλeff (K)/δ ≤ smax (dλeff (K), δ) columns.
To finish our proof of Theorem 8, we still need a bound on the number of kernel function
evaluations used by the algorithm and on its overall runtime.
Kernel evaluations are performed both during the recursive call at Step 5 and when computing
approximate leverage scores at Step 7. Let s̃ be the number of columns in S̃, and hence in Ŝ. At
Step 7, K needs to be evaluated n · (s̃ + 1) times: ns̃ times to compute KŜ and n times to compute
the diagonal of K. Additionally, by the 3rd guarantee that comes from our inductive assumption,
we need at most c1 · |S̄| · smax (dλeff (KS̄ ), δ/3) kernel evaluations for the recursive call. We claim that:
2
P
smax (dλeff (KS̄ ), δ/3) ≤ 1.317smax (dλeff (K), δ).
(21)
This follows from Lemma 19: since KS̄ = S̄T KS̄ and S̄S̄T I for any sampling matrix, dλeff (KS̄ ) ≤
dλeff (K). Additionally, we use that log(3/δ) ≤ 1.317 log(1/δ) when δ ≤ 1/32.
Using this bound and (15) we see that our total number of kernel evaluations is bounded by:
n · (s̃ + 1) + c1 · |S̄| · smax (dλeff (KS̄ ), δ/3) ≤ n · (smax (dλeff (KS¯), δ/3) + 1) + c1 · .56n · smax (dλeff (KS̄ ), δ/3)
≤ (2.317 + .74c1 ) n · smax (dλeff (K), δ).
13
As long as c1 > 9, the above is < c1 nsmax (dλeff (K), δ), so we see that RecursiveRLS-Nyström
run on a data set of size n performs no more kernel evaluations than that allowed by Theorem 8.
We finally bound runtime, accounting for the recursive call to RecursiveRLS-Nyström and
all other steps. Again, using the 3rd guarantee from our inductive assumption, (21), and (15) to
bound |S̄|, the recursive call that computes S̃ has runtime at most:
c2 · |S̄| · smax (dλeff (KS̄ ), δ/3)2 ≤ .972c2 n · smax (dλeff (K), δ)2 .
In addition to the recursive call, the remaining runtime of the algorithm is dominated by the
−1
time to compute ŜT KŜ + λI
and then to multiply this matrix by the n × s̃ matrix KŜ at Step
7. Both of these operations and all other steps can be performed in O(s̃3 + ns̃2 ) time. Since s̃ ≤ n,
there is a constant c such that the number of steps required for the algorithm besides the recursive
call is cns̃2 ≤ cnsmax (dλeff (KS̄ ), δ/3)2 . Again applying (21), our runtime is bounded by:
.972c2 n · smax (dλeff (K), δ)2 + cnsmax (dλeff (KS̄ ), δ/3)2
which is ≤ c2 n · smax (dλeff (K), δ)2 as long as c2 ≥ 40c.
The proof of our statements above relied on three events succeeding: (15), (17), that the recursive call satisfies (16) and the two following guarantees. Each of these events fails with probability
at most δ, so we conclude via a union bound that they all succeed with probability 1 − 3δ.
Accordingly, we have proven that Theorem (8) holds for fixed universal constants c1 and c2 for
any input data set of size n as long as it holds for any input data set of size m with 1 ≤ m < n.
Along with our base case, this establishes the theorem for all input sizes.
Proof of Theorem 7. Theorem 7 is nearly a direct corollary of Theorem 8. In our proof of Theorem
3 we show that if
3
1 T
(B B + λI) (BT SST B + λI) (BT B + λI)
2
2
for a weighted sampling matrix S, then even if we remove the weights from S so that it has all unit
entries (they don’t effect the Nyström approximation), K̃ = KS(ST KS)+ ST K satisfies:
K̃ K K̃ + λI.
The runtime bounds alsofollow nearly directly from Theorem 8. In particular,
we have established
2
λ
λ
that O nsmax (deff (K), δ) kernel evaluations and O nsmax (deff (K), δ) additional runtime are required by RecursiveRLS-Nyström. We only needed the upper bound to prove Theorem 8, but
along the way (20) actually
showed that in a successful run of RecursiveRLS-Nyström, S has
Θ dλeff (K) log dλeff (K)/δ columns. Additionally, we may assume that deff (K) ≥ 1/2. If it is not,
then it’s not hard to check (see proof of Lemma 19) that λ must be ≥ kKk. If this is the case,
the guarantee of Theorem 7 is vacuous: any
Nyström approximation K̃ satisfies K̃ K K̃ + λI.
With deff (K) ≥ 1/2, dλeff (K) log dλeff (K)/δ and thus s are Θ(smax (dλeff (K), δ) so we conclude that
Theorem 7 uses O(ns) kernel evaluations and O(ns2 ) additional runtime.
14
5
Empirical Evaluation
We conclude with an empirical evaluation of our recursive Nyström method. We first introduce
a variant of Algorithm 2 where, instead of choosing a regularization parameter λ, the user sets a
sample size s and λ is automatically determined such that s = Θ(dλeff · log(dλeff /δ)). This variant is
practically appealing as it essentially yields the best possible approximation to K for a fixed sample
budget. Additionally, it is necessary in applications to kernel rank-k PCA and k-means clustering,
when λ is unknown, but where we set s ≈ k log k (see Appendices B and C).
5.1
Recursive RLS-Nyström algorithm for fixed sample size
Given a fixed sample size s, we will control λ using the following fact:
Fact 9 (Proven in (31), Appendix B). For any K and integer k, for λ =
1
k
Pn
i=k+1 σi (K),
dλeff ≤ 2k.
If we choose k such that s ≈ k log k then setting λ as above will yield an RLS-Nyström approximation with approximately s sampled columns. The details are given in Algorithm 3.
Algorithm 3 Recursive RLS-Nyström sampling, fixed sample size.
input: x1 , . . . , xm ∈ X , kernel function K : X × X → R, sample size s, failure prob. δ ∈ (0, 1/32)
′
output: sampling matrix S ∈ Rm×s .
1: if m ≤ s then
2:
return S := Im×m .
3: end if
4: Let S̄ be a random subset of {1, ..., m}, with each i included independently with probability
1
2 . ⊲ Let X̄ = {xi1 , xi2 , ..., xi|S¯| } for ij ∈ S̄ be the data sample corresponding to S̄.
⊲ Let S̄ = [ei1 , ei2 , ..., ei|S̄| ] be the sampling matrix corresponding to S̄.
5:
6:
7:
8:
9:
S̃ := RecursiveRLS-Nyström(X̄, K, s, δ/3).
Ŝ := S̄ · S̃.
Set k toPthe maximum integer with ck log(2k/δ) ≤ s, where c is some fixed constant.
⊲ Approximate λ
λ̃ := k1 ni=k+1 σi (ŜT KŜ)
−1
ŜT K
for each i ∈ {1, ..., m}.
Set ˜liλ := λ̃5 K − KŜ ŜT KŜ + λ̃I
i,i
⊲ By Lemma 6, equals 23 (B(BT ŜŜT B+ λ̃I)−1 BT )i,i . K denotes the kernel matrix for datapoints
{x1 , . . . , xm } and kernel function K.
10: Set pi := min{1, ˜
liλ · 16 log(2k/δ)} for each i ∈ {1, ..., , }.
11: Initially set weighted sampling matrix S to be empty. For each i ∈ {1, . . . , m}, with probability
pi , append the column √1pi ei onto S.
12: return S
Theorem 10. For sufficiently
large universal constant c, let k be any positive integer with s ≥
P
′
ck log(2k/δ) and λ = k1 ni=k+1 σi (K). Let S ∈ Rn×s be computed by Algorithm 3. With probability
1 − 3δ, s′ ≤ 2s, S is sampled by overestimates of the λ-ridge leverage scores of K, and the Nyström
approximation K̃ = KS(ST KS)+ ST K satisfies the guarantee of Theorem 3. Algorithm 3 uses
O(ns) kernel evaluations and O(ns2 ) runtime.
15
For the λ given in Theorem 10, we have dλeff = Θ(k). Hence, since we set s = Θ(k log k/δ),
additive error λ is essentially the smallest we can obtain using an s sample Nyström approximation.
The proof of Theorem 10 is similar to that of Theorem 7. We defer it to Appendix D.
5.2
Performance of Recursive RLS-Nyström for kernel approximation
We evaluate Algorithm 3 on the datasets listed in Table 2, comparing against the classic Nyström
method with uniform sampling [WS01] and the random Fourier features method [RR07]. Implementations were in MATLAB and run on a 2.6 GHz Intel Core i7 with 16GB of memory.
# of Data Points
n
515345
581012
331152
48842
Dataset
YearPredictionMSD
Covertype
Cod-RNA
Adult
# of Features
d
90
54
8
110
Link
https://archive.ics.uci.edu/ml/datasets/YearPredictionMSD
https://archive.ics.uci.edu/ml/datasets/Covertype
https://www.csie.ntu.edu.tw/~ cjlin/libsvmtools/datasets/
https://archive.ics.uci.edu/ml/datasets/Adult
Table 2: Datasets downloaded from UCI ML Repository [Lic13], except Cod-RNA [UKM06].
For each dataset, we split categorical features into binary indicatory features and mean center
and normalize all features to have variance 1. We use a Gaussian kernel for all tests, with the
width parameter σ selected via cross validation on regression and classification tasks. To compute
kK − K̃k2 , we only process a random subset of 20k data points since otherwise multiplying by the
full kernel matrix K to compute kK − K̃k2 is prohibitively expensive. Experiments on the full
kernel matrices are discussed in Section 5.3.
10
10
-2
10
-4
0
1000
2000
3000
Samples
(a) Adult
4000
5000
10
kK − K̃k2
kK − K̃k2
kK − K̃k2
0
10
10
10
10 2
10
Recursive RLS-Nystrom
Uniform Nystrom
Random Fourier Features
10 4
2
10 4
Recursive RLS-Nystrom
Uniform Nystrom
Random Fourier Features
4
10
2
kK − K̃k2
Recursive RLS-Nystrom
Uniform Nystrom
Random Fourier Features
10 4
10 0
Recursive RLS-Nystrom
Uniform Nystrom
Random Fourier Features
3
10 2
0
-2
0
500
1000
1500
2000
Samples
10
-2
10
-4
10 1
0
1000
2000
3000
4000
Samples
(b) Covertype
(c) Cod-RNA
5000
10
0
0
1000
2000
3000
4000
5000
Samples
(d) YearPredictionMSD
Figure 2: For a given number of samples, Recursive RLS-Nyström yields approximations with lower
error, measured by kK − K̃k2 . Error is plotted on a logarithmic scale, averaged over 10 trials.
Figure 2 confirms that Recursive RLS-Nyström consistently obtains better kernel approximation
error than the other methods. The advantage of Nyström over random Fourier features is substantial
– this is unsurprising as the Nyström methods are data dependent and based on data projection,
as opposed to pointwise approximation of K. Even between the Nyström methods there is a
substantial difference in kernel approximation, especially for large sample sizes.
As we can see in Figure 3, with the exception of YearPredictionMSD, the better quality of the
landmarks obtained with Recursive RLS-Nyström translates into runtime improvements. While the
cost per sample is higher for our method at O(nd + ns) time versus O(nd + s2 ) for uniform Nyström
and O(nd) for random Fourier features, since RLS-Nyström requires fewer samples it more quickly
16
10
0
10
-1
10
-2
Recursive RLS-Nystrom
Uniform Nystrom
10 -3
10
10
2
10
1
10
0
10
10 2
Recursive RLS-Nystrom
Uniform Nystrom
10
-1
10 -2
-4
0
5
10
Runtime (sec.)
(a) Adult
15
10
-3
0
1
2
3
4
10 0
10 -1
10
-2
10
-3
0
5
1
2
3
Runtime (sec.)
Runtime (sec.)
(b) Covertype
10
Recursive RLS-Nystrom
Uniform Nystrom
1
kK − K̃k2
1
kK − K̃k2
10
kK − K̃k2
kK − K̃k2
obtains K̃ with a given accuracy. K̃ will also have lower rank, which can accelerate processing
in downstream applications. For example, to achieve kK − K̃k2 ≤ 1 for the Covertype dataset,
Recursive RLS-Nyström requires 650 samples in comparison to 3800 for uniform Nyström.
(c) Cod-RNA
4
5
3
Recursive RLS-Nystrom
Uniform Nystrom
10 2
10 1
10
0
0
2
4
6
8
10
Runtime (sec.)
(d) YearPredictionMSD
Figure 3: Especially for small error, Recursive RLS-Nyström typically obtains a fixed level of
approximation faster than uniform sampling. It only underperformed uniform sampling for the
YearPredictionMSD dataset. Results for random Fourier features are excluded from this plot:
while the method is faster than Nyström, it never obtained high enough accuracy to be directly
comparable. Error is plotted on a log scale, with results averaged over 10 trials.
5.2.1
Accelerated recursive method
While Recursive RLS-Nyström typically outperforms classic Nyström, on datasets with relatively
uniform ridge leverage scores, such as YearPredictionMSD, it only narrowly beats uniform sampling
in terms accuracy. As a result it incurs a higher runtime cost since it is slower per sample.
To combat this issue we implement a simple heuristic modification of our algorithm. We note
that the final cost of computing the Nyström factors KS and (ST KS)+ is O(ns + s3 ) for both
methods. Recursive RLS-Nyström is only slower because computing leverage scores at intermediate
levels of recursion takes O(ns2 ) time (Step 9, Algorithm 3) . This cost can be improved by simply
adjusting the regularization λ to restrict
pthe sample size on each recursive call to be < s. Specifically,
we can balance runtimes by taking ≈ (ns + s3 )/n samples on lower levels.
Doing so improves our runtime, bringing the per sample cost down to approximately that of
random Fourier features and uniform Nyström (Figure 4a) while nearly maintaining the same approximation quality. For datasets such as Covertype in which Recursive RLS-Nyström performs
significantly better than uniform sampling, so does the accelerated method (see Figure 4b). However, the performance of the accelerated method does not degrade when leverage scores are relatively
uniform – it still offers the best runtime to approximation quality tradeoff (Figure 4c).
We note that further runtime improvements may be possible. Subsequent work extends fast
ridge leverage score methods to distributed and streaming environments [CLV17]. Empirical evaluation of these techniques could lead to even more scalable, high accuracy Nyström methods.
5.3
Performance of Recursive RLS-Nyström for learning tasks
We conclude by verifying the usefulness of our kernel approximations in downstream learning tasks.
We focus on Covertype and YearPredictionMSD, which each have approximately n = 500, 000 data
points. While full kernel methods do not scale in this regime, Recursive RLS-Nyström does since
its runtime depends linearly on n. For example, on YearPredictionMSD the method requires 307
sec. (averaged over 5 trials) to build a 2, 000 landmark Nyström approximation for 463, 716 training
17
2
Recursive RLS-Nystrom
Uniform Nystrom
Random Fourier Features
Acclerated Recursive RLS-Nystrom
1
0.5
0
0
500
1000
Samples
1500
10
2
10 0
10
2000
10
kK − K̃k2
kK − K̃k2
Runtime (sec)
1.5
Recursive RLS-Nystrom
Uniform Nystrom
Random Fourier Features
Accelerated Recursive RLS-Nystrom
10 4
-2
0
500
1000
1500
Recursive RLS-Nystrom
Uniform Nystrom
Random Fourier Features
Accelerated Recursive RLS-Nystrom
3
10 2
10
1
10
0
2000
0
1
(a) Runtimes for Covertype.
(b) Errors for Covertype.
2
3
4
5
Runtime (sec.)
Samples
(c) Runtime/error tradeoff for
YearPredictionMSD.
Figure 4: Our accelerated Recursive RLS-Nyström, which undersamples at intermediate recursive
calls, nearly matches the per sample runtime of random Fourier features and uniform Nyström
while still providing approximation nearly as good as the standard Recursive RLS-Nyström. For
datasets like YearPredictionMSD with relatively uniform kernel leverage scores, the accelerated
version offers the best runtime vs. approximation tradeoff. All results are averaged over 10 trials.
9.8
Recursive RLS-Nystrom
Uniform Nystrom
Random Fourier Features
Recursive RLS-Nystrom
Uniform Nystrom
Random Fourier Features
9.7
9.6
80
9.5
RMSE
Classification Rate
85
75
9.4
9.3
9.2
9.1
9
70
50
100
Samples
8.9
200
500
1000
2000
Samples
(a) Covertype
(b) YearPredictionMSD
Figure 5: Performance of kernel approximation methods for classification and clustering. For
Covertype, classification error is measured in separating Class 2 from the remaining classes. For
YearPredictionMSD, RMSE is for the unnormalized output. Regularization and kernel parameters
are obtained via cross validation on training data. Test results are averaged over 10 trials with a
fixed test set, as all three algorithms are randomized.
points. Ridge regression using the approximate kernel then requires 208 sec. for a total of 515 sec.
In comparison, the fastest method, random Fourier features, required 43 sec. to build a rank 2, 000
kernel approximation and 222 sec. for regression, for a total time of 265 sec.
For Covertype we performed classification using the LIBLINEAR support vector machine library. For all sample sizes the SVM dominated runtime cost, so Recursive RLS-Nyström was only
marginally slower than uniform Nyström and random Fourier features for a fixed sample size.
In terms of classification rate for Covertype and RMSE error for YearPredictionMSD, as can be
seen in Figure 5, both Nyström methods outperform random features. However, we do not see much
difference between the two Nyström methods. We leave open understanding why the significantly
better kernel approximations discussed in Section 5.2 do not necessarily translate to much better
learning performance, or whether they would make a larger difference for other problems.
18
Acknowledgements
We would like to thank Michael Mahoney for bringing the potential of ridge leverage scores to our
attention and suggesting their possible approximation via iterative sampling schemes. We would
also like to thank Michael Cohen for pointing out (and fixing) an error in our original manuscript
and generally for his close collaboration in our work on leverage score sampling algorithms. Finally,
thanks to Haim Avron for pointing our an error in our original analysis.
References
[AM15]
Ahmed Alaoui and Michael W Mahoney. Fast randomized kernel ridge regression
with statistical guarantees. In Advances in Neural Information Processing Systems
28 (NIPS), pages 775–783, 2015.
[AMS01]
Dimitris Achlioptas, Frank Mcsherry, and Bernhard Schölkopf. Sampling techniques
for kernel methods. In Advances in Neural Information Processing Systems 14
(NIPS), 2001.
[ANW14]
Haim Avron, Huy Nguyen, and David Woodruff. Subspace embeddings for the polynomial kernel. In Advances in Neural Information Processing Systems 27 (NIPS),
pages 2258–2266, 2014.
[Bac13]
Francis Bach. Sharp analysis of low-rank kernel matrix approximations. In Proceedings of the 26th Annual Conference on Computational Learning Theory (COLT),
2013.
[BBV06]
Maria-Florina Balcan, Avrim Blum, and Santosh Vempala. Kernels as features: On
kernels, margins, and low-dimensional mappings. Machine Learning, 65(1):79–94,
2006.
[BJ02]
Francis Bach and Michael I. Jordan. Kernel independent component analysis. Journal
of Machine Learning Research, 3(Jul):1–48, 2002.
[BMD09]
Christos Boutsidis, Michael W. Mahoney, and Petros Drineas. Unsupervised feature
selection for the k-means clustering problem. In Advances in Neural Information
Processing Systems 22 (NIPS), pages 153–161, 2009.
[BW09]
Mohamed-Ali Belabbas and Patrick J. Wolfe. Spectral methods in machine learning: New strategies for very large datasets. Proceedings of the National Academy of
Sciences of the USA, 106:369–374, 2009.
[BWZ16]
Christos Boutsidis, David P. Woodruff, and Peilin Zhong. Optimal principal component analysis in distributed and streaming models. In Proceedings of the 48th Annual
ACM Symposium on Theory of Computing (STOC), 2016.
[CEM+ 15]
Michael B. Cohen, Sam Elder, Cameron Musco, Christopher Musco, and Madalina
Persu. Dimensionality reduction for k-means clustering and low rank approximation.
In Proceedings of the 47th Annual ACM Symposium on Theory of Computing (STOC),
pages 163–172, 2015.
19
[CLL+ 15]
Shouyuan Chen, Yang Liu, Michael Lyu, Irwin King, and Shengyu Zhang. Fast
relative-error approximation algorithm for ridge regression. In Proceedings of the 31st
Annual Conference on Uncertainty in Artificial Intelligence (UAI), pages 201–210,
2015.
[CLM+ 15]
Michael B. Cohen, Yin Tat Lee, Cameron Musco, Christopher Musco, Richard Peng,
and Aaron Sidford. Uniform sampling for matrix approximation. In Proceedings of
the 6th Conference on Innovations in Theoretical Computer Science (ITCS), pages
181–190, 2015.
[CLV16]
Daniele Calandriello, Alessandro Lazaric, and Michal Valko. Analysis of Nyström
method with sequential ridge leverage score sampling. In Proceedings of the 32nd
Annual Conference on Uncertainty in Artificial Intelligence (UAI), pages 62–71, 2016.
[CLV17]
Daniele Calandriello, Alessandro Lazaric, and Michal Valko. Distributed adaptive
sampling for kernel matrix approximation. In Proceedings of the 20th International
Conference on Artificial Intelligence and Statistics (AISTATS), 2017.
[CMM17]
Michael B. Cohen, Cameron Musco, and Christopher Musco. Input sparsity time
low-rank approximation via ridge leverage score sampling. In Proceedings of the 28th
Annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1758–1777,
2017.
[CW17]
Kenneth L. Clarkson and David P. Woodruff. Low-rank PSD approximation in inputsparsity time. In Proceedings of the 28th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA), pages 2061–2072, 2017.
[DM05]
Petros Drineas and Michael W Mahoney. On the Nyström method for approximating
a Gram matrix for improved kernel-based learning. Journal of Machine Learning
Research, 6:2153–2175, 2005.
[DMIMW12] Petros Drineas, Malik Magdon-Ismail, Michael W. Mahoney, and David P. Woodruff.
Fast approximation of matrix coherence and statistical leverage. Journal of Machine
Learning Research, 13:3475–3506, 2012.
[DMM08]
Petros Drineas, Michael W Mahoney, and S Muthukrishnan. Relative-error CUR matrix decompositions. SIAM Journal on Matrix Analysis and Applications, 30(2):844–
881, 2008.
[DST03]
Vin De Silva and Joshua B Tenenbaum. Global versus local methods in nonlinear
dimensionality reduction. In Advances in Neural Information Processing Systems 16
(NIPS), pages 721–728, 2003.
[FS02]
Shai Fine and Katya Scheinberg. Efficient SVM training using low-rank kernel representations. Journal of Machine Learning Research, 2:243–264, 2002.
[FSS13]
Dan Feldman, Melanie Schmidt, and Christian Sohler. Turning big data into tiny
data: Constant-size coresets for k-means, PCA, and projective clustering. In Proceedings of the 24th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA),
pages 1434–1453, 2013.
20
[Git11]
Alex Gittens.
The spectral norm error of the naive Nyström extension.
arXiv:1110.5305, 2011.
[GM13]
Alex Gittens and Michael Mahoney. Revisiting the Nyström method for improved
large-scale machine learning. In Proceedings of the 30th International Conference on
Machine Learning (ICML), pages 567–575, 2013. Full version at arXiv:1303.1849.
[HFH+ 09]
Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer, Peter Reutemann,
and Ian H Witten. The WEKA data mining software: an update. ACM SIGKDD
Explorations Newsletter, 11(1):10–18, 2009.
[HKZ14]
Daniel Hsu, Sham M. Kakade, and Tong Zhang. Random design analysis of ridge
regression. Foundations of Computational Mathematics, 14(3):569–600, 2014.
[HTF02]
Trevor Hastie, Robert Tibshirani, and Jerome Friedman. The elements of statistical
learning: data mining, inference and prediction. Springer, 2nd edition, 2002.
[IBM14]
IBM Reseach Division, Skylark Team. Libskylark: Sketching-based Distributed Matrix
Computations for Machine Learning. IBM Corporation, Armonk, NY, 2014.
[KMT12]
Sanjiv Kumar, Mehryar Mohri, and Ameet Talwalkar. Sampling methods for the
Nyström method. Journal of Machine Learning Research, 13:981–1006, 2012.
[LBKL15]
Mu Li, Wei Bi, James T Kwok, and Bao-Liang Lu. Large-scale Nyström kernel matrix
approximation using randomized SVD. IEEE Transactions on Neural Networks and
Learning Systems, 26(1):152–164, 2015.
[Lic13]
M. Lichman. UCI machine learning repository, 2013.
[LJS16]
Chengtao Li, Stefanie Jegelka, and Suvrit Sra. Fast DPP sampling for Nyström with
application to kernel methods. In Proceedings of the 33rd International Conference
on Machine Learning (ICML), 2016.
[LSS13]
Quoc Le, Tamás Sarlós, and Alexander Smola. Fastfood - Computing Hilbert space
expansions in loglinear time. In Proceedings of the 30th International Conference on
Machine Learning (ICML), pages 244–252, 2013.
[MU17]
Michael Mitzenmacher and Eli Upfal. Probability and Computing: Randomization
and Probabilistic Techniques in Algorithms and Data Analysis. Cambridge university
press, 2017.
[PD16]
Saurabh Paul and Petros Drineas. Feature selection for ridge regression with provable
guarantees. Neural Computation, 28(4):716–742, 2016.
[Pla05]
John Platt. FastMap, MetricMap, and Landmark MDS are all Nyström algorithms. In
Proceedings of the 8th International Conference on Artificial Intelligence and Statistics (AISTATS), 2005.
[PVG+ 11]
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau,
M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python.
Journal of Machine Learning Research, 12:2825–2830, 2011.
21
[RCR15]
Alessandro Rudi, Raffaello Camoriano, and Lorenzo Rosasco. Less is more: Nyström
computational regularization. In Advances in Neural Information Processing Systems
28 (NIPS), pages 1648–1656, 2015.
[RR07]
Ali Rahimi and Benjamin Recht. Random features for large-scale kernel machines.
In Advances in Neural Information Processing Systems 20 (NIPS), pages 1177–1184,
2007.
[RR09]
Ali Rahimi and Benjamin Recht. Weighted sums of random kitchen sinks: Replacing
minimization with randomization in learning. In Advances in Neural Information
Processing Systems 22 (NIPS), pages 1313–1320, 2009.
[SS00]
Alex J Smola and Bernhard Schökopf. Sparse greedy matrix approximation for machine learning. In Proceedings of the 17th International Conference on Machine Learning (ICML), pages 911–918, 2000.
[SS02]
Bernhard Schölkopf and Alexander J Smola. Learning with kernels: support vector
machines, regularization, optimization, and beyond. MIT press, 2002.
[SSM99]
Bernhard Schölkopf, Alexander J. Smola, and Klaus-Robert Müller. Advances in
kernel methods. chapter Kernel principal component analysis, pages 327–352. MIT
Press, 1999.
[Tro15]
Joel A. Tropp. An introduction to matrix concentration inequalities. Foundations
and Trends in Machine Learning, 8(1-2):1–230, 2015.
[TRVR16]
Stephen Tu, Rebecca Roelofs, Shivaram Venkataraman, and Benjamin Recht. Large
scale kernel learning using block coordinate descent. arXiv:1602.05310, 2016.
[UKM06]
Andrew V Uzilov, Joshua M Keegan, and David H Mathews. Detection of non-coding
RNAs on the basis of predicted secondary structure formation free energy change.
BMC bioinformatics, 7(1):173, 2006.
[Wan16]
Weiran Wang. On column selection in approximate kernel canonical correlation analysis. arXiv:1602.02172, 2016.
[Woo14]
David P. Woodruff. Sketching as a tool for numerical linear algebra. Foundations
and Trends in Theoretical Computer Science, 10(1-2):1–157, 2014.
[WS01]
Christopher Williams and Matthias Seeger. Using the Nyström method to speed up
kernel machines. In Advances in Neural Information Processing Systems 14 (NIPS),
pages 682–688, 2001.
[WZ13]
Shusen Wang and Zhihua Zhang. Improving CUR matrix decomposition and the
Nyström approximation via adaptive sampling. Journal of Machine Learning Research, 14:2729–2769, 2013.
[YLM+ 12]
Tianbao Yang, Yu-feng Li, Mehrdad Mahdavi, Rong Jin, and Zhi-Hua Zhou. Nyström
method vs random Fourier features: A theoretical and empirical comparison. In
Advances in Neural Information Processing Systems 25 (NIPS), pages 476–484, 2012.
22
[YPW15]
Yun Yang, Mert Pilanci, and Martin J Wainwright. Randomized sketches for kernels:
Fast and optimal non-parametric regression. Annals of Statistics, 2015.
[YZ13]
Martin Wainwright Yuchen Zhang, John Duchi. Divide and conquer kernel ridge
regression. Proceedings of the 26th Annual Conference on Computational Learning
Theory (COLT), 2013.
[Zha06]
Tong Zhang. Learning bounds for kernel regression using effective data dimensionality.
Learning, 17(9), 2006.
[ZTK08]
Kai Zhang, Ivor W. Tsang, and James T. Kwok. Improved Nyström low-rank approximation and error analysis. In Proceedings of the 25th International Conference
on Machine Learning (ICML), pages 1232–1239, 2008.
A
Ridge leverage score sampling bounds
Here we give the primary matrix concentration results used to bound the performance of ridge
leverage score sampling in Theorems 3, 7, and 10.
λ
˜λ
Lemma 11. For any
δ ∈ (0,
n λ > 0 and
o 1/8), given ridge leverage score approximations li ≥ li for
P
all i, let pi = min 1, 16˜liλ log( ˜liλ /δ) . Let S ∈ Rn×s be selected by sampling the standard basis
√
vectors e1 , . . . , en each independently
with P
probability pi and rescaling selected columns by 1/ pi .
P
With probability 1 − δ, 1/2 · i pi ≤ s ≤ 2 i pi and:
1
3
1
1 T
B B − λI BT SST B BT B + λI,
2
2
2
2
(22)
Proof. Let B = UΣVT be the singular value decomposition of B. By Definition 1:
liλ = bTi BT B + λI
−1
bi = bTi VΣ2 VT + λVVT
−1
= bTi VΣ̄2 VT
bi
T
−2 T
= bi VΣ̄ V bi ,
−1
bi
where Σ̄2i,i = σi2 (B) + λ. For each i ∈ 1, . . . , n define the matrix valued random variable:
Xi =
(
1
pi
− 1 Σ̄−1 VT bi bTi VΣ̄−1 with probability pi
−Σ̄−1 VT bi bTi VΣ̄−1 with probability (1 − pi )
P
Let Y = i Xi . We have E Y = 0. Furthermore, BT SST B = VΣ̄Y Σ̄VT + BT B. If we can
show that kYk2 ≤ 12 , then since VΣ̄2 VT = BT B + λI this would give the desired bound:
1
3
1
1 T
B B − λI BT SST B BT B + λI.
2
2
2
2
To prove that kYk2 is small we use an intrinsic dimension matrix Bernstein inequality. This
inequality will bound the deviation of Y from its expectation as long as we can bound each kXi k2
and we can bound the matrix variance E(Y 2 ).
23
Theorem 12 (Theorem 7.3.1, [Tro15]). LetP
X1 , . . . , Xn be random symmetric matrices such that
for all i, E X = 0 and kXi k2 ≤ L. Let Y = ni=1 Xi . As long we can bound the matrix variance:
E(Y 2 ) Z,
then for for t ≥
p
kZk2 + L/3,
P [kYk ≥ t] ≤ 4
2 /2
tr(Z) kZk−t+Lt/3
e 2
.
kZk2
P
If pi = 1 (i.e. c˜liλ log( liλ̃ /δ) ≥ 1) then Xi = 0 so kXi k2 = 0. Otherwise, we use the fact that:
1
1
b bT λ bi bTi BT B + λI.
˜lλ i i
li
i
(23)
This follows because we can write any x as x = (BT B + λI)−1/2 y for some y. We can then write:
xT bi bTi x = yT (BT B + λI)−1/2 bi bTi (BT B + λI)−1/2 y
≤ kyk22 · k(BT B + λI)−1/2 bi bTi (BT B + λI)−1/2 k2 .
Since (BT B + λI)−1/2 bi bTi (BT B + λI)−1/2 is rank 1, we have:
k(BT B + λI)−1/2 bi bTi (BT B + λI)−1/2 k2 = tr (BT B + λI)−1/2 bi bTi (BT B + λI)−1/2
= bTi (BT B + λI)−1 bi = liλ
(24)
where in the last step we use the cyclic property of the trace. Writing y = (BT B + λI)1/2 x and
plugging back into (24) gives:
xT bi bTi x ≤ kyk22 · liλ = xT (BT B + λI)x · liλ .
Rearranging and using that ˜liλ ≥ liλ gives (23). With this bound in place we get:
1
−1 T
T
−1
−1 T
T
·
Σ̄
V
b
b
V
Σ̄
Σ̄
V
B
B
+
λI
VΣ̄−1 = I.
i
i
˜lλ
i
So we have:
Xi
˜lλ
1
1
1 −1 T
P
I
Σ̄ V bi bTi VΣ̄−1 i I =
P λ I.
pi
pi
16 log
li /δ
16 log
liλ̃ /δ
Next we bound the variance of Y.
#
"
2
X
X
1
2
2
− 1 + (1 − pi ) · Σ̄−1 VT bi bTi VΣ̄−2 VT bi bTi VΣ̄−1
E(Y ) =
E(Xi )
pi
pi
X 1
1
P λ Σ̄−1 VT BT BVΣ̄−1
· lλ̃ · Σ̄−1 VT bi bTi VΣ̄−1
pi i
16 log
li /δ
1
1
P λ Σ2 Σ̄−2
P λ D.
(25)
16 log
16 log
li /δ
li /δ
24
where D1,1 = 1 and Di,i = (Σ2 Σ̄−2 )i,i =
σi2 (B)
2
σi (B)+λ
for all i ≥ 2. Note that kDk2 = 1.
P λ
Then applying Theorem 12 with Z = D/16 log
li /δ we see that:
−1/8
1
1
+
P λ
P λ
1
≤ 4 tr(D)e 16 log( li /δ) 192 log( li /δ) .
P kYk2 ≥
2
(26)
Then we observe that:
X
liλ .
tr(D) ≤ 1 + tr(Σ2 Σ̄−2 ) = 1 + tr K(K + λI)−1 = 1 +
i
Plugging into (26), establishes (27):
!
P λ
X
1
λ
≤4 1+
li · e−2 log( li /δ) ≤ δ/2.
P kYk ≥
2
i
P
Note that here we make the extremely mild assumption that i liλ ≥ 1. If not, we can simply use
a smaller λ that makes this condition true, and will have s = O(1).
All that remains to show is that the sample size s is bounded with high probability. If pi = 1, we
always sample i so there is no variance in s. LetPS ⊆ [1, ..., n] be P
the set ofPindices with pi < 1. The
expected number of points sampled from S is i∈S pi = 16 log( ˜liλ /δ) i∈S ˜liλ . Assume without
P
˜λ
loss of generality that
i∈S li ≥ 1 – otherwise can just increase our leverage score estimates
and increase the expected sample size by at most 1. Then, by a standard Chernoff bound, with
probability at least 1 − δ/2,
X
X
X
X
1
˜lλ ≤ s ≤ 2 · 16 log(
˜lλ .
˜lλ /δ)
˜lλ /δ)
· 16 log(
i
i
i
i
2
i∈S
i∈S
Union bounding over failure probabilities gives the lemma.
Lemma 11 yields an easy corollary about sampling without rescaling the columns in S:
Corollary 13. For
0 and δ ∈o(0, 1/8), given ridge leverage score approximations ˜liλ ≥ liλ for
n any λ >P
all i, let pi = min 16˜liλ log( ˜
liλ /δ), 1 . Let S ∈ Rn×s be selected by sampling, but not rescaling,
the standard
basis vectors
e1 , . . . , en each independently with probability pi . With probability 1 − δ,
P
P
1/2 · i pi ≤ s ≤ 2 i pi and there exists some scaling factor C > 0 such that
BT B C · BT SST B + λI.
Proof. By Lemma 11, if we set C ′ =
1
mini pi
we have:
1
1 T
B B − λI C ′ · BT SST B
2
2
BT B 2C ′ · BT SST B + λI
which gives the corollary by setting C = 2C ′ .
25
(27)
B
Projection-cost preserving kernel approximation
In addition to the basic spectral approximation guarantee of Theorem 3, we prove that, with
high probability, the RLS-Nyström method presented in Algorithm 1 outputs an approximation
K̃ satisfying what is known as a projection-cost preservation guarantee. This approximation also
immediately holds for the efficient implementation of sampling in Algorithm 3.
P
Theorem 14 (Projection-cost preserving kernel approximation). Let λ = kǫ ni=k+1 σi (K). For
n×s such that with probability 1 − δ,
any P
ǫ ∈ (0, 1), δ ∈ (0,
P 1/8), RLS-Nyström returns an S ∈ R
1/2 i pi ≤ s ≤ 2 i pi and the approximation K̃ = KS(SKS)+ SK satisfies, for any rank k
orthogonal projection X and a positive constant c independent of X:
tr(K − XKX) ≤ tr(K̃ − XK̃X) + c ≤ (1 + ǫ) tr(K − XKX).
When ridge leverage scores are computed exactly,
P
i pi
=O
k
ǫ
(28)
k
log δǫ
.
Intuitively, Theorem 14 ensures that the distance from K̃ to any low dimensional subspace
closely approximates the distance from K to the subspace. Accordingly, K̃ can be used in place
of K to approximately solve low-rank approximation problems, both constrained (e.g. k-means
clustering) and unconstrained (e.g. principal component analysis). See Theorems 16 and 17.
Proof. Set c = tr(K) − tr(K̃), which is ≥ 0 since K̃ K by Theorem 3. By linearity of trace:
tr(K̃ − XK̃X) + c = tr(K) − tr(XK̃X).
So to obtain (28) it suffices to show:
tr(XKX) − ǫ tr(K − XKX) ≤ tr(XK̃X) ≤ tr(XKX).
(29)
Since X is a rank k orthogonal projection we can write X = QQT where Q ∈ Rn×k has orthonormal
columns. Applying the cyclic property of the trace, and the spectral bound of Theorem 3:
tr(XK̃X) = tr(QT K̃Q) =
k
X
i=1
qTi K̃qi ≤
k
X
qTi Kqi = tr(QT KQ) = tr(XKX).
i=1
This gives us the upper bound of (29). For the lower bound we apply Corollary 4:
tr(XK̃X) =
k
X
i=1
qTi K̃qi
≥
k
X
i=1
qTi Kqi − kǫλ = tr(XKX) − kǫλ.
(30)
P
Pn
Finally, kǫλ = ǫ ni=k+1 σi (K) ≤ ǫ tr(K − XKX) since tr(K) =
i=1 σi (K) and tr(XKX) ≤
Pk
i=1 σi (K) by the Eckart-Young theorem. Plugging into (30) gives (29), completing the proof.
26
P
We conclude by showing that s is not too large. As in the proof ofPTheorem 3,
s
≤
2
P λ
Pi pλi with
probability 1 − δ. When ridge leverage scores are computed exactly i pi ≤ 16 li log( li /δ).
!
n
X
1 X
λ
σi (K I)−1 )
li = tr(K(K + ǫ
k
i
i=k+1
!
n
1
1 X
≤ tr(K(K +
σi (K I)−1 )
ǫ
k
i=k+1
=
n
1X
ǫ
1
=
ǫ
1
≤
ǫ
Accordingly,
C
P
i pi
σ (K) +
i=1 i
k
X
i=1
k+
σi (K)
Pn
1
k
σi (K) +
n
X
1
i=k+1 k
i=k+1 σi (K)
σi (K)
1 Pn
k
i=k+1 σi (K)
σi (K)
Pn
i=k+1 σi (K)
!
+
n
X
i=k+1
=
σi (K) +
2k
.
ǫ
σi (K)
1 Pn
k
i=k+1 σi (K)
!
(31)
k
= 32 kǫ log δǫ
as desired.
Applications to learning tasks
In this section use our general approximation gaurantees from Theorems 3 and 14 to prove that
the kernel approximations given by RLS-Nyström sampling are sufficient for many downstream
learning tasks. In other words, K̃ can be used in place of K without sacrificing accuracy or
statistical performance in the final computation.
C.1
Kernel ridge regression
We begin with a standard formulation of the ubiquitous kernel ridge regression task [SS02]. Given
input data points x1 , . . . , xn ∈ Rd and labels y1 , . . . , yn ∈ R this problem asks us to solve:
def
α = arg min kKc − yk22 + λcT Kc,
c∈Rn
(32)
which can be done in closed form by computing:
α = (K + λI)−1 y.
For prediction, when we’re given a new input x, we evaluate its label to be:
y=
n
X
αi K(xi , x).
(33)
i=1
C.1.1
Approximate kernel ridge regression
Naively, solving for α exactly requires at least O(n2 ) time to compute K, plus the cost of a direct
or iterative matrix inversion algorithm. Prediction is also costly since it requires a kernel evaluation
with all n training points. These costs can be reduced significantly using Nyström approximation.
27
In particular, we first select landmark points and compute the kernel approximation K̃ =
KS(ST KS)+ ST K. We can then compute an approximate set of coefficients:
def
α̃ = (K̃ + λI)−1 y.
(34)
With a direct matrix inversion, doing so only takes O(ns2 ) time when our sampling matrix S ∈ Rn×s
selects s landmark points. This is a significant improvement on the O(n3 ) time required to invert
the full kernel. Additionally, the cost of multiplying by K̃ + λI, which determines the cost of most
iterative regression solvers, is reduced, from O(n2 ) to O(ns).
To predict a label for a new x, we first compute its kernel product with all of our landmark
points. Specifically, let x(1) , . . . , x(s) be the landmarks selected by S’s columns. Define w ∈ Rs as:
def
wi = K(x(i) , x).
and let
y = wT (ST KS)+ ST Kα̃.
(35)
Computationally, it makes sense to precompute (ST KS)+ ST Kα̃. Then the cost of prediction is just
s kernel evaluations to compute w, plus s additional operations to multiply wT by (ST KS)+ ST Kα̃.
This approach is the standard way of applying Nyström approximation to the ridge regression
problem and there are a number of ways to evaluate its performance. Beyond directly bounding
minimization error for (32) (see e.g. [CLL+ 15, YPW15, YZ13]), one particularly natural approach
is to consider how the statistical risk of the estimator output by our approximate ridge regression
routine compares to that of the exactly computed estimator.
C.1.2
Relative error bound on statistical risk
To evaluate statistical risk we consider a fixed design setting which has been especially -popular
[Bac13, AM15, LJS16, PD16]. Note that more complex statistical models can be analyzed as well
[HKZ14, RCR15]. In this setting, we assume that our observed labels y = [y1 , . . . , yn ] represent
underlying true labels z = [z1 , . . . , zn ] perturbed with noise. For simplicity, we assume uniform
Gaussian noise with variance σ 2 , but more general noise models can be handled with essentially
the same proof [Bac13]. In particular, our modeling assumption is that:
yi = zi + ηi
where ηi ∼ N (0, σ 2 ).
Following [Bac13] and [AM15], we want to bound the expected in sample risk of our estimator
for z, which is computed using the noisy measurements y = z+η. For exact kernel ridge regression,
we can check from (33) that this estimator is equal to Kα. The risk R is:
def
R = E kK(K + λI)−1 (z + η) − zk22
η
= k K(K + λI)−1 − I zk22 + E kK(K + λI)−1 ηk22
η
2 T
−2
= λ z (K + λI)
2
z + σ tr(K2 (K + λI)−2 ).
28
The two terms that compose R are referred to as the bias and variance terms of the risk:
def
bias(K)2 = λ2 zT (K + λI)−2 z
def
variance(K) = σ 2 tr(K2 (K + λI)−2 ).
For approximate kernel ridge regression, it follows from (35) that our predictor for z is K̃α̃.
Accordingly, the risk of the approximate estimator, R̃ is equal to:
R̃ = bias(K̃)2 + variance(K̃)
We’re are ready to prove our main theorem on kernel ridge regression.
Theorem 15 (Kernel Ridge Regression Risk Bound). Suppose K̃ is computed using RLS-Nyström
with approximation parameter ǫλ and failure probability δ ∈ (0, 1/8). Let α̃ = (K̃ + λI)−1 y and let
K̃α̃ be our estimator for z computed with the approximate kernel. With probability 1 − δ:
R̃ ≤ (1 + 3ǫ)R.
By Theorem 7, Algorithm
K̃ with just O(ns) kernel evaluations and O(ns2 ) compu 2λ can compute
d
dλ
eff
tation time, with s = O eff
ǫ log δǫ .
In other words, replacing K with the approximation K̃ is provably sufficient for obtaining a
1 + Θ(ǫ) quality solution to the downstream task of ridge regression.
Proof. The proof follows that of Theorem 1 in [AM15]. First we show that:
bias(K̃) ≤ (1 + ǫ)bias(K).
(36)
At first glance this might appear trivial as Theorem 3 easily implies that
(K̃ + λI)−1 (1 + ǫ)(K + λI)−1
However, this statement does not imply that
(K̃ + λI)−2 (1 + ǫ)2 (K + λI)−2
since (K̃ + λI)−1 and (K + λI)−1 do not necessarily commute. Instead we proceed:
1
bias(K̃) = k(K̃ + λI)−1 zk2
λ
≤ k(K + λI)−1 zk2 + k(K̃ + λI)−1 z − (K + λI)−1 zk2
−1
= k(K + λI)
−1
zk2 + k(K̃ + λI)
(triangle inequality)
−1
[(K + λI) − (K̃ + λI)](K + λI)
= k(K + λI)−1 zk2 + k(K̃ + λI)−1 (K − K̃)(K + λI)−1 zk2
≤ k(K + λI)−1 zk2 + k(K̃ + λI)−1 (K − K̃)k2 k(K + λI)−1 zk2
1
= bias(K) 1 + k(K̃ + λI)−1 (K − K̃)k2 .
λ
29
zk2
(submultiplicativity)
(37)
So we just need to bound k(K̃ + λI)−1 (K − K̃)k2 ≤ ǫ. First note that, by Theorem 3, Corollary 4,
K − K̃ ǫλI
and since (K − K̃) and I commute, it follows that
(K − K̃)2 ǫ2 λ2 I.
(38)
Accordingly,
k(K̃ + λI)−1 (K − K̃)k22 = k(K̃ + λI)−1 (K − K̃)2 (K̃ + λI)−1 k2
≤ ǫ2 λ2 k(K̃ + λI)−2 k2
1
≤ ǫ 2 λ2 2 = ǫ 2 .
λ
So k(K̃ +λI)−1 (K− K̃)k2 ≤ ǫ as desired and plugging into (37) we have shown (36), that bias(K̃) ≤
(1 + ǫ)bias(K). We next show that:
variance(K̃) ≤ variance(K),
where variance(K) = σ 2 tr(K2 (K + λI)−2 ) = σ 2
Pn
i=1
σi (K̃) ≤ σi (K) for all i. It follows that, for every i,
σi (K̃)
σi (K̃) + λ
(39)
2
σi (K)
.
σi (K)+λ
≤
σi (K)
.
σi (K) + λ
≤
n
X
Since K̃ K by Theorem 3,
This in turn implies that
n
X
i=1
σi (K̃)
σi (K̃) + λ
!2
i=1
σi (K)
σi (K) + λ
2
,
which gives (39). Combining (39) and (36) we conclude that, for ǫ < 1,
R(fˆK̃ ) ≤ (1 + ǫ)2 R(fˆK ) ≤ (1 + 3ǫ)R(fˆK ).
C.2
Kernel k-means
Kernel k-means
clustering asks us to partition x1 , . . . , xn , into k cluster sets, {C1 , . . . , Ck }. Let
1 P
µi = |Ci | xj ∈Ci φ(xj ) be the centroid of the vectors in Ci after mapping to kernel space. The goal
is to choose {C1 , . . . , Ck } which minimize the objective:
k X
X
i=1 xj ∈Ci
kφ(xj ) − µi k2F
(40)
It is well known that this optimization problem can be rewritten as a constrained low-rank approximation problem (see e.g. [BMD09] or [CEM+ 15]). In particular, for any clustering C =
30
n×k called the cluster indicator
{C1 , . . . , Ck } we can define
p a rank k orthonormal matrix C ∈ R
matrix for C. Ci,j = 1/ |Cj | if xi is assigned to Cj and Ci,j = 0 otherwise. CT C = I, so CCT is
a rank k projection matrix. Furthermore, it’s not hard to check that:
k X
X
i=1 xj ∈Ci
kφ(xj ) − µi k2F = tr K − CCT KCCT .
(41)
Informally, if we work with the kernalized data matrix Φ, (41) is equivalent to
kΦ − CCT Φk2F .
Regardless, it’s clear that solving kernel k-means is equivalent to solving:
min tr K − CCT KCCT
C∈S
(42)
where S is the set of all rank k cluster indicator matrices. From this formulation, we easily obtain:
Theorem
16 (Kernel k-means Approximation Bound). Let K̃ be computed by RLS-Nyström with
ǫ Pn
λ = k i=k+1 σi (K) and δ ∈ (0, 1/8). Let C̃∗ be the optimal cluster indicator matrix for K̃ and let
C̃ be an approximately optimal cluster indicator matrix satisfying:
tr K̃ − C̃C̃T K̃C̃C̃T ≤ (1 + γ) tr K̃ − C̃∗ C̃∗T K̃C̃∗ C̃∗T .
Then, if C∗ is the optimal cluster indicator matrix for K:
tr K − C̃C̃T KC̃C̃T ≤ (1 + γ)(1 + ǫ) tr K − C∗ C∗T KC∗ C∗T
By Theorem 10, Algorithm 3 can compute K̃ with O(ns) kernel evaluations and O(ns2 ) computation
k
.
time, with s = O kǫ log δǫ
In other words, if we find an optimal set of clusters for our approximate kernel matrix, those
clusters will provide a (1 + ǫ) approximation to the original kernel k-means problem. Furthermore,
if we only solve the kernel k-means problem approximately on K̃, i.e. with some approximation
factor (1 + γ), we will do nearly as well on the original problem. This flexibility allows for the use
of k-means approximation algorithms (since the problem is NP-hard to solve exactly).
Proof. The proof is almost immediate from our bounds on RLS-Nyström:
(Theorem 14)
tr K − C̃C̃T KC̃C̃T ≤ tr K̃ − C̃C̃T K̃C̃C̃T + c
(by assumption)
≤ (1 + γ) tr K̃ − C̃∗ C̃∗T K̃C̃∗ C̃∗T + (1 + γ)c
≤ (1 + γ) tr K̃ − C∗ C∗T K̃C∗ C∗T + (1 + γ)c (optimality of C̃∗ )
≤ (1 + γ) tr K̃ − C∗ C∗T K̃C∗ C∗T + c
(since c ≥ 0)
(Theorem 14)
≤ (1 + γ)(1 + ǫ) tr K − C̃∗ C∗T KC∗ C∗T .
31
C.3
Kernel principal component analysis
We consider the standard formulation of kernel principal component analysis (PCA) presented in
[SSM99]. The goal is to find principal components in the kernel space F that capture as much
variance in the kernelized data as possible. In particular, if we work informally with the kernelized
data matrix Φ, we want to find a matrix Zk containing k orthonormal columns such that:
ΦΦT − (ΦZk ZTk )(ΦZk ZTk )T
is as small as possible. In other words, if we project Φ’s rows to the k dimensional subspace spanned
by Vk ’s columns and then recompute our kernel, we want the approximate kernel to be close to
the original.
We focus in particular on minimizing PCA error according to the metric:
tr ΦΦT − (ΦZk ZTk )(ΦZk ZTk )T = kΦ − ΦZk ZTk k2F ,
(43)
which is standard in the literature [Woo14, ANW14]. As with f in kernel ridge regression, to solve
this problem we cannot write down Zk explicitly for most kernel functions. However, the optimal
Zk always lies in the column span of ΦT , so we can implicitly represent it by constructing a matrix
X ∈ Rn×k such that ΦT X = Zk . It is then easy to compute the projection of any new data vector
onto the span of Zk (the typical objective of principal component analysis) since we can multiply
by ΦT X using the kernel function.
By the Eckart-Young theorem the optimal Zk contains the top k row principal components of Φ.
Accordingly, if we write the singular value decomposition Φ = UΣVT we want to set X = Uk Σ−1
k ,
which can be computed from the SVD of K = UΣ2 UT . Zk will equal Vk and (43) reduces to:
tr(K − ΦVk VkT Φ) = tr(K − Vk VkT K)
n
X
=
σi (K).
(cyclic property)
(44)
i=k+1
Theorem
17 (Kernel PCA Approximation Bound). Let K̃ be computed by RLS-Nyström with
ǫ Pn
λ = k i=k+1 σi (K) and δ ∈ (0, 1/8). From K̃ we can compute a matrix X ∈ Rs×k such that if we
set Z = ΦT SX, with probability 1 − δ:
kΦ −
ΦZZT k2F
≤ (1 + 2ǫ)kΦ −
ΦVk VkT k2F
= (1 + 2ǫ)
n
X
σi (K).
i=k+1
By Theorem 10, Algorithm 3 can compute K̃ with O(ns) kernel evaluations and O(ns2 ) computation
k
.
time, with s = O kǫ log δǫ
Note that S is the sampling matrix used to construct K̃. Z = ΦT SX can be applied to vectors
(in order to project onto the approximate low-rank subspace) using only s kernel evaluations.
Proof. Re-parameterizing Zk = ΦT Y, we see that minimizing (43) is equivalent to minimizing
tr(K − KYY T K)
32
over Y ∈ Rn×k such that (ΦT Y)T ΦT Y = Y T KY = I. Then we re-parameterize again by writing
Y = K−1/2 W where W is an n × k matrix with orthonormal columns. Using linearity and cyclic
property of the trace, we can write:
tr(K − KYY T K) = tr(K) − tr(Y T KKY) = tr(K) − tr(WT KW) = tr(K) − tr(WWT KWWT ).
So, we have reduced our problem to a low-rank approximation problem that looks exactly like the
k-means problem from Section C.2, except without constraints.
Accordingly, following the same argument as Theorem 16, if we find W̃ minimizing:
tr(K̃) − tr(W̃W̃T K̃W̃W̃T ),
then:
n
X
tr(K) − tr(W̃W̃T KW̃W̃T ) ≤ (1 + ǫ) min tr(K) − tr(WWT KWWT ) = (1 + ǫ)
σi (K).
W
i=k+1
W̃ can be taken to equal the top k eigenvectors of K̃, which can be found in O(n · s2 ) time.
However, we are not quite done. Thanks to our re-parameterization this bound guarantees
that ΦT K−1/2 W̃ is a good set of approximate kernel principal components for Φ. Unfortunately,
ΦT K−1/2 W̃ cannot be represented efficiently (it requires computing K−1/2 ) and projecting new
vectors to ΦT K−1/2 W̃ would require n kernel evaluations to multiply by ΦT .
Instead, recalling the definition of PS = ΦT S(ST KT S)+ ST Φ from Section 2.1, we suggest using
the approximate principal components:
PS ΦT K̃−1/2 W̃.
Clearly PS ΦT K̃−1/2 W̃ is orthonormal because:
(PS ΦT K̃−1/2 W̃)T PS ΦT K̃−1/2 W̃ = W̃T K̃−1/2 ΦT PS ΦK̃−1/2 W̃
= W̃T IW̃ = I.
We will argue that it is offers nearly as a good of a solution as ΦT K−1/2 W̃. Specifically, substituting
into (43) gives a value of:
tr(K − ΦPS ΦT K̃−1/2 W̃W̃T K̃−1/2 ΦPS ΦT ) = tr(K) − tr(W̃W̃T K̃−1/2 ΦPS ΦT ΦPS ΦT K̃−1/2 )
= tr(K) − tr(W̃W̃T K̃−1/2 K̃2 K̃−1/2 )
= tr(K) − tr(W̃W̃T K̃).
Compare this to the value obtained from ΦT K−1/2 W̃:
h
i h
i
tr(K) − tr(W̃W̃T KW̃W̃T ) − tr(K) − tr(W̃W̃T K̃W̃W̃T )
k
n
X
ǫ X
w̃iT (K − K̃)w̃i ≤ k
= tr W̃W̃T (K − K̃) = tr W̃T (K − K̃)W̃ =
σi (K).
k
i=1
33
i=k+1
(45)
The last
step follows from Theorem 3 which guarantees that (K − K̃) ǫλI. Recall that we set
ǫ Pn
λ = k i=k+1 σi (K) and each column w̃i of W̃ has unit norm.
We conclude that the cost obtained by PS ΦT K̃−1/2 W̃ is bounded by:
T
tr(K − ΦPS Φ K̃
−1/2
T
W̃W̃ K̃
−1/2
T
T
T
ΦPS Φ ) ≤ tr(K) − tr(W̃W̃ KW̃W̃ ) + ǫ
≤ (1 + 2ǫ)
n
X
n
X
σi (K)
i=k+1
σi (K).
i=k+1
This gives the result. Notice that PS ΦT K̃−1/2 W̃ = ΦT S(ST KT S)+ ST ΦΦT K̃−1/2 W̃ so, if we set:
X = (ST KT S)+ ST K̃1/2 W̃,
our solution can be represented as Z = ΦT SX as desired.
C.4
Kernel canonical correlation analysis
We briefly discuss a final application to canonical correlation analysis (CCA) that follows from
applying our spectral approximation guarantee of Theorem 3 to recent work in [Wan16].
Consider n pairs of input points (x1 , y1 ), ..., (xn , yn ) ∈ (X , Y) along with two positive semidefinite kernels, Kx : X × X → R and Ky : Y × Y → R. Let Fx and Fy and φx : X → Fx and
φy : Y → Fy be the Hilbert spaces and feature maps associated with these kernels. Let Φx and Φy
denote the kernelized X and Y inputs respectively and Kx and Ky denote the associated kernel
matrices.
The goal
We consider standard regularized kernel CCA, followingPthe presentation in [Wan16].
P
is to compute coefficient vectors αx and αy such that fx∗ = ni=1 αxi φx (xi ) and fy∗ = ni=1 αyi φy (yi )
satisfy:
(fx∗ , fy∗ ) = arg max fxT ΦTx Φy fy∗
fx ∈Fx ,fy ∈Fy
subject to
fxT ΦTx Φx fx
fyT ΦTy Φy fy
+ λx kfx k2Fx = 1
+ λy kfy k2Fy = 1
In [Wan16], the kernelized points are centered to their means. For simplicity we ignore centering,
but note that [Wan16] shows how bounds for the uncentered problem carry over to the centered
one.
It can be shown that αx = (Kx + λx I)−1 β x and αy = (Ky + λy I)−1 β y where β x and β y are
the top left and right singular vectors respectively of
T = (Kx + λx I)−1 Kx Ky (Ky + λy I)−1 .
The optimum value of the above program will be equal to σ1 (T).
[Wan16] shows that if K̃x and K̃y satisfy:
K̃x Kx K̃x + ǫλx I
K̃y Ky K̃y + ǫλx I
34
then if α̃x and α̃y are computed using these approximations, the achieved objective function value
will be within ǫ of optimal (see their Lemma 1 and Theorem 1). So we have:
Theorem 18 (Kernel CCA Approximation Bound). Suppose K̃x and K̃y are computed by RLSNyström with approximation parameters ǫλx and ǫλy and failure probability δ ∈ (0, 1/8). If we
solve for α̃x and α̃y , the approximate canonical correlation will be within an additive ǫ of the true
canonical correlation σ1 (T).
and
By Theorem 7, Algorithm 2 can compute K̃
x and K̃y with
O(nsx + nsy ) kernel evaluations
O(ns2x + ns2y ) computation time, with sx = O
D
D.1
x
dλ
eff
ǫ
log
λ
x
dλ
eff
δǫ
and sy = O
deffy
ǫ
λ
log
deffy
δǫ
.
Additional proofs
Effective dimension bound
Lemma 19. For any W ∈ Rn×p with WWT I,
n
X
liλ (WT KW)
i=1
≤
n
X
liλ (K),
i=1
or equivalently, by Fact 2,
dλeff (WT KW) ≤ dλeff (K).
Proof. By Definition 1, liλ = K(K + λI)−1
n
X
i=1
i,i
so
n
X
liλ (K) = tr K(K + λI)−1 =
i=1
σi (K)
.
σi (K) + λ
Take any matrix B ∈ Rn×n such that BBT = K. Note that for any matrix Y, σi (YY T ) = σi (Y T Y)
for any non-zero singular values. Accordingly,
σi (WT KW) = σi (WT BBT W) = σi (BT WWT B) ≤ σi (BT B) = σi (BBT ) = σi (K)
The ≤ step follows from WWT I so BT WWT B BT B. We thus have:
n
X
liλ (WT KW) =
i=1
p
X
i=1
n
n
i=1
i=1
X σi (K)
X
σi (WT KW)
liλ (K),
≤
=
σi (WT KW) + λ
σi (K) + λ
giving the lemma.
D.2
Proof of Theorem 10: fixed sample size guarantees
We now prove Theorem 10, which gives the approximation and runtime guarantees for our fixed
sample size algorithm, Algorithm 3. The theorem follows from the recursive invariant:
35
Theorem 20. With probability 1 − 3δ, Algorithm 3 performs O(ns) kernel evaluations, runs in
O(ns2 ) time, and for any integer k with s ≥ ck log(2k/δ) returns S satisfying, for any B with
BBT = K:
3
1 T
(B B + λI) (BT SST B + λI) (BT B + λI)
2
2
for λ =
1
k
(46)
Pn
i=k+1 σi (K).
Proof. Assume by induction that after forming S̄ via uniformly sampling, the recursive call to
Algorithm 3 returns S̃ such that Ŝ = S̄ · S̃ satisfies:
1 T T
3
(B S̄S̄ B + λ′ I) (BT ŜŜT B + λ′ I) (BT S̄S̄T B + λ′ I).
2
2
P
P
where λ′ = k1 ni=k+1 σi (S̄T KS̄). This implies that λ̃ = k1 ni=k+1 σi (ŜT KŜ) satisfies:
!
!
n
n
X
X
3
1
T
′
T
′
σi (S̄ KS̄) + kλ ≤ λ̃ ≤
σi (S̄ KS̄) + kλ
2k
2k
i=k+1
(47)
i=k+1
′
′
λ ≤ λ̃ ≤ 3λ .
Combining with (47) we have:
1 T T
9
(B S̄S̄ B + λ′ I) (BT ŜŜT B + λ̃I) (BT S̄S̄T B + λ′ I).
2
2
So, for all i, ˜liλ (which is computed using (BT ŜŜT B + λ̃I) and oversampling factor 5 in Step 9
of Algorithm 3) is at least as large as the approximate leverage score computed using S̄ instead of
Ŝ. If we sample by these scores, by Lemma 5 and Lemma 11 we will have with probability 1 − δ:
1 T
3
(B B + λ′ I) (BT SST B + λ′ I) (BT B + λ′ I)
2
2
which implies (46) since λ′ ≤ λ since kS̄k2 ≤ 1 so σi (S̄T KS̄) ≤ σi (K) for all i.
It just remains to show that we do not sample too many points. This can be shown using a
similar reweighting argument to that used in the fixed λ case in Lemma 5. Full details appear in
Lemma 13 of [CMM17]. When forming the reweighting matrix W, decreasing Wi,i will decrease
P
n
th ridge
i=k+1 σi (WKW) and hence will decrease λ. However, it is not hard to show that the i
leverage score will stillPdecrease. So we can find W giving a uniform ridge leverage score upper
bound of α. Let λ′ = ni=k+1 σi (WKW).
Using P
the same argument P
as Lemma 5, we can bound the sum of estimated sampling probabilities
′
′
λ
by 64 log( li (WKW)/δ) · liλ (WKW) ≤ s/5 by Fact 9 if we set c large enough. The runtime
and failure probability analysis is identical to that of Algorithm 2 (Theorem 8) – the only extra
step is computing λ̃ which can be done in O(s3 ) time via an SVD of ŜT KŜ.
Proof of Theorem 10. The theorem follows immediately since Theorem 20 guarantees that in the
final level of recussion K is sampled by overestimates of its λ-ridge leverage scores. The runtime
bound follows from Theorem 20 and the fact that it is possible to compute KS using O(ns) kernel
evaluations and (ST KS)+ using O(ns2 + s3 ) = O(ns2 ) additional time.
36
| 8cs.DS
|
arXiv:1710.05012v1 [cs.IT] 13 Oct 2017
Potential Conditional Mutual Information: Estimators and
Properties
Arman Rahimzamani and Sreeram Kannan∗
Department of Electrical Engineering,
University of Washington, Seattle
October 16, 2017
Abstract
The conditional mutual information I(X; Y |Z) measures the average information that X and Y contain
about each other given Z. This is an important primitive in many learning problems including conditional
independence testing, graphical model inference, causal strength estimation and time-series problems. In several
applications, it is desirable to have a functional purely of the conditional distribution pY |X,Z rather than of the
joint distribution pX,Y,Z . We define the potential conditional mutual information as the conditional mutual
information calculated with a modified joint distribution pY |X,Z qX,Z , where qX,Z is a potential distribution,
fixed airport. We develop K nearest neighbor based estimators for this functional, employing importance
sampling, and a coupling trick, and prove the finite k consistency of such an estimator. We demonstrate that
the estimator has excellent practical performance and show an application in dynamical system inference.
1
Introduction
Given three random variables X, Y, Z, the conditional mutual information I(X; Y |Z) (CMI) is the expected value
of the mutual information between X and Y given Z, and can be expressed as follows [1],
CM IX↔Y |Z (pX,Y,Z ) := I(X; Y |Z) = D(pX,Y,Z ||pZ pX|Z pY |Z ).
(1)
Thus CMI is a functional of the joint distribution pX,Y,Z . A basic property of CMI, and a key application, is
the following: I(X; Y |Z) = 0 iff X is independent of Y given Z. This measure depends on the joint distribution
between the three variables pX,Y,Z . There are certain circumstances where such a dependence on the entire joint
distribution is not favorable, and a measure that depends purely on the conditional distribution pY |X,Z is more
useful. This is because, in a way, conditional independence can be well defined purely in terms of the conditional
distribution and the measure pX,Z is extraneous. This motivates the direction that we explore in this paper: we
define potential conditional mutual information as a function purely of pY |X,Z evaluated with a distribution qX,Z
that is fixed a-priori.
An Example: Consider the following causal graph where X → Y , Z → X and Z → Y shown in Figure 1a. Let
us say pY |X,Z has a strong dependence on both X and Z, say defined by the structural equation Y = X + Z. We
would like to measure the strength of the edge X → Y in this causal graphical model. One natural measure in this
context is I(X; Y |Z). However, if we use I(X; Y |Z) as the strength, the strength goes to zero when Z ≈ X and
this is undesirable. In such a case, Janzing et al [2] pointed out that a better strength of causal influence is given
by the following:
C(X → Y ) := D(pX,Y,Z ||pZ pX pY |Z ).
∗ Email:
[email protected] and [email protected]
1
(2)
(a)
(b)
Figure 1: (a): A causal graph, where the interest is in determining the strength of X to Y . (b) A gene expression
trace as a function of time for a few example genes.
This causal measure satisfies certain axioms laid out in that paper and is nonzero in the aforesaid example. However,
in the case that the distribution pX approaches a deterministic distribution (X is approximately a constant), this
measure becomes zero, irrespective of the fact that the relationship from X and Z to Y remains unaltered. We would
like to define a potential dependence measure that is dependent purely on pY |X,Z and which has no dependence on
the observed pX,Z . We note that such a measure should give a (strong) non-zero result if Y = X + Z.
The Measure: We define potential conditional information measure as the conditional mutual information
evaluated under a predefined distribution qX,Z , and denote it as qCM I(X ↔ Y |Z), and express it as follows.
qCM IX↔Y |Z (pY |X,Z ) := CM IX↔Y |Z (qX,Z pY |X,Z ).
(3)
A Simple Property: The main question here is how to choose qX,Z . A simple property that maybe of interest
is the following, which can be easily stated in case that all three variables X, Y, Z are discrete. In such a case,
it will be useful if we can have that qCM IX↔Y |Z (pY |X,Z ) = 0 if and only if pY |X,Z depends purely only on Z.
Such a property will be true for qCMI as long as qX,Z is non-zero for every value of X, Z. In case that all three
variables are real valued, a similar statement can be asserted when qX,Z is a positive everywhere density, under the
assumption that pX,Y,Z induces a joint density.
Instantiations: We will propose some instantiations of the potential CMI by giving examples of the distribution
qX,Z .
• CMI: qX,Z = pX,Z . Here the qX,Z is the factual measure and hence qCMI devolves to pure CMI.
• uCMI: qX,Z = uX,Z , where uX,Z is the product uniform distribution on X, Z. This is well defined when X, Z
is either discrete or has a joint density with a bounded support.
• nCMI: qX,Z = nX,Z , where nX,Z is the i.i.d. Gaussian distribution on X, Z. This is well defined when X, Z
are real-valued (whether scalar or vector).
• maxCMI = maxqXZ CMI(qX,Z pY |X,Z ) is defined as an analog of the Shannon capacity in the conditional
case, where we maximize the CMI over all possible distributions on X, Z. This is akin to tuning the input
distribution to maximize the signal in the graph. Note that uCMI or nCMI is not invariant to invertible
2
functional transformations on X, Z, whereas maxCMI is indeed invariant to such functional transformations.
• iCMI: qX,Z = pX pZ is the CMI evaluated not under the true joint distribution of X, Z but under the product
distribution on X, Z. This measure is related to the causal strength measure proposed in [2], though not
identical.
Note that uCMI, nCMI, maxCMI all satisfy the property that they are zero if and only if pY |X,Z has no
dependence on X, whereas CMI and iCMI do not.
Applications: A key application of the potential information measures is in testing graphical models, where
conditional independence tests are the basic primitive by which models are built [3, 4, 5]. To give a concrete
example of the setting, which motivated us to pursue this line of study, consider the following problem, which
can model gene regulatory network inference from time-series data. We observe a set of n time series, Xi (t)
for t = 1, 2, ...T with i = 1, 2, ..n and wish to infer the graph of the dynamical system. The underlying model
~ is a markov chain with Xi (t) depending only on Xj (t − 1) for j ∈ P a(i) and the goal is to
assumption is that X
determine P a(i), the set of parents of a given node. This was originally studied in the setting when the variables are
jointly Gaussian and hence the dependence is linear (see [6] for the original treatment, and [7, 8] for versions with
latent variables). This problem was generalized to the setting with arbitrary probability distributions and temporal
dependences in [9] and studied further in [10], for one-step markov chains in [11] and deterministic relationships in
[12]. From these works, under some technical condition, we can assert that the following method is guaranteed to
be consistent,
xi → xj ⇐⇒ I{Xi (t − 1); Xj (t)|Xic (t − 1)} > 0.
(4)
Thus to solve this problem, we estimate the CMI between the aforesaid variables. However, we observed while
experimenting with gene regulatory network data (from [13]), that there is a strange phenomenon; the performance
of the inference worsens as we collect more data: the number of data points increases.
An example of a gene expression time series for a few genes is shown in Figure 1b. It is clear that as the number
of time points increases, the system is moving into an equilibrium with very little change in gene expression values.
This induces a distribution on any Xi (t) which looks more and more like a deterministic distribution.
In such a case, an information measure such as CMI which depends on the “input” distribution pxi (t−1) will
converge to zero and thus its performance will deteriorate as the number of samples increases. However a measure
that depends on the conditional distribution pxj t|xi (t−1),xi need not deteriorate with increasing number of samples.
Thus qCMI is more appropriate in this context (see Sec. 4.3 for performance of qCMI on this problem).
Related work: In the case that there is a pair of random variables X, Y , recent work [14] explored conditional
dependence measures which depend only on pY |X . Again in the two-variable case, a measure that had weak
dependence on pX was studied in [15]. The proposal there was to use the strong data processing constant and
hypercontractivity [16, 17] to infer causal strength; this has strong relationships to information bottleneck [18].
In this paper, we extend [19] to conditional independence (rather than independence studied there). In a related
but different direction, Shannon capacity, which is a potential dependence metric, was proposed in [20] to infer
causality from observational data [21].
Main Contributions: In this paper, we make the following main contributions:
1. We propose potential conditional mutual information as a way of quantifying conditional independence, but
depending only on the conditional distribution pY |X,Z .
2. We propose new estimators in the real-valued case that combine ideas from importance sampling, a coupling
trick and k-nearest neighbors estimation to estimate potential CMI.
3. We prove that the proposed estimator is consistent for a fixed k, which does not depend on the number of
samples N .
4. We demonstrate by simulation studies that the proposed estimator has excellent performance when there are
a finite number of samples, as well as an application in gene network inference, where we show that qCMI
can solve the non-monotonicity problem.
3
2
Estimator
In most real settings, we do not have access to either the joint distribution pX,Y,Z or the conditional distribution
pY |X,Z , but need to estimate the requisite information functionals from observed samples. We are given N independent identically distributied samples {(xi , yi , zi )i=1,2,..,N } from pX,Y,Z . In the case of qCMI, the estimator is
also given as input the modified distribution qX,Z . The estimator needs to estimate qCMI from samples.
In the case of discrete valued distributions, it is possible to empirically estimate pX,Y,Z from samples and
calculate the qCMI from this distribution. We focus our attention here on the case of continuous valued alphabet,
where each variable takes on values in a bounded subset of Rd . We assume that X, Y, Z are of dimensions dx , dy , dz
respectively, and let fX,Y,Z denote the joint density of the three variables (we assume that it exists). In such a
case, it is possible to estimate fX,Y,Z using kernel density estimators [22, 23] and then warp the estimate using the
potential measure qX,Z . However, it is known that k-nearest neighbors based estimators perform better even in
the simpler case of mutual information estimation and are widely used in practice [24, 25]. Therefore in this work,
we develop KNN based estimators for qCMI estimation.
2.1
Entropy estimation
Consider first the estimation of the differential entropy of a random variable X with density fX and observed
samples x1 , ..., xN .. A simple method
PN to estimate the differential entropy is to use the re-substitution estimator,
where we calculate ĥ(X) := N1 i=1 log(fˆX (xi )), where fˆX is an estimate of the density of X. We can estimate
the density using a KNN based estimator. To do so, we fix k a-priori, and for each sample xi , find the distance ρk,i
to the nearest neighbor.
k
(5)
fˆX (xi )cd ρdk,i ≈ .
N
This estimator is not consistent when k is fixed, and it was shown in a remarkable result by Kozhachenko and
Loenenko [26] that the bias is independent of the distribution and can be computed apriori. Thus the following
estimator was shown in [26] to be consistent for differential entropy.
ĥKL (X)
=
N
N ρdk,i cd
1 X
log
+ log k − ψ(k).
N i=1
k
While it is possible to have estimators which fix an apriori and then find the number of nearest neighbors to
plug into the formula, such estimators do not adapt to the density (some regions will have many more points inside
an neighborhood than others) and do not have a consistency proof as well. We mention this as fixed estimators
are used for a sub-problem in our estimator.
2.2
Coupling trick
The conditional mutual information can be written as a sum of 4 differential entropies, and one can estimate these
differential entropies independently using KNN estimators and sum them.
I(X; Y |Z) = −h(X, Y, Z) − h(Z) + h(X, Z) + h(Y, Z).
However, even in the case of mutual information, the estimation can be improved by an inspired coupling trick, in
what is called the KSG estimator [24]. We note that the original KSG estimator did not have a proof of consistency
and its consistency and convergence rates were analyzed in a recent paper [27]. Also of interest is the fact that the
coupling trick has been shown to be quite useful in problems where X, Y or both have a mixture of discrete and
continuous distributions or components [28].
This trick was applied in the context of conditional mutual information estimation in [29]. However, we note
that this estimator of CMI does not have a proof of consistency to the best of our knowledge. The CMI estimator
essentially fixes a k for the (X, Y, Z) vector and calculates for each sample (xi , yi , zi ), the distance ρk,i to the k-th
nearest neighbor. The estimator fixes this ρk,i as the distance and calculates the number of nearest neighbors
4
within ρk,i in the Z, (X, Z) and (Y, Z) dimensions as nz,i , nxz,i , nyz,i respectively. The CMI estimator is then given
by,
N
cdx +dz cdy +dz
1 X
ˆ
(ψ(k) − log(nxz,i ) − log(nyz,i ) + log(nz,i )) + log
.
CM I :=
N i=1
cdx +dy +dz cdz
2.3
(6)
qCMI estimator
Here, we adapt this estimator to calculate the qCMI for a given potential distribution qX,Z . The major difference
is the utilization of an importance sampling estimator to get the importance of each sample i estimated as follows,
ωi :=
qXZ (xi , zi )
.
fˆXZ (xi , zi )
(7)
However, importance sampling based reweighting alone is insufficient to handle qCMI estimation, since there
is a logarithm term which depends on the density also. We handle this effect by appropriately re-weighting the
number of nearest neighbors for the (y, z) and z terms carefully using the importance sampling estimators. The
estimation algorithm is described in detail in Algorithm 1.
Algorithm 1: qCMI algorithm
Data: Data Samples (xi , yi , zi ) for i = 1, . . . , N and qX,Z
ˆ I an estimate of qCM I
Result: qCM
Step 1: Calculate weights ωi
for i = 1, . . . , N do
Estimate fˆXZ (xi , zi ) using a Kernel density estimator [23, 22].
i ,zi )
ωi := fqˆXZ (x
, the importance sampling estimate of sample i.
(x ,z )
XZ
i
i
end
Step 2: Calculate information samples Ii
for i = 1, . . . , N do
ρk,i := Distance
of k-th nearest neighbor of (xi , yi , zi ).
P
nxz,i := j6=i:k(xi ,zi )−(xj ,zj )k<ρk,i 1, the number of neighbors of (xi , zi ) within distance ρk,i .
PN
nyz,i := j6=i:k(yi ,zi )−(yj ,zj )k<ρk,i ωj , the weighted number of neighbors of (yi , zi ) within distance ρk,i .
P
nz,i := j6=i:kzi −zj k<ρk,i ωi , the weighted number of neighbors of zi within distance ρk,i .
Ii := ψ(k) − log(nxz,i ) − log(nyz,i ) + log(nz,i ).
end
ˆ I = 1 PN ωi Ii + log cdx +dz cdy +dz .
Return qCM
i=1
N
cd +d +d cd
x
3
y
z
z
Properties
Our main technical result is the consistency of the proposed potential conditional mutual information estimator.
This proof requires combining several elements from importance sampling, and accounting for the correlation
induced by the coupling trick, in addition to handling the fact that the k is fixed and hence introduces a bias into
estimation.
Assumption 1. We make the following assumptions.
R
2
a) fXY Z (x, y, z) (log fXY Z (x, y, z)) dxdydz < ∞.
5
b) All the probability
density functions (PDF) are absolutely integrable, i.e. for all A, B ⊂ {X, Y, Z},
R q
∞ and |fA|B
(a|b)|da < ∞.
R
|fA|B (a|b)|da <
q
c) There exists a finite constant C such that the hessian matrices of fXY Z and fXY
Z exist and it’s true that
q
max{kh(fXY Z )k2 , kh(fXY Z )k2 } < C almost everywhere.
d) All the PDFs are upper-bounded, i.e. there exists a positive constant C 0 such that for all A, B ⊂ {X, Y, Z},
q
< C 0 almost everywhere.
fA|B < C 0 and fA|B
e) fXZ is upper and lower-bounded, i.e. there exist positive constants C1 and C2 such that C1 qXZ (x, z) <
fXZ (x, z) < C2 qXZ (x, z) almost everywhere.
f ) There bandwidth hN of kernel density estimator is chosen as hN = 21 N −1/(2dx +2dz +3) .
z
g) The k for the KNN estimator is chosen satisfying k > max{ dxd+d
,
y
dx +dy dx +dz
dz , dy }
Theorem 1. Under the Assumption 1, the qCMI estimator expressed in Algorithm 1 converges to the true value
qCMI.
p
ˆ I→
qCM
qCM I
(8)
Proof. Please see Section A for the proof.
4
Simulation study
In this section, we describe some simulated experiments we did to test the qCMI algorithm. The reader should
notice that all the tests we have done are taking qXZ as uXZ , i.e. all the tests are done for the special case of
uCMI. So by exploiting the qCMI notations we mean uCMI everywhere.
4.1
qCMI consistency
The first numerical experiment we do is to test the consistency of our qCMI estimation algorithm. We set up a
system of three variables X, Y and Z. The variables X and Z are independent taken from un (0, 1) distribution,
i.e. X and Z are taken form a uniform distribution and then raised to a power of n. When n = 1, the variables
X and Z are already uniform. When n is large, the un (0, 1) distribution skews more towards 0. For simplicity
we apply identical n for both X and Z here. Then Y is generated as Y = (X + Z + W ) mod 1 in which the
noise term W is sampled
from u(0, 0.2). From elementary information theory calculation, we can deduce that
1
I(X; Y |Z) = log .2
= 1.609 if n = 1. Thus I q (X; Y |Z) = 1.609 for all n. We plot the estimated value against the
ground truth.
As the first part of the experiment, we keep the number of samples constant at 1000 and 20000, and change the
degree n from 1 to 10. We compare the results of our KSG-based method with the simple partitioning method,
and the theoretical
√ value of qCMI. For the partitioning method, the number of partitions at each dimension is
determined by 3 100N , so that we observe on average 100 samples inside each quantization bin.
The results are shown in Figure 2a and Figure 2b. Our expectation is that qCMI remains constant as n (degree
of distribution) changes. We see that with relatively high number of samples, the accuracy of proposed qCMI is
satisfactorily high.
As the second part of the experiment, we do the same experiment as the first part, but this time we keep n = 5
and change the number of samples. The result is shown in Figure 2c. We can see convergence of KSG-based qCMI
estimator to the true value and how it outperforms the partitioning-based qCMI method.
As the third part of the experiment, we repeated the process for the first part, but replaced the un (0, 1)
distributions with β(1.5, 1.5) and the noise distribution with N (0, σ 2 ) and repeated the experiment for σ = 0.3, 1.0.
For this part we kept the number of partitions at 25 for each dimension. The results of calculated qCMI values are
shown in Figure 3a and Figure 3b.
6
(a)
(b)
(c)
Figure 2: The qCMI values calculated for un (0, 1) distributions for X and Z, and uniform noise: (a) N=1000
samples and (b) N =20000 samples. (c) Degree n = 5.
4.2
Dealing with discrete components
As we discussed before, the qCMI algorithm replaces the observed distribution fXZ distribution with a distribution
qXZ . This property comes in handy when we want to remove the bias caused by repeated samples. For example, as
discussed earlier, suppose that we want to measure the mutual information of two coupled variables in a dynamical
system evolving through time. Such systems usually start from an initial state, go through a transient state
and eventually reach a steady state. If one takes samples of the system’s state at a constant rate to study the
interaction of two variables, they might end up taking too many samples from the initial and steady states while
the transient phase which usually happens in a relatively short time might be more informative. The conditional
mutual information is not able to deal with this undesirable bias caused by the initial and steady states, while
qCMI inherently deals with the effect by compensating for the samples which are less likely to happen.
To better observe the effect, we repeat the first experiment of the previous section, but this time we generate
1000 samples from the scenario, and then add zeros to the X, Y and Z to create a high probability of occurrence at
(0, 0, 0). The proof of consistency of the estimator holds only when there is a joint density, i.e., the joint measure
is absolutely continuous with respect to the Lebesgue measure, and hence does not directly apply to this case.
We refer the reader to [28] for an analysis of a similar coupled KNN estimator for mutual information in the
discrete-continuous mixture case.
Changing the number of zero points added from 0 to 20000, we apply the conditional MI and qCMI to the data
generated and compare the results. As we can see in figure 3c, with the number of zeros increasing, the value of
conditional MI falls down to zero, unable to capture the inter-dependence of X and Y given Z, while qCMI value
remains unchanged, properly discovering the inter-dependence from the transient values.
4.3
Non-linear Neuron Cells’ Development Process
In this section, we apply the RDI and uRDI algorithms to neuron cells’ development process simulated based on
a model from [13] which can be modeled as a dynamical system. A dynamical System is described as a set of
variables shown by a vector of x which evolve through time starting from an initial state x(0). The evolution can
be described as a vector function g(.) such that x(t) = g (x(t − 1)). Note that g can be a stochastic function in
general, i.e. it may include random coefficients, additive noise and so on.
The dynamical system here describes the evolution of 13 genes through the development process. The nonlinear equations governing the development process approximate a continuous development process, in which ẋ(t) =
g(x(t − 1)). In other words, x(t) = x(t − 1) + dt.g (x(t − 1)) + n(t) in which n are independent Gaussian noises
∼ N (0, σ 2 ).
For this system, we want to infer the true network of causal inferences. In a dynamical system, we say xi
causes xj if xj (t) is a function of xi (t − 1). For this purpose, we first apply the RDI algorithm [12] to extract
7
(a)
(b)
(c)
Figure 3: The qCMI values calculated for a system with beta distribution for X and Z and Gaussian additive
noise: (a) σ = 0.3, and (b) σ = 1.0. (c) The qCMI and CMI values versus the number of zeros added.
the pairwise directed causality between the variables by calculating I (xi (t − 1), xj (t)|xj (t − 1)). Then we apply
the uRDI algorithm, in which the conditional mutual information I(X; Y |Z) in RDI is replaced with qCMI as
I q (X; Y |Z) using qX,Z as a uniform distribution.
This system is a good example of a system in which the genes undergo a rather short transient state compared
to the initial and steady states, and hence we expect an improvement in the performance of causal inference by
applying uRDI (see Figure 1b for an example run of the system). The details of the dynamical system are given in
[13].
We simulated the system for discretization dt = 0.1 and σ = .001, and changed the number of steps until which
the system continues developing, and then applied the RDI and uRDI algorithms to evaluate the performance of
each of the algorithms in terms of the area-under-the-ROC-Curve (AUC). The results are shown in Figure 4a. As
we can see, with the number of steps increasing implying the number of samples captured in the steady state are
increased, the uRDI algorithm outperforms RDI. In another test scenario, we fixed the number of steps at 200, but
concatenated several runs of the same process. The results and the improvement of performance by uRDI can be
seen in the Figure 4b.
4.4
Decaying Linear Dynamical System
In this section, we simulate a linear decaying dynamical system. A dynamical system in the simple case of a
deterministic linear system can be described as:
x(t) = Ax(t − 1)
(9)
In which A is a square matrix.
Here we simulate a system of 13 variables, all of them initialized from a u(0.5, 2) distribution. The first 6
variables (x1 , . . . , x6 ) are evolved through a linear deterministic process as in (9) in which A is a square 6 × 6
matrix initialized as:
u(0.75, 1.25)
0
0
0
u(0.75, 1.25)
0
u(0.75, 1.25) u(0.75, 1.25)
0
0
0
u(0.75, 1.25)
0
u(.75, 1.25) u(.75, 1.25)
0
0
0
A=
(10)
0
u(.75, 1.25)
0
u(.75, 1.25)
0
0
0
0
u(.75, 1.25) u(.75, 1.25) u(.75, 1.25)
0
u(.75, 1.25)
u(.75, 1.25)
0
0
0
u(.75, 1.25)
Then the matrix A is divided by 5 ∗ λmax (A) in which λmax (A) is the greatest eigenvalue of A. It’s done to make
sure that all the variables decay exponentially to 0. After initialization, the matrix A is kept constant throughout
the development process, i.e. it doesn’t change with time t.
8
(a)
(b)
(c)
Figure 4: AUC values for the neuron cells’ development process: a) versus the number of steps. b) versus the
number of runs. (c) for the decaying linear system
The other 7 variables (x7 , . . . , x13 ) are random independent Gaussian variables.
In this experiment, we simulate the system described above for various numbers of time-steps, keeping the
standard deviation of the Gaussian variables at σ = 0.1, and applied both RDI and uRDI algorithms to infer the
true causal inferences. Then we calculate the AUC values, the results are shown in Figure 4c. As we can see, the
uRDI algorithm outperforms RDI by a margin of 0.1 in terms of AUC.
5
Future Directions
In this section, we will describe some promising directions for further investigation.
1. Quantifying causal strength: As pointed out earlier, potential conditional mutual information can be used as
a metric for quantifying causal strength when the graph is a simple three node network (shown in Figure 1a).
However, further work is needed in order to generalize the definition to deduce the causal strength of an edge
or a set of edges in an arbitrary graph, akin to the formulation in [2] and to study the relative advantages
and disadvantages of such a formulation.
2. Discrete qCMI estimators: It has been shown in recent work that such estimators are not optimal even for
determining mutual information in the discrete alphabet case [30, 31, 32]. A very interesting question is how
such minimax-rate optimal estimators can be developed in the potential measures problem.
3. maxCMI estimation: While we have developed efficient estimators for qCMI, in maxCMI, there is a further
maximization over potential distributions q, which leads to some interesting interactions between estimation
and optimization. Recent work has studied estimation of Shannon capacity on continuous alphabets, however,
the formulation is not convex leading to possible local minima [14]. Further work is needed in order to find
provably optimal estimators for maxCMI in the continuous case.
4. Other conditional measures: Recent work [15] has used strong data processing constants as a way for quantifying dependence between two variables, with relationships to information bottleneck. These measures depend
partially on the factual measure pX , and are implicitly regularized. One direction of future work is to develop
multi-variable versions of such estimators to estimate the strength of conditional independence, for example.
5. Multivariable measures: Develop estimators that can handle more general multi-variable information measures
including total correlation [33] and multi-variate mutual information [34].
6. Ensemble estimation: Another approach exploiting k-nearest-neighbors for mutual information is the socalled ensemble estimation approach, where estimators for different k are combined together to get a stronger
9
estimator, with fast convergence [35]. An interesting direction of research is to obtain ensemble estimators
for potential measures.
6
Acknowledgement
The authors would like to thank Xiaojie Qiu and Cole Trapnell for discussions that triggered this work. This work
was supported in part by NSF Career award (grant 1651236) and NIH award number R01HG008164.
A
Proof
Here we’ll try to introduce a proof for qCMI algorithm we devised. As we know, the conditional mutual information
is defined as,
Z
fY |XZ (y|x, z)
I(X; Y |Z) = fXY Z (x, y, z) log
dxdydz.
(11)
fY |Z (y|z)
The qCMI is defined as the mutual information of X and Y given Z when the joint distribution of X and Z is
replaced by a joint uniform distribution qXZ (x, z),
!
Z
fY |XZ (y|x, z)
q
dxdydz.
(12)
I (X; Y |Z) = fY |XZ (y|x, z)qXZ (x, z) log
fYq |Z (y|z)
In which:
fYq |Z (y|z)
=
fYq Z (y, z)
=
fZq (z)
=
fYq Z (y, z)
,
f q (z)
Z Z
q
fXY
Z (x, y, z)dx,
Z
q
fXY
Z (x, y, z)dxdy.
(13)
(14)
(15)
from now on, the superscript U over each distribution function implies that the actual fXZ (x, z) is replaced by
qXZ (x, z). Equivalently, qCMI can be written as,
I q (X; Y |Z) = −hq (X, Y, Z) + hq (X, Z) + hq (X, Y ) − hq (Z),
(16)
where,
Z
q
h (X, Y, Z) ≡
hq (X, Z) ≡
hq (Y, Z) ≡
hq (Z) ≡
−
Z
−
Z
−
Z
−
q
fXY
Z (x, y, z) log fXY Z (x, y, z)dxdydz.
(17)
q
fXY
Z (x, y, z) log fXZ (x, z)dxdydz.
(18)
q
q
fXY
Z (x, y, z) log fY Z (y, z)dxdydz.
(19)
q
q
fXY
Z (x, y, z) log fZ (z)dxdydz.
(20)
R q
Note that −hq (X, Y, Z) + hq (X, Z) = −hq (Y |X, Z) = fXY
Z (x, y, z) log fY |XZ (y|x, z) dxdydz. So the term
inside the logarithm is independent of the distribution over (X, Z) and hence log fXY Z (x, y, z) and log fXZ (x, z)
appear when defining hq (X, Y, Z) and hq (X, Z).
Here we introduce a qCMI estimator, based on the KSG estimator for UMI. Remember the KSG-type estimator
for the conditional MI,
N
1 X
ˆ
(ψ(k) − log(nxz,i ) − log(nyz,i ) + log(nz,i )) + C(dx , dy , dz ).
I(X;
Y |Z) =
N i=1
10
(21)
where
C(dx , dy , dz ) := log
cdx +dz cdy +dz
cdx +dy +dz cdz
.
(22)
The estimator for the qCMI is as below,
N
1 X
Iˆq (X; Y |Z) =
ωi gi (xi , yi , zi ),
N i=1
(23)
where,
ωi
≡
qXZ (xi , zi )
.
fˆXZ (xi , zi )
(24)
gi (xi , yi , zi ) ≡ ψ(k) − log(nxz ) − log(
X
k(yi −yj ,zi −zj )k<ρk,i
B
X
ωj ) + log(
ωj ) + C(dx , dy , dz ).
(25)
kzi −zj k<ρk,i
KSG estimator for qCMI: Proof of convergence
Similar to the [19] we define,
ωi0
≡
n0yz,i
≡
qXZ (xi , zi )
.
fXZ (xi , zi )
X
ωj0 .
(26)
(27)
(i)
j∈Nyz
n0z,i
X
≡
ωj0 .
(28)
(i)
j∈Nz
g 0 (xi , yi , zi ) ≡ ψ(k) − log(nxz ) − log(n0yz,i ) + log(n0z,i ) + C(dx , dy , dz ).
(29)
From the triangle inequality, we can write,
|Iˆq (X; Y |Z) − I q (X; Y |Z)|
≤
+
|Iˆq (X; Y |Z) −
|
1
N
N
X
ωi0 g 0 (xi , yi , zi )|
(30)
(31)
i=1
N
1 X 0 0
ω g (xi , yi , zi ) − I q (X; Y |Z)|.
N i=1 i
(32)
To show the convergence of the KSG estimator for qCMI, we will show that (31) and (32) both converge to
zero. The Lemma 1 proves that (31) converges to zero. The lemma will be proven through the Section C.
Lemma 1. The term (31) converges to 0 as N −→ ∞ in probability.
For (32) we will first write the left hand side term as four entropy terms and show the convergence of each of
the terms to their respecttive true entropy term. This will be done in the Lemmas 2 and 3. So we can write,
N
N
X
ωi0
1 X 0 0
ωi g (xi , yi , zi ) = ĥq (Y, Z) + ĥq (X, Z) − ĥq (X, Y, Z) − ĥq (Z) −
(log(N − 1) + ψ(N )) ,
N i=1
N
i=1
11
(33)
where,
N
1 X 0
ω −ψ(k) + ψ(N ) + log cdx +dy +dz + (dx + dy + dz ) log ρk,i .
N i=1 i
(34)
≡
N
1 X 0
ω (− log(nxz,i ) + log(N − 1) + log cdx +dz + (dx + dz ) log ρk,i ) .
N i=1 i
(35)
ĥqN (Y, Z) ≡
N
1 X 0
ωi − log(n0yz,i ) + log(N − 1) + log cdy +dz + (dy + dz ) log ρk,i .
N i=1
(36)
N
1 X 0
ωi − log(n0z,i ) + log(N − 1) + log cdz + dz log ρk,i .
N i=1
(37)
ĥqN (X, Y, Z) ≡
ĥqN (X, Z)
ĥqN (Z) ≡
PN ω 0
The term i=1 Ni (log(N − 1) + ψ(N )) converges to 0 as N −→ ∞. We will prove the convergence of the rest
of terms in the following lemmas, proving the Theorem 1.
p
Lemma 2. Under the Assumption 1, ĥqN (X, Y, Z) → hq (X, Y, Z) as N −→ ∞.
Proof. It directly follows from the Lemma 2 from [19]. We just need to let X̃ ≡ (X, Z) and then show that
ĥqN (X̃, Y ) −→ hq (X̃, Y ) directly using the Lemma 2 from [19].
Lemma 3. For N −→ ∞,
p
ĥqN (X, Z) + ĥqN (Y, Z) − ĥqN (Z) → hq (X, Z) + hq (Y, Z) − hq (Z).
C
(38)
Proof of lemma 1
Th proof is analog to that of Lemma 1 in [19]. The term (31) is upper-bounded as,
|Iˆq (X; Y |Z) −
= |
N
1 X 0 0
ω g (xi , yi , zi )|
N i=1 i
(39)
N
1 X
(ωi g(xi , yi , zi ) − ωi0 g 0 (xi , yi , zi )) |
N i=1
(40)
≤
N
1 X
|ωi g(xi , yi , zi ) − ωi0 g 0 (xi , yi , zi )|
N i=1
(41)
≤
N
1 X
(|ωi − ωi0 ||g 0 (xi , yi , zi )| − ωi |g(xi , yi , zi ) − g 0 (xi , yi , zi )|)
N i=1
(42)
=
N
1 X
|ωi − ωi0 ||g 0 (xi , yi , zi )| + ωi | log(nyz,i ) − log(n0yz,i )| + ωi | log(nz,i ) − log(n0z,i )|
N i=1
(43)
!
N
1 X
1
1
1
1
0
0
0
0
≤
|ωi − ωi ||g (xi , yi , zi )| + ωi |nyz,i − nyz,i |(
+ 0 ) + ωi |nz,i − nz,i |(
+ 0 ) (44)
N i=1
2nyz,i
2nyz,i
2nz,i
2nz,i
N
N
X
1 X
ωi max1≤j≤N |ωi0 − ωi | max1≤j≤N |ωi0 − ωi |
≤
|ωi − ωi0 ||g 0 (xi , yi , zi )| +
+
(45)
N i=1
N
min1≤j≤N ωi0
min1≤j≤N ωi
i=1
1
1
0
0
≤ max |ωi − ωi | max |g (xi , yi , zi )| +
+
.
(46)
1≤i≤N
1≤i≤N
min1≤j≤N ωi0
min1≤j≤N ωi
12
The term g 0 (xi , yi , zi ) can be easily lower-bounded and upper-bounded as,
− 2 log N ≤ g 0 (xi , yi , zi ) ≤ 2 log N.
(47)
Thus, for any > 0 and sufficiently large N such that log N > max{C2 /3, 3C2 } and, if |ωi − ωi0 | < /(3 log N )
for all i , we have,
1
1
0
0
+
(48)
max |ωi − ωi | max |g (xi , yi , zi )| +
1≤i≤N
1≤i≤N
min1≤j≤N ωi0
min1≤j≤N ωi
!
1
2 log N + C2 +
≤
(49)
3 log N
1/C2 − 3 log
N
(2 log N + C2 + 2C2 ) ≤ .
(50)
≤
3 log N
So for any > 0 and sufficiently large N :
!
N
1 X
0 0
0
P
|ωi g(xi , yi , zi ) − ωi g (xi , yi , zi )| > ≤ P max |ωi − ωi | >
.
(51)
1≤i≤N
N i=1
3 log N
converges to zero as N −→
Following the proof of Lemma 1 in [19], the term P max1≤i≤N |ωi − ωi0 | > 3 log
N
∞. So the desired convergence for the term (31) is obtained.
D
Proof of lemma 3
If we define,
fˆXZ (xi , zi ) ≡
fˆYq Z (yi , zi ) ≡
fˆZq (zi ) ≡
âi
ai
nxz,i
,
x +dz
(N − 1)cdx +dz ρdk,i
n0yz,i
d +dz
y
(N − 1)cdy +dz ρk,i
n0z,i
z
(N − 1)cdz ρdk,i
(52)
,
(53)
,
(54)
≡ log fˆXZ (xi , zi ) + log fˆYq Z (yi , zi ) − log fˆZq (zi ),
≡ log fXZ (xi , zi ) +
log fYq Z (yi , zi )
−
log fZq (zi ).
(55)
(56)
Then,
ĥqN (X, Z) + ĥqN (Y, Z) − ĥqN (Z)
=
−
N
X
ω0
i
i=1
=
−
N
N
X
ω0
i
i=1
N
log fˆXZ (xi , zi ) + log fˆYq Z (yi , zi ) − log fˆZq (zi )
âi .
(57)
(58)
Now we can write,
|ĥqN (X, Z) + ĥqN (Y, Z) − ĥqN (Z) − (hq (X, Z) + hq (Y, Z) − hq (Z)) |
≤
|hq (X, Z) + hq (Y, Z) − hq (Z) −
N
X
i=1
+
N
X
ω0
i
i=1
N
|ai − âi |.
−
ωi0
N
ai |
(59)
(60)
(61)
13
For the term (60), since the terms ai = ωi0 (log fXZ (xi , zi ) + log fYq Z (yi , zi ) + log fZq (zi )) are i.i.d random variables, given the Assumption 1, by the strong law of large numbers, we can write,
N
X
i=1
−
ωi0
ai
N
N
X
=
i=1
−
ωi0
(log fXZ (x, z) + log fYq Z (y, z) − log fZq (z))
N
(62)
=
qXZ (x, z)
E −
(log fXZ (x, z) + log fYq Z (y, z) − log fZq (z))
fXZ (x, z)
Z
− fY |XZ (y|x, z)qXZ (x, z) (log fXZ (x, z) + log fYq Z (y, z) − log fZq (z)) dxdydz
=
hq (X, Z) + hq (Y, Z) − hq (Z).
−→
(63)
(64)
(65)
Therefore, the term (60) converges to 0 almost surely. For the term (61), let Ti = (Xi , Yi , Zi ), thus t = (x, y, z),
and fT (t) = fXY Z (x, y, z). For any fixed > 0, we can write,
!
N
X
ωi0
P
|ai − âi | >
(66)
N
i=1
!
!
N
N
[
X
≤ P
{|ai − âi | > /2} + P
ωi0 > 2N .
(67)
i=1
i=1
The second term converges to zero. The first term can be bounded as,
!
N
[
P
{|ai − âi | > /2}
(68)
i=1
Z
≤ N.P (|ai − âi | > /2) ≤ N
P (|ai − âi | > /2|Ti = t) fT (t)dt.
(69)
The term P (|ai − âi | > /2|Ti = t) can be upper-bounded by I1 (t) + I2 (t) + I3 (t) + I4 (t) + I5 (t), where,
I1 (t)
=
P (ρk,i > r1 |Ti = t) .
(70)
I2 (t)
=
(71)
I3 (t)
=
I4 (t)
=
I5 (t)
=
P (ρk,i < r2 |Ti = t) .
Z r1
P (| log fXZ (xi , zi ) − log fˆXZ (xi , zi )| > /6|ρk,i = r, Ti = t)fρ (r)dr.
r=r2
Z r1
P (| log fU (Y Z) (yi , zi ) − log fˆU (Y Z) (yi , zi )| > /6|ρk,i = r, Ti = t)fρ (r)dr.
r=r2
Z r1
P (| log fU (Z) (zi ) − log fˆU (Z) (zi )| > /6|ρk,i = r, Ti = t)fρ (r)dr.
(72)
(73)
(74)
r=r2
In which,
−1
r1
r2
≡ log N (N fT (t)cdx +dy +dz ) dx +dy +dz
2
≡ max{(log N ) (N fXZ (x, z)cdx +dz )
(75)
−1
dx +dz
2
q
, (log N ) (N f (y, z)cdy +dz )
−1
dy +dz
2
, (log N )
−1
(N fZq (z)cdz ) dz }.(76)
The terms I1 (t) and I2 (t) represent the probability that the value of ρk.i is too large or two small. We will
show that both probabilities converge to 0, i.e. ρi,k obtained lies within a reasonable range. The r1 threshold is
dx +dy +dz
k
determined based on the fact that N
≈ fT (ti )cdx +dy +dz ρk,i
and selecting k = log N is a reasonable choice.
s
The r2 threshold, on the other hand, implies that ρi,k should lie in a range such that kNs ≈ fS (si )cds ρdk,i
for all
subspaces s ∈ {(x, z), (y, z), z}.
The terms I3 (t), I4 (t) and I5 (t) represent the estimation error probability for a reasonable ρk,i .
Considering each of the terms separately, We will prove that each term will go to zero as N goes to infinity.
14
D.0.1
Convergence of I1
The term I1 (t) can be upper-bounded in the same way as explained in [19] for I1 (z). Then we will have,
I1 (t) ≤ kN k−1 exp{−
D.0.2
(log N )dx +dy +dz
}.
4
(77)
Convergence of I2
For the convergence of I2 (t), we are goona take the same steps as [19] for I2 (z). First let BT (t, r) ≡ n : kn − tk < r
−1
be the (dx + dy + dz )-dimensional ball centered at z with radius r. For r2,1 ≡ (log N )2 (N fXZ (x, z)cdx +dz ) dx +dz
and for sufficiently large N , the probability mass within the BT (t, r2,1 ) is given by,
−1
p2,1 ≡ P u ∈ BT (t, (log N )2 (N fXZ (x, z)cdx +dz ) dx +dz
dx +dy +dz
−1
−1
1 + C(log N )4 (N fXZ (x, z)cdx +dz ) dx +dz )2
≤ fT (t)cdx +dy +dz (log N )2 (N fXZ (x, z)cdx +dz ) dx +dz
≤
2fT (t)cdx +dy +dz
dx +dy +dz
dx +dz
(log N )2(dx +dy +dz ) N −
dx +dy +dz
dx +dz
(fXZ (x, z)cdx +dz )
dx +dy +dz
cd +d +d
≤ 2fY |XZ (y|x, z) x y z (log N )2(dx +dy +dz ) N − dx +dz
cdx +dz
dx +dy +dz
c
d
+d
+d
x
y
z
≤ 2C 0
(log N )2(dx +dy +dz ) N − dx +dz .
cdx +dz
(78)
(79)
(80)
Where the last inequalty comes from the assumption that fY |XZ (y|x, z) < C 0 . Similarly, for the second
−1
threshold r2,2 = (log N )2 (N f q (y, z)cdy +dz ) dy +dz ,
p2,2
−1
≡ P u ∈ BT (t, (log N )2 (N fYq Z (y, z)cdy +dz ) dy +dz
dx +dy +dz
−1
−1
≤ fT (t)cdx +dy +dz (log N )2 (N fYq Z (y, z)cdy +dz ) dy +dz
1 + C(log N )4 (N fYq Z (y, z)cdy +dz ) dy +dz )2
≤
2fT (t)cdx +dy +dz
dx +dy +dz
dy +dz
(log N )2(dx +dy +dz ) N
−
dx +dy +dz
dy +dz
(fYq Z (y, z)cdy +dz )
d +dy +dz
fT (t) cdx +dy +dz
− xdy +d
z
≤ 2 q
(log N )2(dx +dy +dz ) N
fY Z (y, z) cdy +dz
d +dy +dz
f q (t) cdx +dy +dz
− xdy +d
z
≤ 2C2 q T
(log N )2(dx +dy +dz ) N
fY Z (y, z) cdx +dz
d +dy +dz
cd +d +d
− xdy +d
z
≤ 2C2 C 0 x y z (log N )2(dx +dy +dz ) N
.
cdy +dz
(81)
(82)
(83)
(84)
The last two inequalities come from the bounds assumed on the distribution functions. Similarly, for the second
15
−1
threshold r2,2 = (log N )2 (N fZq (z)cdz ) dz we can write:
−1
p2,3 ≡ P u ∈ BT (t, (log N )2 (N fZq (z)cdz ) dz
dx +dy +dz
−1
−1
1 + C(log N )4 (N fZq (z)cdz ) dz )2
≤ fT (t)cdx +dy +dz (log N )2 (N fZq (z)cdz ) dz
≤
≤
≤
≤
2fT (t)cdx +dy +dz
dx +dy +dz
dz
(log N )2(dx +dy +dz ) N −
(85)
(86)
dx +dy +dz
dz
(87)
(fZq (z)cdz )
dx +dy +dz
fT (t) cdx +dy +dz
2 q
(log N )2(dx +dy +dz ) N − dz
fZ (z)
cdz
dx +dy +dz
fTq (t) cdx +dy +dz
2C2 q
(log N )2(dx +dy +dz ) N − dz
fZ (z)
cdz
dx +dy +dz
c
d
+d
+d
x
y
z
2C2 C 0
(log N )2(dx +dy +dz ) N − dz
cdz
(88)
(89)
(90)
The last two inequalities come from the bounds assumed on the distribution functions.
Similar to the procedure for I2 (z) in the [19], I2 (t) is the probability that at least k samples lie in BT (t, max{r2,1 , r2,2 , r2,3 }).
Then we have,
I2 (t)
=
=
P (ρk,i < max{r2,1 , r2,2 , r2,3 })
N
−1
X
N-1
N −1−m
max{p2,1 , p2,2 , p2,3 }m (1 − max{p2,1 , p2,2 , p2,3 })
m
(91)
(92)
m=k
≤
N
−1
X
N m max{p2,1 , p2,2 , p2,3 }m
(93)
m=k
≤
N
−1
X
m=k
d +d
dy
z
cdx +dy +dz
− min{ xdz y , dxd+d
,
}
y dx +dz
2CC
(log N )2(dx +dy +dz ) N
min{cdy +dz , cdx +dz , cdz }
0
m
.
(94)
(95)
Similarly, for N sufficently large and applying the sum of geometric series, we have,
I2 (t) ≤ 4CC 0
D.0.3
cdx +dy +dz
min{cdy +dz , cdx +dz , cdz }
k
(log N )2k(dx +dy +dz ) N
−k min{
dx +dy
dz
d
y
z
, dxd+d
,
}
y dx +dz
.
(96)
Convergence of I3
Given Ti = t = (x, y, z), and ρk,i = r and fˆXZ (xi , zi ) =
nxz,i
x +dz
(N −1)cdx +dz ρd
k,i
, we have,
P | log fXZ (Xi , Zi ) − log fˆXZ (Xi , Zi )| > /6|ρk,i = r, Ti = t
= P nxz,i > (N − 1)cdx +dz rdx +dz fXZ (x, z)e/6 |ρk,i = r, Ti = t
+ P nxz,i < (N − 1)cdx +dz rdx +dz fXZ (x, z)e/6 |ρk,i = r, Ti = t .
(97)
(98)
(99)
Given Ti = t, Lemma 4 gives the probability distribution of the nxz,i .
Lemma 4. Given Ti = t = (x, y, z), and ρk,i = r < rN for some deterministic sequance of rN such that
PN −1
limN ←−∞ rN = 0 and for any > 0, the number of neighbors nxz,i −k is distributed as l=k+1 Ul , where Ul are i.i.d
Bernoulli random variables with mean fXZ (x, z)cdx +dz rdx +dz (1 − /8) ≤ E[Ul ] ≤ fXZ (x, z)cdx +dz rdx +dz (1 − /8)
for sufficiently large N .
16
Proof. See the proof of Lemma 5 in [19].
Based on Lemma 4,
P nxz,i > (N − 1)cdx +dz rdx +dz fXZ (x, z)e/6 |ρk,i = r, Ti = t
!
N
−1
X
dx +dz
/6
= P
fXZ (x, z)e − k
Ul > (N − 1)cdx +dz r
(100)
(101)
l=k+1
N
−1
X
= P
!
Ul − (N − 1 − k)E[Ul ] > (N − 1)cdx +dz r
dx +dz
/6
fXZ (x, z)e
− k − (N − 1 − k)E[Ul ] . (102)
l=k+1
The right hand side term inside the probability can be lower bounded as,
(N − 1)cdx +dz rdx +dz fXZ (x, z)e/6 − k − (N − 1 − k)E[Ul ]
≥ (N − 1)cdx +dz r
dx +dz
/6
(103)
dx +dz
− k − (N − 1 − k)fXZ (x, z)cdx +dz r
(1 − /8)
(104)
/6
fXZ (x, z) e − 1 − /8 − k
≥ (N − k − 1)cdx +dz r
(105)
(106)
≥ (N − k − 1)cdx +dz rdx +dz fXZ (x, z) .
48
for sufficiently large N .
2
Applying Bernstein’s inequality, (102) can be upper bounded by exp{− 2304(1+19/144)
(N −k−1)cdx +dz rdx +dz fXZ (x, z)}.
The tail distribution can also be upper-bounded by the same term. Thus,
fXZ (x, z)e
dx +dz
P (| log fXZ (xi , zi ) − log fˆXZ (xi , zi )| > /6|ρk,i = r, Ti = t)
≤ 2 exp{−
2
(N − k − 1)cdx +dz rdx +dz fXZ (x, z)}.
2304(1 + 19/144)
Therefore, the term I3 (t) can be upper-bounded as,
Z r1
I3 (t) =
P (| log fXZ (xi , zi ) − log fˆXZ (xi , zi )| > /6|ρk,i = r, Ti = t)fρ (r)dr
(107)
(108)
r=r2
−1
Z
log N (N fT (t)cdx +dy +dz ) dx +dy +dz
≤
−1
P (| log fXZ (xi , zi ) − log fˆXZ (xi , zi )| > /6|ρk,i = r, Ti = t)fρ (r)dr
r=(log N )2 (N fXZ (x,z)cdx +dz ) dx +dz
−1
Z
≤
log N (N fT (t)cdx +dy +dz ) dx +dy +dz
2
(N − k − 1)cdx +dz rdx +dz fXZ (x, z)}fρ (r)dr
2304(1 + 19/144)
dx +dz
−1
N cdx +dz fXZ (x, z) (log N )2 (N fXZ (x, z)cdx +dz ) dx +dz
}
(109)
−1
r=(log N )2 (N fXZ (x,z)cdx +dz ) dx +dz
2 exp{−
2
4608
2
2(d +d )
≤ 2 exp{−
(log N ) x z }.
4608
For sufficiently large N.
≤
D.0.4
2 exp{−
(110)
Convergence of I4
Given Ti = t = (x, y, z), and ρk,i = r and fˆU (yi , zi ) =
=
+
n0yz,i
d +dz
y
(N −1)cdy +dz ρk,i
, we have,
P | log f q (Yi , Zi ) − log fˆq (Yi , Zi )| > /6|ρk,i = r, Ti = t
P nyz,i > (N − 1)cdy +dz rdy +dz f q (Yi , Zi )e/6 |ρk,i = r, Ti = t
P nyz,i < (N − 1)cdy +dz rdy +dz f q (Yi , Zi )e/6 |ρk,i = r, Ti = t .
17
(111)
(112)
(113)
(1)
(2)
We can write n0yz,i = nyz,i + nyz,i , where,
(1)
nyz,i
X
=
j:kTj −tk<ρi,k
(2)
nyz,i
X
=
j:kTj −tk>ρi,k
qXZ (xj , zj )
fXZ (xj , zj )
(114)
qXZ (xj , zj )
I{k(Yj − Yi , Zj − Zi )k < ρk,i }.
fXZ (xj , zj )
(115)
Given Ti = t, Lemma 5 gives the probability distribution of the n0yz,i .
Lemma 5. Given Ti = t = (x, y, z), and ρk,i = r < rN for some deterministic sequance of rN such that
PN −1
(2)
limN ←−∞ rN = 0 and for any > 0, the distribution of nyz,i is l=k+1 Vl , where Vl are i.i.d random variables with
Vl ∈ [0, 1/C1 ] and mean fXZ (x, z)cdx +dz rdx +dz (1 − /8) ≤ E[Ul ] ≤ fXZ (x, z)cdx +dz rdx +dz (1 − /8) for sufficiently
large N .
Proof. See the proof of Lemma 6 in [19].
According to the Lemma 5 and following the same procedure as I3 (t), the term I4 (t) will also be bounded here
in th same way. We have:
C 1 2
2(d +d )
I4 (t) ≤ 2 exp{−
(log N ) y z }
(116)
4608
D.0.5
Convergence of I5
Similar to the case of I4 (t), given Ti = t = (x, y, z), and ρk,i = r and fˆU (zi ) =
, we have,
P | log f q (Zi ) − log fˆq (Zi )| > /6|ρk,i = r, Ti = t
P nz,i > (N − 1)cdz rdz f q (Zi )e/6 |ρk,i = r, Ti = t
P nz,i < (N − 1)cdz rdz f q (Zi )e/6 |ρk,i = r, Ti = t .
=
+
(1)
n0z,i
z
(N −1)cdz ρd
k,i
(117)
(118)
(119)
(2)
We can write n0z,i = nz,i + nz,i , where,
(1)
nz,i
=
X
j:kTj −tk<ρi,k
(2)
nz,i
=
X
j:kTj −tk>ρi,k
qXZ (xj , zj )
fXZ (xj , zj )
(120)
qXZ (xj , zj )
I{kZj − Zi k < ρk,i }.
fXZ (xj , zj )
(121)
Following the same procedure as for I4 (t), we will obtain the upper bound below for I5 (t):
I5 (t) ≤ 2 exp{−
C1 2
2(d )
(log N ) z }
4608
18
(122)
Now, we can write,
N
X
ω0
i
P
i=1
N
!
|ai − âi | >
(123)
Z
≤ N
(I1 (t) + I2 (t) + I3 (t) + I4 (t) + I5 (t)) fT (t)dt
(124)
(log N )dx +dy +dz
}
≤ kN k exp{−
4
k
d +d
dy
z
cdx +dy +dz
1−k min{ xdz y , dxd+d
,
}
y dx +dz
+
4CC 0
(log N )2k(dx +dy +dz ) N
min{cdy +dz , cdx +dz , cdz }
+
2N exp{−
2
C1 2
2(d +d )
2(d )
(log N ) x z + 4N exp{−
(log N ) z }.
4608
4608
(125)
d
d +d
y
z
, dx +d
} < 0, then all the terms will converge to 0 as
If k is chosen large enough so that 1 − k min{ xdz y , dxd+d
y
z
N goes to infinity, and the proof of Lemma 3 is complete.
References
[1] T. M. Cover and J. A. Thomas, Elements of information theory. John Wiley & Sons, 2012.
[2] D. Janzing, D. Balduzzi, M. Grosse-Wentrup, and B. Schölkopf, “Quantifying causal influences,” The Annals
of Statistics, pp. 2324–2358, 2013.
[3] P. Spirtes, C. N. Glymour, and R. Scheines, Causation, prediction, and search. MIT press, 2000.
[4] P. W. Holland, C. Glymour, and C. Granger, “Statistics and causal inference,” ETS Research Report Series,
vol. 1985, no. 2, 1985.
[5] A. P. Dawid, “Conditional independence in statistical theory,” Journal of the Royal Statistical Society. Series
B (Methodological), pp. 1–31, 1979.
[6] C. W. Granger, “Investigating causal relations by econometric models and cross-spectral methods,” Econometrica: Journal of the Econometric Society, pp. 424–438, 1969.
[7] P. Geiger, K. Zhang, B. Schoelkopf, M. Gong, and D. Janzing, “Causal inference by identification of vector autoregressive processes with hidden components,” in International Conference on Machine Learning,
pp. 1917–1925, 2015.
[8] H. Hosseini, S. Kannan, B. Zhang, and R. Poovendran, “Learning temporal dependence from time-series data
with latent variables,” in Data Science and Advanced Analytics (DSAA), 2016 IEEE International Conference
on, pp. 253–262, IEEE, 2016.
[9] M. Eichler, “Graphical modelling of multivariate time series,” Probability Theory and Related Fields, vol. 153,
no. 1-2, pp. 233–268, 2012.
[10] C. J. Quinn, N. Kiyavash, and T. P. Coleman, “Directed information graphs,” IEEE Transactions on information theory, vol. 61, no. 12, pp. 6887–6909, 2015.
[11] J. Sun, D. Taylor, and E. M. Bollt, “Causal network inference by optimal causation entropy,” SIAM Journal
on Applied Dynamical Systems, vol. 14, no. 1, pp. 73–106, 2015.
[12] A. Rahimzamani and S. Kannan, “Network inference using directed information: The deterministic limit,” in
Communication, Control, and Computing (Allerton), 2016 54th Annual Allerton Conference on, pp. 156–163,
IEEE, 2016.
19
[13] X. Qiu, S. Ding, and T. Shi, “From understanding the development landscape of the canonical fate-switch pair
to constructing a dynamic landscape for two-step neural differentiation,” PloS one, vol. 7, no. 12, p. e49271,
2012.
[14] W. Gao, S. Kannan, S. Oh, and P. Viswanath, “Causal strength via shannon capacity: Axioms, estimators
and applications,” in Proceedings of the 33rd International Conference on Machine Learning, 2016.
[15] H. Kim, W. Gao, S. Kanan, S. Oh, and P. Viswanath, “Discovering potential correlations via hypercontractivity,” arXiv preprint arXiv:1709.04024, 2017.
[16] V. Anantharam, A. Gohari, S. Kamath, and C. Nair, “On maximal correlation, hypercontractivity, and the
data processing inequality studied by erkip and cover,” arXiv preprint arXiv:1304.6133, 2013.
[17] Y. Polyanskiy and Y. Wu, “Strong data-processing inequalities for channels and bayesian networks,” in Convexity and Concentration, pp. 211–249, Springer, 2017.
[18] N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck method,” arXiv preprint
physics/0004057, 2000.
[19] W. Gao, S. Kannan, S. Oh, and P. Viswanath, “Conditional dependence via shannon capacity: Axioms,
estimators and applications,” arXiv preprint arXiv:1602.03476, 2016.
[20] R. Kidambi and S. Kannan, “On shannon capacity and causal estimation,” in Communication, Control, and
Computing (Allerton), 2015 53rd Annual Allerton Conference on, pp. 988–992, IEEE, 2015.
[21] J. M. Mooij, J. Peters, D. Janzing, J. Zscheischler, and B. Schölkopf, “Distinguishing cause from effect using
observational data: methods and benchmarks,” The Journal of Machine Learning Research, vol. 17, no. 1,
pp. 1103–1204, 2016.
[22] L. Devroye and C. S. Penrod, “The consistency of automatic kernel density estimates,” The Annals of Statistics,
pp. 1231–1249, 1984.
[23] S. J. Sheather and M. C. Jones, “A reliable data-based bandwidth selection method for kernel density estimation,” Journal of the Royal Statistical Society. Series B (Methodological), pp. 683–690, 1991.
[24] A. Kraskov, H. Stögbauer, and P. Grassberger, “Estimating mutual information,” Physical review E, vol. 69,
no. 6, p. 066138, 2004.
[25] S. Khan, S. Bandyopadhyay, A. R. Ganguly, S. Saigal, D. J. Erickson III, V. Protopopescu, and G. Ostrouchov,
“Relative performance of mutual information estimation methods for quantifying the dependence among short
and noisy data,” Physical Review E, vol. 76, no. 2, p. 026209, 2007.
[26] L. Kozachenko and N. N. Leonenko, “Sample estimate of the entropy of a random vector,” Problemy Peredachi
Informatsii, vol. 23, no. 2, pp. 9–16, 1987.
[27] W. Gao, S. Oh, and P. Viswanath, “Demystifying fixed k-nearest neighbor information estimators,” in Information Theory (ISIT), 2017 IEEE International Symposium on, pp. 1267–1271, IEEE, 2017.
[28] W. Gao, S. Kannan, S. Oh, and P. Viswanath, “Estimating mutual information for discrete-continuous mixtures,” arXiv preprint arXiv:1709.06212, 2017.
[29] S. Frenzel and B. Pompe, “Partial mutual information for coupling analysis of multivariate time series,”
Physical review letters, vol. 99, no. 20, p. 204101, 2007.
[30] G. Valiant and P. Valiant, “Estimating the unseen: an n/log (n)-sample estimator for entropy and support
size, shown optimal via new clts,” in Proceedings of the forty-third annual ACM symposium on Theory of
computing, pp. 685–694, ACM, 2011.
20
[31] J. Jiao, K. Venkat, Y. Han, and T. Weissman, “Minimax estimation of functionals of discrete distributions,”
IEEE Transactions on Information Theory, vol. 61, no. 5, pp. 2835–2885, 2015.
[32] Y. Wu and P. Yang, “Minimax rates of entropy estimation on large alphabets via best polynomial approximation,” IEEE Transactions on Information Theory, vol. 62, no. 6, pp. 3702–3720, 2016.
[33] S. Watanabe, “Information theoretical analysis of multivariate correlation,” IBM Journal of research and
development, vol. 4, no. 1, pp. 66–82, 1960.
[34] C. Chan, A. Al-Bashabsheh, J. B. Ebrahimi, T. Kaced, and T. Liu, “Multivariate mutual information inspired
by secret-key agreement,” Proceedings of the IEEE, vol. 103, no. 10, pp. 1883–1913, 2015.
[35] K. R. Moon, K. Sricharan, and A. O. Hero III, “Ensemble estimation of mutual information,” arXiv preprint
arXiv:1701.08083, 2017.
21
| 7cs.IT
|
TRACE RECONSTRUCTION WITH VARYING DELETION PROBABILITIES
arXiv:1708.02216v1 [math.PR] 7 Aug 2017
LISA HARTUNG, NINA HOLDEN, AND YUVAL PERES
A BSTRACT. In the trace reconstruction problem an unknown string x = (x0 , . . . , xn−1 ) ∈
{0, 1, ..., m − 1}n is observed through the deletion channel, which deletes each xk with a
e Earlier works have proved that if each
certain probability, yielding a contracted string X.
xk is deleted with the same probability q ∈ [0, 1), then exp(O(n1/3 )) independent copies of
e suffice to reconstruct x with high probability. We extend this upper
the contracted string X
bound to the setting where the deletion probabilities vary, assuming certain regularity
conditions. First we consider the case where xk is deleted with some known probability
qk . Then we consider the case where each letter ζ ∈ {0, 1, ..., m − 1} is associated with
some possibly unknown deletion probability qζ .
1. I NTRODUCTION
N
Let n ∈ , m ∈ {2, 3, . . . }, and [m] = {0, . . . , m − 1}. In trace reconstruction the goal
is to reconstruct an unknown string x = (x0 , . . . , xn−1 ) ∈ [m]n from noisy observations
of x. Here we study the case where the data is noisy due to a deletion channel in which
each bit is deleted independently with a certain probability. In other words, instead of
e obtained by sending x through the
observing x we observe many independent strings X
deletion channel. The probability of deleting xk might depend on either (I) the location
k of xk in the string, or (II) the letter xk .
e is obtained in the following
In Case (I) and given pk ∈ (0, 1] for k ∈ [n], the string X
way for k = 0, 1, . . . , n − 1, starting from an empty string.
e and increase k by one.
• (retention) With probability pk , copy xk to the end of X
• (deletion) With probability qk = 1 − pk , increase k by one.
e is obtained by performing the
In Case (II) and given pζ ∈ (0, 1] for ζ ∈ [m], the string X
following steps for k = 0, 1, . . . , n − 1.
e and increase k by one.
• (retention) With probability px , copy xk to the end of X
k
• (deletion) With probability qxk = 1 − pxk , increase k by one.
In Case (I) we assume the deletion probabilities are known, while we do not make this
assumption in Case (II).
e (1) , . . . , X
e (T ) (called “traces”) from the
For T ∈ we consider T independent outputs X
deletion channel. Our main question is the following. Given ε > 0, how many samples
b ∈ [m]n we have [X
b = x] > 1 − ε? Holenstein,
are needed, such that for some X
Mitzenmacher, Panigrahy, and Wieder [5] proved that exp n1/2 polylog(n) traces are
sufficient in the case where all xk are deleted independently with the same probability
q ∈ [0, 1). See [6] for an alternative proof. This result was recently improved by De,
O’Donnell, and Servedio [3], and by Nazarov and Peres [8]. Using single
bit statistics for
1/3
the traces, they proved that reconstruction is possible with exp O(n ) traces, and that
this is optimal for reconstruction techniques using only single bit statistics. The case of
random strings was studied in [1] and [9].
N
P
Date: August 8, 2017.
1
2
LISA HARTUNG, NINA HOLDEN, AND YUVAL PERES
1.1. Main result. We study the following two settings in Case (I) where the deletion
probabilities depend on the location in the string.
(i) (weak monotonicity) For some δ ∈ (0, 1) the retention probabilities (pk )k∈N satisfy
p` > p2k + δ for all k > `, and pk > δ for all k > 0.
(ii) (periodicity) The deletion probabilities are 2-periodic, meaning that
qk = q for k even,
qk = qe for k odd.
(1.1)
In particular, (i) covers the case where the retention probabilities are monotonically decreasing and bounded away from zero, and the case where all retention probabilities are
in some interval [p + δ, 2p] for p ∈ (0, 1/2] and δ > 0. Also observe that, by reversing the
p
sequence, we can also study strings where the deletion probabilities satisfy p` > 2j + δ
for all ` > j (instead of ` < j).
Theorem 1.1. Let ε > 0 and m ∈ {2, 3, . . . }, and let the deletion probabilities be known
and satisfy Assumption (i) or (ii). There exists a constant C > 0 depending only on ε, m,
and δ (for Assumption (i)), and ε, m, q, and qe (for Assumption (ii)), such that the original
string x ∈ [m]n , can be reconstructed with probability at least 1 − ε from T = dexp Cn1/3 e
i.i.d. samples of the deletion channel applied to x.
For case (II), where the deletion probabilities vary by letter, we prove the following.
Theorem 1.2. For any δ, ε > 0 and m ∈ {2, 3, . . . } there is a constant C > 0 such that if
minζ∈[m] pζ ≥ δ, then the string x ∈ [m]n can be reconstructed with probability at least 1 − ε
from T = dexp Cn1/3 e i.i.d. samples of the deletion channel applied to x. This result holds
also in the case when the deletion probabilities are unknown.
The deletion channel (or variants of it which also allow insertions, substitutions, swaps,
etc.) is relevant for the study of mutations and DNA sequencing errors. In the case of
mutations several studies have revealed that mutation probabilities vary by location in
the genome [4, 7, 10, 11].
Outline of the paper: To prove Theorem 1.1 we will first derive in Section 2.1 an
e as the coefficients of a particular
exact formula expressing the single bit statistics of X
polynomial. In Sections 2.2 and 2.3 we prove that this polynomial cannot be too small
everywhere on a particular boundary arc of the unit disk. In Section 3 we use this to prove
that for any two input strings, at least one bit in the trace has a sufficiently different
expectation so we are able distinguish the two strings. Theorem 1.2 will be proved by
first using the traces to obtain good estimates for the deletion probabilities, and then use
a single bit test (with the estimated probabilities) for the traces sent through a second
deletion channel.
Acknowledgements: We thank Ryan O’Donnell for suggesting the problem of varying
deletion probabilities. Most of this work was done while the first and second author were
visiting Microsoft Research, and they want to thank Microsoft for the hospitality.
2. P REPARATORY L EMMAS
2.1. A polynomial identity. For a ∈
Ψ(w) =
Rn define
n−1
X
k=0
ak pk
k−1
Y
`=0
(p` w + q` ).
(2.1)
TRACE RECONSTRUCTION WITH VARYING DELETION PROBABILITIES
3
e be the output of the deletion channel
Lemma 2.1. Let a = (a0 , a1 , . . . , an−1 ) ∈ Rn and let a
with input a, padded with an infinite sequence of 0’s on the right side. Then
X
e
aj wj = Ψ(w).
(2.2)
E
j≥0
Proof. The expectation on the left side of (2.2) can be written as
!
n−1
k−1
k−1
X
X
X
j
ak p k
w P
B` = j ,
k=0
j=0
(2.3)
`=0
where B` for ` = 0, . . . , n − 1 are independent Bernoulli(p` )-distributed random variables.
Observe that
!
k−1
k−1
Pk−1 k−1
X
X
Y
j
w P
B` = j = E w `=0 B` =
(q` + p` w),
(2.4)
j=0
`=0
`=0
where we used independence in the last step. Plugging (2.4) into (2.3) concludes the
proof of Lemma 2.1.
Next, we will establish that (perhaps after a change of variables) the function |Ψ(w)|
on a small arc of the unit disc is not to small. We use tools from standard complex analysis
(inspired by [2]).
2.2. The deletion probabilities satisfy a monotonicity property.
Lemma 2.2. Assume that a0 = 1, ak ∈ [−1, 1] for k ≥ 1, and that there is some δ ∈ (0, 1/10)
π
such that p0 > p2k + δ and pk > δ for all k > 0. Define the arc γL = {z = eiθ , −π
L ≤ θ ≤ L }.
Then there is a constant c > 0 depending only on δ such that maxγL |Ψ(w)| > e−cL for all
L ≥ 1.
Proof. Throughout the proof c is a positive constant which may depend on δ and which
may change line by line. Define w0 := −q0 /p0 . Let Ω0 be a rectangular domain which
contains w0 and 1/2, has sides parallel to the coordinate axes, and which is such that
|pw + (1 − p)| < 1 − δ/2 for all p ∈ {p0 , p1 , . . . } and w ∈ Ω0 . Observe that an appropriate
domain exists since the required inequality is satisfied for all w ∈ [w0 , 1/2]. Furthermore,
observe that we may assume Ω0 changes continuously as we vary w0 . Let Ω = Ω0 ∪
B1/2 (1/2). For L > 100, let Ω0L be the horizontal translation of Ω by some tL ∈ (0, 1/10),
such that γL is the intersection of Ω0L and the right half of ∂ . There is an L0 > 0
depending only on δ such that w0 ∈ Ω0L for all L > L0 , and it is sufficient to prove the
result of the lemma for L > L0 . Let ΩL = Ω0L ∩ ( ∪ {z : Re(z) < 0}). Then ∂ΩL is the
union of the three disjoint sets γL , γL2 , and γL3 , where
D
D
γL2 := {z ∈ ∂ΩL : Re(z) ≤ 1/2 + tL },
γL3 := {z ∈ ∂ΩL \ γL : Re(z) > 1/2 + tL }.
0
See Figure 1 for an illustration of ∂ΩL . Since log |Ψ| is subharmonic, and letting µw
ΩL
denote harmonic measure of ΩL relative to w0 ,
Z
Z
Z
w0
w0
0
log |Ψ(w0 )| ≤
log |Ψ(z)|dµΩL (z) +
log |Ψ(z)|dµΩL (z) +
log |Ψ(z)|dµw
ΩL (z).
γL
2
γL
3
γL
(2.5)
4
LISA HARTUNG, NINA HOLDEN, AND YUVAL PERES
D
γL2
γL
ΩL
w0
γL3
F IGURE 1. Illustration of objects defined in the proof of Lemma 2.2. The
domain ΩL is bounded by the colored curves.
Without loss of generality we assume |Ψ(z)| < 1 everywhere on γL . For each fixed w0
0
0
there is a constant cw0 such that µw
ΩL (γL ) ≥ cw0 /L. Since we assume Ω varies continuously as we vary w0 ∈ [1 − 1/δ, 0], we can find a c0 > 0 such that cw0 > c0 for all w0 this
gives
Z
c0
w0
0
log |Ψ(z)|dµw
|Ψ(w)| ≤ log max |Ψ(w)|.
ΩL (z) ≤ µΩL (γL ) · log max
γL
γL
L
γL
Observe that log |Ψ(z)| < c for any z ∈ ∂ΩL ∩ {z : Re(z) < 1/2 + tL }, so
Z
0
log |Ψ(z)|dµw
ΩL (z) ≤ c.
For any w ∈
|Ψ(w)| ≤
2
γL
D,
n−1
X
k−1
Y
pk
n−1
X
k=0 `=0
k=0
(qk + pk |w|) ≤
|q` + p` w| ≤
`=0
k=0
n−1
X k−1
Y
(1 − δ + δ|w|)k ≤
1
.
δ(1 − |w|)
(2.6)
It follows that
Z
3
γL
0
log |Ψ(z)|dµw
ΩL (z)
Z
1
1
≤ log
+
log
dµw0 (z).
3
δ
1 − |z| ΩL
γL
(2.7)
A Brownian motion started at w0 must hit the line segment `L := {z ∈ ΩL : Re(z) =
1/4 + tL } before hitting γL3 . Therefore, defining DL = B1/2 (1/2 + tL ), the integral in (2.7)
is bounded above by
Z
1
c sup
log
dµw (z).
(2.8)
3
1 − |z| DL
w∈`L γL
3
Since `L has distance > 1/4 from γL3 , the density of µw
DL on γL is bounded above by a
universal constant. For any z ∈ γL3 in the first quadrant we have 1 − |z| ≥ c−1 (θ − θ0 )2 ,
where θ ∈ (−π, π] is the angle for polar coordinates with origin at 1/2+tL and θ0 ∈ [0, π/2]
corresponds to the point of intersection between ∂DL and ∂ in the first quadrant (see
[8, equation (4.4)]). Therefore we can bound (2.8) from above by
Z π/2
1
log
dθ < ∞.
(2.9)
c
θ
0
D
We conclude that
Z
3
γL
0
log |Ψ(z)|dµw
ΩL (z) < c.
(2.10)
TRACE RECONSTRUCTION WITH VARYING DELETION PROBABILITIES
5
Inserting the above estimates into (2.5), we get
log |p0 | ≤ c +
c0
log max |Ψ(w)|,
γL
L
which implies the lemma.
2.3. The 2-periodic case. Assume without loss of generality that p < pe. If z = pew + qe
then
n−1
k−1
X
Y z − qe
e
Ψ(w) = Ψ(z)
:=
ak p k
(p`
+ q` ).
(2.11)
pe
k=0
`=0
Lemma 2.3. Assume that a0 = 1, ak ∈ [−1, 1] for k ≥ 1, and that (p` )` is 2-periodic. If
π
e
|Ψ(z)|
< β everywhere on the the arc γL = {z = eiθ , −π
L ≤ θ ≤ L } then there exists c
depending on q and qe such that β > e−cL for all L ≥ 1.
Proof. We consider the following cases.
Case 1) p0 = pe.
Then (2.11) gives
z − qe
e
+ q + ....
Ψ(z)
= a0 pe + a1 pz + a2 pez p
pe
e
Setting z = 0 we get Ψ(0)
= pe > 0.
Case 2) p0 = p. Then (2.11) gives
z − qe
z − qe
e
+ q + a2 pz p
+ q + ....
Ψ(z) = a0 p + a1 pe p
pe
pe
(2.12)
(2.13)
Setting again z = 0 gives
pe
q
e
Ψ(0) = a0 p + a1 pe q −
= p + a1 (e
p − p).
pe
The expression in (2.14) can take three values
p
if a1 = 0,
e
Ψ(0) =
pe
if a1 = 1,
−e
p + 2p if a1 = −1.
(2.14)
(2.15)
Case 2.1) If pe 6= 2p, then we see by (2.15) that |Ψ(0)| is bounded from below by a constant
depending only on p and pe.
Case 2.2) p0 = p and pe = 2p. Observe that in this particular case
p
z − qe
1
1
+q = z+ .
pe
2
2
(2.16)
Hence inserting z = −1 into (2.13) gives
e
Ψ(−1)
= a0 p > 0.
(2.17)
e is continuous in some neighborhood of z = −1 (uniformly in a0 , a1 , . . . ).
Moreover, Ψ
e
Hence, there exists δ1 ∈ (0, 1) depending only on p such that |Ψ(1−δ
1 )| is bounded
away from 0.
6
LISA HARTUNG, NINA HOLDEN, AND YUVAL PERES
e
Again |Ψ(z)|
for |z| ≤ 1 is bounded in terms of a geometric series. Set p∗ = ppe and
∗
= 1 − p . Then
1
e
|Ψ(z)|
≤ ∗
.
(2.18)
p (1 − |z|)
Now the claim of Lemma 2.3 follows by letting Ω be a ball contained in such that ∂Ω
is tangent to ∂ at 1 and −1 + δ1 ∈ Ω, and then proceeding as in Lemma 2.2 using the
e
subharmonic function log |Ψ|.
q∗
D
D
3. P ROOF OF MAIN RESULT: D ELETION PROBABILITIES VARYING WITH POSITION
Proof of Theorem 1.1. Throughout the proof all constants should depend only on δ and m.
Let x, y ∈ [m]n be two different strings. We first consider the case m = 2. Let a = x − y.
By Lemma 2.1,
n−1
n−1
k−1
X
X
Y
ej − Yej wj =
E X
ak p k
(q` + p` w).
(3.1)
j=0
k=0
`=0
Let k ∗ = min{k : ak 6= 0}. Then the right side of (3.1) is equal to
Y
(q` + p` w) pk∗ ak∗ +
`<k∗
n−1
X
ak p k
k=k∗ +1
k−1
Y
!
(q` + p` w) .
(3.2)
`=k∗
Let L ∈ N. We first consider deletion probabilities satisfying Assumption (i). First, we
verify that for w ∈ γL
Y
(q` + p` w) ≥ exp(−nC2 /L2 ),
(3.3)
`<k∗
for some universal constant C2 . To see that observe that by writing w = cos(θ) + i sin(θ),
for any 0 < p < 1,
|pw + (1 − p)|2 = 1 − 2(1 − p)p (1 − cos(θ))
≥ 1 − (1 − p)pθ2 + O(θ4 )
= exp −(1 − p)pθ2 + O(θ4 ) ,
(3.4)
where we used the series expansion of cos. Since θ ∈ γL , (3.3) follows. Define
A(w) = pk∗ ak∗ +
n−1
X
ak pk
k=k∗ +1
k−1
Y
(q` + p` w).
(3.5)
`=k∗
By Lemma 2.2 there exists w1 ∈ γL such that |A(w1 )| ≥ e−cL . Hence, taking absolute
values in (3.1) gives for w ∈ γL ,
Y
X
ej − Yej
(3.6)
E X
≥
(q` + p` w1 ) |A(w1 )|
`<k∗
j≥0
≥ exp(−nC2 /L2 )e−cL .
(3.7)
To approximately maximize the term on the right side of (3.6) we choose L of order
and obtain that there is a constant C3 > 0 such that
X
ej − Yej ≥ exp −C3 n1/3 .
E X
(3.8)
n1/3
j≥0
We now conclude the proof similarly as the proof of [8, Theorem 1.1] (see the argument
starting with equation (2.4) of that paper). We first observe that for some j ≥ 0 and
TRACE RECONSTRUCTION WITH VARYING DELETION PROBABILITIES
7
ej − Yej ≥ exp(−C4 n1/3 ). By Hoeffding’s inequality this allows us
C4 > 0 we have E X
to test whether our bit sequence is more likely to equal x or y; in case our string equals
either x or y our test fails with probability at most exp(−C5 n1/3 ). Repeating this for all
possible pairs of bit sequences x, y ∈ {0, 1}n , we can determine the original bit string
with high probability as n → ∞. By increasing the constant C appearing in the statement
of the theorem the result holds for any n. See [8] for a more detailed argument.
For m 6= 2 we proceed similarly. Let x, y ∈ [m]n for m 6= 2. For each fixed ζ ∈ [m],
e and y
e to be equal to x and y, respectively, except that we replace xk by 1 if
we define x
xk = ζ, and we replace xk by 0 if xk 6= ζ. Using the above procedure we can find all k
such that xk = ζ in the original string. Repeating for all ζ ∈ [m] we determine the original
string.
For deletion probabilities satisfying Assumption (ii) let
A(z) = pk∗ ak∗ +
n−1
X
ak pk
k=k∗ +1
k−1
Y
`=k∗
p`
p` qe
.
z + q` −
pe
pe
(3.9)
By Lemma 2.3 there exists z0 ∈ γL such that |A(z0 )| ≥ e−cL . Similarly as above, for
q
z ∈ γL and w = z−e
pe ,
Y
(q` + p` w) ≥ exp(−nC6 /L2 ),
(3.10)
`<k∗
and
|w| ≥ exp −C7 /L2 .
(3.11)
Taking again absolute values in (3.1), this gives for z ∈ γL ,
X
j≥0
ej − Yej |w|j
E X
≥
Y
(q` + p` w) |A(z)|
`<k∗
≥ exp(−nC6 /L2 )e−cL ,
(3.12)
ej − Yej ≥ exp(−C8 n1/3 ) for some j ≥ 0 and C8 > 0.
Using (3.11), this gives that E X
We conclude the proof as above.
4. P ROOF OF MAIN RESULT: D ELETION PROBABILITIES VARYING WITH LETTER
To prove Theorem 1.2, we first observe that the theorem is immediate from [8] and [3]
in the case where the deletion probabilities q0 , . . . , qm−1 are known. This follows since we
can send the traces through a second deletion channel, where each letter ζ ∈ [m] is retained with probability p−1
ζ minζ 0 ∈[m] pζ 0 . The traces obtained in the second deletion channel can be obtained directly from x sent through a single deletion channel with constant
retention probability minζ∈[m] pζ , and x can therefore be reconstructed with exp(O(n1/3 ))
traces. For the case of unknown deletion probabilities, we show in Lemma 4.1 that we
can obtain good estimates for the deletion probabilities by studying the traces. Then we
use Lemma 4.2 to argue that these estimates are sufficiently good, so that the single bit
test still works when we use our estimated values for the deletion probabilities.
Lemma 4.1. Consider the setting of Theorem 1.2, where we assume the deletion probabilities
are unknown, and that each letter in [m] appears at least once in x. Given any δ, C1 > 0,
8
LISA HARTUNG, NINA HOLDEN, AND YUVAL PERES
we can find a C0 > 0 depending only on δ, m, and C1 , such that if we have at least T =
dexp(C0 n1/3 )e traces, then we can use the traces to find an estimate pbζ for each pζ satisfying
max |b
pζ − pζ | > exp(−C1 n1/3 ) < δ.
(4.1)
P
ζ∈[m]
Proof. Fix ζ ∈ [m] and define p := pζ , r := |{k ∈ [n] : xk = ζ}|, µ := rp, and v :=
e t |] : X
e t = ζ}|, then [Yt ] = µ and
rp(1 − p). Observe that 1 − p = v/µ. If Yt = |{k ∈ [|X
k
Var[Yt ] = v. We use the following estimates µ
b, vb, and pb for µ, v, and p, respectively,
E
µ
b=
E
T
1X
Yt ,
T
E
T
vb =
t=1
1 X
(Yt − µ
b)2 ,
T −1
1 − pb = vb/b
µ.
t=1
We have [b
µ] = µ, [b
v ] = v, and Var[b
µ] = v/T < r/T . Also observe that for a universal
constant C > 0, we have Var[b
v ] ≤ Cr4 /T . This means that for an appropriate C0 > 0,
with probability 1 − on (1), we have |b
µ − µ|, |b
v − v| < exp(−C1 n1/3 ). Upon increasing
C0 and using the definition of pb, this gives that |b
p − p| < exp(−C1 n1/3 ) with probability
1 − on (1), which implies the lemma.
For x = (x0 , . . . , xn−1 ) ∈
Rn and p = (p0, . . . , pm−1) ∈ (0, 1]m, define
Φxp (w) =
n−1
X
k=0
xk pxk
k−1
Y
(px` w + qx` ).
(4.2)
`=0
Lemma 4.2. For any C2 > 0, we can find a C1 > 0 such that the following holds for any
p ∈ (0, 1]. Let p = (p, . . . , p) ∈ (0, 1]m and p0 = (p00 , . . . , p0m−1 ) ∈ (0, 1]m satisfy
|p − p0ζ | < exp(−C1 n1/3 ),
∀ζ ∈ [m].
(4.3)
Then for each x ∈ [m]n , each coefficient in the polynomial Φxp (w) − Φxp0 (w) has magnitude
less than exp(−C2 n1/3 ).
Proof. To simplify notation, let ε = exp(−C1 n1/3 ). The magnitude of the coefficient of wj
in Φxp (w) − Φxp0 (w), is bounded above by the following, with B(k, p) denoting a binomial
random variable
n−1
n
o
X
max
[B(k, p + εb) = j] − [B(k, p) = j] : εb ∈ [−ε, ε] + ε,
m
P
P
k=0
which is bounded by a polynomial multiple of ε.
Proof of Theorem 1.2. Throughout the proof we consider constants C0 , C1 , C2 which may
depend on δ, p∗ := minζ∈[m] pζ , and m, but which are independent of all other parameters.
Let T = dexp(C0 n1/3 )e for some C0 > 0 which will be determined later. Consider T
e (1) , . . . , X
e (T ) obtained by sending x through the deletion channel considered in
traces X
the statement of the theorem. Using the T traces, we find an estimate pbζ for each pζ as
described in Lemma 4.1, and let pb∗ = minζ∈[m] pbζ .
e (t) through a second deletion channel, so that we obtain traces
Send each trace X
(1)
(T
)
X̌ , . . . , X̌ . In the second deletion channel the letter ζ is retained with probability
pb∗ /b
pζ . Observe that each trace X̌(t) can be obtained from x by considering a deletion
channel in which the letter ζ is retained with probability p∗ζ := pζ pb∗ /b
pζ . In particular, if
our estimate pbζ for pζ is good for all ζ, then each letter is retained with approximately the
b ∗ = (b
same probability p∗ . Define p∗ = (p∗0 , . . . , p∗m−1 ) ∈ (0, 1]m and p
p∗ , . . . , pb∗ ) ∈ (0, 1]m ;
TRACE RECONSTRUCTION WITH VARYING DELETION PROBABILITIES
9
the first string represents the actual (unknown) retention probabilities, and while the
second string represents our estimated retention probabilities.
Given strings y, y0 ∈ [m]n , let Y̌ (resp. Y̌0 ) denote a string obtained by sending y (resp.
0
y ) through the two deletion channels described above, i.e., the letter ζ is retained with
probability p∗ζ . We first assume m = 2. By Lemma 2.1,
n−1
X
0
E Y̌j − Y̌j0 wj = Φyp∗ (w) − Φyp∗ (w)
(4.4)
j=0
0
0
= Φypb ∗ (w) − Φpb ∗ (w) + Φyp∗ (w) − Φypb ∗ (w) − Φyp∗ (w) − Φypb ∗ (w) .
y0
Let
E(C0 , C1 ) :=
max |b
p∗ − p∗ζ | ≥ exp(−C1 n1/3 ) .
ζ∈[m]
By [8], there is a C2 > 0, such that for some j ∈ [n], the absolute value of the coefficient
0
of wj in Φypb ∗ (w)−Φypb ∗ (w) is at least exp(−C2 n1/3 /2). From the proof in [8] we see that C2
is universal given any lower bound on pb∗ , so on the event E(C0 , C1 )c and for sufficiently
large n, we may assume that the constant C2 depends only on p∗ .
Given C2 > 0, define C1 > 0 as in Lemma 4.2. By Lemma 4.1, we can find C0 > 0 such
that
[E(C0 , C1 )] < δ/2.
(4.5)
P
By Lemma 4.2 and (4.4), on E(C0 , C1 )c and for all sufficiently large n, the absolute value
0
of the coefficient of wj in the polynomial Φyp∗ (w) − Φyp∗ (w) is at least exp(−2C2 n1/3 )
0
and of the same sign as the corresponding coefficient in Φypb ∗ (w) − Φypb ∗ (w). Therefore,
on the event E(C0 , C1 )c , Hoeffding’s inequality gives that with probability at least 1 −
exp(−T exp(−4C2 n1/3 )/2), we can determine if our unknown bit string x equals y or
(t)
y0 , by considering (X̌j )1≤t≤T . After possibly increasing C0 , we see by a union bound
that on E(C0 , C1 )c and for m = 2, we can identify x with probability at least 1 −
2n exp(−T exp(−4C2 n1/3 )/2).
Given y, y0 ∈ [m]n for m 6= 2 we proceed similarly. For each fixed ζ ∈ [m] and with
b and y
b 0 to
x = (x0 , . . . , xn−1 ), we first identify the set Aζ := {k ∈ [n] : xk = ζ}. Define y
0
be equal to y and y , respectively, except that all xk such that xk = ζ has been replaced
by 1, and all xk such that xk 6= ζ has been replaced by 0. Using the approach above with
b and y
b 0 , on the event E(C0 , C1 )c we can determine Aζ except on an event of probability
y
1 − 2n exp(−T exp(−4C2 n1/3 )/2). We repeat the procedure for each ζ ∈ [m], and see that
on the event E(C0 , C1 )c we can determine the sets Aζ , and hence x, except on an event
of probability 1 − on (1). By increasing the constant C0 if necessary, we can reconstruct
the string with probability at least 1 − ε for any n.
R EFERENCES
[1] T. Batu, S. Kannan, S. Khanna, and A. McGregor. Reconstructing strings from random traces. In Proceedings of the fifteenth annual ACM-SIAM symposium on Discrete algorithms, pages 910–918. Society
for Industrial and Applied Mathematics, 2004.
[2] P. Borwein and T. Erdélyi. Littlewood-type problems on subarcs of the unit circle. Indiana University
mathematics journal, 46(4):1323, 1997.
[3] A. De, R. O’Donnell, and R. Servedio. Optimal mean-based algorithms for trace reconstruction. In
Proceedings of the 49th Annual ACM SIGACT Symposium on Theory of Computing, STOC 17, pages 1047–
1056, New York, NY, USA, 2017. ACM.
10
LISA HARTUNG, NINA HOLDEN, AND YUVAL PERES
[4] H. Ellegren, N. G. Smith, and M. T. Webster. Mutation rate variation in the mammalian genome. Current
Opinion in Genetics and Development, 13(6):562 – 568, 2003.
[5] T. Holenstein, M. Mitzenmacher, R. Panigrahy, and U. Wieder. Trace reconstruction with constant deletion probability and related results. In Proceedings of the Nineteenth Annual ACM-SIAM Symposium on
Discrete Algorithms, SODA ’08, pages 389–398, Philadelphia, PA, USA, 2008. Society for Industrial and
Applied Mathematics.
[6] A. McGregor, E. Price, and S. Vorotnikova. Trace reconstruction revisited. In A. S. Schulz and D. Wagner,
editors, Algorithms - ESA 2014: 22th Annual European Symposium, Wroclaw, Poland, September 8-10,
2014. Proceedings, pages 689–700, Berlin, Heidelberg, 2014. Springer Berlin Heidelberg.
[7] G. A. T. McVean, S. R. Myers, S. Hunt, P. Deloukas, D. R. Bentley, and P. Donnelly. The fine-scale
structure of recombination rate variation in the human genome. Science, 304:581–584, 2004.
[8] F. Nazarov and Y. Peres. Trace reconstruction with exp(O(n1/3 )) samples. In Proceedings of the 49th
Annual ACM SIGACT Symposium on Theory of Computing, STOC 17, pages 1042–1046, New York, NY,
USA, 2017. ACM.
[9] Y. Peres and A. Zhai. Average-case reconstruction for the deletion channel: subpolynomially many
traces suffice, 2017. To appear in FOCS.
[10] N. G. Smith, M. T. Webster, and H. Ellegren. Deterministic mutation rate variation in the human
genome. Genome Res., 12:1350–1356, 2002.
[11] F. Supek and B. Lehner. Differential DNA mismatch repair underlies mutation rate variation across the
human genome. Nature, 521(7550):81–84, 2015.
C OURANT I NSTITUTE OF M ATHEMATICAL S CIENCES , N EW Y ORK U NIVERSITY, N EW Y ORK , NY 10012
E-mail address: [email protected]
D EPARTMENT OF M ATHEMATICS , M ASSACHUSETTS I NSTITUTE
E-mail address: [email protected]
M ICROSOFT R ESEARCH , R EDMOND , WA 98052
E-mail address: [email protected]
OF
T ECHNOLOGY , C AMBRIDGE , MA 02139
| 10math.ST
|
Self Organizing Maps to efficiently cluster and functionally
interpret protein conformational ensembles
Domenico Fraccalvieri* 1
Laura Bonati 1
Fabio Stella 2
1 Department
2 Department
1
of Earth and Environmental Sciences, University of Milano Bicocca, Milano, IT
of Informatics, Systems and Communication, University of Milano Bicocca, Milano, IT
*Corresponding Author = [email protected]
Introduction
The wide range of protein biological functions, such as enzymatic activity, ligand- and protein-protein
interactions and allosteric regulation, is strictly related to their flexibility and dynamics[1]. To model
the influence of protein motions across this broad spectrum of events Molecular Dynamics (MD) simulations are now routinely used. The identification of the most functionally relevant conformations is
generally done by grouping the conformations according to a criterion of geometrical similarity and
popular choices include hierarchical clustering, single, complete and average linkage and k-means [10].
These geometrical approaches rely on the assumption that the identified conformational states also correspond to the energetic states [10]. Good candidates to improve this match are Artificial Neural Networks
(ANNs) which are capable to discover the relationships between the measured variables only from the
available dataset [9]. Among the class of ANNs, successfully applied to artificial life systems [2], Self
Organizing Maps (SOMs) [6] represent a particularly powerful data driven model that has been widely
applied for exploration and clustering of high-dimensional datasets [9, 6]. We recently developed an
approach which combines SOMs and hierarchical clustering to efficiently compare conformational ensembles obtained from multiple MD simulations of proteins [3]. To reliably apply the SOM analysis to
these specific input data we identified and optimized a small number of SOM parameters. In particular,
we confirmed that the map size is a crucial parameter and that a well-selected number of neurons is
crucial both to reduce the computational cost of the analysis and to provide an intermediate topological
representation of the input conformational space [3]. As a result the original MD trajectories can be
represented by the few conformations that best represent the clusters obtained. Here we make a step further; the proposed approach consists of processing all the atom positions of each conformation won by
a given SOM neuron using a similarity network. Different similarity measures between atoms behavior
are used to compile a similarity matrix which is inputted to a network model [8]. Network algorithms
are then used to automatically discover and interpret the behavior of the original protein conformational
ensembles exploiting the atomic coordinates enclosed in the SOM neuron.
2
Methods
Self Organizing Map clustering. The details of the proposed SOM approach can be summarized as
follow. First, an ensemble of conformations is extracted from one (or more) MD trajectory with an
optimized sampling rate of 1/100 ps. Each conformation of the trajectories is represented using only
the Cα Cartesian coordinates. The ensemble is encoded in a matrix and is used for the learning stage,
A. Graudenzi, G. Caravagna, G. Mauri, M. Antoniotti (Eds.):
Wivace 2013 - Italian Workshop on
Artificial Life and Evolutionary Computation
EPTCS 130, 2013, pp. 83–86, doi:10.4204/EPTCS.130.13
c D. Fraccalvieri, L. Bonati & F. Stella
This work is licensed under the
Creative Commons Attribution License.
84
SOMs to efficiently cluster and functionally interpret protein conformational ensembles
performed with the set of optimal SOM parameters previously identified by experimental design [3].
The most relevant parameters are: Map size=100, Radius=3, Training length=5000 and Neighbor function=gaussian. The neurons of this trained SOM, i.e. the proto-clusters of the original conformational
ensemble, are grouped using the complete linkage algorithm and the number of clusters is automatically
selected by means of the Mojena’s stopping rule [7]. To provide a graphical representation of the neurons and their connections, each neuron is described by a hexagon and the neurons belonging to each
cluster are homogeneously colored (see Figure 1, central panel). Finally, to provide the most synthetic
structural representation of the original ensemble, only the input conformation closest to the centroid of
each cluster is represented (see Figure 1, left panel). The centroid is defined as the mean vector of all the
prototype vectors of the neurons in that cluster.
Network Analysis. The set of all the Cα’s atomic coordinates of the conformations won by each SOM
neuron, are used to compile a network model. The following similarity measures were analyzed and
empirically compared; x-y-z average absolute correlation, cosine, Pearson, Spearman correlation and
biweight mid-correlation. Networks are constructed from similarity matrices by means of threshold as
described in [5]. The SOM, which summarizes the atomic positions of each conformation associated
with a neuron, allows building a network model which is then exploited to automatically discover and
interpret the behavior of the inputted protein conformational ensembles. Graph partitioning algorithms
[5] to cluster atoms with similar behavior and community detection algorithms [8, 5] were applied to the
network model (see Figure 1, right panel).
Figure 1: Central panel: the clustered SOM. The background colors of the neurons (hexagons) correspond to the cluster attribution. The inner hexagon is proportional to the number of conformations won
by the neuron. Left panel: ribbon representation of one centroid cluster (in orange) superimposed to a
reference structure (in black). Right panel: graphical analysis of one neuron. On the left, each node represents an atom of the neuron conformations and the edges link pairs of correlated atoms. The groups of
nodes (atoms) with similar behavior are homogeneously colored. On the right, the nodes and the edges
are represented on a reference 3D structure (in gray) using the same color code of the graph. The width
of the edge is proportional to the correlation among the atoms (nodes).
D. Fraccalvieri, L. Bonati & F. Stella
3
85
Results and Discussion
The SOM approach developed by our group has been successfully applied showing its potential in the
comparison of MD simulations of different proteins with dimensions ranging from 60 to 500 amino
acids[3, 4]. These applications highlighted key features, related to the topological nature of the SOM.
First, a single map was able to detect both low fluctuations, i.e. little conformational changes with
respect to a reference structure, and large concerted motions, i.e. simultaneous conformational changes
of different region of the protein.[3, 4]. Second, the intrinsic time dependence of the data, not given as
input, was correctly recovered; in fact consecutive conformational transitions were correctly assigned
to adjacent neurons, thus to adjacent clusters, on the map [3]. Moreover, the values of geometrical
descriptors based on a subset of Cα atoms included in the SOM training [3, 4], as well as excluded
features [4], were consistently clustered, indicating the potential of the map to classify unknown features
[4]. Finally the ability of a SOM trained with a set of protein to classify conformations of similar
protein was verified and exploited [4]. The extension of our protocol here presented is aimed at a deeper
inspection of the dynamic information contained in each neuron in the output map. In fact, not only
the subset of conformations won by each neuron, but also the specific motions caught by that neuron
are analyzed. To this end, the matrices composed by the Cα Cartesian coordinates of the original data
in each neuron are analyzed using similarity matrices and represented by graphs (see Figure 1, right
panel). This analysis has a double outcome. First it provides an interesting and informative picture of
the joint behavior of the atoms belonging to the conformations in each neuron, highlighting both local
motions, i.e. motions involving structurally neighbor atoms, and long-range communications mediated
by atoms able to transmit conformational signals to remote sites of the protein. Second this information
can be used to cluster the data according to this graph. Cluster analysis based on similarity of the Cα
dynamic behavior enclosed in each SOM neuron should increase the ability of our approach to compare
the dynamics of different proteins. In conclusion, we already showed that the use of this SOM approach
to cluster protein conformational ensembles increases the cluster quality compared to other methods
routinely used [3]. Moreover both the approaches proposed for the map interpretation allow an efficient
comparison of the dynamics of different proteins. In the former approach, the characteristics of the
conformations belonging to each cluster are summarized and described by the cluster centroid, in the
latter, the characteristic pathways of motion described by each neuron are highlighted by the graphs
(Figure 1). Applications of this approach should range from protein engineering (to design specific
mutation able to reduce, or increase, the function of a protein) to computer-based drug design (to select
specific conformations of a protein suitable to interact with specific ligands).
References
[1] D.D. Boehr, R. Nussinov & P.E. Wright (2009): The role of dynamic conformational ensembles in biomolecular recognition. Nat. Chem. Biol. 5(11), pp. 789–796, doi:10.1038/nchembio.232.
[2] A. Cangelosi, Nolfi S. & Parisi D. (2003): Artificial Life Models of Neural Development. In: On Growth,
form and Computers, Elsevier Academic Press, pp. 339–52. doi:10.1016/B978-012428765-5/50051-7
[3] D. Fraccalvieri, A. Pandini, F. Stella & L. Bonati (2011): Conformational and functional analysis of molecular dynamics trajectories by Self-Organising Maps. BMC Bioinformatics 12, p. 158, doi:10.1186/14712105-12-158.
[4] D. Fraccalvieri, M. Tiberti, A. Pandini, L. Bonati & E. Papaleo (2012): Functional annotation of the
mesophilic-like character of mutants in a cold-adapted enzyme by self-organising map analysis of their
molecular dynamics. Mol. Biosyst. 8, pp. 2680–91, doi:10.1039/C2MB25192B.
86
SOMs to efficiently cluster and functionally interpret protein conformational ensembles
[5] S. Horvath (2011): Weighted Network Analysis: Applications in Genomics and Systems Biology. Springer.
doi:10.1007/978-1-4419-8819-5
[6] T.K. Kohonen (2013): Essentials of the Self-Organizing Map.
doi:10.1016/j.neunet.2012.09.018.
Neural Networks 37, pp. 52–65,
[7] R. Mojena (1977): Hierarchical Grouping Methods and Stopping Rules: An Evaluation. Comput. J. 20(4),
pp. 359–63, doi:10.1093/comjnl/20.4.359.
[8] M.E.J. Newman (2010): Networks: An Introduction. Oxford University Press, New York.
[9] A. Pandini, D. Fraccalvieri & L. Bonati (2013): Artificial Neural Networks for Efficient Clustering of Conformational Ensembles and their Potential for Medicinal Chemistry. Curr. Top. Med. Chem. 13(5), pp. 642–51,
doi:10.2174/1568026611313050007.
[10] J. Shao, S.W. Tanner, N. Thompson & T.E. Cheatham (2007): Clustering molecular dynamics trajectories:
1. Characterizing the performance of different clustering algorithms. J. Chem. Theory Comput. 3(6), pp.
2312–34, doi:10.1021/ct700119m.
| 5cs.CE
|
Decoding-History-Based Adaptive Control of Attention
for Neural Machine Translation
arXiv:1802.01812v1 [cs.CL] 6 Feb 2018
Junyang Lin, Shuming Ma, Qi Su, Xu Sun
MOE Key Laboratory of Computational Linguistics, Peking University
School of Electronics Engineering and Computer Science, Peking University
School of Foreign Languages, Peking University
{linjunyang, shumingma, sukia, xusun}@pku.edu.cn
Abstract
Seq2Seq+Attention: In the Russian capital of Moscow , the
Attention-based sequence-to-sequence model has
proved successful in Neural Machine Translation
(NMT). However, the attention without consideration of decoding history, which includes the past
information in the decoder and the attention mechanism, often causes much repetition. To address this
problem, we propose the decoding-history-based
Adaptive Control of Attention (ACA) for the NMT
model. ACA learns to control the attention by
keeping track of the decoding history and the current information with a memory vector, so that the
model can take the translated contents and the current information into consideration. Experiments
on Chinese-English translation and the EnglishVietnamese translation have demonstrated that our
model significantly outperforms the strong baselines. The analysis shows that our model is capable of generating translation with less repetition
and higher accuracy. The code will be available at
https://github.com/lancopku
Russian capital of Moscow of the Russian capital of Moscow
was killed this year because of the cold war, most of them were
homeless and the elderly, including many people.
Gold: The temperatures in Moscow, capital of Russia, dropped
to such low levels last night that even locals felt freezing cold.
Six people died as a result, bringing up the death toll due to
coldness this year to 239. Most of the dead were the homeless
and the elderly, including many drunk.
1 Introduction
With the development of Deep Learning, Neural Machine
Translation (NMT) has demonstrated outstanding effects, and
the sequence-to-sequence model (Seq2Seq) [Sutskever et al.,
2014] is the most commonly-used model in NMT. The attention mechanism [Bahdanau et al., 2014; Luong et al., 2015]
is often used in the Seq2Seq model, and in many cases it can
significantly improve the performance of the model. In translating, the decoder builds a language model on the target language for semantic coherence, the attention mechanism obtains the source-side information for the word generation at
each time step.
However, the current source-side information that the attention mechanism acquires is often controversial to the translated contents because the attention has no knowledge of the
translated contents. We present a typical example of the
over-translation of the attention-based Seq2Seq model on the
Chinese-English translation in Table 1. From the example, it
can be found that the attention-based Seq2Seq generates the
same phrase “the Russian capital of Moscow” for multiple
times, causing much repetition.
Table 1: An example of the translation of the conventional attentionbased Seq2Seq model on the NIST 2003 Chinese-English translation
task. The text highlighted indicates repetition.
With the motivation to tackle this problem, we propose our
decoding-history-based Adaptive Control of Attention (ACA)
for the attention-based Seq2Seq model. The mechanism controls the output of the attention based on the decoding history, including the past information in the decoder and the
past alignment information in the attention mechanism. The
computation of the attention requires the information of the
memory vector, which is updated based on the decoding history, by manipulating the decoder output and the attention
vector. With the help of the memory, the attention can be
more adaptive to the translated contents so that the repetition
in translation can be reduced.
Our main contributions include:
• We propose a decoding-history-based Adaptive Control
of Attention for the NMT model, which tackles the conflict between the current attention and the decoding history so that the generation can be more adaptive to the
translated contents;
• Experiments on the Chinese-English translation and the
English-Vietnamese translation show that our model
outperforms the strong baselines, with the advantages of
3.61 BLEU score and 1.17 BLEU score over the best
attention-based Seq2Seq model;
• Compared with the strong baselines, the translation of
our model is with less repetition and higher accuracy.
2 Attention-based Seq2Seq
In Figure 1(a), we present a common type of attention-based
Seq2Seq with RNN as its main component, and as we use
-1
-2
-1
+1
-1
+1
-1
-1
+1
+1
-1
+1
-1
+1
ℎ-1
ℎ
ℎ+1
ℎ-1
ℎ
ℎ+1
-1
+1
-1
+1
b. Attention-Based Seq2Seq + ACA
a. Attention-Based Seq2Seq
Figure 1: Structure of the attention-based Seq2Seq and our Seq2Seq with ACA. The left is the structure of the attention-based Seq2Seq
model, and the right is the structure of our model, the attention-based Seq2Seq with the ACA.
LSTM in our model, we introduce the structure of LSTM in
the following.
2.1
Encoder
As words are discrete units, the words in the source sequence
should be sent through an embedding layer to become word
embeddings as the input. On top of the embedding layer,
the encoder turns the embeddings x = {x1 , ..., xn } into a sequence of encoder outputs h = {h1 , ..., hn } and sends out the
final hidden state hn to the decoder.
The encoder in our model is a bidirectional LSTM, which
is defined below:
fi = σ(Wf [xi , hi−1 ] + bf )
ii = σ(Wi [xi , hi−1 ] + bi )
oi = σ(Wo [xi , hi−1 ] + bo )
(1)
(2)
(3)
C̃i = tanh(WC [xi , hi−1 ] + bC )
(4)
2.2
Decoder
The decoder is responsible for decoding the final state of the
encoder hn to a new sequence y = {y1 , ..., ym }. With the
final encoder state as the initial state, the decoder is initialized to decode step by step, with a word embedding at each
time step, until it generates the token representing the end-ofsentence mark.
For the decoder, we implement a unidirectional LSTM.
The output of each time step is sent into a feed-forward
neural network to be projected into the space of vocabulary
Y ∈ R|Y |×dim . At each time step, the decoder generates a
word yt by sampling from a distribution of the target vocabulary Pvocab , where:
Pvocab = sof tmax(Wo vt )
vt = g(Wv [ct ; st ])
st = LST M(yt−1 , st−1 , Ct−1 )
(10)
(11)
(12)
Ci = fi ⊙ Ci−1 + ii ⊙ C̃i
(5)
hi = oi ⊙ tanh(Ci )
(6)
where xi is the input word embedding at each time step from
a minibatch of input sequences. LSTM consists of four gates,
which collectively control the information flow from the last
time step and the current time step. Bidirectional LSTM
contains the same structure of LSTM, but it reads the input in two directions to generate two sequences of hidden
→
−
−
→ −
→ −
→
−
→
←
−
←
− ←
− ←
−
←
−
states h = {h1 , h2 , h3 , ..., hn } and h = {h1 , h2 , h3 , ..., hn },
where:
→
−
−−→
hi = LST M(xi , hi−1 , Ci−1 )
(7)
←
−
←−−
hi = LST M(xi , hi−1 , Ci−1 )
(8)
where g(·) refers to non-linear activation function.
The global attention mechanism [Luong et al., 2015] is
applied to the LSTM output st and the encoders outputs
h = {h1 , ..., hn } in order to obtain the global attention αt,i
and the context vector ct , which is described in the following:
The encoder outputs corresponding to each time step are concatenated as mentioned below:
→ ←
−
−
hi = [ hi ; hi ]
(9)
2.3
ct =
n
X
αt,i hi
(13)
i=1
exp(et,i )
αt,i = Pn
j=1 exp(et,j )
(14)
et,i = s⊤
t−1 Wa hi
(15)
Training
The training for the Seq2Seq model is usually based on maximum likelihood estimation. Given the parameters θ and
-1
R
U
F
Figure 2: Structure of the Recurrent Memory in the ACA. “R”
and “F” refer to the Remove and Feed operations, and “g” refers to
the MLP.
source text x, the model generates a sequence ỹ. The learning
process is to minimize the negative log-likelihood between
the generated text ỹ and reference y, which in our context is
the sequence in target language for machine translation and
summary for abstractive summarization:
1 XX (i) (i) (i)
p(y |ỹ , x , θ)
N i=1 t=1 t <t
N
L=−
T
(16)
where the loss function is equivalent to maximizing the conditional probability of sequence y given parameters θ and
source sequence x.
3 Adaptive Control of Attention
As it is mentioned above, it is easy for the conventional
attention-based Seq2Seq NMT models to suffer from generating incoherent texts due to the conflict between the attention
mechanism and the decoding history. Based on the hypothesis, we propose our decoding-history-based Adaptive Control
of Attention (ACA) mechanism to tackle the problem. Instead of sending the context vector ct directly to the output
layer at each time step, we propose to update the attention
with a recurrent memory that stores the information from the
previous decoding time steps, so that the information from
the attention mechanism can be controlled to be most beneficial to the whole generation. The memory updates itself at
each time step with the information from the current decoder
output and the current context vector, so that it can learn to
remove unnecessary information and store important information at each time step. Moreover, it is responsible for restricting the information flow of the context vector in order to
mitigate the conflict between attention and neural language
model.
3.1
Recurrent Memory
As our objective in this study is to build connection for the
attention at the current time step with the decoding history,
we implement the Recurrent Memory in the decoder for updating the context vector. The recurrent memory in the model
is responsible for controlling the information flow of the attention mechanism, so that the effects of the attention mechanism can be connected with the previous decoding outputs
from the RNN as well as the attention mechanism. Moreover,
the memory should be updated at every decoding time step,
so that it can reflect the development of the decoding history.
Figure 3: Structure of Gated Control. “U” refers to the Update
Gate.
The memory mt is a representation vector at the decoding
time step t, whose initialization m0 is the last hidden state
of the encoder, which is also the initial state for the RNN of
the decoder. At each decoding time step t, the memory mt
is updated with a Remove-Feed operation. The operation is
based on the decision of the decoder output st from the RNN
and the context vector ct from the attention mechanism, so
that the memory can observe the situation at the current time
step and update itself with the guide of the current information. The structure of the Recurrent Memory is presented in
Figure 2.
At the beginning, the previous memory mt−1 experiences
a Remove-Feed operation. The decoder output st and the
context vector ct generate a Remove Gate rt to decide how
to update the memory to be adaptive to the current decoding
and a Feed Gate ft to decide how to update the memory with
new information from the decoder and the attention:
rt = σ(gr ([st ; ct ]))
ft = σ(gf ([st ; ct ]))
(17)
(18)
where g(·) refers to non-linear activation function to generate
a vector of the hidden size.
Next, the previous memory mt−1 passes through the gates
and updates itself to be mt by removing information from
the previous memory and adding new information from the
decoder output and the context vector:
mt = (rt ⊙ mt−1 ) ⊕ (ft ⊙ gi ([st ; ct ]))
(19)
The Remove operation based on st and ct can update the
information stored in the memory based on the decoding and
the attention at the current time step, so that the memory
can be adaptive to the current decoding. The Feed operation based on the two same elements can provide the memory with the new information from the current time step so
that the memory can store the repetition of the translated contents. Next, we introduce how the model makes use of the
time-sensitive memory to improve the decoding.
3.2
Gated Control
After the update, before entering the next time step, the memory mt collaborates with the decoder output st to generate
a gate for the context vector ct . Therefore, the information
from the attention mechanism is controlled by the information from the decoding history and the current state with the
help of the updated memory. The detail operations are illustrated below.
Before entering the next time step, the current memory mt
and the decoder output st generate an Update Gate ut for the
context vector ct :
ut = σ(gu ([mt ; st ]))
(20)
ĉt = ut ⊙ ct
(21)
where ĉt is the final context vector to be concatenated with
the decoder output st to generate the final output v̂t , which is
given by:
v̂t = go ([st ; ĉt ])
(22)
In the final step of decoding, instead of sending vt to the
output layer for the word prediction as mentioned in Equation 10, the model sends v̂t , outputted from the ACA, for the
prediction at each time step.
With the Gated Control, the context vector ct can be rectified based on the decoding history and the current information. The memory storing useful information of the partial
translation can encourage to model to translate contents that
are less repeated compared with the translated contents.Even
if the source-side information in the context vector is in conflict with the decoding history, the conflict can be mitigated
by the gate controlled by the memory.
4 Experiment
This section introduces the details of our experiments, including datasets, setups, baseline models as well as results.
4.1
Datasets
We evaluated our proposed model on the NIST translation
task for Chinese-English translation and provided the analysis on the same task. Moreover, in order to evaluate the performance of our model on the low-resource translation, we
also evaluated our model on the IWLST 2015 [Cettolo et al.,
2015] for the English-Vietnamese translation task.
Chinese-English Translation For the NIST translation
task, we trained our model on 1.25M sentence pairs extracted
from LDC2002E18, LDC2003E07, LDC2003E14, Hansards
portion of LDC2004T07, LDC2004T08 and LDC2005T06,
with 27.9M Chinese words and 34.5M English words. Following Su et al. [2016], we validated our model on the dataset
for the NIST 2005 translation task and tested our model on
that for the NIST 2002, 2003, 2004, 2006, 2008 translation tasks. We used the most frequent 50,000 words for
both the Chinese vocabulary and the English vocabulary.
The evaluation metric is BLEU [Papineni et al., 2002], and
we calculated the case-insensitive NIST BLEU score with
multi-bleu.perl provided by Moses1 .
English-Vietnamese Translation The data is from the
translated TED talks, containing 133K training sentence
pairs provided by the IWSLT 2015 Evaluation Campaign
[Cettolo et al., 2015]. We followed the studies of Huang et al.
[2017], and used the same preprocessing as well as the validation set and the test set. The validation set is the TED tst2012
with 1553 sentences and the test set is the TED tst2013 with
1268 sentences. The English vocabulary is 17.7K words and
the Vietnamese vocabulary is 7K words. The evaluation metric is also BLEU as mentioned above.
1
http://www.statmt.org/moses/.
4.2
Setting
We implement the models using PyTorch, and the experiments are conducted on an NVIDIA 1080Ti GPU. Both the
size of word embedding and hidden size are 512, and the
batch size is 64. We use Adam optimizer [Kingma and Ba,
2014] to train the model with the default setting β1 = 0.9,
β2 = 0.999 and ǫ = 1 × 10−8 , and we initialize the learning
rate to 0.001.
Based on the performance on the development sets, we use
a 3-layer LSTM as the encoder and a 2-layer LSTM as the
decoder. Gradient clipping is applied so that the norm of the
gradients cannot be larger than a constant, which is 10 in our
experiments. Dropout is used with the dropout rate set to 0.2.
Following Xiong et al. [2017], we use beam search with a
beam width of 10 to generate translation for the evaluation
and test, and we normalize the log-likelihood scores by sentence length.
4.3
Baselines
In the following, we introduce our baseline models for
the Chinese-English translation and the English-Vietnamese
translation respectively.
Chinese-English Translation
Following Su et al. [2016] and Zhang et al. [2017], we compare our model with the state-of-the-art NMT systems based
on our implementation and the results directly reported in
their articles, and we report the results of the baselines, Moses
and RNNSearch from the study of Su et al. [2016].
• Moses An open source phrase-based translation system with default configurations and a 4-gram language
model trained on the training data for the target language;
• RNNSearch An attention-based Seq2Seq with finetuned hyperparameters;
• Lattice The Seq2Seq model with a Lattice-based RNN
Encoder [Su et al., 2016];
• Bi-Tree-LSTM A tree-coverage Seq2Seq model which
lets the model depend on the source-side syntax
[Chen et al., 2017];
• Mixed RNN Extending RNNSearch with a mixed RNN
as the encoder [Li et al., 2017];
• CPR Extending RNNSearch with a coverage penalty
[Wu et al., 2016];
• POSTREG Extending RNNSearch with posterior regularization with a constrained posterior set
[Ganchev et al., 2010];
• PKI Extending RNNSearch with posterior regularization to integrate prior knowledge [Zhang et al., 2017].
English-Vietnamese Translation
Following Luong and Manning [2015], Raffel et al. [2017]
and Huang et al. [2017], we compare our model with the
state-of-the-art NMT models, and we present the results of
the baseline directly reported in their studies.
• RNNSearch-1 The attention-based Seq2Seq model by
Luong and Manning [2015];
Model
Moses [Su et al., 2016]
RNNSearch [Su et al., 2016]
Lattice [Su et al., 2016]
Coverage [Tu et al., 2016]
Bi-Tree-LSTM [Chen et al., 2017]
Mixed RNN [Li et al., 2017]
CPR [Zhang et al., 2017]
POSTREG [Zhang et al., 2017]
PKI [Zhang et al., 2017]
Seq2Seq+Attention
+ACA
MT-02
33.19
34.68
35.94
36.57
37.70
33.84
34.37
36.10
35.79
40.25
MT-03
32.43
33.08
34.32
35.64
34.90
31.18
31.42
33.64
35.22
38.31
MT-04
34.14
35.32
36.50
36.63
38.60
33.26
34.18
36.48
36.86
40.20
MT-05
31.47
31.42
32.40
32.73
34.35
35.50
30.67
30.99
33.08
33.14
36.82
MT-06
30.81
31.61
32.77
32.47
30.57
35.60
29.63
29.90
32.90
33.05
36.53
MT-08
23.85
23.58
24.84
25.23
22.38
22.87
24.63
24.56
28.14
Ave.
31.04
31.76
32.95
29.72
30.20
32.51
33.10
36.71
Table 2: Results of our model and the baselines (directly reported in the referred articles) on the Chinese-English translation, tested on the
NIST Machine Translation tasks in 2003, 2004, 2005, 2006 with BLEU score. “-” means that the studies did not test the models on the
corresponding datasets.
BLEU
23.30
26.10
26.80
27.69
28.67
26.93
29.10
Table 3: Results of our model and the baselines (directly reported in
the referred articles) on the English-Vietnamese translation, tested
on the TED tst2013 with the BLEU score.
20
% of the duplicates
Model
RNNSearch-1 [Luong and Manning, 2015]
RNNSearch-2 [Huang et al., 2017]
LabelEmb [Sun et al., 2017b]
NPMT [Huang et al., 2017]
NPMT+LM [Huang et al., 2017]
Seq2Seq+Attention
+ACA
w/o ACA
ACA
15
10
5
0
1-gram
• RNNSearch-2 The implementation of the attentionbased Seq2Seq by Huang et al. [2017];
• LabelEmb Extending RNNSearch with soft target representation [Sun et al., 2017b];
• NPMT The Neural Phrased-based Machine Translation
model by Huang et al. [2017];
• NPMT-LM On the basis of the NPMT, a trained 4thorder language model is added.
4.4
Results
Table 2 shows the overall results of the systems on the
Chinese-English translation task. We compare our model
with the strong baselines with their results directly reported
in their articles. To facilitate fair comparison, we compare
with the baselines that are trained on the same training set
or slightly larger training set as reported in their articles.
Many of the models are from the studies of the recent two
years, which prove to be very strong baselines. The results
have shown that for the six translation tasks, our ACA model
has clear advantage over them, with 4.95 BLEU score over
RNNSearch and 3.76 BLEU score over PKI, which proves
that our model is effective.
Table 3 shows the overall results of the systems on the
English Vietnamese translation. It can be found that on the
low-resource translation, the ACA can also bring significant
improvement for the attention-based Seq2Seq model, with
2-gram
3-gram
4-gram
Figure 4: Percentage of the duplicates at sentence level. Tested on
the NIST 2003 dataset. The red bar is the performance of our ACA,
and the blue bar is the attention-based SeqSeq without ACA.
the advantage of over 2.17 BLEU score over the strongest
attention-based Seq2Seq and 1.41 BLEU score over the
SOTA model NPMT. Moreover, compared with NPMT with
a pretrained language model, our model is still better.
4.5
Analysis
In order to test whether our model can mitigate the problem of
repetition in translation, we tested the repetition on the NIST
2003 dataset, following See et al. [2017]. We evaluated the
proportion of the duplicates of 1-gram, 2-gram, 3-gram and
4-gram in each sentence and calculated the mean value. It
can be found that at all levels, the translation of our model
has less repetition. Moreover, the advantage of ours becomes
clearer with the increase of the number of gram. Especially
for the 4-gram, the proportion of duplicates of our model is
almost only a half of that of the model without ACA. It is normal that there are repeating words in a sentence, but repeating
4-gram in most cases is unreasonable. Compared with the
model without ACA, ACA can help the Seq2Seq model reduce unreasonable repetition and therefore mitigate the problem of over-translation by taking the decoding history into
account.
BLEU (%)
38
36
34
32
30
28
26
24
22
Source: 在此之前一年, 单单手机用户已跃居全球之冠 。
Reference: The year before that, the number of mobile phone
users alone already topped the world.
Seq2Seq+Attention: A year ago, cell phone users of cell phone
users are already in the world.
+ACA: In the past year, cell phone users have leapt to the highest level in the world.
w/o ACA
ACA
0
10
20
30
40
50
60
Sentence length (no shorter than)
Figure 5: Performance on different sentence lengths. Tested on
the NIST 2003 dataset. The red line is the performance of our ACA,
and the blue line is the attention-based SeqSeq without ACA.
Moreover, we choose the NIST 2003 Chinese-English
translation dataset to test the performance of our model and
the conventional attention-based Seq2Seq model without our
ACA. We test the BLEU scores on sentences of length no
shorter than 10, 20, 30, 40, 50, and 60. With the increase
of length, the performance of both models decrease but our
model ACA always has a clear advantage over the attentionbased Seq2Seq. In our hypothesis, the model can adapt to the
decoding history by improving the attention mechanism, so it
is possible that it can perform better on the long-length sentence translation. Our analysis proves that the model can be
more robust to translating sentences of diverse lengths.
4.6
Translation Examples
Table 4 shows two translation examples of our model on
the NIST 2003 dataset, compared with the translation of the
attention-based Seq2Seq model without ACA and the reference. It is obvious that both two translation examples of our
example are similar to the references, outperforming those
of the model without ACA, which has problems of repetition
and meaning inconsistency. For the first sentence, the model
without ACA generates repetition of “cell phone users” and
misses the semantic unit “top”. On the contrary, our translation is closer to literal translation, which is more faithful to
the expression in the source. For the second example, it requires the model to reorder the translation since the name is
followed by an adverbial phrase in the source. The complex
and different structure in Chinese confused the model without
ACA, which can only generate repetition of “we are entering
a new era”. With ACA, our model successfully reorders the
translation by putting the name after the adverbial.
5 Related Work
The
studies
of
encoder-decoder
framework
[Kalchbrenner and Blunsom, 2013; Cho et al., 2014;
Sutskever et al., 2014] for this task launched the Neural
Machine Translation (NMT). To improve the focus on the
information in the encoder, Bahdanau et al. [2014] proposed
the attention mechanism, which greatly improved the perfor-
Source: 佛莱文在谈及推行再生性能源策略已获致成功时
表示: “我们正进入一个新时代。”
Reference: Speaking about the success of promoting the strategy of renewable energies, Flavin said: “we’re entering a new
era.”
Seq2Seq+Attention: “We are entering a new era.” “We are
ente-ring a new era.”
+ACA: Speaking on the success of the renewable energy strategy, Fortuyn said: “we are entering a new era.”
Table 4: Two translation examples of our model, compared with the
translation of the attention-based Seq2Seq model and the reference.
mance of the Seq2Seq model on NMT. Still, the attention
mechanism suffers from prediction failure, and therefore,
a number of studies were proposed to improve the mechanism, which also enhanced the performance of the NMT
model [Luong et al., 2015; Mi et al., 2016b; Jean et al.,
2015; Feng et al., 2016; Tu et al., 2016; Mi et al., 2016a;
Meng et al., 2016; dou; Xiong et al., 2017]. Some of them
[Tu et al., 2016; Meng et al., 2016] incorporated the previous
attention into the current attention for better alignment, but
none of them are based on the decoding history.
Besides improving attention mechanism for NMT, there
are also some more effective neural networks. Gehring et al.
[2017] turned the RNN-based model into CNN-based
model, which greatly improves the computation speed.
Vaswani et al. [2017] removed the CNN and RNN and only
used attention mechanism to build the model and showed outstanding performance. Also, some researches incorporated
external knowledge in their systems and also achieved obvious improvement [Li et al., 2017; Chen et al., 2017].
6 Conclusion
In conclusion, this paper proposes the decoding-historybased Adaptive Control of Attention (ACA) for the NMT
model, which can transmit the significant information in the
decoding history to control the output of the attention mechanism adaptively. Thus, the output of the attention mechanism is based on the the decoding history, including the past
information in the RNN decoder as well as the alignment information in the attention mechanism. With this method, the
conflict between the source-side information from the attention and the translated contents can be mitigated. Compared
with the attention-based Seq2Seq model, our model captures
more correct source information with the help of the decoding history and its translation behaves more adaptive to the
past translation. Experiments on the Chinese-English translation and the English-Vietnamese translation all show that our
model outperforms the strong baselines, which demonstrate
the effectiveness of our model.
References
Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio.
Neural machine translation by jointly learning to align and
translate. CoRR, abs/1409.0473, 2014.
Mauro Cettolo, Jan Niehues, Sebastian Stüker, Luisa Bentivogli, Roldano Cattoni, and Marcello Federico. The iwslt
2015 evaluation campaign. Proc. of IWSLT, Da Nang, Vietnam, 2015.
Huadong Chen, Shujian Huang, David Chiang, and Jiajun
Chen. Improved neural machine translation with a syntaxaware encoder and decoder. In ACL 2017, pages 1936–
1945, 2017.
Kyunghyun Cho, Bart van Merrienboer, Çaglar Gülçehre,
Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and
Yoshua Bengio. Learning phrase representations using
RNN encoder-decoder for statistical machine translation.
In EMNLP 2014, pages 1724–1734, 2014.
Shi Feng, Shujie Liu, Nan Yang, Mu Li, Ming Zhou, and
Kenny Q. Zhu. Improving attention modeling with implicit
distortion and fertility for machine translation. In COLING
2016, pages 3082–3092, 2016.
Kuzman Ganchev, Jennifer Gillenwater, Ben Taskar, et al.
Posterior regularization for structured latent variable models. Journal of Machine Learning Research, 11(Jul):2001–
2049, 2010.
Jonas Gehring, Michael Auli, David Grangier, Denis Yarats,
and Yann N. Dauphin. Convolutional sequence to sequence
learning. In ICML 2017, pages 1243–1252, 2017.
Po-Sen Huang, Chong Wang, Dengyong Zhou, and
Li Deng. Neural phrase-based machine translation. CoRR,
abs/1706.05565, 2017.
Sébastien Jean, KyungHyun Cho, Roland Memisevic, and
Yoshua Bengio. On using very large target vocabulary
for neural machine translation. In ACL 2015, pages 1–10,
2015.
Nal Kalchbrenner and Phil Blunsom. Recurrent continuous
translation models. In EMNLP 2013, pages 1700–1709,
2013.
Diederik P. Kingma and Jimmy Ba. Adam: A method for
stochastic optimization. CoRR, abs/1412.6980, 2014.
Junhui Li, Deyi Xiong, Zhaopeng Tu, Muhua Zhu, Min
Zhang, and Guodong Zhou. Modeling source syntax for
neural machine translation. In ACL 2017, pages 688–697,
2017.
Minh-Thang Luong and Christopher D Manning. Stanford
neural machine translation systems for spoken language
domains. In Proceedings of the International Workshop
on Spoken Language Translation, 2015.
Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based neural machine translation. In EMNLP 2015, pages 1412–1421, 2015.
Shuming Ma and Xu Sun. A semantic relevance based neural network for text summarization and text simplification.
CoRR, abs/1710.02318, 2017.
Shuming Ma, Xu Sun, Jingjing Xu, Houfeng Wang, Wenjie
Li, and Qi Su. Improving semantic relevance for sequenceto-sequence learning of chinese social media text summarization. In Proceedings of the 55th Annual Meeting of
the Association for Computational Linguistics, ACL 2017,
Vancouver, Canada, July 30 - August 4, Volume 2: Short
Papers, pages 635–640, 2017.
Fandong Meng, Zhengdong Lu, Hang Li, and Qun Liu. Interactive attention for neural machine translation. In COLING
2016, pages 2174–2185, 2016.
Haitao Mi, Baskaran Sankaran, Zhiguo Wang, and Abe Ittycheriah. Coverage embedding models for neural machine
translation. In EMNLP 2016, pages 955–960, 2016.
Haitao Mi, Zhiguo Wang, and Abe Ittycheriah. Supervised
attentions for neural machine translation. In EMNLP 2016,
pages 2283–2288, 2016.
Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing
Zhu. Bleu: a method for automatic evaluation of machine
translation. In ACL, 2002, pages 311–318, 2002.
Colin Raffel, Minh-Thang Luong, Peter J. Liu, Ron J. Weiss,
and Douglas Eck. Online and linear-time attention by
enforcing monotonic alignments. In ICML 2017, pages
2837–2846, 2017.
Abigail See, Peter J. Liu, and Christopher D. Manning. Get to
the point: Summarization with pointer-generator networks.
In ACL 2017, pages 1073–1083, 2017.
Jinsong Su, Zhixing Tan, Deyi Xiong, and Yang Liu. Latticebased recurrent neural network encoders for neural machine translation. CoRR, abs/1609.07730, 2016.
Xu Sun, Xuancheng Ren, Shuming Ma, and Houfeng Wang.
meprop: Sparsified back propagation for accelerated deep
learning with reduced overfitting. In Proceedings of
the 34th International Conference on Machine Learning,
ICML 2017, Sydney, NSW, Australia, 6-11 August 2017,
pages 3299–3308, 2017.
Xu Sun, Bingzhen Wei, Xuancheng Ren, and Shuming Ma.
Label embedding network: Learning label representation
for soft training of deep networks. CoRR, abs/1710.10393,
2017.
Ilya Sutskever, Oriol Vinyals, and Quoc V. Le. Sequence to
sequence learning with neural networks. In NIPS, 2014,
pages 3104–3112, 2014.
Zhaopeng Tu, Zhengdong Lu, Yang Liu, Xiaohua Liu, and
Hang Li. Modeling coverage for neural machine translation. In ACL 2016, 2016.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and
Illia Polosukhin. Attention is all you need. CoRR,
abs/1706.03762, 2017.
Bingzhen Wei, Xu Sun, Xuancheng Ren, and Jingjing Xu.
Minimal effort back propagation for convolutional neural
networks. CoRR, abs/1709.05804, 2017.
Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun,
Yuan Cao, Qin Gao, Klaus Macherey, Jeff Klingner,
Apurva Shah, Melvin Johnson, Xiaobing Liu, Lukasz
Kaiser, Stephan Gouws, Yoshikiyo Kato, Taku Kudo,
Hideto Kazawa, Keith Stevens, George Kurian, Nishant
Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa,
Alex Rudnick, Oriol Vinyals, Greg Corrado, Macduff
Hughes, and Jeffrey Dean. Google’s neural machine translation system: Bridging the gap between human and machine translation. CoRR, abs/1609.08144, 2016.
Hao Xiong, Zhongjun He, Xiaoguang Hu, and Hua Wu.
Multi-channel encoder for neural machine translation.
CoRR, abs/1712.02109, 2017.
Jingjing Xu, Xu Sun, Xuancheng Ren, Junyang Lin, Binzhen
Wei, and Wei Li. Dp-gan: Diversity-promoting generative
adversarial network for generating informative and diversified text. CoRR, abs/1802.01345, 2018.
Jiacheng Zhang, Yang Liu, Huanbo Luan, Jingfang Xu, and
Maosong Sun. Prior knowledge integration for neural machine translation using posterior regularization. In ACL
2017, pages 1514–1523, 2017.
| 2cs.AI
|
arXiv:1803.01231v1 [stat.ME] 3 Mar 2018
Bayesian Projected Calibration of Computer
Models
Fangzheng Xie and Yanxun Xu ∗
Department of Applied Mathematics and Statistics, Johns Hopkins University
Abstract
We develop a Bayesian approach called Bayesian projected calibration to address
the problem of calibrating an imperfect computer model using observational data from
a complex physical system. The calibration parameter and the physical system are
parametrized in an identifiable fashion via L2 -projection. The physical process is imposed a Gaussian process prior, which naturally induces a prior on the calibration
parameter through the L2 -projection constraint. The calibration parameter is estimated through its posterior distribution, which provides a natural and non-asymptotic
way for uncertainty quantification. We provide rigorous large sample justifications of
the proposed approach by establishing the asymptotic normality of the posterior of
the calibration parameter with efficient covariance matrix. Through extensive simulation studies and two real-world datasets analyses, we show that Bayesian projected
calibration can accurately estimate the calibration parameters, calibrate the computer
models, and compare favorably to alternative approaches. An R package implementing the Bayesian projected calibration is publicly available at https://drive.google.
com/file/d/1Sij0P-g5ocnTeL_qcQ386b-jfBfV-ww_/view?usp=sharing.
Keywords: Asymptotic normality; Computer experiment; L2 -projection; Semiparametric
efficiency; Uncertainty quantification
∗
Correspondence should be addressed to Yanxun Xu ([email protected])
1
1
Introduction
With the rapid development of computational techniques and mathematical tools, computer
models have been widely adopted by researchers to study large and complex physical systems.
One can think of computer models as complicated nonlinear functions designed by experts
using scientific knowledge (Plumlee, 2017). Compared to physical experiments, computer
models are typically much faster and cheaper to run. Furthermore, computer models can
be used to generate data that are infeasible to collect in practice. For example, a public
available computer model called TITAN2D (Sheridan et al., 2002) was developed to simulate
granular mass flows over digital elevation models of natural terrain, to better understand
the loss of life and disruption of infrastructure due to volcanic phenomena, the data of
which are impossible to collect in real life. For more applications of computer models, we
refer to Fang et al. (2005), Santner et al. (2013), and the April 2018 issue of Statistica
Sinica (http://www3.stat.sinica.edu.tw/statistica), which are devoted to computer
experiments and uncertainty quantification.
In this paper we consider the calibration problem in computer models when they include
not only the variables that can be measured, often referred to as design, but also unknown
parameters that are not directly available in the physical system. These parameters are
referred to as calibration parameters in the literature (Tuo and Wu, 2016). The goal of calibration is to estimate calibration parameters by combining observational data from physical
systems and simulated data from computer models, so that the computer models with the
estimated calibration parameters plugged-in provide decent approximations to the underlying physical systems. Formally, we model the outputs (yi )ni=1 of the physical system η at
design (xi )ni=1 through a nonparametric regression model
yi = η(xi ) + ei ,
i = 1, · · · , n,
where (ei )ni=1 are independent N(0, σ 2 ) noises. The computer model y s (·, θ), also referred to
as the simulator, is a function designed by scientific experts to model the unknown physical
system η(·) when the calibration parameter θ is appropriately estimated.
Despite the success of computer models in many scientific studies, researchers often ask
2
the following question: is the computer model a suitable surrogate for the real physical
system? Compared to the physical system, traditional computer models are rarely perfect or
exact due to their fixed parametric nature or simplifications of complex physical phenomenon
(Tuo and Wu, 2015): i.e., there exists discrepancy between a physical system η(·) and
its corresponding computer model y s (·, θ) even if the computer model is well calibrated.
Kennedy and O’Hagan (2001) first tackle this discrepancy issue under a Bayesian framework,
which has been influential among many other statisticians and quality control engineers. For
an incomplete list of references, we refer to Higdon et al. (2004); Bayarri et al. (2007); Qian
and Wu (2008); Joseph and Melkote (2009); Wang et al. (2009); Chang and Joseph (2014);
Brynjarsdóttir and OHagan (2014); Storlie et al. (2015) among others.
Theoretical properties of calibration problem was not well understood until Tuo and Wu
(2015, 2016), who pointed out that the calibrated computer models estimated by Kennedy
and O’Hagan (2001) could lead to poor approximations to physical systems. Identifiability
issue of the calibration parameter in Kennedy and O’Hagan (2001) was also noticed by Wong
et al. (2017) and H. P. Wynn, among several other discussants, in their written discussion
of Kennedy and O’Hagan (2001). In short, identifiability issue refers to the phenomenon
that the distribution of the observed data from physical system does not uniquely determine
the value of the corresponding calibration parameter given the computer model. In contrast
to the Bayesian methods, which are traditionally applied to computer model calibration
problems, Tuo and Wu (2015, 2016) and Wong et al. (2017) address the identifiability issue
rigorously in frequentist frameworks and provide corresponding theoretical justifications.
We propose a Bayesian method for computer model calibration called Bayesian projected
calibration. To the best of our knowledge, our work is the first one in the literature that
simultaneously achieves the following objectives:
a) Identifiability: The proposed approach is formulated in an identifiable fashion. Tuo
and Wu (2015, 2016) and Wong et al. (2017) define the “true” value of the calibration
parameter to be the one that minimizes the L2 distance between the computer model
y s (·, θ) and the physical system η(·). Following their work, the proposed Bayesian
projected calibration provides a Bayesian method to estimate this “true” value of the
3
calibration parameter.
b) Uncertainty quantification: The Bayesian projected calibration provides a natural
way for uncertainty quantification of the calibration parameter through its full posterior
distribution. Tuo and Wu (2015) showed the asymptotic normality of the L2 -projected
calibration estimate for uncertainty quantification of the calibration parameter, which
may not work in practice, since the amount of physical data is usually very limited
(Tuo, 2017). Hence a Bayesian approach is desired, especially when data are scarce.
c) Theoretical guarantee:
We show that the full posterior distribution of the cali-
bration parameter is asymptotically normal with efficient covariance matrix. Earlier
literatures either only provide asymptotic results of specific point estimators (Tuo and
Wu, 2015, 2016; Wong et al., 2017; Tuo, 2017), or formulate a Bayesian methodology
for calibration problem without large sample evaluation (Plumlee, 2017). Our method
represents the first effort in providing the theoretical guarantee for the full posterior
distribution of Bayesian methods in computer model calibration.
The rest of the paper is organized as follows. In section 2 we formulate the calibration
problem rigorously in an identifiable fashion and introduce the Bayesian projected calibration
method. Section 3 elaborates on the asymptotic properties of the posterior distribution of
the calibration parameter. In section 4, we demonstrate the advantages of the Bayesian
projected calibration in terms of estimation accuracy and uncertainty quantification via
simulation studies and two real-world data examples. Conclusion and further discussion of
the paper are in section 5.
2
2.1
Problem Formulation
Background
We first briefly review the frequentist L2 -projected calibration approach proposed by Tuo and
Wu (2015) before introducing the proposed Bayesian projected calibration method, which
can be regarded as the Bayesian version of the L2 -projected calibration.
4
Suppose one has collected responses (yi )ni=1 from a physical system η on a set of design
points (xi )ni=1 ⊂ Ω ⊂ Rp , where η : Ω → R is a deterministic function, and the design space
Ω is the closure of a connect bounded convex open set in Rp . The physical responses (yi )ni=1
are noisy due to measurement or observational errors, and hence can be modeled by the
following nonparametric regression model:
yi = η(xi ) + ei ,
i = 1, . . . , n,
(2.1)
where ei are independent N(0, σ 2 ) noises. Such a model has been widely adopted in the
literature of calibration (Kennedy and O’Hagan, 2001; Tuo and Wu, 2015; Tuo, 2017; Wong
et al., 2017).
Let Θ be the parameter space of the calibration parameter θ. We assume that Θ ⊂ Rq
is compact. A computer model is a deterministic function y s : Ω × Θ → R that produces
an output y s (x, θ) given a controllable input x ∈ Ω and the calibration parameter θ ∈ Θ.
The goal of calibration is to estimate θ given the computer model y s and the physical data
(yi )ni=1 , such that the calibrated computer model approximates the physical system well.
However, as pointed out by Tuo and Wu (2016) and Wong et al. (2017), the calibration
parameter θ cannot be identified without further restriction, in the sense that θ cannot be
uniquely determined by the distribution of Dn . More precisely, by alternatively expressing
the physical system η in terms of the computer model y s (x, θ) and a discrepancy δ(x) as
follows (Kennedy and O’Hagan, 2001; Wong et al., 2017; Plumlee, 2017; Tuo, 2017):
η(x) = y s (x, θ) + δ(x),
where the discrepancy function δ is completely nonparametric, it is clear that (θ, δ) cannot
be uniquely identified by the physical system η. Therefore, the “true” value of the calibration
parameter that gives rise to the physical data is not well-defined.
To address the identifiability issue, Tuo and Wu (2015) define the “true” value of θ
through the L2 -projection:
∗
θ = arg min kη(·) − y
θ∈Θ
s
(·, θ)k2L2 (Ω)
Z
= arg min
θ∈Θ
[η(x) − y s (x, θ)]2 dx.
(2.2)
Ω
The L2 -projected calibration method provides an estimate θbL2 for θ ∗ using a two-step proce-
5
dure. First, an estimate ηb of the physical system η is obtained via the kernel ridge regression
(Wahba, 1990) given the physical data (xi , yi )ni=1 :
n
ηb = arg min
f ∈HΨ (Ω)
1X
[yi − f (xi )]2 + λkf kHΨ (Ω) ,
n i=1
(2.3)
where Ψ : Ω × Ω → R is a positive definite covariance function, and k · kHΨ (Ω) is the native
norm of the reproducing kernel Hilbert space (RKHS) associated with Ψ. We refer to Wahba
(1990) and Wendland (2004) for detailed treatment of these concepts. Then, the L2 -projected
calibration estimate θbL for θ ∗ is given by
2
θbL2 := arg min kb
η (·) − y s (·, θ)k2L2 (Ω) .
(2.4)
θ∈Θ
√
The L2 -projected calibration has very nice theoretical properties: θbL2 is not only nconsistent for θ ∗ , but is also semiparametric efficient (Tuo and Wu, 2016). In other words,
it provides an optimal estimator to the “true” calibration parameter.
2.2
Bayesian Projected Calibration
The L2 -projected calibration estimate θbL2 enjoys nice asymptotic properties. Nevertheless,
it is a frequentist approach, therefore uncertainty quantification needs to be assessed via
additional procedures, e.g., bootstrap (Wong et al., 2017), etc. In what follows, we address
this issue by formulating its Bayesian counterpart.
We follow the definition of the “true” value θ ∗ of θ given in (2.2), as it minimizes the
uncertainty beyond the computer model for explaining the physical system. There are two
unknown parameters: the physical system η, taking values in some function space F, and
the calibration parameter θ ∈ Θ. The statistical model for calibration can be defined by
∗
s
2
P = Ψσ (y − η(x)) : η ∈ F, θ = arg min kη(·) − y (·, θ)kL2 (Ω) ,
θ∈Θ
where Ψσ (·) is the density function of N(0, σ 2 ). Namely, the parameter (η, θ ∗ ) is constrained
on a manifold in F × Θ defined by
∗
∗
s
2
M = (η, θ ) ∈ F × Θ : θ = arg min kη(·) − y (·, θ)kL2 (Ω) .
(2.5)
θ∈Θ
It is therefore natural to treat the “true” calibration parameter θ ∗ as a functional θ ∗ : F → Θ,
6
η 7→ arg minθ kη(·) − y s (·, θ)k2L2 (Ω) , of a physical system η, and we denote this functional by
θη∗ . To distinguish the parameter (η, θη∗ ) and the truth that generates the data, we denote
η0 to be the true physical system producing physical data (yi )ni=1 , and θ0∗ = θη∗0 .
We now introduce the Bayesian projected calibration. The unknown physical process η
is imposed a mean-zero Gaussian process prior Π = GP(0, τ 2 Ψ), where Ψ : Ω × Ω → R+
is a positive definite covariance function, and τ > 0 is a scaling factor. Let Dn denote the
physical data (xi , yi )ni=1 , and Π(· | Dn ) denote the posterior distribution given Dn . It is
straightforward to show that the posterior distribution of η is also a Gaussian process with
e where
mean function ηe and covariance function Ψ,
ηe(x) = τ 2 Ψ(x1:n , x)T (τ 2 Ψ(x1:n , x1:n ) + σ 2 In )−1 y,
e
Ψ(x,
x0 ) = τ 2 Ψ(x, x0 ) − τ 2 Ψ(x1:n , x)T (τ 2 Ψ(x1:n , x1:n ) + σ 2 In )−1 τ 2 Ψ(x1:n , x0 ).
(2.6)
(2.7)
Here Ψ(x1:n , x) = [Ψ(x1 , x), · · · , Ψ(xn , x)]T ∈ Rn , Ψ(x1:n , x1:n ) = [Ψ(xi , xj )]n×n ∈ Rn×n ,
and y = [y1p , · · · , ynp ]T ∈ Rn . Notice here the predictive mean ηe(x) given physical data Dn
coincides with the kernel ridge regression estimate ηb for some suitably chosen τ (see, for
example, Rasmussen and Williams, 2006). The Gaussian process prior GP(0, τ 2 Ψ) on η naturally induces a prior distribution on θη∗ through the constrained manifold M in (2.5). The
Bayesian projected calibration can be regarded as a Bayesian version of the L2 -projected
calibration method, because both methods estimate the “true” value of θ over the constrained manifold M using Bayesian and frequentist approach, respectively. Furthermore,
in Section 3 we will show that the posterior of θη∗ under the Bayesian projected calibration
is asymptotically centered at the L2 -projected calibration estimate θbL2 .
Remark 1 (Expensive computer model). In the literature of computer experiments, it is
common that the computer model y s is not directly available to us or time-consuming to
run, in which case the computer model can be only computed at given design points. Thus
finding an emulator ŷ s for y s using the data from the computer outputs at given design points
is needed. One first collects a set of data (xsj , θjs , yjs )m
j=1 from m runs of the computer model,
where yjs = y s (xsj , θjs ) is the output at the design point xsj , then estimate the emulator ybs
using the data (xsj , θjs , yjs )m
j=1 . There are varieties of methods for constructing emulators for
7
computer experiments, including Gaussian process models (Santner et al., 2013), radial basis
function interpolation (Wendland, 2004), polynomial chaos approximation (Xiu, 2010), or
the smoothing spline ANOVA (Wahba, 1990). To perform computer model calibration when
the computer model y s is not directly available or time-consuming to run, the calibration
parameter θη∗ can be estimated by replacing y s with the corresponding emulator ybs .
3
Theoretical Properties
In this section we provide large sample justifications of the proposed Bayesian projected
calibration. In particular, asymptotic characterization of the posterior distribtuion Π(θη∗ ∈
· | Dn ) of the calibration parameter θη∗ given the physical data Dn is offered. The posterior
of θη∗ has similar behavior as the L2 -projected calibration estimate θbL2 : Π(θη∗ ∈ · | Dn )
√
is not only n-consistent, but also asymptotically normal with efficient covariance matrix.
The asymptotic normality of Bayesian posterior is also referred to as Bernstein-von Mises
(BvM) limit (see, for example, chapter 10 in van der Vaart, 2000). The development of
semiparametric BvM theorem had not been established until Bickel and Kleijn (2012). For a
thorough treatment of BvM limits of smooth functionals in semiparametric models, we refer
to Castillo and Rousseau (2015a).
Before proceeding to the main results, we introduce some notations and definitions. Given
an integer vector k = [k1 , · · · , kp ]T and a function f (x1 , · · · , xp ) : Ω → R, denote Dk
k
to be the mixed partial derivative operator defined by Dk f = ∂ |k| f /∂xk11 · · · ∂xpp , where
P
|k| := pj=1 kj . Let α > 0 be positive, and α be the greatest integer strictly smaller than α.
The α-Hölder norm of a function f : Ω → R is defined by
kf kCα (Ω) := max
k:|k|≤α
Dk f
+ max sup
L∞ (Ω)
k:|k|=α x6=x0
|Dk f (x) − Dk f (x0 )|
.
kx − x0 kα−α
The α-Hölder space of functions on Ω, denoted by Cα (Ω), is the set of functions with finite α-Hölder norm.
The α-Sobolev space of functions, denoted by Hα (Ω), is the set
of functions f : Ω → R that can be extended to Rp such that the Fourier transform
8
fb(t) = (2π)−p
R
Rp
T
eit x f (x)dx satisfies (van der Vaart and Zanten, 2011)
Z
2
α
1 + ktk2 fb(t) dt < ∞.
Rp
Denote Ψα to be the Matérn correlation function with roughness parameter α
√
α
√
1
0
0
Ψα (x, x ) =
2αkx
−
x
k
K
(
2αkx − x0 k),
α
Γ(α)2α−1
(3.1)
where Kα is the modified Bessel function of the second kind.
To study the asymptotic behavior of the posterior of θη∗ , we first explore convergence
properties of the physical system η. For simplicity we assume that the design space Ω is
the unit hypercube [0, 1]p , and the design points (xi )ni=1 are independently and uniformly
sampled from Ω. The computer model y s is assumed to be directly available to us or cheap
to run. Such a simplified assumption is also adopted in Wong et al. (2017), and it does
no harms to the theoretical analysis, as the amount of data from computer experiments is
typically much larger than the sample size of the physical data. In addition, the computer
s
data (xsj , θjs , yjs )m
bs , when
j=1 are deterministic, and the approximation error between y and y
sufficiently small as m gets large, does not affect the stochastic analysis here. Therefore one
may assume that the approximation error of ybs to y s is negligible. The true but unknown
physical system η0 is assumed to lie in the intersection of the α-Hölder space Cα (Ω) and αSobolev space Hα (Ω) for some α > p/2. We assume that the prior Π for η is the mean-zero
Gaussian process GP(0, τ 2 Ψα ) and without loss of generality, the scaling factor τ is fixed at
1.
Theorem 1 (Convergence of η). Suppose η is imposed the Gaussian process prior Π =
GP(0, Ψα ), and η0 ∈ Cα (Ω) ∩ Hα (Ω), where α > p/2. Then for any sequence Mn → ∞,
E0 Π kη − η0 kL2 (Ω) > Mn n−α/(2α+p) | Dn → 0,
and there exists some constant M > 0 such that
Π kη − η0 kL∞ (Ω) > M | Dn → 0
in P0 -probability.
The resulting rate n−α/(2α+p) is proven to be optimal when the underlying true function
9
η0 is an α-Hölder or α-Sobolev function (See, for example, Stone, 1982, van der Vaart and
Wellner, 1996, and Ghosal and van der Vaart, 2017).
We next discuss the property of θη∗ as a functional: η 7→ θη∗ . Under certain regularity
conditions to be stated next, θη∗ yields a first-order Taylor expansion with respect to η locally
around η0 . Such a smooth property of the functional θη∗ serves as the building block to derive
the asymptotic normality of the posterior of θη∗ .
A1 θη∗ is the unique solution to (2.2) and is in the interior of Θ for η in an L2 -neighborhood
of η0 .
A2 supθ∈Θ ky s (·, θ)kL2 (Ω) < ∞.
A3 The Hessian matrix
Z
Vη =
Ω
∂2
s
∗
[η(x) − y (x, θη )] dx
∂θ∂θ T
is strictly positive definite in an L2 -neighborhood of η0 .
A4 For all j, k = 1, . . . , q, it holds that
∂y s
(·, θ)
∂θj
sup
θ∈Θ
< ∞,
HΨα (Ω)
∂ 2ys
(·, ·) ∈ C1 (Ω × Θ).
∂θj ∂θk
Lemma 1 (Taylor Expansion). Under conditions A1-A4, there exists some > 0 and some
(1)
(1)
(2)
positive constants Lη0 and Lη0 depending on η0 only, such that kθη∗ − θ0∗ k ≤ Lη0 kη − η0 kL2 (Ω)
and
θη∗
−
θ0∗
Z
−2
[η(x) − η0 (x)]V0−1
Ω
∂y s
2
(x, θ0∗ )dx ≤ L(2)
η0 kη − η0 kL2 (Ω)
∂θ
(3.2)
whenever kη − η0 kL2 (Ω) < , where V0 = Vη0 .
It follows immediately from the convergence results of the physical system η (theorem 1)
and the Taylor expansion property of θη∗ (lemma 1) that the posterior of θη∗ is consistent.
Corollary 1 (Consistency of θη∗ ). Suppose η is imposed the Gaussian process prior Π =
GP(0, Ψα ), and η ∈ Cα (Ω)∩Hα (Ω). Then the posterior of θη∗ is consistent, i.e., Π(kθη∗ −θ0∗ k >
| Dn ) → 0 in P0 -probability for any > 0.
10
Now we are in a position to characterize the asymptotic behavior of the posterior distribution of θη∗ , which is the main result of this paper. Under certain regularity conditions, the
√
posterior distribution of n(θ ∗ − θbL ) is asymptotically normal, where θbL is the frequentist
η
2
2
L2 -projected calibration estimator of θ proposed by Tuo and Wu (2015) (see section 2). We
describe the L2 -projected calibration procedure in our context for completeness:
n
1X
[yi − f (xi )]2 + λn kf k2HΨν (Ω) ,
ηb = arg min
n
f ∈HΨν (Ω)
i=1
θbL2 = arg min kb
η (·) − y s (·, θ)k2L2 (Ω) ,
θ∈Θ
where ν = α − p/2, and λn = O(n−2α/(2α+p) ) is a sequence depending on the sample size of
the physical data Dn .
Theorem 2 (Asymptotic Normality). Suppose η is imposed the Gaussian process prior
Π = GP(0, Ψα ), and η0 ∈ Cα (Ω) ∩ Hα (Ω), where α > p/2. Under conditions A1-A4, it holds
that
sup Π
A
√
n(θη∗ − θbL2 ) ∈ A | Dn − N 0, 4σ 2 V0−1 WV0−1 (A) = oP0 (1),
provided that
Z
W=
Ω
s
∂y s
∗ ∂y
∗
(x, θ0 ) T (x, θ0 ) dx
∂θ
∂θ
is strictly positive definite, where the supremum is taken over all measurable subsets in Rq .
Tuo and Wu (2015) proved that the L2 -projected calibration estimate θbL2 is also asymp√
L
totically normal: n(θbL − θ ∗ ) → N(0, 4σ 2 V−1 WV−1 ). Furthermore, the covariance matrix
2
4σ 2 V0−1 WV0−1
0
0
0
achieves semiparametric efficiency in the sense that there does not exist a
regular estimate with a smaller asymptotic covariance matrix (in spectrum). The posterior
of θη∗ possesses a similar optimal behavior, since the covariance matrix of the asymptotic
√
posterior of n(θη∗ − θbL2 ) coincides with that of θbL2 .
√
We finish this section with the following n-consistency result of θη∗ , which is a refinement
of corollary 1. It is a consequence of theorem 2 and the asymptotic normality of θbL .
2
√
Corollary 2 ( n-Consistency of θη∗ ). Suppose η is imposed the Gaussian process prior
Π = GP(0, Ψα ), and η0 ∈ Cα (Ω) ∩ Hα (Ω), where α > p/2. Under the conditions of theo11
√
rem 2, the posterior of θη∗ is n-consistent, i.e., for any sequence Mn → ∞, it holds that
√
E0 Π nkθη∗ − θ0∗ k > Mn | Dn → 0.
We defer the proof of Theorem 2 to the Appendix. The proofs of all other theorems,
lemmas, and propositions are deferred to the supplementary material.
4
Numerical Examples
This section provides extensive numerical examples to evaluate the proposed Bayesian projected calibration. Subsection 4.1 presents simulation studies via three synthetic examples.
Two real-world data examples are included in subsections 4.2 and 4.3, respectively.
4.1
Simulation Studies
We consider three configurations.
• Configuration 1. The computer model is
y s (x, θ) = 7[sin(2πθ1 − π)]2 + 2[(2πθ2 − π)2 sin(2πx − π)],
and the physical system coincides with the computer model when θ0∗ = [0.2, 0.3]T ,
i.e., η0 (x) = y s (x, θ0∗ ). The design space Ω is [0, 1], and the parameter space Θ for θ
is [0, 0.25] × [0, 0.5]. We simulate n = 50 observations from the randomly perturbed
physical system yi = η0 (xi ) + ei , where (xi )ni=1 are uniformly sampled from Ω, and the
variance for the noises (ei )ni=1 is set to 0.22 .
• Configuration 2.
We follow an example provided in Gu and Wang (2017). The
computer model is y s (x, θ) = sin(5θx) + 5x, and the physical system is η0 (x) =
5x cos(15x/2) + 5x. The design space Ω is [0, 1], and the parameter space Θ for θ
is [0, 3]. We simulate n = 30 observations from yi = η0 (xi ) + ei with var(ei ) = 0.22 ,
and (xi )ni=1 are equidistant on Ω. The L2 -discrepancy kη0 (·) − y s (·, θ)kL2 (Ω) between
the computer model y s and the physical system η0 as a function of θ is depicted in
Figure 1. The minimizer of the L2 -discrepancy is at θ0∗ = 1.8771.
12
• Configuration 3. We use the pedagogical example in Plumlee (2017). The physical
system is η0 (x) = 4x + x sin(5x) and the computer model is y s (x, θ) = θx, where
x ∈ Ω = [0, 1] and θ ∈ Θ = [2, 4]. We take (xi )ni=1 = {0, 0.05, 0.1, 0.15, 0.2, · · · , 0.8},
and the responses are given by yi = η0 (xi )+ei with var(ei ) = 0.022 . The L2 -discrepancy
as a function of θ is given by
kη(·) − y s (·, θ)kL2 (Ω) =
p
0.33(4 − θ)2 − 0.2898(4 − θ) + 0.201714,
and is minimized at θ0∗ = 3.5609.
2.2
1.8
L2 Loss
2.6
L2 Loss of Configuration 2
0.0
0.5
1.0
1.5
2.0
2.5
3.0
θ
Figure 1: The L2 -discrepancy kη0 (·) − y s (·, θ)kL2 (Ω) between the computer model y s and the
physical system η0 as a function of θ for configuration 2.
For the three configurations described above, we impose the Gaussian process prior
GP(0, τ 2 Ψα ) on η, where Ψα is the Matérn covariance function given by (3.1) with α = 5/2.
Here the scaling factor τ is set to τ = 1 for all 3 configurations for the ease of implementation.
To draw posterior samples of θη∗ , we first draw posterior samples of η using formula (2.6) and
(2.7), then compute θη∗ by θη∗ = arg minθ kη(·) − y s (·, θ)k2L2 (Ω) . For all three configurations,
1000 samples of θη∗ are drawn from the posterior distribution after burn-in for subsequent
analysis.
For comparison, we implement the calibration method by Kennedy and O’Hagan (2001)
(abbreviated as KO) and the orthogonal Gaussian process method by Plumlee (2017) (abbreviated as OGP). For both methods in all three configurations, Markov chain Monte Carlo
is applied to draw 1000 posterior samples after discarding 1000 burn-in samples.
13
For configuration 1, the summary statistics of the posterior distribution of θη∗ are provided in Table 1, together with those using the KO method and the OGP method. We can
see that both the Bayesian projected calibration and the OGP outperform the KO in terms
of both accuracy of point estimates (posterior means) and uncertainty quantification (length
of credible intervals and standard deviations of posterior samples). Although OGP provides
similar posterior inference compared to the Bayesian projected calibration, the computation
runtime is significantly longer than the other two methods. The computational bottleneck
was also discussed in section 6 of Plumlee (2017). Figure 2(a) presents the scatter plot of
Table 1: Summary Statistics of Posterior of θ for Configuration 1 (the simulation truth is
θ0∗ = [0.2, 0.3]T )
Projected Calibration KO Calibration OGP Calibration
θ
θ1
θ2
θ1
θ2
θ1
θ2
Mean
0.1984
0.3009
0.1927 0.2953 0.2068 0.3024
Standard Deviation 0.0011
0.0013
0.0235 0.01537 0.0005 0.0006
97.5%-Quantile
0.2006
0.3034
0.2474 0.3360 0.2013 0.2999
0.1963
0.2984
0.1653 0.2722 0.1992 0.2975
2.5%-Quantile
Runtime
279s
0.834s
40562s
√
n(θη∗ − θbL2 ). The level curves of the BvM limit shows that the
√
asymptotic distribution of Π( n(θη∗ − θbL2 ) | Dn ) developed in section 3 offers a decent apthe posterior samples of
proximation to the exact posterior. Comparing Figures 2 (a) and (b), the Bayesian projected
calibration outperforms the KO in terms of uncertainty quantification. We also investigate
the performance of the calibrated computer model in Figures 2 (c) and (d). The point-wise
95%-credible bands for the computer model also validate that Bayesian projected calibration
provides a better estimate to y s in contrast to the KO approach.
Similarly, for configuration 2, the advantages of the Bayesian projected calibration in
terms of uncertainty quantification and computational cost can be summarized from the
statistics reported in Table 2. We also provide the densities of the posterior using the
proposed method and KO approach in Figures 3 (a) and (b), respectively. Again, it can be
seen that the Bayesian projected calibration provides smaller uncertainty compared to the
KO calibration. In addition, Figure 3 (c) shows that the asymptotic BvM limit approximates
the exact posterior well even though the sample size is only n = 30.
14
0.04
^
(a) Posterior Distribution of n (θstar
η − θL2)
(b) Posterior Samples of θ based on KO Calibration
●
Posterior Samples of θstar
η
●
●
Level Curve of BvM limit
●
●
●
●
●
● ●
● ●
● ●
● ●●
●
● ● ● ● ●●●
●●
●●
●
● ●● ● ●
●
●
●
●
●
●
●●
●
●
● ● ●
● ●● ●●●●●
●
●●● ●●●
●●● ●
●
●
● ● ● ●●
●●
●●
●●●
●●●
●● ●●●●
● ●
●
●● ●●● ●● ●
●
●●
●●●
●●
● ●●
●
●●●
● ●●●
●
● ●●
●
●
●● ●
●
●●●
●●●
●● ● ●
●●
● ●●
● ●
●●
●●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
● ●● ● ●
● ●●
●
●●
●●
●●
● ●●
●
●●
●
●
●
● ●
●
●
●
●
●
● ●●
●
●
●
●●
●
●
●●●
●
●
●
●●
●
●●●
●
●●
●
●●●
●
●●
●
●
●●●●
●
●●
●●
●
●●
●
●●
●● ●
●
●●
●●● ●
●
●●●
●
●●
●
●●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●●
●
●
●●●● ●
●
●
●●
●
●
●
●● ●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●●●●●
●
●
●
●
●
●
●
●
● ●● ●
●
●●●
●
●
●
●
●
●
●
●
●
●●●
●
●
●●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●●
●●
● ●●
●●
●
● ●●
●●
●
●●●
●●● ●●
●
●●●
●
●●●●
●
●
●●
●
●
●
●●●●
●
●●
●
●
●
●
●●
●
●
●
●●●
●●
●
●
●
●●
●●●
● ●
●
●
●
●
●●
●
●
●
●● ●
●●●
●●
●
●●
●
●
●
●●●
●●
●●
●● ●
●
●●
●●●
●
●
●●
●●
● ●●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●●●
●●
●
●
●●
●
●●●●●
●
● ●●●●
●●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●● ●●
●●
● ●●● ● ●●●
●●
●●
●
● ●● ●
●●
●
●● ●
●●●●
●●●
●●
●●
●●●
●
● ●●
●
● ●●
●
●
●
●
● ●
●●
●
●
●
●●
●
●● ●●●
●
●●
●
●
●●●
●●
●
●●●
● ● ●
●
●●
●●
●
●●●
●●
●
●
●
●● ● ●
●
●
●●●
●●● ●
●
●●
●
●
●●
●
●
●
●
● ●● ●
●
●
● ● ● ● ●
●●
●
● ●● ●● ● ●● ●● ●
● ●
●
●●
●
● ● ● ●● ● ●●●
●● ●
●
●● ●
●●
●●
●
● ●
●●
●
●
●●
●
●
●
−0.02
−0.01
0.00
^
n (θstar
η1 − θL21)
0.01
0.34
θ2
●●
0.30
●
0.26
0.02
0.00
●
●
−0.02
^
n (θstar
η2 − θL22)
●
●
0.02
●
0.16
0.18
●
8
●
●
●
7
●●●
●
●
●
●
●
●●
●
●●
0.2
●
● ●
●
●
●
●
4
6
●
● ●
●●
●●
●
● ●
●
●
●●
● ●
●
●
●
●● ●
●
●
●
●
●●
●
●●●
●
● ●
●
●
●
●
● ●
●
2
5
4
3
●
●
●
●
0.0
●
●
●
●
●
●
●
η
●
●
η
●
●
●
●
Observation
True Process
KO Calibrated Model (mean)
●
● ●
●
●
●
●
0.24
(d) KO Calibration
10
●
●
●
●
0.22
8
●
●
●
●
●
0.20
θ1
6
9
10
Observation
True Process
Projected Calibrated Model (mean)
●
●
●
● ●
● ●
●
● ●
● ●●
●
●
● ●
●
●
●● ●
●
●●● ●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●● ● ●
● ●
●● ●
● ●●● ● ●
●
● ●● ●● ●
●● ● ●
●
●
● ●
●
● ●● ● ●●●
●
●●●
●●
● ●● ●●
● ●● ●
●● ●
●
●● ●● ●
● ●
●
●● ● ● ●
●
●
●
●
●
●
●
●
●
●
●●●●
● ● ●● ●
● ●● ●● ● ● ●
●
●
●●● ●
●● ●●●●
●●● ●
●● ● ● ●
●
●●●●
●●
●●● ●●
● ● ●●●
●
●● ● ●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
● ●●
●●
●●● ● ●
●● ● ● ● ●
●● ●
● ● ● ●●●
●●
●●● ●●●● ●●
●
●●●
● ●●●●●●
●●●●
●
●
● ●●
●● ●
●● ●●●● ●
●
●
●●
●
●●●
●
●●
●● ● ●●
●
● ●●●
●
●
●
●●●
●●●
●●●
● ●●
●●
●
●●
●
●
●
●●●●
●●●
●
●●●●●
●●●●
●●●
●
●
● ●
●● ●
●●
●
●
●
●●
●
● ●
●
●
●
●
●
●
●●
●
●● ●
●
●
●●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ● ●● ●
●
●
●
●●● ●●●●● ●●●●●●● ●●
●
● ●
●
●●
● ●●
●
●
●●
●
●
●●●
● ● ●
●
●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
● ●
●● ● ● ●●●● ●●
●
● ●
● ●●
●
●●●●●
●
●●●
●●● ●●
●●
●●
●●●●
●
●
● ●
●
● ●●
●
●●
●●
●
●●
●●
●
●●●●●
●
●●●●
●●●
●●●●●●●
●●
●● ●
●
●●
●●
●●
●●●●
●●
●●●
●
●
●
●●●
● ●●● ● ●● ●
●●
●
● ●● ●● ●
●
●
● ●●
●●
●●●●●
●
●●
●
●●●
●
●●●
●
●
●
●● ●
●● ●
●
●●
● ●
●
●
●
●
●
● ●
●●
●●
●●●
●
●
●
●●● ●
●
●●
●●●
● ● ●● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●● ● ● ●
● ●●● ● ●
●● ●
● ● ● ● ●● ●
●
●
●
●● ●
●●●
● ●●● ● ●
● ●
● ● ●● ● ●
●● ●
●●
● ●●●●●
●
●
●
● ●
●
●● ● ●●●
● ● ●●●
●●●
● ●
●●
●
●●
●
● ● ● ● ●●
●
●●
●●
●
● ● ●●
● ●● ●
●
●
●●
●●
●●●
●
●
●
●
●
●●
●●
●● ● ● ● ● ●
●
●
●●
●
●
● ● ● ●●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
(c) Bayesian Projected Calibration
●
●
●
● ●
● ●
0.4
0.6
0.8
1.0
0.0
x
0.2
0.4
0.6
0.8
1.0
x
Figure 2: Visualization of the posterior inference for configuration
1 in the simulation studies.
√
Panel (a) shows the scatter plot of the posterior samples of n(θη∗ − θbL2 ) and the level curves
of the corresponding BvM limit. Panel (b) presents the scatter plot of the posterior samples
of θ using the KO approach. Panels (c) and (d) display the calibrated computer models (in
dashed lines) using the Bayesian projected calibration and the KO approach, respectively,
together with their corresponding point-wise 95%-credible intervals (in shaded area).
The scenario for configuration 3 is slightly challenging due to the fact that no physical
data are available on (0.8, 1]. We provide the corresponding summary statistics in Table
3. It can be seen that when the design points are not regularly spread over Ω, the KO
method provides a biased estimate of θ0∗ compared to the Bayesian projected calibration and
the OGP. For the uncertainty quantification performance measured by the width of credible
intervals and standard deviation, the OGP method and the Bayesian projected calibration
are similar, and both outperform the KO approach.
15
Table 2: Summary Statistics of Posterior of θ for Configuration 2 (simulation truth is θ0∗ =
1.8771)
Projected Calibration KO Calibration OGP Calibration
Mean
1.8805
1.8805
1.8721
Standard Deviation
0.0048
0.0593
0.0023
1.8903
1.9832
1.8766
97.5%-Quantile
2.5%-Quantile
1.8712
1.7608
1.8678
Runtime
20.697s
1.034s
31843s
10
BvM Limit Density
0
5
Density
15
Bayesian Projected Calibration: Posterior of n (θstar
η − θL2)
−0.05
0.00
n(
0.05
0.10
)
θstar
η − θL2
Figure
3: Simulation study configuration 2: The histogram of the posterior samples of
√
n(θη∗ − θbL2 ), together with the theoretical BvM limit density (red solid line).
4.2
Ion Channel Example
We apply the Bayesian projected calibration to the ion channel example used in Plumlee
et al. (2016). The dataset involves measurements from experiments concerning ion channels
of cardiac cells. Specifically, the output of the experiment is the current through the sodium
channels in a cardiac cell membrane, and the input is the time. For detailed description
of the experiment, we refer to Plumlee et al. (2016). Here we consider a subset of the
original dataset, which consists of 19 normalized current records needed to maintain the
membrane potential fixed at −35mV, together with the logarithm of the corresponding time
as the inputs. The same dataset is also studied in Plumlee (2017). For the computer model,
16
Table 3: Summary Statistics of Posterior of θ for Configuration 3 (simulation truth is θ0∗ =
3.5609)
Projected Calibration KO Calibration OGP Calibration
Mean
3.4064
3.1109
3.6001
Standard Deviation
0.0614
0.4760
0.0911
3.5964
3.9385
3.7733
97.5%-Quantile
2.5%-Quantile
3.3624
2.1467
3.4167
Clancy and Rudy (1999) suggests the following Markov model for ion channels:
y s (x, θ) = eT
1 exp[exp(x)A(θ)]e4 ,
where ei is the column vector with 1 at the ith coordinate and 0 for the rest components,
the outer exp is the matrix exponential function, and
−θ2 − θ3
θ1
0
0
θ2
−θ1 − θ2
θ1
0
A(θ) =
0
θ2
−θ1 − θ2 θ1
0
0
θ2
−θ1
.
We follow the Monte Carlo methods described in subsection 4.1 to collect 1000 posterior
samples of θ using the Bayesian projected calibration and the KO calibration approach.
The OGP calibration approach, however, fails to provide adequate samples from the posterior distribution for subsequent inference within 20 hours. The roughness parameter α for
the Matérn covariance function is set to 5/2, and we set τ = 0.02, σ = 0.001, as suggested
by Plumlee (2017). Table 4 presents the corresponding comparison of summary statistics.
The calibrated computer models are also visualized in Figure 4. Clearly, the Bayesian projected calibration provides better estimates to both the calibration parameter θ and the
computer model in terms of lower uncertainty (smaller standard deviation and thinner credible intervals). It can also be seen that the Bayesian projected calibrated computer model
approximates to the physical data better than the KO method.
17
Table 4: Summary Statistics of Posterior of θ for the Ion Channel Example
Projected Calibration
KO Calibration
θ
θ1
θ2
θ3
θ1
θ2
θ3
Mean
6.011166 5.578567 3.500813 3.4203734 0.5779939 7.4971337
Standard Deviation 0.000012 0.000006 0.000006 0.2671795 0.5668012 1.5645618
97.5%-Quantile
6.011191 5.578578 3.500824 2.963897 −0.4371212 4.718231
6.011143 5.578556 3.500802 4.023088
1.7700171 10.925112
2.5%-Quantile
KO Calibration
0.15
Bayesian Projected Calibration
0.04
●
●
●
●
−2
−1
0
1
● ●
● ● ●
● ● ●
2
0.10
●
●
●
●
●
●
●
●
0.00
●
Observation
GP Posterior Mean
KO Calibrated Model
●
●
●
0.05
0.12
●
Normalized current
●
●
0.08
Observation
GP Posterior Mean
Calibrated Model
●
●
0.00
Normalized current
●
●
3
−2
Logarithmic time
−1
0
1
●
● ●
● ● ● ●
● ●
2
3
Logarithmic time
Figure 4: Visualization of computer model calibration for the ion channel example. The left
and right panels present the calibrated computer models (dashed lines) using the proposed
approach and the KO calibration approach, respectively. The shaded area is the point-wise
95%-credible intervals for the KO calibrated computer model. The physical data (circles)
and the Gaussian process (GP) estimates of the physical system (dashed lines) are also
displayed.
4.3
Spot Welding Example
Now we consider the spot welding example studied in Bayarri et al. (2007) and Chang and
Joseph (2014). Three control variables in the experiment are the load, current, and gauge.
The physical experiments are only conducted for gauge being 1 and 2. Since the computer
model fails to produce meaningful output when the gauge is set to 1, here we only consider
the case where the gauge is 2, i.e., the control variables are the load and the current only.
The physical data are provided in Table 4 of Bayarri et al. (2007). For each fixed design
point, the mean of the 10 replicates of the output is taken as the response.
The computer model for the spot welding system, on the other hand, is not directly
available to us. In short, the computer model consists of a time-consuming finite element
18
method (FEM) for numerically solving a system of partial differential equations (PDEs).
There are 21 available runs for the computer code, as is presented in Table 3 of Bayarri et al.
(2007). As discussed in section 2, an emulator is needed as a surrogate for the computer
model in such a situation where the code is expensive. Here we apply the RobustGaSP
package (Gu et al., 2018a) to emulate the expensive FEM computer model. For theoretical
background on the RobustGaSP emulator, we refer to Gu et al. (2018b).
We follow similar approach in subsections 4.1 and 4.2 to draw posterior samples using
Bayesian projected calibration and the KO method. The only difference is that the nonavailable computer model y s is replaced by the predictive mean of the RobustGaSP emulator
based on the results of the 21 runs on the FEM computer code. The summary statistics
are presented in Table 5, indicating that the Bayesian projected calibration outperforms the
KO approach in terms of uncertainty quantification, i.e., smaller standard deviation and
thinner credible interval. The calibrated computer models using the Bayesian projected calibration and the KO approach, together with their corresponding point-wise 95%-credible
intervals, are depicted in Figure 5. We can see that in terms of computer model calibration,
both approaches behave similarly. The point-wise credible intervals, however, indicate that
the Bayesian projected calibration method outperforms the KO approach regarding uncertainty quantification. The imperfection of the computer model can also be seen from the
discrepancy presented on the right two panels of Figure 5.
Table 5: Summary Statistics of Posterior of θ for the Spot Welding Example
Projected Calibration KO Calibration
Mean
4.385933
4.107683
Standard Deviation
0.08455849
1.333926
4.505187
6.880466
97.5%-Quantile
2.5%-Quantile
4.183981
1.439716
5
Conclusion and Discussion
We develop a novel Bayesian projected calibration method following the frequentist L2 projected calibration method in Tuo and Wu (2015). The proposed approach is formulated
19
Bayesian Projected Calibration (Load = 4N)
6
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
5
●
●
24
25
26
27
28
23
24
25
26
27
28
29
KO Calibration (Load = 4N)
KO Calibration (Load = 5.3N)
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
5
5
●
●
●
●
7
●
●
●
●
●
●
●
●
●
●
●
6
●
●
30
●
●
8
●
●
●
●
●
●
●
●
●
Nugget diameter
7
30
Observation
GP Posterior Mean
Calibrated Computer Model
Current
●
●
●
●
●
6
29
●
Current
8
23
Observation
GP Posterior Mean
Calibrated Computer Model
4
4
●
Nugget diameter
7
8
●
●
●
●
6
7
●
●
●
●
Nugget diameter
●
●
●
●
●
●
●
5
Nugget diameter
8
●
●
●
●
●
●
●
●
●
Bayesian Projected Calibration (Load = 5.3N)
23
24
25
Observation
GP Posterior Mean
Calibrated Computer Model
26
27
28
29
●
4
4
●
30
23
Current
24
25
Observation
GP Posterior Mean
Calibrated Computer Model
26
27
28
29
30
Current
Figure 5: Visualization of computer model calibration for the spot welding example. The left
and right panels present the calibrated computer models (red dashed lines) as a function of
the current with the load fixed at 4N and 5.3N, respectively. The shaded areas are the pointwise 95%-credible intervals for the corresponding calibrated computer models. The physical
data (circles) and the Gaussian process (GP) estimates of the physical system (solid lines)
are also displayed.
20
in an identifiable way and naturally quantifies uncertainty in the calibration problem through
the posterior distribution. Theoretical justification of the Bayesian projected calibration is
√
provided: the marginal posterior distribution of the calibration parameter is not only nconsistent, but also asymptotically normal with efficient covariance matrix.
The estimation methods in this work and Tuo and Wu (2015) can be viewed as the
following two-step procedure: First estimate the physical system through Gaussian process
models; Then estimate the calibration parameter using the L2 -projection criterion. On the
other hand, it is possible to jointly estimate the calibration parameter and the discrepancy
between the physical system and the computer model. For example, Plumlee (2017) proposes an orthogonal Gaussian process (OGP) model to tackle this joint estimation issue.
The theoretical development for OGP, nevertheless, is only restricted to a point estimate
(Tuo, 2017): the maximum a posteriori (MAP) estimate of θ is asymptotically normal and
semiparametric efficient. It will be non-trivial to apply the technical results developed here
to the OGP calibration approach, and asymptotic characterization of the corresponding full
posterior distribution will be an interesting topic.
Similar to the OGP calibration method, the Bayesian projected calibration also involves
intractable integrals. Plumlee (2017) suggested that the intractable integrals can be numerically computed using Monte Carlo integration. However, this involves drawing N uniform
samples from Ω, and N is typically much larger than n for accurate evaluation of these
integrals, adding heavy computational burden for posterior inference. In particular, the numerical integration yields extra O(N ) operations. It is therefore desired that the computation
barriers for the two approaches can be tackled via efficient algorithms.
APPENDIX
A
Auxiliary Results
In this section we list some auxiliary results that are used to prove theorem 2. The proofs
of the lemmas stated in this section are deferred to the supplementary material. Before
proceeding, we introduce some notions and definitions that are widely used in the study
21
of empirical processes. Suppose F is a function space equipped with metric d. Given two
functions l, u ∈ F, a bracket [l, u] is a set of functions f such that l ≤ f ≤ u everywhere, and
the size of the bracket is defined to be d(l, u). The -bracketing number of F with respect
to the metric d, denoted by N[·] (, F, d), is the minimum number of brackets of size that
are needed to cover F. The bracketing integral J[·] (, F, d) is defined to be the integral of
the logarithmic bracketing number as follows:
Z p
log N (δ, F, d)dδ.
J[·] (, F, d) =
0
Suppose X is the space where random variables take values. Given a sequence (xi )ni=1 of
independent and identically distributed random variables, the empirical measure and the
empirical process of a function f : X → R, denoted by Pn f and Gn f , are defined by
n
1X
Pn f =
f (xi ),
n i=1
n
1 X
Gn f = √
[f (xi ) − Ef (xi )],
n i=1
respectively. For two variables a and b, we use a . b and a & b to denote the inequalities up
to a universal multiplicative constant, and a b if a . b and a & b.
In the empirical processes theory, maximum inequalities are widely adopted to study the
asymptotic behavior of nonparametric estimates. Here we cite one of them that is used in
the proof of theorem 2 (see, for example, lemma 19.36 in van der Vaart, 2000).
Theorem 3. Let (xi )ni=1 be independent and identically distributed according to a distribution
Px over X , and let F be a class of measurable functions f : Y → R. If kf k2L2 (Px ) < δ 2 and
kf k∞ ≤ M for all f ∈ F, where δ and M does not depend on F, then
M
E sup |Gn f | . J[·] δ, F, k · kL2 (Px ) 1 + 2 √ J[·] δ, F, k · kL2 (Px ) .
δ n
f ∈F
The following lemma is the modification of a standard probabilistic theorem for Gaussian
processes. For the related literature, we refer to van der Vaart and van Zanten (2008) and
Ghosal and van der Vaart (2017).
Lemma 2. Suppose η is imposed the Matérn Gaussian process with roughness parameter
α, and η0 ∈ Cα (Ω) ∩ Hα (Ω), where α > p/2. Let n = n−α/(2α+p) . Then there exists a
22
measurable set Bn in C(Ω) such that for sufficiently large n, the following hold:
Π(Bn | Dn ) = 1 − oP0 (1),
√
J[·] (n log n, Bn , k · kL2 (Ω) ) . (log n)2α/(2α+p) n2n .
Now denote
`n (η) =
n
X
log pη (yi , xi ) =
i=1
n
X
log Ψσ (yi − η(xi ))
i=1
to be the log-likelihood function of η given the physical data (xi , yi )ni=1 . Define the event
An = kη − η0 kL2 (Ω) ≤ Mn n ∩ kη − η0 kL∞ (Ω) ≤ M ∩ Bn ,
where Mn = log n, M is given by theorem 1, and Bn is given by lemma 2. Then by theorem
1 and lemma 2 we know that Π(An | Dn ) = 1 − oP0 (1).
Lemma 3. Suppose the conditions of theorem 2 hold. For each vector t ∈ Rq and each
η ∈ F define
2σ 2
∂y s
ηt (x) = η(x) − √ tT V0−1
(x, θ0∗ ).
∂θ
n
Then it holds that
Z
Z
exp [`n (ηt ) − `n (η0 )] Π(dη) = [1 + oP0 (1)]
exp [`n (η) − `n (η0 )] Π(dη) .
An
The asymptotic normality result of the L2 -projected calibration estimate θbL2 from Tuo
√
and Wu (2015) is also useful to study the asymptotic behavior of Π( n(θη∗ − θbL2 ) ∈ · | Dn ).
Theorem 4. Under the conditions of theorem 2, it holds that
" n
#
X ∂y s
1
θbL2 − θ0∗ = 2V0−1
ei
(xi , θ0∗ ) + oP0 (n−1/2 ).
n i=1 ∂θ
B
Proof of Theorem 2
By theorem 1 and lemma 2 we know that Π(An | Dn ) = 1 − oP0 (1). Let Θn = θη∗ : η ∈ An .
It follows directly that Π(θη∗ ∈ Θn | Dn ) = 1 − oP0 (1). Denote
Π(θη∗ ∈ · | Dn , Θn ) =
Π(θη∗ ∈ · ∩ Θn | Dn )
.
Π(θη∗ ∈ Θn | Dn )
23
Then we obtain
√
b
sup Π
− θL2 ) ∈ A | Dn − N 0, 4σ 2 V0−1 WV0−1
A
√
√
n(θη∗ − θbL2 ) ∈ A | Dn − Π
n(θη∗ − θbL2 ) ∈ A | Dn , Θn
≤ sup Π
A
√
+ sup Π
n(θη∗ − θbL2 ) ∈ A | Dn , Θn − N 0, 4σ 2 V0−1 WV0−1 .
n(θη∗
A
√
Denote Bn (A) = {θη∗ : n(θη∗ − θbL2 ) ∈ A}. Since
√
√
sup Π
n(θη∗ − θbL2 ) ∈ A | Dn − Π
n(θη∗ − θbL2 ) ∈ A | Dn , Θn
A
Π θη∗ ∈ Bn (A) ∩ Θn | Dn
∗
= sup Π θη ∈ Bn (A) | Dn −
Π(θη∗ ∈ Θn | Dn )
A
1
∗
c
∗
c
≤
sup Π θη ∈ Bn (A) ∩ Θn | Dn + Π(θη ∈ Θn | Dn ) → 0,
Π(θη∗ ∈ Θn | Dn ) A
it suffices to show that
√
∗
b
n(θη − θL2 ) ∈ A | Dn , Θn − N 0, 4σ 2 V0−1 WV0−1 .
sup Π
A
We prove the result by the method of moment generating function, namely, showing that for
any fixed vector t ∈ Rq , it holds that
Z
i
h √
1 T
∗
2 −1
−1
T
b
exp t n θη − θL2 Π(dη | Dn ) → exp t 4σ V0 WV0 t
2
An
in P0 -probability. The rest part of the proof is completed by lemma 1 and lemma 2 in Castillo
and Rousseau (2015b).
Let n = n−α/(2α+p) . The left-hand side of the preceeding display can be re-written as
Z
−1 Z
h √
i
∗
T
exp [`n (η) − `n (η0 )] Π(dη)
exp t n θη − θbL2 + `n (η) − `n (η0 ) Π(dη) .
An
q
For the vector t ∈ R , define
2σ 2 T −1 ∂y s
ηt (x) = η(x) − √ t V0
(x, θ0∗ ),
∂θ
n
and for each η, define the remainder
Rn (η, η0 ) =
n
n
kη − η0 k2L2 (Ω) − Pn (η − η0 )2 .
2
2
24
Then simple algebra shows
[`n (ηt ) − `n (η0 )] − [`n (η) − `n (η0 )]
n
n
2 X T −1 ∂y s
2
2
(xi , θ0∗ )
= − 2 kηt − η0 kL2 (Ω) − kη − η0 kL2 (Ω) − √
ei t V0
2σ
∂θ
n i=1
1
[Rn (ηt , η0 ) − Rn (η, η0 )]
σ2 Z
√
∂y s
1
(x, θ0∗ )dx − tT 4σ 2 V0−1 WV0−1 t
= 2 n [η(x) − η0 (x)]tT V0−1
∂θ
2
Ω
n
s
2 X T −1 ∂y
1
−√
ei t V0
(xi , θ0∗ ) + 2 [Rn (ηt , η0 ) − Rn (η, η0 )].
∂θ
σ
n i=1
+
Denote the remainder of the Taylor expansion of θη∗ at θ0∗ by
Z
s
∗
∗
−1 ∂y
(x, θ0∗ )dx.
r(η, η0 ) = θη − θ0 − 2 [η(x) − η0 (x)]V0
∂θ
Ω
Then by theorem 4 we have
√ ∗
T
b
t n θη − θL2 + `n (η) − `n (η0 )
√
n
2 X T −1 ∂y s
√
ei t V0
(xi , θ0∗ ) + oP0 (1) + `n (η) − `n (η0 )
=t n
−
−
∂θ
n i=1
Z
√
√
∂y s
(x, θ0∗ )dx
= tT n θη∗ − θ0∗ + oP0 (1) − 2 n [η(x) − η0 (x)]tT V0−1
∂θ
Ω
1 T
1
+ t 4σ 2 V0−1 WV0−1 t − 2 [Rn (ηt , η0 ) − Rn (η, η0 )] + `n (ηt ) − `n (η0 )
2
σ
√
1 T
1
= t 4σ 2 V0−1 WV0−1 t + ntT r(η, η0 ) + 2 [Rn (η, η0 ) − Rn (ηt , η0 )]
2
σ
T
θη∗
θ0∗
+ `n (ηt ) − `n (η0 ) + oP0 (1).
Now set Mn = log n. By lemma 1 we see that
sup
η∈An
√
√
2 −2α/(2α+p)
ntT r(η, η0 ) ≤ L(2)
. Mn2 n(p/2−α)/(2α+p) = o(1).
η0 ktk nMn n
In addition, simple algebra and the law of large numbers imply that
Rn (η, η0 ) − Rn (ηt , η0 )
2
n
s
2σ 4 X T −1 ∂y s
2
T −1 ∂y
∗
4 T −1
−1
∗
=
t V0
(xi , θ0 ) − 2σ t V0 WV0 t − 2σ Gn (η − η0 )(·)t V0
(·, θ0 )
n i=1
∂θ
∂θ
25
s
T −1 ∂y
∗
= −2σ Gn (η − η0 )(·)t V0
(·, θ0 ) + oP0 (1).
∂θ
2
We now claim that
sup |Rn (η, η0 ) − Rn (ηt , η0 )| = oP0 (1).
η∈An
Since over An , we know that kη − η0 kL2 (Ω) ≤ Mn n , kη − η0 kL∞ (Ω) ≤ M , and by lemma 2 it
holds that
√
√
J[·] (Mn n , An , k · kL∞ (Ω) ) . Mn2α/(2α+p) n2n = (log n)2α/(2α+p) n2n ,
it follows by the maximal inequality for empirical process (theorem 3) that
s
T −1 ∂y
∗
E0 sup Gn (η − η0 )(·)t V0
(·, θ0 )
∂θ
η∈An
J[·] (Mn n , An , k · kL2 (Ω) )
√
. J[·] (Mn n , An , k · kL2 (Ω) ) 1 + M
Mn2 2n n
#
"
2α/(2α+p) √ 2
√ 2
nn
Mn
2α/(2α+p)
√
. Mn
nn 1 +
Mn2 n2n
√
. Mn n2n = o(1),
and hence, it holds that supη∈An |Rn (η, η0 ) − Rn (ηt , η0 )| = oP0 (1). Therefore by applying
lemma 3 we obtain
Z
h √
i
exp tT n θη∗ − θbL2 + `n (η) − `n (η0 ) Π(dη)
An
Z
1 T
2 −1
−1
exp [`n (ηt ) − `n (η0 )] Π(dη)
= exp t 4σ V0 WV0 t + oP0 (1)
2
An
Z
1 T
2 −1
−1
= exp t 4σ V0 WV0 t + oP0 (1) [1 + oP0 (1)] exp [`n (η) − `n (η0 )] Π(dη)
2
Z
1 T
2 −1
−1
exp[`n (η) − `n (η0 )]Π(dη).
= exp t 4σ V0 WV0 t + oP0 (1)
2
The proof is thus completed.
26
References
Bayarri, M. J., Berger, J. O., Paulo, R., Sacks, J., Cafeo, J. A., Cavendish, J., Lin, C.H., and Tu, J. (2007). A framework for validation of computer models. Technometrics,
49(2):138–154.
Bickel, P. and Kleijn, B. (2012). The semiparametric Bernstein-von Mises theorem. The
Annals of Statistics, 40(1):206–237.
Brynjarsdóttir, J. and OHagan, A. (2014). Learning about physical parameters: The importance of model discrepancy. Inverse Problems, 30(11):114007.
Castillo, I. (2012). A semiparametric Bernstein–von Mises theorem for Gaussian process
priors. Probability Theory and Related Fields, 152(1):53–99.
Castillo, I. and Rousseau, J. (2015a). A Bernstein-von Mises theorem for smooth functionals
in semiparametric models. The Annals of Statistics, 43(6):2353–2383.
Castillo, I. and Rousseau, J. (2015b). Supplement to “A Bernstein-von Mises theorem for
smooth functionals in semiparametric models”.
Chang, C.-J. and Joseph, V. R. (2014). Model calibration through minimal adjustments.
Technometrics, 56(4):474–482.
Clancy, C. E. and Rudy, Y. (1999). Linking a genetic defect to its cellular phenotype in a
cardiac arrhythmia. Nature, 400(6744):566–569.
Edmunds, D. E. and Triebel, H. (2008). Function spaces, entropy numbers, differential
operators, volume 120. Cambridge University Press.
Fang, K.-T., Li, R., and Sudjianto, A. (2005). Design and modeling for computer experiments.
CRC Press.
Ghosal, S. and van der Vaart, A. (2017). Fundamentals of nonparametric Bayesian inference,
volume 44. Cambridge University Press.
27
Gu, M., Palomo, J., and Berger, J. O. (2018a). RobustGaSP: Robust Gaussian stochastic
process emulation in R. arXiv preprint arXiv:1801.01874.
Gu, M. and Wang, L. (2017). An improved approach to bayesian computer model calibration
and prediction. arXiv preprint arXiv:1707.08215.
Gu, M., Wang, X., and Berger, J. O. (2018b). Robust gaussian stochastic process emulation.
The Annals of Statistics, In Press.
Higdon, D., Kennedy, M., Cavendish, J. C., Cafeo, J. A., and Ryne, R. D. (2004). Combining
field data and computer simulations for calibration and prediction. SIAM Journal on
Scientific Computing, 26(2):448–466.
Hoffman, A. J. and Wielandt, H. W. (2003). The variation of the spectrum of a normal
matrix. In Selected Papers Of Alan J Hoffman: With Commentary, pages 118–120. World
Scientific.
Joseph, V. R. and Melkote, S. N. (2009). Statistical adjustments to engineering models.
Journal of Quality Technology, 41(4):362.
Kennedy, M. C. and O’Hagan, A. (2001). Bayesian calibration of computer models. Journal
of the Royal Statistical Society: Series B (Statistical Methodology), 63(3):425–464.
Plumlee, M. (2017). Bayesian calibration of inexact computer models. Journal of the American Statistical Association, 112(519):1274–1285.
Plumlee, M., Joseph, V. R., and Yang, H. (2016). Calibrating functional parameters in
the ion channel models of cardiac cells. Journal of the American Statistical Association,
111(514):500–509.
Qian, P. Z. and Wu, C. F. J. (2008). Bayesian hierarchical modeling for integrating lowaccuracy and high-accuracy experiments. Technometrics, 50(2):192–204.
Rasmussen, C. E. and Williams, C. K. (2006). Gaussian processes for machine learning,
volume 1. MIT press Cambridge.
28
Santner, T. J., Williams, B. J., and Notz, W. I. (2013). The design and analysis of computer
experiments. Springer Science & Business Media.
Sheridan, M., Bloebaum, C., Kesavadas, T., Patra, A., and Winer, E. (2002). Visualization
and communication in risk management of landslides. WIT Transactions on Modelling
and Simulation, 31.
Stone, C. J. (1982). Optimal global rates of convergence for nonparametric regression. The
Annals of Statistics, pages 1040–1053.
Storlie, C. B., Lane, W. A., Ryan, E. M., Gattiker, J. R., and Higdon, D. M. (2015).
Calibration of computational models with categorical parameters and correlated outputs
via Bayesian smoothing spline ANOVA. Journal of the American Statistical Association,
110(509):68–82.
Tuo, R. (2017). Adjustments to computer models via projected kernel calibration. arXiv
preprint arXiv:1705.03422.
Tuo, R. and Wu, C. F. J. (2015). Efficient calibration for imperfect computer models. The
Annals of Statistics, 43(6):2331–2352.
Tuo, R. and Wu, C. F. J. (2016). A theoretical framework for calibration in computer
models: parametrization, estimation and convergence properties. SIAM/ASA Journal on
Uncertainty Quantification, 4(1):767–795.
van der Vaart, A. and Wellner, J. (1996). Weak Convergence and Empirical Processes: With
Applications to Statistics. Springer Science & Business Media.
van der Vaart, A. and Zanten, H. v. (2011). Information rates of nonparametric gaussian
process methods. Journal of Machine Learning Research, 12(Jun):2095–2119.
van der Vaart, A. W. (2000). Asymptotic statistics, volume 3. Cambridge university press.
van der Vaart, A. W. and van Zanten, J. H. (2008). Rates of contraction of posterior
distributions based on gaussian process priors. The Annals of Statistics, pages 1435–1463.
29
Wahba, G. (1990). Spline models for observational data. SIAM.
Wang, S., Chen, W., and Tsui, K.-L. (2009). Bayesian validation of computer models.
Technometrics, 51(4):439–451.
Wendland, H. (2004). Scattered data approximation, volume 17. Cambridge university press.
Wong, R. K., Storlie, C. B., and Lee, T. (2017). A frequentist approach to computer model
calibration. Journal of the Royal Statistical Society: Series B (Statistical Methodology),
79(2):635–648.
Xie, F., Jin, W., and Xu, Y. (2017).
A theoretical framework for Bayesian nonpara-
metric regression: Orthonormal random series and rates of contraction. arXiv preprint
arXiv:1712.05731.
Xiu, D. (2010). Numerical methods for stochastic computations: a spectral method approach.
Princeton university press.
30
Supplementary Material for “Bayesian Projected
Calibration of Computer Models”
A
Proof of theorem 1
We first present a classic result regarding convergence rate of the Matérn Gaussian process
regression from van der Vaart and Zanten (2011).
Theorem A.1. Suppose η is imposed the Matérn Gaussian process with roughness parameter
α, and η0 ∈ Cα (Ω) ∩ Hα (Ω), where α > p/2. Then there exists some constant C > 0, such
that
Z
E0
kη − η0 k2L2 (Ω) Π(dη | Dn )
≤ Cn−2α/(2α+p) .
(A.1)
Ω
The first assertion follows immediately from the Markov’s inequality:
E0 Π kη − η0 kL2 (Ω) > Mn n−α/(2α+p) | Dn
Z
1
2
kη − η0 kL2 (Ω) Π(dη | Dn )
≤ 2 −2α/(2α+p) E0
Mn n
Ω
C
≤ 2 → 0.
Mn
The posterior distribution of η can be expressed by
"Z n
# "Z n
#−1
Y pη (yi , xi )
Y pη (yi , xi )
Π(η ∈ U | Dn ) =
Π(dη)
Π(dη)
,
p
(y
,
x
)
0
i
i
U i=1 p0 (yi , xi )
i=1
where p0 (yi , xi ) = Ψσ (yi − η0 (xi )) is the density of the true distribution. To prove the second
assertion, we need the following result from Xie et al. (2017) to bound the denominator of
the preceeding display.
Lemma A.1. Assume the conditions of theorem 1 hold. For any D > 0, define the event
(Z n
)
Y pη (yi , xi )
1
Hn =
Π(dη) ≥ Π(kη − η0 kL∞ (Ω) < n ) exp − D + 2 n2n
.
p
(y
,
x
)
σ
0
i
i
i=1
2
c
Suppose (n )∞
n=1 is a sequence such that nn → ∞ and n → 0. Then P0 (Hn ) → 0.
Since α > p/2, we can some positive β such that β ∈ (max{α, p/2}, α). Define n =
n−β/(2β+p) . Since the Matérn Gaussian process assigns prior probability one to the space
31
Cβ (Ω) (see, for example, section 3.1 in van der Vaart and Zanten, 2011), then the Gaussian
process prior on η can be regarded as a mean-zero Gaussian random element in the Banach
space Cβ (Ω) equipped with the β-Hölder norm k · kCβ (Ω) . Therefore by the Borell’s inequality
(see, for example, Ghosal and van der Vaart, 2017) it holds that
1/2 !
Z
2
≤ 2e−2x .
Π kηkCβ (Ω) > 4x
kηk2Cβ (Ω) Π(dη)
(A.2)
for any positive x.
By lemma 15 in van der Vaart and Zanten (2011) there exists a constant M̃ > 0 such that
p/(2β+p)
kf kL∞ (Ω) ≤ M̃ kf kCβ (Ω)
2β/(2β+p)
kf kL2 (Ω)
for any function f ∈ Cβ (Ω). Let s > 0 be a constant
determined later. Then
(
kη − η0 kL2 (Ω) ≤ Mn n−α/(2α+p) ∩
kηkCβ (Ω)
√
≤ 4s nn
Z
kηk2Cβ (Ω) Π(dη)
1/2 )
n
o
p/(2β+p)
⊂ kη − η0 kL∞ (Ω) ≤ M̃ kη − η0 kCβ (Ω) Mn2β/(2β+p) n−(2αβ)/[(2α+p)(2β+p)]
(
Z
1/2 )
√
∩ kηkCβ (Ω) ≤ 4s nn
kηk2Cβ (Ω) Π(dη)
n
p/(2β+p) 2β/(2β+p) −2αβ/[(2α+p)(2β+p)] o
⊂ kη − η0 kL∞ (Ω) ≤ M̃ kηkCβ (Ω) + kη0 kCβ (Ω)
Mn
n
)
(
1/2
Z
√
2
kηkCβ (Ω) Π(dη)
∩ kηkCβ (Ω) ≤ 4s nn
n
⊂ kη − η0 kL∞ (Ω) ≤
2
2
M1 Mn2β/(2β+p) n−2αβ/[(2α+p)(2β+p)] np /[2(2β+p) ]
o
for some constant M1 > 0 depending on η0 only when n is sufficiently large. Noticing that
−α/(2α + p) < −β/(2β + p), then taking Mn = log n yields
n
o
2
2
kη − η0 kL∞ (Ω) ≤ M1 Mn2β/(2β+p) n−2αβ/[(2α+p)(2β+p)] np /[2(2β+p) ]
n
o
2β/(2β+p) −(2β 2 −p2 /2)/(2β+p)2
⊂ kη − η0 kL∞ (Ω) ≤ M1 (log n)
n
⊂ kη − η0 kL∞ (Ω) ≤ M
for some constant M > 0, where the fact β > p/2 is applied. Since by the first assertion
Π(kη −η0 kL2 (Ω) ≤ Mn n−α/(2α+p) | Dn ) = 1−oP0 (1), it suffices to show that E0 [Π(Un | Dn )] →
32
0, where Un is the event
(
Un =
√
Z
kηkCβ (Ω) > 4s nn
1/2 )
kηk2Cβ (Ω) Π(dη)
.
Let Hn be defined as in lemma A.1 with the constant D be determined later. Then P0 (Hnc ) →
0, and we directly compute by Fubini’s theorem
E0 [Π (Un | Dn )] ≤ E0 [1(Hn )Π (Un | Dn )] + P0 (Hnc )
#−1 "Z
#
"Z n
n
Y pη (yi , xi )
Y pη (yi , xi )
Π(dη)
Π(dη)
+ o(1)
= E0 1(Hn )
p (y , xi )
Un i=1 p0 (yi , xi )
i=1 0 i
" n
#
Z
Y pη (yi , xi )
exp[(D + 1/σ 2 )n2n ]
≤
E0
Π(dη) + o(1)
Π(kη − η0 kL∞ (Ω) < n ) Un
p
(y
,
x
)
0
i
i
i=1
exp[(D + 1/σ 2 )n2n ]Π(Un )
+ o(1).
≤
Π(kη − η0 kL∞ (Ω) < n )
By lemma 3 and lemma 4 in van der Vaart and Zanten (2011) we know that
Π(kη − η0 kL∞ (Ω) ≤ n ) ≥ exp −C−p/α
≥ exp −Cnpβ/[α(2β+p)] .
n
for some constant C > 0. Now take D = 1/(2σ 2 ), s = 1/σ, and we conclude
3
2
pβ/[α(2β+p)]
Π(Un ) + o(1)
E0 {Π(Un | Dn )} ≤ exp
n + Cn
2σ 2 n
3
2 2
2
(pβ/[α(2β+p)]
≤ 2 exp
n + Cn
− 2 nn + o(1) → 0,
2σ 2 n
σ
where the last inequality is due to (A.2) and the fact β < α is applied.
B
Proof of Lemma 1
Recall that θη∗ = arg minθ∈Θ kη(·) − y s (·, θ)k2L2 (Ω) . Since by condition A4 it is permitted to
interchange the differentiation with respect to θ and the integral, it follows that
Z
∂y s
∂
s
2
kη(·) − y (·, θ)kL2 (Ω)
= −2
η(x) − y s (x, θη∗ )
(x, θη∗ )dx.
0=
∂θ
∂θ
Ω
θ=θη∗
Now define the function F : F × Θ → Rq by
Z
∂y s
s
F(η, θ) = −2 [η(x) − y (x, θ)]
(x, θ)dx.
∂θ
Ω
33
It follows immediately that F(η, θη∗ ) = 0. The partial derivative of F with respect to θ is
given by
∂
F(η, θ) =
Fθ (η, θ) :=
∂θ
Z
Ω
∂2
s
2
[η(x) − y (x, θ)] dx,
∂θ∂θ T
and the partial Fréchet derivative of F with respect to η is a function Fη : F → Rq given by
Z
∂y s
[Fη (η, θ)](h) = −2 h(x)
(x, θ)dx,
∂θ
Ω
since F is linear with respect to η. Therefore by the implicit function theorem on Banach
space, there exists some > 0 such that over {η ∈ F : kη − η0 kL2 (Ω) < }, the functional
θη∗ : η 7→ arg minθ∈Θ kη(·) − y s (·, θ)k2L2 (Ω) is continuous, the Fréchet derivative θ̇η∗ : F → Rq
for θη∗ exists, and can be computed by
θ̇η∗ (h)
=−
−1
Fθ (η, θη∗ )
Fη (η, θη∗ )
(h) =
2Vη−1
Z
h(x)
Ω
∂y s
(x, θη∗ )dx.
∂θ
Therefore we obtain by the fundamental theorem of calculus and the mean-value theorem
that
θη∗
−
θ0∗
Z
1
d ∗
θη[u] du
0 du
Z 1
d
∗
θ̇η[u]
=
η[u] du
du
0
Z 1
Z
∂y s
−1
∗
(x, θη[u]
)dxdu
=2
Vη[u] [η(x) − η0 (x)]
∂θ
0
Ω
Z
s
−1 ∂y
∗
= 2 [η(x) − η0 (x)]Vη[u
(x, θη[u
0 ] )dx,
0]
∂θ
Ω
=
where η[u] = η0 + (η − η0 )u for 0 ≤ u ≤ 1 and u0 ∈ [0, 1]. By condition A3, we know that the
smallest eigenvalue λmin (Vη ) of Vη is strictly positive in an L2 -neighborhood of η0 , and we
can without loss of generality require that inf kη−η0 kL2 (Ω) ≤ λmin (Vη ) > 0. Hence we proceed
by condition A4 and Jensen’s inequality that
kθη∗
−
θ0∗ k
≤2
sup
kη−η0 kL2 (Ω) ≤
Vη−1
sup
(x,θ)∈Ω×Θ
−1
≤2
∂y s
(x, θ)
∂θ
inf
kη−η0 kL2 (Ω) ≤
λmin (Vη )
sup
(x,θ)∈Ω×Θ
= L(1)
η0 kη − η0 kL2 (Ω)
34
Z
|η(x) − η0 (x)|dx
Ω
∂y s
(x, θ)
∂θ
Z
2
[η(x) − η0 (x)] dx
Ω
1/2
(1)
for some constant Lη0 > 0 depending on η0 only.
We now analyze the property of Vη as a functional {η ∈ F : kη − η0 kL2 (Ω) < } →∈ Rq×q ,
η 7→ Vη . For a matrix A ∈ Rq×q , denote [A]ij to be the (i, j)-th element of A. Directly
compute
s
s
∂y s
∂y s
∗ ∂y
∗
∗ ∂y
∗
=2
(x, θη )
(x, θη ) −
(x, θ0 )
(x, θ0 ) dx
∂θk
∂θj
∂θk
Ω ∂θj
2 s
Z
∂ y
∂ 2ys
s
∗
∗
∗
[η(x) − y (x, θη )]
(x, θη ) −
(x, θ0 ) dx
−2
∂θj ∂θk
∂θj ∂θk
Ω
Z
∂ 2ys
s
∗
s
∗
∗
(x, θ0 ) dx
−2
[η(x) − η0 (x) + y (x, θ0 ) − y (x, θη )]
∂θj ∂θk
Ω
Z
[Vη ]jk − [V0 ]jk
:= 2V1 − 2V2 − 2V3 .
For V1 , by condition A4 we know that ∂y s /∂θ is Lipschitz continuous on Ω×Θ, and therefore
Z
∂y s
∂y s
∂y s
|V1 | ≤
(x, θη∗ )
(x, θη∗ ) −
(x, θ0∗ ) dx
∂θk
∂θk
Ω ∂θj
Z
s
s
∂y
∂y
∂y s
(x, θη∗ ) −
(x, θ0∗ )
(x, θ0∗ ) dx
+
∂θ
∂θ
∂θ
j
j
k
Ω
"
#
s
s
s
∂y s
∂y
∂y
∂y
∂y s
(x, θ)
≤ sup
(·, θη∗ ) −
(·, θ0∗ )
+
(·, θη∗ ) −
(·, θ0∗ )
∂θ
∂θ
∂θ
∂θ
∂θ
k
k
j
j
(x,θ)∈Ω×Θ
L1 (Ω)
L1 (Ω)
≤2
sup
(x,θ)∈Ω×Θ
∂y s
(x, θ)
∂θ
sup
(x,θ)∈Ω×Θ
∂ 2ys
(x, θ) kθη∗ − θ0∗ k
∂θ∂θ T
. kη − η0 kL2 (Ω) .
Condition A4 also implies that ∂ 2 y s /(∂θj ∂θk ) is Lipschitz continuous on Ω × Θ. Hence
Z
|V2 | . [|η(x) − η0 (x)| + |η0 (x) − y s (x, θη∗ )]kθη∗ − θ0∗ kdx
Ω
≤
L(1)
η0 kη
Z
1/2
Z
2
s
∗ 2
− η0 kL2 (Ω) 2 [η(x) − η0 (x)] dx + 2 [η0 (x) − y (x, θη )] dx
Ω
Ω
1/2
(1)
2
2
s
2
≤ Lη0 kη − η0 kL2 (Ω) 2 + 4kη0 kL2 (Ω) + 4 sup ky (·, θ)kL2 (Ω)
θ∈Θ
. L(1)
η0 kη − η0 kL2 (Ω) .
Now we consider V3 :
|V3 | ≤
sup
(x,θ)∈Ω×Θ
∂ 2ys
∂θj ∂θk
Z
[|η(x) − η0 (x)| + |y s (x, θ0 ) − y s (x, θη∗ )|]dx
Ω
35
≤
sup
(x,θ)∈Ω×Θ
≤
sup
(x,θ)∈Ω×Θ
∂ 2ys
∂θj ∂θk
"
∂ 2ys
∂θj ∂θk
"
kη − η0 kL2 (Ω) +
sup
(x,θ)∈Ω×Θ
1+
sup
(x,θ)∈Ω×Θ
∂y s
(x, θ) kθη∗ − θ0∗ kdx
∂θ
#
#
∂y s
(x, θ) L(1)
kη − η0 kL2 (Ω) .
η0
∂θ
We conclude that |[Vη ]jk − [V0 ]jk | ≤ Cη0 kη − η0 kL2 (Ω) for all j, k = 1, . . . , q for some constant
Cη0 > 0 depending on η0 only. By the fact that
q
X
|λj (A) − λj (B)| ≤ kA − Bk2F
j=1
holds for any positive definite matrices A, B ∈ Rq×q (see, for example, Hoffman and Wielandt,
2003), we obtain
|λmin (Vη ) − λmin (V0 )| ≤ kVη −
V0 k2F
=
q
q
X
X
|[Vη ]jk − [V0 ]jk |2 ≤ q 2 Cη20 kη − η0 k2L2 (Ω) .
j=1 k=1
We may also assume without loss of generality that is sufficiently small such that |λmin (Vη )−
λmin (V0 )| ≤ λmin (V0 )/2 whenever kη − η0 kL2 (Ω) ≤ , in which case it holds that kVη−1 k ≥
2kV0−1 k. Hence
Vη−1 − V0−1 = V0−1 (V0 − Vη )Vη−1
≤ V0−1 kkV0 − Vη kkVη−1
≤ 2 V0−1
2
kVη − V0 kF
≤ 2qCη0 V0−1 kη − η0 kL2 (Ω)
whenever kη − η0 kL2 (Ω) < . Hence
Z
∂y s
∗
∗
r(η, η0 ) = θη − θ0 − 2 [η(x) − η0 (x)]V0−1
(x, θ0∗ )dx
∂θ
Ω
Z
s
s
−1 ∂y
−1 ∂y
∗
∗
= 2 [η(x) − η0 (x)] Vη[u0 ]
(x, θη[u0 ] ) − V0
(x, θ0 ) dx
∂θ
∂θ
Ω
Z
s
−1
∗
−1 ∂y
= 2 [η(x) − η0 (x)] (Vη[u0 ] − V0 )
(x, θη[u0 ] ) dx
∂θ
Ω
s
Z
∂y s
−1 ∂y
∗
∗
+ 2 [η(x) − η0 (x)]V0
(x, θη[u0 ] ) −
(x, θ0 ) dx,
∂θ
∂θ
Ω
36
and hence,
"
Z
|η(x) − η0 (x)|
kr(η, η0 )k ≤ 2
Ω
−1
−1
Vη[u
0 ] − V0
∂y s
(x, θ)
∂θ
sup
(x,θ)∈Ω×Θ
#
dx
∂y s
∂y s
−1
∗
∗
+ 2 |η(x) − η0 (x)| kV0 k
(x, θη[u0 ] ) −
(x, θ0 ) dx
∂θ
∂θ
Ω
Z
. kη − η0 kL2 (Ω) q 2 Cη20 kV0−1 kkη − η0 kL2 (Ω) + kV0−1 kkη − η0 k2L2 (Ω) ,
(2)
(2)
implying that kr(η, η0 )k ≤ Lη0 kη − η0 k2L2 (Ω) for some constant Lη0 depending on η0 only.
The proof is thus completed.
C
Proof of Lemma 2
Before proceeding, we introduce the notion of covering number for a metric space (X, d).
The -covering number of (X, d) for > 0, is the smallest number of -balls (with respect to
the metric d) that are needed to cover X.
Since η is imposed the Matérn Gaussian process with roughness parameter α, it follows
that the concentration function
ϕη0 () =
1
kηk2H − log Π(kηkL∞ (Ω) < )
η∈H:kη−η0 kL∞ (Ω) ≤ 2
inf
satisfies ϕη0 () ≤ C−p/α for some constant C > 0 for sufficiently small > 0. Then by
theorem 2.1 in van der Vaart and van Zanten (2008), it holds that
Π(kη − η0 kL∞ (Ω) < n ) ≥ exp(−C 2 n2n ),
(C.1)
where n = n−α/(2α+p) . Pick β > 0 such that β ∈ (max{α, p/2}, α). Then we know that
the Matérn Gaussian process GP(0, Ψα ) assigns prior probability one to Cβ (Ω). Now set
Bn = n C1β (Ω) + mn H1Ψα (Ω), where
C1β (Ω) = f ∈ Cβ (Ω) : kf kCα (Ω) ≤ 1 ,
H1Ψα (Ω) = f ∈ HΨα (Ω) : kf kHΨα (Ω) ≤ 1 ,
mn is some sequence determined later, and HΨα (Ω) is the reproducing kernel Hilbert space
(abbreviated as RKHS) associated with the Matérn covariance function Ψα . Denote Φ to
be the cumulative distribution function of the standard normal distribution and set mn =
37
−2Φ−1 (exp[−(2C + 1/σ 2 )n2n ]). Since η ∼ GP(0, Ψα ) can be viewed as a Gaussian random
element in the Banach space Cβ (Ω) with the norm k · kCβ (Ω) , then by the Borell’s inequality
(van der Vaart and van Zanten, 2008) we have
Π(Bn ) ≥ Φ Φ−1 exp −Cn2n + mn
1
2
−1
−1
2
exp − 2C + 2 nn
exp −Cnn − 2Φ
=Φ Φ
σ
1
≥ Φ −Φ−1 exp − 2C + 2 n2n
σ
1
2
= 1 − exp − 2C + 2 nn .
σ
Hence
Π(η ∈
Bnc )
1
2
≤ exp − 2C + 2 nn .
σ
(C.2)
Now we prove the first inequality using (C.1) and (C.2). Let Hn be defined as in lemma A.1.
Denote Mn = log n. Then
E0 [Π(Bnc | Dn )] ≤ E0 [1(Hn )Π(Bnc | Dn )] + P0 (Hnc )
#−1 "Z
#
"Z n
n
Y
Y pη (yi , xi )
pη (yi , xi )
+ o(1)
Π(dη)
Π(dη)
= E0 1(Hn )
p0 (yi , xi )
p0 (yi , xi )
c
Bn
i=1
i=1
exp[(D + σ −2 )n2n ]
Π(Bnc ) + o(1)
Π(kη − η0 kL∞ (Ω) < n )
1
1
2
2
2
≤ exp D + 2 nn + Cnn − 2C + 2 nn + o(1)
σ
σ
≤ exp (D − C) n2n + o(1).
≤
Hence taking D = C/2 yields E0 [Π(Bnc | Dn )] → 0.
Finally we prove the second inequality involving the bracketing integral. Since HΨα (Ω)
is the RKHS of the Matérn covariance function with roughness parameter α, then HΨα (Ω)
coincides with the Sobolev space Hα+p/2 (Ω) (see, for example, corollary 1 of Tuo and Wu,
2016). The logarithm of the covering number of ρH1Ψα (Ω) is bounded by (Edmunds and
Triebel, 2008)
ρ 2p/(2α+p)
log N , ρH1Ψα (Ω), k · kL∞ (Ω) .
38
for sufficiently small > 0. The metric entropy for the α-Hölder space n C1α (Ω) is also known
in the literature (see, for example, van der Vaart and Wellner, 1996):
n p/β
log N , n C1β (Ω), k · kL∞ (Ω) .
.
Hence for sufficiently small > 0,
log N (, Bn , k · kL∞ (Ω) ) .
m 2p/(2α+p)
n
+
p/β
n
,
and it follows by simple algebra that
Z Mn n q
J[·] (Mn n , Bn , k · kL2 (Ω) ) ≤
log N , Bn , k · kL∞ (Ω) d
0
. mp/(2α+p)
(Mn n )2α/(2α+p) + p/2β
(Mn n )(2β−p)/(2β)
n
n
√
Mn2α/(2α+p) n2n + Mn(2β−p)/(2β) n
√
. Mn2α/(2α+p) n2n
for sufficiently large n.
D
Proof of Lemma 3
Before proceeding, we establish the following fact: if (Wn )∞
n=1 is a sequence of event such
that Π(Wn | Dn ) = oP0 (1), then
Z
Z
exp(`n (η) − `n (η0 ))Π(dη) = Π(Wn | Dn ) exp(`n (η) − `n (η0 ))Π(dη)
Wn
= oP0 (Dn ),
(D.1)
where
Z
Dn :=
exp(`n (η) − `n (η0 ))Π(dη).
Recall that the RKHS HΨα (Ω) of the Matérn Gaussian process with roughness parameter
α > p/2 coincides with the Sobolev space Hα+p/2 (Ω) (Wendland, 2004; Tuo and Wu, 2016),
and the RKHS norm k · kHΨα (Ω) is equivalent to the Sobolev norm k · kHα+p/2 (Ω) . Given
a realization η of the Matérn Gaussian process GP(0, Ψα ), define the following isometry
39
associated to η:
U : H0 =
(K
X
)
ak Ψ(·, tk ) : tk ∈ Ω, ak ∈ R, K ∈ N+
→ L2 (P0 ),
k=1
K
X
ak Ψ(·, tk ) 7→
k=1
K
X
ak η(tk ),
k=1
and extend U from H0 to H0 = HΨα (Ω) continuously. Then under the prior distribution Π,
for any h ∈ HΨα (Ω), U (h) ∼ N 0, khkHΨα (Ω) . Hence by lemma 17 in Castillo (2012), for
any measurable function T : C(Ω) → R, any g, h ∈ HΨα (Ω), and any ρ > 0,
EΠ [1{|U (g)| ≤ ρ}Ψ(η − h)]
1
2
= EΠ 1 U (g) + hg, hiHΨα (Ω) ≤ ρ T (η) exp U (−h) − khkHΨα (Ω) .
2
(D.2)
Let n = n−α/(2α+p) . Denote A1n = {kη − η0 kL2 (Ω) ≤ Mn n }, A2n = {kη − η0 kL∞ (Ω) ≤ M },
and take
2σ 2
∂y s
∂y s
(x, θ0∗ ), h(x) = √ tT V0−1
(x, θ0∗ ),
∂θ
∂θ
n
√
Cn = |U (g)| ≤ L nn kgkHΨ (Ω) ,
g(x) = 2σ 2 tT V0−1
where L is some constant to be specified later. Since U (g/kgkHΨα (Ω) ) follows the standar
normal distribution under the prior, it follows that for sufficiently large L,
√
g
L 2
c
> L nn ≤ 2 exp − nn .
Π(Cn ) = Π U
kgkHΨα (Ω)
2
Then by the proof of lemma 2, we know that Π(Cnc | Dn ) = oP0 (1) by taking a sufficiently
large L. Furthermore, we know that
hg, hiHΨα (Ω)
∂y s
4σ 4
(·, θ0∗ )
= √ tT V0−1
∂θ
n
4
q
2
HΨα (Ω)
X
4σ
∂y s
= √ kV0−1 tk2
sup
(·, θ)
∂θ
n
θ∈Θ
j
j=1
2
√
= o( nn ),
HΨα (Ω)
which implies that for sufficiently large n,
√
√
|U (g)| ≤ (L/2) nn kgkHΨα (Ω) ⊂ U (g) + hg, hiHΨα (Ω) ≤ L nn kgkHΨα (Ω)
√
⊂ |U (g)| ≤ 2L nn kgkHΨα (Ω) .
40
(D.3)
On the other hand,
khkL2 (Ω)
∂y s
2qσ 2
−1
≤ √ kV0 tk max sup
(·, θ)
j=1,··· ,q θ∈Θ ∂θj
n
= o(n ),
L2 (Ω)
implying that
A1n = kηt − η0 + hkL2 (Ω) ≤ Mn n
⊂ kηt − η0 kL2 (Ω) ≤ Mn n + khkL2 (Ω)
⊂ kηt − η0 kL2 (Ω) ≤ 2Mn n := Au1n (t)
(D.4)
for sufficiently large n, where the fact n−1/2 ≤ n is applied. Similarly, for sufficiently large
n it holds that
A1n ⊃ {kηt − η0 kL2 (Ω) ≤ Mn n /2} := Al1n (t).
(D.5)
R
Hence we can bound An exp(`n (ηt ) − `n (η0 ))Π(dη) from above as follows:
Z
exp(`n (ηt ) − `n (η0 ))Π(dη)
An
Z
=
1(Cn ) exp(`n (ηt ) − `n (η0 ))Π(dη) + oP0 (Dn )
An
Z
√
≤ 1 |U (g)| ≤ L nn kgkHΨα (Ω) 1(Au1n (t)) exp(`n (ηt ) − `n (η0 ))Π(dη) + oP0 (Dn ).
The first term in the preceeding display can be further upper bounded using the change of
measure formula (D.2), (D.3), and (D.4):
Z
√
1 |U (g)| ≤ L nn kgkHΨα (Ω) 1(Au1n (t)) exp(`n (ηt ) − `n (η0 ))Π(dη)
Z
√
≤ 1 U (g) + hg, hiHΨα (Ω) ≤ L nn kgkHΨα (Ω) 1(kη − η0 kL2 (Ω) ≤ 2Mn n )
#
"
2
2σ 4 T −1 ∂y s
Π(dη)
× exp(`n (η) − `n (η0 )) exp U (−h) −
t V0
(·, θ0∗ )
n
∂θ
HΨα (Ω)
Z
√
≤ 1 |U (g)| ≤ 2L nn kgkHΨα (Ω) 1(kη − η0 kL2 (Ω) ≤ 2Mn n )
g
× exp(`n (η) − `n (η0 )) exp U − √
Π(dη)
n
Z
≤
exp(`n (η) − `n (η0 )) exp 2Ln kgkHΨα (Ω) Π(dη)
{kη−η0 kL2 (Ω) ≤2Mn n }
41
Z
≤ [1 + o(1)]
exp(`n (η) − `n (η0 ))Π(dη).
Therefore we conclude that
Z
Z
exp(`n (ηt ) − `n (η0 ))Π(dη) ≤ [1 + o(1)] exp(`n (η) − `n (η0 ))Π(dη) + oP0 (Dn )
An
Z
= [1 + oP0 (1)] exp(`n (η) − `n (η0 ))Π(dη).
(D.6)
R
On the other hand, we need to bound An exp(`n (ηt ) − `n (η0 ))Π(dη) from below:
Z
exp(`n (ηt ) − `n (η0 ))Π(dη)
An
Z
≥
1(Cn ) exp(`n (ηt ) − `n (η0 ))Π(dη)
An
Z
=
1(Cn ) exp(`n (ηt ) − `n (η0 ))Π(dη)
A1n
Z
−
1(Ac2n ∪ Bnc )1(Cn ) exp(`n (ηt ) − `n (η0 ))Π(dη)
Z A1n
=
1(Cn ) exp(`n (ηt ) − `n (η0 ))Π(dη) − oP0 (Dn )
A1n
Z
√
≥ 1 |U (g)| ≤ L nn kgkHΨ (Ω) 1(Al1n (t)) exp(`n (ηt ) − `n (η0 ))Π(dη) − oP0 (Dn ).
We lower bound the first term in the preceeding display using (D.2), (D.3), and (D.5):
Z
√
1 |U (g)| ≤ L nn kgkHΨα (Ω) 1(Al1n (t)) exp(`n (ηt ) − `n (η0 ))Π(dη)
Z
√
= 1 U (g) + hg, hiHΨα (Ω) ≤ L nn kgkHΨα (Ω) 1 kη − η0 kL2 (Ω) ≤ Mn n /2
"
#
2
2σ 2 T −1 ∂y s
g
× exp(`n (η) − `n (η0 )) exp U − √
−
t V0
(·, θ0∗ )
Π(dη)
n
∂θ
n
HΨα (Ω)
Z
√
≥ 1 |U (g)| ≤ (L/2) nn kgkHΨα (Ω) 1 kη − η0 kL2 (Ω) ≤ Mn n /2
1
× exp(`n (η) − `n (η0 )) exp − √ |U (g)| [1 − o(1)]Π(dη)
n
Z
√
≥ [1 − o(1)] 1 |U (g)| ≤ (L/2) nn kgkHΨα (Ω) 1 kη − η0 kL2 (Ω) ≤ Mn n /2
× exp(`n (η) − `n (η0 ))Π(dη).
Since Π(kη − η0 kL2 (Ω) > Mn n /2 | Dn ) = oP0 (1), and for sufficiently large L, Π(|U (g)| >
42
√
(L/2) nn kgkHΨ (Ω) | Dn ) = oP0 (1), the last display can be further computed
Z
√
1 |U (g)| ≤ (L/2) nn kgkHΨα (Ω) 1 kη − η0 kL2 (Ω) ≤ Mn n /2 exp(`n (η) − `n (η0 ))Π(dη)
Z
Z
≥ exp(`n (η) − `n (η0 ))Π(dη) − n
o exp(`n (η) − `n (η0 ))Π(dη)
√
|U (g)|>(L/2) nn kgkHΨ
α
(Ω)
Z
−
{kη−η0 kL2 (Ω) >Mn n /2}
exp(`n (η) − `n (η0 ))Π(dη)
Z
=
exp(`n (η) − `n (η0 ))Π(dη) − oP0 (Dn ).
Hence we conclude that
Z
Z
exp(`n (η) − `n (η0 ))Π(dη) ≥ [1 − o(1)] exp(`n (η) − `n (η0 ))Π(dη) − oP0 (Dn )
An
Z
= [1 − oP0 (1)] exp(`n (η) − `n (η0 ))Π(dη).
(D.7)
The proof is completed by combining (D.6) and (D.7).
43
| 10math.ST
|
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS ASSOCIATED WITH
DISCRETE SUBGROUPS OF SEMISIMPLE LIE GROUPS
arXiv:1607.07830v1 [math.GR] 26 Jul 2016
ADRIEN BOYER
Abstract. We prove that the Harish-Chandra–Schwartz space associated with a discrete subgroup of a semisimple Lie group is a dense subalgebra of the reduced C ∗ algebra of the discrete subgroup. Then, we prove that for the reduced C ∗ -norm is controlled by the norm of the Harish-Chandra–Schwartz space. This inequality is weaker
than property RD and holds for any discrete group acting isometrically, properly on a
Riemannian symmetric space.
1. Introduction
A length function on a locally compact group G is a measurable function L : G → R∗+
satisfying L(e) = 0 where e is the neutral element of the group, L(g −1 ) = L(g) and
L(gh) ≤ L(g) + L(h) for all g, h in G. Define the Sobolev norm of parameter d > 0 as
1
Z
2
2
2d
|f (g)| (1 + L(g)) dg
kf kH d (G) :=
L
G
where f is in Cc (G) the space of complex-valued continuous functions on G with compact
support and dg a (left) Haar measure on G. The Sobolev space associated with G denoted
by HLd (G) is the completion of Cc (G) with respect to k · kH d (G) .
L
A locally compact group G has property RD (rapid decay) with respect to L if there
exists d > 0 such that HLd (G) convolves L2 (G) in L2 (G) in the following manner: there
exists C > 0 such that for all compactly supported functions f in Cc (G), we have
(1)
kλG (f )kop ≤ Ckf kH d (G) ,
L
where λG denotes the left regular representation of G, namely λG (f )v = f ∗ v where f in
Cc (G), where v ∈ L2 (G) and k · kop denotes the operator norm. Notice that kλG (f )kop
is nothing but kf kCr∗ (G) i.e. the norm of f in the reduced C ∗ -algebra of G denoted by
Cr∗ (G).
For an introduction to property RD we refer to the recent and exhaustive survey of
Chatterji [8] and to the survey of Sapir [26] for different methods of proving or disproving property RD.
The semisimple Lie groups satisfy property RD and this fact could be traced back to the
work of Herz [17] (of course he did not use this terminology) and it is essentially based
2010 Mathematics Subject Classification. Primary 46H15, 43A90, 22E40; Secondary 22D20, 22D25,
46L80.
Key words and phrases. Harish-Chandra–Schwartz’s spaces, semisimple Lie groups, Harish-Chandra
functions, Furstenberg boundary, property RD, K-theory, Baum-Connes conjecture.
This work is supported by ERC Grant 306706.
1
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
2
on the estimates of the Harish-Chandra function [5]. This suggests that the essence of
property RD is in harmonic analysis. For a characterization of Lie groups having property RD we refer to [11].
As for discrete groups, it has been first established for the free groups by Haagerup in
[16]. The terminology Property RD has been introduced and studied as such by Jolissaint
in [18], who notably established it for groups of polynomial growth and some hyperbolic
gloups although the general case of hyperbolic groups is due to de la Harpe [13]. In [19],
Jolissaint showed in particular that the space of rapidly decreasing functions associated
with a discrete group Γ endowed with a length function L, namely the space ∩d>0 HLd (Γ),
is a dense subalgebra of Cr∗ (Γ) and the inclusion induces isomorphisms in topological
K-theory.
The first example of higher rank discrete groups having property RD is due to Ramagge, Robertson and Steger [24]. Then, Lafforgue, inspired by the methods of [24],
proved that cocompact lattices in SL3 (R) and SL3 (C) satisfy property RD. Note that
Lafforgue proved in [22] that for d large enough, the Sobolev space HLd (Γ) associated
with any discrete group Γ is a dense Banach subalgebra of Cr∗ (Γ) and the inclusion
induces isomorphisms in topological K-theory. For examples of other groups satisfying
property RD we refer to [1], [2], [9], [10], [14].
This property also plays a crucial role in the proof of the Novikov conjecture for hyperbolic groups by Connes and Moscovici in [12], and Lafforgue used it to prove BaumConnes conjecture for some groups having property (T). The major problem related to
property RD is the Valette conjecture:
Conjecture 1.1. (The Valette conjecture)
Property RD holds for any discrete group acting isometrically, properly and cocompactly
either on a Riemannian symmetric space or on an affine building.
This problem in this generality is considered very difficult. In this paper we restrict
our attention to discrete subgroups of semisimple Lie groups. We investigate another
inequality involving convolutor norm of finitely supported functions and the decay of
the Harish-Chandra function.
Let G be a (non compact) connected semisimple real Lie group with finite center. Let
g be its Lie algebra. Let θ be a Cartan involution. Define the bilinear form denoted:
for
p
all X, Y ∈ g, hX, Y i = −B(X, θ(Y )) where B is the Killing form. Set |X| = hX, Xi.
Write g = l ⊕ p the eigenvector space decomposition associated to θ (l for the eigenvalue
1). Let K be the compact subgroup defined as the connected subgroup whose Lie algebra
l is the set of fixed points of θ. Fix a ⊂ p a maximal abelian subalgebra of p. Consider
the root system Σ associated to a and let Σ+ be the set of positive roots, and define the
corresponding positive Weyl chamber as
a+ := H ∈ a such that α(H) > 0, ∀α ∈ Σ+ .
Let A+ = exp(a+ ) be the closure of exp(a+ ). Consider the corresponding Cartan
decomposition KA+ K, with the Cartan projection H : g ∈ G 7→ H(g) ∈ a+ . From now
on let L be the length function
L(g) = L(keH(g) k′ ) := |H(g)|
where g = keH(g) k′ with eH(g) ∈ A+ and k, k′ ∈ K. Notice that L is K bi-invariant.
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
3
Let (G/P, ν) be the Furstenberg boundary where P denotes a minimal parabolic
subgroup of G and ν denotes the unique K-invariant probability measure which is Gquasi-invariant. Then consider the Harish-Chandra function
1
Z
dg∗ ν 2
(b)dν(b) ∈ R+ .
Ξ : g ∈ G 7→
dν
G/P
Define now the Harish-Chandra–Schwartz norm associated with a subgroup H of G as
kf kS d (H) := sup
L
g∈H
|f (g)|(1 + L(g))d
,
Ξ(g)
where f is in Cc (H). The Harish-Chandra–Schwartz space associated with H denoted
by SLd (H) is defined as the completion of Cc (H) with respect to the norm k · kS d (H) .
L
It is a well known fact that for G there exist C, d > 0 so that for all f in Cc (G) the
following inequality holds:
kλG (f )kop ≤ Ckf kS d (G) ,
(2)
L
and we refer to [23, Chapitre 4] for proofs. The aim of this note is to give a proof of the
following result.
Theorem 1.2. Let Γ be a discrete group of a connected real semisimple Lie group. Then
there exist d > 0 and C > 0 such that
(1) SLd (Γ) is a Banach convolution algebra.
(2) kλΓ (f )kop ≤ Ckf kS d (Γ) for all f ∈ SLd (Γ).
L
In particular
SLd (Γ)
is a dense subalgebra of Cr∗ (Γ).
Note that this algebra has already been studied in the context of hyperbolic groups
and rank one semisimple Lie groups in [7].
Property RD and Inequality (2). One of the fundamental properties of the HarishChandra function is the following integrability condition
Z
Ξ2 (g)
dg < ∞,
2d
G (1 + L(g))
for d > 0 large enough depending on G, see Subsection 2.2 for more details. Moreover,
it is still true that for any discrete group Γ in the ambient semisimple Lie group G we
have for all d > 0 large enough
X
Ξ2 (γ)
< ∞,
(1 + L(γ))2d
Γ
see [6, Proposition 3.1] for further details. In the context of semisimple Lie groups it is
clear that there exist d, d′ and C > 0 such that kf kH d ≤ Ckf kS d′ for either f in Cc (G)
L
L
or in Cc (Γ) of a discrete subgroup Γ of G. Hence Inequality (1), namely Property RD,
implies Inequality (2) for G and its discrete subgroups. The converse is wrong since all
non-uniform lattices satisfy Inequality (2) whereas they cannot satisfy property RD.
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
4
The Baum-Connes conjecture and the work of V. Lafforgue. We freely rely on
[28] where the reader could consult for further details about property RD, its connection
with K-theory and the Baum-Connes conjecture. We refer also to the Bourbaki seminar
of Skandalis [27] for a summary of the work of Lafforgue.
Let Γ be a countable group. The Baum-Connes conjecture identifies two objects
associated with Γ, one analytical and one geometrical/topological. The right-hand side
of the conjecture, or analytical side, involves the K-theory of Cr∗ (Γ). The K-theory used
here denoted by Ki (Cr∗ (Γ)) for i = 0, 1; is the usual topological K-theory for Banach
algebras. The left-hand side of the conjecture, or geometrical/topological side, is the Γequivariant K-homology with Γ- compact supports of the classifying space EΓ for proper
actions of Γ.
The link between both sides of the conjecture is provided by the analytic assembly
map, or index map or Baum-Connes map which is a well-defined group homomorphism:
µΓi : RKiΓ (EΓ) → K i (Cr∗ (Γ)).
(3)
The definition of the assembly map can be traced back to [20] and to [3].
Conjecture 1.3. (The Baum-Connes conjecture)
For i = 0, 1 the assembly map µΓi is an isomorphism.
In the important paper [23] of Lafforgue, containing breakthrough ideas about the
Baum-Connes conjecture, a Banach algebra AΓ is an unconditional completion of the
group algebra CΓ if it contains CΓ as a dense subalgebra and if, for f1 , f2 in CΓ such
that |f1 (γ)| ≤ |f2 (γ)| for all γ in Γ, we have kf1 kAΓ ≤ kf2 kAΓ . If AΓ is an unconditional
completion one may define an assembly map
µAΓ
: RKiΓ (EΓ) → K i (AΓ).
i
(4)
In the proof of Lafforgue of the Baum-Connes conjecture for discrete groups having
property (T), one of the fundamental step is the following:
Theorem 1.4. (V. Lafforgue)
If Γ is a group acting properly isometrically either on a simply connected complete Riemannian manifold with non-positive curvature, bounded from below, or on a Euclidean
building then for i = 0, 1 the map µAΓ
: RKiΓ (EΓ) → K i (AΓ) is an isomorphism.
i
Since the Harish-Chandra–Schwartz algebras are unconditional completions, we obtain
as an immediate corollary:
Corollary 1.5. If Γ is a discrete subgroup of G then for i = 0, 1 there exists d > 0 such
S d (Γ)
that the map µi L
: RKiΓ (EΓ) → K i (SLd (Γ)) is an isomorphism.
After having observed that the assembly maps µΓi and µAΓ
are compatible, it remains
i
to determine when one can find an embedding AΓ ֒→ Cr∗ (Γ) that induces an isomorphism of K-theory of Ki (AΓ) and Ki (Cr∗ (Γ)). It turns out that if a discrete group
Γ satisfies property RD the inclusion of the unconditionnal algebra HLd (Γ) for d large
enough induces isomorphism in topological K-theory and enables Lafforgue to prove the
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
5
Baum-Connes conjecture for discrete group having property RD, as the cocompact lattices in rank one semisimple Lie group and in SL3 (k) where k is a local field.
Unfortunately, the Harish-Chandra–Schwartz algebra is not stable by holomorphic functional calculus and does not automatically induce isomorphism in topological K-theory.
Consider an element γ in Γ so that lim L(γ n )/n → 0. It is clear that the unit Dirac mass
centered at γ does not have the same spectral radius in Cr∗ (Γ) and in SLd (Γ).
Nevertheless, it might be true that the embedding SLd (Γ) ֒→ Cr∗ (Γ), for d large enough,
induces isomorphism of K-theory of Ki (SLd (Γ)) and Ki (Cr∗ (Γ)). It is probably worth noting that in some cases it is true (e.g. uniform lattices in SL(3, R)), thanks to Lafforgue’s
work, and this is equivalent to the Baum-Connes conjecture.
Structure of the paper. In Section 2 we recall some standard facts about the quasiregular representation, the Harish-Chandra function and we state a lemma due to
Shalom. In Section 3 we prove several elementary inequalities. Finally in Section 4
we give the proof of Theorem 1.2.
Acknowledgements. I am grateful to Nigel Higson who suggested that I investigate the
structure of convolution algebra of Harish-Chandra–Schwartz’s spaces of discrete groups
and for helpful discussions concerning the work of Vincent Lafforgue and the BaumConnes conjecture. I would like to thank Vincent Lafforgue for fruitful discussions and
for pointing out that these algebras are not stable by holomorphic functional calculus.
Finally, I thank Vladimir Finkelshtein for helpful comments on this paper.
2. Preliminaries: Quasi-regular representation, the Harish-Chandra
function and Shalom’s lemma
2.1. Quasi-regular representation. A connected semisimple real Lie group with finite
center can be written G = KP , where K is a compact connected subgroup and P is a
minimal parabolic subgroup. We denote by ∆P the right-modular function of P . The
map ∆P extends to ∆ : G → R∗+ by ∆(g) = ∆(kp) := ∆P (p). It is well defined because
K ∩ P is compact (observe that ∆P |K∩P = 1). The quotient G/P , called the Furstenberg
boundary, carries a unique quasi-invariant probability measure ν which is K-invariant,
such that the Radon-Nikodym derivative at (g, b) ∈ G × G/P denoted by
c(g, b) =
∆(gb)
dg∗ ν
(b) =
,
dν
∆(b)
∗
where g∗ ν(A) = ν(g −1 A) (notice that for all g ∈ G, the function b ∈ G/P 7→ ∆(gb)
∆(b) ∈ R+
is well defined). We refer to [4, Appendix B, Lemma B.1.3, p. 344-345] for more details.
Let L2 (G/P, ν) the Hilbert space endowed with its inner product denoted by h·, ·i and
consider the quasi-regular representation πν : G → U (L2 (G/P, ν)) associated to P ,
defined by
(5)
1
(πν (g)ξ)(b) = c(g, b) 2 ξ(g−1 b).
Denote by dk the Haar measure on K, and under the identification G/P = K/(K ∩ P ),
denote by d[k] the measure ν on G/P .
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
6
Let Γ be a discrete subgroup of a semisimple Lie group G. In this note we consider the
following unitary representation of Γ
(6)
πν : γ ∈ Γ 7→ πν (γ) ∈ U (L2 (G/P, ν)),
which is nothing but the restriction of the quasi-regular representation πν of G to Γ.
Observe that if L2+ (G/P ) denotes the cone of positive functions observe that πν (g)L2+ (G/P ) ⊂
L2+ (G/P ) for all g in G.
2.2. The Harish-Chandra function. The well-known Harish-Chandra function can
also be defined as
(7)
Ξ : g ∈ G 7→ πν (g)1G/P , 1G/P ∈ R+ ,
where 1G/P denotes the characteristic function of the space G/P .
Observe that Ξ is bi-K-invariant and so Ξ(g) = Ξ(eH(g) ) and Ξ(g) = Ξ(g−1 ) for all
g ∈ G.
Let r be the number of indivisible positive roots in a. We know that there exists C > 0
such that for all H ∈ a where eH ∈ A+ we have
r
Ξ(eH ) ≤ Ce−ρ(H) 1 + L(eH ) ,
P
with ρ = 12 α∈Σ+ nα α ∈ a+ , see [15, Chap 4, Theorem 4.6.4, p.161]. Hence for 2d >
dim(a) + 2r, we set
Z
Ξ2 (eH )
(8)
Cd :=
J(H)dH < ∞,
H 2d
a+ (1 + L(e ))
where J denotes the density that occurs in the disintegration of the Haar measure on
G according to the Cartan decomposition. More precisely the Haar measure dg can be
written dg = dkJ(H)dHdk, where dk is the normalized Haar measure on K, dH the
Lebesgue measure on a+ , and
!n
Y eα(H) − e−α(H) α
,
J(H) =
2
+
α∈Σ
where nα denotes the dimension of the root space associated to α. See [21, Chap.V,
section 5, Proposition 5.28, p.141-142], [15, Chap. 2, §2.2, p.65] and [15, Chap 2, Proposition 2.4.6, p.73] for more details.
2.3. Shalom’s Lemma. The following lemma due to Shalom [26, Lemma 2.3] will be a
tool to study convolution operator norm via the Furstenberg boundary in the semisimple
Lie groups.
Lemma 2.1. Let π : G → U (H) be a unitary representation of a locally compact group.
Assume that there exists a non-zero positive vector ξ in the following sense: hπ(g)ξ, ξi ≥
0 for all g ∈ G. Then for any bounded positive measure µ on G we have kλG (µ)k ≤
kπ(µ)k.
It is obvious that 1G/P is a positive vector for πν .
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
7
3. Elementary inequalities
A function f : G → C is called L-radial if f (g) = f (g′ ) whenever g and g′ satisfy
L(g) = L(g ′ ). Note that the space of L-radial compactly supported functions is identified
by definition of L to the space of compactly supported bi-K-invariant functions denoted
by Cc (K\G/K).
The following proposition is based on standard facts in representation theory. Nevertheless, we give a proof for the convenience of the reader. In the following, we denote by
h·, ·iL2 (G) the scalar product of L2 (G) with respect to the Haar measure.
Proposition 3.1. For all ξ, η in L2 (G/P, ν) and f in Cc (K\G/K) we have
Z
G
f (g)hπν (g)ξ, ηidg = hξ, 1G/P ih1G/P , ηihf, ΞiL2 (G) .
In particular for all d satisfying 2d > dim(a) + 2r we have for all ξ, η in L2 (G/P, ν) and
for all L-radial functions f in HLd (G)
Z
G
f (g)hπν (g)ξ, ηidg ≤ Cd kf kH d kξk1 kηk1 .
L
Proof. Let ξ in L2 (G/P, ν) and observe that the function ξ K : x ∈ G/P 7→
C is constant. By integration we compute its value and we have
R
K
ξ(k−1 x)dk ∈
ξ K = hξ, 1G/P i1G/P .
Thus, for all functions f ∈ Cc (K\G/K) and for all vectors ξ, η ∈ L2 (G/P ) we have:
Z
G
Z Z
Z
f (keH k′ )hπν (keH k′ )ξ, ηidkdk′ J(H)dH
f (g)hπν (g)ξ, ηidg =
K a+ K
Z Z Z
f (eH )hπν (eH )πν (k′ )ξ, πν (k−1 )ηidkdk′ J(H)dH
=
K a+ K
Z
Z
Z
πν (k−1 )ηdk iJ(H)dH
πν (k′ )ξdk ,
f (eH )hπν (eH )
=
+
K
K
Za
f (eH )hπν (eH )ξ K , η K iJ(H)dH
=
a+
Z
= hξ, 1G/P ihη, 1G/P i
f (eH )hπν (eH )1G/P , 1G/P iJ(H)dH
a+
= hξ, 1G/P ih1G/P , ηihf, ΞiL2 (G) .
Then for all d satisfying 2d > dim(a) + 2r we have by taking the absolute value of the
above expression
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
Z
G
f (g)hπν (g)ξ, ηidg = |hξ, 1G/P i||h1G/P , ηi|
Ξ2 (eH )
J(H)dH
H 2d
a+ (1 + L(e ))
= Cd kf kH d kξk1 kηk1 ,
≤ kξk1 kηk1
Z
Z
1 Z
8
f (eH )Ξ(eH )J(H)dH
a+
2
|f |2 (eH )(1 + L(eH ))2d J(H)dH
a+
1
2
L
where Cd is the constant defined in (8) and where the last inequality follows from CauchySchwarz inequality. The proof is complete.
3.1. Stability. Let Γ be a discrete subgroup of G. A positive function g ∈ G 7→ f (g) ∈
R+ is stable on G relative to Γ if there exists a relatively compact neighborhood U of
the neutral element e ∈ G and C > 0 such that
f (γ) ≤ Cf (g)
for all γ ∈ Γ and for all g in γU .
The next lemma gives examples of interesting stable functions.
Lemma 3.1. Let Γ be a discrete group in a semisimple Lie group G. Let d be a real
number and consider the function on G defined as
g ∈ G 7→
hπν (g)1G/P , ξi
∈ R+ ,
(1 + L(g))d
with ξ ∈ L1+ (G/P, ν) a positive function. Then this function is stable on G relative to Γ.
Proof. Observe that the triangle inequality implies for all γ ∈ Γ and x ∈ G
1 + L(x)
1
≤
.
1 + L(γx)
(1 + L(γ))
Then combine the above observation with [6, Lemma 3.8] to complete the proof.
The notion of stability introduced above enables us to discretize inequalities given by
integrals on the ambient group to sums on discrete subgroups. More precisely we have
the following proposition.
Proposition 3.2. Let f be a positive L-radial function satisfying hf, ΞiL2 (G) < ∞ and
assume that f is stable relative to Γ. Assume also that ξ is a positive vector in L1+ (G/P ).
There exists a constant C > 0 such that
X
f (γ)hπν (γ)1G/P , ξi ≤ Chf, ΞiL2 (G) kξk1 .
γ
Proof. Since g 7→ hπν (g)1G/P , ξi is stable by Proposition 3.1, there exists U a neighbourhood of the identity e, small enough so that γU ∩ γ ′ U = ∅ for all γ 6= γ ′ ∈ Γ and a
constant C > 0 such that for all γ in Γ and for all x in U we have
hπν (γ)1G/P , ξi ≤ Chπν (γx)1G/P , ξi and f (γ) ≤ Cf (γx).
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
9
It follows by integration that there exists C > 0 such that
Z
C
f (γ)hπν (γ)1G/P , ξi ≤
f (γx)hπν (γx)1G/P , ξidx,
Vol(U ) U
where dx is the Haar measure. Hence, we have
Z
X
C X
f (γx)hπν (γx)1G/P , ξidx
f (γ)hπν (γ)1G/P , ξi ≤
Vol(U )
γ∈Γ U
γ∈Γ
Z
C X
f (g)hπν (g)1G/P , ξidg
=
Vol(U )
γ∈Γ γU
Z
C
≤
f (g)hπ(g)1G/P , ξidg
Vol(U ) G
C
hf, ΞiL2 (G) kξk1 ,
≤
Vol(U )
where the last inequality follows from Proposition 3.1.
3.2. Inequalities. We start by stating and proving an elementary lemma using CauchySchwarz inequality.
Lemma 3.2. For all ξ, η ∈ L2 (G/P, ν) we have:
1
1
2
2
2
−1
2
|hπν (g)ξ, ηi| ≤ hπν (g)1G/P , |η| i
hπν (g )1G/P , |ξ| i .
Proof. Since πν preserves the cone of positive functions, we may assume that ξ and η
are in L2+ (G/P ). Then the Cauchy-Schwarz inequality implies
!2
Z
1
hπν (g)ξ, ηi2 =
ξ(g−1 x)c(g, x) 2 η(x)dν(x)
G/P
≤
Z
G/P
Z
1
ξ(g−1 x)2 c(g, x) 2 dν(x)
G/P
2
πν (g)1G/P , η
= πν (g)ξ , 1G/P
= πν (g
−1
)1G/P , ξ
2
2
1
c(g, x) 2 η(x)2 dν(x)
πν (g)1G/P , η 2 .
4. Proofs
Before proceeding to the proof of Theorem 1.2, let us define for all d > 0 the function
φd as
(9)
φd : g ∈ G 7→
Ξ(g)
∈ R+ .
(1 + L(g))d
Observe that for all d > 0, the function φd is a bi-K-invariant function and observe that
whenever 2d > dim(a) + 2r, we have
(10)
hφ2d , ΞiL2 (G) = Cd < +∞.
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
10
Observe also that Proposition 3.1 implies that φd is stable on G relative to any discrete
subgroup Γ of G. This function will play a major role in the proof.
Proof of Item (1) of Theorem 1.2. Pick d > 0 such that 2d > dim(a) + 2r and let f1 , f2
be in SL2d (Γ). For all g ∈ Γ we have
X
|f1 ∗ f2 |(g) =
f1 (γ)f2 (γ −1 g)
γ∈Γ
≤ kf1 kS 2d (Γ) kf2 kS 2d (Γ)
L
L
X
γ∈Γ
Ξ(γ)
Ξ(γ −1 g)
.
2d
(1 + L(γ)) (1 + L(γ −1 g))2d
Write now
X
Ξ(γ −1 g)
Ξ(γ)
=
(1 + L(γ))2d (1 + L(γ −1 g))2d
γ∈Γ
X
n
γ|L(γ)≤ L(g)
2
Ξ(γ −1 g)
Ξ(γ)
+
o (1 + L(γ))2d (1 + L(γ −1 g))2d
n
γ|L(γ)> L(g)
2
−1 g) ≥
Observe that L(γ) ≤ L(g)
2 ⇒ L(γ
We have for the first term:
X
Ξ(γ −1 g)
Ξ(γ)
o
n
γ|L(γ)≤ L(g)
2
X
Ξ(γ −1 g)
Ξ(γ)
.
o (1 + L(γ))2d (1 + L(γ −1 g))2d
L(g)
2 .
(1 + L(γ))2d (1 + L(γ −1 g))2d
≤
≤
22d
(1 + L(g))2d
X
n
γ|L(γ)≤ L(g)
2
Ξ(γ)Ξ(γ −1 g)
o (1 + L(γ))2d
X Ξ(γ)Ξ(γ −1 g)
22d
(1 + L(g))2d
(1 + L(γ))2d
γ∈Γ
=
with ξ := πν (g)1G/P ∈ L2 (G/P, ν).
We have for the second term:
X
Ξ(γ −1 g)
Ξ(γ)
o
n
γ|L(γ)> L(g)
2
(1 + L(γ))2d (1 + L(γ −1 g))2d
≤
≤
22d
(1 + L(g))2d
22d
(1 + L(g))2d
X
φ2d (γ)hπν (γ)1G/P , ξi,
γ∈Γ
X
n
γ|L(γ)> L(g)
2
Ξ(γ)Ξ(γ −1 g)
o (1 + L(γ −1 g))2d
X Ξ(γ)Ξ(γ −1 g)
22d
(1 + L(g))2d
(1 + L(γ −1 g))2d
γ∈Γ
=
=
22d
(1 + L(g))2d
X Ξ(gγ −1 )Ξ(γ)
γ∈Γ
(1 + L(γ))2d
X
22d
φ2d (γ)hπν (γ)1G/P , ηi
2d
(1 + L(g))
γ∈Γ
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
11
where the first equality comes from the change of variable γ ′ = γ −1 g and where η :=
πν (g−1 )1G/P ∈ L2 (G/P, ν).
Since φd is a stable bi-K-invariant function, Proposition 3.2 combined with the fact
kπν (g)1G/P k1 = Ξ(g) = kπν (g−1 )1G/P k1 imply for the first term that there exists C > 0
so that for all g ∈ Γ
X
n
γ|L(γ)≤
Ξ(γ)
22d
Ξ(γ −1 g)
≤
C
C Ξ(g).
2d d
(1
+
L(g))
o (1 + L(γ))2d (1 + L(γ −1 g))2d
L(g)
2
A similar inequality holds for the second term. Therefore there exist d, C > 0 such that
for all g ∈ Γ
Ξ(g)
,
|f1 ∗ f2 |(g) ≤ C × Cd kf1 kS 2d (Γ) kf2 kS 2d (Γ)
L
L
(1 + L(g))2d
hence
kf1 ∗ f2 kS 2d (Γ) ≤ C × Cd kf1 kS 2d (Γ) kf2 kS 2d (Γ)
L
L
L
and the proof is complete.
Proof of Item (2) of Theorem 1.2. It is enough to prove that there exist C, d > 0 so that
for all finitely supported positive functions f on Γ
kλΓ (f )k ≤ Ckf kS d (Γ) .
L
By Shalom’s Lemma 2.1 it is sufficient to prove that there exist C, d > 0 so that for all
finitely supported positive functions f on Γ
kπν (f )k ≤ Ckf kS d (Γ) .
L
Fix d > 0 such that 2d > dim(a) + 2r and let f be a positive function in SL2d (Γ). We
have for all positive vectors ξ ∈ L1+ (G/P, ν)
X
f (γ)hπν (γ)1G/P , ξi ≤ kf kS 2d (Γ)
L
γ
X
γ
Ξ(γ)
hπν (γ)1G/P , ξi.
(1 + L(γ))2d
Notice also that the change of variable γ ′ = γ −1 combined with the symmetry properties
of Ξ and L imply
X
f (γ)hπν (γ −1 )1G/P , ξi ≤ kf kS 2d (Γ)
L
γ
X
γ
Ξ(γ)
hπν (γ)1G/P , ξi.
(1 + L(γ))2d
Therefore Proposition 3.2 implies that there exist d, C > 0 such that for all f ∈ SL2d (Γ)
and for all positive vectors ξ ∈ L1+ (G/P, ν)
X
(11)
f (γ)hπν (γ)1G/P , ξi ≤ C × Cd kf kS 2d (Γ) kξk1 .
γ
L
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
12
We have for all ξ and η in L2 (G/P )
X
|hπν (f )ξ, ηi| ≤
|f (γ)|hπν (γ)|ξ|, |η|i
γ∈Γ
≤
X
γ∈Γ
≤
1
1
|f (γ)| hπν (γ)1G/P , |η|2 i 2 hπν (γ −1 )1G/P , |ξ|2 i 2
X
γ∈Γ
1
1 X
2
2
−1
2
|f (γ)|hπν (γ )1G/P , |ξ| i ,
|f (γ)|hπν (γ)1G/P , |η| i
2
γ∈Γ
where the second inequality follows from Lemma 3.2 and the last inequality follows from
Cauchy-Schwarz inequality.
Then Equation (11) implies the existence of C > 0 such that
1
1
|hπν (f )ξ, ηi| ≤ (Ckf kS 2d (Γ) Cd kξk22 ) 2 (Ckf kS 2d (Γ) Cd kηk22 ) 2
L
L
= C × Cd kf kS 2d (Γ) kξk2 kηk2 .
L
By the definition of k · kop we have fund d, C > 0 such that for all finitely supported
functions f on Γ we have
kπν (f )k ≤ Ckf kS d (Γ) ,
L
and the proof is done.
References
[1] S. Barré; M. Pichot, The 4-string braid group B4 has property RD and exponential mesoscopic rank.
Bull. Soc. Math. France 139 (2011), no. 4, 479-502.
[2] S. Barré; M. Pichot, La propriété de décroissance rapide pour le groupe de Wise. (French) [The
rapid decay property for the Wise group] Ann. Inst. Fourier (Grenoble) 65 (2015), no. 2, 709-724.
[3] P. Baum, A. Connes, N. Higson, Classifying space for proper actions and K-theory of group C∗ algebras. C∗ -algebras: 1943-1993 (San Antonio, TX, 1993), 240-291, Contemp. Math., 167, Amer.
Math. Soc., Providence, RI, 1994.
[4] B. Bekka, P. de la Harpe, and A. Valette, Kazhdan’s property (T), New Mathematical Monographs,
vol. 11, Cambridge University Press, Cambridge, 2008.
[5] A. Boyer, Semisimple Lie groups satisfy property RD, a short proof. C. R. Math. Acad. Sci. Paris
351 (2013), no. 9-10, 335-338.
[6] A. Boyer, Quasi-regular representations and property RD, Potential Anal. 44 (2016), no. 2, 355-372.
[7] A. Boyer, On the square root of Poisson kernel in δ-hyperbolic spaces and some aspects of boundary
representations., preprint, arXiv:1505.00739.
[8] I. Chatterji, Introduction to the Rapid Decay property, preprint, arXiv:1604.06387.
[9] I. Chatterji, Property (RD) for cocompact lattices in a finite product of rank one Lie groups with
some rank two Lie groups. Geom. Dedicata 96 (2003), 161-177.
[10] I. Chatterji, K. Ruane, Some geometric groups with rapid decay. Geom. Funct. Anal. 15 (2005), no.
2, 311-339.
[11] I. Chatterji, C. Pittet, and L. Saloff-Coste, Connected Lie groups and property RD, Duke Math. J.
137 (2007), no. 3, 511-536.
[12] A. Connes; H. Moscovici, Cyclic cohomology, the Novikov conjecture and hyperbolic groups. Topology
29 (1990), no. 3, 345-388.
[13] P. de la Harpe, Groupes hyperboliques, algbres d’oprateurs et un thorme de Jolissaint. (French)
[Hyperbolic groups, operator algebras and Jolissaint’s theorem] C. R. Acad. Sci. Paris Sr. I Math.
307 (1988), no. 14, 771-774.
HARISH-CHANDRA–SCHWARTZ’S ALGEBRAS
13
[14] C. Drutu; M. Sapir, Relatively hyperbolic groups with rapid decay property. Int. Math. Res. Not.
2005, no. 19, 1181-1194.
[15] R. Gangolli, V.S. Varadarajan, Harmonic Analysis of Spherical Functions on Real Reductive Groups,
Springer-Verlag, New-York, 1988.
[16] U. Haagerup, An example of a nonnuclear C*-algebra which has the metric approximation property,
Invent. Math. 50 (1978/79), no. 3, 279-293.
[17] C. Herz, Sur le phénomène de Kunze-Stein, C. R. Acad. Sci. Paris Sr. A-B 271 (1970), A491-A493.
[18] P. Jolissaint, Rapidly decreasing functions in reduced C*-algebras of groups, Trans. Amer. Math.
Soc. 317 (1990), no. 1, 167-196.
[19] P. Jolissaint, K-theory of reduced C∗ -algebras and rapidly decreasing functions on groups. K-Theory
2 (1989), no. 6, 723-735.
[20] G. G. Kasparov, Index of invariant elliptic operators, K-theory and representations of Lie groups.
(Russian) Dokl. Akad. Nauk SSSR 268 (1983), no. 3, 533-537.
[21] A-W. Knapp, Representation theory of semisimple groups, Princeton landmarks in mathematics,
2001.
[22] V. Lafforgue, A proof of property (RD) for cocompact lattices of SL(3,R) and SL(3,C). J. Lie Theory
10 (2000), no. 2, 255-267.
[23] V. Lafforgue, K-théorie bivariante pour les algèbres de Banach et conjecture de Baum-Connes,
Invent. Math. 149 (2002), no. 1, 1-95 (French).
[24] J. Ramagge; G. Robertson; T. Steger, A Haagerup inequality for Ã1 × Ã1 and Ã2 buildings. Geom.
Funct. Anal. 8 (1998), no. 4, 702-731.
[25] M. Sapir, The rapid decay property and centroids in groups. J. Topol. Anal. 7 (2015), no. 3, 513-541.
[26] Y. Shalom, Rigidity, unitary representations of semisimple groups, and fundamental groups of manifolds with rank one transformation group, Ann. of Math. (2) 152 (2000), no. 1, 113-182.
[27] G. Skandalis, Progrès récents sur la conjecture de Baum-Connes. Contribution de Vincent Lafforgue.
(French) [Recent progress on the Baum-Connes conjecture: the contribution of Vincent Lafforgue]
Séminaire Bourbaki, Vol. 1999/2000. Astérisque No. 276 (2002), 105-135.
[28] A. Valette, Introduction to the Baum-Connes conjecture, Lectures in Mathematics ETH Zurich.
Birkhauser Verlag, Basel, (2002).
Weizmann Institute of Science, Rehovot, Israel
E-mail address: [email protected]
| 4math.GR
|
1
Maximizing the Link Throughput between
Smart meters and Aggregators as Secondary
arXiv:1506.04830v3 [cs.SY] 16 Feb 2016
Users under Power and Outage Constraints
Pedro H. J. Nardelli, Mauricio de Castro Tomé,
Hirley Alves, Carlos H. M. de Lima and Matti Latva-aho
Abstract
This paper assesses the communication link from smart meters to aggregators as (unlicensed)
secondary users that transmit their data over the (licensed) primary uplink channel. The proposed scenario
assumes: (i) meters’ and aggregators’ positions are fixed so highly directional antennas are employed,
(ii) secondary users transmit with limited power in relation to the primary, (iii) meters’ transmissions
are coordinated to avoid packet collisions, and (iv) the secondary links’ robustness is guaranteed by
an outage constraint. Under these assumptions, the interference caused by secondary users in both
primary (base-stations) and other secondary users can be neglected. As unlicensed users, however,
meter-aggregator links do experience interference from the mobile users of the primary network, whose
positions and traffic activity are unknown. To cope with this uncertainty, we model the mobile users
spatial distribution as a Poisson point process. We then derive a closed-form solution for the maximum
achievable throughput with respect to a reference secondary link subject to transmit power and outage
constraints. Our numerical results illustrate the effects of such constraints on the optimal throughput,
evincing that more frequent outage events improve the system performance in the scenario under study.
We also show that relatively high outage probabilities have little effect on the reconstruction of the
average power demand curve that is transmitted from the smart meter to the aggregator.
Index Terms
cognitive network, Poisson point process, smart grid communication, spectrum sharing
P. H. J. Nardelli, M. C. Tomé, H. Alves and M. Latva-aho are with the Centre for Wireless Communications (CWC), University
of Oulu, Finland. Contact: [email protected]. C. H. M. de Lima is with São Paulo State University (UNESP), São João da
Boa Vista, Brazil. This work is partly funded by Finnish Academy and CNPq/Brazil (n.490235/2012-3) as part of the joint
project SUSTAIN, by Strategic Research Council/Aka BC-DC project (n.292854), and by the European Commission through
the P2P-SmarTest project (n.646469).
2
I. I NTRODUCTION
Over the last few years, cognitive radios have appeared as the solution for more effective
use of the frequency spectrum [1], [2]. Following the concept proposed by Haykin [3], the
radio nodes should understand their environment to establish a wireless network “(...) with
two primary objectives in mind: highly reliable communication whenever and wherever needed;
efficient utilization of the radio spectrum.” Owing to their cognitive ability, radios would be then
capable of sensing the environment to decide on their transmission.
An interesting approach to the cognitive radio concept is the so-called spectrum sharing [4],
where unlicensed – secondary – users want to transmit some information without disturbing the
licensed – primary – users over the same frequency band. Secondary users then need to sense
the spectrum usage to decide about their transmissions. The transmission occurs if the channel
is sensed free, which depends on the primary user activity [5]. Otherwise, the secondary user
either searches for a different band or postpones its transmission.
Although the basic idea is simple, both analysis and implementation are challenging due to
interactive dynamics of the decision-making procedures [6]. The solution gets even harder when
we consider different possible applications and their specific requirements, yielding no universal
solution [7]. Notwithstanding, the idea of spectrum sharing is now widespread in both new
generations of cellular system (allowing for co-existence of macro- and small-cells as in [8])
and different sensor applications (e.g. [9]–[11]).
In this paper, we focus on the specific application of spectrum sharing in the deployment
of part of the communication network in the modern electric power grids [12], [13] – the socalled smart grids. As one would expect, the different applications of the electric power grid
have different requirements from the communication network perspective [14]. For instance,
control operations in the high-voltage grid must be close to real-time (scale of milliseconds) so
the communication link must have both extremely low latency and very high reliability (greater
than 99.9%). In contrast, remote reading of meters would allow for less stringent communication
requirements (latency of minutes and a reliability of 98%). As a consequence, the communication
network design should be fine-tuned with the specific power grid application and its needs. An
informative survey about the different applications and requirements is found in [15].
3
Looking specifically at how to implement cognitive radio for smart grid applications, the
authors in [16] presented an interesting hardware study case using software defined radio in
a micro-grid testbed. In [17], a Lyapunov-drift framework was proposed to differentiate traffic
priorities and then use the cognitive radio strategy called “dynamic spectrum access” to improve
the communications in smart grids. Additionally, [18] provides a recent survey on how to combine
cognitive radio approach into smart grid scenarios.
In our case, references [19] and [20] introduce the idea of employing a spectrum sharing
scheme within households where home appliances use unlicensed channel bands to build a
home area network (HAN). In neighborhood area networks (NANs), different scenarios may be
considered for the smart meters: (i) licensed users within the cellular systems directly connected
with the distribution operator [21], (ii) users of unlicensed bands of unused TV frequencies (TV
white spaces) [22], (iii) secondary users of licensed band [19], [20] and (iv) hybrid licensedunlicensed users [20]. When using solutions (i) and (iv), the smart grid elements are subscribers
(licensed users) of cellular networks. In this case, although quality of service shall be guaranteed,
they may involve prohibitive high costs and a dramatic increase in data traffic [21], making them
not so attractive for electricity providers if compared to (ii) and (iii). And yet, the latter solutions
will have a strong dependence on specific country legislation.
Despite such drawbacks, these solutions are interesting and related to ours. We believe,
however, that the analysis provided by the aforementioned works are focused on network-layer
considerations without a dedicated performance study involving specific topological considerations as node positions and mobility. Herein, we attempt to cope with such limitations by
studying a novel application of spectrum sharing technique for distribution grids considering
that some elements are fixed and others are mobile.
In specific terms, a spectrum sharing scheme (e.g [4]) will be assumed such that secondary
users transmit in the up-link channel of the cellular system. As meters and aggregators are
generally static nodes, their communication link may be built using directional antennas (e.g.
[23]) while respecting a transmit power constraint. In this way, the harmful effects caused by
the secondary user transmissions on the primary users, as well as on other secondary users,
due to co-channel interference are limited. By properly designing the antenna beamforming and
4
setting the power constraint, the probability that the secondary users interfere in the primary
transmissions would be low as far as directional antennas with limited power have a restrict
radiation pattern. Besides, by using the up-link channel, the cellular base-stations are disturbed
by the secondary transmissions and, due to their more robust reception procedures [24], the
interference can be further mitigated.
On the other hand, the secondary users do experience the interference caused by the primary
users. Since the up-link is employed by the former, the interference normally comes from
mobile devices, whose positions and traffic are unknown. To account for such uncertainties,
the interferers’ spatial distribution and traffic characteristics will be modeled using point process
theory [25]. Thereby, it is possible to derive closed-form expressions for important performance
metrics of wireless systems as outage probability and link throughput. It is worth mentioning that
the communication between different smart meters and their respective aggregator is coordinated
so they do not interfere to each other.
Our goal in this work is to optimize the link throughput (which is defined as the spectral
efficiency, given in bits per second per hertz, times the probability that the packet is successfully
decoded by the aggregator) of the secondary link under power (to not affect the primary users)
and outage (secondary link reliability) constraints. Using a similar optimization procedure as in
[26]–[28], we find the signal-to-interference ratio threshold and the transmit power employed by
the secondary link so as to maximize its throughput while respecting the imposed constraints.
Then, we found a closed-form equation for the optimal link throughput as a non-linear function
of the system constraints and the density of interferers, as well as other system parameters.
Interestingly, our results show that frequent outage situations have an unexpected positive effect
on the system performance for the scenario of interest (maximum value of outage constraint is
25%). We use actual data from the average power demand of a household (obtained from “The
Reference Energy Disaggregation Data Set” database [29], [30]) to show that relatively high
outage probabilities do not lead to a poor signal reconstruction if the information is sampled and
sent periodically, for example, every 15 minutes. All in all, our main contribution and novelty
are the theoretical evidences that up-link spectrum sharing is a good candidate for deploying the
communication network in the distribution electricity grids.
5
The rest of this paper is divided as follows. Section II explains the scenario under analysis,
justifying our assumptions and presenting the optimization problem to be solved. In Section III,
we focus on the solution of the optimization problem and illustrate how the system performance
changes with the configuration parameters. Section IV shows how the outage events affect the
reconstruction of the average power demand curve. In section V, we discuss some implications
of our theoretical results, indicating how they might be used in actual deployments. Section VI
concludes this paper.
II. S YSTEM MODEL
In this section, we describe the basic assumptions used to build our model and their implications. The assumptions are stated as follows:
‚
Assumption 1: Spectrum-sharing scenario where licensed (primary) and unlicensed (secondary) users share the frequency bands allocated to the up-link channel.
‚
Assumption 2: Primary link is established between static cellular base-stations and mobile
users. Secondary terminals are smart meters that need to forward data to a given aggregator
through the uplink channel and their positions are fixed.
‚
Assumption 3: The smart meters transmit with limited power Ws such that Ws ď Ws,max
where Ws,max is the maximum power allowed for the secondary users (which can be seen
as an imposition from the primary network).
‚
Assumption 4: Smart meters associated with the same aggregator are able to perfectly
coordinate their transmissions using time scheduling.
In this case, Assumption 2 indicates the possibility of employing directional antennas in the
secondary links as far as their positions are fixed. Orientation errors as defined in [23] can be then
completely avoided when deploying the secondary network, making highly directional antennas
worth. In its turn, Assumption 3 imposes the maximum range that the signal transmitted by the
smart meters can reach. Putting all together, the radiation pattern generated by the secondary
transmission can be seen as a line segment starting in the smart meter, passing through the
aggregator and ending in a point related to Ws,max .
6
Let us now look at the interference related to the proposed spectrum sharing. Assumptions 1, 2
and 4 indicate the co-channel interference occurs: (i) from smart meters to cellular base-stations,
(ii) from smart meters to aggregators that they are not associated, and (iii) from mobile users
to aggregators. From the implications discussed in the previous paragraph, the cases (i) and (ii)
can be neglected by designing the specific locations when the deploying either the secondary or
primary networks. Even if the positions are considered random in two-dimensions, the chance of
having a base-station or an aggregator in the line segment related to the smart meter transmitted
signal approaches zero, which further indicated that such cases should not be considered.
Hence, only case (iii) is relevant for our analysis. To evaluate its impact in the system
performance, we first need to model uncertainty of the mobile users’ positions and traffic activity.
We assume here a Poisson field of interferers [31] such that the interferer nodes are distributed
over an infinite plane following a 2-dimensional Poisson point process Φ with density λ, given
in interferers per square-meter. The wireless channel model employed in this paper consists of
two components: one related to the distance-dependent path-loss such that the received power
decays with the distance and other related to fast-fading [25]. The received power at the node
of interest can be computed as gi ri´α , where ri is the distance between the reference receiver
and the ith node, gi is the channel gain between them, and α ą 2 the path-loss exponent.
Hereafter, we focus our analysis on a reference smart meter-aggregator link, as shown in
Fig. 1. During transmissions intervals, we assume that the interferers’ positions and the channel
gains do not change. We also consider an interference-limited scenario wherein the noise effects
can be neglected. As pointed in [32], the inclusion of the noise power leads to a more complex
analysis without providing any significant qualitative difference.
If the primary users are equipped with omni-directional antennas and transmit with the same
fixed power Wp , the signal-to-interference ratio (SIR) at the aggregator can be computed as [25]:
SIR0 “
Ws g0 r0´α
ÿ
,
Wp gi ri´α
(1)
iPΦ
where the index 0 denotes the reference link.
We assume that the reference link employs both point-to-point Gaussian codes and the interferenceas-noise decoding rule [28], [33] so that a spectral efficiency of log2 p1`βq in bits/s/Hz is achieved
7
Fig. 1.
An illustration of the proposed scenario, where primary and secondary users share the up-link channel. The reference
smart meter (secondary transmitter) is depicted by the house, the aggregator (secondary receiver) by the the CPU , the handsets
are the mobile primary users (interferers to the aggregator) and the big antenna is the cellular base-station. As the smart meter
uses directional antenna with limited transmit power (bold arrow), its interference towards the base-station can be ignored. The
thin black arrows represent the primary users’ desired signal, while the red ones their interference towards the aggregator.
only if the SIR is greater than β. In this case, the probability Psuc that a packet is successfully
decoded by the aggregator is the probability that SIR0 ą β. Then, an outage event happens with
probability 1 ´ Psuc . In our scenario, retransmissions are not allowed so that the information
contained in packets received in outage is lost.
To compute Psuc , we assume quasi-static channel gains (squared envelops) g that are independent and identically distributed exponential random variables (Rayleigh fading), and also a
dynamic topology where interferers’ positions change every transmission interval. Therefore,
every transmission attempt can be viewed as a different realization of the point processes Φ
and the channel gains g. We consider here that the distance between the reference meter and
aggregator is known and has a fixed value r0 “ d. Finally, the success probability is computed
as [25]:
Psuc “ e
´λκπd2 β 2{α
´
Wp
Ws
¯2{α
,
(2)
where κ “ Γp1 ` 2{αqΓp1 ´ 2{αq with Γp¨q being the Gamma function. It is also worth noting
that the primary users might employ different power control strategies, but it will not affect the
8
qualitiative results of our analysis due to the stochastic geometry approach employed (e.g. [34]).
We are now ready to define the performance metric and carry out the optimization problem
under consideration.
Definition 1 (Link throughput): The throughput T of the reference link using the system
model described in this section is defined as:
´λκπd2 β 2{α
´
T “ logp1 ` βq Psuc “ log2 p1 ` βq e
Wp
Ws
¯2{α
.
(3)
Our goal in this paper is to find the setting of parameters for the secondary users to maximize
their link throughput T while respecting the imposed power limit and outage constraints. In our
case, the variables in hand are the coding rate β and the transmit power Ws of the reference
link. Mathematically, we have the following problem:
max
pβ,Ws q
s.t.
log2 p1 ` βq e
´λκπd2 β 2{α
Ws ď Wmax
´
Wp
Ws
¯2{α
,
(4)
1 ´ Psuc ď
where is the maximum acceptable outage probability, reflecting the reference link robustness.
III. M AXIMUM THROUGHPUT UNDER POWER AND OUTAGE CONSTRAINTS
In this section, we solve the optimization problem previously stated. We then provide some
numerical results to illustrate how the constraints imposed to our smart meter-aggregator reference link will affect the maximum achievable throughput. Before we start, we still need to
present a Lemma that tells us how the throughput behaves as a function of the secondary transmit
power Ws and the SIR threshold β when no constraint is considered.
Lemma 1: Let us consider the throughput equation, given by (3), as a function of the variables
Ws ą 0 and β ą 0, i.e. T “ f pWs , βq. The function f is monotonically crescent in respect to
Ws , and it is concave in respect to β if B 2 T {Bβ 2 ă 0.
Outline of proof: The proof of this Lemma is straightforward from the analysis of the
function in terms of the (strictly positive) variables Ws and β, and knowing that the function T
is twice differentiable in terms of β. Any increase in Ws leads to an increase in the exponential
term of (3) and then in T . Increasing β, on the other hand, has a two-fold effect: it increases the
9
logarithmic term while decreases the exponential one. The function T , however, is not always
concave in relation to β ą 0; nevertheless, in the region that B 2 T {Bβ 2 ă 0, T is concave.
Since T is concave for some values of β, we may try to find the value of β that leads to the
maximum T . This is shown in the next Lemma.
˚
˚
Lemma 2: Let βun
denote the value of β that maximizes T assuming that βun
is in the region
˚
where B 2 T {Bβ 2 ă 0. Then, βun
is the solution of the following (transcendental) equation:
αβ “ k β 2{α p1 ` βq lnp1 ` βq,
(5)
where k “ 2λκπd2 pWp {Ws q2{α .
˚
Outline of proof: Let us first consider that βun
is in the region where the inequality
B 2 T {Bβ 2 ă 0 holds. Then, from Lemma 1, the value of β that maximizes T is the solution
of the derivative equation BT {Bβ “ 0. By doing so in (3), we end up in (5), which has no
closed-form solution. If a solution does not exist, then our initial assumption the inequality
˚
B 2 T {Bβ 2 ă 0 does not hold and βun
cannot be obtained.
˚
Nonetheless, in the cases of interest, (5) has a solution and therefore βun
is in the region
where B 2 T {Bβ 2 ă 0 as we shall see next. It is noteworthy that even though (5) does not have
a closed-form solution, it can be easily evaluated numerically through standard mathematical
software as, for instance, [35].
Fig. 2 exemplifies the behavior of the link throughput T with the transmit power Ws and
the threshold β. The curve behaves as predicted by Lemmas 1 and 2. Intuitively, when all
parameters are kept the same, it is advantageous for the smart meter to increase its transmit
power so that the SIR experienced by the aggregetor tends to increase. This behavior, although
individually optimal, is not good for the other links as transmitting with more power would
increase the interference level throughout the network (refer to [27] for a deeper assessment of
an optimization problem wherein the optimal individual solution can be socially harmful). For
this reason, the power constraint that our problem assumes is needed, as we will discuss later.
When analyzing the effects of the threshold β, there is a trade-off involved. An increase of
β leads to a more efficient transmission where more bits/s/Hz can be transmitted in the same
message. This gain, however, comes at expense of more frequent outage events, which in turn
10
β =1
0.45
Ws =1
0.45
0.40
0.35
0.35
0.30
0.30
0.25
0.25
0.20
0.20
0.15
0.15
0.10
0.10
0.05
0.05
T
0.40
0.00
0.0
0.5
1.0
Ws
1.5
2.0
0.00
0.0
0.5
1.0
β
1.5
2.0
2.5
Fig. 2. Link throughput T given in (3) as a function of the transmit power Ws (left) and the SIR threshold β (right) for α “ 4,
d “ 1, λ “ 0.25 and Wp “ 1. The black dot in the right plot is the optimal operating point predicted by Lemma 2.
decreases the link throughput. From Lemma 2, the value of β that leads to the optimal operating
point can be found by numerically solving (5). In our case, we use the function fsolvep q from
the Python library SciPy [35].
We are now almost ready to present the main result of this paper, which is the solution of the
constrained optimization problem presented in the previous section. But first, we still need to
state a lemma about the relation between the optimal values of the constrained and unconstrained
SIR thresholds β.
Lemma 3: Let β ˚ denote the value of β that leads to the maximum constrained throughput
˚
given in (4) and βun
, given in Lemma 2, denote the value that optimizes the unconstrained
throughput. Then:
´
1´e
˚
αβun
˚ q lnp1`β ˚ q
2p1`βun
un
˚
ą ùñ β ˚ ă βun
.
(6)
˚
Proof: In the optimal unconstrained operating point βun
given in Lemma 2, the following
equality must hold:
β 2{α “
αβ
.
kp1 ` βq lnp1 ` βq
(7)
11
˚
Then, although we cannot analytically compute the actual value of βun
, we do know that the
´
outage probability 1 ´ Psuc , given in (2), related to it will be given by: 1 ´ e
˚
αβun
˚ q lnp1`β ˚ q
2p1`βun
un
,
˚
.
which is a monotonically decreasing function of βun
In this case, if that probability is smaller than the constraint , then the constrained optimal
˚
threshold is the unconstrained one, i.e. β ˚ “ βun
. Otherwise, the optimal constrained threshold
˚
˚
β ˚ must be smaller than βun
since the outage probability is a decreasing function of βun
.
Proposition 1: Let us assume that the pair pWs˚ , β ˚ q is the solution of the optimization problem
given by (4). If 1 ´ e
´
˚
αβun
˚ q logp1`β ˚ q
2p1`βun
un
ą , the pair pWs˚ , β ˚ q is computed as:
Ws˚ “ Wmax ,
ˆ
˙α{2
Wmax
lnp1 ´ q
˚
β “
´
.
Wp
λκπd2
(8)
(9)
The optimal throughput T ˚ is then:
˜
T ˚ “ p1 ´ q ˆ log2
Wmax
1`
Wp
ˆ
lnp1 ´ q
´
λκπd2
˙α{2 ¸
.
(10)
Proof: Let us start by considering the variable Ws . From Lemma 1, we know that the
throughput is a monotonically increasing function of Ws , regardless of β. Then, Ws must assume
its highest possible value: Ws˚ “ Wmax .
To find β ˚ , we first use the assumption that 1 ´ e
´
˚
αβun
˚ q lnp1`β ˚ q
2p1`βun
un
ą . From Lemma 3,
˚
the inequality β ˚ ă βun
holds. Then, we use the fact that T is also a monotonically crescent
˚
function of β in the range 0 ă β ă βun
so that β ˚ should be the highest value that satisfies
the inequality 1 ´ Psuc ď . Manipulating the constraint by assuming Ws˚ “ Wmax , we obtain
´
¯α{2
lnp1´q
Wmax
β ď Wp ´ λκπd2
. In this case, equality gives β ˚ .
´
Remark 1: This result is only valid if the initial assumption 1 ´ e
˚
αβun
˚ q lnp1`β ˚ q
2p1`βun
un
ą holds,
which is true for the cases of interest, namely α P p2, 6s and P p0, 0.25q. Notice that α P p2, 6s
comprises indoor and outdoor scenarios as well as rural and urban areas [36]. For example, when
˚
α “ 4 and β “ 1.24 (the value of βun
in the example used in Fig. 2), the outage probability is
˚
0.75. If that inequality does not hold, the optimal solution is β ˚ “ βun
and the optimal power
Ws˚ should be computed accordingly. In other words, the value of β ˚ is kept fixed, while the
optimal power Ws˚ is the variable used to optimize the link throughput.
12
0.012
Wmax =0.3,² =0.15
Wmax =0.1,² =0.15
Wmax =0.3,² =0.05
Wmax =0.1,² =0.05
0.010
T∗
0.008
0.006
0.004
0.002
0.000
0.2
Fig. 3.
0.3
0.4
λ
0.5
0.6
0.7
Maximum link throughput T ˚ given in (10) as a function of the density of active mobile primary users (interferers) λ
for α “ 4, d “ 1, Wp “ 1 and different values of the constraints Wmax and .
Corollary 1: The optimal throughput T ˚ can be approximated by:
˙α{2
ˆ
p1 ´ q Wmax
lnp1 ´ q
˚
,
T «
´
lnp2q Wp
λκπd2
´
¯α{2
lnp1´q
when WWmax
´
is small.
λκπd2
p
(11)
Remark 2: Corollary 1 holds for the cases under study. For example, when the system variables are set as follows: λ “ 0.25, α “ 4, d “ 1, Wp “ 1, Wp “ 0.5 and “ 0.05, the term is
´
¯α{2
lnp1´q
Wmax
´ λκπd2
“ 0.00086, therefore the approximation lnp1 ` xq « x works well.
Wp
Next we will illustrate the analytic results just presented to get more insights on how the
maximum link throughput T ˚ is affected by the activity of the primary users, as well as the
system constraints. Fig. 3 shows how the optimal constrained throughput T ˚ in the smart meterto-aggregator link behaves in relation to the density λ of active mobile users that interfere in its
communication. As expected, increasing the density of interfering nodes decreases the maximum
throughput achieved by the reference link, regardless of values of Wmax and assumed. It is
13
0.012
0.010
T∗
0.008
λ =0.25,² =0.15
λ =0.50,² =0.15
λ =0.25,² =0.05
λ =0.50,² =0.05
0.006
0.004
0.002
0.000
0.0
Fig. 4.
0.1
0.2
Wmax
0.3
0.4
0.5
Maximum link throughput T ˚ given in (10) as a function of the secondary power constraint Wmax for α “ 4, d “ 1,
Wp “ 1 and different values of λ and .
interesting to see that T ˚ exponentially decays with λ, which indicates that the secondary link
performance is dramatically affected by the primary users’ increase of activity.
We can also see from Fig. 3 that the values of the constraints Wmax and affect the maximum
throughput curves. Higher values of either Wmax or leads to higher T ˚ , when λ is fixed. While
the result is intuitive for Wmax (by increasing the transmission power, we obtain higher SIR and
link throughput), it is not so when the outage constraint is considered: a less strict constraint
leads to higher throughputs.
To better understand those behaviors, we present in Figs. 4 and 5 the maximum throughput
T ˚ versus Wmax and , respectively. Fig. 4 shows that the maximum throughput T ˚ linearly
grows with Wmax , which was predicted by Corollary 1. This means that any relaxation in the
power constraint Wmax provides a linear gain in the secondary link throughput, whose slope
is directly defined by the system variables; therefore, a combination between a relatively high
outage constraint and low density of interferers provides the best performance.
14
0.012
0.010
T∗
0.008
λ =0.25,Wmax =0.3
λ =0.25,Wmax =0.1
λ =0.50,Wmax =0.3
λ =0.50,Wmax =0.1
0.006
0.004
0.002
0.000
0.00
Fig. 5.
0.05
0.10
²
0.15
0.20
Maximum link throughput T ˚ given in (10) as a function of the outage constraint for α “ 4, d “ 1, Wp “ 1 and
different values of λ and Wmax .
The behavior of T ˚ as a function of the outage constraint is more complicated since there
is a trade-off involved, as shown in (10) and (11). However, from our assumption that is a
relatively small probability, i.e. P p0, 0.25q, then T ˚ is a (non-linear) crescent function of .
Fig. 5 illustrates this growth, evincing that an increase ∆ for smaller values of leads to lower
variation ∆T ˚ in the maximum throughput. In this case, allowing for more outage events is
more advantageous for the link: the spectral efficiency gains obtained by setting higher SIR
thresholds dominates the system performance. It is worth reinforcing that this behavior – which
is somehow counter-intuitive – is only valid for lower values of P p0, 0.25q. If outages are not
a constraint in our system, the results of Proposition 1 should be reviewed under the perspective
of Lemmas 1 and 2, and Remark 1.
15
IV. O UTAGE EVENTS AND SIGNAL RECONSTRUCTION
We just showed in the previous section that allowing for more frequent outage events increases
the link throughput. However, outage events will affect the information that needs to be sent by
the smart meter to the aggregator.
Let xrns be the discrete signal transmitted by the smart meter and x̂rns be the signal received
by the aggregator, where n “ 1, ..., N with N being the last sample. We consider the xrns is
the average power demand over a fixed period of time τ . The aggregator needs then to use
x̂rns to reconstruct the signal xrns. We assume here that the signal is reconstructed via linear
interpolation between two adjacent points. If the communication is perfect (i.e. x̂rns “ xrns) the
interpolation is always between x̂rks and x̂rk ´ 1s, with k “ 2, ..., N .
This, however, is not the case in our model since outage events may occur due to the primary
users activity. If a sample is lost, the aggregator will interpolate the missing value(s) using the
latest two received samples. Consider the transmitted sequence: xrk ´ 2s, xrk ´ 1s, xrks with
k “ 2, ..., N . If the samples xrk ´ 2s and xrks are successfully received but xrk ´ 1s is not, the
reconstruction is based on the linear interpolation of x̂rks “ xrks and x̂rk ´ 2s “ xrk ´ 2s. The
estimation of the missing point is then x̂rk ´ 1s “ pxrks ` xrk ´ 2sq{2.
To carry out this procedure, we use the “The Reference Energy Disaggregation Data Set”
database [29], [30] from where we build our signal xrns, which is the 15-minute average
power demand over a timespan of 24 hours (one day). The information from the smart meter is
transmitted to the aggregator every 15 minutes (yielding τ “ 0.25 hour and N “ 96 samples),
which reconstructs the signal as previously described.
Fig. 6 illustrates the effects of the outage probability on the signal reconstruction by showing
the original signal and a snapshot of outputs assuming the constraints “ 0.05 and “ 0.15,
where the missing points in the “ 0.05 and “ 0.15 curves are identified by the red squares. At
first sight, one cannot notice a big difference between the three signals, and the most noticeable
differences are in the “ 0.05 curve (specifically in the “peaks”), not in the “ 0.15 one. This
interesting fact happens due to the nature of the signal itself and the randomness of outage events.
The power demand signal seems to have a burst nature with a floor level and few peaks, which
is related to personal habits when the house is occupied and people are engaged in activities
16
Original signal
Average power demand
6000
5000
4000
3000
2000
1000
0
0
1
2
3
4
5
6
7
8
9
10
11
Average power demand
13
14
15
16
17
18
19
20
21
22
23
24
13
14
15
16
17
18
19
20
21
22
23
24
13
14
15
16
17
18
19
20
21
22
23
24
² =0.05
6000
5000
4000
3000
2000
1000
0
0
1
2
3
4
5
6
7
8
9
10
11
12
² =0.15
6000
Average power demand
12
5000
4000
3000
2000
1000
0
0
Fig. 6.
1
2
3
4
5
6
7
8
9
10
11
12
Time [h]
Average power demand of a house measured in watts during a period of 24 hours. The signal is reconstructed in
the aggregator as a linear interpolation between two subsequent points. If an outage happens, the point related to the power
demand at that time is lost. On the top: perfect reconstruction. At the center: outage probability “ 0.05. On the bottom: outage
probability “ 0.15. The red squares in the last two plots are the missing points (samples).
like cooking or showering.
For example, we can see in Fig. 6 that the peaks are around 8:00 in the morning and noon.
These peaks are probably related to people getting ready to work and having lunch. Other than
this, specially between 10:00 in the evening and 6:00 in the morning of the following day,
17
we can see that the energy consumption is quite low and constant, most probably related to
appliances in stand-by and refrigerator cycles [37]. In this way, most of the samples xrks will
have similar values. If independent and identically distributed erasures occur, the probability that
the estimated point approximates the missing one is high.
Consider again the transmitted sequence: xrk ´2s, xrk ´1s, xrks with k “ 2, ..., N and that the
samples xrk ´ 2s and xrks are successfully received but xrk ´ 1s is not. Then, the reconstruction
is x̂rk ´ 1s “ pxrks ` xrk ´ 2sq{2. As in most of the cases xrks « xrk ´ 1s « xrk ´ 2s, then
x̂rk ´ 1s “ pxrks ` xrk ´ 2sq{2 « xrk ´ 1s. However, during the peaks, this does not hold and
errors become evident. The snapshot presented for “ 0.05 is an unlucky one as far as the
transmissions failure happened in the peak periods. Conversely, although more samples were
lost when “ 0.15, they were mostly in the floor-level periods and the signal reconstruction
was not affected in this specific snapshot.
To statistically analyze the effects of the outage events in the signal reconstruction, we use the
root-mean-square deviation (RMSD) such that the reconstruction error of the x̂rns is computed
as:
g
f
N
f1 ÿ
e
RMSD “
px̂rks ´ xrksq2 .
N k“1
(12)
Fig. 7 presents how the RMSD changes with the maximum allowed outage probability .
As expected, more frequent outage events (indicated by higher ) leads to a worse signal
reconstruction (indicated by higher RMSDs). Our results show, however, that has relatively
little effect on reconstructing the signal. Let us first analyze the example presented in Fig. 6
(black curve in Fig. 7). For the highest value considered “ 0.25, the RMSD is 227 watts for a
signal with mean of 587 watts and with range max xrns ´ min xrns “ 4895 watts. In this case,
the normalized RMSD with respect to the mean is approximately 0.39 while with respect to the
range is less than 0.05. Looking back to the signal itself in top plot of Fig. 6, one can see that
there are few points that, if erased, would cause a significant distortion on its reconstruction.
In order to have a more robust analysis, we also present in Fig. 7 the best and the worst cases
in terms of RMSD among 53 households, whose data is available in [30]. For this scenario,
18
350
300
RMSD
250
Worst
Example
Best
200
150
100
50
0
0.00
Fig. 7.
0.05
0.10
²
0.15
0.20
0.25
Root-mean-square deviation (RMSD) versus the outage constraint . The results are presented for the example given
in Fig. 6 together with the best and worst cases in terms of RMSD among 53-household demand data from [30]. The curves
are the result of Monte Carlo simulations considering 105 realizations for each point.
our example is closer to the worst case than to the best one. We further show in Fig. 8 the
frequency diagram (histogram) of RMSD for the scenario where “ 0.25. As one can see, most
households perform better than the example presented Fig. 6 (RMSD “ 227 W); they are in
fact much closer to the best case presented in Fig. 7. All in all, these results reinforces even
more our argument that, for the scenario under consideration, the signal reconstruction is weakly
affected by relatively frequent outage events.
V. D ISCUSSIONS
Throughout the last two sections, we studied a spectrum sharing scenario where a given smart
meter-aggregator pair communicates using a licensed cellular uplink channel. For this scenario,
we analytically assessed how the system constraints for being a secondary user affect the link
throughput and the signal reconstruction. In this section, we start from the presented results to
19
16
Number of households
14
12
10
8
6
4
2
0
Fig. 8.
50
100
150
200
RMSD
250
300
Frequency diagram of the root-mean-square deviation (RMSD) considering data from 53 households from [30] for
“ 0.25 versus the outage constraint . The results are based on Monte Carlo simulations using 105 realizations for each point.
discuss how the proposed spectrum-sharing scheme could be implemented in actual systems.
A. Relation between licensed and unlicensed users
In our theoretical model, we assume that the secondary, unlicensed, users do not affect the
primary users (base-stations). As discussed before, this can indeed be the case when deploying
our strategy in a real system. Since the positions of the smart meters, aggregators and basestations are fixed, the use of highly directional antennas in the smart meter-aggregator link with
low transmit power will decrease the interference level at the base-station, which in turn still
has capabilities of dealing with such residual interference.
Looking at the interference caused by the mobile users in the aggregator, our analysis assume
that the density of these nodes are fixed. This, however, will probably not hold because the
primary user activity probably would intensify during periods that match with power demand
peaks. Then, in a practical scenario, this traffic variations is somehow predicted and the smart
20
meter may adaptively change its operation point either by setting a pre-determined, predicted,
density of interferers to optimize its communication link (simpler solution) or by sensing and
estimating such a density from time to time (more complex solution).
Another important point that our model can be useful relates to Denial-of-Service (DoS)
attacks in the physical layer [38]. In this scenario, a node or a group of nodes intentionally
interfering in the aggregator reception could characterize a DoS. From the analysis presented
here, we could infer how frequent outage events need to be to affect the signal reconstruction.
A dedicate study would be also possible by modeling the attackers as another Poisson point
process.
B. Outage constraint and link robustness
While the power constraint is required by the secondary link to not interfere in the primary
users, the outage constraint is set to guarantee a minimum robustness at the secondary link.
Conversely to what one would expect, our results show that allowing for more frequent outage
events improves the link throughput due to the contradictory effects of the SIR requirement on
the system performance (lower outage constraints lead to higher SIR constraints, which results
in higher spectral efficiency, while it decreases the success probability).
The question that arises from this result is how robust against outage events the secondary
link should be? One can only reply this answer knowing the information that is sent to the
aggregator. Our example shows that, if the information to be sent is the average power demand,
the signal reconstruction is possible even with relatively loose outage constraints. While this
happens due to the nature of the input signal as discussed in the previous section, higher outage
probabilities might be not desirable for other kind of signals or if the aggregator should provide
some kind of feedback to the smart meter change the power demand behavior, as in strategies
of demand-side management [39].
C. Power demand signal processing and transmission
The signal presented in Fig. 6 exemplifies a 15-minute sampling interval of the power demand
of a household. The signal characteristic indicates that time-based sampling might not be the
21
most efficient way of collecting and then send the data to the aggregator. For instance, by using
event-based sampling [40], [41], the communication link should be much more robust (i.e. lower
values of the outage constraint) since there will be much less redundant data and therefore the
loss of any sample will have a more dramatic effect on the signal reconstruction.
Although this is not the focus of the present paper, we would like to mention that there is a
trade-off between the sampling strategies and communication. A more efficient way of sampling
leads to less points for reconstructing the original signal and vice-versa. If this is the case, the
transmission strategies and then the outage constraint should be evaluated in combination with
the sampling procedure. In the scenario used here, the time-based sampling generate redundant
information about the signal so that outage events do not have drastic effects on the signal
reconstruction.
For this new study, however, it is important to have a stochastic characterization of power demand signal as in [42], [43]. By doing so, the signal processing framework would be generalized
and adapted to different consumption patterns. This would allow us to study ways to optimize
different sampling strategies (time-based, event-based and hybrid) and transmission system designs (coding rate, medium access protocol and retransmissions), while considering the different
requirements as reconstruction error and information privacy. Looking at the reconstruction
procedure, more advanced strategies like machine learning would probably offer better options
to reconstruct the data than linear interpolation, but at expense of higher computational costs
[44]. From this perspective, it would be worth evaluating the cost of the sampling-reconstruction
and the benefits of link throughput.
In this paper, however, we have chosen to employ a two-dimensional analysis so as to
understand the relations between optimal throughput and the reconstruction error itself; our
goal is provide the knowledge of what combination of requirements is possible to achieve. For
example, an application that requires a very high reliability would imply in a lower throughput.
Consequently, a cost-benefit analysis involving those aspects would require more information
about the application requirement.
Another different scenario might consider (hybrid) Automatic Repeat Request (ARQ) strategies
in order to enhance the communication link, which reduces outage events [26]. This scenario may
22
also include MAC protocols in order to control re-transmissions as well as coordination among
nodes. One recent example of a MAC protocol designed in a similar scenario was developed
in [45], where the authors introduced a centralized strategy that resorts to a specialized frame
structure that support co-existence between a cognitive and primary networks. Along the same
line, the idea to include a routing protocol that protects the primary users while meeting the
utility requirements of smart grid network was proposed in [46]. Another scenario of interest
deals with security and privacy of the transmitted data, since there is a trade-off between security
and reliability as pointed out in [47], where the authors resort to information theoretical tools
in order to guarantee security already at the physical layer. All in all, combining these more
advanced communication techniques to decrease the outage events with a more efficient sampling
and enhanced security at the physical layer would be an interesting next step for the present
work.
VI. C ONCLUSIONS
This paper assesses a spectrum sharing scenario where smart meters send periodic information
to an aggregator over licensed cellular uplink channels. We assume that the secondary link uses
directional antennas with limited transmit power so its interference in the primary users can be
neglected. Mobile primary users, on the other hand, interfere with the aggregator reception. Modeling the interferers’ spatial distribution as a Poisson point process, we analyzed the secondary
link throughput, finding then its optimal value under power and outage constraints.
Our results show that relatively high outage constraints surprisingly improve the link throughput for the cases of interest, even though more samples will be lost. It is worth mentioning the
fact that smart meter reading application with scheduled intervals requires nowadays a reliability
of at least 98% [15, Table 3]; this is a fairly high value if compared to our results. In fact, the
discrepancy between the actual requirement and our results was a surprise. We also show that, due
to the burst nature of the power demand signal that is transmitted in the smart meter-aggregator
link, outage events do not have a dramatic effect in the signal reconstruction in comparison to
the perfect transmission.
We plan to study in future works how outage events will affect the signal reconstruction
23
under different sampling strategies. In this way, we plan to build a joint sampling-transmission
technique that can improve the system efficiency, as discussed in Section V-C. We also expect
to implement the ideas proposed here in an actual demonstration to verify the validity of our
assumptions and proposed optimization strategy.
R EFERENCES
[1] I. F. Akyildizet et al., “Next generation/dynamic spectrum access/cognitive radio wireless networks: A survey,” Computer
Netw., vol. 50, no. 13, pp. 2127–2159, 2006.
[2] C. Cormio and K. R. Chowdhury, “A survey on MAC protocols for cognitive radio networks,” Ad Hoc Netw., vol. 7, no. 7,
pp. 1315 – 1329, 2009.
[3] S. Haykin, “Cognitive radio: brain-empowered wireless communications,” IEEE J. Sel. Areas Commun., vol. 23, no. 2, pp.
201–220, Feb. 2005.
[4] J. Peha, “Sharing spectrum through spectrum policy reform and cognitive radio,” Proc. IEEE, vol. 97, no. 4, pp. 708–719,
Apr. 2009.
[5] Y. Saleem and M. H. Rehmani, “Primary radio user activity models for cognitive radio networks: A survey,” J. Netw.
Comput. Appl., vol. 43, pp. 1–16, 2014.
[6] B. Wang et al., “Game theory for cognitive radio networks: An overview,” Computer Netw., vol. 54, no. 14, pp. 2537–2561,
2010.
[7] A. Ahmad et al., “A survey on radio resource allocation in cognitive radio sensor networks,” Commun. Surveys Tuts.,
vol. 17, no. 2, pp. 888–917, 2015.
[8] C. H. de Lima et al., “Coordination mechanisms for self-organizing femtocells in two-tier coexistence scenarios,” IEEE
Trans. Wireless Commun., vol. 11, no. 6, pp. 2212–2223, 2012.
[9] L. D. Xu et al., “Internet of things in industries: A survey,” IEEE Trans. Ind. Informat., vol. 10, no. 4, pp. 2233–2243,
Nov. 2014.
[10] S. D. T. Kelly et al., “Towards the implementation of IoT for environmental condition monitoring in homes,” IEEE Sensors
J., vol. 13, no. 10, pp. 3846–3853, Oct. 2013.
[11] A. Bradai et al., “EMCOS: Energy-efficient mechanism for multimedia streaming over cognitive radio sensor networks,”
Pervasive Mob. Comput., vol. 22, pp. 16–32, 2015.
[12] S. F. Bush, Smart grid: Communication-enabled intelligence for the electric power grid.
John Wiley & Sons, 2014.
[13] P. H. J. Nardelli et al., “Models for the modern power grid,” EPJ-ST, vol. 223, no. 12, pp. 2423–2437, 2014.
[14] Y. Yan et al., “A survey on smart grid communication infrastructures: Motivations, requirements and challenges,” Commun.
Surveys Tuts., vol. 15, no. 1, pp. 5–20, 2013.
[15] M. Kuzlu et al., “Communication network requirements for major smart grid applications in HAN, NAN and WAN,”
Computer Netw., vol. 67, pp. 74–88, 2014.
[16] R. C. Qiu et al., “Cognitive radio network for the smart grid: Experimental system architecture, control algorithms, security,
and microgrid testbed,” IEEE Trans. Smart Grid, vol. 2, no. 4, pp. 724–740, Dec. 2011.
24
[17] G. Shah et al., “A cross-layer QoS-aware communication framework in cognitive radio sensor networks for smart grid
applications,” IEEE Trans. Ind. Informat., vol. 9, no. 3, pp. 1477–1485, Aug. 2013.
[18] A. A. Khan et al., “Cognitive radio for smart grids: Survey of architectures, spectrum sensing mechanisms, and networking
protocols,” Commun. Surveys Tuts., 2015, to appear.
[19] R. Yu et al., “Cognitive radio based hierarchical communications infrastructure for smart grid,” IEEE Netw., vol. 25, no. 5,
pp. 6–14, 2011.
[20] V. C. Gungor and D. Şahin, “Cognitive radio networks for smart grid applications: A promising technology to overcome
spectrum inefficiency,” IEEE Veh. Technol. Mag, vol. 7, no. 2, pp. 41–46, Jun. 2012.
[21] J. J. Nielsen et al., “What can wireless cellular technologies do about the upcoming smart metering traffic?” IEEE Commun.
Mag., vol. 53, no. 9, pp. 41–47, Sep. 2015.
[22] Y. Zhang et al., “Cognitive machine-to-machine communications: Visions and potentials for the smart grid,” IEEE Netw.,
vol. 26, no. 3, pp. 6–13, May 2012.
[23] J. Wildman et al., “On the joint impact of beamwidth and orientation error on throughput in directional wireless Poisson
networks,” IEEE Trans. Wireless Commun, vol. 13, no. 12, pp. 7072–7085, Dec. 2014.
[24] J. Andrews, “Interference cancellation for cellular systems: A contemporary overview,” IEEE Wireless Commun., vol. 12,
no. 2, pp. 19–29, April 2005.
[25] M. Haenggi, Stochastic geometry for wireless networks.
Cambridge University Press, 2012.
[26] P. H. J. Nardelli et al., “Optimal transmission capacity of ad hoc networks with packet retransmissions,” IEEE Trans.
Wireless Commun., vol. 11, no. 8, pp. 2760 –2766, Aug. 2012.
[27] ——, “Throughput optimization in wireless networks under stability and packet loss constraints,” IEEE Trans. Mobile
Comput., vol. 13, no. 8, pp. 1883–1895, Aug. 2014.
[28] P. H. J. Nardelli et al., “Throughput analysis of cognitive wireless networks with Poisson distributed nodes based on
location information,” Ad Hoc Netw., 2015.
[29] J. Z. Kolter and M. J. Johnson, “REDD: A public data set for energy disaggregation research,” in Workshop on Data
Mining Applications in Sustainability (SIGKDD), San Diego, CA, vol. 25.
Citeseer, 2011, pp. 59–62.
[30] ——. [Online]. Available: http://redd.csail.mit.edu/
[31] P. Cardieri, “Modeling interference in wireless ad hoc networks,” Commun. Surveys Tuts., vol. 12, no. 4, pp. 551–572,
2010.
[32] S. Weber et al., “An overview of the transmission capacity of wireless networks,” IEEE Trans. Commun., vol. 58, no. 12,
pp. 3593–3604, Dec. 2010.
[33] F. Baccelli et al., “Interference networks with point-to-point codes,” IEEE Trans. Inf. Theory, vol. 57, no. 5, pp. 2582–2596,
May 2011.
[34] S. Weber et al., “The effect of fading, channel inversion, and threshold scheduling on ad hoc networks,” IEEE Trans. Inf.
Theory, vol. 53, no. 11, pp. 4127–4149, Nov. 2007.
[35] [Online]. Available: http://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.optimize.fsolve.html
[36] A. Goldsmith, Wireless Communications.
Cambridge University Press, 2005.
[37] M. Pipattanasomporn et al., “Load profiles of selected major household appliances and their demand response opportunities,” IEEE Trans. Smart Grid, vol. 5, no. 2, pp. 742–750, March 2014.
25
[38] E.-K. Lee et al., “Physical layer security in wireless smart grid,” IEEE Commun. Mag., vol. 50, no. 8, pp. 46–52, 2012.
[39] G. Strbac, “Demand side management: Benefits and challenges,” Energy policy, vol. 36, no. 12, pp. 4419–4426, 2008.
[40] Y. Tsividis, “Event-driven data acquisition and digital signal processing: A tutorial,” IEEE Trans. Circuits Syst. II, Exp.
Briefs, vol. 57, no. 8, pp. 577–581, 2010.
[41] M. Simonov et al., “Gathering process data in low-voltage systems by enhanced event-driven metering,” IEEE Syst. J,
vol. PP, no. 99, pp. 1–12, 2015.
[42] J. Munkhammar et al., “Characterizing probability density distributions for household electricity load profiles from highresolution electricity use data,” Appl. Energ., vol. 135, pp. 382–390, 2014.
[43] I. Sajjad et al., “Probabilistic generation of time-coupled aggregate residential demand patterns,” IET Gener. Transm. Dis.,
vol. 9, no. 9, pp. 789–797, 2015.
[44] M. Abu Alsheikh et al., “Machine learning in wireless sensor networks: Algorithms, strategies, and applications,” Commun.
Surveys Tuts, vol. 16, no. 4, pp. 1996–2018, 2014.
[45] A. Aijaz and H. Aghvami, “PRMA based cognitive machine-to-machine communication in smart grid networks,” IEEE
Trans. Veh. Technol, vol. PP, no. 99, pp. 1–1, 2014.
[46] A. Aijaz et al., “CORPL: A routing protocol for cognitive radio enabled AMI networks,” IEEE Trans. Smart Grid, vol. 6,
no. 1, pp. 477–485, Jan 2015.
[47] H. Alves et al., “Enhanced Transmit Antenna Selection Scheme for Secure Throughput Maximization Without CSI at the
Transmitter and its Applications on Smart Grids,” IEEE Trans. Inf. Forensics and Security, pp. Submitted – Online:, Nov
2015. [Online]. Available: http://arxiv.org/pdf/1511.06518
| 3cs.SY
|
Low-complexity Image and Video Coding Based on an Approximate Discrete Tchebichef Transform
Paulo A. M. Oliveira∗
Renato J. Cintra†
Fábio M. Bayer‡
Sunera Kulasekera§
Arjuna Madanayake§
arXiv:1609.07630v2 [cs.MM] 20 Apr 2017
Abstract
The usage of linear transformations has great relevance for data decorrelation applications, like image and video compression.
In that sense, the discrete Tchebichef transform (DTT) possesses useful coding and decorrelation properties. The DTT transform
kernel does not depend on the input data and fast algorithms can be developed to real time applications. However, the DTT
fast algorithm presented in literature possess high computational complexity. In this work, we introduce a new low-complexity
approximation for the DTT. The fast algorithm of the proposed transform is multiplication-free and requires a reduced number of
additions and bit-shifting operations. Image and video compression simulations in popular standards shows good performance of
the proposed transform. Regarding hardware resource consumption for FPGA shows 43.1% reduction of configurable logic blocks
and ASIC place and route realization shows 57.7% reduction in the area-time figure when compared with the 2-D version of the
exact DTT.
Keywords
Approximate transforms, discrete Tchebichef transform, fast algorithms, image and video coding
multiplications. Such multiplications are known to be more demanding computational structures than additions or fixed-point
Discrete variable orthogonal polynomials emerge as solutions of multiplications, both in software and hardware. Thus, the higher
several hypergeometric difference equations [1]. Classic applica- computational complexity of the DTT precludes its applications
tions of this class of orthogonal polynomials include functional in low power consumption systems [22, 23] and/or real-time proanalysis [2] and graphs [3]. Additionally, such polynomials are cessing, such as video streaming [24, 25]. Therefore, fast algoemployed in the computation of moment functions [4], which rithms for the DTT could improve its computational efficiency.
are largely used in image processing [5–7]. For instance, the A comprehensive literature search reveals only two fast algodiscrete Tchebichef moments [8], which are derived from the dis- rithms for the 4-point DTT [14, 26] and one for the 8-point
crete Tchebichef polynomials, form a set of orthogonal moment DTT [15]. Although these fast algorithms possess lower arithfunctions. Such functions are not discrete approximation based metic complexities when compared with the direct DTT calcuon continuous functions; they are naturally orthogonal over the lation, they still possess high arithmetic complexity, requiring a
discrete domain.
significant amount of additions and bit-shifting operations.
The Tchebichef moments have been used for quantifying image
In a comparable scenario, the computation of DCT-based
block artifact [9], image recognition [10–12], blind integrity vertransforms—which has been employed in several popular coding
ification [13], and image compression [14–18]. In the data comschemes such as JPEG [27], MPEG-2 [28], H.261 [29], H.263 [30],
pression context, bi-dimensional (2-D) moments are computed
H.264 [31], HEVC [32, 33], and VP9 [34]—has profited from
by means of the 2-D discrete Tchebichef transform (DTT). In
matrix approximation theory [35–41]. In this context, discrete
fact, the 8-point DTT can achieve better performance when comtransforms are not exactly calculated, but instead an approxiparison with the discrete cosine transform (DCT) [19], in terms
mate, low-cost computation, is performed. The approximations
of average bit length as reported in [16, 20, 21]. Moreover, the
are designed in such a way to allow similar spectral and coding
8-point DTT-based embedded encoder proposed in [18], shows
characteristics as well as lower arithmetic complexity. Usually,
improved image quality and reduced encoding/decoding time in
approximations are multiplierless, requiring only addition and
comparison with state-of-the-art DCT-based embedded coders.
bit-shifting operations for its computation. In [42], a multiplierThe 8-point DTT has also been employed in blind forensics, as
less approximation for the 8-point DTT is proposed. To the best
a tool to determine the integrity of medical imagery subject to
of our knowledge, this is the only DTT aproximation archived
filtering and compression [13].
in literature.
However, the exact DTT possesses high arithmetic complexThe aim of this work is to introduce an efficient low-complexity
ity, due to its significant amount of additions and float-point
approximation for the 8-point DTT capable of outperform∗
Paulo A. M. Oliveira is with the Signal Processing Group, Departa- ing [42]. To derive multiplierless approximate DTT matrix, a
mento de Estatı́stica, Universidade Federal de Pernambuco, Recife, PE,
Brazil; Multimedia Communications and Signal Processing, University of multicriteria optimization problem is sought, combining differErlangen–Nuremberg, Erlangen, BY, Germany.
ent coding metrics: coding gain and transform efficiency. Addi†
Renato J. Cintra is with the Signal Processing Group, Departamento tionally, a fast algorithm for efficient computation of the sought
de Estatı́stica, Universidade Federal de Pernambuco, Recife, PE, Brazil;
Équipe Cairn, INRIA-IRISA, Université de Rennes, Rennes, France; and approximation is also pursued. For coding performance evaluLIRIS, Institut National des Sciences Appliquées, Lyon, France (e-mail: ation, we propose two computational experiments: (i) a JPEG
[email protected]).
image compression simulation and (ii) a video coding experiment
‡
Fábio M. Bayer is with the Departamento de Estatı́stica and LACESM,
which
consists of embedding the sought approximation into the
Universidade Federal de Santa Maria, Santa Maria, RS, Brazil (e-mail:
[email protected]).
H.264/AVC standard.
1
Introduction
§
Sunera Kulasekera and Arjuna Madanayake are with the Department
of Electrical and Computer Engineering at the University of Akron, OH
(e-mail: [email protected]).
The paper unfolds as follows. Section 2 reviews the mathematical background of the DTT. Section 3 introduces a parametriza1
tion of the DTT to derive a family of DTT approximations and
sets up an optimization problem to identify optimal approximations. In Section 4, we assess the obtained approximation in
terms of coding performance, proximity with the exact transform, and computation cost. Moreover, a fast algorithm for the
proposed approximate DTT is introduced. Section 5 shows the
results of the image and video compression simulations. Section 6 shows hardware resource consumption comparison with
the exact DTT for both FPGA and ASIC realizations. A discussion and final remarks are shown in Section 7.
2
2.1
where T is the N -point unidimensional DTT matrix given by
te [0]
te [1]
···
te [N−1]
0
T=
0
0
te1 [0]
te1 [1]
..
.
..
.
···
..
te1 [N−1]
..
.
.
teN −1 [0] teN −1 [1] ··· e
tN −1 [N−1]
.
The matrix operations induced by (2) represents the 2-D DTT.
Because of the kernel separation property, the 2-D DTT can
be calculated by means the successive applications of the 1-D
DTT to the rows of f ; and then to columns of the resulting
intermediate matrix. The original intensity distribution f can
be recovered by the inverse procedure:
Discrete Tchebichef Transform
f =T−1 · M · (T−1 )⊤ = T⊤ · M · T.
Discrete Tchebichef Polynomials
The discrete Tchebichef polynomials are a set of discrete variable The last equality above stems from the DTT orthogonality proporthogonal polynomials [43]. The kth order discrete Tchebichef erty: T⊤ = T−1 [14]. Therefore, the same structure can be used
polynomials are given by the following closed form expres- at the forward transform as well in the inverse.
sion [14]:
For N = 4 and N = 8, we have the particular cases of interest
in the context of image and video coding. Thus, the 4- and
tk [n] = (1 − N )k · 3 F2 (−k, −n, 1 + k; 1, 1 − N ; 1),
8-point DTT matrices are, respectively, furnished by:
1 1 1 1
where n = 0, 1, . . . , N − 1, 3 F2 (a1 , a2 , a3 ; b1 , b2 ; z) =
P∞ (a1 )k (a2 )k (a3 )k z k
−3 −1 1 3
T
=F
·
4
4
is the generalized hypergeometric func1 −1 −1 1
n=0
(b1 )k (b2 )k
k!
tion and (a)k = a(a + 1) · · · (a + k − 1) is the descendant factorial. Tchebichef polynomials can be obtained according to the and
following recursion [14]:
2k − 1
k−1
tk [n] =
tk [1] tk−1 [n] −
N 2 − (k − 1)2 tk−2 [n],
k
k
−1
T8 =
1
· F8 ·
2
3 −3
1
1
1
1
1
1
−7 −5 −3 −1
1
−5
−77 51 −37 −53 −3
7 −13 −3 9 9
−7 23 −17 −15 15
1 −5
9 −5 −5
−1
7 −21 35 −35
1
1
3
5
−3
1
−7 −5
−3 −13
17 −23
9 −5
21 −7
1
7
7
7
7,
7
1
1
for t0 [n] = 1 and t1 [n] = 2n − N + 1. Indeed, the set {tk [n]},
k = 0, 1, . . . , N − 1, is an orthogonal basis in respect with the
1 √1
1 √1
,
,
where
F
=
diag
,
and F8
=
4
2
2
20
20
unit weight. Consequently, the discrete Tchebichef polynomials
1
1
1
1
1
1
1
1
satisfy the following mathematical relation:
We observe
diag √2 , √42 , √42 , √66 , √142 , √546 , √66 , √858 .
that
T
and
T
are
written
as
a
result
from
the
product
of an
4
8
N−1
X
integer matrix and a diagonal matrix which requires float-point
ti [n]tj [n] = ρ(j, N ) · δi,j ,
representation.
i=0
(N+k)!
and δi,j is the Kronecker delta
where ρ(k, N ) = (2k+1)·(N−k−1)!
function which yields δi,j = 1, if i = j, and δi,j = 0, otherwise.
2.2
3
DTT Approximations and Coding Optimality
In this section, we aim at proposing an extremely low-complexity
DTT approximation. Our methodology consists of generating a
class of parametric approximate matrices and then identify the
optimal class member in terms of coding performance.
2-D Discrete Tchebichef Transform
Let f [m, n], m, n = 0, 1, . . . , N − 1, be an intensity distribution
from a discrete image of size N × N pixels. The 2-D DTT of
f [m, n], denoted by M[p, q], p, q = 0, 1, . . . , N − 1, is given by [8,
3.1
14]:
Related Work
To the best of our knowledge, the only DTT approximation
archived in literature was proposed in [42]. That approximation
was obtained by means of a parameterization of integer funcm,n=0
tions combined with a normalization of transformation matrix
where e
tk [n], k = 0, 1, . . . , N −1, are the orthonormalized
discrete
columns. The derived approximation in [42] furnishes good codp
Tchebichef polynomials given by e
tk [n] = tk [n]/ ρ(k, N ).
ing capabilities, but it lacks orthogonality or near-orthogonality
Note that the transform kernel described in (1) is separable. properties. As a consequence, the forward and inverse transforHence, the following is relation holds true:
mations are quite distinct and possess unbalanced computational
complexities.
N−1
N−1
X
X
e
e
tq [n] · f [m, n],
tp [m]
M[p, q] =
3.2 Parametric Low-complexity Matrices
n=0
m=0
M[p, q] =
N−1
X
e
tp [m] · e
tq [n] · f [m, n],
(1)
for p, q = 0, 1, . . . , N − 1. Therefore, the transform-domain coef- In [35, 36, 38, 44], DCT approximations were proposed according
ficients of f can be calculated by the following matrix operation: to following operation:
M = T · f · T⊤ ,
int(α · C),
(2)
2
where int(·) is an integer function, α is a real scaling factor, and
C is the exact DCT matrix. Usual integer functions include the
floor, ceiling, signal, and rounding functions [38]. In this work,
these functions operate element-wise when applied to a matrix
argument.
A similar approach is sought for the proposed DTT approximation. However, in contrast with the DCT, the rows of
DTT matrix (basis vectors) have a widely varying dynamic
range. Thus, the integer function may excessively penalize
the rows with small dynamic range. To compensate this
phenomenon, we normalize the rows of T4 √and T√8 accord
ing to left multiplications by D4 = diag 2, 320 , 2, 320 and
√ √
√
√
√
√
√
√
568
D8 = diag
, re8, 168
, 168
, 264
, 13
, 2184
, 264
, 3432
7
7
7
23
9
35
spectively.
The sought approximations are required to possess extremely
low complexity. One way of ensuring this property is to adopt an
integer function whose co-domain is a set of low-complexity integer. In the DCT literature, common sets are: P0 = {±1} [35],
P1 = {0, ±1} [36], and P2 = {0, ±1, ±2} [38]. Note that elements from these sets have very simple realization in hardware; implying multiplierless designs with only addition and bitshifting operations [45].
Adopting P2 , we have that a suitable integer function is given
by:
3.4
Now our goal is to identify in the family TN (α) the matrix that
furnishes the best approximation. We adopted two metrics as
figures of merit to guide the optimal choice: (i) the unified coding gain Cg [51,52] and (ii) the transform efficiency η [53]. These
metrics are relevant, because they quantify the transform capacity of removing signal redundancy, as well as data compression
and decorrelation [53].
Hence, following the methodology in [54], we propose the following multicriteria optimization problem:
n
o
Cg T̂N (α) , η T̂N (α) , N ∈ {4, 8},
α∗ = arg max
0<α<5/2
∗
where α is the scaling parameter that results in the optimal low
complexity matrix T∗N , T̂N (α∗ ) according to (3).
The above optimization problem is not analytically tractable.
Thus, we resort to exhaustive numerical search to obtain α∗ . We
consider linearly spaced values of α with a step of 10−3 in the
interval 0 < α < 5/2. For N = 4 and N = 8, we obtain that op23 69
timality is found in the intervals ( 23 , 25 ) and ( 14
, 34 ), respectively.
Therefore, any value of α in the aforementioned intervals effects
the same approximations. For operational reasons, we selected
α = 2. Thus, the resulting low-complexity matrices are given
below:
1 1 1 1 1 1 1 1
−2 −1 −1 0 0 1 1 2
1 1 1 1
−1 0 2
−22 01 −12 −11 −1
∗
∗
−1 −2 −1 2
−2 −1 1 2
T4 =
and T8 =
1 −1 −1 1
1 −2 0 1 1 0 −2 1 .
round : [−1, 1] → P2 = {0, ±1, ±2} ,
x 7→ round(α · x),
0 < α < 5/2,
−1
where round(x) = sign(x) · ⌊|x| + 21 ⌋ is the rounding function
as implemented in MATLAB [46], Octave [47], and Python [48]
programming languages. Following the methodology described
in [38], we obtain the following parametric class of matrices:
TN (α) = round(α · DN · TN ),
3.3
N ∈ {4, 8}.
Optimization Problem
2 −2
1
−1 2 −1 −1 1
0 −1 2 −1 −1
0 0 −1 2 −2
1 −2
2 −1
1 0
1
0
0
The associate optimal approximations are denoted by T̂∗N ,
T̂N (α∗ ) and can be computed
according
Hence,
to (4).
we obtain: T̂∗4 = diag 12 , √110 , 12 , √110 · T∗4 and T̂∗8 =
diag √18 , √112 , √112 , √120 , √112 , √114 , √112 , √110 · T∗8
(3)
4
DTT Approximation
Evaluation and Computational Complexity
A given low-complexity matrix TN (α) can be used to approxi- 4.1 Discussion
mate the DTT matrix by means of orthogonalization or quasi∗
∗
orthogonalization as described in [36–38]. As a result, an ap- The obtained matrix T4 satisfies (5) and therefore T̂4 is orthogproximation for TN , referred to as T̂N (α), can be obtained by: onal. In fact, the proposed matrix is identical to the 4-point integer transform for H.264 encoding introduced by Malvar et al. [44].
T̂N (α) = SN (α) · TN (α),
(4) Therefore, it is also an optimal approximate DTT. Because the
4-point DCT approximation matrix by Malvar et al. was submitq
where SN (α) = {ediag(TN (α) · TN (α)⊤ )}−1 is a diagonal ma- ted to in-depth analyses in the context of video coding [31], such
∗
trix, ediag(·) returns a√diagonal matrix with the diagonal entries results also apply to T4 . Therefore, hereafter we focus the forthcoming
discussions
to
the
proposed 8-point approximation T∗8 .
of its argument and · is the matrix element-wise square root
operator [38]. If
TN (α) · TN (α)⊤ = [diagonal matrix]
4.2
(5)
Orthogonality and Invertibility
The matrix T∗8 does not satisfy (5). Therefore, the associate
approximation T̂∗8 is not orthogonal, i.e. (T̂∗8 )−1 6= (T̂∗8 )⊤ . As
holds true, then T̂N (α) is an orthogonal matrix [49]. Otherwise,
a consequence, the inverse does not inherit the low-complexity
it is possibly a near orthogonal matrix [38]. An approximation
properties of T∗8 . In other words, the entries of (T∗8 )−1 are not
is said quasi-orthogonal when the deviation from diagonality of
in P2 . Hence, the proposed transform possesses asymmetrical
TN (α) · TN (α)⊤ is considered small. Let A be a square real
computational costs when comparing the direct and inverse opmatrix. The deviation from diagonality δ(A) is given by [50]:
erations [55]. The inverse transformation is a much required tool,
specially for reconstruction encoded images back to the spatial
k ediag(A)kF
δ(A) = 1 −
,
domain [27, 56].
kAkF
However, T∗8 · (T∗8 )⊤ has a low deviation from diagonalwhere k · kF is the Frobenius norm for matrices [49]. In the ity [38, 50]: only 0.024—roughly 4.4 times less than the devicontext of image compression, a deviation from diagonality value ation implied by the SDCT [35], which is taken as the stanbelow 1 − √25 ≈ 0.1056 indicates quasi-orthogonality [35, 38].
dard reference. Therefore, the following approximation is valid:
3
(T̂∗8 )−1 ≈ (T∗8 )⊤ · [diag(8, 3, 3, 5, 3, 7/2, 3, 5/2)]−1 . Moreover, where
since diagonal matrices can be absorbed into other computational steps [38, 40, 42], (T∗8 )−1 can be replaced with the lowcomplexity matrix (T∗8 )⊤ . This has the advantage of using the
same algorithm for both forward and inverse approximations.
4.3
Performance Assessment
A1 =
The proposed approximations were compared with their corresponding exact DTT in terms of coding performance as measured according to the coding gain and the transform efficiency.
For N = 8, we also include in our comparisons the DTT approximation proposed in [42]. The coding performance of the
proposed approximations was evaluated according to the figures
of merit Cg and η. Table 1 displays the results. The proposed
approximations are capable of furnishing coding measures very
close to the exact transformations. For comparison purposes,
the exact DCT has its coding gain and transform efficiency of
8.83 dB and 93.99, respectively.
Although our goal is to derive good approximations for coding,
we also analyzed the resulting approximations in terms of proximity metrics. We separated the mean square error (MSE) [57],
the total energy error ǫ [36], and the transform distortion d [58].
All these measures aim at quantifying the distance between the
exact transformations and their respective approximations. Analytic expressions for the MSE and ǫ are detailed in [54]. We also
evaluated the proximity of the proposed approximations with respect to the exact DTT according to the transform distortion
measure suggested in [58]. This metric was originally proposed
as the DCT distortion in the context of DCT approximations and
quantifies in percentage a distance between exact and approximate DCT. Adapting it for the DTT, we obtain the transform
distortion as follows:
d(T̂∗N ) =
1−
n
o
1
· ediag TN · (T̂∗N )⊤
N
2
2
A2 =
P=
0
00
0
0
0
1
0
1
00
2
0
0
0
0
0
1
0
00
00
0
0
1
0
0
0
0
0
0
0
0
1
0
0
0
0
1
0
0
0
1
−1
0
0
0
0
0
0
1
1
0
0
0
0
0
0
0 0
0 0
0 0
0 0
1 0
0 0
0 1
0 0
0 0 0 0 0 1
0 0 0 0 1 0
1 0 0 1 0 0
0 1 1 0 0 0
0 1 −1 0 0 0 ,
1 0 0 −1 0 0
0 0 0 0 −1 0
0 0 0 0 0 −1
1 0 0 0 0 0
0 1 0 0 0 0
0 0 0 0 0 0
2 −1 0 0 0 0
−1 −1 0 0 0 0
0 0 2 −1 0 0
0 0 1 1 0 0
0 0 0 1 1 0
0 0 0 0 2 −1
0 0 0 0 0 −2
1 0 0 0 0 0 0
−2 0 0 0 0 0 0
0 1 0 0 0 0 0
0 0 1 0 0 0 0
0 0 0 1 0 0 0
0 0 0 0 1 1 0
0 0 0 0 0 −1 0
0 0 0 0 −1 0 1
0 0 0 0 0
0 0 0 1 0
1 0 0 0 0
0 0 1 0 0
0 0 0 0 0.
0 0 0 0 1
0 0 0 0 0
0 1 0 0 0
0
0
0
0
0,
1
1
0
Matrix B8 represents a layer of butterfly structures, A1 and
A2 denote additive matrices with bit-shifting operations, and P
represents a final permutation, which is cost-free. The resulting algorithm preserves all algebraic and coding properties of
the direct computation, while requiring less arithmetic operations. Moreover, the factor of 2 of the first matrix row can be
absorbed into the diagonal matrix. The obtained factorization
is represented by the signal flow graph shown in Figure 1. Such
algorithm reduces the arithmetic cost of the proposed approximation to only 24 additions and six bit-shifting operations.
Table 2 compares the arithmetic complexity of the discussed
methods evaluated according to both their respective fast algorithms. The additive and total arithmetic complexity of the
proposed approximation are 45.5% and 58.9% lower than the exact transform, respectively. Although, the computational cost of
the proposed approximation is slightly higher than the forward
DTT approximation in [42], it is important to notice that the
inverse transformation in [42] is relatively more complex. Considering the combination of forward and inverse transformation,
the design in [42] requires 49 additions, whereas the proposed
design requires 48 additions. Bit-shifting costs are virtually
null, because in a hardware implementation they represent only
wiring. For comparison, the popular Loeffler DCT algorithm [61]
requires 11 floating-point multiplications and 29 additions. Although the actual approximation consists of the multiplication
of the low-complexity matrix and a diagonal matrix as shown
in (4), the multiplications introduced by diagonal matrices represent no additional arithmetic complexity in image compression applications. This is because they can be absorbed into the
image quantization step [27, 56] of JPEG-like image compression [36–41]. Furthermore, the new approximation is capable of
a better coding performance and possesses one order of magnitude lower proximity measure as shown in Table 1. Therefore,
both performance and arithmetic complexity measures are favorable to the proposed approximation.
× 100%,
where k·k2 is the euclidean norm for matrices [49]. Low values of
distortion indicates proximity with the DTT. As a comparison
for N = 4, the 4-point DCT approximation proposed in [59]
has a distortion of 7.32%. Proximity results are also shown in
Table 1.
4.4
B8 =
1
Fast Algorithm and Arithmetic Complexity
Now we aim at deriving fast algorithms for the obtained DTT approximations. Being identical to the H.264 4-point DCT approximation, the derived matrix T∗4 was given fast algorithms in [44].
Although T∗8 is multiplication free, without a fast algorithm, its
direct implementation requires 44 additions and 24 bit-shifting
operations. Thus we focus our efforts on the efficient computation of T∗8 . To such end, a sparse matrix factorization is sought,
where the number of additions and bit-shifting operations can
be significantly reduced [45, 60].
The sparse matrix factorization proposed in the manuscript
was derived from scratch based on usual butterfly structures [45]. 5 Image and Video Compression Experiments
We obtained the following decomposition:
In this section, we perform two computation experiments. The
first one consists of still image compression considering a JPEG
T∗8 = P · A2 · A1 · B8 ,
procedure. The second simulation assess the effectiveness of the
4
Table 1: Performance assessment
Cg (dB)
η
MSE
ǫ
N
Method
d
δ
4
Exact DTT [26]
Proposed
7.55
7.55
97.25
97.33
0.001
0.13
0.29%
0
0
8
Exact DTT [15]
DTT Approx. [42]
Proposed
8.68
6.60
8.57
92.86
83.50
89.52
0.015
0.002
3.32
0.77
12.61%
3.03%
0
0.09
0.024
x0
X0
x1
X4
x2
X6
x3
X2
x4
X7
x5
X3
x6
X1
x7
X5
Figure 1: Signal flow graph for T∗8 . Input data xn , n = 0, 1, . . . , 7, relates to output Xm , m = 0, 1, . . . , 7. Dashed arrows and
black nodes represent multiplications by −1 and 2, respectively.
Table 2: Fast algorithm arithmetic complexity comparison
Method
Mult.
Adit.
Shifts
Total
0
0
0
0
44
20
29
24
29
0
8
6
73
20
37
30
Exact DTT [15]
Forward DTT approx. [42]
Inverse DTT approx. [42]
Proposed
5
proposed approximations under realistic video encoding condi- high compression ratios. It could outperform the DTT approximation in [42] in terms of SSIM and SR-SIM for QF < 80
tions.
and QF < 55, respectively. It shows that the proposed approximation is more efficient in the scenario of high and moderate
5.1 Image Compression
compression, which are the very common cases [67], suitable for
We adopted the image compression method according to the low-power devices. The approximation in [42] could attain a
JPEG standard [27, 56]. A set of 45 512×512 8-bit images were better performance at low compression ratios because it satisfies
obtained from a public image bank [62] and submitted to process- the perfect reconstruction property—albeit at the expense of an
ing. The selected images encompass a wide range of categories, inverse transformation with higher arithmetic complexity. On
including 13 textures, 12 satellite images, three human faces and the other hand, the proposed approximation explores the nearseveral other miscellaneous scenarios. For each image, the lumi- orthogonality property which could excel in moderate to high
nance component was extracted and subdivided into 8×8 blocks, compression scenarios—which are often more relevant [67].
Ii,j , i, j = 1, 2, . . . , 64. After preprocessing, each block was subFor qualitative evaluation purposes, Figure 3 shows the
mitted to the the following operation:
compressed Lena image according to the exact DTT and
the proposed approximation. We adopted the scenario of
Mi,j = P · Ii,j · P⊤ ,
high/moderate compression with QF = 15 and QF = 50, respectively. In both cases, the approximate transform was capable of
where Mi,j is the 2-D transform-domain data and P is a given producing comparable results to the exact DTT with visually
1-D transformation matrix, such as the exact or approximate similar images.
DTT. Then each subblock Mi,j was element-wise divided by a
quantization matrix, yielding the quantized JPEG coefficients
5.2 Video Compression Simulation
Ji,j , as follows:
To evaluate the proposed transform performance in video codJi,j = round (Mi,j ⊘ Q) ,
ing, we have embedded the DTT approximation in the widely
used x264 software library [68] for encoding video streams into
where Q = ⌊(S · Q0 + 50)/100⌋ is the quantization matrix, ⌊·⌋
the H.264/AVC standard [31]. The default 8-point transform emdenotes the floor function, the default quantization table is
ployed in H.264 is the following integer DCT approximation [69]:
16 11 10 16 24 40 51 61
8 8 8 8 8 8 8 8
12 12 14 19 26 58 60 55
12 10
6
3 −3 −6 −10 −12
13 16 24 40 57 69 56
14
−4 −8 −8 −4
4
8
14 17 22 29 51 84 80 62
108 −34 −12
1
,
Q0 =
−6
6 12
3 −10
18
22
37
56
68
109
103
77
Ĉ = ·
8 −8 −8
8
8 −8 −8
8.
24 35 55 64 81 104 113 92
8
49 64 78 87 103 121 120 101
6 −12
3 10 −10 −3 12 −6
72 92 95 98 112 100 103
99
4
3
S = 5000/QF , if QF < 50, and 200 − 2 · QF otherwise, and QF
is the quality factor [56]. If QF = 50, then Q = Q0 . Decreasing
values of QF lead to higher compression ratios (with image total
destruction at QF = 0); whereas increasing values leads to lower
compression ratios (with best possible quality at QF = 100). In
our experiments, we adopted QF varying from 10 to 90 in steps
of 5. In the JPEG decoder, each sub-block is initially arithmetic
decoded and dequantized according to: M̂i,j = Ji,j ⊙ Q. Then,
the sub-blocks are inverse transformed: Îi,j = P−1 · M̂i,j · P−⊤ .
Original and compressed images were compared for image
degradation. The structural similarity index (SSIM) [63] and the
spectral residual based similarity (SR-SIM) [64] were separated
as image quality measures. The SSIM takes into account luminance, contrast, and the image structure to quantify the image
degradation, being consistent with subjective quality measurements [65]. On its turn, the SR-SIM is based on the hypothesis
that the visual saliency maps of natural images are closely related to their perceived quality. This measure could outperform
several state-of-the-art figures of merit in experiments with standardized datasets [64]. We did not consider the peak signal-tonoise ratio (PSNR) as a quality measure, because it is not a suitable metric to capture the human perception of image fidelity
and quality [57]. For each value of QF , we considered average
measure values instead of values from particular images. This approach is less prone to variance effects and fortuitous data [36,66].
For direct comparison, we selected the exact DTT [15], the approximation proposed in [42], and the proposed approximation.
As an extra reference, we also included the results from the standard JPEG, which is based on the exact DCT. Figure 2 displays
the results. For both selected measures, the proposed approximation performed very closely to the exact DTT, specially at
−8
−6
8 −4
10 −12
−4
8
12 −10
−8
6
4
−3
The fast algorithm for the above transformation requires 32 additions and 14 bit-shifting operations [69]. Therefore, the proposed
8-point transform requires 25% less additions and 57% less bitshifting operations than the fast algorithm for Ĉ.
Eleven 300-frame common intermediate format (CIF) videos
obtained from an online test video database [70] were encoded
with the standard software and then with the modified software.
We employed the software default settings and conducted the
simulation under two scenarios: (i) target bitrate varying from
50 to 500 kbps with steps of 50 kbps and (ii) quantization parameter (QP) varying from 5 to 50 with steps of 5. Psychovisual
optimization was disabled in order to obtain valid SSIM values.
Besides PSNR evaluation, the discussed software library [68] offers natively SSIM measurements for video quality assessment.
Average SSIM of the luma component were computed for all reconstructed frames. The results are shown in Figure 4 in terms
of the absolute percentage error (APE) [38] of the SSIM with respect to the standard DCT-based transformation in the original
H.264/AVC codec. This measure is given by:
APE(SSIM) =
SSIMH.264 − SSIMP
,
SSIMH.264
where SSIMH.264 returns the SSIM figures as computed according to the H.264 standard and SSIMP represents the SSIM when
the exact DTT, the approximation in [42], or the proposed
approximation are considered. SSIM curves for the DCT are
absent, because they were employed as performance references.
The use of the proposed transform effects a minor degradation in
the video quality. It also could perform better than the previous
approximation in all cases.
6
1.00
1.000
0.95
0.995
Average Y-SR-SIM
Average Y-SSIM
0.990
0.90
0.85
0.80
JPEG
Exact DTT [15]
Approx. [42]
Proposed
0.75
0.70
10
20
30
40
50
60
Quality factor
70
80
0.985
0.980
0.975
JPEG
Exact DTT [15]
Approx. [42]
Proposed
0.970
0.965
0.960
90
10
20
30
40
50
60
Quality factor
(a)
70
80
90
(b)
Figure 2: Average SSIM (top) and SR-SIM (bottom) measurements for image compression for the considered transforms at several
values of QF .
(a) Exact DTT [15], QF = 15
(b) Exact DTT [15], QF = 50
(c) Proposed, QF = 15
(d) Proposed, QF = 50
Figure 3: Compressed ‘Lena’ image for QF = 15 and QF = 50.
0.04
Exact DTT [15]
Approx. [42]
Proposed
0.035
Average Y-SSIM APE
Average Y-SSIM APE
0.040
Exact DTT [15]
Approx. [42]
Proposed
0.05
0.03
0.02
0.030
0.025
0.020
0.015
0.010
0.01
0.005
0.00
100
200
300
Target bitrate (kbps)
400
0.000
500
10
20
30
QP
(a)
(b)
Figure 4: Video quality assessment in terms of target bitrate and QP.
7
40
50
DTT approximation in [42], respectively. Moreover, the proposed transform exhibited similar coding performance with the
exact DTT and outperformed previous approximations [42] according to computational experiments with popular visual image compression standards. In terms of video coding, the results
from the proposed tool were virtually indistinguishable from the
ones furnished by the approximation in [42]. Thus, the new tool
outperform the competing methods both in computational cost
and coding performance. The proposed method was embedded
into the JPEG standard and the standard software library for
H.264/AVC video coding. Obtained results showed negligible
degradation when compared to the standard DCT-based compression methods in both cases. The 2-D versions were realized
in FPGA using ROACH2 hardware platform and ASIC place
and route was realized using Cadence encounter with AMS standard cells and the results show a 43.1% reduction in the number
of CLB for the FPGA realization and a 57.7% reduction in areatime figure for the ASIC place and route realization when compared with the exact DTT. The proposed design could excel in
providing high operation frequency and very low power consumption. Therefore, the proposed approximation offers low computational complexity while maintaining good coding performance.
Systems that operate under low processing constraints and require video streaming can benefit of the proposed low-complexity
codecs and low-power hardware. In particular, applications in
the following contexts meet such requirements that need lowcomplexity [72]: environmental monitoring, habitat monitoring,
surveillance, structural monitoring, equipment diagnostics, disaster management, and emergency response [73].
Figure 5 displays the first encoded frame of two standard video
sequences at low bitrate (200 kbps). The compressed frames resulting from the original and modified codecs are visually indistinguishable.
6
Hardware Section
To compare the hardware resource consumption of the proposed approximate DTT against the exact DTT fast algorithm
proposed in [15], algorithm proposed in [42] and the Loeffler
DCT [61], the 2-D version of both algorithms were initially modeled and tested in Matlab Simulink and then were physically
realized on a Xilinx Virtex-6 XC6VSX475T-1FF1759 Reconfigurable Open Architecture Computing Hardware-2 (ROACH2)
board [71]. The ROACH2 board consists of a Xilinx Virtex 6
FPGA, 16 complex analog-to-digital converters (ADC), multigigabit transceivers and a 72-bit DDR3 RAM.
The 1-D versions were initially modeled and the 2-D versions
were generated using two 1-D designs along with a transpose
buffer. Designs were verified using more than 10000 test vectors
with complete agreement with theoretical values. Results are
shown in Table 3. Metrics, including configurable logic blocks
(CLB) and flip-flop (FF) count, critical path delay (Tcpd , in ns),
and maximum operating frequency (Fmax , in MHz) are provided.
The percentage reduction in the number of CLBs and FFs were
43.2% and 25.0%, respectively, compared with the exact DTT
fast algorithm proposed in [15]. In is important to emphasize
that the approximation in [42] is asymmetric; the forward and inverse transform possess different structures, being the inverse operation more complex (cf. Table 2). For comparisons, we adopt
the average measurement between forward and inverse realizations. The proposed approximation could provide higher maximum operating frequency with improvements of 85.9%, 43.5%,
and 9.7% when compared to the Loeffler DCT [61], the exact
DTT [15], and the design in [42], respectively.
The ASIC realization was done by porting the hardware description language code to 0.18 um CMOS technology and was
subjected to synthesis and place-and-route according to the
Cadence Encounter Digital Implementation (EDI) for AMS libraries. Libraries for the best case scenario were employed in
getting the place-and-route results with gate voltage of 1.8 V.
The adopted figures of merit for the ASIC synthesis were: area
(A) in mm2 , area-time complexity (AT ) in mm2 · ns, area-timesquared complexity (AT 2 ) in mm2 · ns2 , dynamic (Dp ) power
consumption in mW/MHz, critical path delay (Tcpd ) in ns, and
maximum operating frequency (Fmax ) in MHz. Results are displayed in Table 4. The figures of merit AT and AT 2 had percentage reductions of 57.7% and 57.4% when compared with the
exact DTT. Thus, the proposed design could attain reductions
of 17.3%, 20.6%, and 82.1% for area, AT 2 , Tcpd , and dynamic
power consumption, respectively, when compared to [42].
7
Acknowledgment
Arjuna Madanayake thanks the Xilinx University Program
(XUP) for the Xilinx Virtex-6 Sx475 FPGA device installed in
on the ROACH2 board.
References
[1] A. F. Nikiforov, S. K. Suslov, and V. B. Uvarov, Classical Orthogonal
Polynomials of a Discrete Variable, ser. Springer Series in Computational Physics. Springer Berlin Heidelberg, 1991.
[2] P. D. Dragnev and E. B. Saff, “Constrained energy problems
with applications to orthogonal polynomials of a discrete variable,”
Journal d’Analyse Mathematique, vol. 72, pp. 223–259, 1997.
[Online]. Available: http://dx.doi.org/10.1007/BF02843160
[3] M. Câmara, J. Fábrega, M. A. Fiol, and E. Garriga, “Some families of
orthogonal polynomials of a discrete variable and their applications to
graphs and codes,” The Electronic Journal of Combinatorics, vol. 16,
pp. 1–30, 2009.
[4] H. Zhu, M. Liu, H. Shu, H. Zhang, and L. Luo, “General form for obtaining discrete orthogonal moments,” IET Image Processing, vol. 4,
pp. 335–352, Oct 2010.
[5] A. Goshtasby, “Template matching in rotated images,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 7, no. 3,
pp. 338–344, May 1985.
[6] M. I. Heywood and P. D. Noakes, “Fractional central moment method
for movement-invariant object classification,” IEE Proceedings–
Vision, Image and Signal Processing, vol. 142, no. 4, pp. 213–219,
Aug 1995.
Discussion and Conclusion
In this work, a low-complexity near-orthogonal 8-point DTT
approximation suitable for image and video coding was proposed. A fast algorithm for proposed DTT approximation which
requires only 24 additions and six bit-shifting operations was
also introduced. This fast algorithm can be used for both forward and near inverse transformations. The additive arithmetic
cost of the proposed approximation is 45.5% and 2.05% lower
when compared with the exact DTT fast algorithm and the
[7] V. Markandey and R. I. P. de Figueiredo, “Robot sensing techniques
based on high-dimensional moment invariants and tensors,” IEEE
Transactions on Robotics and Automation, vol. 8, no. 2, pp. 186–
195, Apr 1992.
[8] R. Mukundan, S. H. Ong, and R. A. Lee, “Image analysis by
Tchebichef moments,” IEEE Transactions on Image Processing,
vol. 10, pp. 1357–1364, 2001.
[9] L. Leida, Z. Hancheng, Y. Gaobo, and Q. Jiansheng, “Referenceless
measure of blocking artifacts by Tchebichef kernel analysis,” IEEE
Signal Processing Letters, vol. 21, no. 1, pp. 122–125, Jan 2014.
8
(a) Foreman, H.264
(b) Foreman, Approx. [42]
(c) Foreman, Proposed
Figure 5: First frame of the compressed ‘Foreman’ sequence, with a target bitrate of 200 kbps.
Table 3: Hardware resource consumption using Xilinx Virtex-6 XC6VSX475T 1FF1759 device
Method
CLB
FF
Tcpd (ns)
Fmax (MHz)
Exact DTT [15]
2941
7271
7.688
130.07
Approximation in [42]
1515
6058
5.596
178.69
Inverse Approximation in [42]
1713
4834
6.184
161.71
Loeffler DCT [61]
3250
4413
9.956
100.44
Proposed DTT
1671
5455
5.356
186.70
Table 4: Hardware resource consumption for CMOS 0.18 um ASIC place and route
Area (mm2 )
AT
AT 2
Tcpd (ns)
Fmax (MHz)
Dp (mW/MHz)
Exact DTT [15]
0.872
3.84
16.92
4.405
227.01
0.182
Approximation in [42]
0.237
1.34
7.52
5.635
177.46
0.171
Inverse Approximation in [42]
0.323
1.79
9.89
5.536
180.64
0.724
Loeffler DCT [61]
0.684
4.20
25.85
6.148
162.65
1.961
Proposed
0.366
1.62
7.20
4.434
225.53
0.080
FRS Method
9
[10] J.-L. Rose, C. Revol-Muller, D. Charpigny, and C. Odet, “Shape prior
criterion based on Tchebichef moments in variational region growing,”
in 2009 16th IEEE International Conference on Image Processing
(ICIP), Nov 2009, pp. 1081–1084.
[30] ——, “ITU-T recommendation H.263 version 1: Video coding for low
bit rate communication,” Technical Report, ITU-T, 1995.
[31] I. Richardson, The H.264 Advanced Video Compression Standard,
2nd ed. John Wiley and Sons, 2010.
[11] H. Zhang, X. Dai, P. Sun, H. Zhu, and H. Shu, “Symmetric image
recognition by Tchebichef moment invariants,” in 2010 17th IEEE
International Conference on Image Processing (ICIP), Sept 2010,
pp. 2273–2276.
[32] G. J. Sullivan, J. Ohm, W.-J. Han, and T. Wiegand, “Overview of the
high efficiency video coding (HEVC) standard,” IEEE Transactions
on Circuits and Systems for Video Technology, vol. 22, pp. 1649–
1668, 2012.
[12] Q. Li, H. Zhu, and Q. Liu, “Image recognition by combined affine
and blur Tchebichef moment invariants,” in 2011 4th International
Congress on Image and Signal Processing (CISP), vol. 3, Oct 2011,
pp. 1517–1521.
[33] F. Bossen, B. Bross, K. Suhring, and D. Flynn, “HEVC complexity
and implementation analysis,” IEEE Transactions on Circuits and
Systems for Video Technology, vol. 22, no. 12, pp. 1685–1696, Dec
2012.
[13] H. Huang, G. Coatrieux, H. Shu, L. Luo, and C. Roux, “Blind integrity verification of medical images,” IEEE Transactions on Information Technology in Biomedicine, vol. 16, no. 6, pp. 1122–1126,
Nov 2012.
[34] Google
Inc.,
“VP9,”
The
http://www.webmproject.org/vp9/, 2015.
[14] S. Ishwar, P. K. Meher, and M. N. S. Swamy, “Discrete Tchebichef
transform–a fast 4 × 4 algorithm and its application in image/video
compression,” in 2008 IEEE International Symposium on Circuits
and Systems (ISCAS), 2008, pp. 260–263.
[15] S. Prattipati, S. Ishwar, P. K. Meher, and M. N. S. Swamy, “A fast
8×8 integer Tchebichef transform and comparison with integer cosine
transform for image compression,” in 2013 IEEE 56th International
Midwest Symposium on Circuits and Systems (MWSCAS), 2013, pp.
1294–1297.
[16] N. A. Abu, S. L. Wong, N. Herman, and R. Mukundan, “An efficient
compact Tchebichef moment for image compression,” in 2010 10th
International Conference on Information Sciences Signal Processing and their Applications (ISSPA), May 2010, pp. 448–451.
[17] Q. Li and H. Zhu, “Block-based compressed sensing of image using directional Tchebichef transforms,” in 2012 IEEE International Conference on Systems, Man, and Cybernetics (SMC), Oct 2012, pp.
2207–2212.
[18] R. K. Senapati, U. C. Pati, and K. K. Mahapatra, “Reduced memory,
low complexity embedded image compression algorithm using hierarchical listless discrete Tchebichef transform,” IET Image Processing,
vol. 8, no. 4, pp. 213–238, Apr 2014.
[19] N. Ahmed, T. Natarajan, and K. R. Rao, “Discrete cosine transform,”
IEEE Transactions on Computers, vol. C-23, no. 1, pp. 90–93, Jan.
1974.
WebM
Project,
[35] T. I. Haweel, “A new square wave transform based on the DCT,”
Signal Processing, vol. 81, pp. 2309–2319, 2001. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0165168401001062
[36] R. J. Cintra and F. M. Bayer, “A DCT approximation for image
compression,” IEEE Signal Processing Letters, vol. 18, no. 10, pp.
579–582, Oct 2011.
[37] F. M. Bayer and R. J. Cintra, “DCT-like transform for image compression requires 14 additions only,” Electronics Letters, vol. 48, no. 15,
pp. 919–921, July 2012.
[38] R. J. Cintra, F. M. Bayer, and C. J. Tablada, “Low-complexity
8-point DCT approximations based on integer functions,”
Signal Processing, vol. 99, pp. 201–214, 2014. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0165168413005161
[39] S. Bouguezel, M. O. Ahmad, and M. N. S. Swamy, “Low-complexity
8×8 transform for image compression,” Electronics Letters, vol. 44,
no. 21, pp. 1249–1250, Oct 2008.
[40] ——, “A low-complexity parametric transform for image compression,” in 2011 IEEE International Symposium on Circuits and Systems (ISCAS), May 2011, pp. 2145–2148.
[41] ——, “Binary discrete cosine and hartley transforms,” IEEE Transactions on Circuits and Systems I: Regular Papers, vol. 60, no. 4,
pp. 989–1002, Apr 2013.
[42] P. A. M. Oliveira, R. J. Cintra, F. M. Bayer, S. Kulasekera, and
A. Madanayake, “A discrete Tchebichef transform approximation for
image and video coding,” IEEE Signal Processing Letters, vol. 22,
no. 8, pp. 1137–1141, Aug 2015.
[20] F. Ernawan, N. A. Abu, and N. Suryana, “TMT quantization table
generation based on psychovisual threshold for image compression,”
in 2013 International Conference of Information and Communication Technology (ICoICT), Mar 2013, pp. 202–207.
[43] H. Bateman and A. Erdélyi, Higher transcendental functions.
McGraw-Hill,
1953,
vol.
2.
[Online].
Available:
http://books.google.com.br/books?id=p lQAAAAMAAJ
[21] R. K. Senapati, U. C. Pati, and K. K. Mahapatra, “A low complexity
embedded image coding algorithm using hierarchical listless DTT,”
in 2011 8th International Conference on Information, Communications and Signal Processing (ICICS), Dec 2011, pp. 1–5.
[44] H. S. Malvar, A. Hallapuro, M. Karczewicz, and L. Kerofsky, “Lowcomplexity transform and quantization in H.264/AVC,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 13, no. 7,
pp. 598–603, Jul 2003.
[22] F. Ernawan, E. Noersasongko, and N. A. Abu, “An efficient 2×2
Tchebichef moments for mobile image compression,” in 2011 International Symposium on Intelligent Signal Processing and Communications Systems (ISPACS), Dec 2011, pp. 1–5.
[45] R. Blahut, Fast Algorithms for Signal Processing.
versity Press, 2010.
[23] L. W. Chew, L.-M. Ang, and K. P. Seng, “Survey of image compression algorithms in wireless sensor networks,” in 2008 International
Symposium on Information Technology (ITSim), vol. 4, Aug 2008,
pp. 1–9.
[47] J. W. Eaton, D. Bateman, S. Hauberg, and R. Wehbring, GNU Octave version 3.8.0 Documentation, 3rd ed. Free Software Foundation, Inc., Feb 2011.
[24] M. Guo, M. H. Ammar, and E. W. Zegura, “V3: a vehicle-tovehicle live video streaming architecture,” in 2005 3rd IEEE International Conference on Pervasive Computing and Communication
(PerCom), Mar 2005, pp. 171–180.
[25] D. H. Friedman, “Streaming implementation of video algorithms on
a low-power parallel architecture,” in 2013 IEEE Global Conference
on Signal and Information Processing (GlobalSIP), Dec 2013, pp.
650–653.
[26] K. Nakagaki and R. Mukundan, “A fast 4×4 forward discrete
Tchebichef transform algorithm,” IEEE Signal Processing Letters,
vol. 14, pp. 684–687, 2007.
[27] G. K. Wallace, “The JPEG still picture compression standard,” IEEE
Transactions on Consumer Electronics, vol. 38, no. 1, pp. xviii–
xxxiv, Feb 1992.
[28] International Organisation for Standardisation, “Generic coding of
moving pictures and associated audio information part 2: Video,
ISO/IEC JTC1/SC29/WG11 coding of moving pictures and audio,”
1994.
[29] International Telecommunication Union, “ITU-T recommendation
H.261 version 1: Video codec for audiovisual services at p × 64 kbits,”
Technical Report, ITU-T, 1990.
10
Cambridge Uni-
[46] MATLAB, “version 8.1 (R2013a) documentation,” Natick, MA, 2013.
[48] Python, “version 2.7.6 documentation,” Delaware, US, 2015.
[49] G. A. F. Seber, A Matrix Handbook for Statisticians, ser. Wiley
Series in Probability and Mathematical Statistics.
Hoboken, NJ:
John Wiley and Sons, Inc., 2008.
[50] B. N. Flury and W. Gautschi, “An algorithm for simultaneous
orthogonal transformation of several positive definite symmetric
matrices to nearly diagonal form,” SIAM Journal on Scientific and
Statistical Computing, vol. 7, no. 1, pp. 169–184, Jan. 1986. [Online].
Available: http://dx.doi.org/10.1137/0907013
[51] V. K. Goyal, “Theoretical foundations of transform coding,” IEEE
Signal Processing Magazine, vol. 18, no. 5, pp. 9–21, Sept 2001.
[52] J. Katto and Y. Yasuda, “Performance evaluation of subband coding and optimization of its filter coefficients,”
Journal of Visual Communication and Image Representation,
vol. 2,
pp. 303–313,
1991. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/1047320391900114
[53] V. Britanak, P. C. Yip, and K. R. Rao, Discrete Cosine
and Sine Transforms. Academic Press, 2007. [Online]. Available:
http://books.google.com.br/books?id=iRlQHcK-r kC
[54] C. J. Tablada, F. M. Bayer, and R. J. Cintra, “A class of DCT approximations based on the Feig–Winograd algorithm,” Signal Processing,
vol. 113, pp. 38–51, 2015.
[55] R. J. Cintra, H. M. Oliveira, and C. O. Cintra, “The rounded Hartley
transform,” in Proceedings of the IEEE International Telecommunications Symposium–ITS’2002, Sept 2002, pp. 1357–1364.
[64] L. Zhang and H. Li, “SR-SIM: A fast and high performance IQA
index based on spectral residual,” in 2012 19th IEEE International
Conference on Image Processing (ICIP), Sep 2012, pp. 1473–1476.
[56] W. B. Pennebaker and J. L. Mitchell, JPEG: Still Image Data Compression Standard, ser. Chapman & Hall digital multimedia standards series. Springer, 1993.
[65] Z. Wang and A. C. Bovik, “Reduced- and no-reference image quality
assessment,” IEEE Signal Processing Magazine, vol. 28, no. 6, pp.
29–40, Nov 2011.
[57] Z. Wang and A. C. Bovik, “Mean squared error: Love it or leave
it? a new look at signal fidelity measures,” IEEE Signal Processing
Magazine, vol. 26, no. 1, pp. 98–117, Jan 2009.
[66] S. M. Kay, Fundamentals of Statistical Signal Processing, Volume I:
Estimation Theory, ser. Prentice Hall Signal Processing Series. Upper Saddle River, NJ: Prentice-Hall, 1993, vol. 1.
[58] C.-K. Fong and W.-K. Cham, “LLM integer cosine transform and its
fast algorithm,” IEEE Transactions on Circuits and Systems for
Video Technology, vol. 22, no. 6, pp. 844–854, Jun 2012.
[67] R. Pandit, N. Khosla, G. Singh, , and H. Sharma, “Image compression and quality factor in case of JPEG image format,” International
Journal of Advanced Research in Computer and Communication
Engineering, vol. 2, pp. 2578–2581, Jul 2013.
[59] F. M. Bayer, R. J. Cintra, A. Madanayake, and U. S. Potluri, “Multiplierless approximate 4-point DCT VLSI architectures for transform
block coding,” Electronics Letters, vol. 49, no. 24, pp. 1532–1534,
Nov 2013.
[60] A. V. Oppenheim and R. W. Schafer, Discrete-time signal processing,
3rd ed., ser. Prentice-Hall signal processing series.
Prentice Hall,
2010.
[61] C. Loeffler, A. Ligtenberg, and G. S. Moschytz, “A practical fast 1D DCT algorithms with 11 multiplications,” in IEEE International
Conference on Acoustics, Speech, and Signal Processing, vol. 2, May
1989, pp. 988–991.
[62] University
of
Southern
California,
Signal
and
Image
Processing
Institute,
“The
USC-SIPI
image
database,”
http://sipi.usc.edu/database/, 2015.
[63] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,”
IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600–612,
Apr 2004.
11
[68] x264 team, “x264,” http://www.videolan.org/developers/x264.html,
2015.
[69] S. Gordon, D. Marpe, and T. Wiegand, “Simplied use of 8×8
transform–updated proposal and results,” Joint Video Team (JVT)
of ISO/IEC MPEG and ITU-T VCEG, doc. JVT–K028, Munich, Germany, Mar 2004.
[70] “Xiph.org Video Test Media,” https://media.xiph.org/video/derf/,
2015.
[71] (2015) ROACH2. https://casper.berkeley.edu.
[72] I. F. Akyildiz, T. Melodia, and K. R. Chowdhury, “A survey on wireless multimedia sensor networks,” Computer Networks, vol. 51, pp.
921–960, 2007.
[73] N. Kimura and S. Latifi, “A survey on data compression in wireless
sensor networks,” in 2005 International Conference on Information
Technology: Coding and Computing (ITCC), vol. 2, Apr 2005, pp.
8–13.
| 8cs.DS
|
ON THE FINITENESS OF BASS NUMBERS OF LOCAL
COHOMOLOGY MODULES AND COMINIMAXNESS
KAMAL BAHMANPOUR, REZA NAGHIPOUR∗ AND MONIREH SEDGHI
arXiv:1309.0431v1 [math.AC] 2 Sep 2013
DEDICATED TO PROFESSOR HYMAN BASS
Abstract. In this paper, we continue the study of cominimaxness modules with respect
to an ideal of a commutative Noetherian ring (cf. [2]), and Bass numbers of local cohomology modules. Let R denote a commutative Noetherian local ring and I an ideal of
R. We first show that the Bass numbers µ0 (p, HI2 (R)) and µ1 (p, HI2 (R)) are finite for all
p ∈ Spec R, whenever R is regular. As a consequence, it follows that the Goldie dimension of HI2 (R) is finite. Also, for a finitely generated R-module M of dimension d, it is
shown that the Bass numbers of HId−1 (M ) are finite if and only if ExtiR (R/I, HId−1 (M ))
be minimax for all i ≥ 0. Finally, we prove that if dim R/I = 2, then the Bass numbers
of HIn (M ) are finite if and only if ExtiR (R/I, HIn (M )) be minimax, for all i ≥ 0, where
n is a non-negative integer.
1. Introduction
Throughout this paper, let R denote a commutative Noetherian ring (with identity)
and I an ideal of R. For an R-module M, the ith local cohomology module of M with
respect to I is defined as
HIi (M) = lim ExtiR (R/I n , M).
−→
n≥1
We refer the reader to [10] and [6] for more details about local cohomology. An important
problem in commutative algebra is to determine when the Bass numbers of the ith local
cohomology module HIi (M) is finite. In [12] Huneke conjectured that for any ideal I in a
regular local ring (R, m, k), the Bass numbers
i
(Rp ))
µi (p, HIi (R)) = dimk(p) ExtiRp (k(p), HIR
p
are finite for all i and j, where k(p) := Rp /pRp . In particular the injective resolution of
the local cohomology has only finitely many copies of the injective hull of R/p for any p.
There is evidence that this conjecture is true. It is shown that by Huneke and Sharp
[13] and Lyubeznik [16, 17] that the conjecture holds for a regular local ring containing
Key words and phrases. Bass numbers, cominimax modules, Krull dimension, local cohomology, minimax modules.
2000 Mathematics Subject Classification: 13D45, 14B15, 13E05.
This research was in part supported by a grant from IPM (No. 89130053, 89130048).
∗
Corresponding author: e-mail: [email protected] (Reza Naghipour).
1
2
KAMAL BAHMANPOUR, REZA NAGHIPOUR AND MONIREH SEDGHI
a field. We remark that the Bass numbers might be infinite if R is not regular. For
example, if I := (x, y)R ⊆ R := k[x, y, z, w]/(xz − yw), then µ0 (m, HI2 (R)) = ∞ for
m := (x, y, z, w)R (see [11]).
We say that M is a minimax module if there is a finitely generated submodule N of M,
such that M/N is Artinian. The interesting class of minimax modules was introduced by
H. Zöshinger in [23] and he has in [23] and [24] given many equivalent conditions for a
module to be minimax. Finally, the R-module M is said to be an I-cominimax if support
of M is contained in V (I) and ExtiR (R/I, M) is minimax for all i ≥ 0. The concept of
the I-cominimax modules were introduced in [2] as a generalization of important notion
of I-cofinite modules.
The main subject of the paper is to continue the study of I-cominimaxness properties
and the finiteness of Bass numbers of local cohomology modules. First, we provide a
partial answer to Huneke’s conjecture. Namely, it will be shown that the Bass numbers of
µ0 (p, HI2 (R)) and µ1 (p, HI2 (R)) are finite for all ideals I in a regular local ring R and for
all p ∈ Spec R. Pursuing this point of view further we establish the R-module HI2 (R) has
finite Goldie dimension. Recall that an R-module M is said to have finite Goldie dimension
if M does not contain an infinite direct sum of non-zero submodules, or equivalently the
injective hull E(M) of M decompose as a finite direct sum of indecomposable (injective)
submodules.
Since the Bass numbers of the local cohomology modules HIi (M) are not finite in
general, several attempts made to find some conditions for the ideal I to have finiteness
for the Bass numbers of the local cohomology modules with support in I in terms of
I-cominimaxness properties. The following result is in this way:
Theorem 1.1. Let (R, m) be a local (Noetherian) ring, I an ideal of R and M a finitely
generated R-module of dimension d. Then the Bass numbers of HId−1 (M)/Γm (HId−1 (M))
are finite.
One of our tools for proving Theorem 1.1 is the following:
Proposition 1.2. Let (R, m), I and M be as in Theorem 1.1. Then Hm1 (HId−1 (M)) is an
Artinian R-module, where d = dim M.
Pursuing this point of view further we derive the following consequence of Theorem 1.1.
Theorem 1.3. Let (R, m) be a local (Noetherian) ring, I an ideal of R and M a finitely
generated R-module of dimension d. Then the following conditions are equivalent:
(i) The Bass numbers of HId−1 (M) are finite.
(ii) Soc(HId−1 (M)) is finitely generated.
(iii) HId−1 (M) is I-cominimax.
As the second main result of this paper we obtain a characterization of the finiteness
of Bass numbers of ith local cohomology modules of M with respect to I of dimension 2,
i.e. dim R/I = 2. More precisely we shall show that:
Theorem 1.4. Let (R, m) be a local (Noetherian) ring and let I be an ideal of R with
dim R/I = 2. Let M be a finitely generated R-module and i a non-negative integer. Then
the Bass numbers of HIi (M) are finite if and only if the R-module HIi (M) is I-cominimax.
BASS NUMBERS AND COMINIMAXNESS
3
The proof of Theorem 1.4 is given in 2.17. As an application, we derive the following
consequences of Theorem 1.4, which is a characterization of I-cominimaxness of HIi (M)
in terms of finiteness of Bass numbers of local cohomology modules HIi (M) for certain
ideal I of R.
Theorem 1.5. Let (R, m) be a local (Noetherian) ring, and let I be an ideal of R such
that dim R/I = 2. Let M be a finitely generated R-module and t a non-negative integer.
Then the following conditions are equivalent:
(i) Soc HIi (M) is finitely generated for all i ≤ t.
(ii) ExtjR (R/m, HIi (M)) is finitely generated for all j ≥ 0 and all i ≤ t − 1.
(iii) HIi (M) is I-cominimax for all i ≤ t − 1.
Using Theorem 1.5 we obtain some results as following:
Corollary 1.6. Let (R, m) be a regular local ring of dimension d ≥ 2 containing a field
and let I be an ideal of R with dim R/I = 2. Then the R-module HIi (R) is I-cominimax
for all i ≥ 0.
Corollary 1.7. Let R be a Noetherian ring, M a finitely generated R-module and I an
ideal of R such that dim M/IM = 2. Let
Σ := {p ∈ Spec R : µj (p, HIi (M)) = ∞ for some integers i ≥ 0 and j ≥ 0}.
Then Σ is countable and Σ ⊆ Max(R).
Corollary 1.8. Let (R, m) be a local (Noetherian) ring, M a finitely generated R-module
and I an ideal of R such that dim M/IM = 3. Then the set
Σ := {p ∈ Spec R : µj (p, HIi (M)) = ∞ for some integers i ≥ 0 and j ≥ 0},
is countable.
Throughout this paper, R will always be a commutative Noetherian ring with non-zero
identity and I will be an ideal of R. For each R-module L, we denote by AsshR L (resp.
mAssR L) the set {p ∈ AssR L : dim R/p = dim L} (resp. the set of minimal primes of
AssR L). We shall use Max R to denote the set of all maximal ideals of R. Also, for any
ideal a of R, we denote {p ∈ Spec R : p ⊇ a} by V (a). Finally, for any ideal b of R, the
radical of b, denoted by Rad(b), is defined to be the set {x ∈ R : xn ∈ b for some n ∈ N}.
For any unexplained notation and terminology we refer the reader to [6] and [18].
2. The Results
Let (R, m) be a local (Noetherian) ring and I an ideal of R. As noted in the introduction,
it is well known that the Bass numbers µi (p, HI2 (R)) are finite whenever R is a regular
ring containing a field. Our first result shows that if R is regular, then µi (p, HI2 (R)) is
finite for i ∈ {0, 1}.
Theorem 2.1. Let (R, m) be a regular local ring and I an ideal of R. Then the Bass
numbers µ0 (p, HI2 (R)) and µ1 (p, HI2 (R)) are finite for all p ∈ Spec R.
4
KAMAL BAHMANPOUR, REZA NAGHIPOUR AND MONIREH SEDGHI
i
Proof. In view of [5, Proposition 2.20] the Rp -modules ExtjRp (κ(p), HIR
(Rp )) are finitely
p
generated for all j ≥ 0 and i = 0, 1. Hence it follows from [15, Corollary 3.5] that
µ0 (p, HI2 (R)) and µ1 (p, HI2(R)) are finite.
As a corollary of the previous theorem, we deduce that the Goldie dimension of HI2 (R)
is finite.
Corollary 2.2. Let (R, m) be a regular local ring and I an ideal of R. Then the R-module
HI2 (R) has finite Goldie dimension.
Proof. As AssR (HI2 (R)) is finite, by [4, Theorem 2.4], it follows from Theorem 2.1 and
the definition that HI2 (R) has finite Goldie dimension.
For an arbitrary R-module M, the next reult gives us a necessary and sufficient condition for the finiteness of µi (p, M), the Bass numbers of M, in terms of the minimaxness
of ExtiR (R/I, M). Recall that for any module L over a local ring (R, m), the socle of L,
denoted by Soc(L), is defined to be the R-module HomR (R/m, L).
Theorem 2.3. Let (R, m) be a local (Noetherian) ring and I an ideal of R such that
dim R/I = 1. Suppose that M is an R-module and n ≥ 0 an integer. Then the following
conditions are equivalent:
(i) µi (p, M) is finite for all p ∈ V (I) and for all i ≤ n.
(ii) ExtiR (R/I, M) is minimax for all i ≤ n.
Proof. First we show (i)=⇒(ii). Let q ∈ mAssR (R/I) and i ≤ n. Then q 6= m, and so
Rad(q + Rx) = m for all x ∈ m\q. Hence it follows from [7, Proposition 1] that the Rmodule ExtiR (R/q + Rx, M) is finitely generated for all i ≤ n. (Note that ExtiR (R/m, M)
is finitely generated for all i ≤ n.) Now, for x ∈ m\q the exact sequence
x
0 → R/q −→ R/q −→ R/q + Rx → 0,
provides the following exact sequence:
x
ExtiR (R/q + Rx, M) −→ ExtiR (R/q, M) −→ ExtiR (R/q, M),
which implies that the R-module Soc(ExtiR (R/q, M)) is finitely generated, for all i ≤ n.
On the other hand, if L := ExtiR (R/q, M), then as by assumption the Rq -module Lq is
finitely generated, it follows that there exists a finitely generated submodule K of L, such
that (L/K)q = 0, and so Supp(L/K) ⊆ V (m). Finally, the exact sequence
0 → K −→ L −→ L/K −→ 0.
induces the exact sequence
Soc(L) −→ Soc(L/K) −→ Ext1R (R/m, K).
Hence the R-module Soc(L/K) is finitely generated. Consequently in view of [19, Proposition 4.1] the R-module L/K is Artinian, that is L is a minimax R-module. Now, the
assertion follows from [2, Corollary 2.8].
In order to prove (ii)=⇒(i), use [2, Corollary 2.8] and the definition of the minimax
BASS NUMBERS AND COMINIMAXNESS
modules.
5
Corollary 2.4. Let R, I and M be as in Theorem 2.3. Then the following conditions are
equivalent:
(i) µi (p, M) is finite for all i ≥ 0 and all primes p ∈ V (I).
(ii) ExtiR (R/I, M) is minimax for all i ≥ 0.
Proof. The result follows from Theorem 2.3.
Corollary 2.5. Let (R, m) be a local (Noetherian) ring. Let I and J be tow ideals of R
such that dim R/I ≤ 1 and dim R/J = 1. Then the R-modules ExtiR (R/J, HIj (R)) are
minimax for all integers i, j ≥ 0.
Proof. The assertion follows from [5, Corollary 2.10] and Theorem 2.3.
Before we state the next corollary, recall that for an ideal I of a commutative ring R,
the ideal transform DI (R) of I defined by
DI (R) := lim HomR (I n , R).
−→
n≥1
See [6, Section 2.2] for the basic properties of ideal transforms.
Corollary 2.6. Let (R, m) be a regular local ring and let I, J be two ideals of R such that
dim R/I = 1. Then
(i) For any integer i ≥ 0, the R-modules ExtiR (R/I, DJ (R)) and ExtiR (R/I, HJ1 (R)) are
minimax.
(ii) The R-modules HomR (R/I, HJ2 (R)) and Ext1R (R/I, HJ2 (R)) are minimax.
(iii) For all integers i, j ≥ 0, the R-module ExtiR (R/I, HJj (R)) is minimax, whenever R
contains a field.
Proof. (i) follows from [5, Proposition 2.20] and Theorem 2.3. In order to prove (ii) use
Theorems 2.3 and 2.1. Finally, (iii) follows from Corollary 2.4 and [13, 17].
In order to state next results, we need here some preliminary results about the Artinianess of local cohomology modules.
Theorem 2.7. Let (R, m) be a local (Noetherian) ring and I an ideal of R such that
dim R/I = 2. Then for any finitely generated R-module M, the R-module Hm1 (HIi (M)) is
Artinian for all i ≥ 0.
Proof. Since dim R/I = 2, there exists x ∈ m such that dim R/I + Rx = 1. Let J :=
I + Rx. As Hm1 (HI0 (M)) is Artinian, we may assume that i ≥ 1. Then by [21, Corollary
1.4], there exists the following exact sequence,
0 → HJ1 (HIi−1 (M)) → HJi (M) → HJ0 (HIi (M)) → 0.
6
KAMAL BAHMANPOUR, REZA NAGHIPOUR AND MONIREH SEDGHI
Now, as dim HJ1 (HIi−1 (M)) ≤ 1, it follows from this exact sequence that the sequence
Hm1 (HJi (M)) → Hm1 (HJ0 (HIi (M))) → 0,
is exact, and so by [5, Corollary 2.16] the R-module Hm1 (HJ0 (HIi (M))) is Artinian. On the
other hand, since dim R/J = 1 there exists y ∈ m such that J + Ry is m-primary and so
1
HJ+Ry
(HJ0 (HIi (M))) = Hm1 (HJ0 (HIi (M))),
is Artinian R-module. Moreover, using again [21, Corollary 1.4] it follows that the sequences:
0 → HJ1 (HIi (M)) → HJi+1 (M) → HJ0 (HIi+1(M)) → 0,
1
0
0 → HJ+Ry
(HJ0 (HIi (M))) → Hm1 (HIi (M)) → HJ+Ry
(HJ1 (HIi (M))) → 0,
0
are exact. Now, since by [5, Corollary 2.16], HJ+Ry
(HJi+1 (M)) is Artinian, it follows that
0
HJ+Ry
(HJ1 (HIi (M))) is Artinian. Consequently the R-module Hm1 (HIi (M)) is Artinian, as
required.
Theorem 2.8. Let (R, m) be a local (Noetherian) ring, I an ideal of R and M a finitely
generated R-module of dimension d ≥ 1. Then Hm1 (HId−1 (M)) is an Artinian R-module.
Proof. We use induction on d. When d = 1, there is nothing to prove. Now suppose
that d > 1 and the result has been proved for non-zero finitely generated R-modules of
dimension d − 1. By replacing M by M/ΓI (M) we may assume that M is a non-zero
finitely generated I-torsion-free R-module. Then, by [6, Lemma 2.1.1], the ideal I contains
an element x which is a non-zerodevisor on M. Hence the exact sequence
x
0 −→ M −→ M −→ M/xM −→ 0
induces an exact sequence
x
x
· · · −→ HIj (M) −→ HIj (M) −→ HIj (M/xM) −→ HIj+1(M) −→ HIj+1(M) −→ · · · .
Therefore we have the following exact sequence,
0 −→ HIj (M)/xHIj (M) −→ HIj (M/xM) −→ (0 :H j+1 (M ) x) −→ 0,
I
from dim M/xM = d − 1 that HId−1 (M)/xHId−1 (M) is
inductive hypothesis, the R-module Hm1 (HId−2 (M/xM)) is
[20, Corollary 3.3], Supp HId−2 (M/xM) is finite. Hence
and so it follows
Moreover, by the
Next, in view of
dim HId−2 (M/xM) ≤ 1. Therefore the exact sequence
Artinian.
Artinian.
we have
0 −→ HId−2 (M)/xHId−2 (M) −→ HId−2 (M/xM) −→ (0 :H d−1 (M ) x) −→ 0,
I
provides the exact sequence
Hm1 (HId−2 (M)/xHId−2 (M)) −→ Hm1 (HId−2 (M/xM)) −→ Hm1 ((0 :H d−1 (M ) x)) −→ 0.
I
Hence the R-module
Hm1 ((0 :H d−1 (M )
I
x)) is Artinian. Now, from the exact sequences
f
0 → (0 :H d−1 (M ) x) → HId−1 (M) → xHId−1 (M) → 0,
I
BASS NUMBERS AND COMINIMAXNESS
7
g
0 → xHId−1 (M) → HId−1 (M) → HId−1 (M)/xHId−1 (M) → 0,
we obtain the following exact sequences
H 1 (f )
m
Hm1 ((0 :H d−1 (M ) x)) −→ Hm1 (HId−1 (M)) −→
Hm1 (xHId−1 (M)),
I
H 1 (g)
m
Hm0 (HId−1 (M)/xHId−1 (M)) −→ Hm1 (xHId−1 (M)) −→
Hm1 (HId−1 (M)).
Then Ker(Hm1 (f )) and Ker(Hm1 (g)) are Artinian (note that Hm1 ((0 :H d−1 (M ) x)) and
I
HId−1 (M)/xHId−1 (M) are Artinian), and the sequence
0 → ker(Hm1 (f )) → ker(Hm1 (g) ◦ Hm1 (f )) → ker(Hm1 (g)),
is exact. Since ker(Hm1 (g ◦ f )) = (0 :Hm1 (H d−1 (M )) x) and ker(Hm1 (g) ◦ Hm1 (f )) is Artinian, it
I
follows that (0 :Hm1 (H d−1 (M )) x) is also Artinian. Whence according to Melkersson’s TheI
orem [19, Proposition 1.4] the R-module Hm1 (HId−1 (M)) is Artinian, and this completes
the inductive step.
Theorem 2.9. Let (R, m) be a local (Noetherian) ring, I an ideal of R and M a finitely
generated R-module of dimension d. Then the Bass numbers of HId−1 (M)/Γm (HId−1 (M))
are finite.
Proof. In view of Theorem 2.8, the R-module Hm1 (HId−1 (M)/Γm (HId−1 (M))) is Artinian.
On the other hand, since Supp HId−1 (M) is finite (see [20, Corollary 3.3]), it follows that
dim HId−1 (M) ≤ 1, and so for i ≥ 2, Hmi (HId−1 (M)/Γm (HId−1 (M))) = 0. Therefore for
all i 6= 1 we have Hmi (HId−1 (M)/Γm (HId−1 (M))) = 0. Consequently, in view of [19,
Corollary 3.10] the R-modules ExtiR (R/m, HId−1 (M)/Γm (HId−1 (M))) are finitely generated. (Note that a module is m-cofinite if and only if it is Artinian.) Furthermore, because
of dim HId−1 (M) ≤ 1, it follows that dim R/p = 1 for all p ∈ Supp HId−1 (M)/Γm (HId−1 (M))
with p 6= m. Since
(HId−1 (M)/Γm (HId−1 (M)))p ∼
= (HId−1 (M))p ,
d−1
for all p ∈ Supp HId−1 (M)/Γm (HId−1 (M)) with p 6= m, it follows that HIR
(Mp ) 6= 0, and
p
d−1
so dim Mp = d − 1. Consequently HIRp (Mp ) is an Artinian Rp -module. Hence for each
i ≥ 0 and each p ∈ SpecR, the ith Bass number
µi (p, HId−1 (M)/Γm (HId−1 (M))) = dimκ(p) ExtiRp (κ(p), (HId−1 (M)/Γm (HId−1 (M)))p ),
is finite, where κ(p) = Rp /pRp .
The following lemma and proposition will be sueful in the proof of the next main result
of this paper. An R-module M is said to be weakly Laskerian if the set of associated
primes of any quotient module of M is finite (cf. [8, 9]).
Lemma 2.10. Let R be a Noetherian ring and I an ideal of R. Then, for any R-module
T , the following conditions are equivalent:
(i) ExtnR (R/I, T ) is weakly Laskerian for all n ≥ 0.
8
KAMAL BAHMANPOUR, REZA NAGHIPOUR AND MONIREH SEDGHI
(ii) For any finitely generated R-module N with support in V (I), ExtnR (N, T ) is weakly
Laskerian for all n ≥ 0.
Proof. For proving the assertion we use the proof of [14, Lemma 1] and [9, Lemma 2.2].
Proposition 2.11. Let (R, m) be a local (Noetherian) ring and M a finitely generated R-module. Let I ⊆ J be two ideals of R such that dim R/I = 2. Then, for all
i, j ≥ 0, there exists a finitely generated submodule L of ExtjR (R/J, HIi (M)) such that
Supp(ExtjR (R/J, HIi (M))/L) ⊆ V (m).
Proof. In view of [5, Corollary 3.3] and Lemma 2.10 the set AssR ExtjR (R/J, HIi (M)) is
finite. Let
AssR ExtjR (R/J, HIi (M))\{m} = {p1 , . . . , pn }.
Since pk 6= m for all k = 1, . . . , n and height m/I = 2, it follows that height pk /I ≤ 1,
and so dim Rpk /IRpk ≤ 1 for all k = 1, . . . , n. Hence by [5, Corollary 2.7] the Rpk modules of (ExtjR (R/I, HIi (M)))pk are finitely generated for all k = 1, . . . , n. Therefore,
by [14, Lemma 1], the Rpk -modules of (ExtjR (R/J, HIi (M)))pk are finitely generated for
all k = 1, . . . , n. Consequently for every k = 1, . . . , n, there exists a finitely generated
submodule Tk of ExtjR (R/J, HIi (M)) such that (Tk )pk = (ExtjR (R/J, HIi (M)))pk . Let K :=
T1 + · · · + Tn . Then K is a finitely generated submodule of ExtjR (R/J, HIi (M)) and
Supp(ExtjR (R/J, HIi (M))/K) ∩ {p1 , . . . , pn } = ∅.
So that dim ExtjR (R/J, HIi (M))/K ≤ 1. Now if Supp(ExtjR (R/J, HIi (M))/K) ⊆ {m},
then the result follows for L = K. We may therefore assume that
dim ExtjR (R/J, HIi (M))/K = 1.
Let AssR ExtjR (R/J, HIi (M))/K\{m} = {q1 , . . . , qs }. Then, as the above, the Rqt -module
of (ExtjR (R/J, HIi (M))/K)qt is finitely generated for t = 1, . . . , s. Hence there exists a
finitely generated submodule Lt /K of ExtjR (R/J, HIi (M))/K such that
(ExtjR (R/J, HIi (M))/K)qt = (Lt )qt .
Let L := L1 +· · ·+Ls +K. Then L is a finitely generated submodule of ExtjR (R/J, HIi (M))
and
Supp(ExtjR (R/J, HIi (M))/L) ∩ {q1 , ..., qs } = ∅.
Since Supp ExtjR (R/J, HIi (M))/L ⊆ Supp ExtjR (R/J, HIi (M))/K, it follows that
Supp ExtjR (R/J, HIi (M))/L ⊆ {m},
as required.
We are now ready to state and prove the second main theorem of the paper, which is a
characterization of the finiteness of the Bass numbers of (d − 1)th local cohomology of M
with respect to an arbitrary ideal I in terms of the I-cominimaxness of HId−1 (M), where
BASS NUMBERS AND COMINIMAXNESS
9
d = dim M. Before we state theorem, let us recall that a sequence x1 , . . . , xn of element
of an ideal I of R is called an I-filter regular M-sequence, provided:
Supp((x1 , . . . , xi−1 )M :M xi )/(x1 , . . . , xi−1 )M) ⊆ V (I),
for all i = 1, . . . , n. In the case of (R, m) a local ring and I = m, we recover filter regular
sequences as studied in several papers (cf. e.g. [22]).
Theorem 2.12. Let (R, m) be a local (Noetherian) ring, I an ideal of R and M a finitely
generated R-module of dimension d. Then the following conditions are equivalent:
(i) The Bass numbers of the R-module HId−1 (M) are finite.
(ii) Soc(HId−1 (M)) is finitely generated.
(iii) HId−1 (M) is I-cominimax.
Proof. The conclusion (i)=⇒(ii) is obviously true. In order to prove (ii)=⇒ (i), let
Soc HId−1 (M) is a finitely generated R-module. Then, according to Melkersson’s Theorem [19, Proposition 1.4], Γm (HId−1 (M)) is Artinian. Therefore the Bass numbers of
Γm (HId−1 (M)) are finite. Now, it follows from Theorem 2.9 that the Bass numbers of
HId−1 (M) are finite, as required.
In order to prove the implication (iii)=⇒(ii), it follows from [2, Corollary 2.8] that the Rmodule ExtjR (R/m, HId−1 (M)) is minimax for all j ≥ 0. Now it easily seen from definition
that ExtjR (R/m, HId−1 (M)) is of finite length for all j ≥ 0. In particular the R-module
HomR (R/m, HId−1 (M)) is finitely generated, as required.
Finally for the proof of (i)=⇒ (iii), let dim R/I = n. For n ≤ 1 this follows by [5,
Corollary 2.7]. Hence we may assume that n ≥ 2. Let x1 , ..., xn−1 ∈ m be a filter
regular sequence on R/I. Let Ks := I + (x1 , . . . , xs ) for s = 0, 1, . . . , n − 1. We first
show by induction on t, that the R-module ExtjR (R/Kn−t−1 , HId−1 (M)) is minimax for all
j ≥ 0. Since x1 + I, . . . , xn−1 + I is a part of system of parameters for R/I, it follows
that dim R/Kn−1 = 1, and so the result is evidently true for t = 0, by Theorem 2.3.
Suppose that n − 1 ≥ t > 0 and the case t − 1 is settled. Since Γm (R/Kn−t−1 ) is a
finitely generated R-module with Supp(Γm (R/Kn−t−1 )) ⊆ {m}, it follows that from [14,
Lemma 1] and hypothesis (i) that the R-module ExtjR (Γm (R/Kn−t−1 ), HId−1 (M)) is finitely
generated for all j ≥ 0. Now, as Γm (R/Kn−t−1 ) = J/Kn−t−1 for some ideal J of R, from
the exact sequence
0 −→ J/Kn−t−1 −→ R/Kn−t−1 −→ R/J −→ 0,
we get the following long exact sequence,
j−1
· · · −→ ExtR
(J/Kn−t−1 , HId−1 (M)) −→ ExtjR (R/J, HId−1 (M)) −→
−→ ExtjR (R/Kn−t−1 , HId−1 (M)) −→ ExtjR (J/Kn−t−1 , HId−1 (M)) −→ · · · .
Hence we get that the R-module ExtjR (R/Kn−t−1 , HId−1(M)) is minimax if and only if the
R-module ExtjR (R/J, HId−1 (M)) is minimax for all j ≥ 0. Now, as x1 , . . . , xn−1 is a filter
regular sequence on R/I in m, it follows from definition that
[
xn−t 6∈
p.
p∈AssR (R/Kn−t−1 )\{m}
10
KAMAL BAHMANPOUR, REZA NAGHIPOUR AND MONIREH SEDGHI
Therefore, it follows from
AssR (R/J) = AssR (R/Kn−t−1 )\{m}
that xn−t 6∈ ∪p∈AssR R/J p, and so xn−t is an R/J-regular element in m. Consequently, the
exactness
xn−t
0 −→ R/J −→ R/J −→ R/J + Rxn−t −→ 0,
implies that the sequence
j−1
j−1
0 −→ ExtR
(R/J, HId−1 (M))/xn−t ExtR
(R/J, HId−1(M)) −→
−→ ExtjR (R/J + Rxn−t , HId−1 (M)) −→ (0 :Extj (R/J,H d−1 (M )) xn−t ) −→ 0,
R
I
is exact. Since Supp(R/J + Rxn−t ) ⊆ V (Kn−t ), (note that Kn−t ⊆ J + Rxn−t ,) the
induction hypothesis and [14, Lemma 1] yield that the R-module
ExtjR (R/J + Rxn−t , HId−1 (M)),
is minimax, for all j ≥ 0. Therefore the R-module (0 :Extj (R/J,H d−1 (M )) xn−t ) is minimax.
R
I
Now, let p ∈ Supp HId−1 (M)\{m}. Then, by [19, Proposition 5.1] and Vanishing theorem
the Rp -module (HId−1 (M))p , is IRp -cofinite. As I ⊆ J ⊆ J + Rxn−t , it follows from [14,
Lemma 1] that the Rp -module (ExtjR (R/J, HId−1 (M)))p is finitely generated. Moreover,
as in view of [20, Corollary 3.3] the set Supp HId−1 (M) is finite, it follows that the set
AssR ExtjR (R/J, HId−1 (M)) is finite. Therefore, by the proof of Lemma 2.11, there exists
a finitely generated submodule L of ExtjR (R/J, HId−1 (M)) such that
Supp ExtjR (R/J, HId−1 (M))/L ⊆ {m}.
Next, from the exact sequence
0 −→ L −→ ExtjR (R/J, HId−1 (M)) −→ ExtjR (R/J, HId−1 (M))/L −→ 0,
we get the exact sequence
(0 :Extj (R/J,H d−1 (M )) xn−t ) −→ (0 :Extj (R/J,H d−1 (M ))/L xn−t ) −→ Ext1R (R/Rxn−t , L).
R
I
R
I
Hence, it yields that the R-module (0 :Extj (R/J,H d−1 (M ))/L xn−t ) is minimax. Since
R
I
Supp(0 :Extj (R/J,H d−1 (M ))/L xn−t ) ⊆ {m}, it follows from the definition that the R-module
R
I
(0 :Extj (R/J,H d−1 (M ))/L xn−t ) is Artinian. As (0 :Extj (R/J,H d−1 (M ))/L xn−t ) is Rxn−t −torsion,
R
I
R
I
according to Melkersson [19, Proposition 1.4] ExtjR (R/J, HId−1 (M))/L is an Artinian Rmodule. That is ExtjR (R/J, HId−1(M)) is a minimax R-module, as required.
Corollary 2.13. Let (R, m) be a regular local ring of dimension d ≥ 1 containing a field
and let I be an ideal of R. Then the R-module HId−1 (R) is I-cominimax .
Proof. The result follows from Theorem 2.12 and the fact that the Bass numbers of the
R-module HId−1 (R) are finite (see [13, 17]).
BASS NUMBERS AND COMINIMAXNESS
11
Corollary 2.14. Let (R, m) be a local (Noetherian) ring and I an ideal of R. Let M be
a finitely generated R-module of dimension d. Then the Bass numbers of the R-module
HId−1 (M) are finite, whenever m 6∈ AssR HId−1 (M).
Proof. Since AssR (Soc HId−1 (M)) = V (m) ∩ AssR HId−1 (M), it follows from
m 6∈ AssR (HId−1 (M))
that AssR (Soc HId−1 (M)) = ∅. Hence Soc HId−1 (M) = 0, and so the result now follows
from Theorem 2.12.
Corollary 2.15. Let (R, m) be a local (Noetherian) ring. Let I be an ideal of R and x ∈ m
such that I ⊆ Rx. Then for each non-zero finitely generated R-module M of dimension
d ≥ 1, the R-module HId−1 (M) is I-cominimax. In particular, the Bass numbers of
HId−1 (M) are finite.
Proof. When d = 1 there is nothing to prove. In the case that d = 2, it follows from [3,
Theorem 2.3] that the R-module HomR (R/I, HI1 (M)) is finitely generated and so the Rmodule Soc(HI1 (M)) is finitely generated. Now the assertion follows from Theorem 2.12.
x
Finally, for d ≥ 3, it follows from [1, Lemma 2.5] that the map HId−1 (M) −→ HId−1 (M)
is an isomorphism, and so Soc(HId−1 (M)) = 0. Now the assertion again follows from
Theorem 2.12.
Corollary 2.16. Let (R, m) be a regular local ring and I an ideal of R such that
height(I) = 1. Then for each non-zero finitely generated R-module M of dimension
d ≥ 1, the R-module HId−1 (M) is I-cominimax. In particular, the Bass numbers of
HId−1 (M) are finite.
Proof. Since R is a UFD and height(I) = 1, it follows that I is contained in a principal
prime ideal. Now the assertion follows from Corollary 2.15.
We are now ready to show that the subjects of the previous results can be used to
prove a characterization of the finiteness of the Bass numbers of ith local cohomology of
M with respect to an ideal I of dimension 2, i.e., where dim R/I = 2, in terms of the
I-cominimaxness of HIi (M). The main result is Theorem 2.19. The following theorem
will serve to shorten the proof of the main theorem.
Theorem 2.17. Let (R, m) be a local (Noetherian ) ring and let I be an ideal of R with
dim R/I = 2. Let M be a finitely generated R-module and i a non-negative integer. Then
the following conditions are equivalent:
(i) The Bass numbers of HIi (M) are finite.
(ii) The R-module HIi (M) is I-cominimax.
Proof. First we show the implication (i)=⇒ (ii). Suppose that Γm (R/I) := J/I. Then
Supp J/I ⊆ {m}, and so it follows from [14, Lemma 1] and the assumption (i) that the
12
KAMAL BAHMANPOUR, REZA NAGHIPOUR AND MONIREH SEDGHI
R-module ExtjR (J/I, HIi (M)) is finitely generated for all j ≥ 0. Now, the exact sequence
0 −→ J/I −→ R/I −→ R/J −→ 0
induces a long exact sequence
j−1
· · · −→ ExtR
(J/I, HIi (M)) −→ ExtjR (R/J, HIi (M)) −→
−→ ExtjR (R/I, HIi (M)) −→ ExtjR (J/I, HIi (M)) −→ · · · .
Therefore HIi (M) is I-cominimax if and only if the R-module ExtjR (R/J, HIi (M)) is minimax for all j ≥ 0.
It is easy to see that Γm (R/J) = 0. Hence, by [6, Lemma 2.1.1] m contains an R/Jregular element x. Then dim R/J/x(R/J) = dim R/J − 1, and so dim R/J + Rx = 1.
Furthermore, the exact sequence
x
0 −→ R/J −→ R/J −→ R/J + Rx −→ 0
gives us an exact sequence
j−1
j−1
0 −→ ExtR
(R/J, HIi (M))/x ExtR
(R/J, HIi (M)) −→
−→ ExtjR (R/J + Rx, HIi (M)) −→ (0 :Extj (R/J,H i (M )) x) −→ 0
R
I
Now, since dim R/J +Rx = 1, it follows from the assumption (i) and Theorem 2.3 that the
R-module ExtjR (R/J + Rx, HIi (M)) is minimax, for all j ≥ 0. On the other hand, in view
of Proposition 2.11, there exists a finitely generated submodule L of ExtjR (R/J, HIi (M))
such that Supp(ExtjR (R/J, HIi (M))/L) ⊆ {m}. Now, from the exact sequence
0 −→ L −→ ExtjR (R/J, HIi (M)) −→ ExtjR (R/J, HIi (M))/L −→ 0
we obtain the exact sequence
(0 :Extj (R/J,H i (M )) x) −→ (0 :Extj (R/J,H i (M ))/L x) −→ Ext1R (R/Rx, L),
R
I
I
R
which implies that the R-module (0 :Extj (R/J,H i (M ))/L x) is minimax. As
I
R
Supp(0 :Extj (R/J,H i (M ))/L x) ⊆ {m},
R
I
it follows from the definition that the R-module (0 :Extj (R/J,H i (M ))/L x) is Artinian. Since
I
R
(0 :Extj (R/J,H i (M ))/L x) is Rx-torsion, it follows from Melkersson’s result (see [19, ProposiR
I
tion 1.4]) that ExtjR (R/J, HIi (M))/L is an Artinian R-module, and so ExtjR (R/J, HIi (M))
is a minimax R-module, as required.
For prove (ii)=⇒(i) first observe that the R-module ExtjR (R/m, HIi (M)) is minimax,
for all j ≥ 0, and so by the definition, ExtjR (R/m, HIi (M)) has finite length. Therefore µj (m, HIi (M)) is finite, for all j ≥ 0. Now, let p ∈ Supp HIi (M)\V (m). Then
dim Rp /IRp ≤ 1 as easily seen. Therefore the Rp -module of HIi (M)p is IRp -cofinite,
and so the Bass numbers of HIi (M)p are finite. Since the Bass numbers are stable under the localization, it follows that the Bass numbers of HIi (M) are finite, as required.
BASS NUMBERS AND COMINIMAXNESS
13
The next result is an immediate consequence of Theorem 2.17 and Huneke-SharpLyubeznik’ results.
Corollary 2.18. Let (R, m) be a regular local ring of dimension d ≥ 2 containing a field
and let I be an ideal of R with dim R/I = 2. Then the R-module HIi (R) is I-cominimax
for all i ≥ 0.
Proof. The result follows from Theorem 2.17 and the fact that the Bass numbers of the
R-module HIi (R) are finite (see [13, 17]), for all i ≥ 0.
We are now ready to state and prove the third main theorem of the paper, which is a
characterization of the finiteness of the Bass numbers of ith local cohomology of M with
respect to an ideal I of dimension 2, in terms of the I-cominimaxness of HIi (M).
Theorem 2.19. Let (R, m) be a local (Noetherian) ring, and let I be an ideal of R such
that dim R/I = 2. Let M be a finitely generated R-module and t a non-negative integer.
Then the following conditions are equivalent:
(i) SocHIi (M) is finitely generated for all i ≤ t.
(ii) ExtjR (R/m, HIi (M)) is finitely generated for all j ≥ 0 and for all i ≤ t − 1.
(iii) HIi (M) is I-cominimax, for all i ≤ t − 1.
Proof. The implication (ii)=⇒ (i) follows from [15, Corollary 3.5]. In order to prove (i)=⇒
i
(ii), as Ass
4.3]), we can find an element x ∈ m, such
SR HI (M) is finite (see [5, Corollary
St
that x 6∈ p∈Assh(R/I) p, and V (Rx) ∩ ( i=0 AssR HIi (M)) ⊆ V (m).
Now, let L := I + Rx. Then it is easy to see that dim R/L = 1, and
0
HL0 (HIi (M)) = HRx
(HIi (M)) = Hm0 (HIi (M)),
is an Artinian R-module for all i ≤ t, by [19, Proposition 4.1] and assumption (i). Now,
for all i ≤ t − 1, there exists the short exact sequence
1
0
0 −→ HRx
(HIi (M)) −→ HLi+1 (M) −→ HRx
(HIi+1 (M)) −→ 0,
(cf. for instance [21, Corollary 3.5]). Since HLi+1 (M) is L-cofinite and
0
HRx
(HIi+1 (M)) = Hm0 (HIi+1 (M))
1
is Artinian, it follows from [5, Lemma 2.1] that ExtjR (R/L, HRx
(HIi (M))) is minimax for
k
all j ≥ 0 and i ≤ t − 1. Furthermore, since HRx
(HIi (M)) = HLk (HIi (M)), for all k ≥ 0,
it follows that HLk (HIi (M)) = 0 for all k ≥ 2. Hence ExtjR (R/L, HLk (HIi (M))) is minimax
for all j ≥ 0, k ≥ 0 and i ≤ t − 1, and so in view of [19, Proposition 3.9], the R-module
ExtjR (R/L, HIi (M)) is minimax for all j ≥ 0 and i ≤ t − 1. Consequently, by [2, Theorem
2.7] the R-module Ki, j := ExtjR (R/m, HIi (M)) is minimax for all j ≥ 0 and i ≤ t − 1.
Hence there is a finitely generated submodule Ni, j of Ki, j , such that Ki,j /Ni, j is Artinian.
As mKi, j = 0 and m(Ki, j /Ni, j ) = 0, it follows that Ki, j has finite length, as required.
For proving (ii)=⇒(iii), first we use the proof of implication (ii)=⇒(i) in Theorem
2.17, to see that the Bass numbers of HIi (M) are finite, for all i ≤ t − 1. Now, the
14
KAMAL BAHMANPOUR, REZA NAGHIPOUR AND MONIREH SEDGHI
assertion follows from Theorem 2.17. Finally the implication (iii)=⇒(ii) follows again
from Theorem 2.17.
We end the paper by drawing some concequences of Theorem 2.19. The first consequence will help us to construct a certain subset of Spec R which is countable.
Corollary 2.20. Let (R, m) be a local (Noetherian) ring and let I be an ideal of R with
dim R/I = 2. Let M be a finitely generated R-module such that m 6∈ ∪i≥0 AssR (HIi (M)).
Then ExtjR (R/m, HIi (M)) is finitely generated for all j ≥ 0 and all i ≥ 0.
Proof. Since m 6∈ ∪i≥0 AssR (HIi (M)), it follows that SocHIi (M) = 0 for all i ≥ 0. Now
the assertion follows from Theorem 2.19.
Proposition 2.21. Let R be a Noetherian ring and I an ideal of R such that dim R/I = 2.
For a finitely generated R-module M, let
Σ := {p ∈ Spec(R) : µj (p, HIi (M)) = ∞ for some integers i ≥ 0 and j ≥ 0}.
Then Σ is countable and Σ ⊆ Max(R).
Proof. In order to show that Σ ⊆ Max(R), let p ∈ Σ\Max(R). Then it follows from the
definition that p ∈ V (I), and it is easy to see that height(p/I) ≤ 1. Hence it follows from
[5, Corollary 2.10] that p 6∈ Σ, which is a contradiction. Thus Σ ⊆ Max(R).
i
n
On the other hand, as for all integers n ≥ 1 and i ≥ 0, the sets AssS
R ExtR (R/I , M)
are finite and HIi (M) = lim ExtiR (R/I n , M), it follows that the set i≥0 AssR HIi (M)
−→
n≥1
is countable. Now, in order to complete the proof, it is enough to show that Σ ⊆
S
Ass H i (M). To do this suppose that the contrary is true. Then there is m ∈
i≥0
S R I i
Σ\ i≥0 AssR HI (M). Hence, by [5, Corollary 2.10] we have height(m/I) = 2, and so it
follows from Corollary 2.20 that m 6∈ Σ, which is a contradiction.
Corollary 2.22. Let R be a Noetherian ring, M a finitely generated R-module and I an
ideal of R such that dim M/IM = 2 . Let
Σ := {p ∈ Spec(R) : µj (p, HIi (M)) = ∞ for some integers i ≥ 0 and j ≥ 0}.
Then Σ is countable and Σ ⊆ Max(R).
Proof. As dim M/IM = 2 it follows from that dim R/I + AnnR M = 2. Now the result
i
(M) and Proposition 2.21.
follows from HIi (M) ∼
= HI+Ann
R (M )
Corollary 2.23. Let (R, m) be a local (Noetherian) ring, and let I be an ideal of R such
that dim R/I = 3. Let M be a finitely generated R-module. Then the set
Σ := {p ∈ Spec(R) : µj (p, HIi (M)) = ∞ for some integers i ≥ 0 and j ≥ 0},
is countable.
BASS NUMBERS AND COMINIMAXNESS
15
Proof. Let p ∈ Σ\{m}. Then by [5,SCorollary 2.10] we have height(p/I)
S = 2, and so it
follows from Corollary 2.20 that p ∈ i≥0 AssR HIi (M). Thus Σ\{m} ⊆ i≥0 AssR HIi (M),
and so the set Σ is countable.
Corollary 2.24. Let (R, m) be a local (Noetherian) ring, M a finitely generated R-module
and I an ideal of R such that dim M/IM = 3. Then the set
Σ := {p ∈ Spec(R) : µj (p, HIi (M)) = ∞ for some integers i ≥ 0 and j ≥ 0}
is countable.
Proof. As dim M/IM = 3 it follows that dim R/I +AnnR (M) = 3. Now the result follows
i
from HIi (M) ∼
(M) and Corollary 2.23 .
= HI+Ann
R (M )
Acknowledgments
The authors are deeply grateful to the referee for a very careful reading of the manuscript
and many valuable suggestions. We also would like to thank Professors Hossein Zakeri
and Kamran Divaani-Aazar for their careful reading of the first draft and many helpful
suggestions.
References
[1] N. Abazari and K. Bahmanpour, On the finiteness of Bass numbers of local cohomology modules, J.
Alg. Appl. 10(2011), 783-791.
[2] J. Azami, R. Naghipour and B. Vakili, Finiteness properties of local cohomology modules for aminimax modules, Proc. Amer. Math. Soc. 137(2009), 439-448.
[3] K. Bahmanpour and R. Naghipour, On the cofiniteness of local cohomology modules, Proc. Amer.
Math. Soc. 136(2008), 2359-2363.
[4] K. Bahmanpour and R. Naghipour, Associated primes of local cohomology modules and Matlis duality, J. Algebra, 320(2008), 2632-2641.
[5] K. Bahmanpour and R. Naghipour, Cofiniteness of local cohomology modules for ideals of small
dimension, J. Algebra, 321(2009), 1997-2011.
[6] M.P. Brodmann and R.Y. Sharp, Local cohomology; an algebraic introduction with geometric applications, Cambridge University Press, Cambridge,1998.
[7] D. Delfino and T. Marley, Cofinite modules and local cohomology, J. Pure and Appl. Algebra
121(1997), 45-52.
[8] K. Divaani-Aazar and A. Mafi, Associated primes of local cohomology modules, Proc. Amer. Math.
Soc. 133(2005), 655-660.
[9] K. Divaani-Aazar and A. Mafi, Associated primes of local cohomology modules of weakly Laskerian
modules, Comm. Algebra 34(2006), 681-690.
[10] A. Grothendieck, Local cohomology, Notes by R. Hartshorne, Lecture Notes in Math., 862, Springer,
New York, 1966.
[11] R. Hartshorne, Affine duality and cofiniteness, Invent. Math. 9(1970), 145-164.
[12] C. Huneke, Problems on local cohomology, Free resolutions in commutative algebra and algebraic
geometry, Res. Notes Math. 2(1992), 93-108.
[13] C. Huneke and R.Y. Sharp, Bass numbers of local cohomology modules, Trans. Amer. Soc. 339(1993),
765-779.
[14] K.I. Kawasaki, On the finiteness of Bass numbers of local cohomology modules, Proc. Amer. Math.
Soc. 124(1996), 3275-3279.
16
KAMAL BAHMANPOUR, REZA NAGHIPOUR AND MONIREH SEDGHI
[15] K. Khashyarmanesh, On the finiteness properties of extension and tortion functors of local cohomology modules, Proc. Amer. Math. Soc. 135 (2007), 1319-1327.
[16] G. Lyubeznik, Finiteness properties of local cohomology modules (an application of D-modules to
commutative algebra), Invent. Math. 113(1993), 41-55.
[17] G. Lyubeznik, F-modules: applications to local cohomology modules and D-modules in characteristic
p > 0, J. Reine Angew. Math. 491(1997), 65-130.
[18] H. Matsumura, Commutative ring theory, Cambridge Univ. Press, Cambridge, UK, 1986.
[19] L. Melkersson, Modules cofinite with respect to an ideal, J. Algebra, 285(2005), 649-668.
[20] R. Naghipour and M. Sedghi, A characterization of Cohen-Macaulay modules and local cohomology,
Arch. Math., 87(2006), 303-308.
[21] P. Schenzel, Proregular sequences, local cohomology, and completion, Math. Scand. 92 (2003), 161180.
[22] P. Schenzel, N.V. Trung and N.T. Cuong, Verallgemeinerte Cohen-Macaulay-Moduln, Math. Nachr.
85 (1978), 57-73.
[23] H. Zöschinger, Minimax moduln, J. Algebra. 102 (1986), 1-32.
[24] H. Zöschinger, Ü ber die maximalbedingung für radikalvolle untermoduln, Hokkaido Math. J. 17
(1988), 101-116.
Department of Mathematics, Ardabil Branch, Islamic Azad University, Ardabil, Iran.
E-mail address: [email protected]
Department of Mathematics, University of Tabriz, Tabriz, Iran; and School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O. Box: 19395-5746,
Tehran, Iran.
E-mail address: [email protected]
E-mail address: [email protected]
Department of Mathematics, Azarbaijan University of Shahid Madani, Tabriz, Iran.
E-mail address: [email protected]
| 0math.AC
|
NUMERICAL INTEGRATION ON GRAPHS:
WHERE TO SAMPLE AND HOW TO WEIGH
arXiv:1803.06989v1 [math.ST] 19 Mar 2018
GEORGE C. LINDERMAN AND STEFAN STEINERBERGER
Abstract. Let G = (V, E, w) be a finite, connected graph with weighted edges. We are interested in the problem of finding a subset W ⊂ V of vertices and weights aw such that
X
1 X
f (v) ∼
aw f (w)
|V | v∈V
w∈W
for functions f : V → R that are ‘smooth’ with respect to the geometry of the graph. The
main application are problems where f is known to somehow depend on the underlying graph
but is expensive to evaluate on even a single vertex. We prove an inequality showing that
the integration problem can be rewritten as a geometric problem (‘the optimal packing of heat
balls’). We discuss how one would construct approximate solutions of the heat ball packing
problem; numerical examples demonstrate the efficiency of the method.
1. Introduction
1.1. Introduction. The purpose of this paper is to report on a very general idea in the context
of sampling on graphs. It is a companion paper to [16] dealing with a problem in Spectral Graph
Theory. We extend some of the ideas from [15, 16] to sampling, prove an inequality bounding the
integration error in terms of the geometry of the sampling points and give several examples. We
will, throughout the paper, use G = (V, E, w) to denote a connected graph with weighted edges.
Problem (Quadrature). If we are allowed to sample in a set W with |W | = k
vertices, which vertices and weights aw should we pick so that
X
1 X
f (v) ∼
aw f (w)
|V |
v∈V
w∈W
for functions f : V → R that are ‘smooth with respect to to the geometry’ of G?
In many cases, this question does not make a lot of sense: since the graph is finite, one can simply
compute the true average of f by summing over all n vertices. This question is only interesting
whenever sampling f is difficult or expensive. A toy example is the following: suppose we have
a medical database of n 1 people containing all sorts of information and are interested in the
average blood pressure (not in the database) of those n people. Actually going out and measuring
the blood pressure of all n people would take a very long time. However, it is known that blood
pressure is strongly correlated with some of the factors we have on file (say, age, weight, and
smoking habits) and weakly connected or unconnected to others (say, eye color). We can then
build a weighted graph on n vertices where the weight on the edge connecting two people depends
on how similar they are with regards to relevant factors – the hope is that blood pressure, as a
function on the graph, is then smoothly varying. Which of the, say, n/1000 people should we
do a blood pressure measurement on so that the sample average is representative of the global
average? Is there a way to pick them in a way that decreases the expected error over a pure random
selection? It is likely possible to make stronger statements if one restricts to certain classes of
graphs and functions and this could be of substantial interest; our paper will only address the
most general case.
2010 Mathematics Subject Classification. 05C50, 05C70, 35P05, 65D32.
Key words and phrases. Graph, Sampling, Graph Laplacian, Sampling, Heat Kernel, Packing.
1
2
1.2. Formal Setup. We will now make these notions precise. As is not surprising, the precise
definition of ‘smoothness’ of the function is crucial. If f has no particular structure, then there
is little hope of being able to achieve anything at all. However, a weighted graph does induce a
natural notion of smoothness: we want the function to vary little over edges with a large weight
(since vertices connected by a large weight should be ‘similar’) whereas a large variation over edges
with a smaller weight would not be that surprising. We now introduce a notion of a Laplacian on
a Graph; it slightly deviates from more classical notions and could be of independent interest (see
Section §2.4 for a more detailed discussion). Let A denote the (weighted) adjacency matrix of G
(
w(eij )
if i ∼E j
Aij =
0
otherwise,
where w(eij ) ≥ 0 is the weight of the edge eij = eji connecting vertices i and j. A is a symmetric
matrix. We introduce the maximum sum of any of the rows of this symmetric matrix (coinciding,
by symmetry, with the maximum sum of any of the columns) and use it to define a normalized
adjacency matrix: more precisely, we have
dmax = max
1≤i≤n
n
X
Aij
and set
j=1
A0 =
1
A.
dmax
Finally, we introduce the (diagonal) degree matrix D0 associated to the renormalized adjacency
matrix and use it to define a Laplacian: we set
0
Dii
=
1
n
X
dmax
j=1
Aij
and define
L = A0 − D 0 .
We will never work directly with the Laplacian: our main object of interest is the associated
diffusion process whose generator is given by
P = L + Idn×n ,
where Idn×n is the identity matrix of size n. P is a symmetric stochastic matrix and represents
a lazy random walk where the probability of “staying put” depends on the vertex (as opposed
to being, say, 0.5 as in the classical lazy random walk). We denote the eigenvalues of P , which
are merely the eigenvalues of the Laplacian L shifted by 1, by λ1 , . . . , λn . Since P is a stochastic
matrix, we have |λi | ≤ 1. The eigenvectors whose eigenvalues are close to 0 ‘diffuse quickly’ and are
thus the natural high-frequency objects on the graph. This motivates an ordering of eigenvalues
from low frequency to high frequency
1 = |λ1 | ≥ |λ2 | ≥ |λ3 | ≥ · · · ≥ |λn | ≥ 0.
We denote the corresponding orthogonal eigenvectors of P , which clearly coincide
√ with the eigenvectors of the Laplacian L, by φ1 , . . . , φn ∈ Rn (normalized in `2 , φ1 = 1/ n is the constant
vector). We define a function space Xλ , the canonical analogue of trigonometric polynomials on
Td or spherical harmonics on Sd , via
X
X
2
Xλ = f : V → R f =
|hf, φk i| ,
hf, φk i
with norm
kf k2Xλ =
|λk |≥λ
|λk |≥λ
where 0 ≤ λ ≤ 1 is a parameter controlling the degree of smoothness. If λ > µ, then Xλ ⊂ Xµ .
Moreover, X0 = {f : V → R} contains all functions while, at least on generic graphs, X1 contains
only the constant functions (depends on whether the induced random walk is ergodic; this is
not important for our purposes). The norm is just the classical L2 −norm on the subspace – in
more classical terms, we are simply considering the L2 −space obtained via a Littlewood-Paley
projection. This function space Xλ is natural: if the graph G approximates a torus Td , then this
function space will indeed approximate trigonometric functions. If G is close to a discretization
of a sphere Sd , then the space Xλ approximates the space of low-degree spherical harmonics.
3
2. Main Results
2.1. Main Result. Our main result bounds the integration error in terms of kf kXλ and a purely
geometric quantity (explained in detail below) formulated in terms of the quadrature points and
independent of f . This has the nice effect of multiplicatively separating the size of the function
kf kXλ and a quantity that can be interpreted as the ’quality’ of the quadrature scheme.
Theorem. Let W ⊂ V be equipped with weights aw summing to 1. Then, for ` ∈ N, 0 < λ < 1,
12
2
X
X
X
1
kf kXλ
1
(Idn×n + L)`
− .
aw δw
f (v) −
aw f (w) ≤
∀ f ∈ Xλ
n
λ`
n
v∈V
w∈W
w∈W
L2
The proof shows that the inequality is essentially close to sharp if there is a spectral gap close
to λ. If there is no spectral gap, there is little reason to assume that the inequality is sharp. If
one has access to the eigenfunctions of the Laplacian, then it is possible to use a much stronger
inequality discussed below in §2.4. We observe that the inequality only holds for f ∈ Xλ : this
does not mean that the integration scheme will perform poorly on functions outside this function
class. Indeed, it is a fairly common technique to design quadrature rules for a certain class of
very smooth functions – they will usually still be somewhat effective on functions that are not as
smooth as the functions used in the design of the rule. This is, of course, the main idea in the
construction of spherical designs [6, 14, 16] but also appears in other contexts: the Simpson rule for
the integration of a real function on an interval is designed to be exact on quadratic polynomials.
2.2. Geometric interpretation: Heat Ball Packing. Altogether, this suggests that we should
use the quantity on the right-hand side in the inequality in Theorem 1, depending only on the set
W , the weights aw and the free parameter ` (but not on the function f ) as a guideline for how to
construct the quadrature rule. This motivates studying the minimization problem
2
min min (Idn×n + L)
W ⊂V
|W |=k
aw
`
X
w∈W
aw δw
subject to
L2
X
aw = 1.
w∈W
Note that we do not actually need to find the precise minimizer of the solution, indeed, we just
want the quantity to be small. The topological structure of the minimization problem suggests
that there will be many local minimizer that are not global: this has been empirically observed in
related problems in the continuous setting (see e.g. [7]) and there is little reason to assume the
discrete case will be different. Moreover, if the continuous analogies continue to hold, then the
actual numerical value of minimizing configuration can be expected to be quite close to that of
almost-minimizers which is good news: there are many almost-minimizers (thus easy to find) and
they are almost as good as the global minimizer. (Idn×n + L)` δw is the probability distribution
of a random walker starting in w after ` jumps which lends itself to a geometric interpretation.
Guideline. If we manage to find a placement of vertices W with the property
that the random walks, weighted by aw , overlap very little, then we have found a
good quadrature rule on the graph G = (V, E).
We observe that this can be reinterpreted as a ‘packing problem for heat balls’. This principle
has already appeared naturally in the continuous setting of Riemannian manifolds in work of the
second author [15]. It is even meaningful if we can only sample in a single vertex: we should pick
the most central vertex and that is equivalent to heat diffusing quickly to many other points. In
the ideal setting with perfectly distributed heat balls all the weights would be identical (as can
be seen in examples with lots of symmetry, see [16]); this nicely mirrors an unofficial guideline in
numerical integration stating that how far the weights deviate from constant weights can be used
as measure of quality and stability of the method. We summarize that
(1) it is desirable to construct W ⊂ V equipped with weights aw such that random walks,
starting in w ∈ W and weighted by aw , intersect each other as little as possible.
(2) if we are not allowed to chose W , we can still use the procedure above to find weights that
yield a better result than just uniform sampling.
4
We emphasize that we do not attempt to solve the heat ball packing problem here – nor do we
expect it to be easily solvable at this level of generality. The main contribution of this paper is to
introduce the heat ball packing problem as a fundamental issue with implications for sampling.
Problem. How does one find effective heat ball packing configurations quickly?
Are there algorithms leading to effective almost-minimizing configurations? What
theoretical guarantees can be proven?
2.3. Another interpretation. There is a nice bit of intuition coming from a more abstract
perspective: the reason why the above method works is that the heat propagator et∆ is selfadjoint. More precisely, let us assume (M, g) is a compact manifold normalized to vol(M ) = 1 and
let µ be a probability measure (our quadrature rule) on it. Smooth functions f have the property
that they do not change substantially if we apply the heat propagator for a short time t > 0 (this
is one way of quantifying smoothness) and therefore
hf, µi ∼ et∆ f, µ = f, et∆ µ .
In order for this quadrature rule to be effective, we want that et∆ µ is very close to the Lebesgue
measure dx. However, the heat propagator preserves the L1 −mass and thus
1 = kµk = ket∆ µkL1 ≤ ket∆ µkL2 .
The Cauchy-Schwarz inequality is only sharp if et∆ µ coincides with the constant function and
thus minimizing ket∆ µkL2 is a natural way to obtain good quadrature rules. If µ is a weighted
sum of Dirac measures, then et∆ turns this, roughly, into a sum of Gaussians (‘heat balls’) and
finding locations to minimize the L2 −norm becomes, essentially, a packing problem. We observe
that the argument does not single out L2 and minimizing ket∆ µkLp for p > 1 will lead to a very
similar phenomenon – it remains to be seen whether there is any advantage to that perspective
since L2 is usually easiest to deal with in practice.
2.4. The Laplacian: a second method. We believe that our notion of Laplacian may be quite
useful in practice (it already appeared in [11]). It combines the two desirable properties of
(1) having a symmetric matrix (and thus orthogonal eigenvectors)
(2) inducing a diffusion operator that preserves the mean value of the function.
The Kirchhoff matrix L1 = D − A has the second property but not the first; the normalized
Laplacian L2 = Idn×n − D−1/2 AD−1/2 satisfies the first property but not the second. We believe
that, for this reason alone, our notion of a Laplacian may be useful in other contexts as well. We
observe that if we can compute eigenvectors directly, then there is a very direct way of approaching
the problem directly.
Proposition. Let W ⊂ V be equipped with weights aw summing to 1. Then, for all 0 < λ < 1,
sup
f ∈Xλ
f 6=0
X
X
1 X
1
1
f (v) −
aw f (w) =
−
aw δw
kf kXλ n
n
v∈V
w∈W
w∈W
Xλ
2
This statement follows easily from L −duality and will be obtained in the proof of the Theorem as
a by-product. The result is true in general but, of course, one cannot compute the quantity on the
right-hand side unless one has access to the eigenvectors of the Laplacian. If one has indeed access
to either the Laplacian eigenfunctions or at least some of the leading eigenvectors φ1 , . . . , φk where
k n is chosen so that λk ∼ λ (simpler to obtain in practice), then optimizing the functional in
W ⊂ V and weights aw is essentially equivalent to finding good quadrature points. This simple
observation is very effective in practice, we refer to numerical examples below.
3. How to use the Theorem
The result discussed above has the nice effect of cleanly separating the problem of numerical
integration on a graph from the actual graph structure: the geometry of the graph is encoded
implicitly in the geometry of the random walk. This has the nice effect of providing a uniform
treatment but, as a downside, does not provide an immediate method on how to proceed in
5
particular instances. The purpose of this section is to comment on various aspects of the problem
and discuss approaches. Recall that our main result can be written as
12
2
X
X
X
1
1
1
aw δw
−
∀ f ∈ Xλ
f (v) −
aw f (w) ≤ kf kXλ min ` (Idn×n + L)`
`∈N λ
n
n
v∈V
w∈W
w∈W
L2
3.1. The parameters λ and `. A priori we have no knowledge about the degree of smoothness
of the function and have no control over it. The parameter `, on the other hand, is quite important
and has a nontrivial impact on the minimizing energy configurations for the quantity on the righthand side. In practice, we have to choose ` without knowing λ (fixing an allowed number of sample
points implicitly fixes a scale for λ). We propose the following basic heuristic.
Heuristic. If ` is too small, then there is not enough diffusion and heat balls
interact strongly with themselves. If ` is too large, the exponentially increasing
weight λ−` is too large. There is an intermediate regime when heat balls starting
to interact with nearby heat balls.
The heuristic is accurate if the graph is locally (at the scale of typical distance between elements
of W ) close to Euclidean. On general graphs, the optimal scale of ` might be more nontrivial to
estimate – we have observed that, in practice, there is a wide range of ` yielding good results.
3.2. The placement of W . Naturally, to avoid intersections of random walkers, we want to place
the elements of W as far from each other as possible. In particular, if the graph is close to having
a Euclidean structure, we would expect fairly equi-spaced points to do well. A method that was
used in [16] is to start with a random set of k vertices {v1 , v2 , . . . , vk } and compute the
total mutual distance =
k
X
d(vi , vj ),
i,j=1
where d is a metric on G = (V, E). The algorithm then goes through all the vertices and checks
whether moving one of them to a neighboring vertices increases the total mutual distance and,
if so, moves the vertex. This is repeated as long as possible. The simple numerical examples
in [16] all have edges with equal weight and the standard combinatorial graph distance can be
used; the optimal type of metric could strongly depend on the type of graph under consideration.
There is a fairly natural reason why an algorithm of this type has the ability to successfully
produce sets of vertices that are very well spread out. We quickly return to the sphere Sd where
a particularly spectacular justification exists. Let σ be the normalized measure on Sd . Then, for
any set X = {x1 , . . . , xn } ⊂ Sd , we have Stolarsky’s invariance principle [1, 17]
Z Z
n
2
1 X
kxi − xj k = cd DL2 ,cap (X) ,
kx − ykdσ(x)dσ(y) − 2
n
Sd Sd
i,j=1
where the quantity on the right is the L2 −based spherical cap discrepancy and cd is a constant
only depending on the dimension. The L2 −based spherical cap discrepancy is a measure that has
been studied in its own right: if the points are evenly distributed, then it is small. This may
be a somewhat peculiar case. However, it is not difficult to see that on fairly generic manifolds
functionals along the lines of
n
X
1
→ min
d(xi , xj )α
i,j=1
or
n
X
e−αd(xi ,xj ) → min
i,j=1
converge to the uniform distributions if the number of points is large. Moreover, and this is
particularly useful, these types of functionals tend to produce minimal energy configurations that
only weakly depend on the functional being used. On two-dimensional manifolds, the hexagonal
lattice seems to be particularly universal (see [2]). We do not know what kind of interaction
functional is optimal on graphs. In practice, one would like to have fast and reliable algorithms
that scale well and this seems like a problem of substantial interest.
6
3.3. The weights aw . Once we are given a set W ⊂ V and a parameter `, the optimization of
the weights is completely straightforward. Observe that
2
(Idn×n + L)
`
X
=
aw δw
w∈W
L2
X
aw1 aw2 (Idn×n + L)` δw1 , (Idn×n + L)` δw2 .
w1 ,w2 ∈W
This is merely a quadratic form indexed by a |W | × |W | matrix – we thus need to solve the
semidefinite program
X
X
aw1 aw2 (Idn×n + L)` δw1 , (Idn×n + L)` δw2 → min
subject to
aw = 1.
w1 ,w2 ∈W
w∈W
These weights aw play an important role in further fine-tuning an existing set of vertices W ⊂ V .
This is the reason why minimizing the functional can be used to find appropriate weights for any
given set of vertices W : if two vertices in W happen to be somewhat close to each other, then the
quadratic form will take this into account when distributing the weights. Conversely, if one the
vertices in W is surprisingly isolated from the other vertices in W , then the quadratic form will
increase the weight assigned to that vertex. This is exactly how things should be: points that are
oversampling a region in the graph should be given a smaller weight whereas isolated vertices cover
a wider range and are, correspondingly, more important. We refer to §4 for numerical examples
illustrating this point.
3.4. Related results. Sampling on graphs is a fundamental problem and a variety of approaches
have been discussed in the literature [9]. Sampling is usually done for the purpose of compression
or visualization and not numerical integration (in particular, vertices are usually not equipped
with weights). Our approach seems to be very different from anything that has been proposed.
The closest construction seems to be a line of research using biased random walks [8, 10, 13] based
on the idea of sending random walkers and accepting the vertices they traverse with certain biased
probabilities as sampling points. In contrast, we select points so that random walkers starting
thereavoid each other. Other results seem related in spirit [3]. Our approach is motivated by a
recent approach [15] to the study of spherical t−designs of Delsarte, Goethals & Seidel [6] (and
Sobolev [14]). These are sets of points defined on Sd with the property that they integrate a large
number of low-degree polynomials exactly (we refer to a survey of Brauchart & Grabner). [4]. The
second author recently extended some of these results to weighted points and general manifolds
[15]. These ideas are shown to be an effective source of good quadrature points in the Euclidean
setting in a paper of Lu, Sachs and the second author [12].
3
9
1
7
5
8
10
6
11
12
2
4
Figure 1. (Left:) The Icosahedron integrates all polynomials up to degree 5
on S2 exactly (this space is 36-dimensional). (Right:) a subset of 8 vertices
integrates 21 of 24 eigenvectors of the McGee graph exactly. Such examples
require extraordinary amounts of symmetry and are not generic.
The second author recently proved [16] a generalized Delsarte-Goethals-Seidel bound for graph
designs (the analogue of spherical t−designs on combinatorial graphs). The main condition in that
7
paper is algebraic (exact integration of a certain number of Laplacian eigenvectors) as opposed to
quantitative (small integration error). [16] shows a number of truly remarkable quadrature rules
on highly structured graphs that were found by numerical search: one of these rules is depicted
in Figure 1 and manages with only 8 evaluations to integrate 21 out of a total of 24 eigenvectors
exactly. However, these examples are very non-generic, the consequence of strong underlying
symmetries and not likely to be representative of what can be achieved in a typical setting.
4. Numerical Examples
4.1. Importance of weights. We start with a toy example shown in the Figure below to illustrate
the importance of the weights to counterbalance bad geometric distributions. We assume we are
given two clusters and an uneven distribution of six sampling points: five end up in one cluster while
the other cluster contains a total of five points. We construct a graph based on nearest neighbor
distances weighted with a Gaussian kernel. The spreading of heat is rather uneven if all points
are given equal weight. Conversely, by adjusting the weight to minimize the L2 −norm (constraint
to summing to 1 and being nonnegative), a much more balanced distribution is achieved.
Figure 2. Six points selected unevenly in two clusters (left), the heat flow emanating from weighting them all evenly (middle) and the optimal coefficients (0.11,
0.24, 0.05, 0.00, 0.18 and 0.43) for the heat ball packing problem (right).
The weights show that, in particular, one point is given weight 0 and another point is given a
rather small weight (0.05). This is to counterbalance the clustering of points. The isolated point
is given almost half the weight (0.43). We see that the heat distribution in the second cluster is
still highly uneven: this shows that it would be preferable to pick another point since the single
sampling point is actually quite far from the center of the cluster: if it was closer to the center, it
would have received an even larger weight.
0
0.4
0.81
0.19
0.4
0.2
Figure 3. Two optimal configurations for ` = 3 on 3 vertices.
8
Figure 3 shows an example
P on a small graph with 10 vertices and 15 edges. More precisely, we
optimize k(Id10×10 + L)3 w aw δw kL2 over all sets W with three vertices. In the first example, we
see that there is one very central node that distributes well throughout the network, another weight
is actually set to 0. This is a consequence of constraining the optimization to non-negative weights
aw ≥ 0. This constraint is not required by the Theorem, but makes the optimization easier, and
is well-motivated by classical methods in numerical integration. If we move the point that was
assigned weight 0, then the weight splits evenly (the value of the functional barely changes).
4.2. MNIST. Our explicit example is as follows: we consider the data set MNIST, a collection of
handwritten digits represented as 28 × 28 pixels (each pixel being either 0 or 1). For simplicity, we
only consider the subset comprised of the digits 0 and 1. The challenge problem is to figure out
the proportion of elements in the set that are handwritten digits that are 1’s (which is one half).
This ties in to our example in the beginning: suppose we did not know the precise proportion of
1’s and the data is unlabeled. Labeling the data is expensive: the function evaluation would be
one human being looking at a picture and labeling it, which is costly. However, these pictures
Relative error
0.3
0.2
0.1
0
0
50
100
150
200
Number ` of diffusion steps.
Figure 4. Estimating digits in MNIST: the relative integration error for 20 different sets of points of size 50 and how it evolves as a function depending on `
(leading to different selection of weights). The average of these 20 curves is shown
as the bold black line, sampling error for randomly chosen points is the red line.
are merely {0, 1} −vectors in R784 . As is commonly done, we reduce the dimensionality of the
data by projecting onto its first ten principal components. We build a graph by connecting every
element to its 10-nearest neighbors (in Euclidean distance) weighted with a Gaussian kernel and
then symmetrize by averaging the resulting adjacency matrix with its transpose. It is reasonable to
assume that the indicator function of 1’s is smooth over a graph defined by that notion of distance:
handwritten digits looking like a 1 should be close to other handwritten digits that look like 1. We
then proceed as outlined above: we sample random points, move them iteratively so that they are
far away from each other and then adjust weights by solving the semidefinite program. The result
is plotted against the parameter ` and compared to uniform weights on random points (red); the
picture shows 20 different sets of points, the evolution of their integration error depending on `
as well as their average (black). We observe that for the right parameter range of `, the obtained
9
numerical integration scheme performs much better but the precise performance depends on the
points chosen. This highlights the need for fast, stable and guaranteed ways of approximately
solving the heat ball packing problem.
4.3. Using Eigenvectors of Laplacian. This section studies the same example as above, estimating the proportion of handwritten digits ‘1’ in MNIST, but assumes additionally that we
are able to access the eigenvectors of the Laplacian associated to the largest few eigenvalues exactly. We set λ = 0.994 close to 1 leading to a space Xλ spanned by very few of the smoothest
eigenvectors, sample random points, move them far apart and make use of
sup
f ∈Xλ
f 6=0
X
X
1 X
1
1
f (v) −
aw f (w) = − +
aw δw
kf kXλ n
n
v∈V
w∈W
w∈W
Xλ
to explicitly optimize the weights. Our method (blue) is shown to perform exceedingly well; the
0.6
0.5
Relative error
0.4
0.3
0.2
0.1
0
0
20
40
60
80
100
Sampling size
Figure 5. Numerical Integration with access to eigenvectors: our method (blue)
compared to sampling in random points (yellow) and sampling in the location in
which we move the points (red).
weights are crucial, sampling over the same points (red) is even worse than sampling over random
points (yellow) which decays as (#size of subset)−1/2 (averaged over 200 random samples). As
mentioned above, having direct access to the eigenvectors implies that the bound is often close
to sharp. This is illustrated in the subsequent Figure 6 where we replace the indicator function
of the ‘1’-digits in the subset of MNIST comprised of digits 0 and 1 by its mollification obtained
from projecting onto the first 6 eigenvectors. We optimize in the weights and obtain
X
1
−
aw δw
n
w∈W
as an upper bound (red)
Xλ
that is then compared to the error on the function f (blue). We observe that as soon as the
sample size exceeds a certain limit, integration becomes exact. It is quite desirable to obtain a
better understanding of the interplay of parameters involved: suppose wePare given a set of k welldistributed points and optimize the weights so as to minimize −n−1 + w∈W aw δw X , what is
λ
the interplay between k, the Xλ space and the performance of the arising quadrature rule? Or,
put differently, how does the integration error in a Xµ space depend on the space Xλ that was
used to determine the weights? This question is clearly of great relevance in applications.
10
100
Integration error
10-2
10-4
10-6
10-8
10-10
10
20
30
40
50
60
70
80
90
100
Sampling size
Figure 6. Error for the smoothed indicator function on the 1’s (blue) and the
theoretical upper bound on the integration error (red). It depends strongly on W
and, after some initial fluctuation settles down to essentially exact integration;
the theoretical upper bound matches the performance on the particular instance.
The phenomenon, which seems to be generic and easily observed in most examples, is illustrated
in Figure 7. We downsampled the MNIST dataset of digits ‘0’ and ‘1’ to consist of a total of 1000
points, and constructed the graph as before. We then choose a subset of points W of size 100,
increase their mutual distance and fix them as quadrature points.
Integration error
10
0
10-5
10-10
10
20
25
10
-15
0
10
20
30
40
50
Eigenvector
Figure 7. Integration error of three quadrature rules with weights fine-tuned
in three different Xλ −spaces on the first 50 eigenfunctions.
Finally, we optimize their weights in three different Xλ spaces where λ is chosen such that the
dimensions of the spaces are 10, 20 and 25 (i.e. they contain the first 10, 20 and 25 eigenfunctions,
11
respectively). We then plot the integration error of these three quadrature rules on the first
50 eigenfunctions. If we optimize the weights according to the Xλ space containing the first 10
eigenfunctions, then the first 10 eigenfunctions are essentially integrated exactly, the subsequent
integration error is small. The same is true for optimization in the space containing the first 20
eigenfunctions. Then the behavior changes abruptly: if we optimize over the first 25 eigenfunctions,
then the error on those 25 eigenfunctions is small (∼ 10−4 ) and, as in the other examples, increases
afterwards. This seems to be typical: for any given set W ⊂ V , there seems to be a range
of Xλ −spaces such that optimizing parameters leads to exact integration in Xλ . Once their
dimension exceeds a certain (sharp) threshold, the error is still small in Xλ but many orders
of magnitude larger than before. This sharp phase transition could serve as another measure of
quality of W that may be useful in judging the quality of algorithms finding W (the largest number
of eigenfunctions that can be integrated exactly using W for some choice of weights, a measure
already studied in [16]).
5. Proof of the Theorem
Proof. We write the integration error as the inner product of two vectors and decompose it as
!
X
X 1
X
1 X
f (v) −
aw f (w) =
−
aw δw (v) f (v)
n
n
v∈V
w∈W
v∈V
w∈W
*
+
n
X
X
1
=
−
aw δw , φk hf, φk i.
n
w∈W
k=1
Since f ∈ Xλ , we have that hf, φk i = 0 unless |λk | ≥ λ. A simple application of the CauchySchwarz inequality then shows that
X
X
1 X
1
−
f (v) −
aw f (w) ≤
aw δw
n
n
v∈V
w∈W
w∈W
kf kXλ .
Xλ
More precisely, L2 −duality implies that this step is not lossy since
sup
f ∈Xλ
f 6=0
X
X
1
1 X
1
f (v) −
aw f (w) =
−
aw δw
kf kXλ n
n
v∈V
w∈W
w∈W
.
Xλ
For any function g : V → R, we have
2
kgkXλ =
X
|λk |≥λ
2
|hg, φk i| ≤
n
1 X
2
|λk |2` |hg, φk i|
λ2`
k=1
2
n
1 X `
= 2`
λk hg, φk iφk
λ
k=1
1
= 2` (Idn×n + L)` g
λ
L2
2
L2
.
We observe that this inequality is also valid if g ∈
/ Xλ since ever step is a valid bound from above
and k · kXλ is defined on all functions as a semi-norm (we note, however, that if g ∈
/ Xλ , then the
inequality will usually be far from sharp). We use this inequality for
X
1
aw δw
g= −
n
w∈W
to conclude that
X
1
−
aw δw
n
w∈W
kf kXλ
Xλ
kf kXλ
≤
(Idn×n + L)`
λ`
X
1
−
aw δw
n
w∈W
!
.
L2
12
We observe that Idn×n + L is the generator of the diffusion. It is a linear operator for which
constant functions are invariants. This implies
!
X
X
1
1
`
(Idn×n + L)
−
aw δw = − (Idn×n + L)`
aw δw
n
n
w∈W
w∈W
We now show that the operator (Idn×n + L) preserves the average value of the function. It suffices
to show that L maps the average value of a function to 0. We use the definition
1
(A − D) .
L = A0 − D 0 =
dmax
It thus suffices to show that A − D maps every function to a function with mean value 0. This
follows from changing the order of summation
X
XX
XX
X
X
(Af )(i) =
eij f (j) =
eij f (j) =
Dii f (i) =
(Df )(i).
i∈V
i∈V j∈V
j∈V i∈V
i∈V
i∈V
This implies that if we normalize the weights so that constants are being integrated exactly, i.e.
X
aw = 1,
w∈W
then the mean value of
(Idn×n + L)`
X
aw δw
is exactly
w∈W
1
.
n
Squaring out implies
X
1
− (Idn×n + L)`
aw δw
n
w∈W
2
L2
2
X
1
= + (Idn×n + L)`
aw δw
n
w∈W
2
*
+L
X
1
−2
, (Idn×n + L)`
aw δw
n
w∈W
2
`
= (Idn×n + L)
X
−
aw δw
w∈W
L2
1
.
n
Altogether, we have shown
X
X
1 X
kf kXλ
sup
f (v) −
f (w) ≤
(Idn×n + L)`
aw δw
`
λ
f ∈Xλ n
v∈V
w∈W
w∈W
2
L2
21
1
−
.
n
Acknowledgement GCL was supported by NIH grant #1R01HG008383-01A1 (PI: Yuval Kluger)
and U.S. NIH MSTP Training Grant T32GM007205.
References
[1] D. Bilyk, F. Dai and R. Matzke, Stolarsky Principle and Energy Optimization on the Sphere, Constructive
Approximation, to appear
[2] X. Blanc and M. Lewin, The crystallization conjecture: a review. EMS Surv. Math. Sci. 2 (2015), no. 2,
225–306.
[3] A. Bermanis, A. Averbuch, and R. Coifman. ”Multiscale data sampling and function extension.” Applied and
Computational Harmonic Analysis 34.1 (2013): 15-29.
[4] J. Brauchart and P. Grabner, Distributing many points on spheres: minimal energy and designs. J. Complexity
31, no. 3, 293–326, (2015).
[5] F. Chung, Spectral Graph Theory, CBMS Regional Conference Series in Mathematics 92, American Mathematical Society, 1996.
[6] P. Delsarte, J. M. Goethals and J. J. Seidel, Spherical codes and designs. Geometriae Dedicata 6, no. 3,
363–388, (1977).
13
[7] T. Erber and G. Hockney, Complex systems: Equilibrium configurations of n equal charges on a sphere
(2 ≤ n ≤ 112), Advances in Chemical Physics 98 (1997), 495–594.
[8] M. Gjoka, M. Kurant, C. T. Butts, and A. Markopoulou. Walking in facebook: A case study of unbiased
sampling of osns. In INFOCOM, 2010 Proceedings IEEE, pages 1–9. IEEE, 2010.
[9] P. Hu and W.C.Lau, A Survey and Taxonomy of Graph Sampling, arXiv:1308.5865
[10] L. Jin, Y. Chen, P. Hui, C. Ding, T. Wang, A. V. Vasilakos, B. Deng, and X. Li. Albatross sampling: robust
and effective hybrid vertex sampling for social graphs. In Proceedings of the 3rd ACM international workshop
on MobiArch, pages 11–16. ACM, 2011.
[11] G. Linderman and S. Steinerberger, Clustering with t-SNE, Provably, arXiv:1706.02582
[12] J. Lu, M. Sachs and S. Steinerberger, Quadrature Points via Heat Kernel Repulsion, in preparation.
[13] A. H. Rasti, M. Torkjazi, R. Rejaie, N. Duffield, W. Willinger, and D. Stutzbach. Respondent-driven sampling
for characterizing unstructured overlays. In INFOCOM 2009, IEEE, pages 2701–2705. IEEE, 2009.
[14] S. Sobolev, Cubature formulas on the sphere which are invariant under transformations of finite rotation
groups. Dokl. Akad. Nauk SSSR 146, 310–313, (1962).
[15] S. Steinerberger, Spectral Limitations of Quadrature Rules and Generalized Spherical Designs,
arXiv:1708.08736
[16] S. Steinerberger, Designs on Graphs: Sampling, Spectra, Symmetries, arXiv:1803.02235
[17] K. B. Stolarsky. Sums of distances between points on a sphere. II. Proc. Amer. Math. Soc., 41, 575–582, 1973.
Program in Applied Mathematics, Yale University, New Haven, CT 06511, USA
E-mail address: [email protected]
Department of Mathematics, Yale University, New Haven, CT 06511, USA
E-mail address: [email protected]
| 10math.ST
|
.
arXiv:1710.02505v1 [math.NT] 6 Oct 2017
RIGID LOCAL SYSTEMS AND ALTERNATING
GROUPS
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
Contents
1. Introduction
2. The local systems in general
3. The candidate local systems for Alt(2q)
4. Basic facts about Hn
5. basic facts about H2q−1
6. Basic facts about the group Ggeom for F (k, 2q − 1, ψ)
7. The third moment of F (k, 2q − 1, ψ) and of G(k, 2q − 1, ψ)
8. Exact determination of Garith
9. Identifying the group
References
1
2
4
5
6
7
8
11
12
25
1. Introduction
In earlier work [Ka-RLSFM], Katz exhibited some very simple one
parameter families of exponential sums which gave rigid local systems
on the affine line in characteristic p whose geometric (and usually, arithmetic) monodromy groups were SL2 (q), and he exhibited other such
very simple families giving SU3 (q). [Here q is a power of the characteristic p, and p is odd.] In this paper, we exhibit equally simple families
whose geometric monodromy groups are the alternating groups Alt(2q).
We also determine their arithmetic monodromy groups. See Theorem
3.1. [Of course from the resolution [Ray] of the Abhyankar Conjecture, any finite simple group whose order is divisible by p will occur as
the geometric monodromy group of some local system on A1 /Fp ; the
interest here is that it occurs in our particularly simple local systems.]
In the earlier work of Katz, he used a theorem to Kubert to know
that the monodromy groups in question were finite, then work of Gross
[Gross] to determine which finite groups they were. Here we do not
The first author was partially supported by NSF grant DMS-1600056 and the
third author was partially supported by NSF grant DMS-1665014. The first author
would also like to thank the Institute for Advanced Study, Princeton for its support.
1
2
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
have, at present, any direct way of showing this finiteness. Rather,
the situation is more complicated and more interesting. Using some
basic information about these local systems (cf. Theorem 6.1), the
first and third authors prove a fundamental dichotomy: the geometric
monodromy group is either Alt(2q) or it is the special orthogonal group
SO(2q − 1). The second author uses an elementary polynomial identity
to compute the third moment as being 1 (cf. Theorem 7.1), which
rules out the SO(2q − 1) case. This roundabout method establishes the
theorem. It would be interesting to find a “direct” proof that these local
systems have integer (rather than rational) traces; this integrality is in
fact equivalent to their monodromy groups being finite, cf. [Ka-ESDE,
8.14.6]. But even if one had such a direct proof, it would still require
serious group theory to show that their geometric monodromy groups
are the alternating groups.
2. The local systems in general
Throughout this paper, p is an odd prime, q is a power of p, k is a
finite field of charactertistic p, ℓ is a prime 6= p,
ψ = ψk : (k, +) → µp ⊂ Qℓ
×
is a nontrivial additive character of k, and
χ2 = χ2,k : k × → ±1 ⊂ Qℓ
×
is the quadratic character, extended to k by χ2 (0) := 0. For L/k a
finite extension, we have the nontrivial additive character
ψL/k := ψk ◦ TraceL/k
of L, and the quadratic character χ2,L = χ2,k ◦NormL/k of L× , extended
to L by χ2,L (0) = 0.
On the affine line A1 /k, we have the Artin-Schreier sheaf Lψ(x) . On
Gm /k we have the Kummer sheaf Lχ2 (x) and its extension by zero
j! Lχ2 (x) (for j : Gm ⊂ A1 the inclusion) on A1 /k.
For an odd integer n = 2d + 1 which is prime to p, we have the rigid
local system
F (k, n, ψ) := F Tψ (Lψ(xn ) ⊗ j! Lχ2 (x) )
on A1 /k. Let us recall the basic facts about it, cf. [Ka-NG2, 1.3 and
1.4].
It is lisse of rank n, pure of weight one, and orthogonally self dual,
with its geometric monodromy group
Ggeom ⊂ SO(n, Qℓ ).
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
3
Recall that Ggeom is the Zariski closure in SO(n, Qℓ ) of the image of
the geometric fundamental group π1 (A1 /k) in the representation which
“is” the local system F (k, n, ψ). For ease of later reference, we recall
the following fundamental fact.
Lemma 2.1. For any lisse local system on A1 /k, the subgroup Γp of
its Ggeom generated by elements of p-power order is Zariski dense.
Proof. Denote by N the Zariski closure of Γp . Then N is a normal
subgroup of Ggeom . We must show that the quotient M := Ggeom /N
is trivial. We first note that M/M 0 is trivial, as it is a finite, prime
to p quotient of π1 (A1 /k). Thus M is connected. We next show that
M red := M/Ru , the quotient of M by its unipotent radical, is trivial.
For this, it suffices to show that M has no nontrivial irreducible representations. But any such representation is a local system on A1 /k
which is tamely ramified at ∞, so is trivial. Thus M is unipotent.
But H 1 (A1 /k, Qℓ ) vanishes, so any unipotent local system on A1 /k is
trivial, and hence M is trivial.
Let us denote by A(k, n, ψ) the Gauss sum
X
ψ(x)χ2 (x).
A(k, n, ψ) := −χ2 (n(−1)d )
x∈k ×
By the Hasse-Davenport relation, for L/k an extension of degree d, we
have
A(L, n, ψL/k ) = (A(k, n, ψ))d .
The twisted local system
G(k, n, ψ) := F (k, n, ψ) ⊗ A(n, k, ψ)−deg
is pure of weight zero and has
Ggeom ⊂ Garith ⊂ SO(n, Qℓ ).
Concretely, for L/k a finite extension, and t ∈ L, the trace at time t
of G(k, n, ψ) is
X
ψL/k (xn +tx)χ2,L (x) =
Trace(F robt,L |G(k, n, ψ)) = −(1/A(L, n, ψL/k ))
x∈L×
= −(1/A(L, n, ψL/k ))
X
ψL/k (xn + tx)χ2,L (x),
x∈L
the last equality because the χ2 factor kills the x = 0 term.
Let us recall also [Ka-NG2, 3.4] that the geometric monodromy group
of F (k, n, ψ), or equivalently of G(k, n, ψ), is independent of the choice
of the pair (k, ψ).
4
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
To end this section, let us recall the relation of the local system
F (k, n, ψ) to the hypergeometric sheaf
Hn := H(!, ψ; all char′ s of order dividing n; χ2 ).
According to [Ka-ESDE, 9.2.2], F (k, n, ψ)|Gm is geometrically isomorphic to a multiplicative translate of the Kummer pullback [n]⋆ Hn . [An
explicit descent of F (k, n, ψ)|Gm through the n’th power map is given
by the lisse sheaf on Gm whose trace function at time t ∈ L× , for L/k
a finite extension, is
X
ψL/k (xn /t + x)χ2,L (x/t).
t 7→ −
x∈L×
The structure theory of hypergeometric sheaves show that this descent
is, geometrically, a multiplicative translate of the asserted Hn .]
3. The candidate local systems for Alt(2q)
In this section, we specialize the n of the previous section to
n = 2q − 1 = 2(q − 1) + 1.
The target theorem is this.
Theorem 3.1. Let p be an odd prime, q a power of p, k a finite field of
characteristic p, ℓ a prime 6= p, and ψ a nontrivial additive character
of k. For the ℓ-adic local system G(k, 2q − 1, ψ) on A1 /k, its geometric
and arithmetic monodromy groups are given as follows.
(1) Ggeom = Alt(2q) in its unique irreducible representation of dimension 2q − 1.
(2a) If −1 is a square in k, then Ggeom = Garith = Alt(2q).
(2b) If −1 is not a square in k, then Garith = Sym(2q), the symmetric group, in its irreducible representation labeled by the partition (2, 12q−2 ), i.e.
(the deleted permutation representation of Sym(2q)) ⊗ sgn.
Remark 3.2. The traces of elements of Alt(n) (respectively of Sym(n))
in its deleted permutation representation (respectively in every irreducible representation) are integers. One sees easily (look at the action
of Gal(Q(ζp )/Q)) that the local system G(k, 2q − 1, ψ) has traces which
all lie in Q, but as mentioned in the introduction, we do not know a
direct proof that these traces all lie in Z.
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
5
4. Basic facts about Hn
In this section, we assume that n ≥ 3 is odd and that n(n − 1)
is prime to p. The geometric local monodromy at 0 is tame, and a
topological generator of the tame inertia group I(0)tame , acting on Hn ,
has as eigenvalues all the roots of unity of order dividing n.
The geometric local monodromy at ∞ is the direct sum
Lχ2 ⊕ W ; W has rank n − 1, and all slopes 1/(n − 1).
Because n is odd, the local system Hn is (geometrically) orthogonally self dual, and det(Hn ) is geometrically trivial (because trivial at
0, lisse on Gm , and all ∞ slopes are ≤ 1/(n − 1) < 1). Therefore
det(W ) is geometrically Lχ2 . From [Ka-ESDE, 8.6.4 and 8.7.2], we
see that up to multiplicative translation, the geometric isomorphism
class is determined entirely by its rank n − 1 and its determinant Lχ2 .
Because n − 1 is even and prime to p, it follows that up to multiplicative translation, the geometric isomorphism class of W is that of the
I(∞)-representation of the Kloosterman sheaf
Kln−1 := Kl(ψ; all char′ s of order dividing n − 1).
By [Ka-GKM, 5.6.1], we have a global Kummer direct image geometric isomorphism
Kln−1 ∼
= [n − 1]⋆ Lψn−1 ,
where we write ψn−1 for the additive character x 7→ ψ((n−1)x). Therefore, up to multiplicative translation, the geometric isomorphism class
of W is that of [n − 1]⋆ Lψ . Pulling back by [n − 1], which does not
change the restriction of W to the wild inertia group P (∞), we get
[n − 1]⋆ W ∼
= ⊕ζ∈µn−1 Lψ(ζx) .
A further pullback by n’th power, which also does not change the
restriction of W to P (∞), gives
[n − 1]⋆ [n]⋆ W ∼
= ⊕ζ∈µn−1 Lψ(ζxn ) .
Thus we find that the I(∞) representation attached to a multiplicative
translate of [n − 1]⋆ F (k, n, ψ) is the direct sum
M
M
Lψ(ζxn ) =
Lψ(αxn ) .
1
ζ∈µn−1
α∈µn−1 ∪{0}
This description shows that the image of P (∞) in the I(∞)-representation
attached to F (k, n, ψ) is an abelian group killed by p.
6
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
Lemma 4.1. Let L/Fp be a finite extension which contains the n−1’st
roots of unity. Denote by V ⊂ L the additive subgroup of L spanned by
the n − 1’st roots of unity. Denote by V ⋆ the Poincaré dual of V :
V ⋆ := HomFp (V, µp (Qℓ )).
Then the image of P (∞) in the I(∞)-representation attached to F (k, n, ψ)
is V ⋆ , and the representation restricted to V ⋆ is the direct sum of
M
M
1
(evaluation at ζ) =
(evaluation at α).
ζ∈µn−1 (L)
α∈µn−1 (L)∪{0}
Proof. Each of the characters Lψ(αxn ) of I(∞) has order dividing p.
Given an n-tuple of elements (aα )α∈µn−1 (L)∪{0} , consider the character
Λ := ⊗α∈µn−1 (L)∪{0} (Lψ(αxn ) )⊗aα =
= Lψ((Pα∈µ
n−1 (L)∪{0}
aα α)xn ) .
The following conditions are equivalent.
P
(1) α∈µn−1 (L)∪{0} aα α = 0.
(2) The character Λ is trivial on I(∞).
(3) The character Λ is trivial on P (∞).
Indeed,
P it is obvious that (1) =⇒ (2) =⇒ (3). If (3) holds, then for
A := α∈µn−1 (L)∪{0} aα α, we have that Lψ(Ax) is trivial on P (∞), so is
a character of I(∞)/P (∞) = I(∞)tame , a group of order prime to p.
But Lψ(Ax) has order dividing p, so is trivial on I(∞), hence A = 0.
This equivalence shows that the character group of the image of
P (∞) is indeed the Fp span of the α’s, i.e., it is V . The rest is just
Poincaré duality of finite abelian groups.
5. basic facts about H2q−1
Taking n = 2q − 1, the geometric local monodromy at 0 of H2q−1 is
tame, and a topological generator of the tame inertia group I(0)tame ,
acting on Hn , has as eigenvalues all the roots of unity of order dividing
2q − 1.
Turning now to the action of P (∞), we have
Lemma 5.1. Denote by ζ2q−2 ∈ Fq2 a primitive 2q − 2’th root of unity.
In the I(∞)-representation attached to F (k, 2q − 1, ψ), the character
group V of the image of P (∞) is the Fp -space
V = Fq ⊕ ζ2q−2 Fq .
Fix a nontrivial additive character ψ0 of Fq , and denote by ψ1 the
nontrivial additive character of Fq2 given by
ψ1 := ψ0 ◦ TraceFq2 /Fq .
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
7
Then the image V ⋆ of P (∞) is itself isomorphic to V , and the representation of P (∞) is the direct sum of the characters
M
⊕β∈F×q ψ1 (ζ2q−2 βx).
⊕α∈Fq ψ1 (αx)
Proof. When n = 2q − 1, then n − 1 = 2(q − 1). The field Fq2 contains
the 2(q − 1)’th roots of unity. The group µ2(q−1) (Fq2 ) contains the
×
subgroup µq−1 (Fq2 ) = F×
q with index 2, the other coset being ζ2(q−1) Fq .
Thus the Fp span of µ2(q−1) (Fq2 ) inside the additive group of Fq2 is
indeed the asserted V . The characters ψ1 (αx), as α varies over Fq , are
each trivial on ζ2q−2 Fq (because TraceFq2 /Fq (ζ2q−2 ) = 0) and give all
the additive characters of Fq (on which TraceFq2 /Fq is simply the map
x 7→ 2x). The characters ψ1 (ζ2q−2 βx), as β varies over Fq , are trivial
on Fq (because TraceFq2 /Fq (ζ2q−2 ) = 0) and give all the characters of
2
ζ2q−2 Fq (because ζ2(q−1)
lies in F×
q ).
Corollary 5.2. The image of P (∞) in the I(∞)-representation attached to
F (k, 2q − 1, ψ) ⊕ 1
is the direct sum
V = Fq ⊕ ζ2q−2 Fq
acting through the representation
RegFq ⊕ Regζ2q−2 Fq .
6. Basic facts about the group Ggeom for F (k, 2q − 1, ψ)
Recall that Ggeom is the Zariski closure in SO(2q −1, Qℓ ) of the image
of π1geom := π1 (A1 /Fp ) in the representation attached to F (k, 2q −1, ψ).
Thus Ggeom is an irreducible subgroup of SO(2q − 1, Qℓ ).
Theorem 6.1. We have the following two results.
(1) Ggeom is normalized by an element of SO(2q − 1, Qℓ ) whose
eigenvalues are all the roots of unity of order dividing 2q − 1 in
Qℓ .
(2) Ggeom contains a subgroup isomorphic to Fq ⊕Fq , acting through
the virtual representation
Regfirst ⊕ Regsecond − 1.
Proof. The local system F (k, 2q − 1, ψ) is, geometrically, a multiplicative translate of the Kummer pullback [2q − 1]⋆ H2q−1 . Therefore Ggeom
is a normal subgroup of the group Ggeom for H2q−1 , so is normalized
by any element of this possibly larger group. As already noted, local
monodromy at 0 for H2q−1 is an element of the asserted type. This
8
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
proves (1). Statement (2) is just a repeating of what was proved in the
previous lemma.
7. The third moment of F (k, 2q − 1, ψ) and of G(k, 2q − 1, ψ)
Let us recall the general set up. We are given a lisse G on a lisse,
geometrically connected curve C/k. We suppose that G is ι-pure of
weight zero, for an embedding ι of Qℓ into C. We denote by V the Qℓ representation given by G, and by Ggeom the Zariski closure in GL(V )
of the image of π1geom (C/k). For an integer n ≥ 1, the n’th moment of
G is the dimension of the space of invariants
Mn (G) := dim((V ⊗n )Ggeom ).
Recall [Ka-MMP, 1.17.4] that we have an archimedean limit formula
for Mn (G) as the lim sup over finite extensions L/k of the sums
X
(1/#L)
(Trace(F robt,L |G))n ,
t∈C(L)
which we call the empirical moments.
Theorem 7.1. For the lisse sheaf G(k, 2q − 1, ψ) on A1 /k, we have
M3 (G(k, 2q − 1, ψ)) = 1.
Proof. Fix a finite extension L/k. For t ∈ L, we have
Trace(F robt,L |G(k, 2q−1, ψ)) = (−1/A(L, 2q−1, ψL/k ))
X
ψL/k (x2q−1 +tx)χ2,L (x),
x∈L
with the twisting factor given explicitly as
X
ψL/k (x)χ2,L (x).
A(L, 2q − 1, ψL/k ) = −χ2,L (−1)
x∈L×
Write gL for the Gauss sum
gL :=
X
ψL/k (x)χ2,L (x).
x∈L×
Then the empirical M3 is the sum
X X
(1/#L)(χ2,L (−1)/gL )3
(ψL/k (x2q−1 +y 2q−1 +z 2q−1 +t(x+y+z))χ2,L (xyz) =
t∈L x,y,z∈L
X
= (χ2,L (−1)/gL )3
ψL/k (x2q−1 + y 2q−1 + z 2q−1 )χ2,L (xyz) =
x,y,z∈L,x+y+z=0
= (χ2,L (−1)/gL )3
X
ψL/k (x2q−1 +y 2q−1 +(−x−y)2q−1 )χ2,L (xy(−x−y)).
x,y∈L
The key is now the following identity.
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
9
Lemma 7.2. In Fq [x, y], we have the identity
x2q−1 + y 2q−1 + (−x − y)2q−1 = xy(x + y)
Y
(x − αy)2.
α∈Fq \{0,−1}
If we write q = pf , then collecting Galois-conjugate terms this is
Y
h(x, y)2 ,
xy(x + y)
h∈Pf
where Pf is the set of irreducible h(x, y) ∈ Fp [x, y] which are homogeneous of degree dividing f , monic in x, other than x or x + y.
Proof. Because x2q−1 + y 2q−1 + (−x − y)2q−1 is homogeneous of odd
degree 2q − 1 and visibly divisible by y, it suffices to prove the inhomogenous identity, that in Fq [x] we have
Y
x2q−1 + 1 − (x + 1)2q−1 = x(x + 1)
(x − α)2 .
α∈Fq \{0,−1}
The left side
P (x) := x2q−1 + 1 − (x + 1)2q−1
has degree 2q.
So it suffices to show that for each α ∈ Fq \ {0, −1},P (x) is divisible
by (x − α)2 . The key point is that for β ∈ Fq , we have
β 2q−1 = β,
and for α ∈ F×
q we have
α2q−2 = 1.
Thus for any β ∈ Fq , we trivially have P (β) = 0. The derivative P ′ (x)
is equal to
P ′ (x) = −x2q−2 + (x + 1)2q−2 .
′
So if both α and α + 1 lie in F×
q , then P (α) = −1 + 1 = 0.
With this identity in hand, we now return to the calculation of the
empirical moment, which is now
X
Y
h(x, y)2 )χ2,L (xy(−x − y)).
(χ2,L (−1)/gL )3
ψL/k (xy(x + y)
x,y∈L
h∈Pf
The set of (x, y) ∈ A2 (L) at which h∈Pf h(x, y) = 0 has cardinality at
most (q − 2)(#L − 1). So the empirical sum differs from the modified
empirical sum
Y
X
Y
h(x, y)2)
h(x, y)2 )χ2,L (xy(−x−y)
ψL/k (xy(x+y)
(χ2,L (−1)/gL )3
Q
x,y∈L
h∈Pf
h∈Pf
10
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
by a difference which is
(χ2,L (−1)/gL )3 (a sum of at most (q−2)(#L−1) terms, each of absolute value 1).
p
So the difference in absolute value is at most q/ (#L), which tends
to zero as L grows (remember q is fixed). The modified empirical sum
we now rewrite as
X
ψL/k (t)χ2,L (−t)NL (t),
(χ2,L (−1)/gL )3
t∈L×
with NL (t) the number of L-points on the curve Ct given by
Y
h(x, y)2 = t.
Ct : xy(x + y)
h∈Pf
Because xy(x+y) h∈Pf h(x, y)2 is homogeneous of degree 2q −1 prime
to p and is not a d’th power for any d ≥ 2, the curves Ct are smooth
and geometrically irreducible for all t 6= 0, cf. [Ka-PES, proof of 6.5].
Moreover, by the homogeneity, these curves are each geometrically isomorphic to C1 , indeed the family become constant after the tame Kummer pullback [2q − 1]⋆ . Thus for the structural map π : C → Gm /Fp ,
R2 π! (Qℓ ) = Qℓ (−1), R1 π! Qℓ is lisse of some rank r, tame at both 0 and
∞, and mixed of weight ≤ 1, and all other Ri π! (Qℓ ) = 0.
So our modified empirical moment is
X
ψL/k (t)χ2,L (−t)(#L − Trace(F robt,L |R1 π! Qℓ ) =
(χ2,L (−1)/gL )3
Q
t∈L×
= (χ2,L (−1)/gL )3
X
ψL/k (t)χ2,L (−t)(#L)
t∈L×
−(χ2,L (−1)/gL )3
X
t∈L×
Trace(F robt,L |Lχ(t) ⊗ Lχ2 (t) ⊗ R1 π! Qℓ ).
gL2
Remembering that
= χ2,L (−1), we see that
√ the first sum is χ2,L (−1).
We now show that the second sum is O(1/ #L), or equivalently that
the sum
X
Trace(F robt,L |Lψ ⊗ Lχ2 ⊗ R1 π! Qℓ )
t∈L×
is O(#L). By the Lefschetz trace formula [Gro-FL], the second sum is
Trace(F robL |Hc2 (Gm /Fp , Lψ ⊗ Lχ2 ⊗ R1 π! Qℓ )
−Trace(F robL |Hc1(Gm /Fp , Lψ ⊗ Lχ2 ⊗ R1 π! Qℓ ).
The Hc2 group vanishes, because the coefficient sheaf is totally wild
at ∞ (this because it is Lψ tensored with a lisse sheaf which is tame
at ∞). The second sum is O(#L), by Deligne’s fundamental estimate
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
11
[De-Weil II, 3.3.1] (because the coefficient sheaf of mixed of weight ≤ 1,
so its Hc1 is mixed of weight ≤ 2).
Thus the empirical
√ moment is χ2,L (−1) plus an error term which, as
L grows, is O(1/ #L). So the lim sup is 1, as asserted.
8. Exact determination of Garith
Theorem 8.1. Suppose known that G(k, 2q−1, ψ) has Ggeom = Alt(2q).
Then its Garith is as asserted in Theorem 3.1, namely it is Alt(2q) if
−1 is a square in k, and is Sym(2q) if −1 is not a square in k.
Proof. For q > 3, the outer automorphism group of Alt(2q) has order 2, induced by the conjugation action of Sym(2q). Therefore the
normalizer of Alt(2q) in SO(2q − 1) (viewed there by its deleted permutation representation) is the group Sym(2q) (viewed in SO(2q − 1)
by (deleted permutation representation)⊗sgn). If q = 3, the automoprhism group is slightly bigger but the stabilizer of the character of
the deleted permutation module is just Sym(2q). [Indeed, either of the
exotic automorphisms of Alt(6) maps the cycle (123) to an element
which in Sym(6) is conjugate to (123)(456). The element (123) has
trace 2, whereas (123)(456) has trace −1 (both viewed in SO(5) by the
deleted permutation representation)]. Since we have a priori inclusions
Ggeom = Alt(2q) ⊳ Garith ⊂ SO(2q − 1),
the only choices for Garith are Alt(2q) or Sym(2q).
Denoting by V the representation of Garith given by G(k, 2q − 1, ψ),
the action of Garith on line
L := (V ⊗3 )Ggeom
is a character of Garith /Ggeom . We claim that this character is the sign
character sgn of Garith ⊂ Sym(2q). To see this, we argue as follows.
For any n ≥ 3, denoting by Vn the deleted permutation representation of Sym(n + 1), one knows that
(Vn⊗3 )Sym(n+1) = (Vn⊗3 )Alt(n+1)
is one dimensional, [Indeed, if S λ denotes the complex irreducible representation of Sym(n + 1) labeled by the partition λ ⊢ (n + 1), then
n+1
Vn = S (n,1) and sgn = S (1 ) . An application of the LittlewoodRichardson rule to
Sym(n+1)
S λ ⊗ IndSym(n
(S (n) ) = S λ ⊕ (S λ ⊗ Vn )
yields
2)
Vn ⊗ Vn = S (n+1) ⊕ S (n,1) ⊕ S (n−1,2) ⊕ S (n−1,1
12
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
see [FH, Exercise 4.19]. Further similar applications of the LittlewoodRichardson rule then show that Vn ⊗ Vn ⊗ Vn contains the trivial representation S (n+1) once but does not contain sgn.] Hence that the action
of Sym(n + 1) on
((Vn ⊗ sgn)⊗3 )Alt(n+1)
is sgn3 = sgn. Taking n = 2q − 1, we get the claim.
Now apply Deligne’s equidistribution theorem, in the form [Ka-Sar,
9.7.10]. It tells us that if Garith /Ggeom has order 2 instead of 1, then
the Frobenii F robt,L as L runs over larger and larger extensions of k of
even (respectively odd) degree become equidistribued in the conjugacy
classes of Garith lying in Ggeom (respectively lying in the other coset
Garith \ Ggeom ). Then as L/k runs over finite extensions
of odd degree,
√
the empirical third moment will be −1 + O(1/ #L). On the other
hand,√if Garith = Ggeom , then every empirical moment will be 1 +
O(1/ #L).
9. Identifying the group
In this section, we use the information obtained earlier to identify
the group. We choose a field embedding Qℓ ⊂ C, so that we may view
G := Ggeom as an algebraic group over C.
So let p be an odd prime with q a power of p. We start by assuming
that G is an irreducible, Zariski closed subgroup of SO(2q − 1, C) =
SO(V ) such that G contains Q, an elementary abelian subgroup of
order q 2 . Moreover, we assume that we may write Q = Q1 × Q2 so
that V = V0 ⊕ V1 ⊕ V2 where V0 is a trivial Q-module, V0 ⊕ Vi is
the regular representation for Qi and Qi acts trivially on the other
summand. Moreover, we assume that G is a quasi-p group (in the
sense that the subgroup generated by its p-elements is Zariski dense),
cf. Lemma 2.1.
Lemma 9.1. V is tensor indecomposable for Q1 . More precisely, V 6=
X1 ⊗ X2 where the Xi are Q1 -modules each of dimension ≥ 2.
Proof. We argue by contradiction. Suppose V = X1 ⊗ X2 with each
Xi of (necessarily odd) dimensional
≥ 2 . Let χXi beP
the character of
P
Q1 on Xi . So χX1 = a0 1 + aχ χ and χX2 = b0 1 + bχ χ where the
χ are the nontrivial characters of Q1 .
We first reduce to the case when both a0 , b0 are nonzero. The multiplicity of the trivial character of Q1 in V is q, so we have
X
q = a0 b0 +
aχ bχ .
χ
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
13
So either a0 b0 is nonzero, and we are done, or for some nontrivial χ we
have aχ bχ nonzero. In this latter case, replace X1 by X1 ⊗ χ and X2
by X2 ⊗ χ.
Since each nontrivial character χ of Q1 occur exactly once in V , for
each such χ we have
X
1 = a0 bχ + aχ b0 +
aρ bχρ .
ρ6=χ
In particular we have the inequalities
a0 bχ ≤ 1, aχ b0 ≤ 1.
Because a0 , b0 are both nonzero, we infer that if aχ 6= 0, then aχ =
b0 = 1 (respectively that if bχ 6= 0, then a0 = bχ = 1). It cannot be the
case that all aχ vanish, otherwise X1 is the trivial module of dimension
> 1. This is impossible so long as X2 is nontrivial, as each nontrivial
character of Q1 occurs in V exactly once. But if all aχ and all bχ vanish,
then V is the trivial Q1 module, which it is not. Therefore a0 = 1 and,
similarly, b0 = 1, and all aχ , bχ are either 0 or 1. Now use again that
the multiplicity of the trivial character of Q1 in V is q, so we have
X
aχ bχ .
q = a0 b0 +
χ
This is possible only if all aχ and all bχ are 1. But then each Xi has
dimension q, which is impossible, as the product of their dimensions is
2q − 1.
Lemma 9.2. The following statements hold for G.
(1) G preserves no nontrivial orthogonal decomposition of V .
(2) V is not tensor induced for G.
Proof. We first prove (1). We argue by contradiction. Suppose that
V = W1 ⊥ . . . ⊥ Wr with r > 1. Because G acts irreducibly, G transitively permutes the Wi , and all the Wi have the same odd dimension
d (because 2q − 1 = rd). Since r divides 2q − 1, gcd(r, p) = 1, so the
p-group Q fixes at least one of the Wi , say W1 . Because r > 1, there
are other orbits of Q on the set of blocks. Any of these has cardinality
some power of p, so the corresponding direct sum of Wi ’s has odd dimension. As 2q − 1 is odd, there must be evenly many other orbits, so
at least three orbits in total. In each Q-stable odd-dimensional orthogonal space, Q lies in a maximal torus of the corresponding SO group,
so has a fixed line. Hence dim V Q ≥ 3, contradiction.
We next show that V is not tensor induced. We argue by contradiction. If V is tensor induced, write V = W ⊗ . . . ⊗ W (with f ≥ 2
14
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
tensor factors, dim W < dim V ). Then Q1 must act transitively on
the set of tensor factors (otherwise the representation for Q1 is tensor
decomposable and the previous lemma gives a contradiction).
So by Jordan’s theorem, there exists an element y ∈ Q1 that acts
fixed point freely on the set of the f tensor factors. All such elements
are conjugate in GL(W ) ≀ Sym(f ) and we have
χV (y) = (dim W )f /p .
[Indeed, after replacing y by a GL(W )≀Sym(f )-conjugate, the situation
is this. Each orbit of y on the set of tensor factors has length p, and y
acts on each corresponding p-fold self product of W , indexed by Fp , by
mapping ⊗i wi to ⊗i wi+1 . In terms of a basis B := {ej }j=1,...,dim W of W,
the only diagonal entries of the matrix of y on this W ⊗p are given by
the dim W vectors e ⊗ e... ⊗ e with e ∈ B.] On the other hand, we have
χV (y) = q − 1 for any nonzero element y of Q1 . Thus, if d = dim W ,
we have df /p = q − 1. Thus, dim V = df = (q − 1)p > 2q − 1, a
contradiction.
Corollary 9.3. Let L ≤ SO(V ) be any subgroup containing G and let
1 6= N ⊳ L. Then N acts irreducibly on V .
Proof. We argue by contradiction. Note that the conclusions of Lemmas 9.1 and 9.2 also hold for L.
(i) Because N is normal in L, V is completely reducible for N. Let
V1 , . . . , Vr be the distinct N-isomorphism classes of N-irreducible submodules of V . Because V is L-self dual, it is a fortiori N-self dual.
Therefore the set of Vi is stable by passage to the N-dual, Vi 7→ Vi⋆ .
The group L acts transitively on the set of the Vi . Either every Vi is Nself-dual, or none is (the L-conjugates of an N-self dual representation
are N-self dual).
When we write V as the direct sum of its the N-isotypic (“homogeneous” in the terminology of [C-R, 49.5]) components,
V = W1 ⊕ . . . ⊕ Wr ,
then for some integer e ≥ 1 we have N-isomorphisms
Wi ∼
= eVi := the direct sum of e copies of Vi .
If r > 1 and all the Wi are self dual, then this is an orthogonal
decompositon (because for i 6= j, the inner product pairing of (any) Vi
with (any) Vj is an N-homomorphism from Vi to Vj⋆ ∼
= Vj , so vanishes).
This contradicts Lemma 9.2.
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
15
Suppose r > 1 and no Vi is self-dual. Then the Vi occur in pairs of
duals. Therefore both r and dim V are even, again a contradiction.
(ii) We have shown that r = 1 and e > 1, i.e. V ∼
= eV1 . Now we
apply Clifford’s theorem, cf. [C-R, Theorem 51.7]. Thus L preserves
the N-isomorphism class of V1 , and so we get an irreducible projective
representation L 7→ PGL(V1 ) = PSL(V1 ), and V as a projective representation of L is V1 ⊗ X with L acting (projectively) irreducibly on X
through L/N, and X of dimension e. Furthermore, the two factor sets
associated to these two projective representations are inverses to each
other (because the tensor product V1 ⊗X = V is a linear representation
of Q1 ). Since e dim V1 = dim V = 2q − 1, both e and n = dim V1 are
coprime to p.
We now claim that, restricted to Q1 , each of the tensor factors V1
and X lifts to a genuine linear representation. Indeed, using the fact
that PGL(n, C) = PSL(n, C) and the short exact sequence
1 → µn → SL(n, C) → PSL(n, C) → 1,
the obstruction for (V1 )|Q1 , which is given by the first factor set restricted to Q1 , lies in the cohomology group H 2 (Q1 , µn ). As p ∤ n while
Q1 is a p-group, this cohomology group vanishes; and so the first factor set restricted to Q1 is trivial, whence so is the second. Thus the
Q1 -module V is tensor decomposable, contradicting Lemma 9.1.
We next show that G is finite. It is convenient to use one more
fact about G. There is a subgroup A (namely the group Ggeom for the
hypergeometric sheaf H2q−1 ) of SO(V ) such that G is normal in A,
A/G is cyclic of order dividing 2q − 1 and A contains an element x of
order 2q − 1 with distinct eigenvalues on V .
We also use the fact that G has a nontrivial fixed space on V ⊗V ⊗V
(Theorem 7.1).
Theorem 9.4. G is finite.
Proof. Suppose not. Let N be any nontrivial normal (closed) subgroup
of G. By Corollary 9.3, N is irreducible on V .
(i) Let G0 be the identity component of G. We now show that G0
is a simple algebraic group. Taking N = G0 , we have that G0 acts irreducibly and hence is semisimple (as it lies in SO(V )). Moreover, the
center of G0 is trivial (because it consists of scalars in SO(V )). Therefore if G0 is not simple, it is the product of adjoint groups Lj , 1 ≤ j ≤ t
(namely the adjoint forms of the factors of its universal cover), and V
is the (outer) tensor product V = ⊗tj=1 Vj of nontrivial irreducible Lj modules Vj . By [GT, Corollary 2.7], G permutes these tensor factors
16
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
Vj . This action is transitive, otherwise we contradict Lemma 9.1. But
this implies that V is tensor induced for G, contradicting Lemma 9.2.
Thus G0 is a simple algebraic group.
(ii) Because the subgroup of G generated by its p-elements is Zariski
dense, the finite group G/G0 is generated by its p-elements. As p is
odd, it follows that either G = G0 is a simple algebraic group or p = 3
and G0 = D4 (C). [In all other cases, the outer automorphism group,
i.e., the automorphism group of the Dynkin diagram of G0 , has order
at most 2.] Since A/G has odd order, it follows that A ≤ G as well
unless G0 = D4 (C) and 3 divides 2q − 1.
Suppose first that A is connected and so a simple algebraic group.
Then it contains a semisimple element x acting with distinct eigenvalues. This implies that a maximal torus has all weight spaces of
dimension at most 1. Moreover, the module is in the root lattice (since
it is odd dimensional and orthogonal). By a result of Howe [Ho] (see
also [Pa, Table]), it follows if G 6= SO(V ), then either G = G2 (C) with
dim V = 7 or G = PGL2 (C). If dim V = 7, then q = 4, a contradiction.
If G = PGL2 (C), then any finite abelian subgroup of odd order is cyclic
and so Q does not embed in G.
So G = SO(V ). However, SO(V ) has no nonzero fixed points on
V ⊗ V ⊗ V and this contradicts Theorem 7.1.
Thus, it follows that A is disconnected. So the connected component
is D4 (C) and this acts irreducibly on V . If D4 (C) contains the element
of order 2q−1, then a maximal torus has all weight space of dimension 1
and again using [Ho], we obtain a contradiction. If not, then 3 divides
2q − 1, whence p ≥ 5 and Q ≤ D4 (C). Any elementary abelian psubgroup of D4 (C) is contained in a torus and so again we see that the
connected component has all weight spaces of dimension at most 1 and
we obtain the final contradiction using [Ho].
Let F ∗ (X) denote the generalized Fitting subgroup of a finite group
X (so X is almost simple precisely when F ∗ (X) is a non-abelian simple
group).
Corollary 9.5. A and G are almost simple and F ∗ (A) = F ∗ (G) acts
irreducibly on V .
Proof. Let N be a minimal normal subgroup of G. By Corollary 9.3,
N acts irreducibly, and so by Schur’s lemma CA (N) = Z(N) = 1 as
A < SO(V ) with dim V odd. So N is a direct product of non-abelian
simple groups. Arguing as in p. (i) of the proof of Theorem 9.4, we see
that N is non-abelian simple (otherwise the module V would be tensor
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
17
induced). As CG (N) = 1, we see that N ⊳ G ≤ Aut(N), and so G is
almost simple and F ∗ (G) = N.
Now, as G ⊳ A, A normalizes N. Again since CA (N) = 1 we have
that N ⊳ A ≤ Aut(N), and so A is almost simple and F ∗ (A) = N.
We next observe that:
Lemma 9.6. F ∗ (G) is not a sporadic simple group.
Proof. Notice that both G and A are generated by elements of odd
order (p-elements for G, these and elements of order 2q − 1 for A).
On the other hand, we have S ≤ G ≤ A ≤ Aut(S) for S = F ∗ (G).
One knows [Atlas] that if S is sporadic, then |Out(S)| ≤ 2. Therefore,
if S is a sporadic simple group, then G = A = S. The result now
folllows easily from information in [Atlas]. Namely, we observe that if
q is an odd prime power with q 2 dividing |G|, then G has no irreducible
representation of dimension 2q − 1.
We next consider the case F ∗ (G) = Alt(n). First note Alt(5) contains no noncyclic elementary abelian groups of odd order and so is
ruled out. Since 2q − 1 is odd, we see that if G = Alt(n), then
A = G = Alt(n) (as the outer automorphism group of Alt(n) is a
2-group).
Theorem 9.7. Let Γ = Alt(n) with n ≥ 6. Suppose that x ∈ Γ has
odd order and V is an irreducible C[Γ]-module such that x acts as a
semisimple regular element on V . Then one the following holds:
(1) V is the deleted permutation module of dimension n−1 (i.e. the
Alt(n)
nontrivial irreducible constituent of CAlt(n−1) ), and x is either
an n-cycle (for n odd) or a product of two disjoint cycles of
coprime lengths (for n even); or
(2) n = 8, x has order 15 and dim V = 14.
Proof. First note that if V is the deleted permutation module of dimension n − 1, an element with 3 or more disjoint cycles has at least
a two dimensional fixed space on V . Similarly, if x has two disjoint
cycles of lengths which are not coprime, then x has a two dimensional
eigenspace on V .
Next we observe that if x is semisimple regular on V , then the order
of x is at least dim V . This proves the result for 6 ≤ n ≤ 14 by
inspection of the odd order elements and dimensions of the irreducible
modules, aside from the case n = 8 and dim V = 14 (note that Alt(8)
contains an element of order 15). Recall that Alt(8) ∼
= GL4 (2) and it
acts 2-transitively on the nonzero vectors. The only irreducible module
18
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
of dimension 14 is the irreducible summand of the permutation module
of dimension 15. In this case x has a single orbit in the permutation
representation and so x is semisimple regular on V .
Now assume that n ≥ 15.
Suppose first that x has at most three nontrivial cycles. Then the
order of x is less than (n/3)3 = n3 /27 and so dim V < n3 /27. Let W
be a complex irreducible Sym(n)-module whose restriction to Alt(n)
contains V |Alt(n) . Since 2 ≤ dim W < 2n3 /27, it follows by [Ra, Result
3] that W ∼
= S λ or S λ ⊗ sgn, where S λ is the Specht module labeled
by the partition λ of n, with λ = (n − 1, 1), (n − 2, 2), or (n − 2, 1, 1).
Restricting back to Alt(n), we see that V |Alt(n) = S λ |Alt(n) .
Note that
dim S (n−2,1,1) = (n − 1)(n − 2)/2, dim S (n−2,2) = n(n − 3)/2.
It is straightforward to see that the dimension of the fixed space of x
on either of these modules is at least two dimensional, a contradiction.
Hence λ = (n − 1, 1) and V |Alt(n) is the deleted permutation module of
dimension n − 1.
We now induct on n. The base case n ≤ 14 has already done. We
may assume that x has at least four nontrivial cycles (each of odd
length, as x has odd order). View x ∈ J := Alt(a) × Alt(b). where
the projection into Alt(b) is a b-cycle and so the projection into Alt(a)
is a product of at least three disjoint cycles. Thus, a ≥ 9. Let W be
an irreducible J-submodule of V with Alt(a) acting nontrivially. So
W = W1 ⊗ W2 with W1 an irreducible Alt(a)-module. Then x must be
multiplicity free on each Wi and by induction x can have at most two
cycles in Alt(a), a contradiction.
Note that the previous result does fail for n = 5. Alt(5) has a
5-dimensional representation in which an element of order 5 has all
eigenvalues occurring once. Thus if G = Alt(n), we see that n = 2q
and V is the deleted permutation module.
Corollary 9.8. If G = Ggeom is an alternating group Alt(n) for some
n, then n = 2q.
Finally, we consider the case where G is an almost simple finite group
of Lie type, defined over Fs , where s = sf0 is a power of a prime s0 . Let
us denote
S := F ∗ (G) = F ∗ (A).
Recall that S is simple, irreducible on V , and Z(S) = 1 by Corollary
9.5. We will freely use information on character tables of simple groups
available in [Atlas, GAP], as well as degrees of complex irreducible
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
19
characters of various quasisimple groups of Lie type available in [Lu].
Finally, we will also use bounds on the smallest degree d(S) of nontrivial
complex irreducible representations of S as listed in [T, Table 1].
Theorem 9.9. Suppose s0 6= p. Then S ∼
= Alt(m) with m ∈ {5, 6, 8}.
Proof. (i) Assume the contrary. We will exploit the existence of the
subgroup Q ≤ G. Recall that the p-rank mp (G) is the largest rank of
elementary abelian p-subgroups of G. Furthermore,
(9.9.1)
Aut(S) ∼
= Inndiag(S) ⋊ ΦS ΓS ,
where Inndiag(S) is the subgroup of inner-diagonal automorphisms of
S, ΦS is a subgroup of field automorphisms of S and ΓS is a subgroup
of graph automorphisms of S, as defined in [GLS, Theorem 2.5.12]. As
F ∗ (G) = S, we can embed G in Aut(S). Now, given an elementary
abelian p-subgroup P < G of rank mp (G), we can define a normal series
1 ≤ P1 ≤ P2 ≤ P,
where P1 = P ∩ Inndiag(S) and P2 = P ∩ (Inndiag(S) ⋊ ΦS ). As ΦS is
cyclic and P is elementary abelian, P2 /P1 has order 1 or p. Set e = 1
e
if S ∼
= P Ω+
8 (s) and p = 3, and e = 0 otherwise. Then |P/P2 | ≤ p .
Next we bound |P1 | when S is not a Suzuki-Ree group. Let Φj (t)
denote the j th cyclotomic polynomial in the variable t, and let m denote
the multiplicative order of s modulo p, so that p|Φm (s). Note that we
can find a simple algebraic group G of adjoint type defined over Fs and
a Frobenius endomorphism F : G → G such that Inndiag(S) ∼
= GF .
Letting r denote the rank of G, then one can find r positive integers
k1 , . . . , kr and ǫ1 , . . . , ǫr = ±1 such that
r
Y
Y
N
nj
N
(ski − ǫi )
|Inndiag(S)| = s
Φj (s) = s
j≥1
i=1
for suitable integers N, nj . Then, according to [GLS, Theorem 4.10.3(b)],
|P1 | ≤ pnm . Let ϕ(·) denote the Euler function, so that deg(Φm ) =
ϕ(m). Inspecting the integers k1 , . . . , kr , one sees that nm ≤ r/ϕ(m).
It follows that
|P1 | ≤ Φm (s)nm ≤ ((s + 1)ϕ(m) )r/ϕ(m) ≤ (s + 1)r .
In fact, one can verify that this bound on |P1 | also holds for Suzuki-Ree
groups. Putting all the above estimates together, we obtain that
(9.9.2)
q 2 = |Q| ≤ |P | ≤ (s + 1)r+1+e .
We will show that this upper bound on q contradicts the lower bound
(9.9.3)
2q − 1 = dim V ≥ d(S)
20
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
in most of the cases. Let f ∗ denote the odd part of the integer f .
(ii) First we handle the case when S is of type D4 or 3D4 . Here,
q ≤ (s + 1)3 by (9.9.2). On the other hand, d(S) ≥ s(s4 − s2 + 1),
contradicting (9.9.3) if s ≥ 3. If s = 2, then ΦS ΓS = C3 , and so instead
of (9.9.2) we now have that q 2 ≤ 35 , whence q ≤ 13, 2q−1 ≤ 25 < d(S),
again a contradiction.
From now on we may assume e = 0.
Next we consider the case S = PSL2 (s). Then Out(S) = Cgcd(2,s−1) ×
Cf , and mp (S) ≤ 1. It follows that Q is not contained in S but in
S ⋊ Cf and 3 ≤ p|f ∗ , and furthermore q 2 = |Q| ≤ (s + 1)f ∗ . As
d(S) ≥ (s − 1)/2, (9.9.3) now implies that
s + 1 = sf0 + 1 ≤ 16f ∗ ,
a contradiction if s0 ≥ 5, or s0 = 3 and f ≥ 5, or s0 = 2 and f ≥ 7. If
s0 = 3 and f ≤ 4, then f ∗ = 3 = f = p, forcing p = s0 , a contraction.
Suppose s0 = 2 and f ≤ 6. If p = 5, then f ∗ = 5 and mp (G) = 1,
ruling out the existence of Q. If p = 3, then f = 3, 6, whence q 2 ≤ 9
and 2q − 1 ≤ 5 < d(S).
Suppose that S = 2B2 (s) or 2G2 (s) with s ≥ 8. Since mp (S)p
≤ 1, we
see that q 2 ≤ (s + 1)f , contradicting (9.9.3) as d(S) ≥ (s − 1) s/2.
Now we consider the remaining cases with r = 2. Then q ≤ (s+1)3/2
by (9.9.2). This contradicts (9.9.3) if S = G2 (s) (and s ≥ 3), as
6= PSU3 (s)
d(S) ≥ s3 − 1. Similarly, S ∼
6= PSL3 (s) with s ≥ 5 and S ∼
with s ≥ 8. If S = PSp4 (s), then the case 2 ∤ s ≥ 19 is ruled out since
d(S) ≥ (s2 − 1)/2, and similarly the case 2|s ≥ 8 is ruled out since
d(S) = s(s − 1)2 /2. In the remaining cases, ΦS ΓS is a 2-group, and
so Q ≤ S, q 2 ≤ (s + 1)2 , q ≤ s + 1. Now PSL3 (s) and PSU3 (s) with
s ≥ 4 are ruled out by (9.9.3), and the same for PSp4 (s) with s ≥ 4.
Note that when s = 3, q ≥ 4 and so gcd(q, 2s) 6= 1, a contradiction. If
S = SL3 (2), then q = 3 and S has no irreducible character of degree
2q − 1. Finally, Sp4 (2)′ ∼
= Alt(6).
Next we handle the groups with r = 3. Here q ≤ (s + 1)2 by (9.9.2).
Then (9.9.3) implies that s ≤ 5. In this case, Out(S) is a 2-group, and
so Q ≤ S and q ≤ (s + 1)3/2 by (9.9.2). Using (9.9.3), we see that
s ≤ 3. The remaining groups S cannot occur, since S does not have a
real-valued complex irreducible character of degree 2q − 1.
(iii) From now we may assume that r ≥ 4 (and S is not of type D4
or 3D4 ). First we consider the case s = 2. If S = SLn (2) with n ≥ 5,
then since Out(S) = C2 , the arguments in (i) show that q 2 ≤ 3n−1 .
This contradicts (9.9.3), since d(S) = 2n − 2. Suppose S = SUn (2)
with n ≥ 7. Note by [TZ, Theorem 4.1] that the first three nontrivial
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
21
irreducible characters of S are Weil characters and either non-realvalued or of even degree, and the next characters have degree at least
(2n − 1)(2n−1 − 4)/9. Hence (9.9.3) can be improved to
2q − 1 ≥ (2n − 1)(2n−1 − 4)/9,
which is impossible since q 2 ≤ 3n by (9.9.2). If S = PSUn (2) with
n = 5, 6, then q 2 ≤ 36 , and S has no nontrivial real-valued irreducible
character of odd degree ≤ 2q − 1 ≤ 53. If S = 2F4 (2)′ or F4 (2),
then q 2 ≤ 35 , q ≤ 13, and S has no nontrivial real-valued irreducible
character of odd degree ≤ 2q − 1 ≤ 25.
Suppose S = Sp2n (2) or Ω±
2n (2). Then Out(S) is a 2-group (recall
S is not of type D4 ), and so q 2 ≤ 3n . On the other hand, d(S) ≥
(2n − 1)(2n−1 − 2)/3, contradicting (9.9.3). Finally, if S is of type E8 ,
E7 , E6 , or 2E6 , then q 2 ≤ 38 whereas d(S) > 210 , again contradicting
(9.9.3).
(iv) Suppose that S = PSp2n (s) with n ≥ 4 and 2 ∤ s ≥ 3. Then by
(9.9.2) and (9.9.3) we have
(sn − 1)/2 ≤ 2q − 1 ≤ 2(s + 1)(n+1)/2 − 1,
implying n ≤ 5 and s = 3. In this case, inspecting the order of PSp10 (3)
we see that q 2 ≤ 121, and so 2q − 1 ≤ 21 < d(S), a contradiction.
Next suppose that S = PSUn (3) with n ≥ 5. Then q ≤ 2n and
d(S) ≥ (3n − 3)/4, and so (9.9.3) implies that n = 5. In this case,
inspecting the order of SU5 (3) we see that q 2 ≤ 61, and so 2q − 1 ≤
13 < d(S), again a contradiction.
Now we may assume that r ≥ 4, s ≥ 3, S ∼
6= PSp2n (s) if 2 ∤ s, and
moreover s ≥ 4 if S ∼
= PSUn (s). Then one can check that d(S) ≥
sr · (51/64) (with equality attained exactly when S ∼
= PSU5 (4)). Hence
(9.9.2) and (9.9.3) imply that
(51/64)2 · s2r ≤ d(S)2 ≤ (2q − 1)2 < 4(s + 1)r+1 ≤ 4 · (4s/3)r+1,
and so
(3s/4)r−1 < 4 · (64/51)2 · (4/3)2 ,
which is impossible for r ≥ 4.
Theorem 9.10. Suppose s0 = p. Then S ∼
= Alt(6).
Proof. (i) Assume the contrary. We now exploit the existence of the
element x ∈ A of order 2q − 1 which has simple spectrum on V . As
before, we can embed A in Aut(S) and again use the decomposition
(9.9.1). Let hyi = hxi ∩ Inndiag(S). We also view S = G F for some
Frobenius endomorphism F : G → G of a simple algebraic group G of
adjoint type, defined over Fp . Note that y is an F -stable semisimple
22
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
element in G, hence it is contained in an F -stable maximal torus T of
G by [DM, Corollary 3.16]. It follows that |y| ≤ |T F | ≤ (s + 1)r , if r is
the rank of G. Set e = 3 if S is of type D4 or 3D4 , and e = 1 otherwise.
Then the decomposition (9.9.1) shows that
|x|/|y| ≤ ef ∗ ,
where f ∗ denotes the odd part of f as before (and s = pf ). We have
thus shown that
(9.10.1)
2q − 1 = |x| ≤ (s + 1)r ef ∗ .
We will frequently use the following remark:
(9.10.2)
Either f = 1 and s ≥ 3f ∗ , or s ≥ 9f ∗ .
We will show that in most of the cases (9.10.1) contradicts (9.9.3). First
we handle the case S is of type D4 or 3D4 , whence d(S) ≥ s(s4 −s2 + 1).
Hence (9.10.1) and (9.10.2) imply that
s(s4 − s2 + 1) ≤ 2q − 1 ≤ s(s + 1)4 /3
if f > 1, a contradiction. If f = 1, then since 2q −1 = dim V is coprime
to 2s, we see by [Lu] that
2q − 1 > s7 /2 > 3(s + 1)4 ,
contradicting (9.10.1).
(ii) From now on we may assume that e = 1. Next we rule out
the case where V |S is a Weil module of S ∈ {PSLn (s), PSUn (s)} with
n ≥ 3, or S = PSp2n (s) with n ≥ 2. Indeed, in this case, if S = PSLn (s)
then
dim V = (sn − s)/(s − 1), (sn − 1)/(s − 1)
is congruent to 0 or 1 modulo p and so cannot be equal to 2q − 1.
Similarly, if S = PSUn (s), then V |S can be a Weil module of dimension
2q − 1 only when 2|n and dim V = (sn − 1)/(s + 1). In this case,
q = (2q − 1)p = ((sn + s)/(s + 1))p = s
(where Np denotes the p-part of the integer N), and so 2s − 1 = (sn +
s)/(s + 1), a contradiction. Likewise, if S = PSp2n (s), then V |S can
be a Weil module of dimension 2q − 1 only when p = 3 and dim V =
(sn + 1)/2. In this case,
sn = (2 dim V − 1)p = (4q − 3)p ,
and so q = 3 and n = 2. One can show that PSp4 (3) does possess a
complex irreducible module of dimension 2q − 1 = 5, with an element
x of order 5 with simple spectrum on V and a subgroup Q ∼
= C32 with
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
23
desired prescribed action on V ; however, any such module is not selfdual. Henceforth, for the aforementioned possibilities for S we may
assume that dim V ≥ d2 (S), the next degree after the degree of Weil
characters. Note that d2 (S) for these simple groups S is determined in
Theorems 3.1, 4.1, and 5.2 of [TZ].
(iii) Suppose S = PSL2 (s); in particular, s 6= 9. Assume f ≥ 4.
As Out(S) = C2,s−1 × Cf , we see that q 2 ≤ sfp < s2 /20, whereas
2q − 1 ≥ d(S) ≥ (s − 1)/2, a contradiction. If f ≤ 3 but fp > 1,
then f = p = 3, s = 33 , q 2 ≤ sf = 34 , forcing q = 9. But then
S = PSL2 (27) has no irreducible character of degree 2q − 1 = 17. Thus
fp = 1, q 2 ≤ s, and so (9.9.3) implies that s ≤ 17. As s 6= 9, we see
that mp (G) = mp (S) = 1, contradicting the existence of Q.
Next we consider the case S = PSL3 (s) or PSU3 (s). If f > 1, then
(9.9.3)–(9.10.2) imply
(s − 1)(s2 − s + 1)/3 ≤ d2 (S) ≤ 2q − 1 ≤ (s + 1)2 s/9,
which is impossible. Thus f = 1, whence
(s − 1)(s2 − s + 1)/3 ≤ d2 (S) ≤ 2q − 1 ≤ (s + 1)2 ,
yielding s ≤ 5. But if s = 3 or 5, then any nontrivial χ ∈ Irr(S) of odd
degree coprime to s is a Weil character, which has been ruled out in
(ii).
Suppose now that S = PSL4 (s) or PSU4 (s). For s ≥ 5 we have
(s − 1)(s3 − 1)/2 ≤ d2 (S) ≤ 2q − 1 ≤ (s + 1)3 s/3,
which is possible only when s ≤ 11. Thus 3 ≤ s ≤ 11, whence f ∗ = 1,
and so
(s − 1)(s3 − 1)/2 ≤ d2 (S) ≤ 2q − 1 ≤ (s + 1)3 ,
leading to s = 3. If s = 3, then any odd-order element in G has order
≤ 13, whereas d(S) = 21, contradicting (9.9.3).
To finish off type A, assume now that S = PSLn (s) or PSUn (s) with
n ≥ 5. Then (9.9.3)–(9.10.2) imply
whence
(sn + 1)(sn−1 − s2 )
≤ d2 (S) ≤ 2q − 1 ≤ (s + 1)n−1 s/3,
2
(s + 1)(s − 1)
s2n−3 < (s + 1)n s/3 < s51n/40
(because (s + 1)/s ≤ 4/3 < 311/40 ), a contradiction as n ≥ 5.
(iv) Suppose S = P Ω±
2n (s) with n ≥ 4. For n ≥ 5 we get that
(sn − 1)(sn−1 − s)
≤ d(S) ≤ 2q − 1 ≤ (s + 1)n f ≤ (s + 1)n s/3,
2
s −1
24
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
whence
s2n−3.1 < (s + 1)n s/3 < s51n/40 ,
a contradiction. If n = 4, then S = P Ω−
8 (s). In this case, since 2q − 1
is coprime to 2s, [Lu] implies that
2q − 1 ≥ (s4 + 1)(s2 − s + 1)/2 > (s + 1)4 s/3,
again a contradiction.
Suppose S = PSp2n (s) with n ≥ 2 or Ω2n+1 (s) with n ≥ 3. Using the
bound 2q − 1 ≥ d2 (S) for S = PSp2n (s) and 2q − 1 ≥ d(S) otherwise,
we get for n ≥ 3 that
whence
(sn − 1)(sn − s)
≤ 2q − 1 ≤ (s + 1)n f ≤ (s + 1)n s/3,
s2 − 1
s2n−2.1 < (s + 1)n s/3 < s51n/40 ,
a contradiction. If n = 2, then S = PSp4 (s), and we have
s(s − 1)2 ≤ 2q − 1 ≤ (s + 1)s /3,
forcing q ≤ 9. If 5 ≤ q ≤ 9, then since the degree 2q − 1 = dim V
is coprime to 2s, we again get 2q − 1 > 300 ≥ (s + 1)2 s/3. Finally,
PSp4 (3) has no nontrivial non-Weil character of degree coprime to 6.
(v) If S is of type E6 , 2E6 , E7 , or E8 , then
(s5 + s)(s6 − s3 + 1) ≤ d(S) ≤ 2q − 1 ≤ (s + 1)8 f ≤ (s + 1)8 s/3,
a contradiction. Similarly, if S = F4 (s), then
s8 − s4 + 1 = d(S) ≤ 2q − 1 ≤ (s + 1)4 s/3,
which is impossible. Likewise, if S = G2 (s) with s ≥ 5, then
s3 − 1 ≤ d(S) ≤ 2q − 1 ≤ (s + 1)2 s/3,
again a contradiction. Next, if S = G2 (3), then 2q − 1 ≤ 16 cannot be
a degree of an irreducible character of S. Finally, if S = 2G2 (s), then
s2 − s + 1 = d(S) ≤ 2q − 1 ≤ (s + 1)f ≤ (s + 1)s/3,
again a contradiction since s ≥ 27.
Our proof is now concluded by applying Theorem 9.7.
RIGID LOCAL SYSTEMS AND ALTERNATING GROUPS
25
References
[As] Aschbacher, M., Maximal subgroups of classical groups, On the maximal subgroups of the finite classical groups, Invent. Math. 76 (1984), 469–514.
[Atlas] Conway, J. H., Curtis, R. T., Norton, S. P., Parker, R. A. and Wilson, R.
A., Atlas of Finite Groups, Clarendon Press, Oxford 1985.
[C-R] Curtis, C.W. and Reiner, I., Representation Theory of Finite Groups and
Associative Algebras, Interscience Publishers, New York 1962, xiv+689 pp.
[De-Weil II] Deligne, P., La conjecture de Weil II, Pub. Math. I.H.E.S. 52 (1981),
313–428.
[DM] Digne, F. and Michel, J., Representations of Finite Groups of Lie Type,
London Mathematical Society Student Texts 21, Cambridge University Press,
1991.
[FH] W. Fulton and J. Harris, Representation Theory, Springer-Verlag, New York,
1991.
[GAP] The GAP group, GAP - groups, algorithms, and programming, Version 4.4,
2004, http://www.gap-system.org.
[GLS] Gorenstein, D., Lyons, R., and Solomon, R.M., The Classification of the Finite Simple Groups, Number 3. Part I. Chapter A, volume 40 of Mathematical
Surveys and Monographs, American Mathematical Society, Providence, RI,
1998.
[Gross] Gross, B. H., Rigid local systems on Gm with finite monodromy, Adv.
Math. 224 (2010), no. 6, 2531–2543.
[Gro-FL] Grothendieck, A., Formule de Lefschetz et rationalité des fonctions L,
Seminaire Bourbaki 1964-65, Exposé 279, reprinted in: Dix Exposés sur la
cohomologie des schémas, North-Holland, 1968.
[GT] Guranick, R. M. and Tiep, P. H., Symmetric powers and a conjecture of
Kollar and Larsen, Invent. Math. 174 (2008), 505–554.
[Ho] Howe, R., Another look at the local θ-correspondence for an unramified dual
pair. Festschrift in honor of I. I. Piatetski-Shapiro on the occasion of his sixtieth
birthday, Part I (Ramat Aviv, 1989), 93–124, Israel Math. Conf. Proc., 2,
Weizmann, Jerusalem, 1990.
[Ka-ESDE] Katz, N., Exponential Sums and Differential Equations, Annals of
Mathematics Studies, 124. Princeton Univ. Press, Princeton, NJ, 1990. xii+430
pp.
[Ka-GKM] Katz, N., Gauss Sums, Kloosterman Sums, and Monodromy Groups,
Annals of Mathematics Studies, 116. Princeton Univ. Press, Princeton, NJ,
1988. ix+246 pp.
[Ka-MMP] Katz, N., Moments, Monodromy, and Perversity, Annals of Mathematics Studies, 159. Princeton University Press, Princeton, NJ, 2005. viii+475 pp.
[Ka-NG2] Katz, N., Notes on G2 , determinants, and equidistribution, Finite Felds
and their Applications 10 (2004), 221–269.
26
ROBERT M. GURALNICK, NICHOLAS M. KATZ, AND PHAM HUU TIEP
[Ka-PES] Katz, N., Perversity and exponential sums, in: Algebraic Number Theory
- in honor of K. Iwasawa, Advanced Studies in Pure Mathematics 17, 1989,
209–259.
[Ka-RLS] Katz, N., Rigid Local Systems, Annals of Mathematics Studies, 139.
Princeton University Press, Princeton, NJ, 1996. viii+223 pp.
[Ka-RLSFM] Katz, N., Rigid local systems on A1 with finite monodromy, preprint
available at www.math.princeton.edu/~nmk/gpconj87.pdf.
[Ka-Sar] Katz, N., and Sarnak, P., Random matrices, Frobenius eigenvalues, and
monodromy, American Mathematical Society Colloquium Publications, 45.
American Mathematical Society, Providence, RI, 1999. xii+419 pp.
[LS] Liebeck, M. W. and Seitz, G. M. On the subgroup structure of classical groups,
Invent. Math. 134 (1998), 427–453.
[Lu] Lübeck,
F.,
Character
degrees
and
their
multiplicities
for
some
groups
of
Lie
type
of
rank
<
9,
http://www.math.rwth-aachen.de/~Frank.Luebeck/chev/DegMult/index.html.
[Pa] Panyushev, D. I., Weight multiplicity free representations, g-endomorphism
algebras, and Dynkin polynomials, J. London Math. Soc. 69 (2004), 273–290.
[Ra] Rasala, R., On the minimal degrees of characters of Sn , J. Algebra 45 (1977),
132–181.
[Ray] Raynaud, M. Revêtements de la droite affine en caractéristique p > 0 et
conjecture d’Abhyankar, Invent. Math. 116 (1994), no. 1-3, 425–462.
[T] Tiep, P. H., Low dimensional representations of finite quasisimple groups, Proceedings of the London Math. Soc. Symposium “Groups, Geometries, and
Combinatorics”, Durham, July 2001, A. A. Ivanov et al eds., World Scientific,
2003, N. J. 277–294.
[TZ] Tiep, P. H. and Zalesskii, A. E., Minimal characters of the finite classical
groups, Comm. Algebra 24 (1996), 2093–2167.
Department of Mathematics, University of Southern California,
Los Angeles, CA 90089-2532, USA
E-mail address: [email protected]
Department of Mathematics, Princeton University, Princeton, NJ
08544-1000, USA
E-mail address: [email protected]
Department of Mathematics, Rutgers University, Piscataway, NJ
08854-8019, USA
E-mail address: [email protected]
| 4math.GR
|
arXiv:1412.1393v1 [cs.PL] 3 Dec 2014
CLAZY: Lazy Calling for Common Lisp
Marco Antoniotti
Lab Linguaggi Programmazione – Programming Language Lab < LLP ΠΛΛ >
Dipartimento di Informatica, Sistemistica e Comunicazione,
Università Milano Bicocca
U14 - Viale Sarca 336, I-20126 Milan, Italy
Abstract
This document contains a description of a Common Lisp extension that allows a programmer to
write functional programs that use normal order evaluation, as in non-strict languages like Haskell.
The extension is relatively straightforward, and it appears to be the first one such that is integrated
in the overall Common Lisp framework1 .
1
Introduction
Common Lisp is a functional language (and also an imperative, object-oriented one, which, moreover, can
be used in a declarative fashion). As a functional language it falls in the category of strict languages like
ML [7] and OCaml [5], unlike Haskell [6], which is in the category of normal-order or lazy languages.
That is to say that the following code will enter an infinite loop, should it be executed at the Common Lisp prompt.
cl-prompt> (defun si (condicio ergo alternatio)
(if condicio ergo alternatio))
SI
cl-prompt> (si t 42 (loop))
In a lazy language the function si (if in Latin) would return 42 instead of waiting for the form (loop)
to produce a value.
In a bout of Haskell envy, I decided to look into some extensions to Common Lisp that would introduce
ways to program in a lazy way. The result may sound crazy, and, in fact, a little bit it is.
The notion of lazy evaluation dates back to the Algol days and the notion of by-name parameter
passing. In the Lisp camp, the best known way to introduce a form of lazy evaluation is to implement
streams as described in Structure and Interpretation of Computer Programs (SICP) [1]; incidentally this
form of lazy evaluation is also used by Okasaki [8] in his exposition of functional data structures in ML.
In SICP, streams are implemented using two primitives, force and delay, which can then be used to
build a lazy container (the “stream”) using a macro cons-stream, and two accessors head and tail. A
sufficient implementation in Common Lisp is the following:
1A
version of this note was presented at the 1st European Lisp Symposium 2008, Bordeaux, France
1
Antoniotti, M.
< LLP ΠΛΛ >
CLAZY: Lazy Calling. . .
(defmacro delay (expr) ‘(lambda () ,expr))
(defun force (thunk) (funcall thunk))
(defmacro cons-stream (head tail) ‘(cons ,head (delay ,tail)))
(defun head (lazy-stream) (car lazy-stream))
(defun tail (lazy-stream) (force (cdr lazy-stream)))
At this point there are several Common Lisp packages floating around the net, that implement this flavor
of lazy evaluation. E.g., Heresy [4], funds [2] and FSet [3] are exemplars of this approach. CLAZY goes
off a (different) tangent and provides a more fundamental way to build such lazy constructions.
1.1
Limits of the delay/force Duo
Given delay and force, one could always implement the operator si as a macro using delay, as in
(defmacro si (condicio ergo alternatio)
‘(if (force ,condition)
(force (delay ,ergo))
(force (delay ,alternatio))))
but this is a bit unsatisfactory as far as Haskell envy is concerned. si cannot be funcalled in any meaningful
way and cannot be passed around as we would expect a regular function to be. A different solution is
needed.
2
Defining and Calling Lazy Functions
It is possible to come up with a more satisfactory solution that will allow us to bypass delay and force,
at the price of tweaking the “calling convention”. Then we can write si as:
(deflazy (condicio ergo alternatio)
(if condicio ergo alternatio))
where deflazy defines both lazy and strict versions of the operator.
The lazy function si can now be called as
CL prompt> (lazy:call #’si t 42 (loop))
42
2
Antoniotti, M.
< LLP ΠΛΛ >
CLAZY: Lazy Calling. . .
CL prompt> (lazy:call (lazy #’(lambda (condicio ergo alternatio)
(if condicio
ergo
alternatio)))
t
(+ 20 20 2)
(loop))
42
Figure 1: An example of the use of the special operator lazy.
I.e., lazy:call is the lazy version of funcall. The complexity of writing lazy code is thus moved to the
call points. This may or may not be desirable, but it can be argued that this is a slightly better way than
having to manually force expressions. In any case, the CLAZY approach still uses the delay/force duo
under the hood, and they are available for more manual intervention.
From the example above, it should be apparent that lazy:call is a macro that does something special
with the call, recognizing functions that are defined via deflazy. As a matter of facts, the expansion of
lazy:call looks like this:
(lazy:call <op> <arg1> <arg2> ... <argN>)
=⇒
(funcall <lazyfied op>
<thunked arg1>
<thunked arg2>
...
<thunked argN>)
The “lazy” version of <op> is defined by deflazy and each <thunked argi > is a closed over version of
the argument as if delay was invoked on it.
Of course, a simple version of such idea can be easily implemented with a few macros, however, a
well integrated version within the overall Common Lisp environment requires a few more bits and pieces.
As example, CLAZY wants to make the analogy between lazy:call and funcall as tight as possible.
This means that we need a way to pass (almost) regular lambda’s to lazy:call. This can be done the
special operator lazy, which acts as function; moreover, it does wrap around the function operator as
expected. See Figure (1) for an example.
Extra work is needed to handle &optional and &key parameters, but the overall design lies in this
tweaking of the calling point and in allowing lazy functional objects to be passed around as regular
functions (of course to be called via lazy:call).
3
Antoniotti, M.
2.1
< LLP ΠΛΛ >
CLAZY: Lazy Calling. . .
Example: Lazy Functional Conses
Another example which turns out to be more easily realizable with CLAZY is the standard “conses are
functions” one.
(deflazy conc (head tail)
(lambda (selector)
(ecase selector
(car head)
(cdr tail))))
(deflazy head (cons) (funcall cons ’car))
(deflazy tail (cons) (funcall cons ’cdr))
Now, we can build truly lazy lists2
CL prompt> (defparameter ll
(lazy:call ’conc
1
(lazy:call ’conc
(loop)
(lazy:call ’conc
3
(loop)))))
LL
CL prompt> (head (tail (tail ll)))
3
Or the usual streams from SICP as the integers here below.
(defun integers-from (n)
(lazy:call ’conc n (integers-from (1+ n))))
(defparameter integers (integers-from 0))
Yet, it must be noted that having normal order evaluation at one’s disposal naturally leads to the implementation of much more complex and sophisticated functional software, as in the case of the integrators
in Section 3.5 of [1].
2 Note
where the (loop) calls appear.
4
Antoniotti, M.
< LLP ΠΛΛ >
CLAZY: Lazy Calling. . .
(lazy:call (lazy (lambda (x &key (y (loop) y-supplied-p))
(if y-supplied-p y (+ x 21))))
21)
(lazy:call (lazy (lambda (x &key ((:y yy) (loop)))
(if x (+ x 21) yy)))
21)
(lazy:call (lazy (lambda (x &key ((:y yy) (loop)))
(if x (+ x 21) yy)))
nil :y 42))
Figure 2: &key arguments are dealt with as expected. The answer is always, as expected, 42.
2.2
Extra Considerations
CLAZY is supposed to be used in a very controlled way. While it is true that it adds normal order
evaluation to Common Lisp, the user must remember that s/he is not using Haskell or a similar language.
At its core, Common Lisp is a strict language, which allows side-effects; not a good mix to produce lazy
code in a careless way. See also the note on normal order evaluation in Section 3.5 on streams of [1].
3
Reference Implementation
The CLAZY reference implementation can be found at common-lisp.net. The implementation lies within
a package nicknamed LAZY and is based on the macros lazy:call, lazy:deflazy, and lazy:lazy.
The lazy:call macro is used at calling time (as the name implies). The deflazy macro is used
to define functions. The lazy “special operator” returns a functional object that should be called in a
lazy way, although the system is set up in such a way to “pass through” constant values (as tested by
constanp).
The reference implementation is based on the pre-processing of lambda list arguments by deflazy:
each argument is substituted by an internal name, which is expected to be bound to a thunk generated by
lazy:call as per delay. In the body of a lazy function (or of a lazy lambda) each lambda list argument
is actually re-defined as a a symbol-macrolet, which expands in the appropriate force call. deflazy
installs the lazy version of the function being defined in the property list of the function name.
Ordinary Lambda List Processing. As noted before, CLAZY pre-processes &optional and &key
arguments in such a way to preserve the expected Common Lisp semantics. E.g., the calls in Figure (2)
yield 42 as expected. On the contrary, the implementation does not treat &rest arguments in a special
way (i.e., they are not thunked ), this is because there is no way to access the list forming machinery in
Common Lisp when &rest arguments are present; in a lazy piece of code, the list in the &rest argument
will contain the actual thunks generated as if by delay.
5
Antoniotti, M.
4
< LLP ΠΛΛ >
CLAZY: Lazy Calling. . .
Conclusions
CLAZY is an a exercise in Common Lisp style, which is also useful. The CLAZY library shows how, at
the price of introducing a special call operator (lazy:call), it is possible to introduce normal order or
lazy evaluation in Common Lisp. The extension has the following desirable characteristics: (i) it does not
require the construction of a full blown interpreter implementing lazy evaluation, and (ii) thanks to the
deflazy macro it allows a programmer to write code in the most natural way. It is much more difficult
to achieve the same effect in any other language than Common Lisp, even when the language has macros.
It is the under-the-hood interaction of macros and symbol-macrolet that makes CLAZY possible.
Of course, once this basic machinery is in place, extra Common Lisp incantations can be made and
reader macros put in place as desired.
CLAZY is not perfect of course. The main open issue to complete the integration within the frame
provided by Common Lisp is to work out a way to deal with CLOS methods. One way to achieve this
would be to automatically define a method specializing on thunks for a given generic function. While this
may work, it does open up typing issues3 that need to be worked out in details before proceeding with a
full blown proposal.
References
[1] H. Abelson, J. Sussman, and J. Sussman. Structure and Interpretation of Computer Programs. MIT
Press, second edition, 1996.
[2] A. Baine. Funds: Functional Data Structures in Common Lisp.
Project page at http://common-lisp.net/project/funds, 2007.
[3] S Burson. FSet: a functional set-theoretic collections library.
Project page at http://common-lisp.net/project/fset, 2007.
[4] M. Lamari. Heresy: Haskellesque lazy-list and functional tools with a Common Lisp slant.
Project page at http://cl-heresy.sourceforge.net/Heresy.htm, 2007.
[5] X. Leroy. The OCaml system release 4.02.
Published online at http://caml.inria.fr/pub/docs/manual-ocaml/, September 2014.
[6] S. Marlow (editor). Haskell 2010 Language Report.
Published online at: https://www.haskell.org/onlinereport/haskell2010/, November 2009.
[7] R. Milner, M. Tofte, R. Harper, and D. MacQueen. The Definition of Standard ML (Revised). MIT
Press, 1997.
[8] C. Okasaki. Purely Functional Data Structures. Cambridge University Press, 1998.
[9] K. M. Pitman. The Common Lisp Hyperspec.
Published online at http://www.lisp.org/HyperSpec/FrontMatter/index.html, 1994.
3 lazy:call would need to know the actual resulting type of the argument expressions to meaningfully set up a discrimination for the underlying method.
6
| 6cs.PL
|
Stepwise Transmit Antenna Selection in Downlink
Massive Multiuser MIMO
Ali Bereyhi, Saba Asaad, Ralf R. Müller
arXiv:1802.05148v2 [cs.IT] 19 Feb 2018
Friedrich-Alexander Universität Erlangen-Nürnberg (FAU)
[email protected], [email protected], [email protected]
Abstract—Due to the large power consumption in RF-circuitry
of massive MIMO systems, practically relevant performance measures such as energy efficiency or bandwidth efficiency are neither
necessarily monotonous functions of the total transmit power nor
the number of active antennas. Optimal antenna selection is however computationally infeasible in these systems. In this paper,
we propose an iterative algorithm to optimize the transmit power
and the subset of selected antennas subject to non-monotonous
performance measures in massive multiuser MIMO settings. Numerical results are given for energy efficiency and demonstrate
that for several settings the optimal number of selected antennas
reported by the proposed algorithm is significantly smaller than
the total number of transmit antennas. This fact indicates that
antenna selection in several massive MIMO scenarios not only
reduces the hardware complexity and RF-costs, but also enhances
the energy efficiency of the system.
I. I NTRODUCTION
The ever increasing growth in the number of digital devices
such as smart phones, tablets and personal computers as well
as broadband wireless applications put a heavy strain on wireless community to design the next generation of cellular networks. Considering this backdrop, wireless technologies are
shifting towards systems with higher transmission rates and
lower power consumption; or more precisely systems with
high power efficiency [1]. In this respect, massive MultipleInput Multiple-Output (MIMO) has been considered as one of
leading candidates for future wireless communications [2], [3].
The so-called “power scaling law” in massive MIMO systems demonstrates that for a given transmission rate, the
required transmit power is effectively scaled. More precisely,
for a Base Station (BS) with a large transmit antenna array,
the transmit power of each user is reduced substantially
proportional to the number of antennas [4]. Despite the power
scaling law, the large antenna arrays employed in massive
MIMO systems increase the consumed power at the Radio
Frequency (RF)-chains. In other words, although in massive
MIMO systems the radiated power per each transmit antenna is
extremely low, the growth in the number of antennas increases
the power consumed in the RF circuits which is in general
linearly scaled with the array size [5]. In addition to the large
power consumption, massive MIMO systems with a dedicated
RF-chain at each antenna suffer from multiple other issues
such as high implementation cost and hardware complexity.
This work was supported by the German Research Foundation, Deutsche
Forschungsgemeinschaft (DFG), under Grant No. MU 3735/2-1.
Considering this fact, it is concluded that despite the appealing theoretical benefits of massive MIMO, it still needs to
be developed from the implementational viewpoints. Recent
studies have addressed some of these issues by investigating
approaches such as antenna selection [6], hybrid analog and
digital precoding schemes [7] and spatial modulation [8], in
the context of massive MIMO.
In this study, we focus on antenna selection which we believe
is a promising approach for cost and complexity alleviation in
massive MIMO settings. In general, the optimal antenna selection approach is computationally infeasible in these systems,
since it requires an exhaustive search. Consequently, studies
in this direction either investigate the asymptotic performance
of optimal approaches by means of advanced mathematical
tools such as the replica method [9]–[12], or develop and
analyze suboptimal algorithms with feasible complexity [13].
A body of recent work has demonstrated that by using effective antenna selection strategies, several asymptotic properties
of massive MIMO systems, such as energy efficiency [14] and
channel hardening [13], are essentially maintained with significantly fewer RF chains. Antenna selection has been moreover
shown to be enhancing in several scenarios. For example, the
study in [15] demonstrated that the secrecy performance of a
wiretap channel with large antenna arrays can be significantly
improved via antenna selection compared to the case with
all antennas being active. Accounting for power consumption
in RF-circuitry, energy efficiency, bandwidth efficiency, and
other practically relevant performance measures are neither
necessarily monotonous functions of the total transmit power
nor the number of active antennas. This fact depicts that
antenna selection, as well as transmit power control in general
not only can reduce the hardware cost, but also may enhance
the performance of the system. In this paper, we propose an
iterative method to optimize the total transmit power and the
subset of selected antennas for a desired performance measure.
From mathematical points of view, optimal antenna selection is equivalent to the problem of finding an efficient subset
of integers which maximizes an objective function subject to
a set of constraints. Such problems also arise in other contexts
such as pattern recognition and data mining [16], [17]. There
are therefore a variety of iterative approaches which determine
a suboptimal, but effective, subset with low computational
complexity among which stepwise regression methods have
obtained more credit [16]. Based on these greedy methods,
several studies invoked the results in the literature to develop
iterative antenna selection algorithms for conventional singleuser MIMO systems1 [18]–[21] considering the spectral efficiency as the performance measure. Regarding the promising
results in the literature of conventional MIMO and considering
the large nature of massive MIMO systems, the stepwise regression methods can propose effective selection algorithms in
these systems with feasible computational complexity.
Contributions
In this paper, we consider the downlink massive multiuser
MIMO scenario with linear precoding and develop a class
of iterative antenna selection algorithms based on stepwise
regression methods to optimize the total transmit power and
the subset of selected antennas subject to a non-monotonous
performance measure. Given a set of constraints on the number
of possible active antennas and transmit power, these algorithms iteratively select a subset of antennas and tune the
transmit power such that the desired performance measure is
maximized. In addition to significant gains reported by the
proposed algorithm, numerical results are given for energy
efficiency and demonstrate that the antenna selection can also
enhance the energy efficiency compared to the case of full
antenna selection in some massive MIMO settings.
Notation
Throughout the paper the following notations are adopted.
Scalars, vectors and matrices are represented with non-bold,
bold lower case and bold upper case letters, respectively. HH ,
HT and H∗ indicate the Hermitian, transpose and conjugate
of H, respectively, and IN is the N × N identity matrix.
The Euclidean norm of x is denoted by kxk and log (·)
indicates the binary logarithm. Moreover, E {·} represents the
mathematical expectation. The subset of A containing the
members of A which do not belong to B is moreover denoted
by A\B. For the sake of compactness, we abbreviate the set
of integers {1, . . . , N } with [N ].
II. P ROBLEM F ORMULATION
Consider a downlink transmission in a multiuser MIMO system with K single-antenna users. The base station is equipped
with a transmit antenna array of length N and employs the
Transmit Antenna Selection (TAS) protocol S to select a
subset of L antennas for transmission. Let yk ∈ C denote
the symbol received at user k for k ∈ [K] and G ∈ CN ×K
represent the matrix of channel coefficients between user
terminals and the transmit antenna array. It is assumed that the
system operates in standard Time-Division Duplexing (TDD)
mode meaning that the channel is reciprocal in the uplink and
doswnlink. In this case, by using the reciprocity of the channel,
the received vector y := [y1 , . . . , yK ]T is written as
y = HT
Lx + z
(1)
where x ∈ CL×1 denotes the transmit vector and contains the symbols being transmitted over L active antennas,
1 By conventional MIMO systems we mean systems with few number of
antennas at the receive and transmit sides.
z ∈ CK×1 is circularly symmetric zero-mean and unitvariance complex Gaussian noise, and HL ∈ CL×K describes
the channel between users and L transmit antennas selected
by S, i.e., HL = S(G; L). The transmit vector x is moreover
constructed at the base station by linear precoding over active
antennas, i.e.,
√
(2)
x = P AL s
where s := [s1 , . . . , sK ]T with sk ∈ C being the data symbol
for user k, and AL ∈ CL×K represents the precoding matrix
for HL , i.e., AL = Pre(HL ). It is moreover assumed that
E |sk |2 = 1 and the total transmit power is P , or equivalently
E Tr{AL AH
L } = 1.
A. Achievable Rates and Power Consumption Model
The achievable rate at user k is bounded from below by [22]
Rk (L, P ) = log (1 + SINRk (L, P ))
(3)
with SINRk (L, P ) being defined for k ∈ [K] as
SINRk (L, P ) :=
tk (HL , AL )P
1 + uk (HL , AL )P
(4)
2
where tk (HL , AL ) := |hT
k (L)ak (L)| with hk (L) and ak (L)
denoting the kth column vector of HL and AL , respectively,
and uk (HL , AL ) is the kth multiuser interference coefficient
defined as
uk (HL , AL ) :=
K
X
j=1,j6=k
2
|hT
k (L)aj (L)| .
The total power consumed in this system, including the receive
side, is moreover modeled as [4]
Rx
Q(L, P ) = ξP + LQTx
cr + KQcr + (K + 1) Qsync
(5)
where ξ represents the inverse efficiency factor of the power
Rx
amplifiers used at the base station, QTx
cr and Qcr are respectively the circuit power consumed at each transmit and receive
RF-chain, and Qsync denotes the power consumption at the
local oscillators utilized at the base station and user terminals
for frequency synthesis.
B. Nonuniform Average Spectral Efficiency
Let wk ∈ R+ be a weighting coefficient which describes the
service quality required for k. The average spectral efficiency
with respect to the vector w = [w1 , . . . , wK ]T is then given by
R(L, P |w) :=
K
1 X
wk Rk (L, P ).
K
(6)
k=1
This nonuniform average spectral efficiency recovers the conventional definition of the average spectral efficiency by setting
wk = 1 for k ∈ [K]. Moreover, nonuniform choices of weights
address cases in which different users enjoy different priorities.
C. Nonuniform Average Energy Efficiency
The average energy efficiency for a given w is defined as
E(L, P |w) :=
R(L, P |w)
Q(L, P )
(7)
which determines the average data rate per each unit of energy
consumed in this system.
III. I TERATIVE TAS AND P OWER C ONTROL A LGORITHM
In this setting, we deal with two correlated challenges of
TAS and transmit power control. More precisely, the measure of performance in general does not necessarily grow
monotonically in terms of the number of active antennas L
and the total transmit power P . For several measures, there
are some scenarios in which the performance is optimized
at some L < Lmax and P < Pmax where Lmax and Pmax
are the maximum number of selected antennas and maximum
possible transmit power constrained by hardware restrictions.
An example of such a scenario is the case with the average
energy efficiency as a measure of performance. In this case, the
increase in the number of active antennas as well as the growth
in the total transmit power can increase both the average
spectral efficiency and the consumed power, and consequently,
the energy efficiency could be maximized at some L and P
such that L < Lmax and P < Pmax .
In order to formulate this problem, we denote the measure
of performance for a weighting vector w with M(SL , L, P |w)
when L transmit antennas, whose indices are in SL ⊆ [N ],
are selected, and the total transmit power is set to P . In this
case, the problem of optimal TAS and transmit power control
is to find L⋆ , the set S⋆L⋆ and the transmit power P ⋆ such that
(S⋆L⋆ , L⋆ , P ⋆ ) =
argmax
SL ⊆[N ],L≤Lmax
M(SL , L, P |w)
The performance measure M(L, P |w) is linearly written
in terms of Rk (L, P ), e.g., M(L, P |w) = R(L, P |w),
M(L, P |w) = E(L, P |w) or some linear combination of
spectral and energy efficiency.
The algorithms start from ℓ = 1 and select antennas up to
some L in which L ≤ Lmax . For initialization, we set the
antenna with strongest channel to be the first selected antenna
which means H1 ∈ C1×K is chosen to be the row of G with
maximum norm. Consequently, the initial precoding matrix is
given by A1 = Pre(H1 ). In this case, we set the initial value
for the transmit power to be P1 such that
•
(8)
0≤P ≤Pmax
for some Lmax and Pmax . Some examples for M(SL , L, P |w)
are R(L, P |w) and E(L, P |w). For the sake of brevity, we
further drop the argument SL and represent the performance
measure with M(L, P |w) in the rest of this manuscript.
The optimization problem in (8) is computationally infeasible due to the large number of searches needed for TAS. In
fact, even for cases in which M(SL , L, P |w) is convex with
respect to P , antenna selection needs an exhaustive search of
size N
L which grows significantly large in N . We therefore
take an alternative approach based on the stepwise regression
method and design an iterative algorithm which jointly selects
transmit antennas and tunes the total transmit power P .
A. Primary Assumptions and Notations
In the following sections, we develop iterative algorithms
for a large class of performance measures. For deriving the
algorithms, we consider the following set of assumptions.
• The number of selected transmit antennas is limited by
Lmax , i.e., L ≤ Lmax antennas are selected.
• The maximum possible transmit power is Pmax which
means that P ≤ Pmax in each transmission interval.
P1 = argmax M(1, P |w).
(9)
0≤P ≤Pmax
Here, M(1, P |w) represents the performance measure, e.g.,
R(1, P |w) or E(1, P |w), considering the channel matrix to be
H1 and the precoding matrix to be A1 . With this initialization,
we derive a stepwise update rule for each step.
B. Stepwise Update Rule
Assume that we have already selected ℓ < Lmax transmit
antennas, and now intend to select a new antenna. We represent
the index of the selected antenna in step ℓ + 1 with n⋆
where n⋆ ∈ [N ]\S(ℓ) with S(ℓ) containing the indices of
the antennas selected in steps [ℓ]. Let us denote the n⋆ th row
of G with the vector gT (n⋆ ) = [g1 (n⋆ ), . . . , gk (n⋆ )]. Without
loss of generality, one can write
hk (ℓ + 1)=[hT
(ℓ), gk (n⋆ )]T
(10a)
pk
T
T
⋆
⋆ T
⋆
ak (ℓ + 1)=[ µ(ℓ, n ) ak (ℓ) + dk (ℓ, n ), bk (ℓ, n )] (10b)
where µ(ℓ, n⋆ ), dk (ℓ, n⋆ ) ∈ Cℓ×1 and bk (ℓ, n⋆ ) are derived in
terms of Hℓ , Aℓ and g(n⋆ ) depending on the precoding rule
Pre(·). We later derive these parameters for different linear
precodings, namely Maximum Ratio Transmission (MRT),
Zero Forcing (ZF) and Regularized Zero Forcing (RZF).
By substituting the new channel and precoding coefficients
in (4), we have
SINRk (ℓ + 1, P ) =
µ(ℓ, n⋆ )tk (Hℓ , Aℓ )P + ǫk (ℓ, n⋆ )P
1 + µ(ℓ, n⋆ )uk (Hℓ , Aℓ )P + ψk (ℓ, n⋆ )P
(11)
where ǫk (ℓ, n⋆ ) and ψk (ℓ, n⋆ ) are given by
ǫk (ℓ, n⋆ ) := |δkk (ℓ, n⋆ )|2
np
o
∗
⋆
+ 2Re
µ(ℓ, n⋆ ) hT
(12a)
k (ℓ)ak (ℓ)δkk (ℓ, n )
⋆
ψk (ℓ, n ) :=
K
X
j=1,j6=k
+ 2Re
|δkj (ℓ, n⋆ )|2
o
np
∗
⋆
µ(ℓ, n⋆ ) hT
(12b)
k (ℓ)aj (ℓ)δkj (ℓ, n )
with δkj (ℓ, n⋆ ) being defined as
⋆
⋆
⋆
δkj (ℓ, n⋆ ) := hT
k (ℓ)dj (ℓ, n ) + gk (n )bj (ℓ, n ).
(13)
Here, the argument n⋆ indicates the dependency on the coefficients of the new selected channel. By substituting (11) into
(3), one can finally write
Rk (ℓ + 1, P ) = Rk (ℓ, P ) + log θk (ℓ, n⋆ , P )φk (ℓ, P )
(14)
Algorithm 1 Iterative TAS and Power Control
Input: G, Pmax , Lmax , precoding rule Pre(·) and performance measure M(ℓ, P |w)
Initiate Let ℓ = 1 and
n⋆ = argmaxkG(n, :)k
where φk (ℓ, P ) is given by
φk (ℓ, P ) :=
1/P + uk (Hℓ , Aℓ )
1/P + uk (Hℓ , Aℓ ) + tk (Hℓ , Aℓ )
(15)
and set H1 = G(n⋆ , :) and A1 = Pre(H1 ). Set
P1 = argmax M(1, P |w)
and θk (ℓ, n⋆ , P ), which depends on the selected antenna, reads
θk (ℓ,n⋆ , P ) :=
µ(ℓ, n⋆ )tk (Hℓ , Aℓ ) + ǫk (ℓ, n⋆ )
1+
.
1/P + µ(ℓ, n⋆ )uk (Hℓ , Aℓ ) + ψk (ℓ, n⋆ )
M(ℓ + 1, P |w) = M(ℓ, P |w) + Θ(ℓ, n⋆ , P )
(16)
(17)
⋆
where Θ(ℓ, n , P ) is derived as a function of θk (ℓ, n , P ) and
φk (ℓ, P ). We refer to (17) as the stepwise update rule.
C. Iterative Approach for TAS and Power Control
Based on the stepwise update rule, we now propose an iterative TAS and power control algorithm. Here, we state the
algorithm for a general precoding and performance measure
and give detailed derivations for different measures and precoding matrices in the next sections.
The algorithm follows the stepwise regression method. In
each step, we select the antenna which results in the maximum
possible growth in the performance measure. Although this
approach does not necessarily lead to the optimal solution,
it enjoys low computational complexity while demonstrating
a better performance compared to other low complexity approaches. Using the stepwise regression method, one starts
with the initialization H1 , A1 and P1 for ℓ = 1 and chooses
the transmit antenna n⋆ in step ℓ as
n⋆ = argmax Θ(ℓ, n, Pℓ ).
n∈[N ]\S(ℓ)
(18)
T
The channel matrix is then updated as Hℓ+1 = [HT
ℓ , gn⋆ ]
1
and the transmit power for the next step is set to be
Pℓ+1 = argmax M(ℓ + 1, P |w).
(19)
0≤P ≤Pmax
In general, the algorithm can select antennas until it reaches
Lmax . Nevertheless, the stepwise update rule may stop growing at some step ℓ < Lmax . We therefore stop the algorithm
at step L⋆ ≤ Lmax when either L⋆ = Lmax or
Θ(L⋆ , n⋆ , PL⋆ ) ≤ 0.
(22)
0≤P ≤Pmax
and S(1) = {n⋆ }.
The stepwise variation of Rk (ℓ, P ) in (14) lets us to select a
new antenna in step ℓ + 1 given the selected antennas in steps
[ℓ] and the power P , such that Rk (ℓ, P ) has the maximum
growth. Although such a selection does not necessarily give
the optimal selection, it leads to an efficient algorithm with
low complexity. Starting from (14), the performance measure
is updated in each step as
⋆
(21)
n∈[N ]
(20)
1 It is worth to note that this optimization involves only a single argument
P , and its dimension does not grow with the system size.
while ℓ < Lmax
n⋆ = argmax Θ(ℓ, n, Pℓ )
n∈[N ]\S(ℓ)
(23)
let Θ⋆ = Θ(ℓ, n⋆ , Pℓ ).
if Θ⋆ ≤ 0 then
break
end if
⋆
T T
Set Hℓ+1 = HT
and
ℓ , G(n , :)
D(ℓ, n⋆ ) = [d1 (ℓ, n⋆ ), . . . , dK (ℓ, n⋆ )]
(24a)
T
(24b)
⋆
⋆
⋆
b(ℓ, n ) = [b1 (ℓ, n ), . . . , bK (ℓ, n )] ,
and update the precoding matrix as
p
µ(ℓ, n⋆ )Aℓ + D(ℓ, n⋆ )
.
Aℓ+1 =
bT (ℓ, n⋆ )
(25)
Define the scalar function
M(ℓ + 1, P |w) = M(ℓ, P |w) + Θ(ℓ, n⋆ , P )
(26)
and update the transmit power as
Pℓ+1 = argmax M(ℓ + 1, P |w).
(27)
0≤P ≤Pmax
Update S(ℓ + 1) = S(ℓ) ∪ {n⋆ } and ℓ = ℓ + 1.
end while
Output: L⋆ = ℓ, PL⋆ = Pℓ and S(L⋆ ) = S(ℓ).
The outputs of this algorithm are then L⋆ , PL⋆ and
The algorithm is illustrated in Algorithm 1 in details.
S(L⋆ ).
The proposed algorithm considers a general setting. In the
sequel, we derive the detailed formulation for several performance measures and linear precoders. Namely we determine
the stepwise update rule for the average spectral and energy
efficiency while considering MRT, ZF and RZF precoders.
IV. D ERIVATIONS
FOR
D IFFERENT M EASURES
As mentioned before, the average spectral efficiency and
the average energy efficiency are some common choices for
the performance measure. We therefore derive explicitly the
stepwise update rule for each of these measures.
A. Average Spectral Efficiency
B. Regularized Zero Forcing
For this performance measure M(ℓ, P |w) is given by
K
1 X
M(ℓ, P |w) = R(ℓ, P |w) =
wk Rk (ℓ, P ).
K
For RZF precoding, we have
(28)
k=1
⋆
Therefore, Θ(ℓ, n , P ) in this case reads
K
1 X
wk log θk (ℓ, n⋆ , P )φk (ℓ, P ).
Θ(ℓ, n⋆ , P ) =
K
(29)
B. Average Energy Efficiency
R(ℓ, P |w)
M(ℓ, P |w) = E(ℓ, P |w) =
Q(ℓ, P )
(30)
with Q(ℓ, P ) being defined in (5). Noting that
Q(ℓ + 1, P ) = Q(ℓ, P ) +
QTx
cr
(31)
the stepwise growth term is found as in (32) on the top of the
next page.
(41)
∗
with Jℓ (λ) := HT
ℓ Hℓ + λ IK . Noting that for any λ
Jℓ+1 (λ)−1 = Jℓ (λ)−1 − r(ℓ, n⋆ ; λ)rH (ℓ, n⋆ ; λ),
(42)
⋆
For different linear precoders, the terms µ(ℓ, n ), dk (ℓ, n )
and bk (ℓ, n⋆ ) are of different forms. In this section, we derive
these terms for the MRT, ZF and RZF precoders. For the sake
of compactness, we define the matrix D(ℓ, n⋆ ) ∈ Cℓ×K and
the vector b(ℓ, n⋆ ) ∈ CK×1 as
⋆
⋆
D(ℓ, n ) = [d1 (ℓ, n ), . . . , dK (ℓ, n )]
(34a)
(43)
where we define
Jℓ (λ)−1 g(n⋆ )
r(ℓ, n⋆ ; λ) := p
1 + gH (n⋆ )Jℓ (λ)−1 g(n⋆ )
(44)
Note that gH (n⋆ )Jℓ (λ)−1 g(n⋆ ) ≥ 0 for any λ, since Jℓ (λ)−1
is a positive semi-definite matrix. Consequently, one can derive
the rank-one update for βrzf (ℓ; λ) as
D IFFERENT L INEAR P RECODERS
⋆
⋆
where the factor βrzf (ℓ; λ) reads
−1/2
βrzf (ℓ; λ) = Tr{Jℓ (λ)−2 Jℓ (0)}
(40)
one utilizes the Sherman-Morrison formula [23] and writes
In this case, we have
FOR
−1
Jℓ+1 (λ) = Jℓ (λ) + g(n⋆ )gH (n⋆ ),
k=1
V. D ERIVATIONS
∗
Aℓ = βrzf (ℓ; λ) H∗ℓ HT
ℓ Hℓ + λ IK
2 (ℓ
βrzf
1
1
= 2
+ ∆(ℓ, n⋆ ; λ),
+ 1; λ)
βrzf (ℓ; λ)
(45)
where ∆(ℓ, n⋆ ; λ) is defined in (33) on the top of the next
page with E(ℓ, n⋆ ; λ) ∈ CK×K being defined as
E(ℓ, n⋆ ; λ) :=
kr(ℓ, n⋆ ; λ)k2 IK − Jℓ (λ)−1 Jℓ (0) − Jℓ (0)Jℓ (λ)−1 . (46)
Consequently, one can conclude that
−1
2
µ(ℓ, n⋆ )= 1 + βrzf
(ℓ; λ)∆(ℓ, n⋆ ; λ)
(47a)
The single-rank updates in (10b) can then be represented as
p
∗
⋆
H
⋆
⋆
#
"p
⋆
D(ℓ, n )=βrzf (ℓ; λ) µ(ℓ, n )Hℓ r(ℓ, n ; λ)r (ℓ, n ; λ) (47b)
µ(ℓ, n⋆ )Aℓ + D(ℓ, n⋆ )
s
.
(35)
Aℓ+1 =
2
βrzf
(ℓ; λ)µ(ℓ, n⋆ )
T
⋆
⋆
b (ℓ, n )
r∗ (ℓ, n⋆ ; λ)
(47c)
b(ℓ, n )=
H
1 + g (n⋆ )Jℓ (λ)−1 g(n⋆ )
A. Maximum Ratio Transmission
where we further drop λ in the left hand sides of (47a)-(47c)
The MRT precoder reads
for the sake of compactness.
Aℓ = βmrt (ℓ) H∗ℓ
(36)
C. Zero Forcing
∗ T −1/2
where βmrt (ℓ) = Tr{Hℓ Hℓ }
. Consequently, βmrt (ℓ) is
The ZF precoder deduces from RZF precoding by setting
updated in each step by
λ = 0. In this case,
⋆
⋆
⋆
T
b(ℓ, n ) = [b1 (ℓ, n ), . . . , bK (ℓ, n )] .
(34b)
1
1
⋆ 2
2 (ℓ + 1) = β 2 (ℓ) + kg(n )k .
βmrt
mrt
(37)
From (37), one can show that
Aℓ
p
2 (ℓ)kg(n⋆ )k2
1 + βmrt
Aℓ+1 =
.
βmrt (ℓ) gH (n⋆ )
p
2
⋆
2
1 + βmrt (ℓ)kg(n )k
(38)
As a result, it is concluded that D(ℓ, n⋆ ) = 0 and
−1
2
µ(ℓ, n⋆ ) = 1 + βmrt
(ℓ)kg(n⋆ )k2
p
b(ℓ, n⋆ ) = βmrt (ℓ) µ(ℓ, n⋆ ) g∗ (n⋆ ).
(39a)
(39b)
Aℓ = βzf (ℓ) H∗ℓ Jℓ (0)−1
(48)
in which βzf (ℓ) = βrzf (ℓ; 0). By setting λ = 0 in the results
derived for RZF, we have
1
1
= 2
− kr0 (ℓ, n⋆ )k2 ,
(49)
2
βzf (ℓ + 1)
βrzf (ℓ)
where we define r0 (ℓ, n⋆ ) := r(ℓ, n⋆ ; 0). Moreover,
−1
2
µ(ℓ, n⋆ ) = 1 − βzf
(ℓ)kr0 (ℓ, n⋆ )k2
p
⋆
D(ℓ, n⋆ ) = βzf (ℓ) µ(ℓ, n⋆ ) H∗ℓ r0 (ℓ, n⋆ )rH
0 (ℓ, n )
s
2
βzf
(ℓ)µ(ℓ, n⋆ )
r∗ (ℓ, n⋆ ).
b(ℓ, n⋆ ) =
H
1 + g (n⋆ )Jℓ (0)−1 g(n⋆ ) 0
(50a)
(50b)
(50c)
"
#
K
1
1 X
⋆
Tx
Θ(ℓ, n , P ) =
wk log θk (ℓ, n , P )φk (ℓ, P ) − Qcr E(ℓ, P |w) .
Q(ℓ + 1, P ) K
⋆
(32)
k=1
∆(ℓ, n⋆ ; λ) =
kr(ℓ, n⋆ ; λ)k2
+ rH (ℓ, n⋆ ; λ)E(ℓ, n⋆ ; λ)r(ℓ, n⋆ ; λ)
1 + gH (n⋆ )Jℓ (λ)−1 g(n⋆ )
0.9
Proposed Algorithm
EEavg in [bits/Joule]
0.8
Random TAS with L = L⋆
Random TAS with L = Lmax
0.7
0.6
0.5
0.4
0.3
20
40
60
80
100
120
Lmax
Fig. 1: Energy efficiency versus the maximum number of selected
antennas Lmax considering N = 128 transmit antennas and K = 4
users. Here, the power constraint is Pmax = 0 dB and the parameters
Tx
of RF-chains are ξ −1 = 0.4, QRx
cr = Qcr = 48 mW and Qsync =
62 mW. As the figure shows, the proposed algorithm significantly
outperforms the random TAS approaches with power control.
VI. N UMERICAL R ESULTS
Throughout the numerical investigations, we have considered independent and identically distributed (i.i.d.) Rayleigh
fading channels. Hence, the entries of G are i.i.d. zero-mean
and unit-variance complex Gaussian random variables. The
system parameters, such as the energy efficiency and number
of selected antennas, are then averaged over multiple realizations numerically. Due to the linear computational complexity
of the proposed algorithm, Monte-Carlo simulations are feasible even for large dimensions. It is worth moreover to indicate
that the optimal search for the dimensions considered here are
practically infeasible due to its high computational complexity.
For the MRT precoder, we have sketched the average energy
efficiency defined as
EEavg = E {E(L⋆ , P ⋆ |w)}
(51)
versus Lmax where L⋆ and P ⋆ are the outputs of Algorithm 1
for the MRT precoder and E(ℓ, P |w). The performance of the
proposed algorithm is moreover compared to the performance
of random TAS algorithms with power control.
Fig. 1 shows the energy efficiency as a function of Lmax for
K = 4 users and N = 128 available transmit antennas when
K ≤ Lmax ≤ N . The transmit power is considered to be
(33)
limited by Pmax = 0 dB and the transmit and receive power
amplifiers are assumed to have efficiency factor ξ −1 = 0.4.
The consumed power at the RF chains and synchronization
Rx
oscillators are moreover set to be QTx
cr = Qcr = 48 mW
and Qsync = 62 mW. In this figure, the blue line shows the
performance of the proposed algorithm in which the selection
subset as well as the transmit power is being iteratively tuned
via the stepwise approach. For the sake of comparison, we
have also plotted the energy efficiency of two random TAS
algorithms with power control. The first random algorithm,
whose performance is depicted by the yellow line, selects
Lmax out of N transmit antennas and chooses the transmit
power such that the energy efficiency is maximized. In the
other random approach, shown in the figure with the red
line, the number of selected transmit antennas for each Lmax
is set to L⋆ in which L⋆ denotes the number of selected
antennas via the proposed algorithm and is a function of
Lmax . The latter random algorithm then selects L⋆ transmit
antennas at random and optimize the energy efficiency with
respect to P subject to P ≤ Pmax . As Fig. 1 depicts, the
proposed algorithm significantly outperforms the random TAS
approaches with power control1 . The figure furthermore shows
that even in the case of random selection, the increase in the
number of selected antennas does not necessarily improve the
performance. In fact for large values of Lmax , the random
algorithm with L⋆ < Lmax selected antennas enhances the
energy efficiency compared to the random TAS approach with
Lmax active transmit antennas.
In order to illustrate the latter observation, we have further
sketched in Fig. 2 the average number of selected antennas
given by the proposed algorithm, i.e., E {L⋆ }, versus Lmax
for the setting considered in Fig. 1. As it demonstrates, the
algorithm up to Lmax = 23 selects the almost Lmax transmit
antennas. Nevertheless, for larger values of Lmax it stops
to select more antennas around L⋆ ≈ 24. This means that
further growth in the number of selected antennas degrades
the energy efficiency of the system. Such an observation is
intuitive, due to the fact that the growth in the number of active
antennas can both increase the rate and consumed power. In
order to observe the latter intuition further, we have plotted
the energy efficiency of a modified version of the proposed
algorithm, for the setting considered in previous figures, in
Fig. 3. In this version of the algorithm, the proposed algorithm
is enforced to select exactly Lmax transmit antennas via the
1 Note that when L
max grows with N , random selection is a good bound
on the performance of several practical TAS algorithms; see [9].
0.9
120
Proposed Algorithm
Upper bound E {L⋆ } = Lmax
0.8
EEavg in [bits/Joule]
100
E {L⋆ }
80
60
40
0.7
0.6
0.5
0.4
20
Proposed Algorithm
Random TAS with L = Lmax
Stepwise TAS with L = Lmax
0.3
20
40
60
80
100
120
Lmax
20
40
60
80
100
120
Lmax
Fig. 2: Average number of active antennas selected via the proposed
algorithm versus the maximum number of selected antennas considering N = 128 and K = 4. Here, it is assumed that Pmax = 0 dB,
Tx
ξ −1 = 0.4, QRx
cr = Qcr = 48 mW and Qsync = 62 mW. The
figure depicts that the algorithm stops selecting transmit antennas at
L⋆ ≈ 24 which is significantly smaller than N .
Fig. 3: Energy efficiency versus the maximum number of selected
antennas considering the proposed algorithm as well as the modified
version of the stepwise approach in which exactly Lmax antennas
are selected. The parameters are set similar to those in Fig. 1. As it
depicts, the energy efficiency in the modified algorithm is maximized
at Lmax = 24 which agrees with the results reported in Fig. 2.
stepwise regression method. In fact in this case, the algorithm
does not break at L⋆ and uses forward selection to collect a
subset of Lmax transmit antennas. As the figure shows, the
modified algorithm meets the maximum around Lmax ≈ 24
which agrees with the observation in Fig. 2. At Lmax = N ,
moreover, it lies on the random TAS since they both select
all transmit antennas and optimize the transmit power subject
to the energy efficiency. Noting that the proposed algorithm
needs L⋆ number of iterations, the result indicates that the
complexity is further reduced for Lmax growing with N , since
in the large limit L⋆ does not necessarily grow with Lmax .
[2] E. G. Larsson, O. Edfors, F. Tufvesson, and T. L. Marzetta, “Massive
MIMO for next generation wireless systems,” IEEE Communications
Magazine, vol. 52, no. 2, pp. 186–195, 2014.
[3] J. Hoydis, S. Ten Brink, and M. Debbah, “Massive MIMO in the UL/DL
of cellular networks: How many antennas do we need?” IEEE Journal
on selected Areas in Communications, vol. 31, no. 2, pp. 160–171, 2013.
[4] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and spectral
efficiency of very large multiuser MIMO systems,” IEEE Transactions
on Communications, vol. 61, no. 4, pp. 1436–1449, 2013.
[5] S. Cui, A. J. Goldsmith, and A. Bahai, “Energy-efficiency of MIMO
and cooperative MIMO techniques in sensor networks,” IEEE Journal
on Selected Areas in Communications, vol. 22, no. 6, pp. 1089–1098,
2004.
[6] A. F. Molisch, M. Z. Win, Y.-S. Choi, and J. H. Winters, “Capacity of
MIMO systems with antenna selection,” IEEE Transactions on Wireless
Communications, vol. 4, no. 4, pp. 1759–1772, 2005.
[7] L. Liang, W. Xu, and X. Dong, “Low-complexity hybrid precoding
in massive multiuser MIMO systems,” IEEE Wireless Communications
Letters, vol. 3, no. 6, pp. 653–656, 2014.
[8] M. Di Renzo, H. Haas, A. Ghrayeb, S. Sugiura, and L. Hanzo, “Spatial modulation for generalized MIMO: Challenges, opportunities, and
implementation,” Proceedings of the IEEE, vol. 102, no. 1, pp. 56–103,
2014.
[9] A. Bereyhi, M. A. Sedaghat, S. Asaad, and R. R. Müller, “Nonlinear
precoders for massive MIMO systems with general constraints,” International ITG Workshop on Smart Antennas (WSA), pp. 1–8, 2017.
[10] A. Bereyhi, M. A. Sedaghat, and R. R. Müller, “Asymptotics of nonlinear
LSE precoders with applications to transmit antenna selection,” in IEEE
International Symposium on Information Theory (ISIT), pp. 81–85, 2017.
[11] M. A. Sedaghat, A. Bereyhi, and R. R. Müller, “A new class of nonlinear precoders for hardware efficient massive MIMO systems,” IEEE
International Conference on Communications (ICC), pp. 1–7, 2017.
[12] M. A. Sedaghat A. Bereyhi, and R. R. Müller, “Least Square Error
Precoders for Massive MIMO With Signal Constraints: Fundamental
Limits,” IEEE Transactions on Wireless Communications, vol. 17, no. 1,
pp. 667–679, 2018.
[13] S. Asaad, A. Bereyhi, R. R. Müller, and A. M. Rabiei, “Asymptotics of
transmit antenna selection: Impact of multiple receive antennas,” IEEE
International Conference on Communications (ICC), pp. 1–6, 2017.
[14] Z. Wang and L. Vandendorpe, “Antenna selection for energy efficient
MISO systems,” IEEE Communications Letters, 2017.
[15] S. Asaad, A. Bereyhi, R. R. Müller, R. F. Schaefer, and A. M. Rabiei,
“Optimal number of transmit antennas for secrecy enhancement in mas-
VII. C ONCLUSION
This paper has proposed a class of iterative algorithms for
TAS and transmit power control in downlink massive multiuser
MIMO systems with linear precoders. The algorithms follow
the approach of stepwise regression methods and enjoy low
computational complexity. The results have shown significant
gains obtained via the proposed algorithm. It has been also
demonstrated that antenna selection, in addition to RF-cost
reduction, could enhance the performance in several massive
MIMO settings. The derivations given in this paper can be
more enlightened by further numerical investigations which
will be added in the forthcoming version of the manuscript.
Using tools from large-system analysis, the asymptotic outputs
of the algorithm can be derived in terms of explicit fixed-point
equations. Such investigations are possible extensions of this
work and are currently ongoing.
R EFERENCES
[1] S. Verdú, “Spectral efficiency in the wideband regime,” IEEE Transactions on Information Theory, vol. 48, no. 6, pp. 1319–1343, 2002.
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
sive MIMOME channels,” IEEE Global Communications Conference
(GLOBECOM), pp. 1–6, 2017.
R. O. Duda, P. E. Hart, and D. G. Stork, Pattern classification. John
Wiley & Sons, 2012.
J. Han, J. Pei, and M. Kamber, Data mining: concepts and techniques.
Elsevier, 2011.
A. Gorokhov, D. A. Gore, and A. J. Paulraj, “Receive antenna selection
for MIMO spatial multiplexing: Theory and algorithms,” IEEE Transactions on Signal Processing, vol. 51, no. 11, pp. 2796–2807, 2003.
M. Gharavi-Alkhansari and A. B. Gershman, “Fast antenna subset
selection in MIMO systems,” IEEE Transactions on Signal Processing,
vol. 52, no. 2, pp. 339–347, 2004.
S. Sanayei and A. Nosratinia, “Capacity maximizing algorithms for joint
transmit-receive antenna selection,” in IEEE Thirty-Eighth Asilomar
Conference on Signals, Systems and Computers, vol. 2., pp. 1773–1776,
2004.
X. Zhou, B. Bai, and W. Chen, “An iterative algorithm for joint antenna
selection and power adaptation in energy efficient MIMO,” in IEEE
International Conference on Communications (ICC), pp. 3812–3816,
2014.
G. Caire, N. Jindal, M. Kobayashi, and N. Ravindran, “Multiuser MIMO
achievable rates with downlink training and channel state feedback,”
IEEE Transactions on Information Theory, vol. 56, no. 6, pp. 2845–
2866, 2010.
M. S. Bartlett, “An inverse matrix adjustment arising in discriminant
analysis,” The Annals of Mathematical Statistics, vol. 22, no. 1, pp.
107–111, 1951.
| 7cs.IT
|
1
Gray Box Identification of State-Space Models
Using Difference of Convex Programming
arXiv:1611.04359v1 [cs.SY] 14 Nov 2016
Chengpu Yu, Lennart Ljung and Michel Verhaegen
Abstract—Gray-box identification is prevalent in modeling
physical and networked systems. However, due to the nonconvex nature of the gray-box identification problem, good initial
parameter estimates are crucial for a successful application. In
this paper, a new identification method is proposed by exploiting
the low-rank and structured Hankel matrix of impulse response.
This identification problem is recasted into a difference-of-convex
programming problem, which is then solved by the sequential
convex programming approach with the associated initialization
obtained by nuclear-norm optimization. The presented method
aims to achieve the maximum impulse-response fitting while
not requiring additional (non-convex) conditions to secure nonsingularity of the similarity transformation relating the given
state-space matrices to the gray-box parameterized ones. This
overcomes a persistent shortcoming in a number of recent
contributions on this topic, and the new method can be applied
for the structured state-space realization even if the involved
system parameters are unidentifiable. The method can be used
both for directly estimating the gray-box parameters and for
providing initial parameter estimates for further iterative search
in a conventional gray-box identification setup.
Index Terms—Structured state-space model, convex-concave
procedure, nuclear-norm optimization.
I. I NTRODUCTION
Nowadays, the control and identification of structured statespace system model have attracted great attention in the control
community. There are two main sources of structured statespace models: the modeling of practical physical systems [1]–
[3] and the description of networked systems [4], [5]. When
modeling physical systems, the non-zero entries of the system
matrices always have physical meanings such as masses,
velocity, acceleration, and so on. Identification of the physical
parameters can provide us a better understanding of the inner
physical mechanism of the investigated object. On the other
hand, a network connected system often can be represented
as a structured system with the structure straightforwardly
determined by the interconnections among the involved subsystems. Identification of such kind of structured system
models provides the foundation for the model-based network
control.
In the literature, there are two kinds of methods to identify
structured state-space models. One is the traditional graybox set-up, to identify the parameterized state-space models
C. Yu and M. Verhaegen are with the Delft Center for Systems and
Control, Delft University, Delft 2628CD, Netherlands ([email protected],
[email protected])
L. Ljung is with the Division of Automatic Control, Department of
Electrical Engineering, Linkoping University, Sweden ([email protected])
This work is sponsored by the European Research Council, Advanced Grant
Agreement No. 339681.
directly from the input-output (IO) data using the predictionerror method [2], [3]. Since the involved identification problem
is always a non-convex optimization problem, the conventional
nonlinear optimization methods, such as regularized GaussNewton method [2, Section 10.2], and the gradient project
method [3, Chapter 7], are sensitive to the initial parameter
estimate. This traditional setup thus requires reasonable knowledge of the parameters and structures to be identified. Since
the gray box situation starts from some physical insight, this
knowledge may be sufficient in some cases, but too demanding
in other. Resorting to testing random initial parameters may
not be feasible for large problems.
The other approach to structured state-space models is to
first estimate an unstructured, black-box model using, e.g.,
subspace identification methods, followed by the recovery of
the physical parameters embedded in the concerned structured
model. Using the classical subspace identification methods,
such as MOESP and N4SID [2], [3], the system matrices in
the first step can be consistently estimated under some mild
conditions. The parameter recovery in the second step turns
out to be a small-scale bilinear optimization problem.
To solve the bilinear optimization problem involved with the
gray-box identification, an alternating minimization algorithm
was developed in [6] and a null-space based method was proposed in [7]. In order to prevent the singular similarity transformation, a non-smooth optimization approach was presented
in [8]. Furthermore, in order to avoid estimating the similarity
transformation, an H∞ -norm-based identification algorithm
was proposed in [9]. The above mentioned algorithms are
sensitive to initial conditions. To cope with this problem, the
bilinear optimization problem was reformulated into a sumof-squares of polynomials which is then solved by the semidefinite programming method [10]; however, this technique
is limited to solving small-scale problems having only a few
unknown variables.
In this paper, a difference-of-convex programming (DCP)
based method is developed for the identification of structured
state-space models. This approach estimates the system parameters by the structured factorization of a block Hankel
matrix of system impulse response, which is inspired by the
Ho-Kalman decomposition method [3]. More explicitly, the
proposed method boils down to solving a low-rank structured
matrix factorization problem. In this paper, this non-convex
optimization problem is transformed into a difference-ofconvex programming (DCP) problem which is then tackled
by the sequential convex programming technique [11].
The advantages of the proposed method against many
recently developed methods are as follows. Different from the
2
identification method in [8], the proposed algorithm framework avoids the non-singularity constraint on the similarity
transformation and can be applied to the realization of nonidentifiable gray-box models. Unlike the model-matching H∞
method [9] which requires to solve an infinite-dimensional
optimization problem, the proposed identification method is
finite-dimensional so that it is more computational amenable.
Moreover, compared with other gradient-based or alternating
minimization methods [3], [6], the proposed identification
method performs well in practice thanks to the high-quality
initial conditions obtained by solving the convexified low-rank
and structured matrix factorization problem.
The rest of the paper is organized as follows. Section II
formulates the identification problem of gray-box models.
Section III reviews the gray-box identification using the classical prediction-error method. Section IV gives an alternative
way for the gray box identification, which is to identify the
black-box model first, following the identification of system
parameters by solving a bilinear optimization problem. Section
V provides a new method for solving the bilinear optimization
problem. Section VI demonstrates the performance of the
proposed identification method, followed by some conclusions
in Section VII.
II. P ROBLEM
STATEMENT
We consider a parameterized state-space model as follows
ẋ(t) = A(θ)x(t) + B(θ)u(t)
y(kT ) = C(θ)x(kT ) + w(kT ),
(1)
where u(t) ∈ Rm , x(t) ∈ Rn , y(t) ∈ Rp and w(k) ∈ Rp
are system input, state, output, and measurement noise,
respectively; θ ∈ Rl is the parameter vector; t and k represent
continuous and discrete time indices, respectively; T is the
sampling period.
The parameter vector θ in (1) typically represents unknown
values of physical coefficients. Here, we assume that the
structured system matrices are affine with respect to θ, i.e.
A(θ) = A0 +
B(θ) = B0 +
C(θ) = C0 +
l
X
i=1
l
X
i=1
l
X
Ai θi
H(q, θ) = C(θ) (qI − AT (θ))
−1
BT (θ),
(3)
where
AT = e
A(θ)T
,
BT =
Z
T
eA(θ)τ B(θ)dτ.
τ =0
Given the sampled IO data {u(kT ), y(kT )} for
k = 0, 1, · · · that are generated from model (1) for a
certain value θ∗ , the concerned gray-box identification
problem is to estimate the parameter vector θ∗ from the
measured IO data.
To address the concerned identification problem, the following assumptions are made throughout the paper.
A1 The system in (1) is minimal;
A2 The magnitudes of the imaginary parts of the eigenvalues
of A(θ) are less than the Nyquist frequency Tπ .
A3 The input sequence u(kT ) is persistently exciting of any
finite order [2, Chapter 13];
A4 The measurement noise w(t) is uncorrelated with the
system input u(t).
Assumption A2 ensures that the corresponding discrete-time
model of (1) is minimal when Assumption A1 is satisfied
[12]. Assumptions A3-A4 are standard assumptions for the
consistent identification of the discrete-time system model
H(q, θ).
III. G RAY- BOX APPROACH
The estimation of the parameter vector θ∗ using the sampled IO data {u(kT ), y(kT )} is typically a gray-box identification problem. The traditional identification method is
the prediction-error method [2] in which the predicted or
simulated outputs ŷ(kT |θ) are computed using the discretetime model H(q, θ) for any θ. The corresponding prediction
error criterion is written as
N −1
1 X
ky(kT ) − ŷ(kT |θ)k2
min
θ
N
k=0
s.t. x̂(kT + T ) = AT (θ)x̂(kT ) + BT (θ)u(kT )
ŷ(kT |θ) = C(θ)x̂(kT ) + w(kT ) for k = 0, 1, · · · , N − 1.
(4)
Bi θi
Ci θi ,
i=1
where the coefficient matrices Ai , Bi and Ci are known.
Besides the structures of the system matrices, the system
order of (1) is known as a priori knowledge as well.
Denote the corresponding true continuous-time transfer
function by:
G(s, θ) = C(θ) (sI − A(θ))−1 B(θ).
period T . Denoting the discrete-time system, obtained by the
sampling period T with the system input u(t) being piecewise
constant between the sampling instants kT , as
(2)
Although state-evolution equation in (1) is continuous, we
can only obtain sampled IO data in practice with sampling
This general method has the best possible asymptotic accuracy,
but the main drawback is that the optimization problem is
(highly) non-convex and may have many local minima. The
gradient-based optimization algorithms such as Gauss-Newton
method [2, Section 10.2], and gradient projection method [3,
Chapter 7] can be used to solve (4). However, the performance
mainly relies on the selection of initial parameter estimate. The
gray-box structure information may be sufficient to provide
such initial estimates that are in the domain of attraction of
the global minimum but otherwise one may have to resort to
random initial parameters. It is shown in [10] that the chances
to reach the global minimum of (4) from random starting
points may be very slim for problems of realistic sizes.
3
IV. B LACK - BOX + A LGEBRAIC A PPROACH
Besides the gray-box approach, there exist other routes to
estimate the parameter vector θ∗ from the sampled IO data.
Even though the gray-box approach may end up in local
minima, it is still possible to find the true system from data
by a black-box approach. Subspace approaches like N4SID
and MOESP [2], [3] can, under mild conditions, obtain the
true discrete-time system H(q, θ∗ ) as the length of the IO
data tends to infinity. That discrete-time system can be easily
transformed to continuous-time using the zero-order hold
interpolation approach [13]. As a result, the continuous-time
transfer function G(s, θ∗ ) will be known, but in an unknown
state-space basis:
G(s, θ∗ ) = C ∗ (sI − A∗ )−1 B ∗ .
(5)
The identification problem has now been transformed to an
algebraic problem:
Given the values of A∗ , B ∗ , C ∗ , determine the parameter
vector θ satisfying
C ∗ (sI − A∗ )−1 B ∗ = C(θ)(sI − A(θ))−1 B(θ).
(6)
The estimate of θ obtained in this way can then be used as
initial estimate in the minimization of (4). This approach was
discussed in [6], [8], [10].
V. S OLVING
THE
A LGEBRAIC
A. Using Similarity Transformation
Equation (6) means that there exists a similarity transformation Q such that
QB ∗ = B(θ),
C ∗ = C(θ)Q.
(7)
V (Q, θ) = kQA∗ − A(θ)Qk2F + kQB ∗ − B(θ)k2F
+ kC ∗ − C(θ)Qk2F
(8)
The optimization problem in (8) is a bilinear estimation
problem and an alternating minimization method was
proposed in [6]. In [10], the optimization problem in (8)
was minimized by a convex sum-of-squares method in case
A(θ), B(θ), C(θ) are affine in θ; however, this method is
limited to solving small-scale problems having only rather
few unknown variables.
M(θ)
can be equivalently written in a
Another way to deal with the possible mismatch between
equations (6) and (7) is to minimize the model-matching
criterion kG(s, θ) − G(s, θ∗ )k using either H2 norm or H∞
norm, as suggested in [10]. The H∞ -norm based modelmatching method has been investigated in [9]. Compared with
the minimization of (8), the H∞ -norm based model-matching
method reduces the number of unknown variables but at the
price of the introduction of a semi-infinite and non-smooth
program [9].
B. Using the Hankel Matrix of Impulse Response
In this section, aiming at dealing with the possible
drawback of minimizing equation (8), a new identification
approach is developed by exploiting the structured and
low-rank factorization of the block Hankel matrix of impulse
response.
After obtaining a full-parameterized state-space realization
G(s, θ∗ ) = C ∗ (sI − A∗ )−1 B ∗ , we can obtain the associated
impulse response sequence denoted by
Mi (θ∗ ) = C(θ∗ )Ai (θ∗ )B(θ∗ ) = C ∗ (A∗ )i B ∗
From that we can form the criterion
The equation group in (7)
vector form:
(A∗ )T ⊗ I − I ⊗ A(θ)
(B ∗ )T ⊗ I
I ⊗ C(θ)
{z
|
Even if we can find a global optimal solution (Q⋆ , θ⋆ )
by one of the above mentioned methods, it might not be
meaningful for the identification purpose stated in Section II.
The reason for this is that the optimal solution Q⋆ might be
singular and the obtained transfer function C ⋆ (sI − A⋆ )−1 B ⋆
might not be equal to C ∗ (sI − A∗ )−1 B ∗ . In fact, equations
(6) and (7) are equivalent if and only if Q is nonsingular [15].
To deal with this problem, a condition-number constraint on
Q was considered in [8], which turns out to be a non-smooth
and highly non-convex optimization problem.
PROBLEM
To solve the algebraic problem in (6), two routes are
provided here: one is the similarity transformation of the statespace realization and the other is the low-rank and structured
factorization of the block Hankel matrix of impulse response.
QA∗ = A(θ)Q,
To solve the above bilinear equation, a gradient projection
method was given in [3, Chapter 7.5.4], a null-space-based
optimization method was developed in [7] and a differenceof-convex based method was proposed in [14].
0
vec(Q) = vec(B(θ)) .
vec(C ∗ )
}
{z
}
|
N (θ)
(9)
for i = 0, 1, · · · . Let Hv,h (θ∗ ) be
constructed by Markov parameters
M0 (θ∗ )
M1 (θ∗ )
M1 (θ∗ )
M2 (θ∗ )
Hv,h (θ∗ ) =
..
..
.
.
Mv−1 (θ∗ )
Mv (θ∗ )
a block Hankel matrix
···
···
.
..
Mh−1 (θ∗ )
Mh (θ∗ )
..
.
···
Mv+h−2 (θ∗ )
,
(10)
where the subscripts v, h, satisfying v, h ≥ n, denote the number of block rows and number of block columns, respectively.
Given the block Hankel matrix Hv,h (θ∗ ), the concerned graybox identification problem is formulated as
min kHv,h (θ∗ ) − Hv,h (θ)k2F
θ
C(θ)B(θ)
..
s.t. Hv,h (θ) =
.
C(θ)Av−1 (θ)B(θ)
···
.
..
···
C(θ)Ah−1 (θ)B(θ)
..
.
C(θ)Av+h−2 (θ)B(θ)
(11)
.
4
In the above equation, the block Hankel matrix Hv,h (θ) has
a low-rank factorization as
C(θ)
C(θ)A(θ)
Hv,h (θ) =
..
.
C(θ)Av−1 (θ)
{z
}
|
Ov (θ)
×
B(θ)
A(θ)B(θ)
|
···
{z
Ch (θ)
Ah−1 (θ)B(θ) ,
}
(12)
where Ov (θ) and Ch (θ) denote the extended observability
and controllability matrix, respectively.
Denote Y = Hv,h (θ∗ ). By exploiting the shift properties
embedded in extended observability and controllability matrices, the optimization problem (11) can be recasted into a
low-rank structured matrix factorization problem:
min
θ,Ov ,Ch ,X
kY −
Xk2F
s.t. X = Ov Ch
Ov (1 : p, :) = C(θ)
Ch (:, 1 : m) = B(θ)
(13)
Ov (1 : (v − 1)p, :) A(θ) = Ov (p + 1 : vp, :)
A(θ)Ch (:, 1 : (h − 1)m) = Ch (:, m + 1 : hm) .
Step 1: The first constraint, X = Ov Ch , in (13) can be
equivalently written as a rank constraint.
Lemma 1: [16] The bilinear equation X = Ov Ch is
equivalent to the rank constraint
X Ov
= n.
(14)
rank
Ch In
The equivalent rank constraints for the last two constraints
of (13) are derived below. To simplify the notation, we denote
Ōv = Ov (1 : (v − 1)p, :), Ov = Ov (p + 1 : vp, :) and C¯h =
Ch (:, 1 : (h − 1)m), C h = Ch (:, m + 1 : hm). The last two
constraints can be represented as
q
X
i=1
A0 C¯h +
q
X
i=1
Ōv θi Ai = Ov
Ai C¯h θi = C h .
Ōv A0 +
A0 C¯h +
q
X
i=1
q
X
Γi Ai = Ov
Ai Υi = C h
(15)
An equivalent form of the combination of the fourth and fifth
constraints is given in the following lemma.
(16)
i=1
1
θ1
···
rank vec(Ōv ) vec(Γ1 ) · · ·
vec(C¯h ) vec(Υ1 ) · · ·
θq
vec(Γq ) = 1.
vec(Υq )
Proof: To prove the lemma, it is sufficient to prove that
the variables in (15) and those in (16) are one-to-one mapping.
On the one hand, the variables in (16) can be uniquely
determined from those in (15) by assigning Γi = Ōv θi and
Υi = C¯h θi . On the other hand, the variables in (15) can be
uniquely determined by the SVD decomposition of the matrix
on the left-hand side of the following equation
#
# "
"
1
vec(Ōv )
vec(C¯h )
θ1
vec(Γ1 )
vec(Υ1 )
···
···
···
θq
vec(Γq )
vec(Υq )
=
1
vec(Ōv )
vec(C¯h )
[ 1 θT ] .
By Lemmas 1 and 2, the bilinear optimization problem
in (13) can be equivalently formulated as a rank-constrained
optimization problem as follows:
min
X
s.t. rank
Ch
θ,Ov ,Ch ,X,Γ,Υ
In the above optimization problem, the first and the last
two constraints in the above equation are bilinear. To solve
this problem, the DCP-based identification framework [14]
will be adopted, which contains the following three steps:
(i) the bilinear optimization problem is transformed into
a rank constrained optimization problem; (ii) the rank
constrained problem is recasted into a DCP problem; (iii)
the DCP problem is then solved using the sequential convex
programming technique.
Ōv A0 +
Lemma 2: The constraint equation (15) is equivalent to
kY − Xk2F
Ov
=n
In
Ov (1 : p, :) = C(θ)
Ch (:, 1 : m) = B(θ)
q
X
Γi Ai = Ov
Ōv A0 +
(17)
i=1
Ōv = Ov (1 : (v − 1)p, :), O v = Ov (p + 1 : vp, :)
q
X
¯
Ai Υi = C h
A0 Ch +
i=1
C¯h = Ch (:, 1 : (h − 1)m), C h = Ch (:, m + 1 : hm)
1
θ1
···
θq
rank vec(Ōv ) vec(Γ1 ) · · · vec(Γq ) = 1.
vec(C¯h ) vec(Υ1 ) · · · vec(Υq )
The above optimization contains two rank constraints. To
deal with the above rank constrained optimization, we shall
further formulate it as a difference of convex optimization
problem.
Step 2: For notational simplicity, we denote
1
θ1
···
θq
T = vec(Ōv ) vec(Γ1 ) · · · vec(Γq ) .
vec(C¯h ) vec(Υ1 ) · · · vec(Υq )
Let σi (T ) be the i-th largest singular value of T for i =
1, · · · , q + 1. Define
fκ (T ) =
κ
X
i=1
σi (T ) .
5
It is remarked that fκ (·) is a Ky Fan κ-norm, which is a convex
function [17].
Inspired by the truncated nuclear norm method in [14], [18],
the rank constraint rank (T ) = 1 can be replaced by
fq+1 (T ) − f1 (T ) = 0.
(18)
The above equation means that all the singular values except
the largest one of T are zero. Since fq+1 (T ) = kT k∗, the
above equation can be represented as
kT k∗ − f1 (T ) = 0.
Using the above strategy, instead of directly solving the rank
constrained optimization problem in (17), we try to solve the
following regularized optimization problem:
min
θ,Ov ,Ch ,X,Γ,Υ
Based on the linearizations in (22) and (24), the convex
optimization problem to be solved at the (j + 1)-th iteration
is as follows:
kY − Xk2F + λ1 kΓk∗ − tr U1j,T ΓV1j
θ,Ov ,Ch ,X,Γ,Υ
j
+ λ2 kT k∗ − tr Lj,T
+ ρ kΓ − Γj k2F + kT − T j k2F
1 T R1
X Ov
s.t. Γ =
Ch In
+ λ2 (kT k∗ − f1 (T ))
X Ov
s.t. Γ =
Ch In
i=1
where Lj1 and R1j are respectively the left and right singular
vectors corresponding to the largest singular value. Then, the
linearization of f1 (T ) at the point T = T j is
j
j
T
−
T
R1 .
(24)
f1 (T ) ≈ f1 (T j ) + tr Lj,T
1
min
kY − Xk2F + λ1 (kΓk∗ − fn (Γ))
Ov (1 : p, :) = C(θ)
Ch (:, 1 : m) = B(θ)
q
X
Ōv A0 +
Γi A i = O v
Let T j be the estimate of T at the j-th iteration and its
SVD decomposition be given as
j,T
j
Σj1
R1
j
j
T = L1 L2
,
(23)
j
Σ2
R2j,T
(19)
Ōv = Ov (1 : (v − 1)p, :), O v = Ov (p + 1 : vp, :)
q
X
A0 C¯h +
A i Υi = C h
i=1
C¯h = Ch (:, 1 : (h − 1)m), C h = Ch (:, m + 1 : hm)
1
θ1
···
θq
T = vec(Ōv ) vec(Γ1 ) · · · vec(Γq ) ,
vec(C̄h ) vec(Υ1 ) · · · vec(Υq )
where λ1 , λ2 are regularization parameters. It is remarked
that all the constraints in (19) are linear functions with
respect to the unknown variables and the objective function
is a difference-of-convex function. Although the formulations
(17) and (19) may not be strictly equivalent, they have the
same global optimum.
Step 3: We shall develop a sequential convex programming
method to solve the DC optimization problem in (19). In
order to develop a sequential convex programming method,
it is essential to linearize the concave terms in the objective
function of (19). Let Γj be the estimate of Γ at the j-th
iteration and its SVD decomposition be given as
j,T
S1j
j
V1
j
j
,
(20)
Γ = U1 U2
S2j
V2j,T
where U1j and V1j are respectively the left and right singular
vectors corresponding to the largest n singular values. It can
be established that [17]
(21)
U1j V1j,T ∈ ∂fn Γj .
Then, the linearization of fn (Γ) at the point Γ = Γj is
(22)
fn (Γ) ≈ fn (Γj ) + tr U1j,T Γ − Γj V1j .
Ov (1 : p, :) = C(θ)
Ch (:, 1 : m) = B(θ)
q
X
Ōv A0 +
Γi A i = O v
i=1
Ōv = Ov (1 : (v − 1)p, :), Ov = Ov (p + 1 : vp, :)
q
X
A0 C¯h +
A i Υi = C h
i=1
C¯h = Ch (:, 1 : (h − 1)m), C h = Ch (:, m + 1 : hm)
1
θ1
···
θq
T = vec(Ōv ) vec(Γ1 ) · · · vec(Γq ) ,
vec(C̄h ) vec(Υ1 ) · · · vec(Υq )
(25)
where ρ is a very small positive regularization parameter
and
the proximal term ρ kΓ − Γj k2F + kT − T j k2F is added to
ensure the convergence of the sequential convex programming
approach, as shown in Lemma 3.
To ease the reference, the above sequential convex programming procedure is summarized in Algorithm 1.
Algorithm 1 Sequential convex programming method for (19)
1) Set Γ0 = 0 and T 0 = 0.
2) Repeat
2-1): Compute respectively the left and right singular vectors of Γj
and T j as shown in (20) and (23).
2-2): Obtain the estimates Γj+1 and T j+1 by solving (25).
3) until
kθ j+1 −θ j k2
kθ j k2
≤ ε with ε a small value.
By applying the iterative optimization method in Algorithm
1, the convergence is guaranteed.
Lemma 3: For the difference-of-convex optimization problem in (19), by implementing Algorithm 1, the obtained
estimate θj satisfies that limj→∞ θj+1 − θj = 0. Any
accumulation point of {θj } is a stationary point of (19).
The above lemma can be proven by the results of Theorems
1-2 in [19].
6
Since the difference-of-convex optimization problem in (19)
is still non-convex, the performance of the provided sequential
convex programming procedure relies on the initial conditions.
However, by setting T 0 = 0 and Γ0 = 0, we can find that
the optimization problem in (19) is a nuclear-norm relaxation
of the rank-constrained optimization problem in (17). Due
to the fact that the nuclear norm is the convex envelope
of the low-rank constraint on the unit spectral norm ball
[20], the associated nuclear-norm optimization usually yields a
good candidate for the starting point of the sequential convex
programming procedure.
100
90
80
Success rate (%)
70
60
50
40
30
PEM
Algorithm 1
DCP
20
10
0
1
2
3
In this section, two simulation examples will be carried
out to demonstrate the performance of the proposed method
- Algorithm 1. For comparison purposes, the prediction-error
method (PEM) [2], [10] and the difference-of-convex programming (DCP) method [14] are simulated. The implementation
details of these three methods are given below.
1) Algorithm 1 is simulated by empirically setting the regularization parameters in (25) to λ1 = 10−4 , λ2 = 10−5
and ρ = 10−10 . The tolerance of relative error is set to
ǫ = 10−4 .
2) PEM is simulated by firstly configuring the structure
object using the Matlab command idgrey, see [21]
and then implementing the identification method using
the Matlab command pem. The initial conditions are
randomly generated following the standard Gaussian
distribution.
3) DCP is simulated by setting the regularization parameter
λ in equation (17) of [14] to λ = 10−4 . The tolerance
of relative error is set to 10−4 .
In the simulations, the maximum number of iterations for the
these three methods is set to 100.
A. Randomly generated structures
The first simulation example is conducted following the way
in [10]. The state-space model of (1) is randomly generated
by the Matlab command rss, and the system parameters to
be estimated are randomly picked from the generated models.
Since the system model is randomly generated, it is difficult
to find a unified sampling period for all systems. Therefore,
when simulating DCP and Algorithm 1, the system matrices
A∗ , B ∗ and C ∗ in (6) are assumed to be known.
To ensure the identifiability of the system parameters, the
number of unknown parameters cannot be larger than (p +
m)n; however, system parameters less than (p + m)n may not
always be identifiable [2], [3]. Therefore, we use the impulseresponse fitting to measure the identification performance. In
the simulation, we choose the system order n = 5 and the
input/output dimension m = p = 1. For each fixed number
of free parameters, we carry out 100 Monte-Carlo trials by
randomly generating the system model and randomly picking
a fixed number of free parameters. The success rate is obtained
by counting the number of successful trials using the criterion
IRFr ≤ 10−3 . Denote by θr the estimate of θ at the r-th
4
5
6
7
8
9
10
Number of free parameters
VI. N UMERICAL S IMULATIONS
Fig. 1. Example 1: identification performance of PEM, DCP and Algorithm
1.
Monte-Carlo trial. The impulse-response fitting (IRF) of the
r-th Monte-Carlo trial is defined as
Pv+h−2
kC(θr )Ai (θr )B(θr ) − C ∗ (A∗ )i B ∗ kF
,
IRFr = i=0
Pv+h−2
kC ∗ (A∗ )i B ∗ kF
i=0
(26)
where the dimension parameters v and h are defined in (10).
The identification performance of PEM, DCP and Algorithm
1 is shown in Fig. 1 from which we can draw the following
conclusions.
1) When the number of parameters is larger than 3, DCP
and Algorithm 1 perform much better than PEM. This
is because DCP and Algorithm 1 can find good initializations by nuclear-norm regularized optimization. In
contrast, when the number of parameters are less than
or equal to 3, PEM has a slightly better performance
than DCP and Algorithm 1. This might be relevant to
the selection of the regularization parameters of DCP
and Algorithm 1.
2) When the number of parameters is larger than 6, the
success rate of Algorithm 1 is higher than that of DCP
up to 20%. This might be caused by the fact that DCP
does not consider the non-singularity constraint of the
similarity transformation, while Algorithm 1 implicitly
guarantees the non-singularity of the similarity transformation. However, when the number of parameters is less
than or equal to 6, DCP and Algorithm 1 have similar
performance. This might be because, when the number
of free parameters becomes smaller, singular similarity
transformations are less likely to occur.
B. Compartmental structures
The second simulation example is to identify the compartmental structure having the following form [4]
−θ
1
θ1
A(θ) =
B=
0
0
θ2
−(θ2 + θ3 )
..
.
···
0
1
T
..
.
..
.
..
.
,
..
.
−(θ2n−4 + θ2n−3 )
θ2n−3
0 0 ···
C =
θ2n−2
−θ2n−2
0 1 .
(27)
7
100
90
80
PEM
Algorithm 1
DCP
Success rate (%)
70
60
50
40
30
20
10
0
3
3.5
4
4.5
5
5.5
6
6.5
7
Number of free parameters
Fig. 2. Example 2: identification performance of PEM, DCP and Algorithm
1.
For each fixed system order, 100 Monte-Carlo trials are
carried out by randomly generating the system parameters. The
success rate is obtained by counting the number of successful
trials using the criterion IRFr ≤ 10−3 .
Fig. 2 shows the identification performance of three investigated methods in terms of IRF. It can be found that the success
rates of Algorithm 1 stay around 90% for different system orders, which demonstrates the better performance of Algorithm
1 with relation to DCP and PEM. This can be explained by
that Algorithm 1 can obtain good initializations and guarantee
the non-singularity of the similarity transformation.
VII. C ONCLUSIONS
In this paper, we have proposed a new gray-box identification method by exploiting the low-rank and structured
factorization of the Hankel matrix of impulse response. This
method uses the system impulse-response fitting as the objective function while avoiding the explicit non-singularity
constraint on similarity transformation; thus, it can be applied
to the state-space realizations of non-identifiable gray-box
models. Compared with the classical prediction-error method
initialized at random parameter values, the proposed method
can yield better performance since it can find a good initialization by nuclear-norm based optimization.
Although the proposed identification algorithm demonstrates good performance in terms of system impulse-response
fitting, its computational complexity is higher than the classical
prediction-error method. Thus, investigation will be made in
our future work on improving the computational efficiency of
the proposed identification method.
R EFERENCES
[1] R. C. Dorf and R. H. Bishop, Modern control systems. Pearson, 2011.
[2] L. Ljung, System Identification: Theory for the User. Pearson Education,
1999.
[3] M. Verhaegen and V. Verdult, Filtering and System Identification: A
Least Squares Approach. Cambridge University Press, 2007.
[4] R. Bellman and K. J. Åström, “On structural identifiability,” Mathematical biosciences, vol. 7, no. 3, pp. 329–339, 1970.
[5] J. Van den Hof, “Structural identifiability of linear compartmental
systems,” Automatic Control, IEEE Transactions on, vol. 43, no. 6,
pp. 800–818, 1998.
[6] L.-L. Xie and L. Ljung, “Estimate physical parameters by black box
modeling,” in Proceedings of the 21st Chinese Control Conference,
pp. 673–677, 2002.
[7] O. Prot, G. Mercère, and J. Ramos, “A null-space-based technique for
the estimation of linear-time invariant structured state-space representations,” in Proceedings of the IFAC Symposium on System Identification,
2012.
[8] G. Mercere, O. Prot, and J. Ramos, “Identification of parameterized
gray-box state-space systems: From a black-box linear time-invariant
representation to a structured one,” Automatic Control, IEEE Transactions on, vol. 59, pp. 2873–2885, Nov 2014.
[9] D. Vizer, G. Mercre, O. Prot, and E. Laroche, “h∞ -norm-based
optimization for the identification of gray-box lti state-space model
parameters,” Systems & Control Letters, vol. 92, pp. 34 – 41, 2016.
[10] L. Ljung and P. Parrilo, “Initialization of physical parameter estimates,”
2003.
[11] S. Boyd, “Sequential convex programming,” Lecture Notes, Stanford
University, 2008.
[12] T. Chen and B. Francis, Optimal Sampled-Data Control Systems. Communications and Control Engineering, Springer London, 1994.
[13] G. F. Franklin, J. D. Powell, and M. L. Workman, Digital control of
dynamic systems, vol. 3. Addison-wesley Menlo Park, 1998.
[14] C. Yu, M. Verhaegen, S. Kovalsky, and R. Basri, “Identification of structured lti mimo state-space models,” arXiv preprint arXiv:1509.08692,
2015.
[15] T. Kailath, Linear systems, vol. 1. Prentice-Hall Englewood Cliffs, NJ,
1980.
[16] R. Doelman and M. Verhaegen, “Sequential convex relaxation for convex
optimization with bilinear matrix equalities,” in 2016 European Control
Conference, 2016.
[17] L. Qi and R. S. Womersley, “On extreme singular values of matrix valued
functions,” Journal of Convex Analysis, vol. 3, pp. 153–166, 1996.
[18] Y. Hu, D. Zhang, J. Ye, X. Li, and X. He, “Fast and accurate matrix
completion via truncated nuclear norm regularization,” Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 35, no. 9,
pp. 2117–2130, 2013.
[19] C. Lu, J. Tang, S. Yan, and Z. Lin, “Nonconvex nonsmooth low rank
minimization via iteratively reweighted nuclear norm,” IEEE Transactions on Image Processing, vol. 25, no. 2, pp. 829–839, 2016.
[20] B. Recht, M. Fazel, and P. A. Parrilo, “Guaranteed minimum-rank
solutions of linear matrix equations via nuclear norm minimization,”
SIAM review, vol. 52, no. 3, pp. 471–501, 2010.
[21] L. Ljung, The System Identification Toolbox: The Manual. Natick, MA,
USA: The MathWorks Inc. 1st edition 1986, Edition 8.3 2013, 2013.
| 3cs.SY
|
On the Weighted Quartet Consensus problem∗
Manuel Lafond1 and Celine Scornavacca2
1
Department of Mathematics and Statistics, University of Ottawa, 585 King
Edward Ave., K1N 6N5 Ottawa, Canada
[email protected]
Institut des Sciences de l’Evolution – Université Montpellier, CNRS, IRD,
EPHE, Place Eugène Bataillon, 34095 Montpellier, France
[email protected]
2
arXiv:1610.00505v2 [cs.DS] 10 May 2017
Abstract
In phylogenetics, the consensus problem consists in summarizing a set of phylogenetic trees
that all classify the same set of species into a single tree. Several definitions of consensus exist
in the literature; in this paper we focus on the Weighted Quartet Consensus problem, a
problem with unknown complexity status so far. Here we prove that the Weighted Quartet
Consensus problem is NP-hard and we give a 1/2-factor approximation for this problem. During
the process, we propose a derandomization procedure of a previously known randomized 1/3factor approximation. We also investigate the fixed-parameter tractability of this problem.
1998 ACM Subject Classification F.2.2 G.2.1 G.2.2
Keywords and phrases Phylogenetic tree, consensus tree, quartets, complexity, fixed-parameter
tractability, approximability.
Digital Object Identifier 10.4230/LIPIcs.CPM.2017.23
1
Introduction
Phylogenetics is the branch of biology that studies evolutionary relationships among different
species. These relationships are represented via phylogenetic trees – acyclic connected graphs
with leaves labeled by species – which are reconstructed from molecular and morphological
data [12]. One fundamental problem in phylogenetics is to summarize the information
contained in a set of unrooted trees T classifying the same set of species into a single tree T .
The set T can consist of optimal trees for a single data set, of trees issued from a bootstrap
analysis of a unique data set, or even of trees issued from the analysis of different data sets.
Several consensus methods have been proposed in the past, the probably most known are
the strict consensus [23, 18] and the majority-rule consensus [17, 3]. For a survey, see [7].
In this paper we focus on the Weighted Quartet Consensus (WQC) problem [19],
also called the Median Tree with Respect to Quartet Distance problem [2] and
Quartet Consensus problem in [16]. Roughly speaking, this problem consists in finding a
consensus tree maximizing the weights of the 4-leaf trees – quartets – it displays, where the
weight of a quartet is defined as its frequency in the set of input trees (for a more formal
definition, see next section).
More general versions of this problem, where the input trees are allowed to have missing
species or where the weight of a quartet is not defined w.r.t. a set of trees, are known to be
∗
This work was partially supported by the Natural Sciences and Engineering Research Council of Canada
(NSERC) and the Mitacs Globalink Campus France program.
© Manuel Lafond and Celine Scornavacca;
licensed under Creative Commons License CC-BY
28th Annual Symposium on Combinatorial Pattern Matching (CPM 2017).
Editors: Juha Kärkkäinen, Jakub Radoszewski, and Wojciech Rytter; Article No. 23; pp. 23:1–23:18
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
23:2
On the Weighted Quartet Consensus problem
NP-hard [24] (and in fact, even Max-SNP-Hard), but the complexity of the WQC problem
has remained open so far. This problem has been conjectured to be NP-hard [2, 19], and
heuristics have recently been implemented and widely used, for instance ASTRAL [20], which
is a practical implementation of Bryant and Steel’s work from[8] (in fact, we show that the
ASTRAL algorithm is a 2-approximation for the minimization version of WQC). So far, the
best known approximation algorithm for the WQC problem consists in choosing a random
tree as a solution [16]. This tree is expected to contain a third of the quartets from the input
trees, and so it is a randomized factor 1/3 approximation. In [2], the minimization version
of the problem is studied, where the objective is to find a median tree T minimizing the sum
of quartet distances between T and the input trees (the quartet distance between two trees
T1 and T2 is defined as the number of quartets in T1 that are not in T2 ). A 2-approximation
algorithm is given, based on the fact that the quartet distance is a metric [9, 2].
A related problem that has received more attention is the Complete Maximum Quartet
Compatibility problem (CMQC) (see [5, 4, 16, 14, 25, 26, 10, 21, 22]). In the CMQC problem,
we are given, for each set S of four species, exactly one quartet on S, and the objective
is to find a tree containing a maximum number of quartets from the input. This can be
seen as a version of WQC in which each set of four species has one quartet of weight 1,
and the others have weight 0. The CMQC and WQC are however fundamentally different.
Although one could apply an algorithm for CMQC to WQC (by keeping only the most
frequent quartet on each set of four taxa), maximizing the most-frequent quartets may
enforce the presence of many low-frequency quartets. A better solution may prefer more of
the middle-frequency quartets, and we give an example of this phenomenon. It was shown in
[16] that the CMQC problem admits a polynomial time approximation scheme (PTAS), but
it can only be extended to WQC intances on a constant number of trees. Also, CMQC was
shown in [14, 10] to be fixed-parameter tractable w.r.t. the parameter “number of quartets
to reject from the input”.
The main result of this paper is to establish the NP-hardness of the WQC problem. In
Section 2, we introduce preliminary notions, and in Section 3 we propose a reduction from
the NP-hard Cyclic Ordering problem to WQC. It can be shown that this hardness result
transfers to the rooted setting, in which case we want to optimize triplets (3-leaf rooted trees)
rather than quartets. In Section 4, we discuss how being in a consensus setting, i.e. having
weights based on a set of input trees on the same leaf set rather than arbitrary weights,
does not necessarily make the problem easier, as one could expect: We list some structural
properties that, surprisingly, are not satisfied by optimal solutions of a WQC instance.
Nevertheless, in Section 5 we devise a factor 1/2 approximation algorithm for WQC running
in time O(k 2 n2 + kn4 + n5 ), where k is the number of trees and n the number of species (the
best known randomized algorithm in the non-consensus setting is a factor 1/3 one). As a
matter of fact, our algorithm includes a derandomization of this procedure, which had never
been done before. Finally in Section 6, we show that the known FPT algorithms for the
CMQC problem can be extended to the consensus setting. This yields an FPT algorithm that
is efficient on instances in which there is not too much ambiguity, i.e. when few competing
quartets on the same 4 species appear with the same frequency. We then conclude with some
remarks and open problems related to the quartet consensus problem.
2
Preliminaries
An unrooted phylogenetic tree T consists of vertices connected by edges, in which every pair
of nodes is connected by exactly one path and no vertex is of degree two. The leaves of a tree
M. Lafond and C. Scornavacca
23:3
T , denoted by L(T ) are the set of vertices of degree one. Each leaf is associated to a label;
the set of labels associated to the leaves of a tree T is denoted by L(T ). We suppose that
there is a bijection between L(T ) and L(T ). Due to this bijectivity, we will refer to leaves
and labels interchangeably. We denote |L(T )| as |T |. In the following, we will often omit
the word “phylogenetic” and, unless otherwise stated, all trees are leaf-labeled. A binary
unrooted tree has only vertices with degree three and vertices with degree one. A rooted
(binary) phylogenetic tree is defined in the same way, except that it has exactly one node of
degree two called the root, denoted by r(T ). Note that sometimes in the literature, rooted
trees are seen as directed and such that all arcs are oriented away from the root. Unless
stated otherwise, all trees in this paper are unrooted.
Given an unrooted phylogenetic tree T and a subset Y ⊆ L(T ), we denote by T |Y the tree
obtained from the minimal subgraph of T connecting Y when contracting degree-2 vertices.
A quadset is a set of four labels. For a quadset {a, b, c, d}, there are three non-isomorphic1
unrooted binary trees, called quartets, which are denoted respectively by ab|cd, ac|bd, and
ad|bc, depending on how the central edge splits the four labels. We say that an unrooted
tree T displays the quartet ab|cd if T |{a, b, c, d} is ab|cd. We denote the set of quartets
that
)|
an unrooted tree T displays by Q(T ). Note that if T is binary, then |Q(T )| = |L(T
. A set
4
of quartets Q on a set L is said to be complete if for each quadset {a, b, c, d} ⊆ L, there is in
Q exactly one quartet among ab|cd, ac|bd, and bc|ad.
We are now ready to state our optimization problem. The Weighted Quartet Consensus problem asks for a tree that has as many quartets as possible in common with a
given set of trees on the same set of labels X :
Weighted Quartet Consensus (WQC) problem
Input: a set of unrooted trees T = {T1 , . . . , Tk } such that L(T1 ) = . . . = L(Tk ) = X .
P
Output: a binary unrooted tree M with L(M ) = X that maximizes T ∈T |Q(M ) ∩ Q(T )|.
The problem is weighted as each quartet on X is weighted by frequency in T , see below.
In this paper we will focus on the particular case where the input trees are all binary. In
fact, proving the problem NP-hard for this restricted case implies NP-hardness of the general
problem. Note however that relaxing the requirement of the output M to be binary leads to
a different problem, as one needs to define how unresolved quartets in M are weighted.
In the remainder of the article, we will sometimes consider multi-sets of quartets, that
are sets in which the same quartet can appear multiple times. Denote by fQ (q) the number
of times that a quartet q appears in a multi-set Q (we may write f (q) if Q is unambiguous).
We say that a tree T contains k quartets of Q if there are distinct quartets q1 , . . . , qp ∈ Q(T )
Pp
such that i=1 f (qi ) = k. The Weighted Quartet Consensus problem can be rephrased
as follows: given trees T1 , . . . , Tk , finding a tree M that contains a maximum number of
quartets from Q(T1 ) ] Q(T2 ) ] . . . ] Q(Tk ), where ] denotes multiset union. We will implicitly
work with the decision version of WQC, i.e. for a given integer q, is there a consensus tree
M containing at least q quartets from Q(T1 ) ] Q(T2 ) ] . . . ] Q(Tk )?
Given a quadset {a, b, c, d}, we say that ab|cd is dominant (w.r.t. f ) if f (ab|cd) ≥ f (ac|bd)
and f (ab|cd) ≥ f (ad|bc). We say that ab|cd is strictly dominant if both inequalities are strict.
1
Isomorphism preserving labels.
CPM 2017
23:4
On the Weighted Quartet Consensus problem
3
NP-hardness of the Weighted Quartet Consensus problem
In this section, we present a reduction from the Cyclic ordering problem. This problem
has been used in phylogenetics before in [15] in the context of inferring rooted binary trees
from rooted triplets that are not required to originate from a set of trees on the same leaf set.
But beforehand, we need some additional notation. A caterpillar is an unrooted binary
tree obtained by taking a path P = p1 p2 . . . pr , then adding a leaf `i adjacent to pi for each
1 ≤ i ≤ r, then adding another leaf `01 adjacent to p1 and a leaf `0r adjacent to pr . The two
leaves `01 and `0r inserted last are called the ends of the caterpillar. An augmented caterpillar
T is a binary tree obtained by taking a caterpillar, then replacing each leaf by a binary
rooted tree (replacing a leaf ` by a tree T 0 means replacing ` by r(T 0 )). If T1 , T2 are the two
trees replacing the ends of the caterpillar, then T is called a (T1 , T2 )-augmented caterpillar.
Note that every binary tree is a (T1 , T2 )-augmented caterpillar for some T1 , T2 . Let T be
a caterpillar with leaves (`1 , `2 , . . . , `k ) taken in the order in which we encounter them on
the path between the two ends l1 and lk (more precisely, traverse the `1 − `k path, and
each time an internal node is encountered, append its adjacent leaves to the sequence), and
let T1 , . . . , Tk be rooted binary trees. We denote by (T1 |T2 | . . . |Tk ) the (T1 , Tk )-augmented
caterpillar obtained by replacing `i by r(Ti ), 1 ≤ i ≤ k. This notation gives rise to a natural
ordering of its subtrees, and we say that Ti < Tj if i < j (i.e. Ti appears before Tj in the
given ordering of the caterpillar subtrees). Note that by reversing such an ordering, we
obtain the same unrooted tree. However, in the proofs the given ordering will be important.
Also, since T1 , T2 , and Tk−1 , Tk are interchangeable, we will simply say that these two pairs
are incomparable. If each Ti consists of a single leaf `i for 2 ≤ i ≤ k − 1, then we may write
(T1 |`2 | . . . |`k−1 |Tk ), and `i < `j in T to indicate that `i appears before `j in the ordering.
We are now ready to describe the Cyclic Ordering problem. Let L = (s1 , . . . , sn )
be a linear ordering of a set S of at least 3 elements, and let (a, b, c) be an ordered
triple, with a, b, c ∈ S. We say that L satisfies (a, b, c) if one of the following holds in
L: a < b < c, b < c < a or c < a < b. If C is a set of ordered triples we say that L satisfies C
if it satisfies every element of C. Intuitively speaking, L satisfies (a, b, c) when, by turning L
into a directed cyclic order by attaching sn to s1 , one can go from a to b, then to c and then
to a. This leads to the following problem definition:
Cyclic Ordering problem
Input: A set S of n elements and a set C of m ordered triples (a, b, c) of distinct members
of S.
Question: Does there exist a linear ordering L = (s1 , . . . , sn ) of S that satisfies C?
The Cyclic Ordering problem is NP-hard [13]. In the rest of this section, we let S and
C be the input set and triples, respectively, of a Cyclic Ordering problem instance. We
denote n = |S| and m = |C|. We shall use the following simple yet useful characterization.
I Lemma 1. A linear ordering L of S satisfies C if and only if for each (a, b, c) ∈ C, exactly
two of the following relations hold in L: a < b, b < c, c < a.
Proof. (⇒): let L be a linear ordering satisfying C, and let (a, b, c) ∈ C. Thus in L, one of
a < b < c, b < c < a or c < a < b holds. It is straightforward to verify that in each case,
exactly two of a < b, b < c, c < a hold.
(⇐): suppose that L does not satisfy C. Then there is some (a, b, c) ∈ C such that one of
a < c < b, b < a < c or c < b < a does not hold. Again, one can easily verify that each of
these cases satisfies only one of a < b, b < c and c < a.
J
M. Lafond and C. Scornavacca
23:5
Now, from S and C we construct a set of unrooted binary trees T on the same set of
labels (we will omit the straightforward verification that this construction can be carried out
in polynomial time). First let W and Z be two rooted binary trees each on (nm)100 leaves
(the topology is arbitrary, and the 100 exponent could be optimized). Our trees are defined
on the leaf set X = S ∪ L(W ) ∪ L(Z) (note that S, L(W ), L(Z) are disjoint). Let C ∈ C with
C = (a, b, c). We construct 6 trees from C, that is, 3 pairs of trees:
The “a < b” trees: let (s1 , . . . , sn−2 ) be an arbitrary ordering of S \ {a, b}. Then we build
←
−
the trees TC (ab) = (W |a|b|s1 |s2 | . . . |sn−2 |Z) and T C (ab) = (W |sn−2 |sn−3 | . . . |s1 |a|b|Z).
The “b < c” trees: let (ŝ1 , . . . , ŝn−2 ) be an arbitrary ordering of S \ {b, c}. Then we build
←
−
the trees TC (bc) = (W |b|c|ŝ1 |ŝ2 | . . . |ŝn−2 |Z) and T C (bc) = (W |ŝn−2 |ŝn−3 | . . . |ŝ1 |b|c|Z).
The “c < a” trees: let (s̄1 , . . . , s̄n−2 ) be an arbitrary ordering of S \ {c, a}. Then we build
←
−
the trees TC (ca) = (W |c|a|s̄1 |s̄2 | . . . |s̄n−2 |Z) and T C (ca) = (W |s̄n−2 |s̄n−3 | . . . |s̄1 |c|a|Z).
Denote by T (C) the set of 6 constructed trees for C ∈ C. In this section, the input for our
S
Weighted Quartet Consensus instance constructed from S and C is T = C∈C T (C).
Note that |T | = 6m.
Observe that each tree of T (C) is a (W, Z)-augmented caterpillar. Moreover, note that
the majority of ordered pairs are “balanced” in the pairs of constructed trees: Let a, b ∈ S
←
−
and x, y ∈ S \ {a, b}, and let {TC (ab), T C (ab)} be an “a < b” tree-pair. Then we have x < y
←
−
in TC (ab) if and only if y < x in T C (ab). Similarly for any x ∈ S \ {a, b}, a < x, b < x in
←
−
TC (ab) but x < a, x < b in T C (ab). Only a < b holds in both trees.
Let T ∈ T , and let B(T ) denote the set of quartets of T that have at least two members
of L(W ), or at least two members of L(Z). Thus B(T ) consists in all the quartets of the
form w1 w2 |xy and z1 z2 |xy of T , where w1 , w2 ∈ L(W ), z1 , z2 ∈ L(Z) and x, y ∈ X (note
that no quartet of B(T ) has the form w1 x|yw2 for x, y ∈
/ L(W ), nor the form z1 x|yz2 for
x, y ∈
/ L(Z)). Note that for any tree T 0 ∈ T , B(T ) = B(T 0 ). Let K := 6m|B(T )| be the total
U
number of such quartets in T , i.e. K is the size of T ∈T B(T ). We observe the following:
U
I Remark. Any (W, Z)-augmented caterpillar on X contains the K quartets T ∈T B(T ).
Now, denote Ô := 3m|W ||Z| n−2
+ 2(n − 2) . Let T ∈ T and suppose that T is an “a < b”
2
tree, for some a, b ∈ S. For w ∈ L(W ) and z ∈ L(Z), x, y ∈ S, a quartet wx|yz displayed by
T is called an out-quartet of T if {x, y} =
6 {a, b}, and an in-quartet of T if x = a and y = b
(note that x = b and y = a is not possible, by construction). Let out(T ) and in(T ) denote
the set of out-quartets and in-quartets,
respectively,
of T . Note that each tree Thas |W ||Z|
n−2
in-quartets and |W ||Z|
+ 2(n − 2) out-quartets (because there are n−2
+ 2(n − 2)
2
2
ways to choose {x, y} =
6 {a, b}). Thus Ô is half the total number of out-quartets.
I Lemma 2. Any weighted quartet consensus tree M for T contains at most Ô quartets from
U
(W |s1 | . . . |sn |Z), where
T ∈T out(T ). Moreover, if M is a (W, Z)-augmented caterpillar
U
S = {s1 , . . . , sn }, then M contains exactly Ô quartets from T ∈T out(T ).
←
−
Proof. Let w ∈ L(W ) and z ∈ L(Z). Let {TC (ab), T C (ab)} be an “a < b” tree-pair of T ,
for some a, b ∈ S, and let x, y ∈ S such that {x, y} 6= {a, b}. Because x < y in TC (ab) if
←
−
and only if y < x in T C (ab), we get that the out-quartet wx|yz is in TC (ab) if and only if
←
−
wy|xz is in T C (ab). Since M can only contain one of the two quartets, it follows that M can
←
−
contain at most half of the quartets from out(TC (ab)) ] out( T C (ab)).
at
Thus M contains
U
most half the quartets from T ∈T out(T ), which is 3m|W ||Z| n−2
+
2(n
−
2)
=
Ô.
As
2
for the second assertion, if M = (W |s1 | . . . |sn |Z) then M contains one of wx|yz or wy|xz
for each x, y ∈ S. Thus if M does not contain the out-quartet wx|yz from TC (ab), then it
CPM 2017
23:6
On the Weighted Quartet Consensus problem
←
−
contains the out-quartet wy|xz from T C (ab). We deduce that M contains at least half the
U
←
−
quartets from out(TC (ab)) ] out( T C (ab)), and thus half the quartets from T ∈T out(T ). J
What follows is a key Lemma. The proof is not so straightforward and can be found in
Appendix B.1.
I Lemma 3. Any optimal consensus tree for T is a (W, Z)-augmented caterpillar.
We finally arrive at our main result.
I Theorem 4. The Weighted Quartet Consensus problem is NP-hard.
Proof. We show that there exists a linear ordering of S satisfying C if and only if there
exists a weighted quartet consensus tree M for T that contains at least K + Ô + 4m|W ||Z|
U
quartets from T ∈T Q(T ). For the rest of the proof, we let w ∈ L(W ) and z ∈ L(Z) be
arbitrary leaves of W and Z, respectively.
(⇒): let L = (s1 , s2 , . . . , sn ) be a linear ordering of S satisfying C. Then we claim that
the weighted quartet consensus tree M = (W |s1 |s2 | . . . |sn |Z) contains the desired number of
quartets. Since M is a (W, Z)-augmented caterpillar, M contains K quartets of T that have
two or more elements from L(W ), or two or more elements from L(Z), see remark on page 5.
U
Moreover by Lemma 2, M contains Ô quartets from T ∈T out(T ). As for the in-quartets, let
(a, b, c) ∈ C and let T ((a, b, c)) be the set of 6 trees corresponding to (a, b, c). By Lemma 1, L
satisfies two of the relations a < b, b < c, c < a . This implies that M has exactly two of the
following quartets: wa|bz, wb|cz, wc|az. Since, for every w ∈ L(W ) and z ∈ L(Z), each of
these three quartets appears as an in-quartet in exactly two trees of T ((a, b, c)) (e.g. wa|bz is
←
−
an in-quartet of T(a,b,c) (ab) and T (a,b,c) (ab)), it follows that M contains 4|W ||Z| quartets of
U
in(T ). As this holds for every (a, b, c) ∈ C, M contains 4m|W ||Z| quartets of
UT ∈T ((a,b,c))
T ∈T in(T ). Summing up, we get that M has at least K + Ô + 4m|W ||Z| quartets from T .
(⇐): suppose that no linear ordering of S satisfies C. Let M be an optimal consensus
tree for T . By Lemma 3, we may assume that M is a (W, Z)-augmented caterpillar. We
bound the number of quartets of T that can be contained in M .
First, by Lemma 3, M contains K quartets of T that have at least two elements of
L(W ) or at least two elements of L(Z). As for the quartets with one or zero elements from
L(W ) ∪ L(Z), in any tree T ∈ T there are at most (|W | + |Z|)n3 quartets of the form
wa|bc or za|bc with a, b, c ∈ S, and at most n4 quartets of the form ab|cd with a, b, c, d ∈ S.
Thus M contains at most 6m((|W | + |Z|)n3 + n4 ) < (|W | + |Z|)mn5 quartets of T that are
of the form wa|bc, za|bc or ab|cd with a, b, c ∈ S (the inequality holds because n ≥ 3 and
U
|W | = |Z| = (nm)100 ). Also, by Lemma 2, M contains at most Ô quartets from T ∈T out(T ).
It remains to count the in-quartets.
Let (a, b, c) ∈ C. The following in-quartets appear, each twice, in T ((a, b, c)): wa|bz,
wb|cz, wc|az. It is easy to check that these three quartets are incompatible, i.e. no tree can
contain all three of them, and hence M can have at most two of them. We deduce that
←
−
there must be at least two trees T, T of T ((a, b, c)) such that M contains no quartet from
U
←
−
in(T ) ] in( T ). Therefore M contains at most 4|W ||Z| quartets from T ∈T ((a,b,c)) in(T ),
U
and thus at most 4m|W ||Z| quartets from T ∈T in(T ) assuming that the 4|W ||Z| bound is
attained for every (a, b, c) ∈ C. We will however show that there must be some (a, b, c) ∈ C
U
such that M contains only 2|W ||Z| of the quartets in T ∈T ((a,b,c)) in(T ).
Now, since M is a (W, Z)-augmented caterpillar, we write M = (W |T1 |T2 | . . . |Tk |Z).
For some a ∈ S, let T (a) be the tree of {T1 , . . . , Tk } that contains a as a leaf. Then a
quartet wa|bz is in Q(M ) if and only if T (a) < T (b). Let L be a linear ordering of S such
M. Lafond and C. Scornavacca
23:7
that T (a) < T (b) ⇒ a < b in L. Since no linear ordering of S can satisfy C, by Lemma 1
there must be some (a, b, c) ∈ C such that only one of a < b, b < c, c < a holds in L. This
also means that at most one of T (a) < T (b), T (b) < T (c), T (c) < T (a) holds (because
¬(a < b) ⇒ ¬(T (a) < T (b))). Thus M has at most one of the wa|bz, wb|cz, wc|az quartets.
U
It follows that there are at least 2|W ||Z| quartets from T ∈T ((a,b,c)) in(T ) that M does not
U
contain. Therefore M contains at most 4m|W ||Z| − 2|W ||Z| quartets of T ∈T in(T ).
In total, the number of quartets that M contains from the input is bounded by K + Ô +
(|W | + |Z|)mn5 + (4m − 2)|W ||Z| < K + Ô + 4m|W ||Z|, by our choice of |W | and |Z|. J
The implications of these results for the Weighted Triplet Consensus (WTC) problem are
presented in Appendix A. The same techniques can be used to show that WTC is NP-hard.
4
The (non)-structure of WQC
In the rest of this paper, we aim at designing algorithms building on the fact that the weight
of each quartet is not arbitrary, and is rather based on a set of input trees on the same leaf
set. When designing optimized algorithms for a problem, understanding the relationship
between the input and the optimal solution(s) can be of great help. In phylogenetics, several
problems are harder in the supertree setting, i.e. when the input trees do not all contain
the same species, than in the consensus setting as in the WQC problem. An example is the
problem of finding an unrooted phylogenetic tree containing as minors a set of unrooted
phylogenetic trees – the compatibility problem – which is NP-hard in the supertree setting
[24] and polynomially solvable in the consensus setting [1]. Despite the NP-hardness of WQC,
there may exist some properties inherent to the consensus setting that are useful for devising
efficient FPT algorithm, or for establishing lower bounds on the value of an optimal solution
in order to develop approximation algorithms.
In attempt to establish useful properties of the weights of quartets in the consensus
setting, we initially conjectured that the following relationships between the input trees and
the solution(s) hold. Despite being seemingly reasonable, we prove all these conjectures false.
1. let D be the set of strictly dominant quartets of the input multiset Q, i.e. the quartets
ab|cd such that f (ab|cd) > f (ac|bd) and f (ab|cd) > f (ad|bc). Then there is a constant
α > 0 such that there exists an optimal solution containing at least α|D| quartets of D;
2. if a quartet ab|cd has a higher weight than the sum of the other quartets on the same
quadset, i.e. f (ab|cd) > f (ac|bd) + f (ad|bc), then some optimal solution contains ab|cd;
3. more generally, there exists β > 0 such that if a quartet ab|cd is in a fraction β of the
input trees, then ab|cd must be in some optimal solution. In particular, if ab|cd is in
every input tree, then there is some optimal solution that contains ab|cd;
4. if a quartet ab|cd is in no input tree, then no optimal solution contains ab|cd.
5. call ab|cd a strictly least-frequent quartet if f (ab|cd) < f (ac|bd) and f (ab|cd) < f (ad|bc).
Suppose that there exists a tree T ∗ on leaf set X that contains no strictly least-frequent
quartet, and choose such a T ∗ that contains a maximum number of quartets from the
input. Then T ∗ is an optimal solution for WQC.
Unfortunately, we answered negatively to all conjectures, see Appendix B.2.
5
Approximatibility of WQC
In this section, we show that WQC admits a factor 1/2 approximation algorithm that runs
in polynomial time. Hereafter, the input set of trees is T = {T1 , . . . , Tk } and we denote
CPM 2017
23:8
On the Weighted Quartet Consensus problem
Q = Q(T1 ) ] . . . ] Q(Tk ). We say that a minimization (resp. maximization) problem P can
be approximated within a factor α > 1 (resp. β < 1) if there is an algorithm that, for every
instance I of P , runs in polynomial time and outputs a solution of value AP P (I) such that
AP P (I) ≤ αOP T (I) (resp. AP P (I) ≥ βOP T (I)), where OP T (I) is the optimal value of I.
As mentioned before, the Complete Maximum Quartet Compatibility (CMQC) problem
admits a PTAS, though it can only be applied to the WQC problem when the number
of input trees is constant. There does not seem to exist an easy extension of the PTAS
algorithm for the case of an unbounded number of trees, which makes WQC seem “harder”
than CMQC. Nevertheless, we give a simple factor 1/2 approximation algorithm, which is
better than the (randomized) factor 1/3 approximation, the best known so far, for the general
Maximum Quartet Consistency problem in which the given quartet set is not necessarily
complete. We borrow ideas from [9] to show that this can be achieved by taking the best
solution from either a 1/3 approximation to WQC, or a factor 2 approximation to WMQI,
the minimization version of WQC (see below). For two unrooted binary trees T1 , T2 on leaf
set X , denote dQ (T1 , T2 ) = |Q(T1 ) \ Q(T2 )|. The WMQI problem is defined as follows:
Weighted Minimum Quartet Inconsistency (WMQI) problem
Input: a set of unrooted trees T = {T1 , . . . , Tk } such that L(T1 ) = . . . = L(Tk ) = X .
P
Output: a tree M with L(M ) = X that minimizes T ∈T dQ (M, T ).
Note that the WMQI problem is equivalent to finding a minimum (in the multiset sense)
number of quartets to discard from Q so that it is compatible.
It is not hard to show that dQ is a metric. In particular, dQ satisfies the triangle inequality,
i.e. for any 3 trees T1 , T2 , T3 on the same leaf set, dQ (T1 , T3 ) ≤ dQ (T1 , T2 ) + dQ (T2 , T3 ). This
leads to a factor 2 approximation algorithm for WMQI obtained by simply returning the best
tree from the input. Intuitively, the input tree that is the closest to the others cannot be too
far from the best solution, which is a median tree in the metric space. See [2] for details.
I Theorem 5 ([2]). The following is a factor 2 approximation algorithm for WMQI: output
P
the tree T ∈ T that minimizes Ti ∈T dQ (T, Ti ).
In [2], the authors explain how to compute dQ (T1 , T2 ) in time O(n2 ). Therefore the
factor 2 approximation can be implemented to run in time O(k 2 n2 ), by simply computing
dQ between every pair of trees.
Theorem 5 has a direct implication on the approximation guarantees of the ASTRAL
algorithm in [20], an implementation of the work from Bryant and Steel [8]. This algorithm
finds, in polynomial time, an optimal solution M for a restricted version of WMQI where
every bipartition of M is also a bipartition in at least one of the input trees. The solution T
returned by the algorithm of Theorem 5 above trivially satisfies this condition. Thus, M is
at least as good as T , implying the following.
I Corollary 6. The ASTRAL algorithm is a factor 2 approximation for WMQI.
We do not know whether the factor 2 is tight for the ASTRAL algorithm - we conjecture
that ASTRAL can actually achieve a better approximation ratio. As shown in the rest of
this section, this would have interesting applications for the approximability of WQC.
Indeed, both WQC and WMQI share the same set of optimal solutions, but the two
problems are not necessarily identical in terms of approximability. We show however that
WMQI can be used to approximate WQC. As stated earlier, there is a trivial factor 1/3
randomized approximation for WQC: output a random tree T . Each quartet of Q has a 1/3
M. Lafond and C. Scornavacca
23:9
chance of being contained
by T , and so the expected number of quartets of Q contained by
T is |Q|/3 = k n4 /3 (here |Q| denotes the multiset cardinality). Call this the “random-treealgorithm”. For the sake of having a deterministic algorithm, we show the following:
I Lemma 7. The “random-tree-algorithm” can be derandomized, i.e. there is a deterministic
algorithm that, in time O(kn4 + n5 ), finds a tree containing at least |Q|/3 quartets from Q.
Proof. We derandomize the factor 1/3 algorithm using the standard method of conditional
expectation. For the simplicity of exposition, we will construct a rooted tree T in a top-down
manner (T can be unrooted after the construction). Call a rooted tree T internally binary if
the only nodes of T that have more than two children have only leaves as children. We start
with a fully unresolved internally binary tree T on leaf set X (i.e. T consists of a root whose
n children are in bijection with X ). We then iteratively split each unresolved node v of T
into two subtrees so as to maximize the expected number of quartets that T contains. We
stop when T is a binary tree.
To describe the algorithm more precisely, suppose that T is an internally binary tree on
leaf set X , and let v be a node of T with more than 2 children, say {v1 , . . . , vm } ⊆ X (if no
such v exists, then T is binary and we can stop). We split v by first removing {v1 , . . . , vm }
from T , adding two children x and y to v, and reinserting v1 , . . . , vm one after another, each
as either a child of x or a child of y. We describe how this choice is made. Suppose that
for i ≥ 1, v1 , . . . , vi−1 have been reinserted, resulting in the tree Ti−1 , and that we need
to process vi . Denote by Ti,x (resp. Ti,y ) the tree obtained by inserting vi as a child of x
0
(resp. of y) in Ti−1 . We then define a random binary tree Ti,x
from Ti,x as follows: for each
0
0
v ∈ {vi+1 , . . . , vm }, reinsert v as a child of x with probability 1/2, or as a child of y with
probability 1/2. Then, replace each non-binary node w with children X 0 by a rooted binary
0
tree on leaf set X 0 chosen uniformly at random. We define the random binary tree Ti,y
from
Ti,y using the same process.
For a random tree T 0 obtained by the above process and for q ∈ Q, let I(q, T 0 ) be an
indicator variable for whether q ∈ Q(T 0 ). That is, I(q, T 0 ) = 1 if q ∈ Q(T 0 ), and I(q, T 0 ) = 0
P
otherwise. Let I(T 0 ) = q∈Q I(q, T 0 )fQ (q) 2 . We seek
max
0 ,T 0 }
T 0 ∈{Ti,x
i,y
E [I(T 0 )] =
=
max
E
max
X
0 ,T 0 }
T 0 ∈{Ti,x
i,y
0 ,T 0 }
T 0 ∈{Ti,x
i,y
X
I(q, T 0 )fQ (q)
q∈Q
Pr [q ∈ Q(T 0 )] fQ (q)
q∈Q
0
If Ti,x
attains this maximum, we insert vi below x, and otherwise we insert vi below
y. After every child vi of v has been inserted, we process the next non-binary node. This
concludes the algorithm description (we shall detail how to compute Pr[q ∈ Q(T 0 )] below).
If T is an internally binary tree, by a slight abuse of notation define E [I(T )] = E [I(T 0 )],
where T 0 is the random binary tree obtained by replacing each non-binary node of T on leaf
set X 0 by a random binary tree on leaf set X 0 .
I Claim 1. Let T be an internally binary tree, and suppose that E[I(T )] ≥ |Q|/3. Let v be
a non-binary node of T , and let Tv be the tree obtained after splitting v using the above
algorithm. Then E[I(Tv )] ≥ |Q|/3.
2
Observe that here, q ∈ Q means that there exists at least one occurrence of q in the multisetQ, and so
each quartet present in Q is considered once in the summation, independently of fQ (q).
CPM 2017
23:10
On the Weighted Quartet Consensus problem
Let {v1 , . . . , vm } be the children of v. To prove the claim, we use induction on the
number of processed children of v to show that after each insertion of a child vi , the obtained
0
0
tree Ti ∈ {Ti,x , Ti,y } satisfies E[I(Ti0 )] ≥ |Q|/3, where Ti0 ∈ {Ti,x
, Ti,y
} is the random tree
corresponding to Ti obtained from the above process (i.e. reinserting vi+1 , . . . , vm randomly
under x or y, and resolving non-binary nodes randomly). This proves the statement since
0
Tm = Tv (and thus E [I(Tv )] = E [I(Tm )] = E[I(Tm )] ≥ |Q|/3). As a base case, if i = 1 it
0
0
0
0
is easy to see that T1,x
and T1,y
are identical, and that E[I(T1,x
)] = E[I(T1,y
)] = E[I(T )] ≥
|Q|/3. For i > 1, let Ti−1 be the tree obtained after inserting vi−1 , and suppose without
0
loss of generality that Ti−1 = Ti−1,x . Because, in Ti−1,x
, we insert vi below x or y each with
1
probability 2 , we have
1
1
0
0
0
E I(Ti−1,x
) = E I(Ti−1,x
)|vi is a child of x + E I(Ti−1,x
)|vi is a child of y
2
2
1
0
0
E I(Ti,x ) + E I(Ti,y )
=
2
0
By induction,
we alsohave E[I(T
i−1,x )] ≥ |Q|/3. Combined with the above equality, we
0
0
0
0
obtain 12 E I(Ti,x
) + E I(Ti,y
) ≥ |Q|/3. This implies that one of E[I(Ti,x
)] or E[I(Ti,y
)]
must be at least |Q|/3.
J
Since the fully unresolved tree T from which we start satisfies E[I(T )] ≥ |Q|/3, Claim 1
shows that the algorithm does terminate with a tree containing at least |Q|/3 quartets from
Q. It remains to be show how to compute, when reinserting a node vi , the expectations for
0
0
Ti,x
and Ti,y
.
In fact, it suffices to be able to compute, for a given quartet q = ab|cd, the probability
0
0
0
0
Pr[q ∈ Q(T 0 )] for T 0 ∈ {Ti,x
, Ti,y
}. Moreover, if Pr[q ∈ Q(Ti,x
)] = Pr[q ∈ Q(Ti,y
)], then this
probability does not contribute to determining which scenario maximizes expectation, and in
this case we do not need to consider q. In particular, if none of a, b, c, d is equal to vi , then
0
0
Pr[q ∈ Q(Ti,x
)] = Pr[q ∈ Q(Ti,y
)]. Therefore, it is enough to consider only quartets in which
vi is included. We will assume that vi = a. Moreover, we may assume that two or three of
{b, c, d} are children of v in T (recall that v is the parent of vi in T ), because otherwise the
probability that ab|cd is in T 0 is unaffected by whether a is a child of x or a child of y.
There are still multiple cases depending on which of b, c and d are children of v, and which
have been reinserted or have not, but this probability can be easily found algorithmically.
Let U = {b, c, d} ∩ {vi+1 , . . . , vm }, i.e. the leaves in {b, c, d} that have not been reinserted
yet in T 0 . We obtain new trees S10 , . . . , Sh0 by reinserting, in T 0 , the members of U below x or
y in every possible way – there are only 2|U | ≤ 8 possibilities, so h ≤ 8. Then, for 1 ≤ j ≤ h
denote by Sj0 |q the tree Sj0 restricted to {a, b, c, d} (i.e. obtained by removing every leaf not
in {a, b, c, d}, then contracting degree 2 vertices). Note that Sj0 |q may be non-binary. We get
Ph
Pr[q ∈ Q(T 0 )] = j=1 h1 Pr[q ∈ Q(Sj0 |q )]. This is because every leaf in vi+1 , . . . , vm other
than b, c, d is resinserted independently from the choice for b, c, d, and every non-binary node
remaining after the reinsertions is resolved uniformly. The probability Pr[q ∈ Q(Sj0 |q )] is
straightforward to compute, as only a constant number of cases can occur since Sj0 |q has only
4 leaves. We omit the details.
Time complexity: we must first preprocess the input in order to compute fQ (q) for
each quartet q. This takes time O(kn4 ). As for the computation of Pr[q ∈ Q(T 0 )], assume
that the lowest common ancestor (lca) of two leaves can be found in constant time. This
can be achieved naively by simply storing the lca for each pair of leaves in a table of size
O(n2 ), and updating the table in time O(n) each time a decision on some vi is made (this
does not hinder the total time complexity of the algorithm, though there are more clever
M. Lafond and C. Scornavacca
23:11
ways to handle dynamic tree lca queries [11]). Then the restrictions S10 |q , . . . , Sh0 |q can be
computed in constant time. It is then straightforward to see that, by the above process,
Pr[q ∈ Q(T 0 )] can be computed in constant time. Each time a node vi needs to be reinserted,
this probability must be computed for the O(n3 ) quartets containing vi . There are n−1 splits
to be performed, and each split requires inserting O(n) nodes. Thus the “binarization” process
takes total time O(n5 ), and altogether the derandomization takes time O(kn4 + n5 ).
J
The above leads to a (deterministic) 1/3- approximation. This can be used to show the
following. The proof is similar to that of [9, Theorem 2] and is relegated to Appendix B.3.
I Theorem 8. If WMQI can be approximated within a factor α, then WQC can be approximated within a factor β = α/(3α − 2).
Combined with Theorem 5 and letting α = 2 in Theorem 8 we get the following.
I Corollary 9. WQC can be approximated within a factor 1/2 in time O(k 2 n2 + kn4 + n5 ).
6
Fixed-parameter tractability of WQC
In this section we describe how, based on previous results on the minimum quartet incompat0
0
0
ibility problem on complete sets, WQC can be solved in time O(4d +k2 +k3 n + n4 ). Here k20
and k30 are the number of quadsets that have 2 and 3 dominant quartets, respectively, and d0 is
the number of strictly dominant quartets that we are allowed to reject. The algorithm makes
direct use of the Gramm-Niedermeyer algorithm [14], henceforth called the GN algorithm.
The GN algorithm solves the following problem: given a complete set of quartets Q, find,
if it exists, a complete and compatible set of quartets Q0 such that at most d quartets of Q0
are not in the input set Q (i.e. |Q0 \ Q| ≤ d). This is accomplished by repeatedly applying
the following theorem:
I Theorem 10 ([14]). Let Q be a complete set of quartets. Then Q is compatible if and only
if for each set of five taxa {a, b, c, d, e} ⊆ X , ab|cd ∈ Q implies ab|ce ∈ Q or ae|cd ∈ Q.
The idea behind the GN algorithm is as follows: find a set of five taxa {a, b, c, d, e} that
does not satisfy the condition of Theorem 10, then correct the situation by branching into
the four possible choices:
1. remove ab|cd from Q and add ac|bd to Q;
2. remove ab|cd from Q and add ad|bc to Q;
3. remove {ac|be,ae|bc} ∩ Q from Q and add ab|ce to Q;
4. remove {ac|de,ad|ce} ∩ Q from Q and add ae|cd to Q.
The quartets added to Q will not be questioned in the following branchings. With some
optimization, this leads to a O(4d n + n4 ) FPT algorithm.
In [14], the authors also note that this algorithm can be extended to sets of quartets Q
that contain ambiguous quadsets, i.e. sets {a, b, c, d} for which 2 or 3 of the possible quartets
on {a, b, c, d} are in Q. Suppose there are k2 and k3 , respectively, quadsets that have 2 and 3
quartets in Q. The modified algorithm then, in a first phase, branches into the 2k2 3k3 ways
of choosing one quartet per such quadset, thereby obtaining a complete set of quartets for
each possibility. The GN algorithm is thus applied to the so-obtained complete sets. This
yields a O(2k2 · 3k3 · 4d n + n4 ) algorithm.
It is not hard to see that this gives an FPT algorithm for WQC, where the parameter k2
(resp. k3 ) is the number of quadsets such that 2 (resp. 3) possible quartets appear in the
input trees, and d is the number of quartets ab|cd that appear in every input tree, and that
CPM 2017
23:12
On the Weighted Quartet Consensus problem
we are allowed to discharge. Note however that, in the consensus setting, there is no reason
to believe that k2 and k3 are low - in we fact we believe that k2 + k3 typically takes values in
Θ(n4 ). One reason is that even the slightest amount of noise on a quadset makes it included
in the count of either k2 or k3 (e.g. if k − 1 trees agree on ab|cd and only one contains ac|bd).
The GN algorithm can, however, be used on a more interesting set of parameters. Define
k20 (resp. k30 ) as the number of quadsets that have exactly 2 (resp. 3) dominant quartets,
and let d0 be the number of strictly dominant quartets that we are allowed to discharge. It is
reasonable to believe that, if each tree of the input is close to the true tree T ∗ , most “true”
quartets will appear as strictly dominant in the input, and there should not be too many
0
0
0
ambiguous quadsets. There is a very simple algorithm achieving time O(4d +k2 +k3 n + n4 ).
Construct a complete set of quartets Q as follows: for each quadset {a, b, c, d}, choose a
dominant quartet on {a, b, c, d} and add it to Q (if multiple choices are possible, choose
arbitrarily). Then, run the GN algorithm on Q with the following modification: each time
a quartet q is removed from Q and replaced by another quartet q 0 , decrement either d0 , k20
or k30 , depending on whether q belongs to a quadset with 1, 2 or 3 dominant quartets. It
follows that if there exists a complete and compatible set of quartets Q0 such that at most d0
strictly dominant quartets are rejected, then the modified algorithm will find it. It should be
noted however that finding such a set Q0 does not guarantee that the corresponding tree is
an optimal solution. Indeed, since quartets are weighted, two solutions Q0 and Q00 may both
reject only d0 strictly dominant quartets, yet one has higher weight than the other. However,
the correctness of the algorithm follows from the fact that the GN algorithm finds the set
of every solution discarding at most d0 dominant quartets - and thus it suffices to pick the
solution from this set that has optimal weight.
We finally mention that the FPT algorithms published in [10] are improved versions of
the GN algorithm, can also return every solution and thus can be modified in the same
0
0
0
manner. These yield FPT algorithms that can solve WQC in time O(3.0446d +k2 +k3 n + n4 )
0
0
0
and O(2.0162d +k2 +k3 n3 + n5 ).
7
Conclusion
In this paper, we have shown that the WQC problem is NP-hard, answering a question of [19]
and [2]. In the latter, the authors also propose a variant of the problem in which the output
tree T is not required to be binary. In this case, one needs to assign a cost p to the unresolved
quartets. Our reduction can be extended to show that hardness holds for high enough p,
but the complexity of the general case remains open. We have also shown that WQC can
be approximated within a factor 1/2. One open question is whether the problem admits a
PTAS as the related CMQC problem. The fixed-parameter tractability aspects of WQC
also deserve further investigation. This would require identifying some structural properties
that are present in the consensus setting and that can be used for designing practical exact
algorithms. But as we have shown, this might not be an easy task, as many properties which
seem reasonable for the consensus setting do not hold.
References
1
2
Alfred V Aho, Yehoshua Sagiv, Thomas G Szymanski, and Jeffrey D Ullman. Inferring a
tree from lowest common ancestors with an application to the optimization of relational
expressions. SIAM Journal on Computing, 10(3):405–421, 1981.
Mukul S Bansal, Jianrong Dong, and David Fernández-Baca. Comparing and aggregating
partially resolved trees. Theoretical Computer Science, 412(48):6634–6652, 2011.
M. Lafond and C. Scornavacca
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23:13
Jean-Pierre Barthélemy and Fred R McMorris. The median procedure for n-trees. Journal
of Classification, 3(2):329–334, 1986.
Vincent Berry, David Bryant, Tao Jiang, Paul Kearney, Ming Li, Todd Wareham, and Haoyong Zhang. A practical algorithm for recovering the best supported edges of an evolutionary
tree (extended abstract). In Proceedings of the Eleventh Annual ACM-SIAM Symposium
on Discrete Algorithms, pages 287–296. Society for Industrial and Applied Mathematics,
2000.
Vincent Berry, Tao Jiang, Paul Kearney, Ming Li, and Todd Wareham. Quartet cleaning:
Improved algorithms and simulations. In Jaroslav Nešetřil, editor, Proceedings of the 7th
Annual European Symposium on Algorithms, pages 313–324. Springer Berlin Heidelberg,
1999.
David Bryant. Building trees, hunting for trees, and comparing trees. PhD thesis, University
of Canterbury, 1997.
David Bryant. A classification of consensus methods for phylogenetics. DIMACS Series in
Discrete Mathematics and Theoretical Computer Science, 61:163–184, 2003.
David Bryant and Mike Steel. Constructing optimal trees from quartets. Journal of Algorithms, 38(1):237–259, 2001.
Jaroslaw Byrka, Sylvain Guillemot, and Jesper Jansson. New results on optimizing rooted
triplets consistency. Discrete Applied Mathematics, 158(11):1136–1147, 2010.
Maw-Shang Chang, Chuang-Chieh Lin, and Peter Rossmanith. New fixed-parameter algorithms for the minimum quartet inconsistency problem. Theory of Computing Systems,
47(2):342–367, 2010.
Richard Cole and Ramesh Hariharan. Dynamic LCA queries on trees. In Proceedings of
the Tenth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 235–244. Society
for Industrial and Applied Mathematics, 1999.
Joseph Felsenstein. Inferring phylogenies. Sinauer Associates Sunderland, 2004.
Zvi Galil and Nimrod Megiddo. Cyclic ordering is NP-complete. Theoretical Computer
Science, 5(2):179–182, 1977.
Jens Gramm and Rolf Niedermeier. Minimum quartet inconsistency is fixed parameter
tractable. In Amihood Amir, editor, Proceedings of the 12th Annual Symposium of Combinatorial Pattern Matching, pages 241–256. Springer Berlin Heidelberg, 2001.
Jesper Jansson. On the complexity of inferring rooted evolutionary trees. Electronic Notes
in Discrete Mathematics, 7:50–53, 2001.
Tao Jiang, Paul Kearney, and Ming Li. A polynomial time approximation scheme for
inferring evolutionary trees from quartet topologies and its application. SIAM Journal on
Computing, 30(6):1942–1961, 2001.
Timothy Margush and Fred R McMorris. Consensusn-trees. Bulletin of Mathematical
Biology, 43(2):239–244, 1981.
Fred R McMorris, David B Meronk, and Dean Neumann. A view of some consensus methods
for trees. In Numerical Taxonomy, pages 122–126. Springer, 1983.
Siavash Mirarab. Novel scalable approaches for multiple sequence alignment and phylogenomic reconstruction. PhD thesis, University of Texas at Austin, 2015.
Siavash Mirarab, Rezwana Reaz, Md. Shamsuzzoha Bayzid, Théo Zimmermann, M Shel
Swenson, and Tandy Warnow. ASTRAL: genome-scale coalescent-based species tree estimation. Bioinformatics, 30(17):i541–i548, 2014.
António Morgado and Joao Marques-Silva. A pseudo-boolean solution to the maximum
quartet consistency problem. arXiv preprint arXiv:0805.0202, 2008.
António Morgado and Joao Marques-Silva. Combinatorial optimization solutions for the
maximum quartet consistency problem. Fundamenta Informaticae, 102(3-4):363–389, 2010.
CPM 2017
23:14
On the Weighted Quartet Consensus problem
23
24
25
26
A
Robert R Sokal and F James Rohlf. Taxonomic congruence in the leptopodomorpha reexamined. Systematic Zoology, 30(3):309–325, 1981.
Michael Steel. The complexity of reconstructing trees from qualitative characters and
subtrees. Journal of Classification, 9(1):91–116, 1992.
Gang Wu, Jia-Huai You, and Guohui Lin. A lookahead branch-and-bound algorithm for
the maximum quartet consistency problem. In International Workshop on Algorithms in
Bioinformatics, pages 65–76. Springer, 2005.
Gang Wu, Jia-Huai You, and Guohui Lin. Quartet-based phylogeny reconstruction with
answer set programming. IEEE/ACM Transactions on Computational Biology and Bioinformatics, 4(1):139–152, 2007.
Implications for the Weighted Triplet Consensus problem
For each set of three labels {a, b, c}, there are three non-isomorphic3 rooted binary trees
called triplets. They are denoted by ab|c, ac|b and bc|a, depending on the leaf having the
root as father (c, b and a respectively). We say that a tree T induces or displays the triplet
ab|c if T |{a, b, c} = ab|c. For a rooted tree R, denote by tr(R) the set of triplets of R.
When the consensus is sought for rooted trees, the objective is to find a rooted tree M
that induces a maximum number of triplets contained in the input trees. The Weighted
Triplet Consensus (WTC) is defined as follows.
Weighted Triplet Consensus (WTC) problem
Input: a set of rooted trees R = {R1 , . . . , Rk } such that L(R1 ) = . . . = L(Rk ) = X .
P
Output: a binary rooted tree M with L(M ) = X that maximizes R∈R |tr(M ) ∩ tr(R)|.
As in the unrooted problem, other versions of WTC where the input trees may have
missing species or where the weight of a triplet is not defined w.r.t. a set of trees, are known
to be NP-hard [6]. The WTC problem is conjectured to be NP-hard in [2] (we note that a
more general version where the output can be non-binary is also conjectured NP-hard).
S
We give the main idea behind the proof of the hardness of WTC. Let T = C∈C T (C)
be the set of unrooted trees constructed in the reduction above. For a tree T ∈ T , let e be
the edge separating Z from the rest of the tree (i.e. by removing e from T , one connected
component is exactly Z). Obtain a rooted tree R from T by rooting T at e, that is subdivide
e, thereby creating a degree 2 vertex which is the root of R. The set of rooted trees R is
obtained by applying this rooting to every T ∈ T (the Z subtree could be removed but we
keep it here to make the correspondence easier to see).
Similarly as above, it can be shown that since every input tree is a rooted (W, Z)caterpillar, then any solution must also have this form. This implies in turn that there exists
a linear ordering of S satisfying C if and only if there is a solution M to WTC containing
every triplet from the input on 2 or 3 members of L(W ),every triplet
containing at least
one member of L(Z), plus at least 4m|W | + 3m|W | n−2
+
2(n
−
2)
triplets of the form
2
wa|b with a, b ∈ S. This is obtained by defining the notions of in-triplets and out-triplets
analogously as in the previous section, but with respect to W only. That is, in a “a < b” tree,
for a, b, c, d ∈ S, w ∈ L(W ) and {c, d} 6= {a, b}, wa|b would be an in-triplet, whereas wc|d or
wd|c would be out-triplets. One can argue that for a cyclic triple (a, b, c) ∈ C and the set of
trees T ((a, b, c)), an optimal consensus tree can contain
4|W | of the 6|W | possible in-triplets,
plus at most half of the 6m|W | n−2
+
2(n
−
2)
possible
out-triplets. The arguments are
2
3
Isomorphism preserving labels and the root node.
M. Lafond and C. Scornavacca
23:15
essentially the same as the ones given in the hardness proof of WQC, and so we omit the
details.
I Theorem 11. The Weighted Triplet Consensus problem is NP-hard.
B
B.1
Deferred proofs
Proof of Lemma 3
Despite the Lemma 3 statement being quite intuitive, it requires a surprising amount of care.
We start by a simple proposition that will be needed.
I Proposition B.1. Let X, Y be two non-empty sets such that Y 6⊆ X. Then |X| · |Y \ X| ≥
|Y | − 1.
Proof. Suppose first that X ∩ Y = ∅. Then clearly |X||Y \ X| = |X||Y | ≥ |Y | − 1. Suppose
otherwise that X ∩Y 6= ∅, and denote X 0 = X ∩Y . Then |Y \X| = |Y |−|X 0 | and since Y 6⊆ X,
we must have |Y | ≥ |X 0 | + 1. We also have |X||Y \ X| = |X|(|Y | − |X 0 |) ≥ |X 0 |(|Y | − |X 0 |);
we claim the latter term to be at least |Y | − 1. Let us assume for contradiction that
|X 0 |(|Y | − |X 0 |) < |Y | − 1. If |X 0 | = 1, this is clearly impossible, so assume |X 0 | > 1.
0 2
| −1
0
Then we get |X 0 ||Y | − |Y | < |X 0 |2 − 1 leading to |Y | < |X
|X 0 |−1 = |X | + 1, contradicting
0
|Y | ≥ |X | + 1.
J
Before proceeding, we must introduce the notion of a rooted subtree of a binary unrooted
tree T . Note that by removing an edge e = {u, v} of T , we obtain two disjoint rooted subtrees
T1 and T2 , respectively rooted at u and v. Call T 0 a rooted subtree of T if T 0 is a rooted tree
that can be obtained by removing an edge of T . For X ⊂ L(T ), a rooted subtree for X is
a rooted subtree T 0 of T such that X ⊆ L(T 0 ). We denote by T [X] the rooted subtree for
X that contains a minimum number of leaves (if there are multiple choices, choose T [X]
arbitrarily among the possible choices). Note that T [X] may contain leaves other than X.
We now prove that any optimal solution to T as constructed in our reduction must be a
(W, Z)-augmented caterpillar. Suppose that M is an optimal solution for T , and that M is
not a (W, Z)-augmented caterpillar. Denote MW = M [L(W )] and MZ = M [L(Z)]. If M is a
(W 0 , Z 0 )-augmented caterpillar (W 0 |T1 | . . . |Tk |Z 0 ) for some trees W 0 , Z 0 with L(W 0 ) = L(W )
and L(Z 0 ) = L(Z), it is not hard to see that M 0 = (W |T1 | . . . |Tk |Z) is a better solution
than M , a contradiction. Thus, M is not such a caterpillar, and this implies that either
L(MW ) 6= L(W ) or L(MZ ) 6= L(Z) (or both). That is, the rooted subtrees containing L(W )
and/or L(Z) have “outsider” leaves. Suppose first that L(MW ) 6= L(W ) holds. Then there
exists a node x with children xl and xr in MW such that all leaves Xl below xl are in L(W )
with L(W ) 6⊆ Xl (otherwise MW = M [L(W )] would be chosen incorrectly), and no leaf Xr
below xr belongs to L(W ) (this can be seen by observing that the minimal node x of MW
having leaves both in W and not in W has this property).
We claim that L(Z) 6⊆ Xr . Suppose otherwise that L(Z) ⊆ Xr . Then |Xr | ≥ |Z| and
so |MW | ≥ |W | + |Z|. However in M , by removing the xxr edge we obtain two rooted
trees, one of which is a rooted subtree for L(W ). Moreover, this subtree has at most
|W | + |S| < |W | + |Z| leaves, which contradicts the minimality of MW = M [L(W )]. We
deduce that L(Z) is not a subset of Xr .
Now, observe that M contains the quartet w1 y|w2 z for each w1 ∈ Xl , y ∈ Xr , w2 ∈ L(W )\
Xl , z ∈ L(Z) \ Xr . There are at least |Xl ||Xr |(|L(W ) \ Xl |)(|L(Z) \ Xr |) ≥ (|W | − 1)(|Z| − 1)
such quartets (the inequality is obtained by applying Proposition B.1 to |Xl | · |L(W ) \ Xl |
and |Xr | · |L(Z) \ Z|). Moreover, each input tree of T contains the quartet w1 w2 |yz instead,
CPM 2017
23:16
On the Weighted Quartet Consensus problem
and hence in total in T there are at least 6m(|W | − 1)(|Z| − 1) quartets of the form w1 w2 |yz
that M does not contain. In the same manner, if the case L(MZ ) 6= L(Z) holds, then there
are at least 6m(|W | − 1)(|Z| − 1) quartets of the form z1 z2 |yw that M does not contain,
where here z1 , z2 ∈ L(Z), y ∈
/ L(Z), w ∈ L(W ).
U
Now, let ρ(M ) be the number of quartets that M contains from T ∈T Q(T ) that have
the form wx|yz, where w ∈ L(W ), z ∈ L(Z), x, y ∈ S. Formally,
X
ρ(M ) =
f (wx|yz)
wx|yz∈Q(M )
x,y∈S
w∈L(W )
z∈L(Z)
where f (wx|yz) denotes the number of trees of T that contain the wx|yz quartet. For a
given u ∈ L(W ) ∪ L(Z), let ρ(M, u) denote the number of quartets counted in ρ(M ) that
contain u. Formally, if w ∈ L(W ), we have
ρ(M, w) =
X
f (wx|yz)
wx|yz∈Q(M )
x,y∈S
z∈L(Z)
The definition of ρ(M, z) is the same for z ∈ L(Z), except that z gets fixed instead
P
P
of w in the summation. Notice that ρ(M ) = w∈L(W ) ρ(M, w) = z∈L(Z) ρ(M, z). Let
w∗ = arg maxw∈L(W ) {ρ(M, w)}. We obtain an alternative solution M 0 from M in the
following manner: remove all leaves of L(W ) \ {w∗ } from M , delete the degree 2 nodes, and
replace w∗ by the W tree. Note that if w∗ x|yz is a quartet of M , then wx|yz is a quartet
of M 0 for all w ∈ L(W ), and so ρ(M 0 , w) ≥ ρ(M, w) for all such w by the choice of w∗ .
Consequently, ρ(M 0 ) ≥ ρ(M ). We repeat the same operation on M 0 for the Z tree and
obtain our final tree M ∗ . That is, we find z ∗ = arg maxz∈L(Z) {ρ(M 0 , z)}, and replace z ∗ by
the Z tree. As above, we obtain ρ(M ∗ ) ≥ ρ(M 0 ). Since M ∗ has W and Z as rooted subtrees,
it follows that M ∗ is a (W, Z)-augmented caterpillar.
We argue that M ∗ contains more quartets from the input trees than M . First observe
that the quartets on which M and M ∗ differ must contain a member of L(W ) ∪ L(Z), since
U
only these leaves switched position. The tree M ∗ contains every quartet of T ∈T Q(T )
that have at least two members of L(W ), or two members of L(Z). This includes the
aforementioned (at least) 6m(|W | − 1)(|Z| − 1) quartets of the form w1 w2 |yz or z1 z2 |yw
that M does not contain. As for the quartets that contain one member of L(W ) and one
U
member of L(Z), M ∗ contains at least as many such quartets as M since in T ∈T Q(T ),
these quartets are all of the form wx|yz, and we have ρ(M ∗ ) ≥ ρ(M ). Finally, each tree of
T has at most (|W | + |Z|)n3 quartets that have exactly one member of L(W ) ∪ L(Z). Thus
at most 6m(|W | + |Z|)n3 quartets of this type are contained by M and not contained by
M ∗ , but since this is smaller than 6m(|W | − 1)(|Z| − 1) for our choice of |W | and |Z|, M ∗
contains more quartets from the input than M .
B.2
Proofs of Section 4
Conjecture 1 is disproved by Theorem 12, and Conjecture 3 by Theorem 13, which implies
that Conjectures 2 and 4 are also false; finally Conjecture 5 is disproved by Theorem 14.
I Theorem 12. There exists an instance of WQC such that every optimal solution contains
none of the strictly dominant quartets.
M. Lafond and C. Scornavacca
c
b
a
e
⇥2
c
a
d
e
⇥9
b
d
d
e
a
d
e
⇥3
a
b
c
e
b
a
c
b
b
c
e
c
⇥1
a
d
⇥4
d
e
d
e
⇥8
c
⇥2
b
e
⇥5
d
b
e
b
a
c
a
d
⇥2
a
e
23:17
c
a
d
d
b
⇥4
c
b
a
⇥4
c
Figure 1 An instance of WQC such that the optimal solution (the third tree on the first row)
contains no strictly dominant quartet. The numbers correspond to the number of times that each
tree appears in the input.
b
d
c
h c
f
g
g
e
d
a
f
a e
b
g h
h
d
g
a
h
c
c
d
c
f
e
b
f
a
b
e
f
b
a
e
g
h
d
Figure 2 The first four trees form an instance of WQC in which every tree contains ab|cd. The
rightmost tree is the unique optimal solution to the WQC instance (every possible solution was
verified computationally).
Figure 1 shows an instance of WQC demonstrating Theorem 12. In this instance, for
every quadset S, there is a strictly dominant quartet appearing 17 times, whereas the
second-most and third-most quartets appear in 16 and 11 trees, respectively. For example,
f (ac|bd) = 17, f (ad|bc) = 16 and f (ab|cd) = 11. One can check that the best tree is the third
one on the top row (the ae|bc with d grafted on the middle branch). Call this tree T ∗ . For
every quadset S, T ∗ contains the second-most frequent quartet on S. The reason why T ∗
is optimal is that, in the particular instance of Figure 1, any other tree T that contains a
strictly dominant quartet for some quadset S must also contain a least frequent quartet on
some other quadset S 0 . Hence, as there are 5 quadsets, T contains at most 4 · 17 + 11 = 79
quartets from the input, whereas T ∗ contains 5 · 16 = 80. Note that this example consists of
trees on only 5 leaves. We do not know if such instances exist for any n > 5 leaves.
I Theorem 13. There exists an instance of WQC such that there is a quartet q that appears
in every input tree, but q is not a quartet of any optimal solution.
Figure 2 shows an instance of WQC proving Theorem 13. Each input tree contains the
ab|cd quartet, whereas the optimal solution, which is unique, does not. The rightmost tree
contains 180 quartets from the input multiset Q, whereas any other tree has at most 176.
CPM 2017
23:18
On the Weighted Quartet Consensus problem
Finally, we note that the main interest behind Conjecture 5 is the following: if it holds,
in cases where the set F of strictly least-frequent quartets is complete we could tell in
polynomial time – using results of [8] – whether there is a tree T ∗ that contains no quartet
from F . Conjecture 5 could then lead to interesting approximations or FPT algorithms.
However, least-frequent quartets cannot be excluded automatically.
I Theorem 14. There exists an instance of WQC such that every optimal solution contains
a strictly least-frequent quartet, even if there exists a tree T ∗ with no such quartet.
The instance corresponding to Theorem 14 is obtained from the instance shown in
Figure 1, by removing all occurrences of the third tree on the top row (i.e. this tree now
appears 0 times instead of 3 times). The second-most frequent quartets now appear 13 times
each, and so the tree T ∗ that contains all these quartets has a total weight of 5 · 13 = 65.
However, there are trees with a total weight of 75, which are optimal (for instance, the tree
of cardinality 9 in the figure). Each such tree contains a strictly dominant quartet, and as
mentioned before, also a strictly least-frequent quartet.
B.3
Proof of Theorem 8
Let N := k n4 , i.e. the total number of quartets in Q, let p be the maximum number of
quartets that can be preserved from Q for compatibility, and let d be the minimum number
of quartets to discard from Q in order to attain compatibility (here p and d refer to multiset
cardinalities). Note that d = N − p. We show that taking the best tree between the one
obtained from the factor α algorithm for WMQI and the one obtained from the “randomtree-algorithm” achieves a factor β for WQC. Suppose first that p ≤ N/(3β). By Lemma 7,
the “random-tree-algorithm” yields a tree containing at least |Q|/3 = N/3 quartets from Q,
and since N/3 = βN/(3β) ≥ βp, it yields a solution to WQC within a factor β from optimal.
Thus we may assume that p > N/(3β) = N (3α − 2)/(3α). Since we have an α approximation
for WMQI, we may obtain a solution discarding at most αd = α(N − p) quartets. This
solution preserves at least N − (α(N − p)) = αp + (1 − α)N quartets from Q. We claim
that this attains a factor β approximation. Suppose instead that αp + (1 − α)N < βp. Then
p < (α − 1)N/(α − β) which, with a little work, yields p < N (3α − 2)/(3α), contradicting
our assumption on p. Thus, the WMQI approximation preserves at least βp quartets.
| 8cs.DS
|
1
Riding on the Primary: A New Spectrum
Sharing Paradigm for Wireless-Powered IoT
Devices
arXiv:1801.06679v1 [cs.IT] 20 Jan 2018
Xin Kang, Member, IEEE, Ying-Chang Liang, Fellow, IEEE, Jing
Yang, Member, IEEE
Abstract
In this paper, a new spectrum sharing model referred to as riding on the primary (ROP) is proposed
for wireless-powered IoT devices with ambient backscatter communication capabilities. The key idea
of ROP is that the secondary transmitter harvests energy from the primary signal, then modulates its
information bits to the primary signal, and reflects the modulated signal to the secondary receiver without
violating the primary system’s interference requirement. Compared with the conventional spectrum
sharing model, the secondary system in the proposed ROP not only utilizes the spectrum of the
primary system but also takes advantage of the primary signal to harvest energy and to carry its
information. In this paper, we investigate the performance of such a spectrum sharing system under
fading channels. To be specific, we maximize the ergodic capacity of the secondary system by jointly
optimizing the transmit power of the primary signal and the reflection coefficient of the secondary
ambient backscatter. Different (ideal/practical) energy consumption models, different (peak/average)
transmit power constraints, different types (fixed/dynamically adjustable) reflection coefficient, different
primary system’s interference requirements (rate/outage) are considered. Optimal power allocation and
reflection coefficient are obtained for each scenario.
Index Terms
X. Kang is with National Key Laboratory of Science and Technology on Communications, University of Electronic Science
and Technology of China, Chengdu, China 611731 (E-mail: [email protected]).
Y. -C. Liang is with University of Electronic Science and Technology of China, Chengdu, China 611731. He is also with
School of Electrical and Information Engineering, The University of Sydney, NSW 2006, Australia. (E-Mail: [email protected]).
J. Yang is with Singapore PowerGrid, Singapore Power Group, 2 Kallang Sector, Singapore 349277. (E-mail:
[email protected]).
Part of this paper has been presented in [1] at IEEE ICC 2017.
2
Cognitive radio, spectrum sharing, backscatter communication, IoT, wireless-powered network,
energy harvesting, ergodic capacity, power allocation, optimization.
I. I NTRODUCTION
Internet of Things (IoT) is a key application scenario of the fifth generation (5G) mobile communication systems. It covers a wide range of use cases, such as smart home, smart wearables,
smart farming, smart manufacturing, smart utilities, and smart city, which enable new business
opportunities and new operational considerations for 5G. With the diverse use cases anticipated
in IoT, the types of IoT devices are expected to diversified, and the characteristics and demands
of different IoT devices are expected to vary a lot. Some of the devices, such as sensors and
Radio-Frequency Identification (RFID) tags, are expected to be simple, small, low power, low
throughput field devices. For this kind of IoT devices, a key requirement from the industry [2] is
that the power consumption should be very low and the battery life should be as long as ten years
for extreme use cases. In these situations, energy harvesting, with potential to provide a perpetual
power supply, becomes an attractive approach to prolong these devices’ battery lifetime. Classic
sources for energy harvesting include solar and wind. Recently, ambient radio signal [3]-[6] is
receiving much research attention as a new viable source for energy harvesting, supported by
the advantage that the wireless signals can carry both energy and information.
The backscatter communications technology used in RFID systems is a real-world application
of energy harvesting from RF signals. In a typical backscatter system [7], [8], the reader
transmits a RF sinusoidal signal to a passive tag. The passive tag harvests RF energy from
the signal to power its circuit, modulates its information bits onto the received sinusoidal signal
by intentionally changing its amplitude and/or phase which is realized by changing its antenna
impedance, and reflects the modulated signal back to the reader. In [9], ambient backscatter,
which is able to harvest energy from and transmits information over the ambient RF signals (e.g.
TV signals), was proposed. In [10], Wi-Fi backscatter that uses the existing Wi-Fi infrastructure to
provide internet connectivity for RF-powered devices was proposed. In [11] and [12], maximumlikelihood detection was studied for an ambient backscatter system in which the tag adopts
differential modulation. In [13], the modulator and the decoder design for backscattering over
the ambient orthogonal frequency division multiplexing (OFDM) signals was studied. It was
shown in recent work [14] and [15] that power harvested from ambient RF signals is sufficient to
support the daily communication of a battery-less sensor through dedicated decoder and signal
3
design. In [16], it was further shown a network architecture with combined wireless power
transfer and backscattering communication technology is able to a large and dense IoT network.
However, these aforementioned works mainly focus on the hardware and decoder design but
lacks fundamental system analysis from theoretical aspects. Besides, the mutual influence (such
as the interference) between the backscatter system and the primary system are not considered.
In this paper, we introduce the ambient backscatter communication technology to the cognitive
radio (CR) system, and propose a new spectrum sharing model based on that. The proposed spectrum sharing model applies to CR systems with conventional primary communication systems
and ambient-backscatter-based secondary systems. The key idea is that the secondary transmitter
harvests energy from the primary signal, then modulates its information bits to the primary signal,
and reflects the modulated signal to the secondary receiver without violating the primary system’s
interference requirement [17]. The main contribution of this paper is summarized as follows:
•
We propose a new spectrum sharing model as Riding on the Primary (ROP). The differences
between the proposed ROP and the existing technologies are as follows: (i) Compared with
conventional backscatter communication systems, the reader (secondary receiver) in our
system does not need to generate and transmit a RF sinusoidal signal, which can reduce its
power consumption and prolongs its battery life. (ii) Compared with ambient backscatter
communication systems, the interference from the wireless-powered tag (secondary transmitter) to the primary system is taken into consideration when designing the system. (iii)
Compared with conventional spectrum sharing systems, the secondary system not only
utilizes the spectrum of the primary system but also takes advantage of the primary system’s
signal transmission to carry its information.
•
We investigate the performance of the proposed ROP system under fading channels. To be
specific, we maximize the ergodic capacity of the secondary system by jointly optimizing the
transmit power of the primary signal and the reflection coefficient of the secondary ambient
backscatter. Different (ideal/practical) energy consumption models, different (peak/average)
transmit power constraints, different types (fixed/dynamically adjustable) reflection coefficient, different primary system’s interference requirements (rate/outage) are considered.
Optimal power allocation and optimal reflection coefficient are obtained for each case. It
is worth pointing out that certain degree of cooperation is needed between the primary and
secondary system in order to do the joint optimization. For example, the primary system
needs the knowledge of the secondary system (e.g., channel information) to optimize the
4
PT
PR
ST
SR
Fig. 1. System Model
transmit power, whereas the secondary system needs the knowledge of the primary system
to optimize the reflection coefficient.
•
We show by numerical examples that dynamically adjustable reflection coefficient will result
in a better system performance than a fixed reflection coefficient. The system performance
under the average transmit power constraint is in general better than that under the peak
transmit power constraint. Applying the primary transmission outage probability constraint
to protect the primary system usually leads to a higher secondary ergodic capacity than
adopting the primary transmission rate constraint. All these findings can serve as the
guidance for designing high-performance practical ROP system.
The rest of this paper is organized as follows. Section II introduces the system model of
the proposed ROP spectrum sharing system. Section III presents the basic problem formulation.
Section IV investigates the capacity maximization problem under the practical energy consumption model. Section V studies the capacity maximization problem for fixed reflection coefficient
and average transmit power constraint. Section VI considers the capacity maximization problem
under the primary transmission outage constraint. Then, numerical results are given in Section
VII to verify the proposed studies. Finally, Section VIII concludes the paper.
II. S YSTEM M ODEL
A. Channel Model
In this paper, we consider a spectrum sharing communication system consists of a primary
communication pair and a secondary transmission pair. The primary communication pair is a
5
conventional communication system consisting of a RF source (e.g. Base Stations, TV towers,
WiFi APs) and a receiver (e.g., cell phones, TV receivers). The secondary communication pair
is an ambient backscatter communication system which consists of a wireless-powered passive
tag and a battery-powered reader. For ease of explanation, we denote the RF source and the
receiver of the primary system as the primary transmitter (PT) and the primary receiver (PR),
respectively. The wireless-powered tag and the reader of the backscatter system are denoted as
secondary transmitter (ST) and secondary receiver (SR), respectively. In this paper, we consider
the block fading channel model [17], where the channel coefficients remain the same for each
block but may change from one block to another. As shown in Fig. 1, the channel power gains
for the fading block n, from the PT to the PR, from the ST to the SR are denoted by h1 (n) and
g1 (n), respectively. The channel power gains for cross channels for fading block n, i.e., from
the PT to the ST, from the PT to the SR, and from the ST to the PR, are denoted by f (n),
h2 (n) and g2 (n), respectively.
B. Transmission Model
Transmitted signal at the PT. Let s(n; k) denote the transmitted signal of the PT at kth
symbol of the nth block where |s(n; k)|2 = 1, and p(n) denote the transmit power for fading
block n. Then, the transmitted signal of the PT for the kth symbol of block n is given by
xP T (n; k) =
p
p(n)s(n; k), ∀k.
(1)
Transmitted signal at the ST. In fading block n, the signal received at the ST from the PT
p
is f (n)xP T (n; k). Note that the noise at the ST (Tag) is neglected as [9], [10], [11], [12],
[14], [18] since the on-tag integrated circuit only includes passive components. The power of
the received signal at the ST from the PT is f (n)p(n). Part energy of the received signal is
absorbed by the ST to power its circuit operation. The remaining part of the received signal is
modified and backscattered to the reader. For convenience, we refer to this splitting factor as
the reflection coefficient, and denote it by α(n) where 0 ≤ α(n) ≤ 1. Then, the energy of the
transmitted signal of the ST can be denoted as α(n)f (n)p(n). Let c(n; k) where |c(n; k)|2 = 1
denote the ST’s own signal, then the transmitted signal of the ST for the kth symbol of block
n is given by
xST (n; k) =
p
p
p
α(n) f (n) p(n)s(n; k)c(n; k), ∀k.
(2)
6
Note that we assume there is no signal processing delay of the backscatter circuit, i.e., there is
no time delay between the transmitted signal and the received signal of the ST. This assumption
is widely used in backscatter communication research literatures [9]-[13].
Received signal at the PR. Let yP R (n; k) denote the received signal at the PR for the kth
symbol of block n, then we have
yP R (n; k) =
+
p
p
h1 (n)xP T (n; k)
g2 (n)xST (n; k) + NP R (n; k), ∀k,
(3)
where NP R (n; k) denotes the Gaussian receiving noise at the PR with zero mean and variance
σP2 R . Then, the instantaneous received signal-to-interference-plus-noise ratio (SINR) at the PR
for block n denoted by γP R (n) is given by
γP R (n) =
h1 (n)p(n)
.
g2 (n)α(n)f (n)p(n) + σP2 R
(4)
Received signal at the SR. Let ySR (n; k) denote the received signal at the SR for the kth
symbol of block n, then we have
p
g1 (n)xST (n; k)
p
+ h2 (n)xP T (n; k) + NSR (n; k), ∀k,
ySR(n; k) =
(5)
where NSR (n; k) denotes the Gaussian receiving noise at the SR with zero mean and variance
2
σSR
. In this paper, we assume that SR decodes the received signal by performing successive
interference cancellation (SIC). SIC is a well-known physical layer technique [19], [20]. Briefly,
SIC is the ability of a receiver to receive two or more signals concurrently. The SIC receiver
decodes the stronger signal first, subtracts it from the combined signal, and extracts the weaker
one from the residue. For the system setup considered in this paper, the secondary system is
an ambient backscatter system. The strength of the signal received from the ST is in general
much lower than that received from the PT (e.g., TV/WiFi signals). Thus, the SIC procedure
at the SR is decoding the primary signal first and then subtracting it from the received signal
before decoding its own signal. Thus, the instantaneous received SNR at the SR for the block
n denoted by γSR (n) is given by
γSR (n) =
g1 (n)α(n)f (n)p(n)
.
2
σSR
(6)
Note that we assume that the SR performs SIC decoding while the PR does not. This is due
the fact that SIC needs to decode the stronger signal first. As mentioned above, the strength of
7
the signal received from the ST is in general much lower than that received from the PT. Thus,
SIC is not applicable at the PR.
III. E RGODIC C APACITY M AXIMIZATION
Under the system model given in Section II, the ergodic capacity of the secondary system can
be written as
CSR
g1 (n)α(n)f (n)p(n)
,
= E log2 1 +
2
σSR
(7)
where E [·] denotes the statistic expectation, and it is taken over the joint fading states of the
fading block n. In this paper, our objective is to maximize the ergodic capacity CSR of the
secondary system while guaranteing the performance of the primary system. In the following,
we introduce the constraints that need to be considered when optimizing this network.
PT’s transmit power constraint. Let Ppk denote the maximum transmit power of the PT,
then the peak transmit power constraint can be written as
0 ≤ p(n) ≤ Ppk , ∀n.
(8)
The reflection coefficient constraint. Since the tag is a passive device, thus the energy
harvested and reflected from the tag must be equal to the energy received from the primary
signal. Thus, the reflection coefficient must satisfy the following constraint
0 ≤ α(n) ≤ 1, ∀n.
(9)
PR’s rate constraint. To guarantee the quality of service (QoS) of the primary system, we
assume that there is a minimum rate requirement, which can be written as
h1 (n)p(n)
≥ γ, ∀n.
log2 1 +
g2 (n)α(n)f (n)p(n) + σP2 R
(10)
where γ is the minimum rate of the primary system.
Tag’s (ST’s) circuit operation power requirement. As aforementioned, the ST harvests
energy from the primary signal to power its circuit operation. Let ǫST be the minimum power
that the ST needs to support its circuit operation, then the following constraint must be satisfied
in order for the ST to work, i.e.,
ηST (1 − α(n))f (n)p(n) ≥ ǫST , ∀n.
where ηST is the energy harvesting efficiency coefficient.
(11)
8
In this paper, the objective is to optimize the performance of such a spectrum sharing system
by jointly optimizing the transmit power p(n) of the PT and the reflection coefficient α(n) of
the ST. The problems can be formulated as
P1:
Max
{p(n), α(n)}
CSR ,
s.t. (8), (9), (10), (11).
(12)
(13)
For notation convenience, the fading block number n is dropped from now on.
P1 is not a convex optimization problem. A problem is a convex optimization problem if its
objective function is either convex or concave, and its feasible set is a convex set. For P1, its
objective function is neither convex nor concave. This can be verified by looking at its Hessian
matrix (A function is convex if its Hessian matrix is positive semi-definite). According to the
composition rule given in [21], for P1, we can determine the convexity of the objective function
by investigating the convexity of α ∗ p. The Hessian matrix of α ∗ p is [0 1; 1 0], which is
not positive semi-definite. Thus, P1 is not a convex optimization problem, and the conventional
convex optimization techniques can not be applied to solve P1. To solve P1, we first introduce
the following Lemma.
Lemma 1. The largest α that makes P1 feasible, denoted by αL , is given by
ǫST
h1
σP2 R
1
, 1−
,
−
αL = max 0,min
g2 f 2γ −1 Ppk
ηST f Ppk
(14)
Proof. The constraint (10) can be rewritten as
α≤
h1
σP2 R
−
.
(2γ − 1)g2 f
g2 f p
(15)
It is observed that the right hand side of (15) is an increasing function of p. Besides, due to the
fact that 0 ≤ p ≤ Ppk , the largest α denoted by αL1 satisfies (15) is
αL1 =
h1
σP2 R
.
−
(2γ − 1)g2 f
g2 f Ppk
(16)
Similarly, the constraint (11) can be rewritten as
α≤1−
ǫST
.
ηST pf
(17)
The right hand side of (17) is an increasing function of p. Besides, due to the fact that 0 ≤ p ≤
Ppk , the largest α denoted by αL2 satisfies (17) is
αL2 = 1 −
ǫST
.
ηST f Ppk
(18)
9
Thus, the largest α that satisfies both (15) and (17) is given by min{αL1 , αL2 }. Combining
with the fact that 0 ≤ α ≤ 1, Lemma 1 follows.
Theorem 1. The optimal solution of P1 is given by
p∗ = Ppk ,
(19)
α ∗ = αL ,
(20)
where αL is given by (14).
Proof. First, it is observed from P1 that all the constraints are instantaneous constraints. Thus,
maximizing the ergodic capacity is equivalent to maximizing the instantaneous transmission
g1 αf p
. For any given α in the feasible region, the instantaneous rate is a
rate, i.e., log2 1 + σ2
SR
monotonically increasing function with respect to p, and it attains the maximum value when
p = Ppk . It is observed that for any given feasible p, the instantaneous rate is a monotonically
increasing function with respect to α. Thus, α should be chosen as the largest α that makes P1
feasible, i.e., αL given by (14), and it is shown in the proof of Lemma 1 that αL is obtained
when p = Ppk . Thus, it is clear that P1 is maximized when p∗ = Ppk and α∗ = αL .
IV. P RACTICAL E NERGY C ONSUMPTION M ODEL
In this section, we consider a more practical energy consumption model of the ST’s backscatter
circuit, which is
ηST (1 − α(n))f (n)p(n) ≥ ǫb + ǫs (rST (n)) ,
(21)
where ǫb denotes the static energy consumption when the circuit is on, and ǫs (rST (n)) denotes
the dynamic energy consumption which is a function of its transmission rate. In practice, the
dynamic energy consumption is in general proportional to the transmission rate. This is due to the
following fact. The backscatter transmitter maps its bit sequence to RF waveforms by adjusting
the load impedance of the antenna. The backscatter can control the rate at which it will generate
the modulation symbols by controlling the switching frequencies on the SPDT (single pole
double throw) switches [14]. In general, to achieve a higher data rate, the backscatter needs
a higher frequency operation on the switches, which will cost more energy. Thus, we model
(n)p(n)
, where u is a
the dynamic energy consumption by ǫs (rST (n)) = u log2 1 + g1 (n)α(n)f
σ2
SR
constant conversion parameter that relates the transmission rate with the energy consumption. In
10
this section, under this energy consumption model, we re-investigate the optimization problem
for this spectrum sharing system.
P2:
Max
{p(n), α(n)}
CSR ,
s.t. (8), (9), (10), (21).
(22)
(23)
The constraint (21) is a hyper-function with respect to α and p, which makes the problem
difficult to solve. Thus, to solve P2, we first present the following two propositions.
Proposition 1. Let α̂ denote the largest α that satisfies the constraint (21) for a given p, then
α̂ can be obtained by solving the following equation:
g1 α̂f p
ηST (1 − α̂)f p = ǫb + u log2 1 + 2
.
σSR
(24)
Proof. It is observed that the left hand side of (21) (i.e., ηST (1 − α)f p) is a monotonically
g1 αf p
)
decreasing function with respect to α, while the right hand side of (21) (ǫb + u log2 1 + σ2
SR
is a monotonically increasing function with respect to α. Thus, it is easy to observe that the
largest α is the intersection point of two curves, which is the solution of (24). Proposition 1 is
thus proved.
Proposition 2. Let αB1 and αB2 be the solution of (24) when p = p1 and p = p2 , respectively.
Then, we have
αB1 p1 < αB2 p2 , if p1 < p2 .
(25)
Proof. Since αB1 and αB2 be the solution of (24) when p = p1 and p = p2 , respectively. We
have
g1 αB1 f p1
ηST (1 − αB1 )f p1 = ǫb + u log2 1 +
.
2
σSR
(26)
g1 αB2 f p2
.
ηST (1 − αB2 )f p2 = ǫb + u log2 1 +
2
σSR
(27)
Then, using (27) to minus (26), we have
2
g1 f αB2 p2 + σSR
.
ηST f (p2 −p1+ αB1 p1 −αB2 p2 ) = u log2
2
g1 f αB1 p1 + σSR
(28)
Then, in the following, we prove Proposition 2 by contradiction. Assume αB1 p1 ≥ αB2 p2
when p1 < p2 . Then, under this presumption, it is clear that the left hand side of (28) is strictly
11
positive, while the right hand side of (28) is zero or negative. This contradicts with the fact that
the left hand side of (28) should be equal to the right hand side of (28). Thus, our presumption
does not hold. Thus, it follows that αB1 p1 < αB2 p2 , if p1 < p2 . Proposition 2 is thus proved.
However, unlike P1, we cannot further prove αB1 < αB2 , if p1 < p2 . Thus, the approach used
to solve P1 can not be applied here. Thus, to solve P2, we first consider the following problem,
which is
P2a: Max CSR ,
(29)
{p, α}
s.t. (8), (9), (21).
(30)
Let αB and αpk be the solution of (24) when p = PB and p = Ppk , respectively. Then,
from Proposition 2, it follows αB PB < αpk Ppk , ∀PB < Ppk . Since the objective function is an
increasing function with respect to αp, it is clear that the objective function attains its maximum
value at αpk Ppk . Thus, the optimal solution of P2a can be obtained as
p∗ = Ppk ,
(31)
α∗ = αpk ,
(32)
Now, we return to P2. It is clear that the constraint (10) can be rewritten as α ≤
σ2
h1
− g2PfRp .
(2γ −1)g2 f
It is observed that its right hand side is an increasing function of p. Besides, due to the fact that
0 ≤ p ≤ Ppk , the largest α satisfying α ≤
αM =
h1
(2γ −1)g2 f
−
2
σP
R
g2 f p
denoted by αM is
h1
σP2 R
−
.
(2γ − 1)g2 f
g2 f Ppk
(33)
Based on these results, we are now able to solve P2, and the solution is summarized in the
following theorem.
Theorem 2. The optimal solution of P2 is given by
p∗ = Ppk ,
(34)
α∗ = min {αM , αpk }
(35)
where αM is given by (33), and αpk is the solution of (24) when p = Ppk .
Proof. To prove Theorem 2, we consider the following two cases.
12
Case 1: αpk ≤ αM . In this case, the optimal solution of P2 is the same as that of P3a, which
is α∗ = αpk and p∗ = Ppk .
Case 2: αpk > αM . In this case, the optimal α∗ must satisfy the condition 0 ≤ α∗ ≤ αM .
Now, we look at the following equation,
ηST (1 − αpk )f Ppk
g1 αpk f Ppk
.
= ǫb + u log2 1 +
2
σSR
(36)
The above equation comes from the fact that αpk is the solution of (24) when p = Ppk . Since
αpk > αM , thus if we replace αpk with αM , the left hand side of (36) will increase, while the
right hand side of (36) will decrease. Thus, it follows
g1 αM f Ppk
,
ηST (1 −αM )f Ppk ≥ ǫb + u log2 1 +
2
σSR
(37)
which indicates αM and Ppk is a feasible solution of P2. Since αM is the largest feasible α∗ and
Ppk is the largest feasible p∗ , thus the optimal solution is α∗ = αM and p∗ = Ppk .
Summarizing the above results, Theorem 2 follows.
V. F IXED R EFLECTION C OEFFICIENT
AND
AVERAGE T RANSMIT P OWER C ONSTRAINT
Fixed Reflection Coefficient. In practice, for the purpose of circuit design simplicity, the
reflection coefficient of the tag is designed to be fixed, i.e., the reflection cannot be dynamically
changed in each fading block. To capture this fact, we introduce the following constraint
α(n) = α, ∀n.
(38)
Average transmit power constraint. In practice, there is always a long-term power budget
of the PT, and an average power constraint usually applies, which can be written as
E [p(n)] ≤ Pav ,
(39)
where the statistic expectation is taken over the joint fading states of the fading block.
In the following subsections, we reinvestigate P1 and P2 under the above two practical
constraints.
A. Ideal Energy Consumption Model
P3:
Max
{p(n), α(n)}
CSR ,
s.t. (9), (10), (11), (38), (39).
(40)
(41)
13
The objective function of P3 is the same as P1, and thus it is easy to verify that P3 is not a
convex optimization problem. Therefore, it cannot be solved directly using convex optimization
techniques. Thus, for solving P3, we first consider P3 under a given α = ᾱ. For given ᾱ, P3
can be rewritten as
P3a:
Max
{p(n)≥0}
g1 (n)ᾱf (n)p(n)
,
E log2 1 +
2
σSR
s.t. E[p(n)] ≤ Pav ,
h1 (n)p(n)
log2 1 +
≥ γ,
g2 (n)ᾱf (n)p(n) + σP2 R
ǫST − ηST (1 − ᾱ)f (n)p(n) ≤ 0.
(42)
(43)
(44)
(45)
Note the fading block number n is dropped from now on for notation convenience.
P3a is a convex optimization problem since its objective function is concave and all the
constraints are affine. To solve this problem, we first at the feasibility of the problem. Note that
the constraint (44) is infeasible if h1 − (2γ − 1)g2 ᾱf < 0, i.e., no matter how p is chosen, (44)
cannot be satisfied for such fading block. Thus, to save power, the optimal power allocation for
such fading block is p∗ = 0. When feasible, it can be shown that constraints (44) and (45) can
be rewritten as p ≥
2 (2γ −1)
σP
R
h1 −(2γ −1)g2 ᾱf
and p ≥
ǫST
,
ηST (1−ᾱ)f
respectively. Thus, (44) and (45) can be
replaced by the following constraint
p ≥ Pm ,
(46)
where
σP2 R (2γ −1)
ǫST
Pm , max
.
,
h1 −(2γ −1)g2 ᾱf ηST (1 − ᾱ)f
Theorem 3. The optimal solution of P3a is given by
0,
if λ ≥ σ2g1 ᾱf
,
SR ln 2
1 ᾱf
if σ2g1 ᾱf
> λ > ln2 σ2 g+g
,
p∗ = P m ,
( SR 1 ᾱf Pm )
SR ln 2
σ2
1
1 ᾱf
λ ln
.
− g1SR
, if λ ≤ ln2 σ2 g+g
2
ᾱf
( SR 1 ᾱf Pm )
where λ can be obtained by solving E [p∗ ] = Pav , and Pm is given by (47).
(47)
(48)
Proof. By introducing the dual variable associated with the average transmit power constraint,
the partial Lagrangian of P3a problem is expressed as
g1 ᾱf p
− λ (E[p] − Pav ) ,
L (p, λ) = E log2 1 + 2
σSR
(49)
14
where λ is nonnegative Lagrange dual variable associated with constraints (43).
Let A denote the set of {p ≥ Pm }. The dual function is then expressed as q(λ) = maxp∈A L (p, λ) .
The Lagrange dual problem is then defined as minλ≥0 q(λ). The duality gap is zero for the convex
optimization problem addressed here, and thus solving its dual problem is equivalent to solving
the original problem. The duality gap is zero if and only if strong duality holds. If the primal
problem is convex and satisfies the Slaters condition [21], then strong duality holds. The Slater’s
condition reduces to feasibility when the constraints are all linear equalities and inequalities [21].
P3a is a convex problem and all its constraints are linear inequalities. It also can be verified
that P3a is feasible under our constraints. Thus, the duality gap is zero for P3a. Therefore,
according to the Karush-Kuhn-Tucker (KKT) conditions, the optimal solutions needs to satisfy
the following equations:
λ∗ (E[p∗ ] − Pav ) = 0,
(50)
λ∗ ≥ 0, E[p∗ ] − Pav ≤ 0,
(51)
For a fixed λ, by dual decomposition, the dual function can be decomposed into a series of
similar sub-dual-functions each for one fading state. For a particular fading state, the problem
can be shown equivalent to
P3b:
Max log2
{p≥0}
g1 ᾱf p
1+ 2
σSR
− λp,
s.t. p ≥ Pm ,
(52)
(53)
It is easy to observe that the optimal solution of this subproblem is p∗ = +∞ if λ = 0. Thus,
in the following, we consider the optimal solution of this subproblem under the condition that
λ 6= 0.
The Lagrangian of this subproblem is
g1 ᾱf p
Lsub (p, µ) = log2 1 + 2
− λp − µ (Pm − p) + νp,
σSR
(54)
where µ and ν are nonnegative Lagrange dual variables associated with the constraints p ≥ Pm
and p ≥ 0, respectively. Since the problem is convex, KKT conditions are sufficient to obtain
15
its optimal solution. Thus, in the following, we investigate its KKT conditions:
g1 ᾱf
∂L (p∗ , µ∗ )
=
−λ + µ∗ + ν ∗ = 0,
2
∂p∗
ln2 (σSR + g1 ᾱf p∗ )
(55)
µ∗ (p∗ − Pm ) = 0,
(56)
ν ∗ p∗ = 0,
(57)
p∗ − Pm ≥ 0,
(58)
p∗ ≥ 0, µ∗ ≥ 0, ν ∗ ≥ 0.
(59)
Now, we derive the optimal solution by solving these KKT conditions. To solve these KKT
conditions, we consider the following two cases:
•
Case 1: µ∗ > 0, ν ∗ = 0. For this case, it follows from (56) that
p∗ = P m .
Then, based on (55), µ∗ can be obtained by solving
(60) holds only when λ −
•
g1 ᾱf
2 +g ᾱf P
ln2(σSR
m)
1
(60)
g1 ᾱf
2 +g ᾱf P
ln2(σSR
m)
1
− λ + µ∗ = 0. Thus,
> 0.
Case 2: µ∗ > 0, ν ∗ > 0. In this case, it follows from (56) that p∗ = Pm . However, it follows
from (57) that p∗ = 0. Thus, by contradictory, this case cannot happen.
•
Case 3: µ∗ = 0, ν ∗ = 0. For this case, it follows from (55) that
p∗ =
σ2
1
− SR .
λ ln 2 g1 ᾱf
(61)
Then, taking (58) into consideration, (62) holds only when
•
g1 ᾱf
2 +g ᾱf P
ln2(σSR
m)
1
∗
1
λ ln 2
−
2
σSR
g1 ᾱf
≥ Pm , i.e., λ −
≤ 0.
Case 4: µ = 0, ν ∗ > 0. It follows from (57) that
p∗ = 0.
Then, based on (55), ν ∗ can be obtained by solving
only when λ −
g1 ᾱf
2 ln 2
σSR
(62)
g1 ᾱf
2 ln 2
σSR
− λ + ν ∗ = 0. Thus, (62) holds
> 0.
Thus, combining the results obtained in Case 1 to 4, the optimal solution for the subproblem
can be summarized as (48). Now, we have to find the optimal λ∗ . As aforementioned, λ∗ has
to satisfy (87) and (89). It is observed that if λ∗ = 0, the optimal solution for the subproblem
is p∗ = +∞. This definitely violates the constraint E[p∗ ] ≤ Pav . Thus, it follows that λ∗ 6= 0.
Then, according to the constraint λ∗ (E[p∗ ] − Pav ) = 0, it follows that the optimal λ∗ must satisfy
E[p∗ ] = Pav . Theorem 2 is thus proved.
16
With the optimal solution of P3a given in Theorem 3, the optimal solution of P3 can be
obtained by performing a one-dimension search for α over the space [0, 1].
B. Practical Energy Consumption Model
P4:
Max
{p(n), α(n)}
CSR ,
(63)
s.t. (9), (10), (21), (38), (39).
(64)
The objective function of P4 is the same as P1, and thus it is easy to verify that P4 is not a
convex optimization problem. Therefore, it cannot be solved directly using convex optimization
techniques. Thus, before solving P4, we first present the following proposition.
Proposition 3. For a given ᾱ, the constraint (21) can be rewritten as
p ≥ Pc ,
(65)
where Pc is the positive solution of ηST (1 − ᾱ)f Pc = ǫb + u log2 1 +
g1 ᾱf Pc
2
σSR
.
Proof. For the convenience of exposition, we introduce the following two functions F1 (p) ,
p
. It is easy to observe that both F1 (p) and
ηST (1 − ᾱ)f p and F2 (p) , ǫb + u log2 1 + gσ1 ᾱf
2
SR
F2 (p) are monotonically increasing functions with respect to p, and the increasing rate of F1 (p)
is larger than that of F2 (p). It is also observed that F1 (0) ≤ F2 (0). Thus, F1 (p) and F2 (p)
must have a positive crossing point Pc . For any p larger than Pc , we have ηST (1 − ᾱ)f p ≥
g1 ᾱf p
ǫb + u log2 1 + σ2
. Proposition 3 is thus proved.
SR
It can be shown that constraints (10) can be rewritten as p ≥
2 (2γ −1)
σP
R
.
h1 −(2γ −1)g2 ᾱf
Thus, (10) and
(21) can be replaced by the following constraint
p ≥ PL ,
(66)
where
PL , max
σP2 R (2γ −1)
, Pc .
h1 −(2γ −1)g2 ᾱf
(67)
Using the same approach as P3, for a given ᾱ, the optimal power allocation of P4 can be
summarized as follows.
17
Theorem 4. For a given ᾱ, the optimal power allocation of P4 is given by
,
0,
if λ ≥ σ2g1 ᾱf
SR ln 2
ᾱf
if σ2g1 ᾱf
> λ > ln2 σ2 g1+g
,
p∗ = P L ,
ln 2
(
1 ᾱf PL )
SR
SR
2
σSR
ᾱf
1
λ ln
−
, if λ ≤ ln2 σ2 g1+g
.
2
g1 ᾱf
( SR 1 ᾱf PL )
(68)
where λ can be obtained by solving E [p∗ ] = Pav , and PL is given by (67).
With the optimal solution of P4 under given ᾱ obtained in Theorem 4, the optimal solution
of P4 can be obtained by performing a one-dimension search for α over the the space [0, 1].
VI. P RIMARY T RANSMISSION O UTAGE C ONSTRAINT
In previous sections, we use the PR’s rate constraint to guarantee the quality of service (QoS)
of the primary system, i.e., there is a minimum rate requirement that the primary transmission has
to fulfil. However, this constraint is too strict. In practice, certain ratio of transmission outage
is usually acceptable. Thus, in this section, we introduce the primary transmission outage
constraint [22], which is mathematically defined as
h1 (n)p(n)
≤ γ ≤ ǫ,
Prob log2 1 +
g2 (n)α(n)f (n)p(n) + σP2 R
(69)
where Prob {·} denotes the probability of the event, and ǫ denotes the maximum outage probability that is acceptable by the primary system.
A. Peak Transmit Power Constraint
We first consider the peak transmit power constraint case. Using the primary transmission outage constraint to replace the PR’s rate constraint, and considering the fixed reflection coefficient,
the problem can be formulated as
P5:
Max
{p(n), α(n)}
CSR ,
s.t. (8), (9), (11), (38), (69).
(70)
(71)
The objective function of P5 is the same as P1, and thus it is easy to verify that P5x is not a
convex optimization problem. Therefore, it cannot be solved directly using convex optimization
techniques. We first consider P5 under a given ᾱ, i.e., α = ᾱ where ᾱ is a constant. For
expression convenience, we denote the problem of P5 under a given ᾱ as P5a.
18
Theorem 5. The optimal power allocation of P5a, i.e., P5 under a given ᾱ, is given by
p∗ (n) = Ppk , ∀n.
(72)
Proof. For a given ᾱ, the constraint (11) can be rewritten as p(n) ≥
fading block n, it is easy to observe that P5a is infeasible when
ǫST
. For a given
ηST (1−ᾱ)f (n)
ǫST
> Ppk . This
ηST (1−ᾱ)f (n)
means that there does not exist a feasible p(n) that can satisfy the ST’s circuit operation power
requirement, which indicates the secondary system does not work in this fading block. Thus,
the contribution of this fading block to the secondary transmission’s ergodic capacity is zero.
The interference from the secondary system to the primary system is also zero. Thus, channel
inversion power allocation or the maximum transmit power should be adopted for the primary
system to minimize the possibility of outage for this fading block. Therefore, for simplicity, the
optimal power allocation for such a fading block can be obtained as
p∗ (n) = Ppk .
In the following, we consider the case that
introduce the following indicator function
1, if log 1 +
2
χ=
0, otherwsie.
ǫST
ηST (1−ᾱ)f (n)
(73)
≤ Ppk . To solve the problem, we
h1 (n)p(n)
2
g2 (n)α(n)f (n)p(n)+σP
R
≤ γ,
(74)
For notation convenience, the fading block number n is dropped from now on. Then, it can be
shown that the constraint given in (69) is equivalent to
E [χ] − ǫ ≤ 0.
(75)
By introducing the dual variable associated with (75), the partial Lagrangian of P5 under given
ᾱ can be expressed as
g1 ᾱf p
− λ (E[χ] − ǫ) ,
L (p, λ) = E log2 1 + 2
σSR
(76)
where λ is nonnegative Lagrange dual variable associated with the constraint E [χ] − ǫ ≤ 0.
o
n
ST
≤
p
≤
P
Let A denote the set of ηST ǫ(1−
pk . The dual function is then expressed as
ᾱ)f
q(λ) = maxp∈A L (p, λ) . According to the Karush-Kuhn-Tucker (KKT) conditions, the optimal
solution needs to satisfy the following equations:
λ∗ (E[χ∗ ] − ǫ) = 0,
(77)
λ∗ ≥ 0, E[χ∗ ] − ǫ ≤ 0,
(78)
19
³
log2 1 +
g1 ᾱfp
2
σSR
´
³
log2 1 +
g1 ᾱfp
2
σSR
´
− λχ(p)
p=
2
(2γ −1)σP
R
h1 −(2γ −1)g2 αf
p
0
Fig. 2. An illustration of function log2 1 +
g1 ᾱf p
2
σSR
− λχ(p).
For a fixed λ, by dual decomposition, the dual function can be decomposed into a series of
similar sub-dual-functions each for one fading state. For a particular fading state, the problem
can be shown equivalent to
P5b:
g1 ᾱf p
− λχ(p),
Max log2 1 + 2
{p≥0}
σSR
ǫST
,
s.t. p ≥
ηST (1 − ᾱ)f
p ≤ Ppk .
Since log2 1 +
λ = 0.
g1 ᾱf p
2
σSR
(79)
(80)
(81)
is an increasing function, the optimal solution of P5b is p∗ = Ppk if
In the following, we consider the optimal solution of this subproblem under the condition that
λ 6= 0. Fisrt, it can be observed that χ(p) can be rewritten as
Case 1: h1 − (2γ − 1)g2 αf ≤ 0.
χ(p) = 1, ∀p.
(82)
20
Case 2: h1 − (2γ − 1)g2 αf > 0.
1,
χ(p) =
0,
if p ≤
if p >
2
(2γ −1)σP
R
,
h1 −(2γ −1)g2 αf
γ
2
(2 −1)σP R
.
h1 −(2γ −1)g2 αf
(83)
Let us first consider Case 1. In this case, the objective function of P5b becomes log2 1 +
g1 ᾱf p
2
σSR
−
λ, which is an increasing function with respect to p. Thus, the optimal solution of P5b is p∗ = Ppk .
Now we consider Case 2. It is observed from (83) that χ(p) is a step function with respect to
p for this case. The critical point is p =
2
(2γ −1)σP
R
.
h1 −(2γ −1)g2 αf
Thus, the objective function of P5b has a
p
−
form as illustrated in Fig. 2. It can be observed from Fig. 2 that the function log2 1 + gσ1 ᾱf
2
SR
∗
λχ(p) is in general an increasing function of p. Thus, the optimal solution of P5b is p = Ppk .
Combining all these results, the optimal solution for P5a can be obtained as (72). Theorem 5
is thus proved.
With the optimal solution of P5a (i.e, P5 under given ᾱ) obtained in Theorem 5, the optimal
solution of P5 can be obtained by performing a one-dimension search for α over the space [0, 1].
B. Average Transmit Power Constraint
Now, we investigate the problem under the average transmit power constraint, which is
P6:
Max
{p(n), α(n)}
CSR ,
s.t. (9), (11), (38), (39), (69).
(84)
(85)
We first consider P6 under a given ᾱ, i.e., α = ᾱ where ᾱ is a constant. For expression
convenience, we denote the problem of P6 under a given ᾱ as P6a.
Same as solving P5, it can be shown that (69) can be rewritten as (75). By introducing the
dual variables associated with (75) and (39), the partial Lagrangian of P6a under given ᾱ can
be expressed as
g1 ᾱf p
− λ (E[χ] − ǫ)
L (p, λ) = E log2 1 + 2
σSR
− µ (E[p] − Pav ) ,
(86)
where λ and µ are nonnegative Lagrange dual variables associated with the constraints E [χ]−ǫ ≤
0 and E [χ] − Pav ≤ 0, respectively.
21
Let A denote the set of
n
ǫST
ηST (1−ᾱ)f
o
≤ p . The dual function is then expressed as q(λ) =
maxp∈A L (p, λ) . According to the Karush-Kuhn-Tucker (KKT) conditions, the optimal solution
needs to satisfy the following equations:
λ∗ (E[χ∗ ] − ǫ) = 0,
(87)
µ∗ (E[p∗ ] − Pav ) = 0,
(88)
λ∗ ≥ 0, E[χ∗ ] − ǫ ≤ 0,
(89)
µ∗ ≥ 0, E[p∗ ] − Pav ≤ 0.
(90)
For a fixed λ, by dual decomposition, the dual function can be decomposed into a series of
similar sub-dual-functions each for one fading state. For a particular fading state, the problem
can be shown equivalent to
P6b:
g1 ᾱf p
− λχ(p) − µp,
Max log2 1 + 2
{p≥0}
σSR
ǫST
s.t. p ≥
.
ηST (1 − ᾱ)f
(91)
(92)
P6a can be solved by iteratively solving P6b for all fading states with fixed λ and µ, and updating
these dual variables via sub-gradient based method, e.g., the ellipsoid method [21], for which
the details are omitted here for brevity.
In the following, we derive the solution to P6b. To solve the problem, we first introduce the
following function
F (p) , log2
g1 ᾱf p
1+ 2
σSR
− µp.
(93)
It is easy to show that F (p) is a concave function with respect to p, and attains its maximum
value when p is equal to
p̃ =
1
σ2
− SR .
µ ln 2 g1 ᾱf
(94)
Note that the objective function in P6b now becomes F (p) − λχ(p). Now we look at χ(p). It
has been shown in the previous subsection that χ(p) can be rewritten as (82) and (83).
Thus, when h1 −(2γ −1)g2 αf ≤ 0, we have χ(p) = 1, ∀p. Then, the objective function of P6b
becomes F (p) − λ, which attains its maximum value at the same p as F (p). Thus, the optimal
solution of P6b for this case is p∗ = p̃.
22
F(p)
F(p) − λχ(p)
0
p′
p̃
p
(a) p ′ ≤ p̃
F(p)
F(p) − λχ(p)
F(p) − λχ(p)
p′
p̃
0
p
(b) p ′ > p̃
Fig. 3. An illustration of different forms of function F(p) − λχ(p).
When h1 − (2γ − 1)g2 αf > 0, as shown in (83), χ(p) is a step function with respect to p,
and its critical point is p =
2
(2γ −1)σP
R
.
h1 −(2γ −1)g2 αf
In the following, we consider P6b for this case. For
the convenience of discussion, we denote the critical point of χ(p) as p′ , i.e.,
p′ =
(2γ − 1)σP2 R
.
h1 − (2γ − 1)g2 αf
(95)
For the convenience of discussion, we rewrite the constraint (92) as p ≥ p′′ , where
p′′ =
ǫST
.
ηST (1 − ᾱ)f
(96)
23
Let p∗ denote the optimal solution of P6b. The following discussions are then made on p∗ :
Case 1: p′ ≤ p̃. An illustration of this case is given in Fig. 3(a). It can be seen from Fig. 3(a)
that F (p) attains it maximum value at p = p̃, and at the value of p = p̃, χ(p̃) is equal to zero.
Thus, F (p) − λχ(p) attains its maximum value at p = p̃. Now, in order to obtain the optimal
solution for P6a, we have to consider the following two subcases based on the relationship
among p′ , p′′ , and p̃.
•
Subcase 1: p′′ ≤ p̃. In this case, p̃ is within the feasible region. Thus, the optimal solution
for P6b is p∗ = p̃.
•
Subcase 2: p̃ < p′′ . In this case, p̃ is not within the feasible region. It is observed from Fig.
3(a) that F (p) − λχ(p) is a decreasing function with respect to p for any p ≥ p̃. Thus, the
optimal solution for P6b is p∗ = p′′ for this case.
Case 2: p′ > p̃. An illustration of this case is given in Fig. 3(b). This case is a little bit
complex than Case 1. As shown in Fig. 3(b), F (p) − λχ(p) may attain its maximum value at
p = p̃ or p = p′ depending on the value of λ. Thus, the following three subcases are considered
for finding the optimal p∗ for P6b.
•
Subcase 1: p′′ < p̃ < p′ . For this case, both p̃ and p′ are within the feasible region. Thus,
•
the optimal solution can be obtained as
p̃, if F (p̃) − λ > F (p′ ),
p∗ =
p′ , if F (p̃) − λ ≤ F (p′ ).
(97)
Subcase 2: p̃ ≤ p′′ ≤ p′ . For this case, only p′ is within the feasible region. However, it is
possible that the value of F (p′′ ) − λ may be larger than that of F (p′ ). Thus, the optimal
solution can be obtained as
•
p′′ ,
p∗ =
p′ ,
if F (p′′ ) − λ > F (p′),
if F (p′′ ) − λ ≤ F (p′ ).
(98)
Subcase 3: p̃ < p′ < p′′ . For this case, both p̃ and p′ are not within the feasible region. It
is observed from Fig. 3(b) that F (p) − λχ(p) is a decreasing function with respect to p for
any p ≥ p′ . Thus, the optimal solution for P6b is p∗ = p′′ for this case.
Combining all the results obtained above, the optimal solution for P6a can be summarized in
the following theorem.
24
Theorem 6. The optimal power allocation of P6a, i.e., P6 under a given ᾱ, is given by
p̃, if R1 holds,
p∗ (n) =
p′ ,
p′′ ,
if R2 holds,
(99)
if R3 holds,
where p̃, p′ , p′′ are given by (94), (95), (96), respectively. R1 , R2 , R3 are defined as R1 ,
{{h1 − (2γ − 1)g2 αf ≤ 0} or {h1 − (2γ − 1)g2 αf > 0 and p̃ > max {p′ , p′′ }} or {p′ > p̃ > p′′
and h1 − (2γ − 1)g2 αf > 0 and F (p̃) − λ > F (p′ )}}.
R2 , {{h1 − (2γ − 1)g2αf > 0 and p′ > p̃ > p′′ and F (p̃) − λ < F (p′ )} or {p′ > p′′ > p̃
and h1 − (2γ − 1)g2 αf > 0 and F (p′′ ) − λ < F (p′ )}}.
R3 , {{h1 − (2γ − 1)g2 αf > 0 and p′′ > max {p′ , p̃}} or {h1 − (2γ − 1)g2 αf > 0 and p′ >
p′′ > p̃ and F (p′′ ) − λ > F (p′)}}.
With the optimal solution of P6a (i.e, P6 under given ᾱ) obtained in Theorem 6, the optimal
solution of P6 can be obtained by performing a one-dimension search for α over the space [0, 1].
VII. N UMERICAL R ESULTS
In this section, several numerical examples are presented to evaluate the performance of the
derived results. We assume i.i.d. Rayleigh fading for all channels involved, and thus the channel
power gains of these channels are exponentially distributed. The stochastic mean of the channel
power gain is assumed to be one. It is worth pointing out that the assumption of particular
distributions of the channel power gains does not affect the structure of the problem studied and
the solution obtained. The power of the noises at the receiver of PR and SR are assumed to be
one. The energy harvest efficiency ηST for ST is assumed to be one. The constant conversion
parameter u is assumed to be one. The results given in these examples are obtained by averaging
over 10000 channel realizations.
A. Ideal Vs. Practical Energy Consumption Model
In this subsection, we investigate the impact of the energy consumption model on the system
performance for two different scenarios given below.
1) Peak transmit power constraint with dynamic reflection coefficient: The curves presented
in Fig 4 are obtained based on the solutions to P1 and P2. It is observed from Fig. 4 that the
ergodic capacity of the secondary system increases with the increasing of Ppk for all curves. It is
Ergodic Capacity of Secondary System (bit/complex dim.)
25
4.5
Ideal Energy Model, γ=1
Ideal Energy Model, γ=2
Practical Energy Model, γ=1
Practical Energy Model, γ=2
4
3.5
3
2.5
2
1.5
1
0.5
0
0
5
10
15
Peak Transmit Power Constraint, Ppk (dB)
20
Ergodic Capacity of Secondary System (bit/complex dim.)
Fig. 4. Peak Transmit Power Constraint with Dynamic Reflection Coefficient
3.5
Ideal Energy Consumption Model, γ=1
Practical Energy Consumption Model, γ=1
3
2.5
2
1.5
1
0.5
0
0
5
10
15
Average Transmit Power Constraint(dB)
Fig. 5. Average Transmit Power Constraint with Fixed Reflection Coefficient
20
26
also observed that the ergodic capacity of the ideal energy consumption model is larger than that
of the practical energy model for the same Ppk . This is due to the fact that for the practical energy
consumption model, the dynamic power consumption is considered, more power is needed to
support the tag’s circuit operation, and thus the power left for transmitting the signal is less,
which results in a lower transmission rate. We also observe that the capacity for γ = 1 is larger
than that for γ = 2 for both idea and practical energy consumption models. This is as expected
since a lower primary system’s rate requirement indicates a high interference tolerance, and thus
the secondary system can transmit with a higher power which results in a higher transmission
rate.
2) Average transmit power constraint with fixed reflection coefficient: In Fig.5, we show
that ergodic capacity for both ideal and practical energy consumption model under the average
transmit power constraint and the fixed reflection coefficient. The curves presented in Fig 5
are obtained based on the solutions to P3 and P4. The trend of the curves are same as that of
Fig.4. Thus, for concise, the explanations are not repeated here. However, comparing Fig.5 with
Fig.4, we observe that the ergodic capacity in Fig.5 is lower than that of Fig.4 when Ppk = Pav
for the same energy consumption model and the same γ. This is due to the fact the reflection
coefficient in Fig.4 can be dynamically adjusted to its optimal value for each fading block while
the reflection coefficient in Fig.5 remains the same for each fading block.
B. Study of the Primary Transmission Outage Constraint
In this subsection, we investigate the impact of the primary transmission outage constraint on
the system performance for two different setups given below.
1) Peak transmit power constraint: The curves presented in Fig 6 are obtained based on
the solution to P5. Firstly, it is trivial to observe from Fig. 6 that the ergodic capacity under
larger Ppk is larger for the same given primary transmission outage probability. Secondly,
the ergodic capacities for all three curves become flat when the primary transmission outage
probability is sufficiently large. This is as expected since the peak transmit power constraint
becomes the bottleneck that limits the system performance when the primary transmission
outage probability is large. It is also observed that all three curves are not starting from the
zero primary transmission outage probability. This is because the zero primary transmission
outage probability is in fact equivalent to the PR’s rate constraint given in (10). For a fixed
α and a given peak power constraint Ppk , there does not exist a power allocation that satisfies
Ergodic Capacity of Secondary System (bit/complex dim.)
27
5
Ppk=10dB
4.5
Ppk=15dB
P =20dB
4
pk
3.5
3
2.5
2
1.5
1
0.5
0
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Primary transmission outage probability
0.4
0.45
Fig. 6. Peak Power Constraint with Primary Transmission Outage Constraint
log2 1 +
h1 (n)p(n)
2
g2 (n)αf (n)p(n)+σP
R
≥ γ for a given positive γ for all the fading slots due to the
continuous distribution of the channel power gains.
2) Average transmit power constraint: In Fig. 7, four curves with different primary transmission outage probability are given. The curves presented in Fig 7 are obtained based on the
solution to P6. It is easy to observe that the ergodic capacities increase with the increasing
of the transmit power constraint. The capacity difference for four curves is small when the
transmit power constraint is small. This is intuitive since the average transmit power constraint
is the bottleneck that limits the system performance when it is very small. It is also observed
that the ergodic capacity under larger primary outage probability is larger for the same given
transmit power constraint. Besides, the curves with 100% and 0% primary outage probability
serve as the upper bound and the lower bound, respectively. This is as expected since the case
under 100% primary outage probability is equivalent to the case without the primary outage
probability constraint. The case under 0% primary outage probability is equivalent to the case
under the primary rate constraint which is studied in P3.
Ergodic Capacity of Secondary System (bit/complex dim.)
28
6
Primary Transmission Outage, 100%
Primary Transmission Outage, 30%
Primary Transmission Outage, 10%
Primary Transmission Outage, 0%
5
4
3
2
1
0
0
5
10
15
Average Transmit Power Constraint(dB)
20
Fig. 7. Average Power Constraint with Primary Transmission Outage Constraint
VIII. C ONCLUSIONS
In this paper, we proposed the Riding on the Primary (ROP) spectrum sharing model for
wireless-powered IoT devices with ambient backscatter communication capabilities. We investigated the performance of such a spectrum sharing system under fading channels. The
ergodic capacity of the secondary system was investigated by jointly optimizing the transmit
power of the primary signal and the reflection coefficient of the secondary system. Different
(ideal/practical) energy consumption models, different (peak/average) transmit power constraints,
different types (fixed/dynamically adjustable) reflection coefficient, different primary system’s
interference requirements (rate/outage) are considered were considered. Closed-form solutions
were obtained for most cases. Performance for different scenarios were studied and compared
through numerical simulations.
R EFERENCES
[1] X. Kang, Y. -C. Liang, and J. Yang, ”Riding on the primary: A new spectrum sharing paradigm for wireless-powered IoT
devices,” in Proc. IEEE ICC, Paris, France, May. 2017.
[2] 3GPP TR 22.861: ”Feasibility Study on New Services and Markets Technology Enablers for Massive Internet of Things;
Stage 1”, Version 14.0.0, Jun. 2016.
29
[3] N. Shinohara, Power without wires, IEEE Microw. Mag., vol. 12, no. 7, pp. S64C-S73, Dec. 2011.
[4] K. Huang and E. Larsson, ”Simultaneous information and power transfer for broadband wireless systems”, IEEE Trans.
Signal Process., vol. 61, no. 23, pp. 5972–5986, Dec. 2013.
[5] X. Kang, Y. K. Chia, C. K. Ho and S. Sun, ”Cost minimization for fading channels with energy harvesting and conventional
energy,” IEEE Trans. Wireless Commun., vol. 13, no. 8, pp. 4586–4598, Aug. 2014.
[6] X. Kang, C. K. Ho and S. Sun, ”Full-duplex wireless-powered communication network with energy causality,” IEEE Trans.
Wireless Commun., vol. 14, no. 10, pp. 5539–5551, Oct. 2015.
[7] C. Boyer and S. Roy, ”Backscatter communication and RFID: coding, energy, and MIMO analysis”, IEEE Trans. Commun.,
vol. 62, no. 3, pp. 770–785, Mar. 2014.
[8] C. Boyer and S. Roy, ”Space time coding for backscatter RFID”, IEEE Trans. Wireless Commun., vol. 12, no. 5, pp.
2272–2280, May. 2013.
[9] V. Liu, A. Parks, V. Talla, S. Gollakota, D. Wetherall, J. R. Smith, ”Ambient backscatter: wireless communication out of
thin air,” in Proc. ACM SIGCOMM, Hong Kong, China, Jun. 2013, pp. 39–50.
[10] B. Kellogg, A. Parks, S. Gollakota, J. R. Smith and D. Wetherall, ”Wi-fi backscatter: Internet connectivity for RF-powered
devices”, in Proc. ACM SIGCOMM, Chicago, USA, Jun. 2014, pp. 1–12.
[11] G. Wang, F. Gao, R. Fan, and C. Tellambura ”Ambient backscatter communication systems detection and performance
analysis,” IEEE Trans. Commun., vol. 64, no. 11, pp. 4836–4846, Nov. 2016.
[12] J. Qian, F. Gao, G. Wang, S. Jin, and H. Zhu ”Noncoherent detections for ambient backscatter system,” IEEE Trans.
Wireless Commun., vol. 16, no. 3, pp. 1412–1422, May. 2017.
[13] G. Yang, and Y. -C. Liang, ”Backscatter communications over ambient OFDM signals: transceiver design and performance
analysis,” in Proc. of IEEE Globecom, Washington DC, USA, Dec. 2016.
[14] D. Bharadia, K. Joshi, M. Kotaru, and S. Katti, ”BackFi: High throughput WiFi backscatter,” in Proc. of ACM SIGCOMM,
London, UK, Aug. 2015, pp. 283–296.
[15] P. Zhang, M. Rostami, P. Hu, and D. Ganesan, ”Enabling practical backscatter communication for on-body sensors,” in
Proc. of ACM SIGCOMM, Florianopolis, Brazil, Aug. 2016, pp. 370–383.
[16] K. Han and K. Huang, ”Wirelessly powered backscatter communication network: Modeling, coverage and capacity,” IEEE
Trans. Wireless Commun., vol. 16, no. 4, pp. 2548–2561, Apr. 2017.
[17] X. Kang, Y. -C. Liang, A. Nallanathan, H. K. Garg, and R. Zhang, ”Optimal power allocation for fading channels in
cognitive radio networks: Ergodic capacity and outage capacity,” IEEE Trans. Wireless Commun., vol. 8, no. 2, pp. 940–
950, Feb. 2009.
[18] F. Fuschini, C. Piersanti, F. Paolazzi and G. Falciasecca, ”Analytical approach to the backscattering from UHF RFID
transponder”, IEEE Antenna Wireless Propag. Lett., vol. 7, pp. 33–35, Feb. 2008.
[19] T. M. Cover and J. A. Thomas, ”Elements of Information Theory”, John Wiley & Sons, Inc., 1991.
[20] S. Verdu, ”Multiuser Detection”, Cambridge University Press, 1998.
[21] S. Boyd and L. Vandenberghe, Convex optimization, Cambridge University Press, 2004.
[22] X. Kang, R. Zhang, Y. -C. Liang, and H. K. Garg, ”Optimal power allocation strategies for fading cognitive radio channels
with primary user outage constraint,” IEEE J. Sel. Areas Commun., vol. 29, no. 2, pp. 374–383, Feb. 2011.
| 7cs.IT
|
arXiv:1612.06958v3 [math.GR] 14 Sep 2017
CONTRACTION GROUPS AND PASSAGE TO SUBGROUPS
AND QUOTIENTS FOR ENDOMORPHISMS OF TOTALLY
DISCONNECTED LOCALLY COMPACT GROUPS
BY
Timothy P. Bywaters
The University of Sydney, School of Mathematics and Statistics
NSW 2006, Australia
e-mail: [email protected]
AND
Helge Glöckner
Universität Paderborn, Institut für Mathematik
Warburger Str. 100, 33098 Paderborn, Germany
e-mail: [email protected]
AND
Stephan Tornier
ETH Zürich, Department of Mathematics
Rämistrasse 101, 8092 Zürich, Switzerland
The University of Newcastle, School of Mathematical and Physical Sciences
University Drive, Callaghan, NSW 2308, Australia
e-mail: [email protected]
ABSTRACT
The concepts of the scale and tidy subgroups for an automorphism of a totally disconnected locally compact group were defined in seminal work by
George A. Willis in the 1990s, and recently generalized to the case of endomorphisms (G. A. Willis, Math. Ann. 361 (2015), 403–442). We show that
central facts concerning the scale, tidy subgroups, quotients, and contraction groups of automorphisms extend to the case of endomorphisms. In
particular, we obtain results concerning the domain of attraction around
an invariant closed subgroup.
1
2
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Statement of Results
Let G be a Hausdorff topological group with neutral element e and α : G → G
be an endomorphism (which we always assume continuous). The contraction
group of α is
n
o
con(α) := x ∈ G : lim αn (x) = e .
n→∞
Given a closed subgroup H ⊆ G with α(H) ⊆ H, we say that a sequence
(xn )n∈N0 in G converges to e modulo H (and we write xn → e modulo H) if, for
each identity neighbourhood V ⊆ G, there exists N ∈ N0 such that xn ∈ V H
for all n ≥ N (cf. [1], [13]). The domain of attraction of H with respect to α is
defined as the set con(α, H) of all x ∈ G such that αn (x) → e modulo H. The
following result extends the partial results [1, Theorem 3.8], [13, Theorem 1],
[4, Theorem 2.4] and [6].
Theorem A. Let α : G → G be an endomorphism of a totally disconnected,
locally compact group G and H ⊆ G be a closed subgroup. If α(H) = H or
α(H) ⊆ H and H is compact, then con(α, H) = con(α)H.
Given x ∈ G, a sequence (xn )n∈N0 in G is called an α-regressive trajectory for x
if x0 = x and α(xn ) = xn−1 for all n ∈ N. Let con− (α) be the set of all x ∈ G
for which there exists an α-regressive trajectory (xn )n∈N0 with
lim xn = e.
n→∞
Then con− (α) is a subgroup of G, called the anti-contraction group of α. If H
is a closed subgroup of G with α(H) ⊆ H, then the anti-contraction set of α
modulo H is defined as the set con− (α, H) of all x ∈ G for which there exists
an α-regressive trajectory (xn )n∈N0 with xn → e modulo H.
Theorem B. Let α : G → G be an endomorphism of a totally disconnected,
locally compact group G and let H ⊆ G be a closed subgroup with α(H) ⊆ H. Then
con− (α, H) = con− (α)H.
For each α-regressive trajectory (xn )n such that xn → e modulo H, there is an
α-regressive trajectory (yn )n such that yn ∈ xn H for all n and limn→∞ yn = e.
Let G be a totally disconnected, locally compact group and α : G → G be an
ENDOMORPHISMS OF T.D.L.C. GROUPS
3
endomorphism. For a compact open subgroup U ⊆ G, the index
[α(U ) : α(U ) ∩ U ] ∈ N
is called the displacement index of U . If U has minimal displacement index
among all compact open subgroups, then U is called minimizing, and one defines
the scale of α as this minimum displacement index, denoted s(α). Equivalently,
U has certain structural properties, summarized as being tidy Let par− (α) be
the anti-parabolic subgroup of α consisting of all x ∈ G for which there exists an
α-regressive trajectory (xn )n∈N0 such that {xn : n ∈ N0 } is relatively compact.
The article [23] analyzes how the scale of an automorphism α behaves with
respect to taking subgroups and quotients. We generalize these results to the
case of endomorphisms. Given a group G, a normal subgroup H of G and
an endomorphism α of G with α(H) ⊆ H, we let α denote the associated
endomorphism of G/H.
Theorem C. Let α : G → G be an endomorphism of a totally disconnected,
locally compact group G and H ⊆ G a closed subgroup with α(H) ⊆ H.
(a) There is a compact open subgroup U of G which is tidy for α and such
that U ∩ H is tidy for α|H . Furthermore
sH (α|H ) ≤ sG (α).
(b) If, in addition, H is normal in G, then
sH (α|H )sG/H (α) divides sG (α).
(c) If, in addition to all the above, H ⊆ par− (α) is a closed normal subgroup
of G and α(H) = H, then
sH (α|H )sG/H (α) = sG (α).
Using Theorem A, we show the following result which generalizes [1, Theorem 3.32] and the corresponding statement in [9].
Theorem D. Let α : G → G be an endomorphism of a totally disconnected,
locally compact group G. Then α has small tidy subgroups if and only if con(α)
is closed in G.
4
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Using Theorem B, we prove a generalization of [1, Proposition 3.21 (3)]:
Theorem E. Let α : G → G be an endomorphism of a totally disconnected,
locally compact group G. Then
(1)
s(α) = s α|con− (α) .
Further consequences of Theorems A and B are obtained. Let par(α) ⊆ G
be the parabolic subgroup associated to an endomorphism α, i.e., the set of all
x ∈ G for which {αn (x) : n ∈ N0 } is relatively compact in G. Let
lev(α) := par(α) ∩ par− (α)
be the Levi subgroup. If G is locally compact and totally disconnected, then
par(α) and lev(α) are closed subgroups of G such that α(par(α)) ⊆ par(α) and
α(lev(α)) = lev(α) (see [26, Proposition 19]). Given a compact open subgroup
T
V ⊆ G, we define V− := n∈N0 α−n (V ) and write V+ for the set of all x ∈ V
for which there exists an α-regressive trajectory (xn )n∈N0 within V .
We study par(α), lev(α) and their connections to tidy subgroups. If α has small
tidy subgroups, then the strongest conclusions can be obtained, as compiled in
the next theorem. For α an automorphism, cf. already [6] for part (a) and [1,
Theorem 3.32] for (b). The semidirect products mean semidirect products of
topological groups.
Theorem F. Let α be an endomorphism of a totally disconnected, locally compact group G. Then α has small tidy subgroups if and only if α|lev(α) has small
tidy subgroups. In this case, the following holds:
(a) Ω := con(α) lev(α) con− (α) is an open identity neighbourhood in G such
that α(Ω) ⊆ Ω, and the product map
(2)
π : con(α) × lev(α) × con− (α) → Ω, (x, y, z) 7→ xyz
is a homeomorphism.
(b) par(α) = con(α) ⋊ lev(α) and par− (α) = con− (α) ⋊ lev(α).
(c) α|par− (α) , α|lev(α) and α|con− (α) are automorphisms.
(d) Every subgroup V of G which is tidy for α is a subset of Ω; it satisfies
V− = (con(α) ∩ V ) ⋊ (lev(α) ∩ V )
and V+ = (con− (α) ∩ V ) ⋊ (lev(α) ∩ V ).
ENDOMORPHISMS OF T.D.L.C. GROUPS
5
(e) A compact open subgroup V ⊆ G is tidy for α if and only if
(3)
V = (con(α) ∩ V )(lev(α) ∩ V )(con− (α) ∩ V ),
(4)
α(con(α) ∩ V )
⊆ con(α) ∩ V,
(5)
α(lev(α) ∩ V )
= lev(α) ∩ V, and
(6)
−
α(con (α) ∩ V ) ⊇ con− (α) ∩ V.
(f) The compact open subgroups W ⊆ lev(α) with α(W ) = W form a basis
of identity neigbourhoods in lev(α).
If G is a Lie group over a totally disconnected local field (as in [3] and [18])
and α : G → G is an analytic endomorphism with small tidy subgroups, then
con(α), lev(α) and con− (α) are Lie subgroups of G (in the strong sense of submanifolds) and the product map in (2) is an analytic diffeomorphism, see [9].
By (e) in Theorem F, the automorphism α|lev(α) is distal, see [14]. Information
concerning contractive automorphisms of locally compact groups can be found
in [19] and [10]; contractive analytic automorphisms of Lie groups over a totally
disconnected local field K are discussed in [21] (for K = Qp ) and [8].
Independenly, studies of endomorphisms of totally disconnected groups have
also been performed in [15] and [5]. Whereas the former deals with profinite
groups, the latter has a view towards topological entropy. According to [5,
Corollary 4.11] (combined with Proposition 9.5), the topological entropy htop (α)
of an endomorphism α of a totally disconnected, locally compact group G is
given by htop (α) = ln s(α), if con(α) is closed. Our Theorem C (c) therefore
implies the following Addition Theorem for topological entropy, which is complementary to a similar, known result ([5, Theorem 1.2]):
If con(α) is closed in G and H is a closed normal subgroup of G such that
H ⊆ par− (G) and α(H) = H, then htop (α) = htop (α|H ) + htop (α).
Acknowledgements. The research was carried out during the ‘Winter of Disconnectedness’ hosted by the University of Melbourne (Matrix Center, Creswick)
and the University of Newcastle (NSW). We owe thanks to both institutions for
their hospitality and are grateful to George Willis for stimulating questions,
discussions, and support. The first and last authors would like to extend their
gratitude to Colin Reid for helpful discussions, as well as their supervisors,
Jacqui Ramagge and Marc Burger respectively, for their ongoing support and
6
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
making attendance to the aforementioned workshop possible. This research
was conducted while the first author was supported by the Australian Government Research Training Program Scholarship and, partly, while the last author
was supported by the SNSF Doc.Mobility fellowship 172120. Finally, we owe
thanks to an anonymous referee whose comments helped us to improve the original manuscript.
Conventions. Write N := {1, 2, . . .} and N0 := N ∪ {0}. All topological groups
and uniform spaces are assumed Hausdorff. The automorphisms considered are
continuous, with continuous inverse. A map q : X → Y between topological
spaces is called a quotient map if it is surjective and Y carries the quotient
topology (the final topology with respect to q). If X is a set and α : X → X a
map, we say that a subset M ⊆ X is α-invariant if α(M ) ⊆ M . If α(M ) = M ,
we say that M is α-stable. This terminology is in line with [1] but differs from
[13]. A sequence (xn )n∈N0 (or two-sided sequence (xn )n∈Z ) in a topological
space X is called bounded if {xn : n ∈ N0 } (resp., {xn : n ∈ Z}) is relatively
compact in X. If G is a totally disconnected, locally compact group, then
COS(G) denotes the set of all compact open subgroups of G. We write |X| for
the cardinality of a set X. If k is an ordinal number, we write [0, k[ for the set
of ordinals j < k. For each cardinal number c, the set of ordinal numbers with
cardinality c has a minimal element k. Thus [0, k[ has cardinality c, while [0, j[
has cardinality less than c for each j ∈ [0, k[. We identify c with k.
Structure of the Article. In a preparatory first section, we recall necessary
concepts from [26]. Auxiliary results concerning convergence modulo H and
generalizations of the concept for self-maps of uniform spaces are compiled in
Section 2. Section 3 is devoted to the proof of Theorem A for metrizable G,
by an adaptation of arguments from [1]. In Section 4, we prove Theorem B for
metrizable G. Also this proof was inspired by [1], but major changes were necessary to enable (unique) α-orbits to be replaced with (not necessarily unique)
α-regressive trajectories. In both cases, the generalization to non-metrizable G
uses a method that was stimulated by Jaworski’s arguments in [13], but deviates
in detail as we use a different cardinal as the parameter in a transfinite induction. In this way, we can do without some of the preparatory lemmas in [13],
which do not have analogues for anti-contraction groups. See Section 6 for the
case of Theorem A. The relevant cardinal invariant is introduced in Section 5.
As to Theorem B, we have to line up regressive trajectories by a transfinite
ENDOMORPHISMS OF T.D.L.C. GROUPS
7
induction over the steps of a projective system, rather than mere elements of
anti-contraction groups (Section 7). This also enables us to get around the problem that although con(α, K) ∩ con(α, L) = con(α, K ∩ L) for compact K and L,
a corresponding formula cannot be expected for anti-contraction groups of an
endomorphism. Section 8 analyzes how the scale of an endomorphism behaves
with respect to subgroups and quotients and proves most of Theorem C. In
Sections 9–13, we give applications of Theorems A and B concerning endomorphisms, their scale and the various subgroups going along with them. Notably,
we complete the proof of Theorem C and obtain proofs for Theorems D, E,
and F.
1. Basic concepts
In this section, we collect basic definitions and properties of the objects appearing in this article. For details, we refer the reader to [26]. Let G be a
totally disconnected, locally compact group and α : G → G be an endomorphism. Recall that by a theorem of van Dantzig (see [20] or [11, (7.7)]), G
admits an identity neighbourhood basis consisting of compact open subgroups.
For a compact open subgroup U ⊆ G, the index
[α(U ) : α(U ) ∩ U ] ∈ N
is called the displacement index of U . If U has minimal displacement index
among all compact open subgroups, then U is called minimizing, and one defines
the scale of α as this minimum displacement index, denoted s(α) or sG (α).
Equivalently, U has certain structural properties, summarized as being tidy.
Write
\
U− :=
α−n (U ) = {x ∈ U : (∀n ∈ N0 ) αn (x) ∈ U },
n∈N0
where α−n (U ) means the preimage (αn )−1 (U ). Let U+ be the set of all x ∈ U
for which there exists an α-regressive trajectory (xn )n∈N0 such that xn ∈ U for
all n. Then
\
U+ =
Un,α with
n∈N0
(7)
U0,α := U and Un+1,α := U ∩ α(Un,α ) for n ∈ N0 ;
moreover, U+ and U− are compact subgroups of G such that
α(U− ) ⊆ U−
and α(U+ ) ⊇ U+
8
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
(see [26], or also Lemma 1.2). The sets
[
[
U−− :=
α−n (U− ) and U++ :=
αn (U+ )
n∈N0
n∈N0
are unions of ascending sequences of subgoups, whence they are subgroups of G.
If U = U+ U− , then U is called tidy above for α. If U++ is closed in G and
[αn+1 (U+ ) : αn (U+ )] ∈ N
is independent of n ∈ N0 , then U is called tidy below. If U is both tidy above
and tidy below for α, then U is called tidy for α (see [26]).
We shall use the fact that U is tidy if and only if U is tidy above and U−− is
closed in G (see [26, Proposition 9]).
By [26, Lemma 5 and Theorem 2], a compact open subgroup U of G is minimizing for α if and only if it is tidy for α, in which case
s(α) = [α(U ) : α(U ) ∩ U ] = [α(U+ ) : U+ ].
If U is a compact open subgroup of G, then the compact open subgroup
ℓ
\
α−j (U )
j=0
is tidy above for large ℓ ∈ N0 (see [26]). For U ⊆ G a compact open subgroup,
U+ ∩ U− is a compact subgroup of G such that α(U+ ∩ U− ) = U+ ∩ U− (see [26]
or Lemma 1.2). Let par− (α) be the anti-parabolic subgroup of α consisting of all
x ∈ G for which there is an α-regressive trajectory (xn )n∈N0 with {xn : n ∈ N0 }
being relatively compact. Then par− (α) is a closed subgroup of G and
α(par− (α)) = par− (α)
(see [26, comment after Proposition 19]). The intersection nub(α) of all tidy
subgroups is called the nub of α. The nub is compact, and α-stable in the sense
that α(nub(α)) = nub(α) (see [26, Proposition 20]). The nub coincides with the
trivial group {e} if and only if α has small tidy subgroups in the sense that the
subgroups which are tidy for α form a basis of identity neighbourhoods. The
bounded iterated kernel bik(α) is defined as the closure
bik(α) := {x ∈ par− (α) : αn (x) = e for some n ∈ N0 }
(as opposed to [26, Definition 12] which is neither intended nor used); it is
an α-stable compact normal subgroup of par− (α). The endomorphism α of
ENDOMORPHISMS OF T.D.L.C. GROUPS
9
par− (α)/ bik(α) induced by α is an automorphism and bik(α) ⊆ nub(α) [26,
Proposition 20]. Then nub(α)/bik(α) = nub(α) (see Proposition 12.2 (b)) and
nub(α) is the largest α-stable closed subgroup of par− (α)/ bik(α) on which α
acts ergodically (see [25, Proposition 4.4]).
For α an endomorphism of a totally disconnected locally compact group, the
following sets were already defined above.
Definition 1.1: If X is a Hausdorff topological space, K ⊆ X a compact
subset and α : X → X a continuous map, we define
\
\
Kn,α ,
α−n (K)
and
K+ :=
K− :=
n∈N0
n∈N0
where K0,α := K and Kn,α := K ∩ α(Kn−1,α ) for n ∈ N.
The following properties are useful in general. Compare [26] for proofs.
Lemma 1.2: For X, α, K, Kn,α , K+ , K− as in Definition 1.1, we have:
(a) K− is a compact subset of K. It is the set of all x ∈ K such that
αn (x) ∈ K for all n ∈ N0 . Moreover, α(K− ) ⊆ K− .
(b) Kn,α is a compact subset of K, for each n ∈ N0 . It is the set of all
x ∈ K for which there exist x0 , x1 , . . . , xn ∈ K such that x0 = x and
α(xj ) = xj−1 for all j ∈ {1, 2, . . . , n}.
(c) K+ is a compact subset of K. It is the set of all x ∈ K for which there
exists an α-regressive trajectory (xn )n∈N0 with x0 = x and xn ∈ K for
all n ∈ N0 . Moreover, K+ ⊆ α(K+ ).
(d) K+ ∩ K− is a compact subset of K. It is the set of all x ∈ K for which
there exists a family (xn )n∈Z of elements xn ∈ K such that x0 = x and
α(xn ) = xn+1 for all n ∈ Z. Moreover, α(K+ ∩ K− ) = K+ ∩ K− .
If X is a topological group, α an endomorphism and K a compact subgroup
of X, then also K+ , K− and each Kn,α are compact subgroups of X.
2. Dynamics on uniform spaces
In this section, we collect generalities about dynamical systems on uniform
spaces. As a reference for the theory of uniform spaces, we recommend [12]. A
simple topological fact will be used (whose proof is left to the reader).
10
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Lemma 2.1: Let X be a topological space and N a filter basis in X. If each
T
K ∈ N is compact and U ⊆ X is an open subset with N ⊆ U , then there
exists some K ∈ N such that K ⊆ U .
2.2: Let X be a set, α : X → X be a map and x ∈ X. As usual, the sequence
(αn (x))n∈N0 is called the α-orbit of x. A sequence (xn )n∈N0 is called an αregressive trajectory for x if x0 = x and α(xn ) = xn−1 for all n ∈ N. Note
the slightly different usage in [26]. A two-sided sequence (xn )n∈Z is called an
α-trajectory (for x0 ) if α(xn ) = xn+1 for all n ∈ Z.
The following observation allows α-regressive trajectories to be created from
a given sequence.
Lemma 2.3: Let X be a Hausdorff topological space, α : X → X be a continuous map and (zn )n∈N0 be a sequence in X. Let x0 ∈ X and, for n, m ∈ N0 ,
(
αm−n (zm ) if n ∈ {0, . . . , m};
(m)
(8)
zn :=
x0
if n > m.
(m)
Thus z (m) := (zn )n∈N0 ∈ X N0 for each m ∈ N0 . If y = (yn )n∈N0 is an
accumulation point of (z (m) )m∈N0 in X N0 with respect to the product topology,
then (yn )n∈N0 is an α-regressive trajectory.
Proof. Let (z (ma ) )a∈A be a convergent subnet of (z (m) )m∈N0 with limit y =
(yn )n∈N0 . For every n ∈ N, there is a0 ∈ A with ma ≥ n for all a ≥ a0 and thus
(m )
α(zn(ma ) ) = zn−1a .
Since α is continuous, passing to the limit we obtain that α(yn ) = yn−1 . Hence
(yn )n∈N0 is an α-regressive trajectory.
(m)
For fixed n ∈ N0 , we have zn = αm−n (zm ) for all m ≥ n. Therefore,
accumulation points of (z (m) )m∈N0 are independent of the choice of x0 .
(m)
Definition 2.4: We call (z (m) )m∈N0 = ((zn )n∈N0 )m∈N0 (as in (8)) a double
sequence associated to (zn )n∈N0 with respect to α. If X is a topological group,
we shall always choose x0 := e.
2.5: Let (X, E) be a uniform space, with filter E of entourages. For V ∈ E, x ∈
S
X and M ⊆ X, we write V [x] := {y ∈ X : (x, y) ∈ V } and V [M ] := z∈M V [z].
ENDOMORPHISMS OF T.D.L.C. GROUPS
11
2.6: Let (X, E) be a uniform space and S ⊆ X. Given a subset S ⊆ X, we
say that a sequence (xn )n∈N0 in X converges to S with respect to E (and write
xn →E S) if, for each V ∈ E, there exists N ∈ N0 such that xn ∈ V [S] for all
n ≥ N . Further, let α : X → X be a continuous map. The domain of attraction
of S with respect to α is defined as
con(α, S) := x ∈ X : αn (x) →E S .
Given a subset S ⊆ X, we write con− (α, S) for the set of all x ∈ X for which
there exists an α-regressive trajectory (xn )n∈N0 such that xn →E S.
2.7: Recall that the right uniform structure R on a topological group G is the
filter generated by the filter basis of the sets
Ve := {(x, y) ∈ G × G : y ∈ V x},
for V ranging through the set of identity neighbourhoods in G (see [11, (4.11)]).
Then Ve [x] = V x and Ve [M ] = V M for subsets M ⊆ G. Thus, if H ⊆ G is a
closed subgroup and (xn )n∈N0 is a sequence in G, then
xn → e modulo H if and only if xn →R H.
If α is an endomorphism of G with α(H) ⊆ H, we deduce that the sets con(α, H)
and con− (α, H) coincide with those obtained for the uniform space (G, R).
2.8: Recall that a subset R of a uniform space (X, E) is called precompact if, for
each V ∈ E, there exist n ∈ N0 and x1 , . . . , xn in X (or, equivalently, in R) with
R ⊆ V [x1 ] ∪ · · · ∪ V [xn ].
Every compact subset K ⊆ X is precompact. If (X, E) is complete (e.g., in the
case of (G, R) with G a locally compact group, see [2, §3.3, Corollary 1]), then
a subset R ⊆ G is precompact if and only if it is relatively compact (cf. [12,
p. 22]).
The following variant of [1, Lemma 3.9] (devoted to locally compact groups)
is very useful for our purposes; [13] already applied the latter lemma also for
Hausdorff topological groups (without proof).
Lemma 2.9: Let (xn )n∈N0 be a sequence in a uniform space (X, E).
(a) If xn →E K for some precompact subset K ⊆ X, then {xn : n ∈ N0 } =:
R is precompact in X. If xn →E K and K is compact, then R is
relatively compact and thus (xn )n∈N0 is bounded.
12
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
(b) If xn →E S for a closed subset S ⊆ X, then each accumulation point of
(xn )n∈N0 is contained in S.
(c) If (xn )n∈N0 is bounded and S ⊆ X is closed, then xn →E S if and only
if each accumulation point of (xn )n∈N0 lies in S.
Proof. (a) If V ∈ E, choose W ∈ E with W ◦ W ⊆ V . Since K is precompact,
there is a finite subset Φ ⊆ K such that K ⊆ W [Φ]. Hence W [K] ⊆ W [W [Φ]] =
(W ◦ W )[Φ] ⊆ V [Φ]. There exists N ∈ N0 such that xn ∈ W [K] for all n ≥ N .
Hence R := {xn : n ∈ N0 } is contained in the finite union
V [Φ] ∪
N[
−1
V [xn ],
n=0
and thus R is precompact.
If, moreover, K is compact, let us show that the closure R is complete in the
uniform structure induced by (X, E). Then R, being a uniform space which
is both precompact and complete, will be compact (see, e.g., [12, p. 22]). Let
y ∈ R such that y 6∈ R; we show that y ∈ K. Let (yj )j∈J be a net in R which
converges to y. Given U ∈ E, pick V ∈ E with V ◦ V ⊆ U . There is N ∈ N with
(9)
xn ∈ V [K] for all n ≥ N .
Since none of x0 , . . . , xN −1 is an accumulation point of (yj )j∈J , we find jV ∈ J
such that yj 6∈ {x0 , . . . , xN −1 } for all j ≥ jV and thus
(10)
yj ∈ V [K] for all j ≥ jV .
Since yj → y, we deduce that y ∈ V [K] ⊆ V [V [K]] ⊆ U [K]. As
K = K, we see that y ∈ K.
T
U∈E
U [K] =
If R was not complete, we could find a Cauchy net (yj )j∈J in R which does not
converge in R, and which therefore does not have a convergent subnet (nor an
accumulation point). Make E × J a directed set by declaring (V1 , j1 ) ≤ (V2 , j2 )
if V1 ⊇ V2 and j1 ≤ j2 . As before, for U ∈ E and V ∈ E with V ◦ V ⊆ U , we
can pick N ∈ N with (9) and find jV ∈ J such that (10) holds. Hence, with the
induced order,
A := {(V, j) ∈ E × J : yj ∈ V [K]}
ENDOMORPHISMS OF T.D.L.C. GROUPS
13
is a directed set, and (yj )(V,j)∈A is a subnet of (yj )j∈J . For a = (V, j) ∈ A, we
find ka ∈ K such that
yj ∈ V [ka ].
(11)
Since K is compact, the net (ka )a∈A has a subnet (ka(i) )i∈I which converges to
some k ∈ K. Write a(i) = (V (i), j(i)). Given an entourage U ∈ E, pick V ∈ E
such that V ◦ V ⊆ U . There exists i0 ∈ I such that
ka(i) ∈ V [k]
and V (i) ⊆ V for all i ≥ i0 in I. Hence, using (11),
yj(i) ∈ V (i)[ka(i) ] ⊆ V [ka(i) ] ⊆ V [V [k]] ⊆ U [k]
for all i ≥ i0 . The subnet (yj(i) )i∈I of (yj )j∈J therefore converges to k, contrary
to our assumption. Thus (yj )j∈J cannot exist and R must be complete.
(b) Assume that a subnet (xn(j) )j∈J of the given sequence converges to x ∈ S.
For all U ∈ E and V ∈ E with V ◦ V ⊆ U , we find j0 ∈ J such that xn(j) ∈ V [x]
for all j ≥ j0 and thus
x ∈ V [x] ⊆ (V ◦ V )[x] ⊆ U [x] ⊆ U [S].
Hence x ∈
T
U∈E
U [S] = S = S.
(c) If not xn →E S, then there exists V ∈ E such that J := {n ∈ N0 : xn 6∈ V [S]}
is an infinite set. After shrinking V , we may assume that V [S] is open. Now
P := {xn : n ∈ J} has compact closure L in X. Since P is a subset of the
closed set X \ V [S], also L ⊆ X \ V [S]. Now L being compact, the subnet
(xn )n∈J of (xn )n∈N0 has a subnet which converges to some c ∈ L. Then c is an
accumulation point of (xn )n∈N0 such that c 6∈ V [S] and hence c 6∈ S. Together
with (b), this establishes (c).
Lemma 2.10: Let (X, E) be a uniform space and (Aj )j∈J be a family of closed
subsets of X, with J 6= ∅. Let (xn )n∈N0 be a sequence in X such that xn →E Aj
for each j ∈ J. If Aj0 is compact for some j0 ∈ J, then
\
(12)
xn →E
Aj .
j∈J
Proof. As Aj0 is compact and xn →E Aj0 , the sequence (xn )n∈N0 is bounded,
by Lemma 2.9 (a). If c ∈ X is an accumulation point of (xn )n∈N0 , then c ∈ Aj
14
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
for each j ∈ J, by Lemma 2.9 (b), and thus c ∈
by Lemma 2.9 (c).
T
j∈J
Aj =: A. Hence xn →E A,
Lemma 2.11: Let (X, E) be a uniform space, α : X → X be a continuous map
and (Aj )j∈J be a family of closed subsets of X, with J 6= ∅. If Aj0 is compact
for some j0 ∈ J, then
\
\
(13)
con(α, Aj ) = con α,
Aj .
j∈J
j∈J
T
Proof. If x ∈ j∈J con(α, Aj ), then αn (x) →E Aj for each j ∈ J, whence
T
αn (x) →E j∈J Aj =: A (by Lemma 2.10) and thus x ∈ con(α, A). Thus
T
j∈J con(α, Aj ) ⊆ con(α, A). The converse inclusion trivially holds.
Lemma 2.12: Let (X, E) be a uniform space, α : X → X be a continuous map
and K ⊆ X be a compact subset. Then
con(α, K) = con(α, K+ ∩ K− ) and
con− (α, K) = con− (α, K+ ∩ K− ).
Moreover, every α-regressive trajectory which converges to K in (X, E) also
converges to K+ ∩ K− .
Proof. To prove the first assertion, we show con(α, K) ⊆ con(α, K+ ∩ K− ) (as
the other inclusion is trivial). If x ∈ con(α, K), then αn (x) → e modulo K,
whence R := {αn (x) : n ∈ N0 } is relatively compact in X (see Lemma 2.9 (a)).
Thus, by Lemma 2.9 (c), we shall have x ∈ con(α, K+ ∩ K− ) if we can show
that each accumulation point c of (αn (x))n∈N0 is an element of K+ ∩ K− .
By Lemma 2.9 (b), we know that each accumulation point of (αn (x))n∈N0 is
an element of K. Let (αna (x))a∈A be a subnet which converges to c. Given
m ∈ N0 , the subnet
(αna +m (x))a∈A = (αm (αna (x)))a∈A
of (αn (x))n∈N0 converges to αm (c), by continuity of αm . Thus αm (c) is an accuT
mulation point of (αn (x))n∈N0 and thus αm (c) ∈ K. Hence c ∈ m∈N0 α−m (K) =
K− . To see that also c ∈ K+ (which shows our first claims), let us show that
c ∈ Km,α for each m ∈ N0 . There is am ∈ A such that na ≥ m for all a ≥ am .
Then (αna −m (x))a≥am is a subnet of (αn (x))n∈N0 and hence has a subnet converging to some b ∈ K. Applying αj to the entries of the latter subnet (for given
j ∈ {0, . . . , m}), we obtain a convergent net with limit αj (b) =: xm−j , which
ENDOMORPHISMS OF T.D.L.C. GROUPS
15
has to be an element of K as the convergent net is a subnet of (αn (x))n∈N0 .
For j = m, the convergent net just constructed is a subnet of (αna (x))a≥am ;
its limit x0 = αm (b) therefore has to coincide with c. Thus x0 , x1 , . . . , xm ∈ K,
x0 = c and
α(xj ) = α(αm−j (b)) = αm−(j−1) (b) = xj−1
for all j ∈ {1, . . . , m}, whence c ∈ Km,α (by Lemma 1.2 (b)).
To complete the proof, it suffices to prove the final statement as it entails
that con− (α, K) ⊆ con− (α, K+ ∩ K− ) (and the other inclusion trivially holds).
Thus, let (xn )n∈N0 be an α-regressive trajectory in G such that xn → e modulo K. By Lemma 2.9 (a), the closure L := {xn : n ∈ N0 } ⊆ X is compact.
Hence, by Lemma 2.9 (c), (xn )n∈N0 will converge to e modulo K+ ∩ K− if
we can show that each accumulation point c of (xn )n∈N0 is an element of
K+ ∩ K− . By Lemma 2.9 (b), we have c ∈ K. Now α(c) is an accumulation point of (α(xn ))n∈N0 , by continuity of α, and hence also an accumulation
point of (α(xn+1 ))n∈N0 = (xn )n∈N0 . Thus, the set C of accumulation points of
(xn )n∈N0 is α-invariant, i.e., α(C) ⊆ C. Hence C ⊆ K− .
For c as before and m ∈ N, let (xnβ )β∈B be a subnet of (xn )n∈N0 which converges
to c. Then tβ := (xnβ , xnβ +1 , . . . , xnβ +m )β∈B is a net in the compact topological space Lm+1 and hence has a convergent subnet, with limit (c0 , c1 , . . . , cm ) ∈
Lm+1 , say. Then c0 = c and α(cj ) = cj−1 for all j ∈ {1, . . . , m} (as the entries
of each tβ have this property and α is continuous). Thus c ∈ Km,α for each
m ∈ N0 and thus
\
c∈
Km,α = K+ .
m∈N0
Hence c ∈ K+ ∩ K− , as required.
We close the section with some straightforward observations, for later use.
Lemma 2.13: Let f : X → Y be a uniformly continuous map between uniform
spaces (X, E) and (Y, F ) Let S ⊆ X be a subset.
(a) If (xn )n∈N0 is a sequence in X such that xn →E S, then f (xn ) →F f (S).
(b) If α : X → X and β : Y → Y are continuous maps such that β◦f = f ◦α,
then f (con(α, S)) ⊆ con(β, f (S)).
Proof. (a) If U ∈ F , then V := (f × f )−1 (U ) ∈ E, whence there exists N ∈ N0
such that xn ∈ V [S] for all n ≥ N . Thus, for such n, there is sn ∈ S such
16
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
that xn ∈ V [sn ] and thus (sn , xn ) ∈ V . Hence (f (sn ), f (xn )) ∈ U and thus
f (xn ) ∈ U [f (sn )] ⊆ U [f (S)]. So f (xn ) →F f (S).
(b) For x ∈ con(α, S), we have β n (f (x)) = f (αn (x)) →F f (S), by (a).
The next lemma is obvious.
Lemma 2.14: Let (X, E) be a uniform space, S ⊆ Y ⊆ X be subsets and
(xn )n∈N0 be a sequence in Y such that xn →E S. Let F be the uniform structure
induced by E on Y . Then also xn →F S.
Lemma 2.15: Let G be a topological group, S ⊆ G be a subgroup and K ⊆ G.
(a) If K ⊆ S and S is α-invariant for an endomorphism α of G, then
con(α, K) ∩ S = con(α|S , K).
(b) If S is open and (xn )n∈N0 is a sequence in S such that xn →R K in G,
then xn →R K ∩ S in G (and in S).
Proof. (a) Using the inclusion map f : S → G and β := α|S , Lemma 2.13
yields con(α|S , K) ⊆ con(α, K) ∩ S. If, conversely, x ∈ con(α, K) ∩ S, then
αn (x) →R K in G and hence in S (by Lemma 2.14), whence x ∈ con(α|S , K).
(b) If V ⊆ G is an identity neighbourhood such that V ⊆ S, then xn ∈ V K if
and only if xn ∈ V (K ∩ S). Hence xn →R K in G if and only if xn →R (K ∩ S)
in G. By Lemma 2.14, then also xn →R (K ∩ S) in S.
3. Proof of Theorem A for metrizable G
In this section we prove Theorem A in the metrizable case. We include two
lemmas the first one of which concerns how tidiness above passes to subgroups.
Lemma 3.1: Let α be an endomorphism of a totally disconnected, locally compact group G. Let H ⊆ G be an α-invariant closed subgroup and W ⊆ G be a
compact open subgroup. Then there exists ℓ0 ∈ N0 such that
ℓ
\
j=0
α−j (W ) ∩ H
is tidy above for α|H , for all ℓ ≥ ℓ0 .
ENDOMORPHISMS OF T.D.L.C. GROUPS
17
Proof. Since α(H) ⊆ H, we have
H∩
ℓ
\
α−j (W )
= {w ∈ H : (∀j = 0, . . . , ℓ) αj (w) ∈ W }
j=0
= {w ∈ H : (∀j = 0, . . . , ℓ) αj (w) ∈ H ∩ W }
(14)
=
ℓ
\
(α|H )−j (H ∩ W ).
j=0
The group in (14) is tidy above for α|H by [26, Proposition 3] for large ℓ.
To prove Theorem A, we shall use the following lemma (which parallels the
case of automorphisms formulated in [1, Lemma 3.10]):
Lemma 3.2: Let α be an endomorphism of a totally disconnected, locally compact group G and H ⊆ G be a closed subgroup. Assume that
(i) α(H) = H; or
(ii) H is compact.
Let x ∈ con(α, H) and O ⊆ G be a compact open subgroup. Then there exists
h ∈ H and N ∈ N0 such that αn (xh) ∈ O for all n ≥ N and αn (xh) → e
modulo (O ∩ H)+ ∩ (O ∩ H)− (which is a compact subgroup of O ∩ H).
Proof. If H is compact, then con(α, H) = con(α, H+ ∩ H− ), where H+ ∩ H− is
compact and α(H+ ∩ H− ) = H+ ∩ H− (see Lemma 1.2 (d) and Lemma 2.12).
After replacing H with H+ ∩ H− , we may therefore assume that α(H) = H.
Thus, it suffices to discuss the case (i).
Assume α(H) = H now. After shrinking O if necessary, we may assume that
O ∩ H = (O ∩ H)− (O ∩ H)+ ,
by Lemma 3.1. Let V ⊆ O be a compact open subgroup such that α(V ) ⊆ O.
Since αn (x) → e modulo H, there exists N ∈ N0 such that αn (x) ∈ V H for all
n ≥ N . As αN (H) = H, we find h0 ∈ H with αN (xh0 ) ∈ V . We now complete
h0 to a sequence (hi )i∈N0 of elements in h0 (O ∩ H) such that
αN +j (xhi ) ∈ V (O ∩ H) for 0 ≤ j ≤ i.
Suppose that h0 , . . . , hk have already been constructed to satisfy this condition.
Using that α(O ∩ H) ⊆ (O ∩ H)α((O ∩ H)+ ), we obtain
αN +k+1 (xhk ) ∈ α(V (O ∩ H)) ⊆ Oα(O ∩ H) ⊆ Oα((O ∩ H)+ ).
18
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Choose ℓk+1 ∈ α((O ∩ H)+ ) such that αN +k+1 (xhk )ℓk+1 ∈ O. There exists an
α-regressive trajectory (gn )n∈N0 for ℓk+1 such that gn ∈ O ∩ H for all n ≥ 1.
We define
hk+1 := hk gN +k+1 .
Then hk+1 ∈ h0 (O∩H) and αN +k+1 (xhk+1 ) ∈ O. Moreover, αN +k+1 (xhk+1 ) =
αN +k+1 (x)αN +k+1 (hk+1 ) ∈ V H and thus
αN +k+1 (xhk+1 ) ∈ O ∩ V H = V (O ∩ H).
For all j ∈ {0, . . . , k}, we have
αN +j (xhk+1 ) = αN +j (xhk )αN +j (gN +k+1 ) = αN +j (xhk )gk−j+1 ∈ V (O ∩ H)
as well, whence hk+1 is as desired. This completes the recursive construction.
Let h be an accumulation point of the sequence (hi )i∈N0 in h0 (O ∩ H) ⊆ H.
Then αn (xh) is in the compact set V (O ∩ H) for all n ≥ N , and so αn (xh) ∈
O. Thus xh ∈ con(α, O). As also xh ∈ con(α, H), Lemma 2.11 shows that
xh ∈ con(α, O ∩ H). Hence αn (xh) → e modulo (O ∩ H)+ ∩ (O ∩ H)− , by
Lemma 2.12.
3.3: For the proofs of Theorem A and Theorem B we recall that a topological
group is metrizable if and only if it is Hausdorff and first-countable, see for
instance [11, (8.3)].
Proof. (Theorem A, for metrizable G). The inclusion con(α)H ⊆ con(α, H)
is trivial. To establish the converse inclusion, we adapt the proof of [1, Theorem 3.8]: Since G is assumed metrizable, using 3.3 we find a sequence O1 ⊇
O2 ⊇ · · · of compact open subgroups of G which form a basis of identity neighbourhoods. Set H0 := H and
Hj := (Oj ∩ Hj−1 )+ ∩ (Oj ∩ Hj−1 )− ⊆ Oj ∩ Hj−1 for j ∈ N.
Let x0 ∈ con(α, H). We claim that, for j ∈ N, there exist elements xj ∈
xj−1 Hj−1 and integers Nj ∈ N0 such that
αn (xj ) → e modulo Hj
and αn (xj ) ∈ Oj for all n ≥ Nj .
Lemma 3.2 (with x0 in place of x and O1 in place of O) provides h0 ∈ H0 = H
and N1 ∈ N0 such that x1 := x0 h0 satisfies αn (x1 ) ∈ O1 for all n ≥ N1 and
αn (x1 ) → e modulo (O1 ∩ H)+ ∩ (O1 ∩ H)− = H1 . This shows the case j = 1.
ENDOMORPHISMS OF T.D.L.C. GROUPS
19
If x1 , . . . , xj have been constructed, we can apply Lemma 3.2 with xj in place
of x, Oj+1 in place of O and Hj in place of H. We obtain an hj ∈ Hj and
Nj+1 ∈ N0 such that xj+1 := xj hj satisfies αn (xj+1 ) ∈ Oj+1 for all n ≥ Nj+1
and αn (xj+1 ) → e modulo (Oj+1 ∩Hj )+ ∩(Oj+1 ∩Hj )− = Hj+1 . This completes
the recursive construction.
Note that
xi ∈ xj Hj · · · Hi−1 ⊆ xj Oj · · · Oi−1 = xj Oj
for all i ≥ j ≥ 1. Hence (xj )j∈N0 is a Cauchy sequence in G (for the left uniform
structure) and hence convergent, as G is locally compact and therefore complete
by [2, Paragraph 3.3, Corollary 1]. Let
y := lim xn
n→∞
and h := x−1
0 y.
Since xj ∈ x0 H0 · · · Hj−1 ⊆ x0 H for all j ∈ N, passing to the limit we deduce
that y ∈ x0 H and thus h ∈ H.
To see that y ∈ con(α), let W ⊆ G be an identity neighbourhood. After
shrinking W , we may assume that W = Oj for some j ∈ N. As
xi ∈ xj Hj · · · Hi−1 ⊆ xj Hj
for all i ≥ j, we deduce that y ∈ xj Hj and thus
αn (y) ∈ αn (xj )αn (Hj ) = αn (xj )Hj ⊆ Oj Oj = Oj = W
for all n ≥ Nj . Hence y ∈ con(α), which completes the proof.
4. Proof of Theorem B for metrizable G
In this section we prove Theorem B for metrizable G, using the following analogue of Lemma 3.2.
Lemma 4.1: Let G be a totally disconnected, locally compact group. Let
α : G → G be an endomorphism and H ⊆ G be an α-invariant closed subgroup. Let O ⊆ G be a compact open subgroup, x ∈ con− (α, H) and (xn )n∈N0
be an α-regressive trajectory for x such that xn → e modulo H. Then there
exist N ∈ N0 and an α-regressive trajectory (yn )n∈N0 such that
yn ∈ O
for all n ≥ N ,
yn ∈ xn H for all n ∈ N0 , and yn → e modulo (H ∩ O)+ ∩ (H ∩ O)− .
20
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
T
Proof. By Lemma 3.1, after replacing O with ℓj=0 α−j (O) for some ℓ ∈ N0 , we
may assume that O ∩ H is tidy above for α|H , i.e.,
(15)
H ∩ O = (H ∩ O)+ (H ∩ O)− .
We choose a compact open subgroup V ⊆ O such that α(V ) ⊆ O. Since xn → e
modulo H, there exists N ∈ N0 such that
xn ∈ V H
for all n ≥ N .
We show by induction that there exist elements zj ∈ xN +j H for j ∈ N0 with
αi (zj ) ∈ V (H ∩ O)
for all i ∈ {0, 1, . . . , j}.
Since xN ∈ V H, we have xN ∈ z0 H for some z0 ∈ V . Then z0 is as required.
If j ∈ N0 and zj has already been found, then zj ∈ xN +j H implies that
zj = xN +j a
for some a ∈ H. Since xN +j+1 ∈ V H, we can write
xN +j+1 = vb
with v ∈ V and b ∈ H. Thus
zj = xN +j a = α(xN +j+1 )a = α(v)α(b)a,
with α(b) ∈ α(H) ⊆ H. Since zj ∈ V (H ∩ O) ⊆ O and α(v) ∈ α(V ) ⊆ O, we
conclude that α(b)a ∈ H ∩ O and thus
α(b)a = h+ h−
for suitable h+ ∈ (H ∩O)+ and h− ∈ (H ∩O)− . Since α((H ∩O)+ ) ⊇ (H ∩O)+ ,
we find h ∈ (H ∩ O)+ such that h+ = α(h). Thus
zj+1 := vh ∈ V (H ∩ O)+ ⊆ V (H ∩ O)
and α(zj+1 ) = α(v)h+ = zj (h− )−1 , entailing that for i ∈ {1, . . . , j + 1}
αi (zj+1 ) =
∈
αi−1 (α(zj+1 )) = αi−1 (zj )αi−1 (h− )−1
V (H ∩ O)(H ∩ O)− = V (H ∩ O).
This completes the induction.
(j)
For j ∈ N0 , define the sequence g (j) = (gi )i∈N0 in V (H ∩ O) via
(
αj−i (zj ) if i ∈ {0, . . . , j};
(j)
gi :=
e if i > j.
ENDOMORPHISMS OF T.D.L.C. GROUPS
21
Then (g (j) )j∈N0 is a sequence in the compact space (V (H ∩ O))N0 and hence
has a convergent subnet (g (ja ) )a∈A , whose limit f = (fi )i∈N0 is an α-regressive
trajectory by Lemma 2.3 (a). Then also
(
fn−N if n ≥ N
yn :=
αN −n (f0 ) if n ∈ {0, 1, . . . , N }
is an α-regressive trajectory. Let i ∈ N0 ; since
(j)
gi
= αj−i (zj ) ∈ xN +i H
for all j ≥ i and xN +i H is closed, we deduce that
yN +i = fi ∈ xN +i H.
In particular, f0 = xN h0 for some h0 ∈ H and thus also
yn = αN −n (fN ) = αN −n (xN )αN −n (h0 ) = xn αN −n (h0 ) ∈ xn H
for n ∈ {0, 1, . . . , N }.
Since yn ∈ V (H ∩ O) ⊆ O for n ≥ N , we have yn → e modulo O. As yn ∈ xn H
and xn → e modulo H, also yn → e modulo H. Hence yn → e modulo
H ∩ O, by Lemma 2.10, and thus yn → e modulo (H ∩ O)+ ∩ (H ∩ O)− , by
Lemma 2.12.
Proof. (Theorem B). The inclusion con− (α)H ⊆ con− (α, H) is trivial. Let
(0)
y0 ∈ con− (α, H). Thus y0 has an α-regressive trajectory (yn )n∈N0 with
yn(0) → e modulo H.
Let (O(j) )j∈N be a sequence of compact open subgroups O(1) ⊇ O(2) ⊇ · · · of G
which forms a basis of identity neighbourhoods. Set H0 := H and
Hj := (O(j) ∩ Hj−1 )+ ∩ (O(j) ∩ Hj−1 )− ⊆ O(j) ∩ Hj−1 for j ∈ N.
We claim that there exist integers Nj ∈ N0 for j ∈ N and α-regressive trajecto(j)
ries y (j) = (yn )n∈N0 such that
yn(j) → e modulo Hj ,
(j)
(j)
(j−1)
yn ∈ W (j) for all n ≥ Nj , and yn ∈ yn
Hj−1 for all n ∈ N0 .
Lemma 4.1 (with y0 in place of x and O(1) in place of O) provides N1 ∈ N0 and
(1)
(0)
(1)
an α-regressive trajectory (yn )n∈N0 with yn ∈ yn H for all n ∈ N0 ,
yn(1) ∈ O(1)
for all n ≥ N1 ,
22
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
and y (1) → e modulo (O(1) ∩H)+ ∩(O(1) ∩H)− = H1 . This shows the case j = 1.
(1)
(j)
If N1 , . . . , Nj and α-regressive trajectories (yn )n∈N0 , . . . (yn )n∈N0 have been
(j)
(j)
constructed, we can apply Lemma 4.1 with y0 in place of x, (yn )n∈N0 in
place of (xn )n∈N0 , O(j+1) in place of O and Hj in place of H. We obtain an
(j+1)
)n∈N0 such
integer Nj+1 ∈ N0 and an α-regressive trajectory y (j+1) = (yn
(j+1)
(j+1)
(j)
that yn
∈ O(j+1) for all n ≥ Nj+1 , and moreover yn
∈ yn Hj for all
(j+1)
n ∈ N0 and yn
→ e modulo (O(j+1) ∩ Hj )+ ∩ (O(j+1) ∩ Hj )− = Hj+1 . This
completes the recursive construction.
(j+1)
(j)
(j)
(j)
We have yn
∈ yn Hj ⊆ yn O(j) ⊆ yn O(1) for all n ∈ N0 and j ∈ N. Hence
Q
(1)
(j)
(1) (1)
yn ∈ yn O
for all j ∈ N and n ∈ N0 , showing that y (j) ∈ n∈N0 yn O(1)
for all j ∈ N. The product being compact and metrizable, we find a convergent
Q
(1)
subsequence of (y (j) )j∈N , with limit z = (zn )n∈N0 ∈ n∈N0 yn O(1) , say. As
each y (j) is an α-regressive trajectory and α is continuous, we see that also
(j)
(0)
(zn )n∈N0 is an α-regressive trajectory. Since y0 ∈ y0 H = y0 H for each j ∈ N
and y0 H is closed, we have z0 ∈ y0 H.
(i)
(j)
If j ∈ N and n ∈ N0 , then yn ∈ yn O(j) for all i ≥ j, entailing that also
(j)
zn ∈ yn O(j) . Hence zn ∈ O(j) for all n ≥ Nj , whence zn → e and thus
z0 ∈ con− (α).
5. Metrizable quotient groups
This section prepares the discussion of endomorphisms of non-metrizable groups.
As in the case of automorphisms treated in [13], we consider metric quotients
of such as a tool.
Lemma 5.1: Let G be a locally compact group and N a compact normal subgroup. Then the following conditions are equivalent:
(a) G/N is metrizable.
(b) There exists a sequence (Vn )n∈N of open identity neighbourhoods Vn ⊆
T
G such that n∈N Vn = N .
Proof. Let q : G → G/N , x 7→ xN be the canonical quotient map, which is continuous and open. (a)⇒(b): If G/N is metrizable, then G/N has a countable basis {Wn : n ∈ N} of open identity neighbourhoods, and the sets Vn := q −1 (Wn )
(n ∈ N) are as desired.
ENDOMORPHISMS OF T.D.L.C. GROUPS
23
(b)⇒(a): Let (Vn )n∈N be a sequence of open identity neighbourhoods in G
with intersection N . As N is compact, the open neighbourhood Vn of N contains
a uniform neighbourhood of the form Kn N , for some identity neighbourhood
Kn ⊆ G. Since G is locally compact, Kn can be chosen compact. We may also
assume that K1 ⊇ K2 ⊇ · · · . Thus
U := {q(Kn ) : n ∈ N}
T
is a filter basis of compact identity neighbourhoods in G/N . Now n∈N q(Kn )
= {e} (as the sets q −1 (q(Kn )) = Kn N have intersection N ), entailing that
U converges to {e}. Thus U is a countable basis of identity neighbourhoods
in G/N and thus G/N is metrizable by 3.3.
Lemma 5.2: Let G be a locally compact group and N be a countable set of
closed normal subgroups N ⊆ G such that G/N is metrizable. Set M :=
T
N ∈N N . Then also G/M is metrizable.
Proof. For N ∈ N , let qN : G → G/N be the canonical quotient map. Also, let
qM : G → G/M be the canonical quotient map. Then
Y
φ := (qN )N ∈N : G →
G/N, x 7→ (xN )N ∈N
N ∈N
is a continuous homomorphism to a metrizable topological group. As the kernel
is ker(φ) = M , the induced continuous homomorphism
Y
φ : G/M →
G/N
N ∈N
(determined by φ ◦ qM = φ) is injective. If K ⊆ G is a compact identity
neighbourhood in G, then qM (K) is a compact identity neighbourhood in G/M .
Since φ(qM (K)) is metrizable and φ restricts to a homeomorphism from the
compact set qM (K) onto φ(qM (K)), we see that qM (K) is metrizable. Since
qM (K) is an identity neighbourhood in G/M , this topological group has a
countable basis of identity neighbourhoods and hence is metrizable.
The following fact will be used in the proofs of the non-metrizable case of
Theorem A and Theorem B.
Remark 5.3: If G is a locally compact group, M ⊆ G a σ-compact subset and
α : G → G an endomorphism, then there exists an α-invariant, σ-compact open
24
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
subgroup S ⊆ G such that M ⊆ S. To see this, let V ⊆ G be a compact identity
neighbourhood. Then the subgroup S of G generated by the σ-compact set
[
αn (M ∪ V )
n∈N0
is σ-compact and open. Moreover, M ⊆ S and α(S) ⊆ S.
Lemma 5.4: Let G be a σ-compact, locally compact group, and α : G → G be
an endomorphism. Let K be the set of all compact, normal subgroups K ⊆ G
T
such that α(K) ⊆ K and G/K is metrizable. Then K = {e}.
Proof. Let U ⊆ G be a neighbourhood of e. By [11, (8.7)], G has a compact,
normal subgroup N such that N ⊆ U and G/N is metrizable. Hence
\
N=
Vn
n∈N
for suitable open identity neighbourhoods Vn ⊆ G, by Lemma 5.1. Then
\ \
\
α−m (N ) =
K :=
α−m (Vn )
m∈N0
m∈N0 n∈N
is an α-invariant, compact normal subgroup of G contained in U . As K is
an intersection of a countable set of open identity neighbourhoods, G/K is
metrizable (by Lemma 5.1).
The following fact, whose proof is left to the reader, is useful for our purposes.
Lemma 5.5: Let (I, ≤) be a directed set and S := ((Gj )j∈I , (qi,j )i≤j ) be a
projective system of topological groups Gj and continuous homomorphisms
qi,j : Gj → Gi . Let G be a topological group and (qj )j∈I be a family of continuous homomorphisms qj : G → Gj such that qi,j ◦ qj = qi for i ≤ j in I. If
T
j∈I ker(qj ) = {e} and qj is a quotient homomorphism with compact kernel for
each j ∈ I, then (G, (qj )j∈I ) is the projective limit of S.
Remark 5.6: (a) In the situation of Lemma 5.5, let (xj )j∈I be a family of
elements xj ∈ Gj such that xi = qi,j (xj ) if i ≤ j. Then (xj )j∈I ∈ P , where P
is the standard realization of the projective limit. Thus there exists a unique
element x ∈ G such that qj (x) = xj for all j ∈ I.
(b) In the situation of Lemma 5.5, assume that H ⊆ G is a closed subgroup.
Then (H, (qj |H )j∈I ) is the projective limit of ((qj (H))j∈I , (qij |qj (H) )i≤j ).
In fact, since qj is a quotient homomorphism with compact kernel, it is a closed
ENDOMORPHISMS OF T.D.L.C. GROUPS
25
mapping (see [11, (5.18)]), whence qj |H : H → qj (H) is a surjective, continuous,
closed mapping and hence a quotient map. We can therefore apply Lemma 5.5
with H in place of G.
Definition 5.7: If α is an endomorphism of a σ-compact, totally disconnected,
locally compact group G, then there exists a non-empty set K of α-invariant,
compact normal subgroups K ⊆ G such that G/K is metrizable for each K ∈ K
T
and K = {e} (see Lemma 5.4). Let κ(G, α) := |K| be the minimum cardinality
of such K.
It is not difficult to see that κ(G, α) equals the character of G in the situation
of Definition 5.7 unless G is first-countable in which case κ(G, α) = 1 as we
may choose K := {e}. Notably, κ(G, α) is independent of α. We suppress the
details as the additional information shall not be used.
6. Proof of Theorem A for non-metrizable G
To pass from metrizable to general groups in Theorem A, we could adapt the arguments from [13]. However, as the treatment of Theorem B requires a different
approach, we prefer to apply this new approach also to Theorem A.
Definition 6.1: We say that an endomorphism α of a topological group G
has the decomposition property if con(α, K) = con(α)K for each α-invariant
compact subgroup K ⊆ G.
Lemma 6.2: Let α be an endomorphism of a topological group G and x ∈
con(α, H) for an α-invariant compact subgroup H ⊆ G. Let q : G → Q be a
quotient homomorphism to a topological group Q with α-invariant kernel N .
Let α : Q → Q be the map determined by α ◦ q = q ◦ α. Let y ∈ con(α) ⊆ Q
with y ∈ q(x)q(H). Then there exists y ∈ con(α, H ∩ N ) such that
(16)
q(y) = y
and y ∈ xH.
If α has the decomposition property, then there exists y ∈ con(α) with (16).
Proof. Let h ∈ H such that y = q(x)q(h). As q(xh) = y ∈ con(α), we have xh ∈
con(α, N ). Since x ∈ con(α, H) = con(α, H)H, we deduce xh ∈ con(α, H).
Hence xh ∈ con(α, H ∩ N ), by Lemma 2.10 (and we can take y := xh). If α
has the decomposition property, we can write xh = yn with y ∈ con(α) and
n ∈ H ∩ N . Then y ∈ xH and q(y) = q(xh) = y.
26
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
6.3: Let I be an interval of ordinal numbers of the form [0, k[ or [0, k]. Let
α be an endomorphism of a topological group G and (Nj )j∈I be a family of
α-invariant, compact normal subgroups Nj ⊆ G such that Nj ⊆ Ni for i ≤ j
in I. For j ∈ I, let qj : G → G/Nj be the canonical quotient map and αj be
the endomorphism of G/Nj determined by αj ◦ qj = qj ◦ α. For i ≤ j, let
qi,j : G/Nj → G/Ni
be the map determined by qi,j ◦ qj = qi . Then ((G/Nj )j∈I , (qi,j )i≤j ) (with
i, j ∈ I) is a projective system of topological groups.
The next lemma is a variant of [13, Lemma 7].
Lemma 6.4: Let 0 6= k be a limit ordinal, α an endomorphism of a topological
group G and (Nj )j∈[0,k[ be a family of α-invariant, compact normal subgroups
T
Nj ⊆ G such that Nj ⊆ Ni for i ≤ j in [0, k[. Let Nk := j∈[0,k[ Nj and
use notation as in 6.3 with I := [0, k]. Let H ⊆ G be an α-invariant closed
subgroup and x ∈ con(α, H). Assume that, for each j ∈ [0, k[, an element
yj ∈ con(αj ) ⊆ G/Nj is given such that
(17)
qi,j (yj ) = yi
(18)
yj ∈ qj (x)qj (H)
for all i ≤ j in [0, k[, and
for all j ∈ [0, k[.
Then there is a unique element yk ∈ G/Nk such that
(19)
qi,k (yk ) = yi
for all i ∈ [0, k[.
Moreover, yk ∈ con(αk ) ⊆ G/Nk and
(20)
yk ∈ qk (x)qk (H).
Proof. By Lemma 5.5, G/Nk is the projective limit of the projective system
((G/Nj )j∈[0,k[ , (qi,j )i≤j ), with the limit maps qi,k . Thus, the identities (17)
determine a unique element yk ∈ G/Nk such that (19) holds. By Remark 5.6 (b),
qk (H) is the projective limit of its quotients qi (H) for i ∈ [0, k[, with the limit
maps qi,k |qk (H) : qk (H) → qi (H). Since
qi,k (qk (x)−1 yk ) = qi (x)−1 yi ∈ qi (H) for all i ∈ [0, k[,
we deduce that qk (x)−1 yk ∈ qk (H). To see that yk ∈ con(αk ), let U ⊆ G/Nk be
−1
an identity neighbourhood. Then qi,k
(V ) ⊆ U for some i ∈ [0, k[ and identity
ENDOMORPHISMS OF T.D.L.C. GROUPS
27
neighbourhood V ⊆ G/Ni . Since limn→∞ (αi )n (qi (yi )) = e, there is n0 ∈ N
such that
qi,k ((αk )n (yk )) = (αi )n (yi ) ∈ V
−1
for all n ≥ n0 and thus (αk )n (yk ) ∈ qi,k
(V ) ⊆ U .
The following lemma is a variant of [13, Lemma 8].
Lemma 6.5: Let k > 0 be an ordinal number, α be an endomorphism of a
topological group G and (Nj )j∈[0,k[ be a family of α-invariant, compact normal
subgroups Nj ⊆ G such that Nj ⊆ Ni for all i ≤ j in [0, k[ and
\
Ni
Nj =
i<j
for each limit ordinal 0 6= j ∈ [0, k[. Let H ⊆ G be an α-invariant compact
subgroup and x ∈ con(α, H) ⊆ G. Use notation as in 6.3 with I := [0, k[.
Assume that αj has the decomposition property for each j ∈ [0, k[. Then there
exist elements yj ∈ con(αj ) ⊆ G/Nj for j ∈ [0, k[, such that conditions (17) and
(18) from Lemma 6.4 are satisfied.
Proof. For all j ∈ [0, k[, we have qj (x) ∈ con(αj , qj (H)), by Lemma 2.13 (b).
Let M be the set of all families (yj )j∈[0,ℓ[ with ℓ ∈ ]0, k] such that (17) and (18)
are satisfied, with ℓ in place of k. Since α0 has the decomposition property and
q0 (x) ∈ con(α0 , q0 (H)), there exists y0 ∈ con(α0 ) such that
y0 ∈ q0 (x)q0 (H).
Then (yj )j∈{0} ∈ M and thus M 6= ∅. For f = (yj )j∈[0,ℓ[ and g = (yj′ )j∈[0,ℓ′ [
in M , write f ≤ g if ℓ ≤ ℓ′ and f = g|[0,ℓ[ . To see that (M, ≤) is inductively
ordered, let Γ be a totally ordered subset of M . If Γ = ∅, then every element
of M (which is non-empty) is an upper bound for Γ. If Γ 6= ∅, write [0, ℓf [
for the domain of f ∈ Γ. Let ℓ = sup{ℓf : f ∈ Γ} ∈ ]0, k]. Then the union of
the graphs of the f ∈ Γ is the graph of a function g ∈ M which is an upper
bound for Γ. Thus (M, ≤) is inductively ordered and thus M has a maximal
element f = (yj )j∈[0,ℓ[ , by Zorn’s Lemma. If ℓ = k, then the assertion of the
lemma holds for α. If ℓ < k, we obtain a contradiction because f could then
be extended to an element of M defined on [0, ℓ] = [0, ℓ + 1[, as we now verify.
It suffices to find an element yℓ ∈ con(αℓ ) ⊆ G/Nℓ such that (19) and (20) are
satisfied, with ℓ in place of k.
If ℓ is a limit ordinal, then Lemma 6.4 yields yℓ ∈ con(αℓ ) as desired.
28
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
If ℓ has a precursor j, we apply Lemma 6.2 with G/Nℓ , the quotient map
qj,ℓ : G/Nℓ → G/Nj , αℓ , qℓ (Nj ), qℓ (H), qℓ (x), and yj in place of G, q, α, N ,
H, x, and y, respectively. Since αℓ has the decomposition property, we obtain
yℓ ∈ con(αℓ ) ⊆ G/Nℓ such that qj,ℓ (yℓ ) = yj and yℓ ∈ qℓ (x)qℓ (H).
Proof. (Theorem A, completed). We show by transfinite induction on cardinals k ≥ 1 that each endomorphism α of a σ-compact, totally disconnected,
locally compact group G with κ(G, α) ≤ k has the decomposition property
(where κ(G, α) is as in Definition 5.7). If k ≤ ℵ0 , then κ(G, α) ≤ ℵ0 implies
that G is metrizable (see Lemma 5.2); thus α has the decomposition property
(as verified in Section 3).
Let k > ℵ0 now and assume that all endomorphisms β of σ-compact, totally disconnected, locally compact groups H with κ(H, β) < k have the decomposition
property. The assertion will hold for k if we can show that every endomorphism α of a σ-compact, totally disconnected, locally compact group G with
κ(G, α) = k has the decomposition property. In this situation, we choose a
set K with |K| = k as in Definition 5.7 and a bijection
[0, k[ → K,
We define
Nj :=
\
Ki
j 7→ Kj .
for all j ∈ ]0, k[ and N0 := N1 .
i<j
This yields a decreasing family (Nj )j∈[0,k[ of α-invariant, compact normal subT
groups of G with Nj = i<j Ni for each limit ordinal 0 6= j ∈ [0, k[. For
j ∈ [0, k[, let qj : G → G/Nj be the canonical quotient map and αj be the
endomorphism of G/Nj induced by α. Then Kj := {qj (Ki ) : i < j} is a set of
cardinality at most j which has the properties of K required in Definition 5.7,
with G/Nj in place of G and αj in place of α. Hence
κ(G/Nj , αj ) ≤ |Kj | ≤ j < k,
and thus αj has the decomposition property. Note that k, being an infinite cardinal, is a limit ordinal. Given x ∈ con(α, H) ⊆ G, we apply Lemma 6.5 to find
yj ∈ con(αj ) ⊆ G/Nj for j ∈ [0, k[ such that (17) and (18) are satisfied. Now
T
Lemma 6.4 provides y ∈ G ∼
= G/ j<k Nj such that αn (y) → e and x−1 y ∈ H.
Hence α has the decomposition property, as required.
Now let α be an endomorphism of a totally disconnected, locally compact
ENDOMORPHISMS OF T.D.L.C. GROUPS
29
group G which need not be σ-compact. Let H ⊆ G be an α-stable closed subgroup or an α-invariant compact subgroup. Given x ∈ con(α, H), Lemma 3.2
provides an h ∈ H such that xh ∈ con(α, K) with K := (O ∩ H)+ ∩ (O ∩ H)− .
As observed in Remark 5.3, G has an α-invariant σ-compact open subgroup S
such that K ∪ {xh} ⊆ S. Now, by Lemma 2.15 (a),
xh ∈ con(α, K) ∩ S = con(α|S , K).
Since α|S has the decomposition property (as S is σ-compact), we find y ∈
con(α|S ) ⊆ con(α) and k ∈ K such that xh = yk. Thus x = ykh−1 ∈
con(α)KH = con(α)H.
7. Theorem B for non-metrizable G
To deal with anti-contraction groups (rather than contraction groups), we need
to consider regressive trajectories rather than individual group elements. This
consideration leads to the following analogue of Definition 6.1.
Definition 7.1: Let α be an endomorphism of a topological group G. We say
that α-regressive trajectories decompose if, for each α-invariant closed subgroup
H ⊆ G and α-regressive trajectory (xn )n∈N0 such that xn → e modulo H, there
exists an α-regressive trajectory (yn )n∈N0 with yn ∈ xn H for all n ∈ N0 and
limn→∞ yn = e. Then
con− (α, H) = con− (α)H
in particular (since x0 = y0 h for some h ∈ H, we have con− (α, H) ⊆ con− (α)H;
the converse inclusion is trivial).
Lemma 7.2: Let α be an endomorphism of a topological group G and (xn )n∈N0
be an α-regressive trajectory in G such that xn → e modulo H for some
α-invariant closed subgroup H ⊆ G. Let N ⊆ G be an α-invariant, compact normal subgroup and q : G → G/N be the canonical quotient map. Let
α : G/N → G/N be the induced map determined by α ◦ q = q ◦ α and (y n )n∈N0
be an α-regressive trajectory in G/N such that y n ∈ q(xn )q(H) for each n ∈ N0 .
Then there exists an α-regressive trajectory (yn )n∈N0 in G such that
q(yn ) = y n
and yn ∈ xn H
for all n ∈ N0 .
If y n → e as n → ∞ and xn → e modulo H, then yn → e modulo H ∩ N .
30
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Proof. For each n ∈ N0 , there is hn ∈ H such that y n = q(xn )q(hn ). We
(m)
let (z (m) )m∈N0 = ((zn )n∈N0 )m∈N0 be the double sequence associated with
(xn hn )n∈N0 (as in Definition 2.4). Given n ∈ N0 , we have
q(zn(m) ) = q(αm−n (xm hm )) = αm−n (q(xm hm )) = αm−n (y m ) = y n = q(xn hn )
(m)
for all m ≥ n. Moreover, zn = αm−n (xm hm ) = xn αm−n (hm ) ∈ xn H. Thus
(m)
zn is contained in the compact set xn (H ∩ hn N ) for all m ≥ n (and always in
xn (H ∩hn N )∪{e}), whence (z (m) )m∈N0 has an accumulation point y = (yn )n∈N0
Q
in n∈N0 xn (H ∩ hn N ). Then q(yn ) = q(xn hn ) = y n and yn ∈ xn H. By
Lemma 2.3, (yn )n∈N0 is an α-regressive trajectory.
If q(yn ) = y n → e, then yn → e modulo N . If, moreover, xn → e modulo H,
then yn ∈ xn H implies that also yn → e modulo H. Hence yn → e modulo
H ∩ N (by Lemma 2.10), as N is compact.
Lemma 7.3: Let 0 6= k be a limit ordinal, α be an endomorphism of a topological group G and (Nj )j∈[0,k[ be a family of α-invariant, compact normal
T
subgroups Nj ⊆ G such that Nj ⊆ Ni for i ≤ j in [0, k[. Let Nk := j∈[0,k[ Nj
and use notation as in 6.3 with I := [0, k]. Let H ⊆ G be an α-invariant,
closed subgroup and (xn )n∈N0 be an α-regressive trajectory in G such that
xn → e modulo H. Assume that, for each j ∈ [0, k[, an αj -regressive trajectory
(j)
(yn )n∈N0 in G/Nj with lim yn(j) = e is given such that, for each n ∈ N0 ,
n→∞
(21)
qi,j (yn(j) ) = yn(i)
(22)
yn(j) ∈ qj (xn )qj (H)
for all i ≤ j in [0, k[,
for all j ∈ [0, k[.
(k)
Then there are unique elements yn ∈ G/Nk such that
qi,k (yn(k) ) = yn(i)
(23)
for all i ∈ [0, k[.
(k)
Moreover, (yn )n∈N0 is an αk -regressive trajectory in G/Nk which converges
to e and
yn(k) ∈ qk (xn )qk (H)
(24)
for all n ∈ N0 .
Proof. By Lemma 5.5, G/Nk is the projective limit of the projective system
((G/Nj )j∈[0,k[ , (qi,j )i≤j ), with the limit maps qi,k . Thus, for n ∈ N0 ,
(k)
(21) determines a unique yn ∈ G/Nk with (23) (see Remark 5.6 (a)). As
(k)
(k)
(i)
qi,k (αk (yn+1 )) = αi (qi,k (yn+1 )) = αi (yn+1 ) = yn(i) = qi,k (yn(k) ) for all i ∈ [0, k[,
ENDOMORPHISMS OF T.D.L.C. GROUPS
(k)
(k)
31
(k)
we have αk (yn+1 ) = yn . Thus (yn )n∈N0 is an αk -regressive trajectory. Now
qk (H) is the projective limit of its quotients qi (H) for i ∈ [0, k[, with the limit
maps qi,k |H : H → qi,k (H) (see Remark 5.6 (b)). Since
qi,k (qk (xn )−1 yn(k) ) = qi (xn )−1 yn(i) ∈ qi (H) for all i ∈ [0, k[,
(k)
(k)
we deduce that qk (x)−1 yn ∈ qk (H). To see that yn → e, let U ⊆ G/Nk be
−1
an identity neighbourhood. Then qi,k
(V ) ⊆ U for some i ∈ [0, k[ and identity
(i)
neighbourhood V ⊆ G/Ni . Since qi (yn ) → e, there is n0 ∈ N such that
qi,k (yn(k) ) = yn(i) ∈ V
(k)
−1
for all n ≥ n0 and thus yn ∈ qi,k
(V ) ⊆ U .
Lemma 7.4: Let k be an ordinal number, α be an endomorphism of a topological group G and (Nj )j∈[0,k[ be a family of α-invariant, compact normal
subgroups Nj ⊆ G such that Nj ⊆ Ni for all i ≤ j in [0, k[ and
\
Ni
Nj =
i<j
for each limit ordinal 0 6= j ∈ [0, k[. Let H ⊆ G be an α-invariant closed
subgroup. Let (xn )n∈N0 be an α-regressive trajectory in G such that xn → e
modulo H. Using notation as in 6.3, assume that αj -regressive trajectories in
G/Nj decompose for each j ∈ [0, k[. Then there exist αj -regressive trajectories
(j)
(yn )n∈N0 in G/Nj which converge to e, for all j ∈ [0, k[, such that conditions
(21) and (22) from Lemma 7.3 are satisfied.
Proof. For each j ∈ [0, k[, the sequence (qj (xn ))n∈N0 is an αj -regressive trajectory in G/Nj . Moreover, qj (xn ) → e modulo qj (H), by Lemma 2.13 (a). Let
(j)
M be the set of all families f = (yn )(j,n)∈[0,ℓ[ ×N0 with ℓ ∈ ]0, k] such that (21)
and (22) are satisfied, with ℓ in place of k. Since α0 -regressive trajectories in
G/N0 decompose and q0 (xn ) → e modulo q0 (H), there exists an α0 -regressive
(0)
trajectory (yn )n∈N0 such that
(0)
(0)
yn ∈ q0 (xn )q0 (H) for all n ∈ N0 and yn → e.
(j)
(j)
Then (yn )(j,n)∈{0}×N0 ∈ M and thus M 6= ∅. For f = (yn )(j,n)∈[0,ℓ[ ×N0 and
(j)
(zn )(j,n)∈[0,ℓ′ [ ×N0
in M , write f ≤ g if ℓ ≤ ℓ′ and f = g|[0,ℓ[ ×N0 . To see that
g=
(M, ≤) is inductively ordered, let Γ be a totally ordered subset of M . If Γ = ∅,
then every element of M (which is non-empty) is an upper bound for Γ. If Γ 6= ∅,
write [0, ℓf [ ×N0 for the domain of f ∈ Γ. Let ℓ = sup{ℓf : f ∈ Γ} ∈ ]0, k]. Then
32
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
the union of the graphs of the f ∈ Γ is the graph of a function g ∈ M which
is an upper bound for Γ. Hence (M, ≤) is inductively ordered and thus M
(j)
has a maximal element f = (yn )(j,n)∈[0,ℓ[ ×N0 , by Zorn’s Lemma. If ℓ = k,
then the assertion of the lemma holds for α and (xn )n∈N0 . If ℓ < k, we reach a
(j)
contradiction, as follows: Since [0, ℓ+1[ = [0, ℓ], we can extend (yn )(j,n)∈[0,ℓ[ ×N0
to a properly larger element in (M, ≤) if we can find an αℓ -regressive trajectory
(ℓ)
(ℓ)
(yn )n∈N0 in G/Nℓ with yn → e such that (23) and (24) are satisfied, with ℓ
in place of k.
(ℓ)
If ℓ is a limit ordinal, then Lemma 7.3 yields (yn )n∈N0 in G/Nℓ as desired
(contradiction).
If ℓ has a precursor j, we apply Lemma 7.2 with G/Nℓ , the quotient map
(j)
qj,ℓ : G/Nℓ → G/Nj , αℓ , qℓ (Nj ), qℓ (H), (qℓ (xn ))n∈N0 , and (yn )n∈N0 in place
of G, q, α, N , H, (xn )n∈N0 , and (y n )n∈N0 , respectively. We obtain an αℓ regressive trajectory (yn )n∈N0 in G/Nℓ such that
yn → e modulo qℓ (H) ∩ qℓ (Nj )
(j)
and qj,ℓ (yn ) = yn as well as yn ∈ qℓ (xn )qℓ (H) for all n ∈ N0 . Since αℓ (ℓ)
regressive trajectories decompose, we find an αℓ -regressive trajectory (yn )n∈N0
(ℓ)
in G/Nℓ such that yn → e and
yn(ℓ) ∈ yn (qℓ (H) ∩ qℓ (Nj )) ⊆ qℓ (xn )qℓ (H),
(k)
(j)
whence qj,ℓ (yn ) = qj,ℓ (yn ) = yn . Again, a contradiction is obtained.
Proof. (Theorem B, completed). We show by transfinite induction on cardinals k ≥ 1 that α-regressive trajectories decompose for each endomorphism α of
a σ-compact, totally disconnected, locally compact group G such that κ(G, α) ≤
k (where κ(G, α) is as in Definition 5.7). If k ≤ ℵ0 , then κ(G, α) ≤ ℵ0 implies
that G is metrizable (see Lemma 5.2); thus α-regressive trajectories decompose
(as verified in Section 4).
Let k > ℵ0 now and assume that β-regressive trajectories decompose for all endomorphisms β of σ-compact, totally disconnected, locally compact groups H
with κ(H, β) < k. The assertion will hold for k if we can show that α-regressive
trajectories decompose for each endomorphism α of a σ-compact, totally disconnected, locally compact group G such that κ(G, α) = k. In this situation,
ENDOMORPHISMS OF T.D.L.C. GROUPS
33
we choose a set K with |K| = k as in Definition 5.7 and a bijection
[0, k[ → K,
j 7→ Kj .
We define
Nj :=
\
Ki
for all j ∈ ]0, k[ and N0 := N1 .
i<j
This yields a decreasing family (Nj )j∈[0,k[ of α-invariant, compact normal subT
groups of G with Nj = i<j Ni for each limit ordinal 0 6= j ∈ [0, k[. For
j ∈ [0, k[, let qj : G → G/Nj be the canonical quotient map and αj be the
endomorphism of G/Nj induced by α. Then Kj := {qj (Ki ) : i < j} is a set
of cardinality ≤ j which has the properties of K required in Definition 5.7,
with G/Nj in place of G and αj in place of α. Hence
κ(G/Nj , αj ) ≤ |Kj | ≤ j < k,
and thus αj -regressive trajectories decompose. Note that k, being an infinite
cardinal, is a limit ordinal. Given an α-regressive trajectory (xn )n∈N0 in G, we
(j)
apply Lemma 7.4 to find αj -regressive trajectories (yn )n∈N0 for j ∈ [0, k[ such
(j)
that limn→∞ yn = e and both (21) and (22) are satisfied. Now Lemma 7.3 proT
vides an α-regressive trajectory (yn )n∈N0 in G ∼
= G/ j<k Nj such that yn → e
and x−1
n yn ∈ H for all n ∈ N0 . Hence α-regressive trajectories decompose, as
required.
Now let α be an endomorphism of a totally disconnected, locally compact
group G which need not be σ-compact, H ⊆ G be an α-invariant closed subgroup and (xn )n∈N0 be an α-regressive trajectory such that xn → e modulo H.
There exists a σ-compact, α-invariant open subgroup S ⊆ G which contains
the countable (hence σ-compact) set {xn : n ∈ N0 } (see Remark 5.3). Then
S ∩ H is an α|S -invariant closed subgroup of S and (xn )n∈N0 an α|S -regressive
trajectory in S such that xn → e modulo S ∩ H (see Lemma 2.15 (b)). Since
α|S -regressive trajectories decompose (by σ-compactness of S), there exists an
α|S -regressive trajectory (yn )n∈N0 in S such that yn → e in S (hence in G)
and yn ∈ xn (S ∩ H) ⊆ xn H for all n ∈ N0 . Thus α-regressive trajectories
decompose. Notably, con(α, H) = con(α)H.
34
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
8. Subgroups and quotients
In this section, we generalize several results of [23] about how tidy subgroups
and the scale behave with respect to taking subgroups and quotients from automorphisms to endomorphisms, paralleling the topological entropy study in [5].
8.1: For many results in this section we rely on the tidying procedure from [26]
which we shall outline for the benefit of the reader: Given a totally disconnected
locally compact group G, an endomorphism α and U ∈ COS(G), the following
steps produce a tidy subgroup for α.
Tn
(1) There exists n ∈ N such that U−n := k=0 α−k (U ) is tidy above for α.
By replacing U with U−n we assume U is tidy above.
(2) Define
LU := {x ∈ G : ∃y ∈ U+ ∃m, n ∈ N with αm (y) = x and αn (x) ∈ U− }
and LU := LU .
e := {x ∈ U : xLU ⊆ LU U }.
(3) Set U
e LU is a compact open subgroup of G which is tidy for α.
(4) Then U
e LU = U .
If the original subgroup U was already tidy, then U
Subgroups. We first explore the effect of taking subgroups on tidiness and
the scale. The following lemma shows that tidy subgroups behave well when
passing to subgroups. It is applied in Proposition 8.3 which deals with the scale.
Lemma 8.2: Let G be a totally disconnected, locally compact group, α an
endomorphism of G and H ⊆ G a closed subgroup with α(H) ⊆ H. Further,
let U ∈ COS(G) be tidy for α. Set V := U ∩ H. Then there is N ∈ N such that
V−N is tidy for α|H .
Proof. Note that V is a compact open subgroup of H. By [26, Proposition 3]
there is N ∈ N such that V−N is tidy above for α. Since U is minimizing, the
same proposition implies that U−N is tidy for α. By Lemma 3.1, replacing U
by U−N , we may assume that V is tidy above for α|H . To see that this V is
tidy, we show that LV ⊆ V where LV is given in 8.1. Since V ⊆ H is closed
this implies that LV = LV ⊆ V and hence V is tidy below and therefore tidy
for α|H by [26, Proposition 8]. First, note that
\
V− =
V−n = U− ∩ H
n≥0
ENDOMORPHISMS OF T.D.L.C. GROUPS
35
(see (14)). Also, since V+ is the collection of all elements in V that admit an
α-regressive trajectory in V = U ∩ H, it follows that V+ ⊆ U+ ∩ H. Now,
suppose that x ∈ LV . Then x ∈ H and there are y ∈ V+ and m, n ∈ N such
that αm (y) = x and αn (y) ∈ V− . By the above, y ∈ U+ and αn (y) ∈ U− .
Therefore, x ∈ LU ∩ H. Since U is tidy for α we have LU ⊆ U and thus
conclude x ∈ U ∩ H = V . This shows LV ⊆ V as required.
Proposition 8.3: Let G be a totally disconnected, locally compact group, α
an endomorphism of G and H ⊆ G a closed subgroup with α(H) ⊆ H. Then
sH (α|H ) ≤ sG (α).
Furthermore, if H is normal in G and U ∈ COS(G) is tidy for α such that U ∩H
is tidy for α|H , then α((U ∩ H)+ )U+ is a subgroup of G and
sH (α|H ) = [α((U ∩ H)+ )U+ : U+ ].
Proof. By Lemma 8.2 there is a tidy subgroup U ∈ COS(G) for α such that
V := U ∩ H is tidy for α|H . In particular, sH (α|H ) = [α(V+ ) : V+ ] and sG (α) =
[α(U+ ) : U+ ]. Define ϕ : α(V+ )/V+ → α(U+ )/U+ by setting ϕ(uV+ ) := uU+ for
all uV+ ∈ α(V+ )/V+ . Clearly ϕ is well defined as V+ ⊆ U+ . For the first claim
it suffices to show that ϕ is injective. Indeed, assume that ϕ(uV+ ) = ϕ(vV+ ) for
some uV+ , vV+ ∈ α(V+ )/V+ . Then it follows that x := v −1 u ∈ α(V+ )∩U+ where
α(V+ ) = α((U ∩H)+ ) ⊆ H. By [26, Lemma 1] we conclude x ∈ U ∩H ∩α(V+ ) =
V ∩ α(V+ ) = V+ .
For the second claim suppose that H is normal in G. It suffices to show that
α((U ∩ H)+ )U+ = U+ α((U ∩ H)+ ): Indeed, this implies that α((U ∩ H)+ )U+
is a group and the assertion then follows from the previous paragraph. Now,
(U ∩ H)0 := U ∩ H is a normal subgroup of U0 := U and (U ∩ H)n+1 :=
α((U ∩ H)n ) ∩ U ∩ H is normal in Un+1 := α(Un ) ∩ U for each n ∈ N0 by the
following inductive argument: By the inductive hypothesis, (U ∩ H)n is normal
in Un , whence α((U ∩ H)n ) is normal in α(Un ). Since U ∩ H is normal in U , it
follows that α((U ∩ H)n ) ∩ U ∩ H is normal in α(Un ) ∩ U which completes the
induction. As a consequence,
\
\
Un .
(U ∩ H)n is normal in U+ :=
(U ∩ H)+ :=
n∈N0
n∈N0
Let u ∈ U+ . Then u = α(w) for some w ∈ U+ . Applying α to (U ∩ H)+ w =
w(U ∩ H)+ , we deduce that α((U ∩ H)+ )u = uα((U ∩ H)+ ).
36
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Quotients. We now turn our attention to quotients. Again, we first consider
tidy subgroups and then apply our findings to gain insight into the scale. Our
first lemma is a tool that provides control over α-regressive trajectories. In the
e are as in the tidying procedure, 8.1.
following, LU and U
Lemma 8.4: Let G be a totally disconnected, locally compact group and α an
endomorphism of G. Further, let U ∈ COS(G) be tidy above for α. Then
e LU = U
e.
U ∩U
e ⊆ U ∩U
e LU as U
e ⊆ U and U
e ⊆ U
e LU . Now
Proof. The definitions imply U
e
suppose x ∈ U ∩ U LU . We need to show xLU ⊆ LU U . Indeed, we have
e LU LU = U
e LU ⊆ LU U .
xLU ⊆ U
There are examples of automorphisms [23] and associated tidy below subgroups which do not behave well when passing to quotients. Lemma 8.6 shows
that although we can not expect a tidy below subgroup to be tidy below when
passing to a quotient, the original subgroup can be chosen such that the quotient is as close as possible to being tidy below using the tidying procedure.
The proof of Lemma 8.6 relies on the following result which is immediate from
the proof of [26, Lemma 16]. We state it for convenience.
Lemma 8.5: Let G be a totally disconnected locally compact group and α an
e where U is tidy above for α. Then for any
endomorphism of G. Suppose u ∈ U
e± .
pair u± ∈ U± with u = u+ u− we have u± ∈ U
Lemma 8.6: Let G be a totally disconnected, locally compact group, α an
endomorphism of G and H ⊆ G a closed normal subgroup with α(H) ⊆ H.
Denote by α the endomorphism induced by α on G/H and by q : G → G/H
the quotient map. Then there is a compact open subgroup U of G such that
(a) U tidy for α,
(b) U ∩ H is tidy for α|H , and
(c) q(U ) is tidy above for α and Lq(U) q(U ) = q(U )Lq(U) .
Proof. Applying Lemma 8.2, choose V ∈ COS(G) which is tidy for α and such
that V ∩ H is tidy for α|H . Then q(V ) is tidy above for α: On the one hand
\
\
\
α−n (q(V )) = q(V )− .
q(V− ) = q
α−n (V ) ⊆
q(α−n (V )) =
n≥0
n≥0
n≥0
ENDOMORPHISMS OF T.D.L.C. GROUPS
37
Also, V+ = {x ∈ V | there is an α-regressive trajectory for x in V } and therefore q(V+ ) ⊆ q(V )+ as α-regressive trajectories descend to the quotient. Combining the above we conclude
q(V ) = q(V+ V− ) = q(V+ )q(V− ) ⊆ q(V )+ q(V )− .
That is, q(V ) is tidy above for α. Now define U := V ∩ q −1 (q(V )e), where
q(V )e is as in Step 3 of the tidying procedure, 8.1. Then q(U ) = q(V )e and
hence q(U ) is tidy above for α by [26, Lemma 16]. In addition, by applying
[26, Proposition 6 (3)] we see that Lq(U) = Lq(V )e = Lq(V ) . It follows from
[26, Lemma 13] and q(U ) = q(V )e that q(U )Lq(U) = Lq(U) q(U ). Furthermore,
V ∩ H ⊆ ker q ⊆ q −1 (q(V )e). Hence
U ∩ H = V ∩ H ∩ q −1 (q(V )e) = V ∩ H
is tidy for α|H .
It remains to show that U is tidy for α. We begin by proving that U is tidy
above for α. Let u ∈ U . Then since V is tidy above, u = v+ v− for some
v± ∈ V± and we aim to show that v± ∈ U± . Note that q(u) = q(v+ )q(v− ) with
q(v± ) ∈ q(V± ) ⊆ q(V )± . Since q(u) ∈ q(V )e, we deduce q(v± ) ∈ (q(V )e)± by
applying Lemma 8.5. Since αn (v− ) ∈ V− and αn (q(v− )) ∈ (q(V )e)− for all n ≥ 0
we have q(αn (v− )) ∈ (q(V )e)− . Therefore, the orbit of v− ∈ V ∩ q −1 (q(V )e) =
U stays in U and we conclude v− ∈ U− .
As to v+ , choose an α-regressive trajectory (vi )i∈N0 for v+ contained in V+ .
We will show this sequence is contained within U . It is clear that v0 = v+ ∈ U .
Suppose for the purpose of induction that vn ∈ U . Applying [26, Lemma 15]
we see that q(vn ) ∈ q(U ) ∩ q(V+ ) ⊆ q(V )e ∩ q(V )+ = (q(V )e)+ . There exists
w ∈ (q(V )e)+ such that
α(w) = q(vn ) = α(q(vn+1 )).
Now w, q(vn ) and q(vn+1 ) are all elements of par− (α). By [26, Proposition 20],
there exists b ∈ bik(α) such that q(vn+1 ) = wb. Since q(V )eLq(V ) is tidy,
b ∈ q(V )eLq(V ) . Hence q(vn+1 ) ∈ q(V )eLq(V ) . By Lemma 8.4, q(vn+1 ) ∈ q(V )e
which implies vn+1 ∈ U . Inductively vi ∈ U for all i ∈ N0 and so v+ ∈ U+ .
To see that U is tidy below, note that V is tidy below and U ⊆ V . Hence
LU ⊆ V+ ∩ V− . Clearly, q(V+ ∩ V− ) ⊆ Lq(V ) and so q(V+ ∩ V− ) ⊆ q(V )e. Hence
V+ ∩ V− ⊆ U . As a consequence, LU ⊆ U which implies that U is tidy below,
see [26, Proposition 8].
38
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
In the following lemma, we factor the subgroup used to calculate the scale.
Later on, we turn this into a factorization of the scale itself.
Lemma 8.7: Let G be a totally disconnected, locally compact group, α an
endomorphism of G and H ⊆ G a closed normal subgroup with α(H) ⊆ H.
Denote by α the endomorphism induced by α on G/H. Then there is a closed
subgroup J of G with α((H ∩ U )+ )U+ ⊆ J ⊆ α(U+ ) and
sG/H (α) = [α(U+ ) : J].
Proof. Let U satisfy the conclusions of Lemma 8.6. Let q : G → G/H denote
the quotient map. Then q(U )Lq(U) is tidy for α and
sG/H (α) = [α(q(U )+ )Lq(U) : q(U )+ Lq(U) ]
using [26, Proposition 4 and Proposition 6 (2)]. Now consider the map
α(q(U )+ )/(α(q(U )+ ) ∩ q(U )+ Lq(U) ) → α(q(U )+ Lq(U) )/q(U )+ Lq(U)
given by
g(α(q(U )+ ) ∩ q(U )+ Lq(U) ) 7→ g(q(U )+ Lq(U) ).
This map is well-defined as α(q(U )+ ) ∩ q(U )+ Lq(U) ) ⊆ q(U )+ Lq(U) . It is injective because any two elements in the domain which have the same image have
coset representatives which differ by an element in α(q(U )+ ) ∩ q(U )+ Lq(U) .
To see surjectivity, simply note that α(Lq(U) ) ⊆ Lq(U) ⊆ q(U+ )Lq(U) by [26,
Lemma 6]. This shows
sG/H (α) =
(25)
=
[α(q(U )+ )Lq(U) : q(U )+ Lq(U) ]
[α(q(U )+ ) : α(q(U )+ ) ∩ q(U )+ Lq(U) ].
We know that α(q(U )+ ) ∩ q(U )+ Lq(U) is closed in G/H because α and q are
continuous, U is compact and Lq(U) is closed. Set
J := q −1 α(q(U )+ ) ∩ q(U )+ Lq(U) ∩ α(U+ ).
By the above, J ⊆ α(U+ ) is a closed subgroup. To see α((H ∩ U )+ )U+ ⊆ J,
note that
(26)
q(α((H ∩ U )+ )U+ ) = q(U+ ) ⊆ q(U )+ ⊆ α(q(U )+ ) ∩ q(U )+ Lq(U) =: S
because α((H ∩ U )+ )U+ = U+ α((H ∩ U )+ ) and α((H ∩ U )+ ) is contained in H.
The formula
x.(yS) := q(x)yS
for x ∈ α(U+ ) and y ∈ q(U+ )
ENDOMORPHISMS OF T.D.L.C. GROUPS
39
defines a left action of α(U+ ) on X := α(q(U+ ))/S which is transitive as
q(α(U+ )) = α(q(U+ )). Since S ∈ X has stabilizer q −1 (S) ∩ α(U+ ) = J under the action, the Orbit Formula (as in [17, 1.6.1 (i)]) shows that
[α(U+ ) : J] = |X| = [α(q(U+ )) : S].
Combining this with (26) and (25) we obtain sG/H (α) = [α(U+ ) : J].
Theorem 8.8: Let G be a totally disconnected, locally compact group, α an
endomorphism of G and H ⊆ G a closed normal subgroup with α(H) ⊆ H.
Denote by α the endomorphism induced by α on G/H. Then sH (α|H )sG/H (α)
divides sG (α).
Proof. Let U satisfy the conclusions of Lemma 8.6. By Lemma 8.7 there is a
closed subgroup J of G such that
U+ ⊆ α((U ∩ H)+ )U+ ⊆ J ⊆ α(U+ ).
Recall that by Proposition 8.3, the set α((U ∩ H)+ )U+ is indeed a subgroup of
G. Applying Lemma 8.7 and Proposition 8.3 yields
sG (α) = [α(U+ ) : U+ ]
= [α(U+ ) : J][J : α((U ∩ H)+ )U+ ][α((U ∩ H)+ )U+ : U+ ]
= sG/H (α)[J : α((U ∩ H)+ )U+ ]sH (α|H ).
which completes the proof.
We end this section by considering the special case of nested subgroups inside
par− (α) for which we achieve equality in Theorem 8.8.
Lemma 8.9: Let G be a a totally disconnected locally compact group, α be
an endomorphism of G and H ⊆ par− (α) a closed α-stable subgroup. Then
par− (α|H ) = H.
Proof. Suppose x ∈ H. We can find an α-regressive trajectory (x = x0 , x1 , . . .)
which is contained in some compact set K. Since α(H) = H we can choose
another α-regressive trajectory (x = y0 , y1 , . . .) such that yn ∈ H for all n ∈ N.
Now yn , xn ∈ par− (α) for all n ∈ N and hence x−1
n yn ∈ bik(α). This shows
yn ∈ xn bik(α) ⊆ K bik(α). Since both K and bik(α) are compact, K bik(α) is
compact and hence K bik(α) ∩ H is a compact subset of H. This shows that
(y0 , y1 , . . .) is bounded and hence x ∈ par− (α|H ).
40
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
The following result is known for automorphisms, see [8, Proposition 3.21 (2)].
In the case of automorphisms, the proof utilizes the modular function which is
not defined for endomorphisms. Instead we consider the factoring of the scale
given by Theorem 8.8.
Proposition 8.10: Suppose α is an endomorphism of a totally disconnected
locally compact group G and H ⊆ par− (α) is a closed α-stable subgroup. Also
suppose that N ⊆ H is a closed normal α-stable subgroup. Denote by α the
endomorphism induced by α|H on H/N . Then
sH (α|H ) = sH/N (α)sN (α|N ).
Proof. For simplicity, we write α for α|H as the enveloping group will play no
further role. By Lemma 8.9, par− (α) = H and so if U ∈ COS(H) is tidy for
α, then U = U+ by [26, Proposition 11]. By Lemma 8.2, we may assume that
U ∩ N is tidy for α|N . Let q : H → H/N denote the quotient map. Choose
U ∈ COS(H) satisfying conditions of Lemma 8.6 with respect to the normal
subgroup N . From the proof of Theorem 8.8 we have
sH (α) = sH/N (α)[J : α((U ∩ N )+ )U+ ]sN (α|N ),
where J is given in the proof of Lemma 8.6 by
J = q −1 (α(q(U )+ ) ∩ q(U )+ Lq(U) ) ∩ α(U+ ).
It suffices to show J ⊆ α((U ∩ N )+ )U+ . Since q(U+ ) ⊆ q(U )+ , as seen in the
proof of Lemma 8.6, and U+ = U we have q(U+ ) ⊆ q(U )+ ⊆ q(U ) = q(U+ ),
which gives equality throughout. Thus J = q −1 α(q(U )) ∩ q(U )Lq(U) ∩ α(U ).
Since q(U ) is an open identity neighbourhood, we obtain
q(U )Lq(U) = q(U )Lq(U) = q(U )Lq(U) .
Suppose that x ∈ q −1 (q(U )Lq(U) ). We can write x = ul where u ∈ U and
l ∈ q −1 (Lq(U) ). Consider q(l) = lN ∈ Lq(U) . There exists n ∈ N with
αn (lN ) = αn (l)N ∈ q(U ).
This implies αn (l)m ∈ U for some m ∈ N . Then αn (l)m has an α-regressive
trajectory contained in U = U+ . Using that fact that N is assumed to be αstable, choose m′ ∈ N such that αn (m′ ) = m. Since [26, Proposition 20] implies
that any two elements in the preimage of an element of par− (α) = H are equal
modulo bik(α), we have lm′ ∈ U bik(α) by comparing αn (lm′ ) = αn (l)m with
ENDOMORPHISMS OF T.D.L.C. GROUPS
41
the α-regressive trajectory for αn (l)m contained in U . But U is tidy and so
bik(α) ⊆ U . Hence l ∈ U N and thus x ∈ U N . This shows that J ⊂ U N ∩α(U ).
Suppose now that x ∈ U N ∩ α(U ). Then we can write x = un where u ∈ U
and n ∈ N . Choose α-regressive trajectories
(u = u0 , u1 , . . .), (un = v0 , v1 , . . .), and (n = n0 , n1 , . . .)
such that ui , vi+1 ∈ U for all i ≥ 0 and ni ∈ N for all i ∈ N. Note that
(un = u0 n0 , u1 n1 , . . .) is also an α-regressive trajectory. For all i ≥ 1 we have
ui ni ∈ vi bik(α). Noting that bik(α) ⊆ U , we have ni ∈ U for all i ≥ 1. Then
n1 ∈ (U ∩ N )+ and so n = n0 = α(n1 ) ∈ α((U ∩ N )+ ). As x = un, this shows
x ∈ U α((U ∩ N )+ ) = α((U ∩ N )+ )U (with equality by Proposition 8.3).
9. Small tidy subgroups and Theorem D
We now turn to Theorem D which characterizes admitting small tidy subgroups
for an endomorphism in terms of its contraction group. The following link between (anti)-contraction groups and the subgroups of a compact open subgroup
associated to tidiness is known for automorphisms (cf. [1, Proposition 3.16]).
Lemma 9.1: Let α be an endomorphism of a totally disconnected, locally compact group G, and V ⊆ G be a compact open subgroup. Then
V−− = con(α, V+ ∩ V− )
and V++ = con− (α, V+ ∩ V− ).
Proof. If x ∈ con(α, V+ ∩ V− ), then there exists n ∈ N0 such that αm (x) ∈
V (V+ ∩ V− ) = V for all m ≥ n. Thus αk (αn (x)) ∈ V for all k ∈ N0 , whence
αn (x) ∈ V− and hence x ∈ α−n (V− ) ⊆ V−− . Thus con(α, V+ ∩ V− ) ⊆ V−−
If x ∈ V−− , there exists m ∈ N0 such that αm (x) ∈ V− and hence αn (x) =
αn−m (αm (x)) ∈ V for all n ≥ m. Thus αn (x) → e modulo V . By Lemma 2.12,
also αn (x) → e modulo V+ ∩ V− .
If x ∈ con− (α, V+ ∩ V− ), let (xn )n∈N0 be an α-regressive trajectory for x such
that xn → e modulo V+ ∩V− . There exists n ∈ N0 such that xm ∈ V (V+ ∩V− ) =
V for all m ≥ n. Then (xn+m )m∈N0 is an α-regressive trajectory for xn in V ,
whence xn ∈ V+ and x0 = αn (xn ) ∈ αn (V+ ) ⊆ V++ .
If x ∈ V++ , then x = αm (y) for some m ∈ N0 and y ∈ V+ . Let (yn )n∈N0 be an
42
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
α-regressive trajectory for y in V . Then
x = αm (y), αm−1 (y), . . . , α(y), y = y0 , y1 , y2 , . . .
is an α-regressive trajectory for x which eventually lies in V and hence converges
to e modulo V . As it also converges to e modulo V+ ∩ V− (by Lemma 2.12), we
see that x ∈ con− (α, V+ ∩ V− ).
Lemma 9.2: Let α be an endomorphism of a totally disconnected, locally compact group G and V ⊆ G by a tidy subgroup for α. Then
con(α) ⊆ con(α, V+ ∩ V− ) and con− (α) ⊆ con− (α, V+ ∩ V− ).
Proof. Since V is tidy, both V−− = con(α, V+ ∩V− ) and V++ = con− (α, V+ ∩V− )
are closed in G (with equalities by Lemma 9.1). So con(α) ⊆ con(α, V+ ∩ V− )
and con− (α) ⊆ con− (α, V+ ∩ V− ).
If α is an endomorphism of a totally disconnected locally compact group, we
write TID(α) for the set of all tidy subgroups of α. The following auxiliary
result will be improved in Proposition 12.2 (d).
Lemma 9.3: Let α be an endomorphism of a totally disconnected, locally compact group G. Then con(α) ⊆ con(α) nub(α).
Proof. For each V ∈ TID(α), we have con(α) ⊆ con(α, V+ ∩ V− ) ⊆ con(α, V ).
and therefore
\
\
V
con(α, V ) = con α,
con(α) ⊆
V ∈TID(α)
=
V ∈TID(α)
con(α, nub(α)) = con(α) nub(α),
using Lemma 2.11 for the first equality and Theorem A for the last.
The following result is known for automorphisms (cf. [1, Lemma 3.31 (3)]).
Proposition 9.4: Let α be an endomorphism of a totally disconnected, locally
compact group G and V ⊆ G be a compact open subgroup which is tidy above
for α. Then V is tidy if and only if V contains the nub of α.
Proof. If V is tidy, then V ⊇ nub(α) by definition of the nub. For the converse,
assume that V is tidy above and V ⊇ nub(α). Since W1 ∩ W2 is tidy for all tidy
subgroups W1 and W2 (see [26, Proposition 12]), TID(α) is a filter basis. By
ENDOMORPHISMS OF T.D.L.C. GROUPS
43
Lemma 2.1, there exists W ∈ TID(α) such that W ⊆ V . Using both Lemma 9.1
and Theorem A twice, we see that
V−−
=
con(α, V+ ∩ V− ) = con(α)(V+ ∩ V− ) = con(α)(W+ ∩ W− )(V+ ∩ V− )
=
con(α, W+ ∩ W− )(V+ ∩ V− ) = W−− (V+ ∩ V− ).
As W−− is closed and V+ ∩ V− is compact, V−− is closed. Hence V is tidy.
The following proposition subsumes Theorem D. For the special case of automorphisms, see already [1, Theorem 3.32] (for metrizable G).
Proposition 9.5: Let α be an endomorphism of a totally disconnected, locally
compact group G. Then the following conditions are equivalent:
(a)
(b)
(c)
(d)
α has small tidy subgroups.
nub(α) = {e}.
Every compact open subgroup V ⊆ G which is tidy above for α is tidy.
con(α) is closed in G.
Proof. The implication “(a)⇒(b)” is trivial, while “(b)⇒(c)” is immediate from
Proposition 9.4. If (c) holds and U ⊆ G is an identity neighbourhood, pick a
T
compact open subgroup V ⊆ U . After replacing V with ℓj=0 α−j (V ) for some
ℓ ∈ N0 , we may assume that V is tidy above for α and hence tidy, as we
assume (c). Hence (c) implies (a).
(b)⇒(d) holds by Lemma 9.3.
(d)⇒(c): By Lemma 9.1 and Theorem A, we have V−− = con(α, V+ ∩ V− ) =
con(α)(V+ ∩ V− ), which is closed in G as V+ ∩ V− is compact and we assume
that con(α) is closed. Hence V is tidy.
Tidy automorphisms were introduced in [6].
Definition 9.6: An endomorphism α of a totally disconnected, locally compact group is called tidy if it satisfies one (hence any) of the conditions (a)–(d)
in Proposition 9.5.
Corollary 9.7: Let α be an endomorphism of a totally disconnected, locally
compact group G and H ⊆ G an α-invariant closed subgroup. If α is tidy, then
also α|H is tidy. If H is also normal and compact, then the endomorphism α
induced by α on G/H is tidy.
44
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Proof. If α is tidy, then con(α) is closed in G by Proposition 9.5, whence
con(α)∩H = con(α|H ) is closed in H. Thus α|H is tidy by the same proposition.
Assume also H is normal in G and compact. Applying Theorem A, we have
q −1 (con(α)) = con(α, H) = con(α)H which is closed since con(α) is closed by
Proposition 9.5 and H is compact. Hence con(α) is closed and thus α is tidy,
by Proposition 9.5.
10. Dynamics on par− (α) and lev(α)
We now prove results involving the restriction of an endomorphism
α : G → G to the anti-parabolic subgroup par− (α) or the Levi subgroup lev(α) =
par(α) ∩ par− (α), to be used in the proof of Theorem E.
Lemma 10.1: Let α be an endomorphism of a totally disconnected, locally
compact group G and K be a compact subgroup of G such that α(K) = K.
Set β := α|par− (α) . Then K ⊆ par− (α) and
con− (α, K) = con− (β, K) ⊆ par− (α).
In particular, con− (α) = con− (β) ⊆ par− (α).
Proof. Since α(K) = K, we find an α-regressive trajectory in K for each
x ∈ K, showing that K ⊆ par− (α). If x ∈ con− (α, K), then there exists
an α-regressive trajectory (xn )n∈N0 for x such that xn → e modulo K. As
K is compact, {xn : n ∈ N0 } is relatively compact (see Lemma 2.9 (a)), entailing that xn ∈ par− (α) for each n ∈ N0 and thus x ∈ con− (β, K) (cf.
Lemma 2.14). Thus con− (α, K) ⊆ con− (β, K). The converse inclusion is trivial
(cf. Lemma 2.13 (a)).
Among other things, the next lemma is useful for the proof of Theorem E.
Independently, part (c) of the lemma was also obtained in [5, Lemma 4.5 (3)].
Lemma 10.2: Let α be an endomorphism of a totally disconnected, locally
compact group G. Then the following holds:
(a) A compact open subgroup U ⊆ G is tidy below for α if and only if
U ∩ par− (α) is tidy below for α|par− (α) .
(b) If α|par− (α) is injective, then α|par− (α) is an automorphism of par− (α).
If U ⊆ G is a tidy subgroup for α, then
ENDOMORPHISMS OF T.D.L.C. GROUPS
45
(c) U+ is a compact open subgroup of par− (α) which is tidy for α|par− (α) ,
and s(α) = s(α|par− (α) ).
(d) U+ is a compact open subgroup of U++ which is tidy for α|U++ , and
s(α) = s(α|U++ ).
Proof. (a) The subgroup V := U ∩ par− (α) is compact and open in P − :=
par− (α). Let β := α|P − and V+ be the group of all x ∈ V for which there exists
a β-regressive trajectory in V . If x ∈ U+ , then there exists an α-regressive
trajectory (xn )n∈N0 for x in U . Since {xn : n ∈ N0 } ⊆ U is relatively compact,
we have xn ∈ U ∩ P − = V for all n ∈ N0 , whence x ∈ V+ . Thus U+ ⊆ V+ and
(27)
U+ = V+ .
If x ∈ U++ , then there exists an α-regressive trajectory (xn )n∈N0 for x and
N ∈ N0 such that xn ∈ U+ = V+ for all n ≥ N . Again, we deduce that
xn ∈ P − for all n ∈ N0 . This entails x ∈ V++ . Thus U++ ⊆ V++ and hence
(28)
U++ = V++ .
Since P − ⊆ G is closed, U++ is closed in G if and only if V++ is closed in P − .
As U+ = V+ ⊆ P − , we have αn (U+ ) = β n (V+ ) for each n ∈ N0 and thus
(29)
[αn+1 (U+ ) : αn (U+ )] = [β n+1 (V+ ) : β n (V+ )] for all n ∈ N0 .
As a consequence, the sequence ([αn+1 (U+ ) : αn (U+ )])n∈N0 is constant if and
only if ([β n+1 (V+ ) : β n (V+ )])n∈N0 is a constant sequence. Thus U is tidy below
for α if and only if V is tidy below for β.
(c) Since U+ = P − ∩ U (see [26, Proposition 11(1)]), U+ is open in P − . As
U+ ⊆ α(U+ ), the subgroup V := U+ = P − ∩ U is tidy above for β := α|P −
(since V ⊆ β(V ) implies V+ = V ). By (a), V is also tidy below. Therefore V is
tidy for β and now (29) yields
s(α) = [α(U+ ) : U+ ] = [β(V+ ) : V+ ] = s(β).
(d) Since U+ = U++ ∩ U by [26, Proposition 11(1)] and U++ is closed, U+ is a
compact open subgroup of U++ . It is tidy above for α|U++ since α(U+ ) ⊇ U+
and tidy below since U++ and the indices [αn+1 (U+ ) : αn (U+ )] for n ∈ N0 are
unchanged if α is replaced with α|U++ . Notably, s(α) = [α(U+ ) : U+ ] remains
unchanged.
(b) The self-map β := α|P − of P − is surjective as α(P − ) = P − , and injective
46
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
by hypothesis. By the proof of (c), P − has a compact open subgroup V such
that V ⊆ β(V ), whence also β(V ) is a compact open subgroup of P − . Since V
is compact, β|V is a quotient homomorphism onto its image and hence open
onto β(V ), which is open in P − . Thus, the endomorphism β is an open map.
Remark 10.3: Note that α|par− (α) is injective if and only if bik(α) = {e}.
Hence α|par− (α) is injective if α is tidy (then bik(α) ⊆ nub(α) = {e}).
The following result is known for automorphisms (cf. [1, Theorem 3.32]).
Proposition 10.4: Let α be an endomorphism of a totally disconnected, locally compact group G. If con(α) is closed in G, then also con− (α) closed.
Proof. As con(α) is closed, β := α|par− (α) is an automorphism (see Remark 10.3
and Proposition 9.5). Since W+ is tidy for β for each compact open subgroup
W ⊆ G which is tidy for α (see Lemma 10.2 (c)), β has small tidy subgroups.
As each of these is also tidy for β −1 , Proposition 9.5 shows that con(β −1 ) is
closed in par− (α) and hence in G. Since con− (α) = con− (β) = con(β −1 ) by
Lemma 10.1, the assertion follows.
Remark 10.5: If con− (α) is closed, then con(α) need not be closed. For example, let F 6= {e} be a finite group and G := F N with the product topology.
Consider the left shift
α : G → G,
(xn )n∈N 7→ (xn+1 )n∈N
with kernel F × {e}{2,3,...} . Then con(α) is the dense proper subgroup of all
finitely supported sequences (hence not closed). Let
σ : G → G,
x = (xn )n∈N 7→ (e, x1 , x2 , . . .)
be the right shift. If x ∈ G, then (σ n (x))n∈N0 is an α-regressive trajectory for
x such that σ n (x) → e, whence x ∈ con− (α). Thus con− (α) = G is closed.
We mention that the proof of Lemma 3.31 in [1] (and hence [1, Theorem 3.32])
uses a tidying procedure for compact open subgroups which was developed in
[24] and is different from the one of Section 8.1. It involves a certain compact
subgroup K ⊆ G which, following [24, §2, Step 2a, p. 4], is defined as
\
KO ,
(30)
K :=
O∈COS(G)
ENDOMORPHISMS OF T.D.L.C. GROUPS
47
where KO := lev(α) ∩ O−− (in [24], the abbreviation KO := lev(α) ∩ O−− is
used). If G is metrizable, then K admits a simpler description, as the closure
K = con(α) ∩ lev(α)
(31)
(i.e., K = con(α) ∩ par− (α)); this was asserted without proof in [24], and left
open for non-metrizable G. It was later shown by W. Jaworski (unpublished,
personal communication, 2008). We now give a shorter argument.
Proposition 10.6: Let α be an endomorphism of a totally disconnected, locally compact group G. Then
\
(32)
KO .
con(α) ∩ lev(α) =
O∈COS(G)
Proof. For O ∈ COS(G), we have O+ ∩ O− ⊆ lev(α) and thus
(33)
O−− ∩ lev(α) = con(α, O+ ∩ O− ) ∩ lev(α) = con(α|lev(α) , O+ ∩ O− ),
using Lemma 9.1 for the first equality and then Lemma 2.15 (a). Hence
\
con(α|lev(α) , O+ ∩ O− )
K ⊇
O∈COS(G)
=
con α|lev(α) ,
\
O∈COS(O)
using Lemma 2.11 and the fact that
T
(O+ ∩ O− ) = con(α|lev(α) )
O∈COS(G) (O+
∩ O− ) ⊆
T
O∈COS(G) O
=
{e}. So K ⊇ con(α|lev(α) ) = con(α) ∩ lev(α), with equality by Lemma 2.15 (a).
To verify the converse inclusion, note that
KO = con(α|lev(α) , O+ ∩ O− ) = con(α|lev(α) )(O+ ∩ O− ) = con(α|lev(α) )(O+ ∩O− )
by (33), Theorem A and compactness of O+ ∩ O− . If
\
\
KO =
x∈K=
con(α|lev(α) )(O+ ∩ O− ),
O∈COS(G)
O∈COS(G)
for each O ∈ COS(G) we can write
x = xO yO
with xO ∈ con(α|lev(α) ) and yO ∈ O+ ∩ O− . Direct COS(G) by declaring
O1 ≤ O2 if O2 ⊆ O1 . Then the net (yO )O∈COS(G) converges to e, whence
−1
xO = xyO
→ x.
48
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Thus x ∈ con(α|lev(α) ), showing that K ⊆ con(α|lev(α) ) = con(α) ∩ lev(α).
We shall see later that K = nub(α) (Lemma 12.2 (e)).
11. Proof of Theorem E
We shall use a lemma concerning induced endomorphisms on quotients. For a
precursor of (a) and (b) for (inner) automorphisms, compare [7, Lemma 3.6]
and its variant in [16, Lemma 4.9 (iv)].
Lemma 11.1: Let α be an endomorphism of a totally disconnected, locally
compact group G and N ⊆ G be an α-stable compact normal subgroup. Let
q : G → G/N be the canonical quotient map and α be the endomorphism of
G/N induced by α. Then the following holds:
(a) Let U ⊆ G be a compact open subgroup such that N ⊆ U . Then U is
tidy for α if and only if q(U ) is tidy for α.
(b) s(α) = s(α).
(c) If N ⊆ nub(α), then q(nub(α)) = nub(α) and nub(α) = q −1 (nub(α)).
Proof. (a) For x ∈ U and n ∈ N0 , we have αn (q(x)) = q(αn (x)) ∈ q(U ) if and
only if αn (x) ∈ U N = U . Hence q(x) ∈ q(U )− if and only if x ∈ U− , entailing
U− = q −1 (q(U )− ) and hence also q(U− ) = q(U )− .
If x ∈ U+ , then there exists an α-regressive trajectory (xn )n∈N0 for x in U .
Then (q(xn ))n∈N0 is an α-regressive trajectory for q(x) in q(U ) and hence q(x) ∈
q(U )+ . Thus
(34)
q(U+ ) ⊆ q(U )+ .
If q(x) ∈ q(U )+ and (yn )n∈N0 is an α-regressive trajectory for q(x) in q(U ), pick
zn ∈ q −1 ({yn }) for n ∈ N0 . Then
q(αi (zn )) = αi (q(zn )) = αi (yn ) = yn−i ∈ q(U )
for i ∈ {0, 1, . . . , n}, showing that αi (zn ) ∈ q −1 (q(U )) = U . Since q(αn (zn )) =
y0 = q(x), we have xαn (zn )−1 ∈ ker(q) = N . Since α(N ) = N , we can choose
a ∈ N such that αn (a) = xαn (zn )−1 . Then azn ∈ U , αi (azn ) = αi (a)αi (zn ) ∈
N U = U for i ∈ {0, . . . , n} and αn (azn ) = αn (a)αn (zn ) = x. Hence x ∈ Un,α
ENDOMORPHISMS OF T.D.L.C. GROUPS
and thus
x∈
\
49
Un,α = U+ .
n∈N0
Thus q −1 (q(U )+ ) ⊆ U+ . Together with (34), this implies U+ = q −1 (q(U )+ )
and q(U+ ) = q(U )+ .
Since U = U N and
U+ U− = U+ U− N,
the final equality in the chain q(U )+ q(U )− = q(U+ )q(U− ) = q(U+ U− ) = q(U )
holds if and only if U+ U− = U . Thus U is tidy above for α if and only if q(U )
is tidy above for α. We have
q(αn (U+ )) = αn (q(U+ )) = αn (q(U )+ )
for each n ∈ N0 . Since αn (U+ ) ⊇ U+ = q −1 (q(U )+ ) ⊇ N , we deduce that
αn (U+ ) = q −1 (αn (q(U )+ )).
Hence
U++
=
[
αn (U+ ) =
n∈N0
n∈N0
= q
−1
[
[
n∈N0
n
q −1 (αn (q(U )+ ))
!
α (q(U )+ )
= q −1 (q(U )++ ),
showing that U++ is closed if and only if q(U )++ is closed. Finally,
[αn+1 (U+ ) : αn (U+ )]
(35)
= [q −1 (αn+1 (q(U )+ )) : q −1 (αn (q(U )+ ))]
= [αn+1 (q(U+ )) : αn (q(U+ ))],
whence the sequence ([αn+1 (U+ ) : αn (U+ )])n∈N0 is constant if and only if so is
([αn+1 (q(U )+ ) : αn (q(U )+ )])n∈N0 . Thus U is tidy if and only if q(U ) is tidy.
(b) Let V ⊆ G/N be a compact open subgroup which is tidy for α. Then
U := q −1 (V ) is a compact open subgroup of G such that N ⊆ U and q(U ) = V
is tidy, whence U is tidy, by (a). Using (35), we get
s(α) = [α(U+ ) : U+ ] = [α(q(U )+ ), q(U )+ ] = [α(V+ ) : V+ ] = s(α).
(c) As N ⊆ nub(α) ⊆ U for each U ∈ TID(α), we deduce from (a) that
TID(α) = {q −1 (V ) : V ∈ TID(α)}.
50
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Hence
nub(α) =
\
V ∈TID(α)
q −1 (V ) = q −1
\
V ∈TID(α)
and thus q(nub(α)) = nub(α).
V = q −1 (nub(α))
Proof. (Theorem E). Let β := α|par− (α) and α be the automorphism induced
by β on par− (α)/ bik(α). Let q : par− (α) → par− (α)/ bik(α) be the canonical
quotient map, whose kernel bik(α) is compact. Then
(36)
s(α) = s(β) = s(α) = s(α|con− (α) ),
using Lemma 10.2 (c), Lemma 11.1 (b) or Proposition 8.10, as well as [1, Proposition 3.21]. The first and second equality were obtained independently in [5,
Lemma 4.5 (3), (4)]. We have con− (β, bik(α)) = con− (β) bik(α) by Theorem B
and thus q(con− (α)) = q(con− (β)) = con− (α). Since bik(α) is compact, we
deduce that
con− (α) bik(α)
is closed. It then coincides with the set con− (α) bik(α), which is mapped onto
con− (α) by q. Hence
p := q|con− (α) : con− (α) → con− (α)
is a continuous surjective homomorphism and hence a quotient morphism, as
both groups are σ-compact (whence the open mapping theorem, [11, (5.29)]
applies): In fact, as the left hand side is a closed subgroup of con− (α) bik(α)
and this group has con− (α) as a quotient with compact kernel, it will be σcompact if con− (α) is σ-compact, which holds by Lemma 11.2 to come.
Since p has compact kernel, by applying Lemma 11.1 (b) or Proposition 8.10
we deduce that s(α|con− (α) ) = s(α|con− (α) ). Together with (36), we get s(α) =
s(α|con− (α) ).
We applied the following lemma to the automorphism α −1 :
Lemma 11.2: For each automorphism α : G → G of a locally compact group,
the closure con(α) of the contraction group is σ-compact.
Proof. Let K be a compact identity neighbourhood in con(α). Then
con(α) = con(α)K.
ENDOMORPHISMS OF T.D.L.C. GROUPS
51
For each x ∈ con(α), there is n ∈ N with αn (x) ∈ K whence x ∈ α−n (K). Thus
[
con(α) =
α−n (K)K
n∈N
is σ-compact.
As an application of Theorem B, Proposition 8.10 and Theorem E, we include the following addition to the study begun in Section 8. Note that since
α-stable compact groups are contained within par− (α), this result is a generalization of Lemma 11.1 (b). It also generalizes Proposition 8.10 which can be
seen by choosing G = par− (α). As a consequence, the result is also a generalization of [1, Proposition 3.21 (2)] which is precisely Proposition 8.10 restricted
to automorphisms.
Theorem 11.3: Let G be a totally disconnected locally compact group and α
an endomorphism of G. Further, let N ⊆ G be a closed normal subgroup which
is α-stable and contained in par− (α). If α is the endomorphism induced by α
on G/N , then
sG (α) = sG/N (ᾱ)sN (α).
Proof. Applying Lemma 10.2 and Proposition 8.10, it suffices to show
sG/N (α) = spar− (α)/N (α),
where the right hand side denotes the scale of the restriction of α to par− (α)/N .
Let xN ∈ con− (α). Then x ∈ con− (α, N ). Thus x ∈ con− (α)N by Theorem B.
Therefore, xN ∈ par− (α)/N as con− (α) ⊆ par− (α). Since par− (α) is closed
we have con− (α) ⊆ par− (α)/N.
By Theorem E we have s(α) = s(α|con− (α) ), so applying Proposition 8.3 we
get sG/N (α) = scon− (α) (α) ≤ spar− (α)/N (α) ≤ sG/N (α) as required.
12. Further results concerning the nub
In this section, we provide further results concerning the nub of an endomorphism, as well as consequences, for example an improved description of the
closure of a contraction group.
Lemma 12.1: If α is an endomorphism of a totally disconnected, locally compact group G, then
lev(α) = lev(α|par− (α) )
and
bik(α) = bik(α|lev(α) ).
52
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
Proof. Abbreviate γ := α|lev(α) . If x ∈ par− (α) and αn (x) = e for some n, then
x ∈ lev(α). Hence
{x ∈ par− (α) : (∃n ∈ N0 ) αn (x) = e} = {x ∈ lev(α) : (∃n ∈ N0 ) γ n (x) = e}.
Since lev(α) is closed, the closures of the preceding set in G and lev(α) coincide.
Thus bik(α) = bik(γ).
If x ∈ lev(α), there exists a bounded α-trajectory (xn )n∈Z with x0 = x. Then
xn ∈ lev(α) for each n, since (xn+m )m∈Z is a bounded α-trajectory for this
element. Thus (xn )n∈Z is a bounded γ-trajectory for x in lev(α) (and hence
also a bounded α|par− (α) -trajectory for x in par− (α)). Thus x ∈ lev(α|par− (α) )
and hence lev(α) ⊆ lev(α|par− (α) ). The converse inclusion is trivial.
The following statements (a) and (b) are implicit in comments concerning the
nub in [26, lines following Remark 4] (where no proof is given); independently,
they were verified in [5, Lemma 4.5 (5)].
Proposition 12.2: Let α be an endomorphism of a totally disconnected, locally compact group G. Let q : par− (α) → par− (α)/ bik(α) =: Q be the canonical quotient map and α be the automorphism of Q induced by α|par− (α) . Then
(a)
(b)
(c)
(d)
nub(α|par− (α) ) = nub(α),
nub(α) = q(nub(α)),
nub(α) = con(α) ∩ nub(α),
T
con(α) = U∈TID(α) U−− = con(α) nub(α), and
T
(e) nub(α) = con(α) ∩ lev(α) = U∈COS(G) lev(α) ∩ U−−
T
T
= lev(α) ∩ U∈TID(α) U−− = V ∈COS(lev(α)) V−− .
Proof. (a) Abbreviate P − := par− (α). We recall from [26, Corollary 5] that
T
T
nub(α) = U∈TID(α) U = U∈TID(α) (U+ ∩ U− ), whence also
(37)
nub(α) =
\
U+ .
U∈TID(α)
The right hand side of (37) contains nub(α|P − ), by Lemma 10.2 (c). Hence
nub(α) ⊇ nub(α|P − ). For the reverse inclusion we show the contrapositive.
Suppose x 6∈ nub(α|P − ). Then x 6∈ W for some compact open subgroup W ⊆
P − which is tidy for α|P − . Let O be a compact open subgroup of G with
ENDOMORPHISMS OF T.D.L.C. GROUPS
O ∩ P − ⊆ W . Then
V :=
\
53
yOy −1
y∈W
is a compact open subgroup of G which is normalized by W , entailing that V W
is a compact open subgroup of G. Note that x 6∈ V W (if we could write x = vw
with v ∈ V and w ∈ W , then v = xw−1 ∈ V ∩ P − ⊆ O ∩ P − ⊆ W , whence
x = vw ∈ W contrary to the choice of W ). For large ℓ ∈ N0 ,
U :=
ℓ
\
α−j (V W )
j=0
is tidy above for α. By Lemma 3.1, after increasing ℓ if necessary, we may
assume that, moreover, P − ∩ U is tidy above for α|P − . Since nub(α|P − ) ⊆ W
and α(nub(α|P − )) = nub(α|P − ), we have nub(α|P − ) ⊆ U and hence
nub(α|P − ) ⊆ U ∩ P − .
As a consequence, U ∩ P − is tidy for α|P − , whence U is tidy below for α (by
Lemma 10.2 (a)) and hence tidy. Thus x 6∈ nub(α) ⊆ U . This shows that
nub(α) = nub(α|P − ).
(b) As ker(q) = bik(α) ⊆ nub(α) = nub(α|P − ) (using (a)), we have q(nub(α)) =
q(nub(α|P − )) = nub(α) by Lemma 11.1 (c).
(c) By (b), the map p := q|nub(α) : nub(α) → nub(α) is onto and thus a quotient
map by the open mapping theorem [11, (5.29)], as nub(α) is compact. Hence
q(con(α) ∩ nub(α)) = q(con(α|nub(α )) = con(α|nub(α) ),
by Theorem A, which is dense in nub(α) by [1, Lemma 3.31, (1) and (2)]. As a
consequence,
(38)
con(α|nub(α) ) ker(p) = con(α|nub(α) ) bik(α)
is dense in nub(α). The subset
{x ∈ par− (α) : (∃n ∈ N0 ) αn (x) = e}
⊆ bik(α) ∩ con(α) ⊆ nub(α) ∩ con(α)
= con(α|nub(α) )
is dense in bik(α). Combining this with the density of the set from (38) in
nub(α), we find that con(α|nub(α) ) is dense in nub(α) (as asserted).
(d) We know from Lemma 9.3 that con(α) = con(α)(con(α) ∩ nub(α)). By (c),
54
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
con(α) ∩ nub(α) is dense in nub(α), and thus con(α) ∩ nub(α) = nub(α). Hence
con(α) = con(α) nub(α). It only remains to recall that
\
U−−
=
\
con(α, U+ ∩ U− ) =
=
con(α, U )
U∈TID(α)
U∈TID(α)
U∈TID(α)
\
con(α, nub(α)) = con(α) nub(α)
by Lemma 9.1, Lemma 2.12, Lemma 2.11, and Theorem A.
(e) As {U ∩lev(α) : U ∈ COS(G)} is a basis of identity neighbourhoods in lev(α)
T
and lev(α) ∩ U−− = (lev(α) ∩ U )−− , we have U∈COS(G) lev(α) ∩ U−− =
T
V ∈COS(lev(α)) V−− . The second and third groups in (e) (from the left) coincide by Lemma 10.6, and we denote them by K (as in (30)). Now K =
con(α) ∩ lev(α) ⊇ con(α) ∩ nub(α) = nub(α), by Proposition 12.2 (c). On the
other hand,
K=
\
U∈COS(G)
lev(α) ∩ U−− ⊆
\
lev(α) ∩ U−− ⊆ nub(α)
U∈TID(α)
as each of the sets lev(α) ∩ U−− is closed, and contained in U+ ∩ U− , as
we now verify: If x ∈ lev(α) ∩ U−− , then x ∈ lev(α) and αn (x) ∈ U− for
some n ∈ N0 . Let (xn )n∈N0 be a bounded α-regressive trajectory for x. Then
αn (x), αn−1 (x), . . . , α(x), x, x1 , x2 , . . . is a bounded α-regressive trajectory for
αn (x) ∈ U , whence αn (x) ∈ U+ by [26, Proposition 11 (a)] and thus αn (x) ∈
U+ ∩ U− . Since α(U+ ∩ U− ) = U+ ∩ U− , there exists and α-regressive trajectory
(yj )j∈N0 for αn (x) inside U+ ∩ U− . Then yn−1 x ∈ (U+ ∩ U− ) lev(α) ⊆ lev(α) ⊆
par− (α) and αn (yn−1 x) = e, showing that yn−1 x ∈ bik(α) ⊆ nub(α). Hence
x = yn (yn−1 x) ∈ (U+ ∩ U− ) nub(α) = U+ ∩ U− .
Thus lev(α) ∩ U−− ⊆ U+ ∩ U− and thus K ⊆
T
U∈TID(α) U+ ∩ U− = nub(α).
T
U∈TID(α) (lev(α)
∩ U−− ) ⊆
Proposition 12.2 (e) provides several characterizations of the nub, and identifies it as the group considered in Proposition 10.6 (which had been used in the
literature in the case of automorphisms, as recalled above). A complementary
characterization of the nub was obtained in [5, Corollary 4.6].
ENDOMORPHISMS OF T.D.L.C. GROUPS
55
13. Theorem F: Dynamics on the ‘big cell’
Before we prove Theorem F, let us collect some tools. For automorphisms, part
(a), (d), (e) and (f) of the following lemma are known (cf. Proposition 3.4,
Corollary 3.17, and Lemma 3.18 in [1]).
Lemma 13.1: Let α be an endomorphism of a totally disconnected, locally
compact group G. Then the following holds:
(a) con(α) is a normal subgroup of par(α) and con− (α) is normal in par− (α).
(b) V− = con(α|V− )(V+ ∩ V− ) for each compact open subgroup V ⊆ G, and
con(α|V− ) is normal in V− . If V is tidy for α, then V+ ∩ V− = lev(α) ∩ V
and con(α|V − ) = con(α) ∩ V .
(c) V+ = (con− (α)∩V+ )(V+ ∩V− ) for each compact open subgroup V ⊆ G,
and con− (α) ∩ V+ is normal in V+ . If V is tidy, then con− (α) ∩ V+ =
con− (α) ∩ V .
(d) par(α) = con(α) lev(α). Moreover, the surjective homomorphism
q : lev(α) → par(α)/con(α), g 7→ g con(α) is continuous and open.
(e) par− (α) = con− (α) lev(α). Moreover, the surjective homomorphism
p : lev(α) → par− (α)/con− (α), g 7→ g con− (α) is continuous and open.
(f) A compact open subgroup V ⊆ lev(α) is tidy for α|lev(α) if and only if
α(V ) = V .
Proof. (a) If (xn )n∈N0 and (gn )n∈N0 are sequences in G such that xn → e and
the set {gn : n ∈ N0 } is relatively compact, then also gn xn gn−1 → e by virtue of
[11, (4.9)]. The assertions follow from this observation.
(b) If x ∈ V− ⊆ V−− , then αn (x) → e modulo V+ ∩ V− in G (see Lemma 9.1)
and hence also in V− (see Lemma 2.14). Thus, using Theorem A,
V− = con(α|V− , V+ ∩ V− ) = con(α|V− )(V+ ∩ V− ).
By (a), con(α|V− ) is normal in par(α|V− ) = V− . If V is tidy and x ∈ con(α)∩V ,
then {αn (x) : n ∈ N0 } is relatively compact and thus x ∈ V− (by [26, Proposition 11 (2)]). Hence x ∈ con(α|V− ) and so con(α) ∩ V ⊆ con(α|V− ). The
converse inclusion is trivial.
Finally, V+ ∩ V− = lev(α) ∩ V by [26, Proposition 11].
(c) As con− (α) is normal in par− (α), the intersection con− (α) ∩ V+ is normal
in the subgroup V+ of par− (α). If x ∈ V+ ⊆ V++ , then x ∈ con− (α, V+ ∩ V− )
56
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
by Lemma 9.1 and thus x ∈ (con− (α) ∩ V+ )(V+ ∩ V− ), by Theorem B. Trivially,
(con− (α) ∩ V+ )(V+ ∩ V− ) ⊆ V+ . If V is tidy, then con− (α) ∩ V ⊆ V+ (by [26,
Proposition 11 (1)]), whence con− (α) ∩ V+ = con− (α) ∩ V .
(d) If x ∈ par(α), then the closure K := {αn (x) : n ∈ N0 } is compact. Since
αn (x) ∈ K for all n, we have αn (x) →R K and thus αn (x) →R K+ ∩ K− , by
Lemma 2.12. Since K+ ∩ K− is compact and α(K+ ∩ K− ) = K+ ∩ K− ,
K+ ∩ K− ⊆ lev(α).
Thus x ∈ con(α, K+ ∩ K− ) ⊆ con(α, lev(α)) = con(α) lev(α), using Theorem A.
Thus par(α) ⊇ con(α) lev(α). The converse inclusion is trivial.
If V ⊆ G is a compact open subgroup which is tidy for α, then V− = V ∩
par(α) is open in par(α) and V+ ∩ V− = V ∩ lev(α) is open in lev(α) (cf. [26,
Proposition 11]). Since V− = (con(α) ∩ V )(V+ ∩ V− ) by (a), we see that
q(V+ ∩ V− ) = V− con(α)/con(α)
is open in par(α)/con(α). Since V+ ∩ V− is compact, the continuous surjection
q|V+ ∩V− : V+ ∩V− → q(V+ ∩V− ) is a quotient map and thus an open map (being
also a homomorphism). The assertion follows.
(e) Every x ∈ par− (α) admits a bounded α-regressive trajectory (xn )n∈N0 . In
particular, K := {xn : n ∈ N0 } is compact. Since xn ∈ K for all n, we have
xn (x) →R K and thus xn →R K+ ∩ K− , by Lemma 2.12. Since K+ ∩ K− is
compact and α(K+ ∩ K− ) = K+ ∩ K− ,
K+ ∩ K− ⊆ lev(α).
Thus x ∈ con− (α, K+ ∩ K− ) ⊆ con− (α, lev(α)) = con− (α) lev(α), using Theorem B. Thus par− (α) ⊇ con− (α) lev(α). The converse inclusion is trivial.
If V ⊆ G is a tidy subgroup for α, then V+ = V ∩ par− (α) is open in par− (α)
and V+ ∩ V− = V ∩ lev(α) is open in lev(α) (cf. [26, Proposition 11]). Since
V+ = (con− (α) ∩ V+ )(V+ ∩ V− ) by (b), we see that
p(V+ ∩ V− ) = V+ con− (α)/con− (α)
is open in par− (α)/con− (α). Since V+ ∩V− is compact, the continuous surjection
p|V+ ∩V− : V+ ∩V− → p(V+ ∩V− ) is a quotient map and thus an open map (being
also a homomorphism). The assertion follows.
ENDOMORPHISMS OF T.D.L.C. GROUPS
57
(f) If α(V ) = V , then V+ = V− = V (whence V = V+ V− ) and V−− is open (and
thus closed); so V is tidy. If, conversely, V is tidy, then every x ∈ V lies on
a bounded α-trajectory (xn )n∈Z (as x ∈ lev(α)), whence x ∈ V+ ∩ V− by [26,
Proposition 11]. Thus V = V+ ∩ V− and thus α(V ) = V .
Proof. (Theorem F). If α has small tidy subgroups, then also αL := α|lev(α) , by
Corollary 9.7. If, conversely, αL has small tidy subgroups, then
{e} = nub(αL ) = con(αL ) ∩ lev(αL ) = con(αL ),
(39)
using Proposition 9.5 for the first equality, Proposition 12.2 (e) for the second,
and the fact that lev(αL ) = lev(α) for the last equality. Thus
nub(α) = con(α) ∩ lev(α) = con(αL ) = {e},
(40)
whence α has small tidy subgroups (by Proposition 9.5).
Assuming now that α has small tidy subgroups, we have:
(c) Since ker(α)∩par− (α) ⊆ bik(α) ⊆ nub(α) = {e}, the induced endomorphism
α|par− (α) of par− (α) is injective and hence an automorphism of the topological group par− (α) (by Lemma 10.2 (b)). We know that the endomorphisms
α|con− (α) and α|lev(α) are surjective. Being restrictions of the automorphism
α|par− (α) , also α|con− (α) and α|lev(α) are automorphisms.
(b) By (40), we have con(α) ∩ lev(α) = {e}. Together with Lemma 13.1 (a)
and (d), this entails that par(α) = con(α) ⋊ lev(α) as an abstract group. By
Proposition 13.1 (d), the map q : lev(α) → par(α)/ con(α), g 7→ g con(α) is
an isomorphism of topological groups. As a consequence, the homomorphism
par(α) → lev(α), xy 7→ y (with x ∈ con(α), y ∈ lev(α)) is continuous and thus
par(α) = con(α) ⋊ lev(α) as a topological group.
As α has small tidy subgroups, so does the automorphism β := α|par− (α) (see
Corollary 9.7). Hence
par− (α)
=
par− (β) = par(β −1 ) = con(β −1 ) ⋊ lev(β −1 )
=
con− (β) ⋊ lev(β) = con− (α) ⋊ lev(α).
(d) Let V ⊆ G be a compact open subgroup which is tidy for α. Then
V
=
V −1 = (V− )−1 (V+ )−1 = V− (V+ )−1
=
(con(α) ∩ V )(lev(α) ∩ V )(con− (α) ∩ V ) ⊆ Ω,
58
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
using Lemma 13.1 (b) and (c). As each of con(α) and con− (α) intersects lev(α)
in {e} (by (b)), we deduce from Lemma 13.1 (b) and (c) that
(41) V− = (con(α) ∩ V ) ⋊ (lev(α) ∩ V ) and V+ = (con− (α) ∩ V ) ⋊ (lev(α) ∩ V )
as abstract groups. Using (b), we see that (41) also holds in the sense of topological groups.
(a) By (d), Ω is an identity neighbourhood. For x ∈ G, the translations
λx , ρx : G → G,
λx (y) := xy, ρx (y) := yx for y ∈ G
are homeomorphisms. If z ∈ Ω, write z = xy with x ∈ con(α) and y ∈
lev(α) con− (α) = par− (α) (cf. Lemma 13.1 (e)). Since
Ω = con(α) par− (α),
we have (ρy ◦ λx )(Ω) ⊆ Ω. Since (ρy ◦ λx )(Ω) is a neighbourhood of
(ρy ◦ λx )(e) = xy = z,
so is Ω. Hence Ω is open. If also z = ab with a ∈ con(α) and y ∈ par− (α), then
a−1 x = by −1 ∈ con(α) ∩ par− (α) = con(α|lev(α) ) = {e}
(where we used (39)). Hence x = a and y = b, showing that π is injective and
hence a bijection. It remains to show that π is an open map. If V ⊆ G is a tidy
subgroup for α and
L := (con(α) ∩ V ) × (lev(α) ∩ V ) × (con− (α) ∩ V ),
then π restricts to a continuous bijection
(42)
πV : L → V− V+ = V,
as a consequence of (d). Since L is compact, πV is a homeomorphism. Hence
π|L is an open map. Given a ∈ con(α), b ∈ lev(α) and c ∈ con− (α), we have
π(x, y, z) = aπ|L (a−1 x, yb−1 , bzc−1 b−1 )bc
for (x, y, z) ∈ con(α) × lev(α) × con− (α) in some open neighbourhood W of
(a, b, c). Since translations and the automorphism con− (α) → con− (α), v 7→
bvb−1 are open maps, π|W is an open map and thus π a homeomorphism.
(e) If V is tidy, then V is the product of the α-invariant set con(α) ∩ V =
con(α|V− ), the α-stable set lev(α) ∩ V = V+ ∩ V− and the set con− (α) ∩ V =
ENDOMORPHISMS OF T.D.L.C. GROUPS
59
con− (α) ∩ V+ whose image under α is contained in con− (α). In view of (a), we
now deduce from
(V+ ∩ V− )(con− (α) ∩ V+ ) = V+ ⊆ α(V+ ) = (V+ ∩ V− )α(con− (α) ∩ V+ )
that con− (α) ∩ V+ ⊆ α(con− (α) ∩ V+ ). Conversely, assume that (3)–(6) are
satisfied by a compact open subgroup V ⊆ G. If v ∈ V− , write π −1 (v) =:
(a, b, c). As con(α), lev(α) and con− (α) are α-invariant, we deduce from
αn (v) = αn (a)αn (b)αn (c) ∈ V
that αn (c) ∈ con− (α) ∩ V for all n ∈ N0 , whence c ∈ par(α) ∩ con− (α) =
con(α) lev(α) ∩ con− (α) = {e} and thus c = e. We now easily deduce that
V− = (con(α) ∩ V )(lev(α) ∩ V ).
If v ∈ V+ and (a, b, c) := π −1 (v), let (vn )n∈N0 be an α-regressive trajectory for v
in V . Write (an , bn , cn ) := π −1 (vn ). Then (an )n∈N0 is an α-regressive trajectory
for a in V , whence a ∈ con(α) ∩ par− (α) = con(α) ∩ lev(α) ⊆ nub(α) = {e} and
thus a = e. We now easily deduce that
V+ = (lev(α) ∩ V )(con− (α) ∩ V ).
Using that lev(α) ∩ V is α-stable and con− (α) ⊆ V++ (cf. Lemma 9.1), we find
V++ = (lev(α) ∩ V ) con− (α).
As con− (α) is closed in G and lev(α) ∩ V is compact, V++ is closed in G.
Since β := α|par− (α) is an automorphism, the indices [αn+1 (V+ ) : αn (V+ )] =
[β n+1 (V+ ) : β n (V+ )] are independent of n ∈ N0 . Hence V is tidy.
(f) As lev(α) has small tidy subgroups, (f) follows from Lemma 13.1 (f).
References
[1] Baumgartner, U. and G. A. Willis, Contraction groups and scales of automorphisms of
totally disconnected locally compact groups, Isr. J. Math. 142 (2004), 221–248.
[2] Bourbaki, N., “General Topology I”, Chapters 1–4, Springer, Berlin, 1988.
[3] Bourbaki, N., “Lie Groups and Lie Algebras”, Chapters 1–3, Springer, Berlin, 1989.
[4] Dani, S. G. and R. Shah, Contraction subgroups and semistable measures on p-adic Lie
groups, Math. Proc. Camb. Philos. Soc. 110 (1991), 299–306.
[5] Giordano Bruno, A. and S. Virili, Topological entropy in totally disconnected locally
compact groups, Ergodic Theory Dyn. Syst. 37(7) (2017), 2163-2186.
60
T.P. BYWATERS, H. GLÖCKNER AND S. TORNIER
[6] Glöckner, H., Contraction groups for tidy automorphisms of totally disconnected groups,
Glasg. Math. J. 47 (2005), 329-333.
[7] Glöckner, H., Locally compact groups built up from p-adic Lie groups, for p in a given
set of primes, J. Group Theorey 9 (2006), 427–454.
[8] Glöckner, H., Contractible Lie groups over local fields, Math. Z. 260 (2008), 889–904.
[9] Glöckner, H., Endomorphisms of Lie groups over local fields, preprint,
arXiv:1701.01804v2.
[10] Glöckner, H. and G. A. Willis, Classification of the simple factors appearing in composition series of totally disconnected contraction groups, J. Reine Angew. Math. 634
(2010), 141–169.
[11] Hewitt, E. and K. A. Ross, “Abstract Harmonic Analysis,” Vol. 1, Springer, Berlin, 2 1979.
[12] Isbell, J. R., “Uniform Spaces,” AMS, Providence, 1964.
[13] Jaworski, W., On contraction groups of automorphisms of totally disconnected locally
compact groups, Isr. J. Math. 172 (2009), 1–8.
[14] Raja, C. R. E. and R. Shah, Some properties of distal actions on locally compact groups,
preprint, arXiv.1201.4287v2.
[15] Reid, C. D., Endomorphisms of profinite groups, Groups Geom. Dyn. 8 (2014), 553–564.
[16] Reid, C. D., Dynamics of flat actions on totally disconnected, locally compact groups,
New York J. Math. 22 (2016), 115–190.
[17] Robinson, D. J. S., “A Course in the Theory of Groups,” Springer, New York, 2 1996.
[18] Serre, J.-P., “Lie Algebras and Lie Groups,” Springer, Berlin, 1992.
[19] Siebert, E., Contractive automorphisms on locally compact groups, Math. Z. 191 (1986),
73–90.
[20] van Dantzig, D., Studien over topologische Algebra, University of Groningen (1931).
[21] Wang, J. S. P., The Mautner phenomenon for p-adic Lie groups, Math. Z. 185 (1984),
403–412.
[22] Willis, G. A., The structure of totally disconnected, locally compact groups, Math. Ann.
300 (1994), 341–363.
[23] Willis, G. A., Further properties of the scale function on a totally disconnected group, J.
Algebra 237 (2001), 142–164.
[24] Willis, G. A., Tidy subgroups for commuting automorphisms of totally disconnected locally compact groups: An analogue of simultaneous triangularisation of matrices, New
York J. Math. 10 (2004), 1–35.
[25] Willis, G. A., The nub of an automorphism of a totally disconnected, locally compact
group, Ergodic Theory Dyn. Syst. 34 (2014), 1365–1394.
[26] Willis, G. A, The scale and tidy subgroups for endomorphisms of totally disconnected
locally compact groups, Math. Ann. 361 (2015), 403–442.
| 4math.GR
|
arXiv:1601.07317v2 [math.GR] 15 Oct 2016
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS IN
LOCALLY COMPACT GROUPS
COLIN D. REID AND PHILLIP R. WESOLEK
Abstract. In The essentially chief series of a compactly generated locally
compact group, an analogue of chief series for finite groups is discovered for
compactly generated locally compact groups. In the present article, we show
that chief factors necessarily exist in all locally compact groups with sufficiently
rich topological structure. We also show that chief factors have one of seven
types, and for all but one of these types, there is a decomposition into discrete
groups, compact groups, and topologically simple groups.
Our results for chief factors require exploring the theory developed in Chief
factors in Polish groups in the setting of locally compact groups. In this context, we obtain tighter restrictions on the factorization of normal compressions
and the structure of quasi-products. Consequently, both (non-)amenability
and elementary decomposition rank are preserved by normal compressions.
Contents
1. Introduction
2. Preliminaries
3. Elementary groups and the decomposition rank
4. Factorization of normal compressions in l.c.s.c. groups
5. Properties invariant under normal compressions
6. Generalized direct products
7. Chief factors and blocks
8. Extensions of chief blocks
9. Examples
References
1
6
8
17
23
27
34
39
48
52
1. Introduction
It is clear that a finite group G admits a normal series in which each of the
factors is a chief factor, that is, a normal factor K/L of G such that there is no
normal subgroup of G lying between K and L. It is then straightforward to show
that chief factors of finite groups are products of simple groups. One thereby arrives
at composition series, which give a complete decomposition of a finite group into
simple groups.
The first author is an ARC DECRA fellow. Research supported in part by ARC Discovery
Project DP120100996.
The second author is supported by ERC grant #278469.
1
2
COLIN D. REID AND PHILLIP R. WESOLEK
The previous work [15] establishes the existence of the essentially chief series for
compactly generated locally compact groups, which is an analogue for compactly
generated locally compact groups of the chief series for finite groups. For the
essentially chief series, we must allow for compact and discrete factors in addition
to chief factors. A primary focus of the work at hand is an attempt to complete the
analogy with finite groups by decomposing chief factors into topologically simple
groups, discrete groups, and compact groups. Theorem 1.18 is our high mark in
this direction, but it does not quite complete the analogy; Question 8.25 remains
an outstanding mystery.
The proofs require exploring the theory developed in [14] in the setting of locally compact Polish (equivalently, locally compact second countable) groups. As
expected, the results we obtain for locally compact Polish groups are much stronger
than those for general Polish groups.
We shall use the abbreviations l.c.s.c. for “locally compact second countable”
and t.d.l.c.s.c. for “totally disconnected locally compact second countable.”
1.1. Normal compressions. We begin by studying normal compressions. A topological group H is a normal compression of a topological group G if there is a
continuous, injective homomorphism ψ : G → H with dense normal image. As
shown in [14], when G and H are Polish groups, there is a Polish group H̃ such
that ψ can be decomposed as β ◦ α with α : G → H̃ a closed embedding with
normal image and β : H̃ → H a quotient map.
In the locally compact setting, this factorization can be refined to obtain better
control over the group H̃.
Theorem 1.1 (See Theorem 4.4). Suppose that G and H are t.d.l.c.s.c. groups
with ψ : G → H a normal compression map. Then there is a t.d.l.c.s.c. group H̃,
a closed embedding α : G → H̃ with normal image, and a quotient map β : H̃ → H
such that
(1) ψ = β ◦ α;
(2) H̃/α(G) is compact;
(3) ker(β) is discrete, lies in the FC-center of H̃, and centralizes α(G);
(4) H̃ = α(G) ker(β).
In the case of compactly presented t.d.l.c.s.c. groups, more can be said; see
Corollary 4.5. For a normal compression of l.c.s.c. groups, the compression map
factorizes as a sequence of closed normal embeddings and quotient maps such that
all kernels and cokernels appearing in the sequence are ‘small’; see Theorem 4.10.
Using these factorizations, we study the extent to which properties pass between
G and H when H is a normal compression of G.
Theorem 1.2. Suppose that G and H are l.c.s.c. groups with H a normal compression of G.
(1) G is amenable if and only if H is amenable. (See Proposition 5.6.)
(2) If G and H are totally disconnected, then G is elementary with decomposition rank α if and only if H is elementary with decomposition rank α. (See
Proposition 5.4.)
The class of elementary groups is the smallest class of t.d.l.c.s.c. groups containing the profinite groups and discrete groups and closed under the elementary
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
3
operations. This class admits a canonical rank function called the decomposition
rank, taking values in the set ω1 of at most countable ordinals. It is convenient to
define non-elementary t.d.l.c.s.c. groups to have decomposition rank ω1 .
A Polish group G is quasi-discrete if the quasi-center QZ(G) := {g ∈ G |
CG (g) is open} is dense. The property of being discrete is not preserved under
normal compressions, but quasi-discreteness is preserved up to metabelian factors.
Theorem 1.3 (See Theorem 5.9). Suppose that G and H are l.c.s.c. groups with
H a normal compression of G.
(1) If H is quasi-discrete, then G/QZ(G) is metabelian.
(2) If G is quasi-discrete, then H/QZ(H) is metabelian.
1.2. Generalized direct products. We next consider generalized direct products
in the locally compact setting.
Definition 1.4. For G a topological group and J a collection of closed normal
subgroups, we set GJ := hN | N ∈ Ji. A collection S of closed normal subgroups is
a quasi-direct factorization of G if GS = G, and S has the following topological
independence property:
\
\
∀X ⊆ P(S) :
X=∅⇒
GA = {1}.
A∈X
In such a case, (G, S) is said to be a quasi-product (or we say G is a quasi-product,
when the factorization is implicit).
Basic examples of quasi-products are local direct products.
Definition 1.5. Suppose that (Gi )i∈I is a set of topological groups and that there
is a distinguished open subgroup Oi ≤ L
Gi for each i ∈ I. The local direct product
of (Gi )i∈I over (Oi )i∈I is denoted by i∈I (Gi , Oi ) and defined to be
(
)
G
f :I→
Gi | f (i) ∈ Gi , and f (i) ∈ Oi for all but finitely many i ∈ I
i∈I
Q
with component-wise multiplication and the group topology such that
L i∈I Oi with
the product topology continuously embeds as an open subgroup of i∈I (Gi , Oi ).
We identify certain quasi-products that are closely related to local direct products.
Definition 1.6. Let G be a locally compact group with S a quasi-direct factorization
direct model
L of G. A local L
L for (G, S) is a continuous homomorphism
φ : S (G) → GLwhere S (G) := N ∈S (N, N ∩ O) for some open subgroup O of
G is such that S (G) is locally compact and φ N = idN for all N ∈ S.
Local direct models for quasi-products, provided they exist, enjoy a universal
property and are in particular unique; see Theorem 6.16 and Corollary 6.17.
Theorem 1.7 (See Corollary 6.20). If (G, S) is a t.d.l.c. quasi-product, then (G, S)
admits a local direct model.
A natural source of quasi-products are groups of strict semisimple type: A Polish
group G is of strict semisimple type if it is a quasi-product of non-abelian
topologically simple groups.
4
COLIN D. REID AND PHILLIP R. WESOLEK
Theorem 1.8 (See Theorem 6.24). For G an l.c.s.c. group of strict semisimple
type,
n
Y
Y
G'D×
Ci ×
Sj
i∈I
j=0
with D a t.d.l.c.s.c. group of strict semisimple type, each Ci a topologically simple compact connected Lie group, and each Sj a non-compact topologically simple
connected Lie group.
Corollary 1.9 (See Corollary 6.25). Suppose that G is an l.c.s.c. group of strict
semisimple type and let S be the set of non-abelian topologically simple closed normal
subgroups of G. Then (G, S) is a quasi-product that admits a local direct model.
1.3. Chief factors and blocks. We finally bring our results together to study
chief factors in locally compact groups. The following results draw on the general
theory of chief factors as developed in [14] as well as the existence of essentially
chief series for compactly generated locally compact groups as shown in [15]. In
fact, we will show (Proposition 8.17) that even without the assumption of compact
generation, chief factors of arbitrary l.c.s.c. groups G account for a large part of
their complexity as topological groups.
A normal factor K/L of closed normal subgroups of a topological group G is a
chief factor if there is no closed M E G such that L < M < K.
Definition 1.10. Given a topological group G, we say that the closed normal
factors K1 /L1 and K2 /L2 are associated if the following equations hold:
K1 L2 = K2 L1 ; K1 ∩ L1 L2 = L1 ; K2 ∩ L1 L2 = L2 .
The association relation is an equivalence relation when restricted to non-abelian
chief factors via [14, Proposition 6.8]. A chief block is an association class of nonabelian chief factors. The set of chief blocks of a group G is denoted by BG .
The association relation is not in general comparable with isomorphism. To
study group-theoretic properties via chief blocks, we identify properties which hold
on the level of blocks. Blocks with compact or discrete representatives have exceptional properties, so it is natural to isolate a class of blocks which excludes
these.
Definition 1.11. Suppose that a ∈ BG . We say that the block a ∈ BG is robust
if none of the representatives of a are compact or quasi-discrete. The collection of
robust chief blocks is denoted by BrG .
Definition 1.12. A property P of groups is a (robust) block property if for
every l.c.s.c. group G and a a (robust) chief block of G, there exists K/L ∈ a with
P if and only if every K/L ∈ a has P . For a (robust) block property P , we say that
a (robust) chief block a of G has property P if some, equivalently all, K/L ∈ a has
property P .
Theorem 1.13 (See Propositions 7.8 and 7.9). The following are block properties: Being elementary with rank α, being amenable, being quasi-discrete, and being
isomorphic to a given connected group H. Compact generation is a robust block
property.
Representatives of non-robust blocks therefore have a restricted structure.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
5
Corollary 1.14. Suppose that G is an l.c.s.c. group and suppose a is a chief block
of G that is not robust. Then either every representative of a is isomorphic to a
fixed connected compact group, or every representative of a is totally disconnected
and quasi-discrete.
To generalize the existence of chief factors from compactly generated groups
to general l.c.s.c. groups, and also to analyze the internal structure of these chief
factors, we employ the concepts of minimally covered and extendable chief blocks as
introduced in [14]. For a Polish group G, a closed normal subgroup N E G covers
a block a ∈ BG if there is K/L ∈ a such that K ≤ N . A block a is minimally
covered if there is a least normal subgroup that covers the block.
Definition 1.15. Let G be a Polish group and H be a closed subgroup of G. Given
a ∈ BH and b ∈ BG , we say that b is an extension of a in G if for every closed
normal subgroup K of G, the group K covers b if and only if K ∩ H covers a.
Theorem 1.16 (See Theorem 8.15). If O is a compactly generated open subgroup
of an l.c.s.c. group G, then all robust blocks of O extend to G.
Blocks of an l.c.s.c. group arising from extending a robust block of a compactly
generated open subgroup are called regionally robust blocks. The set of regionally robust blocks of an l.c.s.c. group G is denoted by Brr
G . These blocks are
minimally covered and robust; see Lemma 8.6 and Propositions 8.8 and 8.10. The
set of regionally robust blocks is setwise invariant under all continuous automorphisms. A l.c.s.c. group also has at most countably many regionally robust blocks
(Lemma 8.12). Moreover, an l.c.s.c. group G has at least one regionally robust
block provided that its structure as a topological group is sufficiently complex:
Proposition 1.17 (See Proposition 8.17). For G an l.c.s.c. group, either
(1) Brr
G 6= ∅, or
(2) G◦ is compact-by-solvable-by-compact, and G/G◦ is elementary with rank
at most ω + 1. If G is compactly generated, then G/G◦ is elementary with
finite rank.
Significant restrictions on the structure of topologically characteristically simple
l.c.s.c. groups then follow; chief factors in l.c.s.c. groups are particular examples of
such groups.
Theorem 1.18 (See Theorem 8.21). Suppose that G is a topologically characteristically simple l.c.s.c. group. Then G is exactly one of the following:
(1) an abelian group;
(2) a non-abelian quasi-discrete t.d.l.c.s.c. group;
(3) an elementary t.d.l.c.s.c. group of decomposition rank 2 with trivial quasicenter;
(4) an elementary t.d.l.c.s.c. group of decomposition rank ω + 1;
(5) a direct product of copies of a simple Lie group;
(6) a quasi-product of copies of a topologically simple t.d.l.c.s.c. group of decomposition rank greater than ω + 1;
(7) a t.d.l.c.s.c. group of stacking type and decomposition rank greater than ω + 1.
In particular, chief factors of l.c.s.c. groups have one of the above forms.
6
COLIN D. REID AND PHILLIP R. WESOLEK
All of the cases (1)-(7) can occur as chief factors of an l.c.s.c. group. Example 9.2
shows case (7) occurs naturally, even for a chief factor of a compactly generated
l.c.s.c. group.
2. Preliminaries
2.1. Notations and definitions. For P a poset, we say that D ⊆ P is a directed
family if for all M, N ∈ D, there is L ∈ D with M ≤ L and N ≤ L. We say F ⊆ P
is a filtering family if for all M, N ∈ F there is L ∈ F such that L ≤ M ∩ N .
Given a group G and a set I, we denote the direct sum by G<I and the direct
product by GI . For a subset K ⊆ G, CG (K) is the collection of elements of G that
centralize every element of K. For the normalizer, we write NG (K). For A, B ⊆ G,
we put
[A, B] := haba−1 b−1 | a ∈ A and b ∈ Bi
For a, b ∈ G, [a, b] := aba−1 b−1 and ab := bab−1 . For a, b, c ∈ G, we set [a, b, c] :=
[[a, b], c].
All topological groups are Hausdorff topological groups and are written multiplicatively. A topological group is Polish if the topology is separable and admits
a complete, compatible metric. A locally compact group is Polish if and only if
it is second countable. Topological group isomorphism is denoted by '. For a
topological group G, the connected component of the identity is denoted by G◦ .
The topological closure of a set K in G is denoted by K. We shall say that G is
topologically perfect if [G, G] is dense in G.
We use “t.d.”, “l.c.”, and “s.c.” for “totally disconnected”, “locally compact”,
and “second countable”, respectively. The symbols Z, Q, Zp , Qp , and R denote the
additive groups of the relevant rings with the discrete topology for Q and Z and
the usual locally compact topology for the others. We write T for the circle group
R/Z and write Cp for the cyclic group of order p.
2.2. Generalities on locally compact groups. A locally compact group is almost connected if the identity component is cocompact. For G a locally compact group, the set of almost connected subgroups is denoted by U(G). Note that
G◦ = U ◦ for all open subgroups U of G, so U(G) consists precisely of those open
subgroups U of G such that U/G◦ is compact. If G is a t.d.l.c. group, then U(G)
is a base of identity neighborhoods, by van Dantzig’s theorem.
A locally compact group is locally elliptic if every compact subset (equivalently,
every finite subset) generates a relatively compact subgroup. The locally elliptic
radical of a locally compact group G, denoted by RadLE (G), is the union of all
closed normal locally elliptic subgroups of G.
Theorem 2.1 (Platonov, [13]). For G a locally compact group, RadLE (G) is the
unique largest locally elliptic closed normal subgroup of G. Additionally,
RadLE (G/ RadLE (G)) = {1}.
The quasi-center of a locally compact group G is
QZ(G) := {g ∈ G | CG (g) is open}.
The quasi-center is a characteristic subgroup, but it is usually not closed. A locally
compact group is quasi-discrete if it has a dense quasi-center.
For a compactly generated t.d.l.c. group G, there exists a locally finite graph
Γ such that G acts vertex-transitively on Γ with compact open stabilizers; such
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
7
a graph is called a Cayley–Abels graph for G. The existence of this graph is
ensured by the next proposition.
Proposition 2.2 (See [1, Beispiel 5.2] or [11, Section 11]). Let G be a compactly
generated t.d.l.c. group, let U ∈ U(G), let A be a compact symmetric subset of G
such that G = hU, Ai and let D be a dense symmetric subset of G.
(1) There exists a finite symmetric subset B of G such that B ⊆ D and
BU = U B = U BU = U AU.
(2) For any subset B satisfying part (1), it is the case that G = hBiU and the coset
space G/U carries the structure of a locally finite connected graph, invariant
under the natural G-action, where gU is adjacent to hU if and only if gU 6= hU
and U g −1 hU ⊆ U BU . Thus, G/U is the vertex set of a Cayley–Abels graph
for G.
A (real) Lie group is a topological group that is a finite-dimensional analytic
manifold over R such that the group operations are analytic maps. Closed subgroups of Lie groups are again Lie groups. A Lie group G can have any number of
connected components in general, but G◦ is always an open subgroup of G.
A Lie group L has a largest connected solvable normal subgroup; it is called the
solvable radical of L and denoted by Sol(L). A connected Lie group with trivial
solvable radical is called semisimple. A semisimple Lie group has discrete center,
and modulo its center, it is a finite direct product of abstractly simple groups.
The manifold structure of a Lie group gives a well-defined dimension dimR (G),
which is additive with respect to extensions. We stress that dimension zero Lie
groups are discrete. The dimension places a restriction on directed families of
normal subgroups.
Lemma 2.3. Let G be a connected Lie group and let D be a directed family of closed
normal subgroups of G. Then there exists N ∈ D such that M N/N is discrete in
G/N for all M ∈ D.
Proof. Let d be the maximum of the set {dimR (N ) | N ∈ D} and take N ∈ D
realizing this maximum. For all M ∈ D, there is K ∈ D extending N and M ,
since D is directed. The group K/N is a dimension zero Lie subgroup of G/N ,
hence K/N is discrete. We thus conclude that M N/N is also discrete, verifying
the lemma.
Lie groups play a fundamental role in the structure theory of connected groups.
Theorem 2.4 (Gleason–Yamabe; see [12, Theorem 4.6]). If G is an almost connected locally compact group, then RadLE (G) is compact, and G/ RadLE (G) is a
Lie group. Moreover, there exist arbitrarily small compact normal subgroups K of
G such that G/K is a Lie group.
2.3. Normal compressions of Polish groups. We conclude our preliminary
discussion by recalling the basic theory of normal compressions developed in [14].
Definition 2.5. For G and H Polish groups, a normal compression map from
G to H is an injective, continuous homomorphism ψ : G → H with dense, normal
image. When such a map exists, we say that H is a normal compression of G.
8
COLIN D. REID AND PHILLIP R. WESOLEK
Given a normal compression map ψ : G → H, there is a canonical action H y G
via
h.g := ψ −1 (hψ(g)h−1 ).
This action is called the ψ-equivariant action, as it is the action with respect to
which ψ is an H-equivariant map for H acting on itself by conjugation.
Proposition 2.6 ([14, Proposition 3.5]). Suppose that G and H are Polish groups
with ψ : G → H a normal compression map. Then the semidirect product G oψ H
via the ψ-equivariant action is a Polish group under the product topology.
Normal compression maps factor through this semidirect product.
Theorem 2.7 ([14, Theorem 3.6]). Suppose that G and H are Polish groups with
ψ : G → H a normal compression map and let O ≤ H be an open subgroup. Then
the following hold:
(1) π : G oψ O → H via (g, o) 7→ ψ(g)o is a continuous, surjective homomorphism
with ker(π) = {(g −1 , ψ(g)) | g ∈ ψ −1 (O)}, and if O = H, then ker(π) ' G as
topological groups.
(2) ψ = π ◦ ι where ι : G → G oψ O is the usual inclusion.
(3) G oψ O = ι(G) ker(π), and the subgroups ι(G) and ker(π) are closed normal
subgroups of G oψ O with trivial intersection.
This factorization allows some properties to pass back and forth under normal
compressions, possibly up to an abelian factor.
Proposition 2.8 ([14, Proposition 3.8]). Suppose that G and H are Polish groups
with ψ : G → H a normal compression map and let K be a closed normal subgroup
of G. Then the following hold:
(1) The image ψ(K) is a normal subgroup of H.
(2) If ψ(K) is also dense in H, then [G, G] ≤ K, and every closed normal subgroup
of K is normal in G.
We shall also make use of two technical results concerning connected components.
Lemma 2.9 ([14, Lemma 3.11]). Let G and H be Polish groups with ψ : G → H
a normal compression map. Let R := ψ −1 (H ◦ ) and let N := ψ(R). Then the
following inclusions hold:
[H ◦ , ψ(G)] ≤ ψ(R); [H ◦ , H] ≤ N ; [H ◦ , ψ(G)] ≤ ψ(G◦ ); [H ◦ , N ] ≤ ψ(G◦ ).
Theorem 2.10 ([14, Theorem 3.12]). Suppose that G and H are Polish groups
with ψ : G → H a normal compression map. Then the following hold:
(1) The group H ◦ /ψ(G◦ ) is nilpotent of class at most two.
(2) If H is connected, then both G/G◦ and H/ψ(G◦ ) are abelian.
(3) If G is totally disconnected, then H ◦ is central in H.
Corollary 2.11 ([14, Corollary 3.13]). Suppose that G and H are centerless topologically perfect Polish groups and that H is a normal compression of G. Then G
is connected if and only if H is connected, and G is totally disconnected if and only
if H is totally disconnected.
3. Elementary groups and the decomposition rank
Elementary groups will appear several times in the present work. We here recall
the relevant definitions and note several new properties.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
9
3.1. Introduction.
Definition 3.1. The class E of elementary groups is the smallest class of
t.d.l.c.s.c. groups such that
(i) E contains all second countable profinite groups and countable discrete groups.
(ii) E is closed under taking closed subgroups.
(iii) E is closed under taking Hausdorff quotients.
(iv) E is closed under forming group extensions.
S
(v) If G is a t.d.l.c.s.c. group and G = i∈N Oi where (Oi )i∈N is an ⊆-increasing
sequence of open subgroups of G with Oi ∈ E for each i, then G ∈ E . We say
that E is closed under countable increasing unions.
The operations (ii) − (v) are often called the elementary operations. It turns
0
out operations (ii) and (iii) follow from the others, and (iv) can be weakened to (iv) :
E is closed under extensions of profinite groups and discrete groups. These results
are given by [20, Theorem 1.3]. The class of elementary groups enjoys additional
closure properties:
Theorem 3.2 (See [20, Theorem 1.3]). The class E enjoys the following properties:
(1) If G ∈ E , H is a t.d.l.c.s.c. group, and ψ : H → G is a continuous, injective
homomorphism, then H ∈ E .
(2) If G is a t.d.l.c.s.c. group that is residually in E , then G ∈ E . In particular, E
is closed under inverse limits that result in a t.d.l.c.s.c. group.
A primary tool to study elementary groups is a canonical rank function.
Definition 3.3. Let T be the class of t.d.l.c.s.c. groups. The decomposition
rank ξ is a partial ordinal-valued function from T to [1, ω1 ) satisfying the following
properties:
(a) ξ(G) = 1 if and only if G ' {1};S
(b) If G ∈ T is non-trivial and G = i∈N Oi with (Oi )i∈N some increasing sequence
of compactly generated open subgroups of G, then ξ(G) is defined if and only
if ξ(Ri ) is defined for Ri := Res(Oi ) for all i ∈ N. If ξ(G) is defined, then
ξ(G) = sup ξ(Ri ) + 1.
i∈N
By [20, Theorem 4.7] and [20, Lemma 4.12], such a function ξ exists, is unique,
has domain of definition exactly E , and is equivalent to the decomposition rank as
defined in [20]. Note that if G is a non-trivial t.d.l.c.s.c. group that is compactly
generated and has no non-trivial discrete quotients, then G is not in E . The class E
is thus strictly smaller than the class of all t.d.l.c.s.c. groups. For convenience, we
extend the definition of ξ to all t.d.l.c.s.c. groups by saying that a non-elementary
group satisfies ξ(G) = ω1 .
By definition, the rank of an elementary t.d.l.c.s.c. group is always a countable successor ordinal. We observe a slightly stronger restriction on the rank of a
compactly generated group.
Lemma 3.4. Let G be a compactly generated elementary t.d.l.c.s.c. group. Then
ξ(G) = ξ(Res(G)) + 1 = β + 2
for some at most countable ordinal β.
Proof. From the rank formula, ξ(G) = ξ(Res(G)) + 1; take Oi = G for all i ∈ N.
On the other hand, ξ(Res(G)) is a successor ordinal, say β + 1, so ξ(G) = β + 2.
10
COLIN D. REID AND PHILLIP R. WESOLEK
3.2. Stability of the decomposition rank. The decomposition rank enjoys several useful stability properties.
Proposition 3.5. For G a non-trivial elementary group, then the following hold:
(1) If H is a t.d.l.c.s.c. group, and ψ : H → G is a continuous, injective homomorphism, then ξ(H) ≤ ξ(G). ([20, Corollary 4.10])
(2) If L E G is closed, then ξ(G/L) ≤ ξ(G). ([20, Theorem 4.19])
We give here a slightly sharper version of [16, Lemma 7.4]. Given a successor
ordinal α, we write (α − 1) to mean the ordinal β such that α = β + 1.
Lemma 3.6. Suppose that G is a t.d.l.c.s.c. group and G lies in a short exact
sequence of t.d.l.c.s.c. groups
{1} → N → G → Q → {1}.
Then ξ(G) ≤ (ξ(N ) − 1) + ξ(Q).
Proof. We argue by induction on ξ(Q) for all such G. In the base case, ξ(Q) = 1
and Q = {1}, so ξ(G) = ξ(N ) = (ξ(N ) − 1) + ξ(Q). Suppose that the lemma
holds up to β and that ξ(Q) = β + 1. Let π : G → Q be the usual projection and
fix (Oi )i∈N an increasing exhaustion of G by compactly generated open subgroups.
Put Wi := π(Oi ) and note that the Wi form an increasing exhaustion of Q by
compactly generated open subgroups.
In view of the decomposition rank formula, it suffices to consider ξ(Res(Oi )) for
an arbitrary i. Fix i ∈ N, form R := Res(Oi ), and put M := RN . It follows
that M/N ≤ Res(Wi ), hence ξ(M/N ) ≤ β. The induction hypothesis implies
ξ(M ) ≤ (ξ(N ) − 1) + β, and since R ,→ M , we conclude that ξ(R) ≤ (ξ(N ) − 1) + β.
Appealing to Proposition 3.5, ξ(G) ≤ (ξ(N ) − 1) + β + 1, verifying the lemma.
Lemma 3.7.
Qn Suppose that H1 , . . . , Hn are t.d.l.c.s.c. groups. If each Hi is elementary, then i=1 Hi is elementary with
!
n
Y
ξ
Hi = max ξ(Hi ).
i=1
1≤i≤n
Qn
Proof. Set G := i=1 Hi . In view of Proposition 3.5, max1≤i≤n ξ(Hi ) ≤ ξ(G).
We argue by induction on ξ(G) for the converse inequality. If ξ(G) = 1, then G
is the trivial group, and the inductive claim is immediate. Suppose that ξ(G) =
β + 1. For each 1 ≤ i ≤ n, let (Oj,i )j∈N be an increasing
Qnexhaustion of Hi by
compactly generated open subgroups. The groups Lj := i=1 Oj,i then form an
increasing exhaustion of G by compactly generated open subgroups. Hence, ξ(G) =
supj∈N ξ(Res(Li )) + 1.
Qn
On the other hand, Res(Lj ) = i=1 Res(Oj,i ), so the inductive hypothesis ensures that
ξ(Res(Lj )) = max ξ(Res(Oj,i )).
1≤i≤n
By passing to a subsequence if necessary, we may assume there is 1 ≤ k ≤ n such
that ξ(Res(Lj )) = ξ(Res(Oj,k )) for all j ∈ N. Therefore,
ξ(G) = sup ξ(Res(Oj,k )) + 1 = ξ(Hk ).
j∈N
Since it must be the case that ξ(Hk ) = max1≤i≤n ξ(Hi ), the induction is complete.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
11
Lemma 3.8. Suppose that G is a t.d.l.c.s.c. group and N is a non-trivial closed
cocompact normal subgroup of G. If N ∈ E , then G ∈ E with ξ(G) = ξ(N ).
Proof. Plainly G ∈ E . To compute the decomposition rank, fix (Oi )i∈N a countable
⊆-exhaustion of G by compactly generated open subgroups of G and put Ni :=
N ∩ Oi . The group Ni is open in N , and since Ni is cocompact in Oi , it is also
compactly generated. We therefore infer that
ξ(N ) = sup ξ(Res(Ni )) + 1.
i∈N
On the other hand, [20, Lemma 4.16] shows that Oi / Res(Ni ) is residually discrete, and it follows that Res(Oi ) = Res(Ni ). We conclude that
ξ(G) = sup ξ(Res(Oi )) + 1 = sup ξ(Res(Ni )) + 1 = ξ(N ),
i∈N
i∈N
verifying the lemma.
At this point it will be convenient to introduce notation for a kind of limit that
often occurs with elementary decomposition rank.
Definition 3.9. Given a set I of ordinals, define sup+ (I) = sup(I) if sup(I) is
either ω1 or a successor ordinal and define sup+ (I) = sup(I) + 1 if sup(I) is a
countable limit ordinal. If I is a singleton {α}, we write sup+ (α). The ordinal
sup+ (I) is the least ordinal that is an upper bound for I and that is a potential
decomposition rank for a t.d.l.c.s.c. group.
It is easily verified, using the definition of the decomposition rank, that if G is
an ascending union of open subgroups Oi , then ξ(G) = supi∈N + ξ(Oi ). In fact, the
same formula applies more generally if G contains a dense ascending union of closed
subgroups, as long as these subgroups each have open normalizer.
Proposition 3.10. Let G be a t.d.l.c.s.c. group and let (Ci )i∈I be a directed family
S
of closed subgroups of G each with open normalizer such that G = i∈I Ci . If each
Ci is elementary, then G is elementary with
ξ(G) = sup+ ξ(Ci ).
i∈I
Proof. By Proposition 3.5, we have ξ(Ci ) ≤ ξ(G) for each i. Thus sup+
i∈I ξ(Ci ) ≤
ξ(G).
For the reverse inequality, fix U ∈ U(G) and let (Oj )j∈N be an increasing exhaustion of GSby compactly generated open subgroups each containing U . For each
j, the union i∈I Oj ∩ Ci is dense in Oj , so in view of Proposition 2.2, we may find
i and A ⊆ Oj ∩ Ci such that U AU = AU and U AU contains a generating set for
Oj . Therefore, Oj = hAiU = (Oj ∩ Ci )U .
The subgroup Ci has an open normalizer, so we may take an open V E U such
that V ≤ NG (Ci ). The group hAV iV is then a finite index subgroup of Oj . Letting
L be the normal core of hAV iV in Oj , we see that L is a finite index normal
subgroup of Oj and of hAV iV . Lemma 3.8 thus ensures that
ξ(hAV iV ) = ξ(L) = ξ(Oj ).
Furthermore, hAV i is a closed cocompact normal subgroup of hAV iV , so a second
application of Lemma 3.8 implies that ξ(hAV i) = ξ(Oj ). Proposition 3.5 now
ensures that ξ(Oj ) ≤ ξ(Ci ).
12
COLIN D. REID AND PHILLIP R. WESOLEK
Letting α := supi∈I ξ(Ci ), there are two possibilities: either α = α0 +1 is achieved
by some ξ(Ci ), or α is a limit ordinal. If α is a limit ordinal, then
ξ(G) = sup ξ(Res(Oj )) + 1 ≤ sup ξ(Ci ) + 1 = sup+ ξ(Ci ).
i∈I
j∈N
i∈I
On the other hand, suppose that α is achieved by some ξ(Ci ); without loss of
generality, ξ(Ci ) = α for all i. For every compactly generated open subgroup Oj ,
we have that ξ(Oj ) ≤ α, so ξ(Res(Oj )) ≤ α0 . Hence, ξ(G) ≤ α0 + 1 = α. In either
case, we have obtained the required upper bound on ξ(G). In particular, it follows
that G is elementary.
There is also a dual result for residually elementary groups.
Lemma 3.11. Let G be a t.d.l.c.s.c. group and α an ordinal. Then {N E G |
N is closed and ξ(G/N ) ≤ α} is a filtering family.
Proof. It suffices to show Σ := {N E G | N is closed and ξ(G/N ) ≤ α} is closed
under pairwise intersections. Taking M, N ∈ Σ, the diagonal map G → G/N ×G/M
induces an injective, continuous map G/(N ∩ M ) → G/N × G/M . Lemma 3.7 and
Proposition 3.5 now ensure that
ξ(G/(N ∩ M )) ≤ ξ(G/N × G/M ) ≤ α.
Hence, N ∩ M ∈ Σ.
Proposition 3.12. Let G be a t.d.l.c.s.c. group and let K be a family of closed
normal subgroups of G with trivial intersection. Then
ξ(G) = sup + ξ(G/K).
K∈K
If G is compactly generated and elementary, then there is K ∈ K such that ξ(G) =
ξ(G/K).
Proof. Every K ∈ K has rank less than or equal sup+
K∈K ξ(G/K). Applying
Lemma 3.11, we may expand K to a filtering family K̃ such that sup+
ξ(G/K) =
K∈K̃
sup+
K∈K ξ(G/K). Replacing K with K̃, we may assume, without loss of generality,
that K is a filtering family.
If G/K is non-elementary for some K ∈ K, then G is also non-elementary,
and there is nothing to prove. We thus assume that G/K is elementary for all
K ∈ K, so by Theorem 3.2(2), G is elementary. Proposition 3.5 then ensures that
ξ(G/K) ≤ ξ(G) for all K ∈ K, and it follows that sup+
K∈K ξ(G/K) ≤ ξ(G).
+
Say that α + 1 = supK∈K G/K. We argue that ξ(G) ≤ α + 1 by induction on
α. The base case, α = 0, is immediate, since in this case G = {1}. Suppose that α
is a successor ordinal; say α = β + 1. For any compactly generated open subgroup
O ≤ G and K ∈ K, ξ(O/(O ∩ K)) ≤ β + 2, so ξ(Res(O/(O ∩ K))) ≤ β + 1. The
quotient of O/(O ∩ K) by Res(O)(O ∩ K)/O ∩ K is a quotient of O/ Res(O), so it
is residually discrete. Hence, Res(O)(O ∩ K)/(O ∩ K) ≤ Res(O/(O ∩ K)). There is
thus an injective, continuous map from Res(O)/(Res(O)∩K) into Res(O/(O ∩K)).
Appealing to Proposition 3.5, we conclude that
ξ(Res(O)/(Res(O) ∩ K)) ≤ β + 1.
Since β < α and β does not depend on K, the inductive hypothesis ensures that
ξ(Res(O)) ≤ β + 1 = α. It now follows that ξ(G) ≤ β + 2 = α + 1, since O is an
arbitrary compactly generated open subgroup.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
13
Let us now consider the case when α is a limit ordinal, and since α > 0, the
ordinal α is transfinite. Take O ≤ G a compactly generated open subgroup. If ξ(O)
is finite, then the inductive hypothesis implies that ξ(O) = sup+
K∈K (ξ(O/O ∩ K)).
Since {ξ(O/O ∩ K)}K∈K is a bounded collection of finite ordinals, there is some
K ∈ K such that ξ(O) = ξ(O/O ∩ K). Let us now suppose that ξ(O) is transfinite.
Since K is a filtering family, we may apply [15, Theorem 3.3] to find K ∈ K such
that there is a compact L ≤ K ∩ O that is normal in O and is open in K ∩ O.
Applying Proposition 3.5 and Lemma 3.6, we deduce that
ξ(O/K ∩ O) ≤ ξ(O) ≤ ξ(L) + ξ(K ∩ O/L) + ξ(O/K ∩ O) ≤ 4 + ξ(O/K ∩ O).
Since ξ(O) is transfinite, the only way to achieve such inequalities is if ξ(O/K ∩O) =
ξ(O). We conclude that ξ(O) ≤ α + 1 for any compactly generated open subgroup
O and hence that ξ(Res(O)) ≤ α. It now follows that ξ(G) ≤ α + 1, completing
the induction.
The second claim follows from the previous paragraph.
3.3. Regionally SIN groups. Elementary groups of decomposition rank 2 have a
natural characterization. A topological group is a small invariant neighborhood
group, abbreviated SIN, if there is a basis at 1 of conjugation invariant neighborhoods. A t.d.l.c. group is a SIN group if and only if there is a basis at 1 of compact
open normal subgroups.
We say that a t.d.l.c. group is regionally SIN if every open compactly generated
subgroup is a SIN group.
Lemma 3.13. For a non-trivial t.d.l.c.s.c. group G, the following are equivalent:
(1) G is regionally SIN;
(2) G is elementary with ξ(G) = 2.
Proof. From the definition of decomposition rank, we see that G is elementary with
ξ(G) = 2 if and only if every compactly generated open subgroup of G is residually
discrete. By [6, Corollary 4.1], a compactly generated t.d.l.c. group is residually
discrete if and only if it is a SIN group, and the conclusion follows.
Lemma 3.14. Let G be a quasi-discrete l.c.s.c. group. Then G◦ is central in G
and G/G◦ is elementary with ξ(G/G◦ ) ≤ 2.
Proof. Every open subgroup contains G◦ , so G◦ centralizes QZ(G). Since QZ(G)
is dense, G◦ is in fact central.
We see that G/G◦ is quasi-discrete, so we may assume G◦ = {1}. The group
G is then a t.d.l.c.s.c. group, and via [6, Proposition 4.3], G is regionally SIN.
Lemma 3.13 now ensures that ξ(G) ≤ 2.
3.4. The bounded rank residuals. Proposition 3.12 suggests a natural source
of closed characteristic subgroups.
Definition 3.15. For a t.d.l.c.s.c. group G and an ordinal α ≥ 1, define the rank-α
residual to be
\
Resα (G) := {K E G | K is closed and ξ(G/K) ≤ α}
The rank-ω residual will be called the finite rank residual of G.
These residuals have many nice properties. First, the rank of the quotient
G/ Resα (G) is restricted.
14
COLIN D. REID AND PHILLIP R. WESOLEK
Proposition 3.16. For G a t.d.l.c.s.c. group and α an at most countable ordinal,
the following hold:
(1) ξ(G/ Resα (G)) ≤ sup+ (α);
(2) If G is compactly generated and α is a limit ordinal, then ξ(G/ Resα (G)) < α.
(3) If G is compactly generated, then Res2 (G) = Res(G).
Proof. Claim (1) is immediate from Proposition 3.12 and Lemma 3.11. Claim (3)
is an easy exercise.
For claim (2), Proposition 3.12 ensures there is a closed normal subgroup K of
G such that ξ(G/R) = ξ(G/K) ≤ α. Since the rank of G/R is a successor ordinal,
we conclude that ξ(G/R) < α.
The rank-α residual of G can be expressed in terms of the rank-α residuals of
compactly generated open subgroups.
Proposition 3.17. Let α ≥ 2 be an at most countable ordinal, G an t.d.l.c.s.c.
group, and (Oi )i∈N be an increasing exhaustion of G by compactly generated open
subgroups. If α is a successor ordinal, then
[
Resα (Oi ).
Resα (G) =
i∈N
If α is a limit ordinal, then
!
Resα (G) =
\
[
β<α
i∈N
Resβ (Oi ) .
Proof. Take β an ordinal and (Oi )i∈N an ⊆-increasing
exhaustion of G by compactly
S
generated openSsubgroups. The union i∈N Resβ (Oi ) is an increasing union of
subgroups, so i∈N Resβ (Oi ) is a subgroup of G. Given a compactly generated
open subgroup L of G, sufficiently
large i are such that L ≤ Oi , hence Resβ (L) ≤
S
Resβ (Oi ). We deduce that i∈N Resβ (Oi ) is in fact the union of S
Resβ (L) as L
ranges over all compactly generated open subgroups of G. Hence, i∈N Resβ (Oi )
is indeed a characteristic subgroup of G.
We are now ready to proveSthe claims. Suppose first that α = β + 1 is a
successor ordinal and set R := i∈N Resα (Oi ). Suppose that K E G is closed with
ξ(G/K) ≤ α. By Proposition 3.5, Oi /(Oi ∩ K) is a compactly generated open
subgroup of G/K of rank at most α for every i. We deduce that Resα (Oi ) ≤ K, so
R ≤ K. Therefore, R ≤ Resα (G).
For the converse inclusion, fix a compactly generated open subgroup L/R of G/R.
For i sufficiently large, L ⊆ Oi R, and there is a quotient map from Oi / Resα (Oi )
to Oi R/R. Applying Proposition 3.5, we deduce that
ξ(L/R) ≤ ξ(Oi R/R) ≤ ξ(Oi / Resα (Oi )) ≤ α.
Thus, ξ(Res(L/R)) ≤ β, and since L is arbitrary, it follows that ξ(G/R) ≤ β + 1 =
α. We conclude that R ≥ Resα (G).
Let us now consider the case when α is a limit ordinal. Set Rβ := Resβ (G).
Since G has no quotients of rank exactly α, we see that
\
Rα =
Rβ .
β<α
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
15
We can now use induction on α and the first part of the proposition applied to Rβ
to write
!
\ [
Resβ (Oi ) ,
Rα =
β<α
i∈N
as required.
Corollary 3.18. Let G be a t.d.l.c.s.c. group and (Oi )i∈N be an increasing exhaustion of G by compactly generated open subgroups. Then
[
Res2 (G) =
Res(Oi ).
i∈N
Remark 3.19. If α is a limit ordinal, it is not necessarily the case that Resα (G) =
S
i∈N Resα (Oi ). For instance, if G is a topologically characteristically simple t.d.l.c.s.c.
group with ξ(G) = ω + 1 (for an example, see §9.3 below), then Resω (G) = G by
Theorem 3.23 below, but the finite rank radical of every compactly generated open
subgroup of G is trivial.
In the case of the rank-2 residual, we also have an easy calculation for the rank
of Res2 (G) in terms of the rank of G. This will put a restriction on the possibilities
for the ranks of topologically characteristically simple t.d.l.c.s.c. groups.
Proposition 3.20. For G a non-compact elementary t.d.l.c.s.c. group, exactly one
of the following occurs:
(1) ξ(G) = ξ(Res2 (G)) = α + 1 where α is a countable transfinite limit ordinal;
(2) Res2 (G) is a non-cocompact subgroup of G, and ξ(G) = ξ(Res2 (G)) + 1.
Proof. Let (Oi )i∈N be an increasing exhaustion of G by compactly generated open
S
subgroups and set R := Res2 (G). By Corollary 3.18, i∈N Res(Oi ) = R . We have
ξ(G) = sup ξ(Res(Oi )) + 1,
i∈N
and by Proposition 3.10, we also have
ξ(R) = sup+ ξ(Res(Oi )).
i∈N
Set β := supi∈N ξ(Res(Oi )). If β is a successor ordinal, then ξ(R) = β, and
ξ(G) = β + 1, so ξ(G) = ξ(R) + 1. In particular, R is a proper subgroup of G with
ξ(G/R) = 2. In view of Lemma 3.8, the only way R can be cocompact is if it is
trivial; however, this would imply that G is compact, contradicting the hypotheses.
We conclude that (2) holds and (1) does not hold.
If β is a limit ordinal, then β is transfinite. We conclude that ξ(R) = ξ(G) =
β + 1. Condition (1) thus holds and (2) does not hold, completing the proof of the
dichotomy.
Corollary 3.21. If G is a non-trivial topologically characteristically simple t.d.l.c.s.c.
group, then exactly one of the following holds:
(1) ξ(G) = 2;
(2) ξ(G) = α + 1 for α some countable transfinite limit ordinal;
(3) G is non-elementary.
16
COLIN D. REID AND PHILLIP R. WESOLEK
Proof. The three cases are clearly mutually exclusive.
Suppose that (3) does not hold, so G is an elementary group. Since G is topologically characteristically simple, Res2 (G) = {1}, or Res2 (G) = G. By Proposition 3.20, Res2 (G) = {1} if and only if G is elementary of rank 2; the case ξ(G) = 1
being eliminated by the hypothesis that G 6= {1}. We thus deduce that (1) holds
if Res2 (G) = {1}. If Res2 (G) = G, then G is non-compact, since ξ(G) > 2. The
dichotomy of Proposition 3.20 thus ensures that ξ(G) = α +1 for α some transfinite
limit ordinal, hence (2) holds.
3.5. The rank-2 series. We conclude this section by isolating a canonical series
of closed characteristic subgroups in a t.d.l.c.s.c. group.
Definition 3.22. For a t.d.l.c.s.c. group G, the (lower) rank-2 series, denoted
by (Gi )i∈N , is defined as follows: G0 := G and Gi+1 := Res2 (Gi ). We call a
t.d.l.c.s.c. group rank-2 resolvable if the rank-2 series stabilizes at {1} in finitely
many steps.
We can think of the rank-2 series as an analogue of the derived series, with rank2 groups playing the role of abelian groups. The next results show this analogy can
be pushed further.
Theorem 3.23. For G a t.d.l.c.s.c. group and (Gi )i∈N the rank-2 series for G, the
following hold:
(1) For each i ∈TN, Gi = Resi+1 (G).
(2) Resω (G) = i∈N Gi .
(3) If G is compactly generated, then Resω (G) = Gj for some j ∈ N; in particular,
the series (Gi )i∈N stabilizes in finitely many steps.
Proof. For claim (1), we argue by induction on i ≥ 1. The base case, i = 1, holds
by definition. Suppose that the claim holds up to i and consider Gi+1 . Let K E G
be such that ξ(G/K) ≤ i+1 and put R := π −1 (Res2 (G/K)) where π : G → G/K is
the usual projection. We see that G1 ≤ R and that ξ(R/K) ≤ i by Proposition 3.20.
The inductive hypothesis now implies that the i-th term of the rank-2 series for R
is contained in K. Since G1 ≤ R, we conclude that the i-th term of the rank-2
series of G1 is contained in K. That is to say, Gi+1 ≤ K, hence Gi+1 ≤ Resi+1 (G).
For the converse inclusion, the induction hypothesis implies that ξ(G1 /Gi+1 ) ≤
i + 1, since Gi+1 is the i-th term of the rank-2 series for G1 . It is easy to see
that Res2 (G/Gi+1 ) = G1 /Gi+1 , hence applying Proposition 3.20, we deduce that
ξ(G/Gi+1 ) ≤ i + 2. Therefore, Gi+1 = Resi+1 (G), completing the induction.
For claim (2), the quotient G/Gj is the largest quotient of G of rank at most
j + 1 by claim (1). In particular,
if G/K is a quotient of G of finite rank, then
T
K ≥ Gj for some j. Hence, j∈N Gj ≤ Resω (G). The other inclusion is obvious,
T
so we conclude that Resω (G) = j∈N Gj .
For the final claim, Proposition 3.16 ensures that G/ Resω (G) has finite rank,
hence Resω (G) = Gj for some j.
The next corollary is now immediate from Theorem 3.23 and Proposition 3.16.
Corollary 3.24. Let G be a t.d.l.c.s.c. group. Then G is an elementary group with
finite rank if and only if G is rank-2 resolvable.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
17
If G does have finite rank, we see that the bound in Lemma 3.6 is achieved by
the extensions coming from the rank-2 series. Thus at least in the case of groups
of finite rank, the bound in Lemma 3.6 is the best possible.
Corollary 3.25. Let G be an elementary t.d.l.c.s.c. group such that ξ(G) < ω, let
(Gi )i∈N be the rank-2 series for G and let i ∈ N. Then
ξ(G) = (ξ(Gi ) − 1) + ξ(G/Gi ).
Proof. By Theorem 3.23 we have Gi = Resi+1 (G). It follows by Proposition 3.16
that ξ(G/Gi ) = j + 1 where 0 ≤ j ≤ i. Then we see that Gi = Resi+1 (G) =
Resj+1 (G) = Gj and that we have a strictly descending chain of subgroups G0 >
G1 > · · · > Gj . In particular, 2 ≤ ξ(Gk ) < ω for all 0 ≤ k < j, so for these k, then
(ξ(Gk ) − 1) is not a limit ordinal. Repeated application of Proposition 3.20 yields
ξ(G) = ξ(Gj ) + j; in turn, ξ(Gj ) + j = (ξ(Gj ) − 1) + ξ(G/Gj ).
4. Factorization of normal compressions in l.c.s.c. groups
Suppose that G and H are l.c.s.c. groups with ψ : G → H a normal compression
map. Since G and H are Polish groups, Proposition 2.6 and Theorem 2.7 supply a
factorization of ψ as
α/
G oψ H
G
ψ
β
#
H
where α is a closed embedding with normal image, β is a quotient map, and ker(β)
is isomorphic to G as a topological group.
Our aim in the present section is to use the structure of locally compact groups
to factorize ψ into a sequence of closed embeddings with normal image and quotient
maps, such that all kernels and cokernels appearing in the sequence are small. This
will allow group-theoretic properties to be transported between G and H.
4.1. The totally disconnected case. We first prove a factorization result for
normal compression maps ψ : G → H of l.c.s.c. groups where ψ restricts to an
isomorphism on the identity components. This is, in particular, the case when G
and H are totally disconnected.
An FC-group is a group in which every conjugacy class is finite; equivalently,
an FC-group is such that every element has a finite index centralizer. The set of
finite conjugacy classes of a group G forms a subgroup called the FC-center of
G. We shall require some basic properties of FC-groups and elements with finite
conjugacy classes.
Lemma 4.1. Let G be a group.
(1) If x ∈ G is an element of finite order such that |G : CG (x)| is finite, then x is
contained in a finite normal subgroup of G. (Dietzmann’s Lemma [18, 14.5.7].)
(2) If G is an FC-group, then [G, G] and G/Z(G) are torsion. Hence, every element
of [G, G] or G/Z(G) is contained in a finite normal subgroup. ([18, Section
14.5])
Proving our factorization theorem requires two technical lemmas.
18
COLIN D. REID AND PHILLIP R. WESOLEK
Lemma 4.2. Let G and H be l.c.s.c. groups with ψ : G → H a normal compression.
For U ≤ H and K ≤ G closed subgroups, form H o := G oψ U where U acts via
the ψ-equivariant action and set
∆K := {(k −1 , ψ(k)) ∈ H o | k ∈ K}.
Then,
(1) The set ∆K is a closed subgroup of H o with ∆K ≤ CH o (G × {1}).
(2) If ψ restricts to a homeomorphism from K to ψ(K) and ψ(K) is a closed
normal subgroup of U , then ∆K is normal in H o , and q ◦ ι is a closed map
where q : H o → H o /∆K is the quotient map and ι : G → H o is the usual
embedding.
Proof. For (1), easy calculations verify that ∆K is a subgroup and that ∆K commutes with G×{1}. That ∆K is closed follows since K is closed and ψ is continuous.
For (2), since ψ(K) is normal in U , the subgroup ∆K is normalized by {1} × U ,
and part (1) implies ∆K is centralized by G×{1}. The subgroup ∆K is thus normal
in H o . To show that q ◦ ι is a closed map, it suffices to show ι(R)∆K is closed in
H o for every closed subset R of G. Fix such an R. The set G × ψ(K) ⊆ H o is
closed, and there is a continuous map γ : G × ψ(K) → G defined by (g, ψ(k)) 7→ gk.
Moreover, ι(R)∆K is the preimage γ −1 (R). We deduce that ι(R)∆K is a closed
subset of H o , proving that q ◦ ι is a closed map.
Lemma 4.3. Let G and H be l.c.s.c. groups with ψ : G → H a normal compression
map. If U ∈ U(G), then NH (ψ(U )) is open in H. Furthermore, for every L ∈ U(H),
there exists W ≤o U such that ψ(W ) is normalized by L.
Proof. Consider U(G) as a collection of subgroups of G × {1} in the semidirect
product H o = Goψ H, which is a Polish group by Proposition 2.6. Since U ∈ U(G)
is a closed subgroup of H o , the normalizer NH o (U ) is closed. On the other hand,
the ψ-equivariant action is continuous, so H o y U(G) as permutations. Recalling
that U(G) is countable, we deduce that NH o (U ) is a closed subgroup of H o with
countable index. The Baire Category Theorem now implies that NH o (U ) is open
in H o , and it follows that NH (ψ(U )) is open in H.
For the second claim, consider L an almost connected subgroup of H and fix
U ∈ U(G). For each l ∈ L, the subgroup ψ −1 (lψ(U )l−1 ) is exactly the image of U
by l under the ψ-equivariant action of H on G, hence ψ −1 (lψ(U )l−1 ) is again an
element of U(G). On the other hand, every open subgroup of L has finite index.
We deduce that |L : NL (ψ(U ))| is finite, and thus,
!
\
−1
−1
W := ψ
lψ(U )l
l∈L
is an element of U(G). Furthermore, ψ(W ) is normalized by L, verifying the claim.
A covering map between Polish groups G and H is a surjective, continuous
homomorphism ψ : G → H such that there exist identity neighborhoods U in G
and V in H for which ψ U : U → V is a homeomorphism. Covering maps between
l.c.s.c. groups can be characterized as follows: ψ : G → H is a covering map if and
only if it is a surjective continuous group homomorphism such that the kernel is
discrete.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
19
We now obtain a decomposition of a normal compression ψ of t.d.l.c.s.c. groups
as ψ = β ◦ α, where α is a closed embedding and β is a covering map.
Theorem 4.4. Suppose that G and H are l.c.s.c. groups with ψ : G → H a normal
compression map and that ψ restricts to a homeomorphism from G◦ to H ◦ . Then
there is an l.c.s.c. group H̃, a closed embedding α : G → H̃ with normal image, and
a covering map β : H̃ → H such that
(1)
(2)
(3)
(4)
ψ = β ◦ α;
H̃/α(G) is compact;
ker(β) lies in the FC-center of H̃ and centralizes α(G); and
H̃ = α(G) ker(β).
If G and H are t.d.l.c.s.c. groups, then H̃ is also a t.d.l.c.s.c. group.
Proof. Fix O ∈ U(H) and form H o := G o O where O acts on G via the ψequivariant action. By Proposition 2.6, H o with the product topology is an l.c.s.c.
group. Fix W ∈ U(G) such that ψ(W ) E O, as given by Lemma 4.3, and observe
that V := W o O is an almost connected open subgroup of H o .
We now define the group H̃ and the maps α and β. For W as fixed in the
previous paragraph, Lemma 4.2 ensures
∆W := {(w−1 , ψ(w)) ∈ H o | w ∈ W }
is a closed normal subgroup of H o . We put H̃ := H o /∆W . For the map α, let
ι : G → H o be the obvious inclusion and let q : H o → H̃ be the usual projection.
We set α := q ◦ ι. To define β, consider the quotient map π : H o → H as in
Theorem 2.7. The kernel of π is exactly ∆G , and since ∆G contains ∆W , the map
π induces the desired map β : H̃ → H.
Let us now check the claimed properties. The maps α and β are clearly continuous. It is also clear that α is injective with normal image. That α is a
closed map is given by Lemma 4.2. To see β is a covering map, we remark
that (g −1 , ψ(g)) ∈ ∆W V implies g ∈ W , so ∆G ∩ (∆W V ) = ∆W . Therefore,
ker(β) ∩ q(V ) = {1}, so ker(β) is discrete, verifying that β is a covering map.
Claims (1) and (4) are immediate from Theorem 2.7. For claim (2), the subgroup
ι(G)∆W contains the connected component of H o , since ψ(G◦ ) = H ◦ . Recalling
that O is almost connected, it follows that the quotient
H o /ι(G)∆W ' H̃/α(G)
is compact.
For claim (3), ker(β) equals ∆G /∆W . The group ∆G is centralized by ι(G),
hence ker(β) is centralized by α(G), verifying the second claim of (3). In view of
(2), every element of ker(β) has a cocompact centralizer in H̃. On the other hand,
ker(β) is a discrete normal subgroup of H̃, so every element indeed has an open
centralizer. We thus deduce every x ∈ ker(β) has a finite index centralizer and so
has a finite conjugacy class. That is to say, ker(β) is in the FC-center of H̃.
A locally compact group G is compactly presented if there is a presentation
for G as an abstract group G = hS | Ri such that S has compact image in G and
R consists of relators with bounded length. For such groups, we obtain a stronger
factorization.
20
COLIN D. REID AND PHILLIP R. WESOLEK
Corollary 4.5. Suppose that G and H are l.c.s.c. groups with ψ : G → H a normal
compression map and that ψ restricts to a homeomorphism from G◦ to H ◦ . If G
is compactly generated and H is compactly presented, then H̃, α, and β given by
Theorem 4.4 can be chosen such that H̃/α(G) is compact abelian and ker(β) is
central, finitely generated, and torsion free.
Proof. Take O ∈ U(H), find W ∈ U(G) such that ψ(W ) E O, and form H̃ =
G o O/∆W as given by Theorem 4.4. The group H̃ is compactly generated since it
contains G as a cocompact subgroup.
In view of [7, Proposition 8.A.10], ker(β) is the normal closure of a compact set.
Applying Theorem 4.4, ker(β) is discrete and contained in the F C-center of H̃,
hence ker(β) is indeed finitely generated by some finite subset X. Each x ∈ X is
centralized by a finite index open subgroup of H̃, so C := CH̃ (ker(β)) is a finite
index subgroup of H̃.
Letting q : G o O → H̃ be the usual projection, we may find O0 ≤ O open with
finite index such that q({1} × O0 ) ≤ C. Since q ◦ ι(G) = ι(G)∆W /∆W centralizes
ker(β), we may assume further that ψ(W ) ≤ O0 . We now apply Theorem 4.4 to O0
and W to form H̃ 0 , α0 , and β 0 . The group H̃ 0 is an open subgroup of H̃ contained
in C, and thus, ker(β 0 ) is central in H̃ 0 . We thus deduce that ker(β 0 ) is a central
finitely generated subgroup.
Observing that ker(β 0 ) = ∆G /∆W , the projection on the second coordinate gives
an injective homomorphism ker(β 0 ) → O0 /ψ(W ). The group O0 /ψ(W ) is a profinite
group since ψ(G◦ ) = H ◦ . We may thus find O00 ≤ O0 a finite index open subgroup
such that ψ(W ) ≤ O00 and O00 /ψ(W ) avoids the torsion elements of the image of
ker(β 0 ) in O0 /ψ(W ). Applying Theorem 4.4 to O00 and W , we form H̃ 00 , α00 , and
β 00 . The group H̃ 00 is an open subgroup of H̃ 0 contained in C, and thus, ker(β 00 )
is central in H̃ 00 . Furthermore, the map ∆G /∆W → O00 /ψ(W ) must have torsion
free image. We thus deduce that ker(β 00 ) is central, finitely generated, and torsion
free. Without loss of generality, we now assume O = O00 .
Appealing to Theorem 4.4 again, there is a normal compression map θ : ker(β) →
H̃/α(G). Since ker(β) is abelian, we conclude that H̃/α(G) is abelian, verifying
the corollary.
4.2. The general case. We now relax the requirement that ψ restricts to an
isomorphism from G◦ to H ◦ . To deal with the connected components, we require
several preliminary results.
For an l.c.s.c. group G, the group of continuous automorphisms, denoted by
Aut(G), carries a Polish topology, namely the modified compact-open topology; this
topology is alternatively called the Braconnier topology or the Arens g-topology.
When we consider Aut(G) as a topological group, we always mean with respect to
this to this topology.
Lemma 4.6. Suppose that G and H are l.c.s.c. groups with ψ : G → H a continuous embedding with dense image and that A ≤ Aut(G) is closed. If every α ∈ A
has (α) ∈ Aut(H) such that ψ ◦ α = (α) ◦ ψ, then the map : A → Aut(H) is
continuous.
Proof. That ψ has a dense image ensures the map is a group homomorphism. To
show that is continuous, it thus suffices to verify that it is Borel measurable, since
A and Aut(H) are Polish groups; see, for example, [9, (9.10)].
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
21
A sub-basis at 1 for Aut(H) is given by the sets
AH (K, U ) := {β ∈ Aut(H) | ∀x ∈ K β(x)x−1 ∈ U and β −1 (x)x−1 ∈ U }
where K ⊆ H is compact and U ⊆ H is an open neighborhood of 1. We may
further restrict to compact sets K which are the closure of an open set. To show
is Borel measurable, it is therefore enough to show the preimage of AH (K, U ) is
Borel-measurable in A for U open and K = O with O open and relatively compact.
Fix such a K and U .
Fix a countable basis B for the topology on H. Since U is open and H is a
regular space, we may cover U by B ∈ B such that B ⊆ U . Let C ⊆ B be a cover of
U by such B; we may additionally assume C is closed under finite unions.
The set
S
ψ −1 (K) is a non-trivial closed set in G, so we may write ψ −1 (K) = i∈N Ji where
the Ji are compact sets.
For B ∈ C and i ∈ N, we now consider
Σ(i, B) := AG Ji , ψ −1 (B) ∩ A.
We claim
−1 (AH (K, U )) =
[ \
Σ(i, B).
B∈C i∈N
Since the Σ(i,SB) are
T open sets in A, this will prove is Borel measurable.
Take β ∈ B∈C i∈N Σ(i, B) and say that B ∈ C is such that β ∈ Σ(i, B) for
all i. Take x ∈ ψ(ψ −1 (K)) and suppose that x = ψ(y) for y ∈ ψ −1 (K). Since
(β)±1 ◦ ψ = ψ ◦ β ±1 , we have
(β)±1 (x)x−1 = ψ(β ±1 (y)y −1 ) ∈ B.
The set K is the closure of an open set, so ψ(ψ −1 (K)) is dense in K. We thus deduce
that (β)±1 (x)x−1 ∈ B ⊆ U for all x ∈ K. That is to say, (β) ∈ AH (K, U ).
On the other hand, take β ∈ −1 (AH (K, U )). Since K is compact, C is closed
under finite unions and the maps K → U by x 7→ (β)±1 (x)x−1 are continuous,
there is some B ∈ C for which B contains the image of both maps. For each i and
y ∈ Ji , we thus have T(β)±1 (ψ(y))ψ(y)−1 ∈ B. Hence, β ±1 (y)y −1 ∈ ψ −1 (B). It
now follows that β ∈ i∈N Σ(i, B), and the claim is demonstrated.
A locally compact group G is a (CA) group if the natural map ρ : G → Aut(G)
has a closed image, where Aut(G) is endowed with the modified compact-open
topology.
Theorem 4.7 (Zerling, [21, Main Theorem]). Suppose that G is a connected locally
compact group that is not a (CA) group. Then there is a (CA) locally compact
connected group N , a torus T that acts continuously on N , and a vector group V
that maps densely into T , such that:
(1) P := N o T is a (CA) locally compact group.
(2) G ' N o V .
(3) Every automorphism of G as a topological group extends to an automorphism
of P as a topological group.
Given a normal compression map ψ : G → H of connected l.c.s.c. groups, we
have control over the derived group of H.
Corollary 4.8. If G and H are connected l.c.s.c. groups with ψ : G → H a normal
compression map, then [H, H] ≤ ψ(G)Z(H).
22
COLIN D. REID AND PHILLIP R. WESOLEK
Proof. Let P and N be as given in Theorem 4.7 for G and let : Aut(G) → Aut(P )
be the group homomorphism, necessarily unique, taking each automorphism of G
to its extension to an automorphism of P . Lemma 4.6 ensures is continuous.
Define ρ : H → Aut(G) to be the map induced by the ψ-equivariant action of H
on G.
The group ψ(G) is dense in H, so by continuity, ρψ(G) is dense in ρ(H). Since
P is a (CA) group, we see that Inn(P ) is a closed subgroup of Aut(P ) that contains
ρψ(G). Hence, ρ(H) ≤ Inn(P ).
We thus obtain a continuous homomorphism θ : H → P/Z(P ). In particular,
θ([H, H]) ≤ [P, P ]Z(P )/Z(P ), and since [P, P ] ≤ N ≤ G, every element of [H, H]
induces an inner automorphism of G. We conclude that [H, H] ≤ ψ(G)CH (ψ(G)).
Since CH (ψ(G)) = Z(H), we indeed have [H, H] ≤ ψ(G)Z(H), as desired.
The following special case is immediate.
Corollary 4.9. Suppose that G and H are connected l.c.s.c. groups with ψ : G → H
a normal compression map. If G is topologically perfect and H is centerless, then
ψ is an isomorphism of topological groups.
We now prove the desired factorization result for arbitrary normal compression
maps between l.c.s.c. groups.
Theorem 4.10. Suppose that G and H are l.c.s.c. groups with ψ : G → H a
normal compression map. Then there are continuous homomorphisms and l.c.s.c.
groups
α
β1
α
β2
α
β3
1
2
3
G = G0 −→
K1 −→ G1 −→
K2 −→ G2 −→
K3 −→ H
such that the following properties hold:
(1) ψ = β3 ◦ α3 ◦ β2 ◦ α2 ◦ β1 ◦ α1 ;
(2) The αi are closed embeddings with normal image, the βi are quotient maps,
and β3 is a covering map;
(3) The groups K1 /α1 (G) and ker(β1 ) are abelian;
(4) The groups K2 /α2 (G1 ) and ker(β2 ) are solvable of derived length at most 3;
(5) K3 /α3 (G2 ) is compact, and K3 = α3 (G2 ) ker(β3 ); and
(6) ker(β3 ) lies in the FC-center of K3 and centralizes α3 (G2 ).
Proof. We build the maps αi and βi and the groups Ki and Gi in stages.
Define R1 := ψ(G◦ ) ≤ H ◦ and K1 := G oψ Z(R1 ) where Z(R1 ) acts via the
ψ-equivariant action. Set
∆1 := {(g −1 , ψ(g)) ∈ K1 | g ∈ G}
and G1 := K1 /∆1 . Let α1 be the natural closed embedding of G into K1 and let
β1 be the natural quotient map from K1 to G1 . The groups K1 and G1 are l.c.s.c.
groups by Proposition 2.6 and Lemma 4.2, and the cokernel of α1 and the kernel
of β1 are abelian.
Let ψ1 : G1 → H be such that ψ1 ◦ β1 ◦ α1 = ψ; that is, ψ1 is the map
(g, r)∆1 7→ ψ(g)r. The map ψ1 is a normal compression map. Additionally, the
map
ψ1 ◦ β1 α1 (G◦ ) : α1 (G◦ ) → R1
is a normal compression map from the connected l.c.s.c. group α1 (G◦ ) to the connected l.c.s.c. group R1 . By Corollary 4.8, we have that [R1 , R1 ] ≤ ψ(G◦ )Z(R1 ).
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
23
Both ψ(G◦ ) and Z(R1 ) are subgroups of ψ1 (G1 ), hence [R1 , R1 ] ≤ ψ1 (G1 ). The
group T1 := ψ1−1 ([R1 , R1 ]) is thus isomorphic as a topological group to [R1 , R1 ] via
ψ1 .
Define A2 := G1 oψ1 H ◦ , set ∆2 := {(t−1 , ψ1 (t)) ∈ A2 | t ∈ T1 }, and put K2 :=
A2 /∆2 . By Lemma 4.2, K2 is an l.c.s.c. group, and the natural map α2 : G1 → K2
is a closed embedding. Letting ρ be the composition H ◦ → A2 → K2 , the image
α2 (G1 ) contains ρ([R1 , R1 ]). Lemma 2.9 ensures [[H, H ◦ ], H ◦ ] ≤ R1 , and thus,
K2 /α2 (G1 ) is solvable of derived length at most 3.
Define ∆3 := {(g −1 , ψ1 (g)) ∈ A2 | g ∈ G1 } and G2 := A2 /∆3 . Define the map
β2 to be the natural quotient map from K2 to G2 . The kernel of β2 is ker(β2 ) =
∆3 /∆2 . The projection onto the second coordinate of ∆3 induces an injective
homomorphism ∆3 /∆2 → H ◦ /[R1 , R1 ]. Since [[H, H ◦ ], H ◦ ] ≤ R1 , we conclude
that H ◦ /[R1 , R1 ] is solvable of derived length at most 3. Hence, ker β2 is solvable
of derived length at most 3.
As with G1 , we obtain a normal compression map ψ2 : G2 → H such that
ψ2 ◦ β2 ◦ α2 ◦ β1 ◦ α1 = ψ.
The group A2 already contains a copy of H ◦ , so we have ψ2 (G2 ) ≥ H ◦ . The map ψ2
thus restricts to an isomorphism from (G2 )◦ to H ◦ . We finally obtain the maps α3
and β3 by applying Theorem 4.4 to ψ2 : G2 → H, completing the construction.
5. Properties invariant under normal compressions
We now explore to what extent various properties are preserved under normal
compressions. Several of the properties studied here and the derived preservation
theorems will be used later to analyze the structure of chief factors.
5.1. Compact generation. Given a t.d.l.c.s.c. group G and a normal compression
H of G with H compactly
generated, it does not followL
that G is compactly generQ
ated. For example, i∈N Fi is a normal compression of i∈N Fi for any sequence of
finite groups (Fi )i∈N , and R is a normal compression of Q. Our next theorem shows
that if H is compactly generated, then profinite quotients and connected central
factors of H are the only potential obstacles to the compact generation of G.
Theorem 5.1. Suppose that G and H are l.c.s.c. groups with ψ : G → H a normal
compression map and set R := ψ −1 (H ◦ ).
(1) If G is compactly generated, then H is compactly generated.
(2) If H is compactly generated and R/[R, G] is compactly generated, then there exists an increasing exhaustion (Ni )i∈N of G by compactly generated open normal
subgroups such that ψ(Ni ) is cocompact and normal in H for each i ∈ N.
Proof. Claim (1) is immediate.
For (2), we see that R ≥ G◦ , and Lemma 2.9 implies [R, G] ≤ G◦ . The quotient
R/[R, G] is compactly generated, so R/G◦ is also compactly generated. It now
follows that G is compactly generated if and only if G/R is compactly generated.
We thus reduce to G/R, and since ψ induces a normal compression map ψ̃ : G/R →
H/H ◦ , we may suppose that both G and H are totally disconnected.
Using Lemma 4.3, we fix W ∈ U(G) and U ∈ U(H) such that ψ(W ) is a normal
subgroup of U . Let {gi }i∈N be a countable dense subset of G with 1 = g0 and
set Si := {ψ(g0 ), . . . , ψ(gi )}. By Proposition 2.2, there is a finite symmetric set
24
COLIN D. REID AND PHILLIP R. WESOLEK
Ai ⊆ H containing Si such that U Ai U = Ai U and Ai U generates H. Since ψ(G)
is dense, we can choose Ai to be a subset of ψ(G); we can also choose (Ai )i∈N to
be an increasing sequence.
The group U acts continuously on G via the ψ-equivariant action. Under this
action, the orbit U.ψ −1 (Ai ) is a compact subset of G, so we obtain a compactly
generated open subgroup Ni := hU.ψ −1 (Ai ), W i that contains {g0 , . . . , gi }. The
sequence (Ni )i∈N is thus an increasing exhaustion of G by compactly generated
open subgroups.
Fix i, set N := Ni , and set A := Ai . The group ψ(N ) is generated by AU ∪ψ(W ).
That ψ(W ) E U ensures that U normalizes ψ(N ), and plainly, A ⊆ ψ(N ). Since
H = hA, U i, we deduce that ψ(N ) E H and hence N E G. The group U maps
surjectively onto H/ψ(N ), so ψ(N ) is also cocompact in H. The proof is now
complete.
Corollary 5.2. Suppose that G and H are t.d.l.c.s.c. groups with H a normal
compression of G. Then G is compactly generated if and only if H is compactly
generated and every discrete quotient of G is finitely generated.
Proof. If G has an infinitely generated discrete quotient, then certainly G is not
compactly generated, and if H is not compactly generated, then G is not compactly generated by Theorem 5.1. Conversely, if H is compactly generated and
every discrete quotient of G is finitely generated, then in particular G/Ni is finitely
generated for all i where (Ni )i∈N is the exhaustion given by Theorem 5.1. It now
follows that G = Nj for j large enough, and thus, G is compactly generated.
Corollary 5.3. Suppose that G and H are non-abelian non-compact topologically
characteristically simple l.c.s.c. groups with H a normal compression of G. Then
G is compactly generated if and only if H is compactly generated.
Proof. Since both G and H are topologically characteristically simple groups, Corollary 2.11 ensures either both are totally disconnected or both are connected. Connected groups are compactly generated, so the corollary is immediate in the latter
case. We thus assume that G and H are t.d.l.c.s.c. groups.
The forward implication is given by Theorem 5.1. Conversely, suppose that H is
compactly generated and let ψ : G → H be a normal compression map. Appealing
to Theorem 5.1, there is N E G open and compactly generated such that ψ(N ) is
cocompact in H. The quotient H/ψ(N ) is therefore a profinite group.
If this quotient is non-trivial, then Res(H) is a proper characteristic closed subgroup of H, hence Res(H) = {1}. We thus deduce that H is residually discrete.
Applying [6, Corollary 4.1], H admits a compact open normal subgroup L. The locally elliptic radical of H is then a non-trivial characteristic closed subgroup, hence
H is a compactly generated locally elliptic group. We conclude that H is compact,
which contradicts the hypotheses.
The group ψ(N ) thus equals H, and Proposition 2.8 implies [G, G] ≤ N . Since G
is non-abelian and topologically characteristically simple, we conclude that N = G.
In particular, G is compactly generated.
5.2. Decomposition rank. The property of being elementary and the associated decomposition rank are outright invariant under normal compressions between
t.d.l.c.s.c. groups.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
25
Proposition 5.4. Suppose that G and H are t.d.l.c.s.c. groups with H a normal
compression of G. Then G is elementary with decomposition rank α if and only if
H is elementary with decomposition rank α.
Proof. Let ψ : G → H be a normal compression map. Let H̃, α : G → H, and
β : H̃ → H be as given by Theorem 4.4.
Suppose first that G is an elementary group. Since G embeds as a cocompact
normal subgroup of H̃, Lemma 3.8 implies that ξ(H̃) = ξ(G). The group H is
a quotient of H̃, so in view of Proposition 3.5, we deduce that H is elementary
with ξ(H) ≤ ξ(G). On the other hand, ψ : G → H is a continuous injective
homomorphism, and Proposition 3.5 thus shows that ξ(G) ≤ ξ(H). We conclude
that ξ(H) = ξ(G).
Conversely, if H is elementary, then Proposition 3.5 ensures that G is elementary.
The forward direction then implies that ξ(G) = ξ(H).
5.3. Amenability. A locally compact group G is amenable if every continuous
affine action of G on a non-empty compact convex subset of a locally convex topological vector space has a fixed point; we direct the reader to [2, Appendix G] for an
account of the basic theory of amenable groups. Amenability is a second property
outright preserved under normal compressions.
Let us recall several basic facts.
Lemma 5.5 (See [2, Appendix G]). Let G be an l.c.s.c. group.
(1) Every compact or virtually solvable l.c.s.c. group is amenable.
(2) If K is a closed normal subgroup of G, then G is amenable if and only if both
K and G/K are amenable.
(3) If G is amenable, then every closed subgroup of G is amenable.
(4) If H is an amenable l.c.s.c. group and ψ : H → G is a continuous homomorphism, then ψ(H) is amenable.
S
(5) If G = α∈I Aα with (Aα )α∈I a directed system of closed amenable subgroups
of G, then G is amenable.
Observe that the FC-groups are also amenable, with respect to the discrete
topology; this is an easy exercise using Lemma 5.5.
Proposition 5.6. Suppose that G and H are l.c.s.c. groups with H a normal
compression of G. Then G is amenable if and only if H is amenable.
Proof. If G is amenable, then H is amenable by Lemma 5.5.
Conversely, suppose that H is amenable. Applying Theorem 4.10, we obtain a
factorization of ψ : G → H as ψ = β3 ◦ α3 ◦ β2 ◦ α2 ◦ β1 ◦ α1 where
α
β1
α
β2
α
β3
1
2
3
G = G0 −→
K1 −→ G1 −→
K2 −→ G2 −→
K3 −→ H
The kernel ker(β3 ) is an FC-group, so it is amenable. Since H is also amenable,
Lemma 5.5 ensures that K3 is amenable. Each γ ∈ {α1 , α2 , α3 , β1 , β2 } is either
a closed embedding or a quotient map with solvable kernel. In either case, if the
range of γ is amenable, then so is the domain. Since K3 is amenable, we deduce
that G is amenable.
5.4. Quasi-discreteness. Just as with compact generation, discreteness is not a
normal compression invariant, even if we restrict to compressions
between
L
Q nonabelian topologically characteristically simple groups; cf.
A
and
5
i∈Z
i∈Z A5 .
26
COLIN D. REID AND PHILLIP R. WESOLEK
There is a more general property that is close to discreteness and admits better
preservation properties: A l.c.s.c. group G is said to be quasi-discrete if G has a
dense quasi-center QZ(G) := {g ∈ G | CG (g) is open}.
Remark 5.7. For an l.c.s.c. group G, the quasi-center QZ(G) is the union of all
countable conjugacy classes in G, so QZ(G) is the union of all countable normal
subgroups of G. This gives the following characterization: For G an l.c.s.c. group,
G is quasi-discrete if and only if G contains a countable dense normal subgroup.
The quasi-discrete l.c.s.c. groups are thus exactly the l.c.s.c. groups that arise as
normal compressions of countable discrete groups.
Our proof requires a classical commutator identity; recall that [x, y, z] := [[x, y], z],
[a, b] = aba−1 b−1 , and ab = bab−1 .
Fact 5.8 (Hall–Witt identity). For G a group and any a, b, c ∈ G, the following
holds:
[a−1 , c, b]a [b−1 , a, c]b [c−1 , b, a]c = 1.
Theorem 5.9. Suppose that G and H are l.c.s.c. groups with H a normal compression of G.
(1) If H is quasi-discrete, then G/QZ(G) is metabelian.
(2) If G is quasi-discrete, then H/QZ(H) is metabelian.
Proof. Let ψ : G → H be a normal compression map.
For (1), set Q := QZ(H). For g ∈ G such that ψ(g) ∈ Q, the injectivity
of ψ ensures that ψ −1 (CH (ψ(g))) = CG (g), so CG (g) is open. We deduce that
ψ −1 (Q) ≤ QZ(G).
Setting R := ψ(G) ∩ Q, the subgroups ψ(G) and Q commute modulo R, and
since ψ(G) and Q are both dense normal subgroups of H, the group H/R is abelian.
The group G/K with K := ψ −1 (R) is therefore abelian. The map ψ : K → R is a
normal compression map, and K ∩ QZ(G) is a closed normal subgroup of K with
dense image in R. Proposition 2.8 ensures that [K, K] ≤ K ∩QZ(G). It now follows
that G/QZ(G) is metabelian, verifying (1).
For (2), each g ∈ QZ(G) centralizes G◦ , so G◦ ≤ Z(G), which implies that
Z := ψ(G◦ ) is central in H. We thus obtain a normal compression map G/G◦ →
H/Z with G/G◦ quasi-discrete. Let K := π −1 (QZ(H/Z)) with π : H → H/Z the
usual projection. For each m, n ∈ K, there is an open subgroup O ≤ H such that
[O, m] ∪ [O, n] ⊆ Z. Therefore, [o−1 , m, n] = [n−1 , o, m] = 1, and the Hall-Witt
identity with a := o, c := m, and b := n implies [m−1 , n, o] = 1 for all o ∈ O. We
deduce that [m−1 , n] ∈ QZ(H) for all m, n ∈ K, and thus, [K, K] ≤ QZ(H).
To show that H/QZ(H) is metabelian, it suffices to verify that H/K is abelian.
We observe that H/K ' (H/Z)/(K/Z), so it is enough to show that (H/Z)/QZ(H/Z)
is abelian. By passing to H/Z, we assume that Z = {1} and argue that H/QZ(H)
is abelian. Given this reduction, G is a t.d.l.c. group.
Fix n, m ∈ QZ(G) and W ∈ U(G) such that W centralizes n and m. Appealing
to Lemma 4.3, we may find V ∈ U(H) such that ψ(W ) E V . We now consider
R := hψ(n), ψ(m), ψ(W ), V i ≤ H. The subgroup L := ψ(W ) is a compact normal
subgroup of R that centralizes ψ(m) and ψ(n). Furthermore, (ψ(G) ∩ R)L/L is
countable and normal in R/L, so it is quasi-central in R/L. We may thus find an
open subgroup V 0 of V such that [V 0 , ψ(n)] ∪ [V 0 , ψ(m)] ⊆ L. Fixing v ∈ V 0 , we
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
27
see that [v −1 , ψ(m), ψ(n)] = 1 and [ψ(n)−1 , v, ψ(m)] = 1. The Hall–Witt identity
with a := v, b := ψ(n), and c := ψ(m) thus implies that [ψ(m)−1 , ψ(n), v] = 1. It
follows that [ψ(m)−1 , ψ(n)] = ψ([m−1 , n]) has an open centralizer in H.
We have now established that ψ([QZ(G), QZ(G)]) ≤ QZ(H). On the other hand,
QZ(G) is dense in G, so by continuity, ψ(QZ(G)) is a dense subgroup of H. We
deduce that
[H, H] ≤ ψ([QZ(G), QZ(G)]).
The group H/QZ(H) is thus abelian, completing the proof of (2).
Corollary 5.10. Suppose that G and H are topologically perfect l.c.s.c. groups
with H a normal compression of G. Then H is quasi-discrete if and only if G is
quasi-discrete.
6. Generalized direct products
We now relate locally compact generalized central products to local direct products, showing in particular that any t.d.l.c. quasi-product is a normal compression of
a local direct product. We then establish a structure theorem for locally compact
groups of semisimple type. Our discussion concludes by considering generalized
central products of elementary groups.
6.1. Preliminaries. For a topological group G and J a set of closed subgroups,
we define GJ := hN | N ∈ Ji.
Definition 6.1. Suppose that G is a topological group and S ⊆ N (G). The set
S is a generalized central factorization of G, if GS = G and [N, M ] = {1} for
any distinct N and M in S. In such a case, (G, S) is said to be a generalized
central product; we will also say that G is a generalized central product, when
the factorization is implicit.
We say S is a quasi-direct factorization of G if GS = G and S has the
following topological independence property:
\
\
∀X ⊆ P(S) :
X=∅⇒
GA = {1}.
A∈X
In such a case, (G, S) (or G) is said to be a quasi-product. A subgroup H of G
is a quasi-factor of G if it is an element of some quasi-direct factorization of G.
A factorization S is non-trivial if |S| ≥ 2.
We shall make use of a general property of generalized central products.
Theorem 6.2 ([14, Theorem 1.7]). Let G be a topological group with S a non-trivial
generalized central factorization of G. Then the diagonal map
Y
d:G→
G/GS\{N }
N ∈S
is a continuous homomorphism such that d(G) ∩ G/GS\{N } is dense in G/GS\{N }
for every N ∈ S and ker(d) is central in G. Furthermore, S is a quasi-direct
factorization of G if and only if d is injective.
Theorem 6.2 has a useful, immediate corollary.
Corollary 6.3. Let G be a topological group with S a non-trivial generalized
central
factorization of G. Then S is a quasi-direct factorization if and only if
T
G
S\{N } = {1}.
N ∈S
28
COLIN D. REID AND PHILLIP R. WESOLEK
At various points, we will take quotients of generalized central products. There
is a natural quasi-product one can extract from such a quotient:
Proposition 6.4 ([14, Proposition 4.7]). Let G be a topological group with S a nontrivial generalized central factorization of G and let N be a closed normal subgroup
of G. Set
\
M :=
GS\{S} N .
S∈S
Then M/N is central in G/N , and
S/M := {SM /M | S ∈ S and S 6≤ M }
is a quasi-direct factorization of G/M .
The canonical examples of quasi-products are the local direct products:
Definition 6.5. Suppose that (Gi )i∈I is a set of topological groups and that there
is a distinguished open subgroup Oi ≤ L
Gi for each i ∈ I. The local direct product
of (Gi )i∈I over (Oi )i∈I is denoted by i∈I (Gi , Oi ) and defined to be
(
)
G
f :I→
Gi | f (i) ∈ Gi and f (i) ∈ Oi for all but finitely many i ∈ I
i∈I
with the group topology such that the natural embedding of
product topology is continuous and open.
Q
i∈I
Oi with the
In the setting of locally compact groups, a notion of convergence becomes useful
to study generalized products.
Definition 6.6. A set S of closed subgroups of a topological group G is compactly
convergent if for every compact set C and every identity neighborhood U in G,
there exists a cofinite subset S 0 ⊆ S such that GS 0 ∩ C ⊆ U .
This notion of convergence arises in quasi-products.
Lemma 6.7. If (G, S) is a quasi-product, then S is compactly convergent in G.
Q
Proof. Via Theorem 6.2, the diagonal map d : G → NQ∈S G/GS\{N } is injective
and continuous. For any subset
Q J ⊆ S, we have d(GJ ) ≤ N ∈J G/GS\{N } . For any
identity neighborhood V in N ∈S G/GS\{N } , there thus exists a cofinite subset S 0
of S such that d(GS 0 ) ⊆ V .
Fix C a compact set and U an identity neighborhood in G; without loss of
generality, we may assume that U is open and that 1 ∈ C. Since d is injective
and continuous, it Q
restricts to a homeomorphism C → d(C). There is then an
open subset V of N ∈S G/GS\{N } such that d(C) ∩ V = d(C ∩ U ), and since
1 = d(1) ∈ d(C ∩ U ), the set V is an identity neighborhood. We deduce there is a
cofinite subset S 0 of S such that d(GS 0 ) ⊆ V , and hence
d(GS 0 ∩ C) = d(GS 0 ) ∩ d(C) ⊆ V ∩ d(C) = d(C ∩ U ) ⊆ d(U ).
Therefore, GS 0 ∩ C ⊆ U , so S is compactly convergent.
With topologically perfect centerless quasi-products G, there is a connection
between the connectedness of G and the connectedness of its quasi-factors.
Lemma 6.8. Let G be a topological group. If A is a centerless closed normal
subgroup of G, then G/CG (A) is centerless.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
29
Proof. Let π : A → G/CG (A) be the restriction of the usual quotient map. For
hCG (A) ∈ Z(G/CG (A)), we see that [a, h] ∈ A ∩ CG (A) for all a ∈ A. Since
Z(A) = {1}, it must be the case that [a, h] = 1. We conclude that h ∈ CG (A), and
thus, hCG (A) is the trivial coset.
Lemma 6.9. Let G be a centerless Polish group with S a quasi-direct factorization
of G. If every N ∈ S is topologically perfect and totally disconnected, then so is G.
Proof. It is immediate that G is topologically perfect and that every N ∈ M is
centerless. For N ∈ M, the quotient G/CG (N ) is a normal compression of N . The
quotient is then topologically perfect, and it is centerless by Lemma 6.8. Applying
Corollary 2.11, we conclude that G/CG (N ) is totally disconnected, and thus, G◦ ≤
CG (N ). It now follows that G◦ ≤ Z(G) = {1}, so G is totally disconnected.
Our primary interest in generalized direct products is their connection with
groups of semisimple type.
Definition 6.10. Let G be a topological group. A component of G is a closed
subgroup M of G such that the following conditions hold:
(a) M is normal in hhM ii.
(b) M/Z(M ) is non-abelian.
(c) Whenever K is a closed normal subgroup of hhM ii such that K < M , then K
is central in hhM ii.
The layer E(G) of G is the closed subgroup generated by the components of G.
Definition 6.11. A topological group G is of semisimple type if G = E(G). We
say G is of strict semisimple type if in addition Z(G) = {1}.
Each component of a semisimple type group is topologically perfect via [14, Observation 5.3], so groups of semisimple type are topologically perfect. Furthermore,
they are very close to products of topologically simple groups.
Theorem 6.12 (See [14, Theorem 1.11]). Let G be a non-trivial topological group
and let M be the collection of components of G.
(1) The group G is of semisimple type if and only if (G, M) is a generalized central
product and each M ∈ M is non-abelian and central-by-topologically simple.
(2) The group G is of strict semisimple type if and only if (G, M) is a quasi-product
and each M ∈ M is non-abelian and topologically simple.
6.2. Local direct models.
Definition 6.13. Let G be a locally compact group with S a generalized central
factorization of G. AL
local direct
L model for (G, S) is a locally compact local direct
product of the form S (G) := N ∈S (N, N ∩O)Lwhere O is an open subgroup of G
together with a continuous homomorphism φ : S (G) → G such that φ N = idN
for all L
N ∈ S. When there is no ambiguity in doing so, we will refer to either the
group S (G) or the homomorphism φ as the local direct model.
Let us make a few basic observations about local direct models.
Lemma 6.14. Let G L
be a locally compact group with S a generalized central factorization of G. If φ : S (G) → G is a local direct model for G, then img(φ) is a
dense normal subgroup of G.
30
COLIN D. REID AND PHILLIP R. WESOLEK
L
Proof. Since S (G) is locally compact, there is a cofinite S 0 ⊆ S such that N ∩ O
is compact for all N ∈ S 0 . It now follows that
D
E
img(ψ) = S ∪ hN ∩ O | N ∈ S 0 i .
Therefore, img(ψ) is dense in G. Furthermore, since O normalizes N ∩ O for all
N ∈ S, we see that O ≤ NG (img(ψ)). The normalizer NG (img(ψ)) is then a dense
open subgroup and thus equals G.
Proposition 6.15. L
Let G be a locally compact group with S a quasi-direct factorization of G. If φ : S (G) → G is a L
local direct model for G, then φ is injective.
Hence, G is a normal compression of S (G).
Proof. Suppose for contradiction that f ∈ ker(φ) is non-trivial. Since φ S is
injective for all S ∈ S, there isLsome N ∈ S such that f (N ) 6= 1 and f S\{N } 6= 1.
Consider the functions g, h ∈ S∈S (N, N ∩ O) defined as follows:
(
(
f (M ) if M 6= N
f (M ) if M = N
; and h(M ) :=
g(M ) :=
1
else.
1
else
By construction, f = gh, and since 1 = φ(f ) = φ(g)φ(h), we deduce that
φ(h) ∈ N ∩ GS\{N } . This contradicts the independence property enjoyed by S.
We now establish the universal property for the local direct model.
Theorem 6.16. Let G be a locally compact group with S a compactly convergent
generalized
L central factorization and suppose that (G, S) admits a local direct model.
If H =
ψ :H →G
N ∈S (N, WN ) is a locally compact local direct product withL
a continuous map suchLthat ψ N = idN for all N ∈ S, then H ≤
S (G), and
the inclusion ι : H → S (G) is a continuous map making the following diagram
commute:
L
φ
9/ G
S O (G)
ι
ψ
H
Proof. If S is finite, the universal property L
is immediate. Let us L
suppose that S is
infinite and let O ≤ G be open such that N ∈S (N, NQ∩ O) = S (G). Since H
is locally compact,
Q there is a cofinite T ⊆ S such that N ∈T WN is compact, and
hence WH := ψ( N ∈T WN ) is compact. That S is compactly convergent ensures
there is a cofinite S 0 ⊆ T such that GS 0 ∩ WH ≤ O. For f ∈ H, itLnow follows that
f (N ) ∈ GS 0 ∩ WHL⊆ O for cofinitely many N ∈ S, hence fL∈ S (G). We thus
deduce that H ≤ S (G). One easily checks that ι : H → S (G) is continuous,
verifying the theorem.
It follows local direct models are unique, provided they exist.
Corollary 6.17. Let G be a locally compact group with S a compactly convergent
generalized central factorization. If there is a local direct model for (G, S), then it
is unique up to isomorphism.
It remains to identify which factorizations of locally compact groups admit a
local direct model.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
31
Lemma 6.18. Let G be a compact group with S a generalized central factorization
of G and let I be the directed set Q
of finite subsets of S. For a sequence (hN )N ∈S ∈
Q
N
and
J
∈
I,
set
h
:=
J
N ∈S
N ∈J hN where the product
Q is taken in G. If S
is compactly convergent, then for any sequence (hN )N ∈S ∈ N ∈S N the associated
net ĥ := (hJ )J∈I in G converges.
Proof. As G is compact, there exists a limit point k ∈ G of a subnet of ĥ. Given
an identity neighborhood U in G, there exists J ∈ I such that GS\J ⊆ U , since S
is compactly convergent and G is compact. It is then the case that hJ U = hJ 0 U
for all J 0 ∈ I such that J ⊆ J 0 , and considering J 0 from our convergent subnet, we
deduce that hJ U = kU . Hence, ĥ converges.
Say that a locally compact group is almost t.d.l.c. if the connected component
is compact.
Theorem 6.19. Let G be an almost t.d.l.c. group. If S is a compactly convergent
generalized central factorization of G, then (G, S) admits a local direct model.
Proof. Fix U ∈ U(G) and consider K := hN ∩ U | N ∈ Si. The subgroup K is
compact and admits a generalized central factorization {N ∩ U | N ∈ S}. This
factorization is also compactly convergent
since S is compactly convergent.
L
Form the local direct product
(N,
N ∩ U ) and let I be the directed set
N ∈S
consisting of the finite subsets of S ordered by inclusion. An element h := (hN )N ∈S
L
Q
of N ∈S (N, N ∩ U ) gives a net ĥ := (hJ )J∈I in G where hJ := N ∈J hN . Since
h has only finitely many coordinates that do not lie in U , there is a finite J ⊆ S
0
−1
such that h−1
hJ 0 for J 0 extending J
J hJ 0 ∈ K ≤ U for all J ⊇ J. The net (hJ )
converges by Lemma 6.18, hence the net ĥ converges.
L
We may now define ψ : N ∈S (N, N ∩ U ) → G via ψ(h) := lim ĥ. It is easily
verified that ψ is a continuous homomorphism, that ψ restricts to the identity map
on each factor N , and that ψ has a dense image in G.
In view of Lemma 6.7, we deduce the following corollary:
Corollary 6.20. Let G be an almost t.d.l.c. group. If S is a quasi-direct factorization of G, then (G, S) admits a local direct model.
There are two classes of groups for which any quasi-product is actually a direct
product, so in particular, the quasi-product is its own local direct model. The first
follows immediately from Theorem 6.19.
Corollary
6.21. If G is a compact group with S a quasi-direct factorization, then
Q
G ' N ∈S N .
Proposition 6.22. Let G be a connected l.c.s.c. group that is topologically
perfect
Q
and centerless. If S is a quasi-direct factorization of G, then G ' S∈S S, and all
but finitely many members of S are compact.
Proof. Fix N ∈ S and set [G, N ] = [N, N ] =: K. The usual projection π : N →
G/GS\{N } =: H is a normal compression map. Since H is topologically perfect,
π(K) must be dense in H, and thus, π([K, K]) is also dense in H. In view of
Proposition 2.8, we deduce that K ≤ [K, K] and therefore that K is topologically
perfect.
32
COLIN D. REID AND PHILLIP R. WESOLEK
We now have a normal compression map π : K → H with K and H topologically
perfect and H connected. Theorem 2.10 ensures the topologically perfect group K
is connected, and applying Corollary 4.8, we infer that π(K)Z(H) = H. On the
other hand, π(N ) ≤ H = π(K)Z(H), so π(N ) = π(K)(Z(H) ∩ π(N )). It is clear
from the definition of a quasi-product that Z(N ) ≤ Z(G), so N is centerless. The
group π(N ) is then also centerless, and thus, Z(H) ∩ π(N ) = {1}. We deduce
that π(N ) = π(K) and therefore that N is topologically perfect. Applying again
Theorem 2.10, N is additionally connected. Every N ∈ S is therefore topologically
perfect, connected, and centerless.
In view of Theorem 2.4, there is a compact normal subgroup R of G such that
G/R is a Lie group. The quotient map π : G → G/R is a closed map, so each
member of {N R/R | N ∈ S} is closed and normal in G/R. By Lemma 2.3, there is
a finite subset J of S such that N GJ R/GJ R is discrete for all N ∈ S. The group
G/GJ R is connected, so each N GJ R/GJ R is indeed abelian. The group G/GJ R
is thus abelian, and since G is topologically perfect, we deduce that GJ R = G.
Each M ∈ S \ J is topologically perfect, so L := GS\J is topologically perfect.
We now see that
L = [L, L] = [G, L] = [GJ R, L] = [R, L] ≤ R.
The group L is then compact. In particular, every
Q element of S \ J is compact.
Applying Corollary 6.21, we conclude that GS\J ' S∈S\J S, and it follows further
Q
that G ' GJ × S∈S\J S.
We now turn our attention to GJ . Fix N ∈ J and form GJ /CGJ (N ). This
quotient is centerless by Lemma 6.8. The restriction of the natural projection
π : N → GJ /CGJ (N ) is a normal compression map. Since N and GJ are connected,
topologically perfect groups, Corollary 4.9 ensures π is indeed
an isomorphism.
Q
We conclude GJ = N CGJ (N ), and it follows that GJ ' N ∈J N , finishing the
proof.
Since we are primarily interested in non-abelian chief factors, which are topologically characteristically simple, the above results will suffice for our purposes.
However, one naturally wonders if a local direct model exists for any locally compact quasi-product. The following example shows that not all locally compact
quasi-products admit a local direct model, and it also illustrates the need to exclude abelian quasi-products from Proposition 6.22.
Example 6.23. Set G = R×TN and let us consider G as the collection of functions
f : {−1} ∪ N → R t T such that f (−1) ∈ R and f (i) ∈ T for i ∈ N. The group G
is a connected locally compact group under the product topology. For each i ∈ N
let fi ∈ G be defined as follows:
n = −1
1√
fi (n) :=
2 n=i
0
else.
We now define subgroups of G:
(
Gi :=
R
i = −1
hfi i else.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
33
It is easy to verify that each Gi is closed and the collection {Gi }i∈{−1}∪N is a
quasi-direct factorization of G.
Since G is connected, it Q
has no proper open subgroups, so the only possible local
direct model for (G, S) is i∈N Gi . However, the direct product fails to be a local
direct model since it is not locally compact.
6.3. Locally compact groups of semisimple type. We here derive a structure
result for locally compact groups of strict semisimple type.
Theorem 6.24. For G an l.c.s.c. group of strict semisimple type,
G'D×
Y
i∈I
Ci ×
n
Y
Sj
j=0
with D a t.d.l.c.s.c. group of strict semisimple type, each Ci a topologically simple compact connected Lie group, and each Sj a non-compact topologically simple
connected Lie group.
Proof. Let M list the components of G. Appealing to Theorem 6.12, each N ∈ M
is topologically simple, and M is a quasi-direct factorization of G. Let C be the
connected elements of M.
The group C := hCi is a connected, topologically perfect locally compact group
with quasi-factorization Q
C, and it is centerless, since G is centerless. Applying
Proposition 6.22, C '
many N ∈ C non-compact.
N ∈C N with only finitely
Q
Qn
It now follows from Theorem 2.4 that C ' i∈I Ci × j=0 Sj with each Sj a
non-compact topologically simple connected Lie group and each Ci a topologically
simple compact connected Lie group.
Setting D := CG (C), the group G/D is centerless by Lemma 6.8, and G/D is
a normal compression of C. Appealing to Corollary 4.9, the natural projection
π : G → G/D restricts to an isomorphism from C to G/D, and thus, G ' C × D.
Finally, since C and D are closed direct factors of G with D ≥ GM\C , the
product CGM\C is closed in G. This product is also dense, so it is indeed the
case that G = CGM\C . We thus deduce that D = GM\C . The group D is then a
quasi-product of the set M\C of non-abelian topologically simple t.d.l.c.s.c. groups.
Theorem 6.12 implies D is of strict semisimple type, and D is totally disconnected
by Lemma 6.9.
Corollary 6.25. Let G be an l.c.s.c. group of strict semisimple type with S the
set of non-trivial topologically simple closed normal subgroups of G. Then (G, S)
is a quasi-product that admits a local direct model, and the local direct model is an
l.c.s.c. group of strict semisimple type.
Proof. Theorem 6.12 ensures that (G, S) is a quasi-product. Let C be the collection
of connected non-compact elements of S and define D := hS \ Ci and L := hCi. By
Theorem 6.24, we can write G = D × L. Additionally, D is almost t.d.l.c., and L
is a finite direct product of topologically simple connected Lie groups. It is now
enough to show that D and L both admit local direct models. Since L is a finite
product, it is its own local direct model. On the other hand, D is a quasi-product
of S \ C, so Corollary 6.20 ensures that D also admits a local direct model.
For the second claim, the local direct model is itself a quasi-product of nonabelian topologically simple groups, so it is of strict semisimple type.
34
COLIN D. REID AND PHILLIP R. WESOLEK
6.4. Products of elementary groups. We conclude this section with a generalization of Lemma 3.7 to generalized central products.
Theorem 6.26. Let G be a t.d.l.c.s.c. group with S a countable generalized central
factorization of G. If each N ∈ S is elementary, then G is elementary with
ξ(G) = sup + ξ(N ).
N ∈S
Proof. Let {Ni }i∈N enumerate S and put Gj := hNi | i ≤ ji. Each Gj is a generalized central product of finitely many elementary
Q groups. For fixed j, there is
a local direct model φ : P → Gj where P = i≤j Ni . In view of Lemma 3.7,
ξ(P ) = maxi≤j ξ(Ni ), and applying Proposition 5.4 to the normal compression
P/ ker(φ) → Gj , we deduce that
ξ(Gj ) ≤ ξ(P/ ker(φ)) ≤ ξ(P ) = max ξ(Ni ) ≤ ξ(Gj ),
i≤j
where the middle inequality follows from Proposition 3.5. Equality thus holds
throughout; in particular ξ(Gj ) = maxi≤j ξ(Ni ). Proposition 3.10 now completes
the proof, showing that G is elementary with
ξ(G) = sup+ ξ(Gj ) = sup+ ξ(Ni ).
j∈N
i∈N
7. Chief factors and blocks
Any Polish group G comes with a collection of chief blocks BG , as defined in [14].
Each a ∈ BG is an equivalence class of non-abelian chief factors under the relation
of association. The chief blocks appear to be basic building blocks of compactly
generated l.c.s.c. groups, as discovered in [15]. It is, however, not straightforward
to interpret the group-theoretic significance of chief blocks. Associated chief factors
are not necessarily isomorphic as abstract groups, and isomorphic chief factors are
not necessarily associated. In this section, we show that the association relation
does in fact preserve many group-theoretic properties, so group-theoretic data can
be recovered from the chief blocks.
7.1. Preliminaries. Given a topological group G, we say the closed normal factors
K1 /L1 and K2 /L2 are associated to one another if the following equations hold:
K1 L2 = K2 L1 ; K1 ∩ L1 L2 = L1 ; K2 ∩ L1 L2 = L2 .
A chief factor of a topological group G is a normal factor K/L with L < K closed
such that there is no closed M E G with L < M < K. By [14, Corollary 6.9],
restricting the association relation to non-abelian chief factors of a Polish group G
produces an equivalence relation.
Definition 7.1. A chief block of a Polish group G is an association class of
non-abelian chief factors. The set of chief blocks of G is denoted by BG .
A chief factor is called negligible if it is associated to a compact or discrete factor.
A chief block is negligible if it contains a negligible chief factor; the collection of
non-negligible chief blocks is denoted by B∗G .
An important normal subgroup associated to a chief block is the centralizer: For
a chief block a ∈ BG in a topological group G, the centralizer of a is
CG (a) := {g ∈ G | [g, L] ⊆ K}
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
35
for some (equivalently, any) representative K/L of a; see [14, Proposition 6.8].
Inclusion of centralizers induces a partial order on BG denoted by ≤. Centralizers
also give rise to canonical representatives.
Proposition 7.2 ([14, Proposition 7.4]). Let G be a Polish group and a ∈ BG .
Then,
(1) G/CG (a) is monolithic, and the socle M/CG (a) of G/CG (a) is a representative
of a; and
(2) If R/S ∈ a, then M/CG (a) is a G-equivariant normal compression of R/S.
The representative isolated in Proposition 7.2 is called the uppermost representative.
For Polish groups, a normal series may always be refined to include a representative for a given chief block.
Theorem 7.3 ([14, Theorem 1.14]). Let G be a Polish group, a ∈ BG , and
{1} = G0 ≤ G1 ≤ · · · ≤ Gn = G
be a series of closed normal subgroups in G. Then there is exactly one i ∈ {0, . . . , n−
1} such that there exist closed normal subgroups Gi ≤ B < A ≤ Gi+1 of G for which
A/B ∈ a.
In the setting of locally compact groups, we have available the essentially chief
series, which restricts the number of chief blocks.
Definition 7.4. An essentially chief series for a locally compact group G is a
finite series
{1} = G0 ≤ G1 ≤ · · · ≤ Gn = G
of closed normal subgroups such that each normal factor Gi+1 /Gi is either compact,
discrete, or a chief factor of G.
Theorem 7.5 ([15, Theorem 4.4]). Let G be a compactly generated locally compact
group and (Gi )m
i=0 be a finite ascending sequence of closed normal subgroups of G.
Then there exists an essentially chief series
{1} = K0 ≤ K1 ≤ · · · ≤ Kl = G
such that {G0 , . . . , Gm } is a subset of {K0 , . . . , Kl }.
The previous two theorems together imply that the set of non-negligible blocks
of a compactly generated l.c.s.c. group is finite.
Corollary 7.6. Let G be an l.c.s.c. group and (Ai )m
i=0 be an essentially chief series
for G. Setting
N := {Ai /Ai−1 | Ai /Ai−1 is a non-negligible chief factor of G, and 1 ≤ i ≤ m},
there is a bijection θ : B∗G → N such that θ(a) is the unique representative of a
∗
occurring as a factor of (Ai )m
i=0 . In particular, BG is finite.
36
COLIN D. REID AND PHILLIP R. WESOLEK
7.2. Chief block properties.
Definition 7.7. A property P of groups is a (non-negligible) block property
if for every l.c.s.c. group G and (a ∈ B∗G ) a ∈ BG there exists K/L ∈ a with P if
and only if every K/L ∈ a has P . For a (non-negligible) block property P , we say
a block (a ∈ B∗G ) a ∈ BG has property P if some, equivalently all, K/L ∈ a has
property P .
Proposition 7.8. Compact generation is a non-negligible block property.
Proof. Suppose that G is an l.c.s.c. group and a ∈ B∗G has a compactly generated
representative K/L. Take A/B another representative and form the uppermost
representative M/C ∈ a as given by Proposition 7.2. The factor M/C is a normal
compression of K/L, so it is compactly generated. Furthermore, M/C is a normal
compression of A/B and is non-compact since a is non-negligible. As both M/C
and A/B are also non-abelian topologically characteristically simple groups, A/B is
compactly generated by Corollary 5.3. All representatives of a are thus compactly
generated.
Proposition 7.9. The following are block properties in l.c.s.c. groups: Being isomorphic to a given connected group, being elementary with rank α, being amenable,
and being quasi-discrete.
Proof. Let G be an l.c.s.c. group, a ∈ BG , and M/C be the uppermost representative of a, as given by Proposition 7.2.
For the first property, suppose that K/L ∈ a is isomorphic to a connected group
H. The group M/C is connected via Corollary 2.11, so Corollary 4.9 ensures that
K/L is isomorphic to M/C. Consider another representative A/B of a. There is a
normal compression map A/B → M/C, and again A/B ' M/C by Corollaries 2.11
and 4.9. All representatives of a are therefore isomorphic to the connected group
H.
Let us now suppose that a ∈ BG has a quasi-discrete representative K/L and
take A/B ∈ a. The factor M/C is a normal compression of K/L, and these chief
factors are topologically perfect. Corollary 5.10 thus ensures that M/C is quasidiscrete. On the other hand, M/C is also a normal compression of A/B. Applying
again Corollary 5.10, we deduce that A/B is quasi-discrete, so all representatives
are quasi-discrete.
The proof that being elementary with rank α is a block property is similar, using
Proposition 5.4 instead of Corollary 5.10. Likewise, amenability is a block property
using Proposition 5.6.
Proposition 7.9 allows us to define the decomposition rank of a chief block.
Definition 7.10. Let G be an l.c.s.c. group and a ∈ BG . For an elementary block
a ∈ BG , we define the decomposition rank of a as ξ(a) := ξ(K/L) for some
(equivalently, any) K/L ∈ a.
If a is negligible and totally disconnected, then ξ(a) = 2, since both profinite
and discrete groups have decomposition rank 2. The converse is false in general;
as shown in Example 9.4, there exist decomposition rank two blocks which are
non-negligible. Elementary blocks with transfinite rank also exist. For instance,
Example 9.3 admits a block of rank ω + 1; in later work, we use the techniques outlined in Example 9.2 to give many more examples. We recall that by Corollary 3.21,
only certain ranks are possible:
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
37
Corollary 7.11. Let G be an l.c.s.c. group and let a be a totally disconnected chief
block of G. Then exactly one of the following holds:
(1) ξ(a) = 2;
(2) ξ(a) = α + 1 for α some countable transfinite limit ordinal;
(3) a is non-elementary.
Let us conclude by giving a method to compute the rank of elementary chief
factors of semisimple type.
Proposition 7.12. Let G be a topologically characteristically simple t.d.l.c.s.c.
group of semisimple type and let L ≤ K be closed normal subgroups of G such that
K/L is non-abelian. Then G is elementary if and only if K/L is elementary. If G
is elementary, then ξ(G) = ξ(K/L).
Proof. The forward implication of the first claim is immediate from Proposition 3.5.
For the converse implication, consider the quotient G/L and let M list the components of G. The group G/L contains a dense subgroup generated by the images of
components. Since K/L is non-abelian, there is some C ∈ M that intersects K/L
non-trivially.
Let π : G → G/L be the usual projection and find C ∈ M such that π(C) ∩
K/L is non-trivial. Since G is non-abelian, it has trivial center and therefore is of
strict semisimple type. Every component N ∈ M is thus topologically simple via
Theorem 6.12. The restriction π : C → G/L is injective, and π(C) is contained
in K/L. The group CL/L ≤ K/L is then a normal compression of C. As K/L is
elementary, C is elementary with ξ(C) = ξ(K/L) via Proposition 5.4.
All components of G are isomorphic to C, since G is topologically characteristically simple. In view of Theorem 6.26, we now deduce that G is elementary
with
ξ(G) = sup + ξ(N ) = ξ(C) = ξ(K/L).
N ∈M
7.3. Negligible chief blocks. By Corollary 7.6, an essentially chief series for
a compactly generated l.c.s.c. group G includes representatives for every nonnegligible chief block of G. Of course, an essentially chief series can also include
negligible chief factors. The utility of essentially chief series as a decomposition into
well-behaved “basic” groups, compact groups, and discrete groups thus depends on
how much structure beyond the compact or discrete cases can be hidden in negligible chief factors. We here analyze negligible chief factors and blocks. This analysis
will demonstrate that negligible chief factors are either compact or very close to
discrete.
Let us first identify three types of chief blocks of an l.c.s.c. group G. Proposition 7.9 ensures the following definition is sensible:
Definition 7.13. Let G be an l.c.s.c. group and a ∈ BG . We say that the block a
is connected compact if all representatives are isomorphic to a single connected
compact group. We say that a is quasi-discrete if all representatives are quasidiscrete. The block a is robust if none of the representatives of a are compact or
quasi-discrete. Denote by BrG the collection of robust blocks.
We now argue that these three types partition the collection of chief blocks.
To this end, recall the special structure of topologically characteristically simple
profinite groups.
38
COLIN D. REID AND PHILLIP R. WESOLEK
Lemma 7.14 (See for instance [17, Lemma 8.2.3]). A profinite group P is topologically characteristically simple if and only if
P ' FI
for some finite simple group F and set I. In particular, every topologically characteristically simple profinite group is quasi-discrete, since it contains the dense
quasi-central subgroup F <I .
Theorem 7.15. Suppose that G is an l.c.s.c. group. If a is a chief block of G, then
a is exactly one of the following: connected compact, quasi-discrete, or robust.
Proof. It is obvious the three cases are exclusive, so we need only to show a is
one of the three types. By Proposition 7.9, if a is not quasi-discrete, then none of
its representatives are quasi-discrete. Suppose that a is neither robust nor quasidiscrete; in other words, there is a representative A/B ∈ a that is compact, but not
quasi-discrete. The factor A/B is either connected or totally disconnected. In the
latter case, A/B is profinite, so it is quasi-discrete by Lemma 7.14. We deduce that
A/B is connected and compact, and hence by Proposition 7.9, every representative
of a is isomorphic to A/B as a topological group. Thus, a is connected compact.
A structure theorem for compactly generated groups without robust blocks is
now in hand.
Corollary 7.16. Suppose that G is a compactly generated l.c.s.c. group. If BrG = ∅,
then every essentially chief series (Gi )ni=0 for G is such that each factor Gi /Gi−1
is either compact or quasi-discrete.
Proof. Let {1} = G0 < · · · < Gn = G be an essentially chief series for G, so each
factor Gi /Gi−1 is either compact, discrete, or chief. Discrete factors and abelian
chief factors are obviously quasi-discrete. Theorem 7.15 implies that every nonabelian chief factor in the series is either compact-connected or quasi-discrete.
7.4. Application: Hopfian property. A group is called Hopfian if every continuous self-surjection is also an injection. Using the machinery of chief blocks, we
show many compactly generated l.c.s.c. groups are Hopfian. Our analysis of the
Hopfian property is via the notion of block morphisms developed in [14, Section 8];
we direct the reader to loc. cit. for further details on this construction.
For G a Polish group, let XG denote the collection of non-abelian chief factors.
The partial order on centralizers induces a preorder on XG , which we denote by
(XG , ≤). We remark that the natural quotient of (XG , ≤) giving a partial ordering
recovers the partially ordered set of chief blocks BG .
Definition 7.17. For G and H Polish groups, a strong block morphism ψ :
(XH , ≤) → (XG , ≤) is an homomorphism of preorders such that ψ(A/B) ' A/B
as topological groups for each A/B ∈ XH . If the induced map ψe : BH → BG is
injective, we say that ψ is a strong block monomorphism.
By [14, Proposition 8.4], a continuous surjective homomorphism φ : G → H
of Polish groups induces a strong block monomorphism ψ : (XH , ≤) → (XG , ≤)
defined by A/B 7→ φ−1 (A)/φ−1 (B). In the setting of locally compact groups, this
morphism also respects robust blocks.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
39
Lemma 7.18. Suppose that G and H are l.c.s.c. groups and φ : G → H is a
continuous surjective homomorphism. Then the strong block monomorphism
ψ : (XH , ≤) → (XG , ≤)
induced by φ is such that ψ̃(BrH ) = BrG ∩ ψ̃(BH ), where ψ̃ is the map induced from
ψ.
Proof. Take a ∈ BH and suppose that ψ̃(a) ∈
/ BrG . Taking A/B ∈ a, that ψ is a
strong block monomorphism implies that A/B ' ψ(A/B) and that ψ(A/B) ∈ ψ̃(a).
The block ψ̃(a) is not robust, so Theorem 7.15 implies that ψ̃(a) is either connectedcompact or quasi-discrete. The factor ψ(A/B) is thus either a connected compact
group or a quasi-discrete group, hence A/B is either a connected compact group or
a quasi-discrete group. We conclude that a is either connected-compact or quasidiscrete. That is to say, a is not robust. This demonstrates that ψ̃(BrH ) ⊆ BrG ∩
ψ̃(BH ).
Conversely, suppose that a 6∈ BrH . We may find A/B ∈ a which is either connected and compact or quasi-discrete. The image ψ(A/B) is then either connected
and compact or quasi-discrete, since ψ is a strong block monomorphism. The block
ψ̃(a) is thus not robust. This demonstrates that ψ̃(BH \ BrH ) ⊆ BG \ BrG , and
hence ψ̃(BrH ) = BrG ∩ ψ̃(BH ).
We stress that it is important to consider robust blocks for the previous lemma.
Negligible blocks can become non-negligible in a quotient, so in general we do not
have ψ̃(B∗H ) ⊆ B∗G , where B∗G is the set of non-negligible blocks.
Theorem 7.19. Suppose that G is a compactly generated l.c.s.c. group. If G has no
non-trivial compact or quasi-discrete closed normal subgroups, then G is Hopfian.
Proof. We assume that G is non-trivial. Suppose for contradiction that φ : G →
G is a continuous surjective homomorphism with a non-trivial kernel. We apply
Theorem 7.5 to find a chief series (Gi )ni=0 which refines the normal series {1} <
ker(φ) < G. The first factor G1 /G0 = G1 must be neither compact nor quasidiscrete since G has no such non-trivial closed normal subgroups. Hence, G1 /G0 is
a non-abelian chief factor, and the chief block [G1 /G0 ] is robust via Theorem 7.15.
On the other hand, Lemma 7.18 supplies a canonical injective map ψ̃ : BrG → BrG
arising from the surjection φ : G → G. The set BrG is finite by Corollary 7.6, so ψ̃
is indeed a bijection. We may thus find a chief factor A/B of G such that
K/L = ψ(A/B) = φ−1 (A)/φ−1 (B)
is an element of [G1 /G0 ]. In the series G0 < G1 < ker(φ) ≤ L < K ≤ G, the
non-abelian chief factor G1 /G0 is associated to the factor K/L. This is absurd
since associated chief factors have the same centralizer and L ≤ CG (K/L).
It is easy to build non-discrete compactly generated locally compact groups which
are non-Hopfian simply by taking the direct product of an infinite profinite group
with a finitely generated non-Hopfian group.
8. Extensions of chief blocks
Corollary 7.16 shows a compactly generated l.c.s.c. group G without robust chief
factors has limited complexity as a topological group. In this section, we show that
except for groups with limited topological structure, every l.c.s.c. group admits
40
COLIN D. REID AND PHILLIP R. WESOLEK
robust chief blocks that are minimally covered. These results are then applied to
investigate the structure of chief factors of compactly generated l.c.s.c. groups.
8.1. Preliminaries.
Definition 8.1. Let a be a chief block and let N/M be a closed normal factor of
a topological group G. We say N/M covers a if CG (a) contains M but not N . If
N/M does not cover a, we say that N/M avoids a. A block can be avoided in one
of two ways: The factor N/M is below a if N ≤ CG (a); equivalently, G/N covers
a. The factor N/M is above a if M 6≤ CG (a); equivalently, M/{1} covers a. For a
closed normal subgroup N E G, we will say that N covers or avoids a chief block
by seeing N as the closed normal factor N/{1}.
It follows easily from Theorem 7.3 that if a closed normal factor N/M of a
Polish group G covers a ∈ BG , then there is a representative K/L ∈ a such that
M ≤ L < K ≤ N.
A block a ∈ BG is minimally covered if there is a least closed normal subgroup
covering the block. This least normal subgroup is denoted by Ga . The lowermost
representative is Ga /CGa (a); this normal factor is a representative of a via [14,
Proposition 7.13]. The collection of minimally covered chief blocks is denoted by
Bmin
G .
Definition 8.2. Let G be a Polish group with H a closed subgroup. Given a ∈ BH
and b ∈ BG , we say that b is an extension of a in G if for every closed normal
subgroup K of G, the group K covers b if and only if K ∩ H covers a.
In general, it is not clear whether extensions exist; however, when an extension
exists, it is unique, [14, Lemma 9.2]. If the extension exists, we write aG for the
extension of a to G. The following observations are made in [14, Section 9.1];
to gain familiarity with the definitions, the reader is encouraged to prove these
observations.
Observation 8.3. Let A ≤ B ≤ G be closed subgroups of a topological group G
and a ∈ BA be extendable to B. Then a is extendable to G if and only if aB is
extendable to G. If a is extendable to G, then aG = (aB )G .
Observation 8.4. Let G be a topological group with H a closed subgroup. If K/L
is a closed normal factor of G such that K ∩ H/L ∩ H covers a ∈ BH , then
CH (K/L) ≤ CH (a).
Observation 8.5. Let H be a closed subgroup of a Polish group G and a ∈ BH be
extendable to G. Then a closed normal factor K/L of G covers aG if and only if
K ∩ H/L ∩ H covers a.
In the case of minimally covered blocks, extensions are easier to identify:
Lemma 8.6 ([14, Lemma 9.7]). Let G be a topological group with H a closed
subgroup of G and a ∈ Bmin
H . Then a is extendable to G if and only if there exists
b ∈ Bmin
such that (Gb ∩ H)/C(Gb ∩H) (b) covers a. If a is extendable to G, then
G
aG = b.
We can say more when the subgroup H is also normal.
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
41
Proposition 8.7 ([14, Proposition 9.8]). Let G be a topological group with H a
closed normal subgroup of G and a ∈ Bmin
H . Then a is extendable to G, and the
extension aG has least representative M/N where
\
M := hhHa iiG and N :=
gCG (a)g −1 ∩ M.
g∈G
The previous results are particularly useful in the setting of compactly generated
groups, where minimally covered blocks are known to exist.
Proposition 8.8 ([15, Proposition 4.10, Corollary 4.9]). If G is a compactly generated l.c.s.c. group, then all non-negligible chief blocks are minimally covered, and
there are only finitely many non-negligible chief blocks.
8.2. Extending robust blocks. We now show that the robust blocks of any compactly generated open subgroup of a l.c.s.c. group G extend to G. Let us first note
that robustness is preserved by extension, provided the extension exists.
Lemma 8.9. Suppose that G is an l.c.s.c. group with K/L a quasi-discrete closed
normal factor of G. If L ≤ B < A ≤ K are such that A/B is a chief factor of G,
then A/B is quasi-discrete.
Proof. The factor K/B is quasi-discrete, and A/B is a normal subgroup of K/B.
If A/B intersects QZ(K/B) trivially, then A/B is central in K/B and a fortiori is
quasi-discrete. If the group A/B has non-trivial intersection with QZ(K/B), then
A/B has a non-trivial quasi-center. Since the quasi-center is characteristic and
A/B is a chief factor of G, we conclude that A/B is quasi-discrete.
Proposition 8.10. Suppose that G is an l.c.s.c. group with O an open subgroup
and let a be a robust chief block of O.
(1) If K/L is a closed normal factor of G such that K ∩ O/L ∩ O covers a, then
K/L is neither compact nor quasi-discrete.
(2) If a extends to G, then aG is robust.
Proof. For (1), the subgroup (K ∩ O)L/L is an open subgroup of K/L, and (K ∩
O)L/L ' (K ∩ O)/(L ∩ O). If K/L is compact, then (K ∩ O)/(L ∩ O) is compact,
and since a representative of a occurs as a closed normal factor of (K ∩ O)/(L ∩ O),
it follows that a is negligible, which is absurd. If instead K/L is quasi-discrete,
then (K ∩ O)L/L is also quasi-discrete since it is an open subgroup of K/L, so
(K ∩ O)/(L ∩ O) is quasi-discrete. In this case, Lemma 8.9 ensures that a has a
quasi-discrete representative, which contradicts the hypothesis that a is robust.
Claim (2) now follows immediately from (1) and Observation 8.5.
We now isolate a set of blocks that arise from robust blocks of compactly generated open subgroups.
Definition 8.11. Let G be an l.c.s.c. group and let a ∈ BG . We say that a is a
regionally robust block if there exists a compactly generated open subgroup O
of G and a robust block b of O such that a is the extension of b in G. The collection
of regionally robust blocks is denoted by Brr
G.
Proposition 8.8 and Lemma 8.6 ensure that every regionally robust block is
minimally covered, and Proposition 8.10 ensures that every regionally robust block
r
is robust. If G is compactly generated, we now see that Brr
G ⊆ BG , so indeed
rr
r
BG = BG .
42
COLIN D. REID AND PHILLIP R. WESOLEK
Lemma 8.12. An l.c.s.c. group has at most countably many regionally robust
blocks.
Proof. Let G be an l.c.s.c. group. A regionally robust block b of G is determined
by a pair (O, a) where O is a compactly generated open subgroup of G and a ∈ BrO
is such that aG = b. An easy counting argument shows there are only countably
many compactly generated open subgroups, and via Proposition 8.8, a compactly
generated open subgroup has finitely many robust blocks. We conclude there are
at most countably many such pairs (O, a), and hence an l.c.s.c. group has at most
countably many regionally robust blocks.
We now set about showing that regionally robust blocks of open subgroups always
extend. We begin by considering compactly generated open subgroups of compactly
generated l.c.s.c. groups; the key here is to exploit the essentially chief series. In the
following proof, it will be important that certain subgroups and homomorphisms
are invariant under some group action. The action is always the obvious action by
conjugation.
Lemma 8.13. Let G be a compactly generated l.c.s.c. group. If O is a compactly
generated open subgroup of G, then all robust blocks of O extend to G.
Proof. Fix a ∈ BrO and let (Hi )ni=0 be an essentially chief series for G. The restriction (Hi ∩ O)ni=0 is a normal series for O. Applying Theorem 7.3, there is some
0 ≤ i < n such that Hi+1 ∩ O/Hi ∩ O covers a. The factor Hi+1 /Hi is neither compact nor quasi-discrete by Proposition 8.10, and thus, Hi+1 /Hi is a non-abelian
chief factor. Theorem 7.15 implies that b := [Hi+1 /Hi ] is indeed robust. In view
of Proposition 8.8, we deduce that b ∈ Bmin
G .
We argue that b is the extension of a to G. Let us first consider the uppermost
representative N/C of b. Since N ∩ O ≥ Hi+1 ∩ O, the subgroup N ∩ O covers a.
On the other hand, Observation 8.4 implies CO (Hi+1 /Hi ) ≤ CO (a), and the construction of the uppermost representative ensures C = CG (Hi+1 /Hi ). We conclude
that C ∩ O ≤ CO (a) and thus that C ∩ O is below a. Hence, N ∩ O/C ∩ O covers a.
We now consider the lowermost representative I/J of b. Proposition 7.2 gives a
G-equivariant normal compression ψ : I/J → N/C. We also have an O-equivariant
isomorphism
ι : N ∩ O/C ∩ O → (N ∩ O)C/C := W.
Since N ∩ O/C ∩ O covers a, there are A and B closed O-invariant subgroups
of N ∩ O/C ∩ O such that A/B is a representative of a. The subgroup W is
open in N/C, so ψ(I/J) ∩ W is dense and normal in W and hence accounts for
a dense normal subgroup of W/ι(B). Since A/B is topologically perfect, setting
D := (ψ(I/J) ∩ ι(A))ι(B), we have
D ≥ [ψ(I/J) ∩ W, ι(A)]ι(B) ≥ [W, ι(A)]ι(B) = ι(A).
In other words, D is a dense subgroup of ι(A).
The map ι is an O-equivariant isomorphism of topological groups, so ι−1 (D) is
a dense O-invariant subgroup of A. Since A/B is not quasi-discrete, it follows that
|ι−1 (D) : B| is uncountable. On the other hand, the group V := (I ∩ O)J/J is open
in I/J, so it is of countable index. It follows that (ψ(V ) ∩ ι(A))ι(B) is of countable
index in D, and thus, |(ψ(V ) ∩ ι(A))ι(B) : ι(B)| is also uncountable; in particular,
ψ(V ) ∩ ι(A) ι(B). The group I ∩ O therefore cannot centralize a. That is to say,
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
43
I ∩ O covers a. Plainly, J ∩ O is below a, since J ≤ C, so I ∩ O/J ∩ O covers a.
Lemma 8.6 now ensures that b is the extension of a to G.
We now upgrade from compactly generated supergroups to arbitrary l.c.s.c. supergroups. The strategy is to apply Lemma 8.13 to an increasing exhaustion by
compactly generated open subgroups. The existence of the extensions ai in the
following statement is ensured by Lemma 8.13.
Proposition 8.14. Let G be an l.c.s.c. group with (Oi )i∈N an increasing exhaustion
of G by compactly generated open subgroups. Fix a0 ∈ BrO0 and set ai := aOi for
each i ∈ N. Then there is a robust chief block b of G that extends each of the ai to
G. Additionally,
[ \
COn (ai ).
CG (b) =
n≥0 i≥n
Proof. Set
D :=
[ \
COn (ai )
n≥0 i≥n
For fixed n and all i ≥ n, Observation 8.4 ensures that COn (ai+1 ) ≤ COn (ai ), hence
\
D ∩ On =
COn (ai ) ≤ COn (an ).
i≥n
The group D is thus closed and such that D ∩ O0 avoids a0 .
For N ∈ N (G) such that N D, there exists n such that N ∩ On D ∩ On .
Therefore, N ∩ On COn (ai ) for some i ≥ n, so N ∩ Oi covers ai . From the
definition of an extension, we infer that N ∩ O0 covers a0 . The subgroup D is thus
the unique maximal
closed normal subgroup of G such that D ∩ O0 avoids a0 .
T
Set M := {N ∈ N (G) | N ∩ O0 covers a0 , and D ≤ N }. Since the block a0
is minimally covered, it follows that M ∩ O0 covers a0 . On the other hand, D is
maximal among closed normal subgroups of G that avoid a0 , hence M/D is a chief
factor of G. Set b := [M/D] and let C = CG (b). Since M ∩ O0 /D ∩ O0 covers a0 ,
we have C ∩ O0 ≤ CO0 (a0 ) by Observation 8.4. The subgroup C therefore avoids
a0 , so C ≤ D by the previous paragraph. As D clearly centralizes M/D, we deduce
that C = D.
Finally, N ∈ N (G) is such that N ∩ O0 covers a0 if and only if N D. Since
D = CG (b), it is also the case that N ∈ N (G) covers b if and only if N D. The
block b is thus the extension of a0 , and it is robust via Proposition 8.10. That b
extends each ai follows from Observation 8.3.
We now prove our general extension theorem.
Theorem 8.15. Suppose that G is an l.c.s.c. group and H is a closed subgroup of
G. If there is a sequence
H = H0 ≤ H1 ≤ · · · ≤ Hn = G
of closed subgroups of G such that Hi either an open subgroup of Hi+1 or a normal
subgroup of Hi+1 for each 0 ≤ i < n, then every regionally robust block of H extends
to a regionally robust block of G.
Proof. By Observation 8.3 and induction on n, it is enough to consider the cases
where H is either open in G or normal in G.
44
COLIN D. REID AND PHILLIP R. WESOLEK
Suppose first that H is open in G and let a be a regionally robust block of H.
The block a is the extension of some robust chief block b0 of O0 , where O0 is a compactly generated open subgroup of H. Starting with O0 , we can find an increasing
exhaustion (Oi )i∈N of G by compactly generated open subgroups. Proposition 8.14
then provides the extension b of b0 to G, and appealing to Observation 8.3, b is
the extension of a to G. We conclude that a is extendable to G, and the extension
is clearly also a regionally robust block of G.
Suppose that H is normal in G and let a be a regionally robust block of H. Say
that a is the extension of b ∈ BrO , where O is a compactly generated open subgroup
of H. There are only countably many compactly generated open subgroups of H,
so NG (O) is open in G. We can then take a compactly generated open subgroup
W of NG (O) such that O E W . By Propositions 8.7 and 8.10, b extends to a
robust chief block bW of W . The previous case now ensures that bW extends to a
regionally robust block c := (bW )G of G, and in view of Observation 8.3, we deduce
that
c = (bW )G = bG = (bH )G = aG .
The block a is thus extendable to G. Since bW is a robust block of a compactly
generated open subgroup of G, we see that aG is also a regionally robust block.
8.3. Further observations on regionally robust blocks. The persistence of
regionally robust blocks from open subgroups shows that regionally robust blocks
are persistent in a more subtle manner: if a chief factor K/L has a regionally robust
block, then [K/L] is a regionally robust block.
Proposition 8.16. Let G be an l.c.s.c. group with K/L a closed normal factor of
G. If Brr
K/L 6= ∅, then K/L covers a regionally robust block of K ,and hence, it
covers a regionally robust block of G. If in addition K/L is a chief factor of G,
then [K/L] is a regionally robust block of G.
Proof. Fix a regionally robust block a of K/L and say that a is the extension of
a robust block b of O where O is an open compactly generated subgroup of K/L.
We may fix W ≤ K compactly generated and open such that W L/L = O. Given
a continuous surjection of l.c.s.c. groups, there is a canonical way to lift blocks
from the range to the domain; see the discussion in Section 7.4. Lemma 7.18
also ensures that robust blocks lift to robust blocks. Considering the surjection
W → W L/L = O, the lift b̃ of b to W is robust. Theorem 8.15 then implies that
b̃ is extendable to K, and the extension is plainly regionally robust. In view of the
definition of the lift, L ∩ W avoids e
b, so the normal factor K/L covers e
bK =: c. By
Theorem 8.15, c is also extendable to a regionally robust block d of G; since K/L
is a normal factor of G, it also covers d, via Proposition 8.7.
The factor K/L covers the regionally robust block d of G, so we may insert closed
normal subgroups L ≤ B < A ≤ K of G such that A/B ∈ d. If K/L is already a
chief factor, we must have A = K and B = L. We deduce that K/L ∈ d, hence
[K/L] is a regionally robust block of G.
We now prove an alternative for l.c.s.c. groups.
Proposition 8.17. For G an l.c.s.c. group, either
(1) Brr
G 6= ∅, or
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
45
(2) G◦ is compact-by-solvable-by-compact, and G/G◦ is elementary with rank
at most ω + 1. If G is compactly generated, then G/G◦ is elementary with
finite rank.
Proof. Suppose that Brr
G = ∅; we argue that (2) holds.
Let us first suppose that G is compactly generated. Set R := RadLE (G◦ ).
The group R is a compact normal subgroup of G, and G◦ /R is a Lie group by
Theorem 2.4. Letting π : G◦ → G◦ /R be the natural projection, put S :=
π −1 (Sol(G◦ /R)) where Sol(G◦ /R) is the solvable radical and take Z to be such
that Z/S := Z(G◦ /S). This produces a series of closed normal subgroups of G:
{1} ≤ R ≤ S ≤ Z ≤ G◦ ≤ G.
Applying Theorem 7.5, we refine this series into an essentially chief series {1} =
G0 < · · · < Gn = G. Any factor given by terms Gk < Gk+1 with Z ≤ Gk <
Gk+1 ≤ G◦ is a product of non-abelian connected simple groups, since G◦ /Z is a
centerless connected semisimple group. Since BrG = Brr
G = ∅, these simple groups
must be compact. The connected component G◦ is therefore compact-by-solvableby-compact.
For the totally disconnected part, Corollary 7.16 ensures that any chief factors
Gk+1 /Gk with G◦ ≤ Gk are quasi-discrete, hence elementary of rank 2. The compact totally disconnected factors and discrete factors are also elementary of rank 2.
The rank is subadditive by Lemma 3.6, hence
ξ(G/G◦ ) ≤ 2n < ω.
This completes the proof in the compactly generated case.
Suppose now that G is an arbitrary l.c.s.c. group. Taking O ≤ G open and
r
compactly generated, we see that Brr
O = BO = ∅, since otherwise G admits a
regionally robust block via Theorem 8.15. Appealing to the compactly generated
case of the proposition, we deduce that O◦ = G◦ is compact-by-solvable-by-compact
and that O/O◦ is elementary with finite rank. The quotient G/G◦ is thus an
increasing union of finite rank elementary groups, hence ξ(G/G◦ ) ≤ ω + 1, via
Proposition 3.10.
As a consequence, sufficiently large chief factors are necessarily regionally robust.
Corollary 8.18. Let G be an l.c.s.c. group with K/L a chief factor of G. Suppose
that one of the following holds:
(1) K/L is connected, non-abelian and non-compact;
(2) K/L is elementary with ξ(K/L) > ω + 1;
(3) K/L is totally disconnected and non-elementary.
Then [K/L] is a regionally robust block of G, and Brr
K/L 6= ∅.
Proof. By Proposition 8.16, it is sufficient to show that K/L has a regionally robust
block. Such a regionally robust block exists in all cases by Proposition 8.17.
8.4. A-simple groups. For a topological group G and A ≤ Aut(G), we say that
G is A-simple if the only A-invariant closed normal subgroups of G are {1} and
G. A trichotomy exists for A-simple topological groups.
Definition 8.19. Let G be a topologically characteristically simple topological
group.
(1) The group G is of weak type if Bmin
G = ∅.
46
COLIN D. REID AND PHILLIP R. WESOLEK
min
(2) The group G is of stacking type if Bmin
G 6= ∅ and for all a, b ∈ BG , there
exists ψ ∈ Aut(G) such that ψ.a < b.
Theorem 8.20 ([14, Theorem 9.16]). Suppose that G is an A-simple topological
group for some A ≤ Aut(G). Then exactly one of the following holds:
(1) The group G is of weak type.
(2) The group G is of semisimple type and A acts transitively on BG .
(3) The group G is of stacking type and for all a, b ∈ Bmin
G , there exists ψ ∈ A such
that ψ.a < b.
Using our results for regionally robust blocks, we obtain a stronger version of
Theorem 8.20 for locally compact groups, by describing the groups that are neither
semisimple nor stacking type.
Theorem 8.21. Suppose that G is a topologically characteristically simple l.c.s.c.
group. Then G is exactly one of the following:
(1) an abelian group;
(2) a non-abelian quasi-discrete t.d.l.c.s.c. group;
(3) a regionally SIN t.d.l.c.s.c. group with trivial quasi-center;
(4) an elementary t.d.l.c.s.c. group of decomposition rank ω + 1;
(5) a direct product of copies of a simple Lie group;
(6) a quasi-product of copies of a topologically simple t.d.l.c.s.c. group of decomposition rank greater than ω + 1;
(7) a t.d.l.c.s.c. group of stacking type and decomposition rank greater than ω + 1.
Proof. We may assume that G is non-abelian. Since G is topologically characteristically simple, the group G is then topologically perfect and centerless.
Plainly, G is either connected or totally disconnected. Let us suppose first that G
is connected. By [8, Theorem 10.25], G has a unique largest pro-solvable connected
normal subgroup, and since G is topologically perfect, G itself cannot be prosolvable. That G is characteristically simple now ensures that G is semisimple in
the sense of [8, Theorem 10.29]. Applying [8, Theorem 10.29], we deduce that G is
a direct product of copies of a simple Lie group. Therefore, case (5) holds.
We now suppose that G is totally disconnected. We consider first the case that
ξ(G) > ω + 1. From Proposition 8.17, it follows that G has a regionally robust
block, so it is either of semisimple type or of stacking type via Theorem 8.20. If
G is of semisimple type, then it is a quasi-product of copies of some topologically
simple t.d.l.c.s.c. group H by Theorem 6.12. If H is elementary, Theorem 6.26
ensures that ξ(G) = ξ(H), so ξ(H) > ω + 1. If H is non-elementary, then plainly
ξ(H) > ω + 1. We thus deduce that exactly one of (6) and (7) holds.
Let us finally consider the case that ξ(G) ≤ ω + 1. By Corollary 3.21, ξ(G) is
either equal to ω + 1 or 2. If ξ(G) = ω + 1, then (4) holds. If ξ(G) = 2, then
Lemma 3.13 implies G is a regionally SIN group. If G additionally has a trivial
quasi-center, then (3) holds; otherwise, since G is topologically characteristically
simple, the quasi-center is dense and (2) holds.
We thus deduce that for any topologically characteristically simple l.c.s.c. group
one of (1)-(7) holds. That the conditions are exclusive we leave as an exercise.
All of the cases (1)-(7) can occur. In particular, Example 9.4 gives an example of
(3), and Example 9.2 gives an example of (7). The techniques sketched in Section 9
can be used to produce elementary chief factors which satisfy (7); these examples
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
47
will appear in a future work. Examples satisfying (6) are easy to produce, but we
do not know if (6) can occur for elementary groups.
Corollary 8.22. Chief factors of l.c.s.c. groups have one of the forms given in
Theorem 8.21. Moreover, if K1 /L1 and K2 /L2 are associated chief factors of an
l.c.s.c. group G, then K1 /L1 and K2 /L2 have the same form.
Proof. Since G1 := K1 /L1 and G2 := K2 /L2 are chief factors, they are topologically
characteristically simple. They are thus each of one of the forms described in
Theorem 8.21.
Applying [14, Lemma 6.6], there is a third normal factor K/L of G such that
K/L is a normal compression of both G1 and G2 . Hence, G1 is abelian if and only
if G2 is abelian. Having dispensed with case (1), we may assume that both G1
and G2 are non-abelian. Cases (2), (3), (4), and (5) are now easy applications of
Proposition 7.9. For the cases (6) and (7), Proposition 7.9 ensures that ξ(G1 ) =
ξ(G2 ), and the trichotomy of Theorem 8.20 is invariant under association via [14,
Proposition 9.19]. Therefore, G1 satisfies (6) if and only if G2 satisfies (6), and G1
satisfies (7) if and only if G2 satisfies (7).
The division into cases of Theorem 8.21 is not exactly a refinement of Theorem 8.20, since elementary t.d.l.c.s.c. groups of rank ≤ ω + 1 can be of semisimple
or stacking type. Here is an alternative version that emphasizes the role of weak
type; after dividing into the three cases given by Theorem 8.20, the proof is essentially the same as that of Theorem 8.21.
Theorem 8.23. Suppose that G is an A-simple l.c.s.c. group for some A ≤ Aut(G).
Then exactly one of the following holds:
(1) The group G is abelian.
(2) The group G is of weak type and is a non-abelian t.d.l.c.s.c. group that is quasidiscrete.
(3) The group G is of weak type and is a regionally SIN t.d.l.c.s.c. group with trivial
quasi-center.
(4) The group G is of weak type and is an elementary t.d.l.c.s.c. group of decomposition rank ω + 1.
(5) The group G is of semisimple type and A acts transitively on BG .
(6) The group G is a t.d.l.c.s.c. group of stacking type and for all a, b ∈ Bmin
G ,
there exists ψ ∈ A such that ψ.a < b.
There are weak type chief factors of rank exactly ω + 1; see Example 9.3. Case
(4) therefore cannot be omitted.
Remark 8.24. Let us pause to consider the seven cases of Theorem 8.21 with
respect to the goal of decomposing chief factors into compact, discrete, or topologically simple groups. This goal is clearly achieved in cases (5) and (6), and it is
straightforward to show in case (1) that G is discrete-by-compact in the connected
case and compact-by-discrete in the totally disconnected case. In cases (2) and (3),
G is a directed union of SIN t.d.l.c.s.c. groups, so in particular a directed union
of compact-by-discrete groups. In case (4), G is a directed union of compactly
generated open subgroups, and each such subgroup admits a canonical finite series
in which all factors are regionally SIN, via Theorem 3.23. Hence, G is constructed
from compact and discrete groups.
48
COLIN D. REID AND PHILLIP R. WESOLEK
This leaves case (7). Stacking type ensures a rich subnormal subgroup structure,
but it is unclear exactly how rich this structure can be. The techniques of Section 9
show that there is a great deal of freedom in a stacking type group. The most
natural decomposition property to hope for in general thus seems to be the following
well-foundedness condition:
Question 8.25. Suppose that G =: G0 is a topologically characteristically simple
l.c.s.c. group. If G0 is abelian, elementary with rank at most ω + 1, or of semisimple
type, we stop. Else, we find G1 := K/L a non-abelian chief factor of G0 which is
robust. Continuing in this fashion produces G0 , G1 , . . . a sequence of l.c.s.c. groups.
Is it the case that any such sequence halts in finitely many steps? What about in
the case that the group G is also elementary?
9. Examples
9.1. Preliminaries. We here describe a technique which produces compactly generated t.d.l.c.s.c. groups with interesting chief factors. We only sketch the construction and recall the relevant facts; the full details will appear in a later article. The
construction is inspired by ideas from [3, 10, 19], and the reader familiar with [10]
can likely prove all claimed facts below.
Let T be the countably regular tree and fix δ an end of T . We orient the edges
of T such that all edges point toward the end δ. The resulting directed graph is
denoted T~ ; we call δ the distinguished end of T~ . Given a countable set X, a
coloring of T~ is a function c : E T~ → X such that for each v ∈ V T~ ,
cv := c inn(v) : inn(v) → X
is a bijection. The set inn(v) is the collection of directed edges with terminal vertex
v. The coloring allows us to define the local action of g ∈ Aut(T~ ) at v ∈ V T~ :
σ(g, v) := cg(v) ◦ g ◦ c−1
v ∈ Sym(X).
We call the coloring ended if there is a monochromatic directed ray which is a
representative of the distinguished end δ; we shall always assume our colorings are
ended.
The tree T~ is also equipped with horoballs and horospheres centered at δ, defined
as follows. There is a function η from V T~ to Z such that, for each of the oriented
edges of e of T~ , we have η(t(e)) = η(o(e)) + 1. One can see that the function η
is unique up to an additive constant: it is determined by choosing a base point
v0 such that η(v0 ) = 0. The set {η −1 (i) | i ∈ Z} of fibers is therefore uniquely
determined. The fibers are the horospheres centered at δ, and the sets Xi :=
{v ∈ V T~ | η(v) ≥ i} for i ∈ Z are the horoballs centered at δ; see Figure 1.
X−1
v−1
X0
v0
X1
v1
Figure 1. The horoballs Xi
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
49
To define the groups, it will be convenient to make a definition: A t.d.l.c.s.c.
permutation group is a pair (G, X) where G is a t.d.l.c.s.c. group and X is a
countably infinite set on which G acts faithfully with compact open point stabilizers.
Definition 9.1. Suppose that (G, X) is a t.d.l.c.s.c. permutation group with U ∈
U(G) and color the tree T~ by X. We define the group EX (G, U ) ≤ Aut(T~ ) as
follows: EX (G, U ) is the set of g ∈ Aut(T~ ) such that σ(g, v) ∈ G for all v ∈ V T~
and that σ(g, v) ∈ U for all but finitely many v ∈ V T~ .
It is easy to verify that EX (G, U ) is an abstract group. With more care, one can
identify a natural t.d.l.c.s.c. topology on EX (G, U ). For a vertex v ∈ V T~ , it is an
easy exercise to see that the point stabilizer EX (U, U )(v) is compact in the topology
inherited from Aut(T~ ). The topology on Aut(T~ ) is that of pointwise convergence;
this topology is Polish but not locally compact.
Proposition 9.2. For (G, X) a t.d.l.c.s.c. permutation group and U ∈ U(G), there
is a t.d.l.c.s.c. group topology on EX (G, U ) such that the inclusion EX (U, U )(v) ,→
EX (G, U ) is continuous and open for any v ∈ V T~ .
The t.d.l.c.s.c. permutation group (G, X) controls the t.d.l.c.s.c. group EX (G, U )
in several important ways.
Proposition 9.3. Suppose that (G, X) is a t.d.l.c.s.c. permutation group. Then,
(1) If (G, X) is transitive and G is compactly generated, then EX (G, U ) is compactly generated for any U ∈ U(G).
(2) If G is elementary, then EX (G, U ) is elementary for any U ∈ U(G).
The group EX (G, U ) admits a canonical open normal subgroup. Let PX (G, U )
denote the set of g ∈ EX (G, U ) such that g fixes a vertex; equivalently, PX (G, U ) is
the set of g ∈ EX (G, U ) such that η(gv) = η(v) for some (equivalently, any) v ∈ V T~ .
It is easy to verify that PX (G, U ) is an open, normal subgroup of EX (G, U ) and
that EX (G, U )/PX (G, U ) ' Z. The subgroup PX (G, U ) also gives a chief factor.
Recall that a topological group is monolithic if the intersection of all non-trivial
closed normal subgroups is non-trivial.
Proposition 9.4. Suppose that (G, X) is a transitive t.d.l.c.s.c. permutation group.
Then, EX (G, U ) is monolithic, and the monolith is
M := [PX (G, U ), PX (G, U )].
If G is also topologically perfect, then PX (G, U ) is the monolith.
The pointwise stabilizer Hi := StabL (Xi ) of a horoball Xi , called a horoball
stabilizer, is a closed normal subgroup of L, and Hi ≤ Hj for i ≤ j. We thus
obtain canonical normal factors Hi+1 /Hi of PX (G, U ). We call these factors the
horoball factors of PX (G, U ).
Proposition 9.5. For (G, X) a t.d.l.c.s.c. permutation group, each
L horoball factor
K/L admits a PX (G, U )-equivariant isomorphism ψ : K/L → ∂X (G, U ) where
X is the horoball corresponding to K.
50
COLIN D. REID AND PHILLIP R. WESOLEK
9.2. Stacking type chief factors. We now exhibit an example of a non-elementary
stacking type chief factor in a compactly generated t.d.l.c.s.c. group. Again technical results are stated without proof.
Fix G a non-discrete compactly generated simple t.d.l.c.s.c. group and U ∈ U(G);
for concreteness, one can take P SL2 (Qp ) = G and P SL2 (Zp ) = U . Set X := G/U
and let G y X by left multiplication. The resulting permutation group (G, X) is a
transitive t.d.l.c.s.c. permutation group, so via Proposition 9.3, the group EX (G, U )
is a compactly generated t.d.l.c.s.c. group. In view of Proposition 9.4, PX (G, U ) is
the unique chief factor of EX (G, U ). In this setting, PX (G, U ) is of stacking type
via the following:
Proposition 9.6. Suppose that (G, X) is a transitive permutation t.d.l.c.s.c. group
and U ∈ U(G) is non-trivial. If G is topologically simple and compactly generated,
then PX (G, U ) is a stacking type chief factor of EX (G, U ), and every chief factor
of PX (G, U ) is a horoball factor.
We can also describe the chief factors of PX (G, U ) explicitly.
Via Propositions 9.5
L
and 9.6, every chief factor of PX (G, U ) is isomorphic to N (G, U ).
The group PX (G, U ) is thus a non-elementary stacking type chief factor, giving
an example of Theorem 8.21(7).
9.3. Chief factors of weak type. We apply our construction a second time to
exhibit rank ω+1 weak type chief factors of a compactly generated t.d.l.c.s.c. group.
Our construction requires an observation on the rank of the groups EX (G, U ).
Proposition 9.7. Suppose that (G, X) is a compactly generated transitive t.d.l.c.s.c.
permutation group. If G is elementary with finite rank, then EX (G, U ) is elementary with
ξ(EX (G, U )) = ω + 2
for any non-trivial U ∈ U(G).
Let D be the infinite dihedral group and C be the order two subgroup which
reverses orientation. Setting X := D/C, we obtain a t.d.l.c.s.c. permutation group
(D, X). We show that G := EX (D, C) has a chief factor of weak type with rank
ω + 1.
By Proposition 9.7, G has rank ω +2. Applying Proposition 9.4, G is monolithic,
and the monolith M is such that G/M is metabelian. Since solvable groups are
finite rank, Lemma 3.6 and Corollary 3.21 ensure that ξ(M ) = ω + 1. We now
argue that M is of weak type.
Proposition 9.8. Bmin
M = ∅.
Proof. Let A/B be a chief factor of M and suppose for contradiction that [A/B] ∈
Bmin
M . Fixing a base vertex v =: v0 , we form the horoballs Xi and set Hi to be the
horoball stabilizer
in PX (G, U ).
S
The union i∈Z Hi is a normal subgroup of EX (G,
S U ), and it is dense in PX (G, U ).
Since M is non-abelian and chief, it follows that i∈Z Hi ∩ M
S is dense and normal
in M . If A ∩ Hi ≤ B for all i, then A/B commutes with i∈Z Hi ∩ M in M/B.
It follows that A/B is abelian, which contradicts our hypothesis. We thus deduce
that A ∩ Hi B for some i, so A ∩ Hi covers the chief block [A/B]. The intersection of all Hi is trivial, hence since [A/B] is minimally covered, we may take
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
51
A ∩ Hi covering [A/B] such that A/ ∩ Hi−1 does not cover [A/B]. The normal
factor A ∩ Hi /A ∩ Hi−1 thus covers the chief block [A/B].
The normal factor A ∩ Hi /A ∩L
Hi−1 admits an injection into Hi /Hi−1 . Proposition 9.5 implies that Hi /Hi−1 ' N (D, C), and therefore, Hi /Hi−1 is metabelian.
The factor A∩Hi /A∩Hi−1 is then also metabelian. We may now find K/L ∈ [A/B]
such that A ∩ Hi−1 ≤ L < K ≤ A ∩ Hi , and since K/L is chief, it follows that K/L
is abelian. This is absurd as we assume K/L is a non-abelian chief factor.
The group M is thus a rank ω + 1 chief factor of weak type, giving an example
of Theorem 8.23(4).
9.4. Rank two chief factors. Let U + (A6 ) be the Burger–Mozes universal simple
group for the alternating group on 6 elements with the usual permutation representation; see [4, 5] for a discussion of these groups. The group U + (A6 ) acts on
the 6-regular rooted tree T6 . Fixing δ an end of T6 , we consider the end stabilizer
G := U + (A6 )(δ) .
Fix a base vertex v ∈ T6 and let v = v0 , . . . , vn , . . . be a geodesic ray representing
the end δ. Fix t ∈ G a translation down the geodesic ray and let E E G be the
collection of elliptic isometries — i.e. the isometries which fix a vertex. It is an
exercise to see that E is a locally elliptic closed normal subgroup and G/E ' Z.
Take the edge e = (v0 , v1 ), where v0 and v1 are the first two elements of our
distinguished ray, and define Tv1 to be the half-tree determined by e which contains
v1 . It now follows that
[
tn StabG (Tv1 )t−n .
E=
n≥1
The stabilizer StabG (Tv1 ) is the profinite group given by taking the inverse limit
of the iterated wreath products of (A5 , [5]). We thus deduce that StabG (Tv1 ) is
a topologically perfect profinite group, and it has trivial quasi-center. It follows
further that E has trivial quasi-center.
Lemma 9.9. The subgroup E is a non-negligible chief factor of G with rank two.
Proof. Suppose that N E G is non-trivial, closed, and contained in E. Fixing
n ∈ N non-trivial, there is some nearest vertex w ∈ T6 to the ray v0 , v1 . . . that is
moved by n. By conjugating n with t±1 and elements of E, we may assume that
w = v0 and n fixes v1 .
Let e be the edge from v0 to v1 and let Tv1 be the half-tree determined by e that
contains v1 . We now apply a familiar trick: for any g, h ∈ StabG (Tv1 ), the commutator [[n−1 , g −1 ], h] equals [g, h] ∈ StabG (Tv1 ). The subgroup N thus contains
all commutators of StabG (Tv1 ). The stabilizer StabG (Tv1 ) is topologically perfect,
so N contains StabG (Tv1 ). Since N is normal, N indeed contains tn StabG (Tv1 )t−n
for all n. We thus deduce that N = E, so E is a chief factor of G.
The group E is locally elliptic, hence ξ(E) = 2. Applying Theorem 7.15, the
chief block [E] is either connected compact, quasi-discrete, or robust. Since G is
totally disconnected, [E] is not connected compact. The representative E is not
quasi-discrete, so [E] is also not quasi-discrete. We thus deduce that [E] is a robust
chief block. A fortiori, E is a non-negligible chief factor of G.
The group E is thus a rank 2 chief factor with trivial quasi-center, giving an
example of Theorem 8.21(3).
52
COLIN D. REID AND PHILLIP R. WESOLEK
References
1. Herbert Abels, Kompakt definierbare topologische Gruppen, Math. Ann. 197 (1972), 221–233.
MR 0315039
2. Bachir Bekka, Pierre de la Harpe, and Alain Valette, Kazhdan’s property (T), New Mathematical Monographs, vol. 11, Cambridge University Press, Cambridge, 2008. MR 2415834
3. Matthew G. Brin, Elementary amenable subgroups of R. Thompson’s group F , Internat. J.
Algebra Comput. 15 (2005), no. 4, 619–642. MR 2160570 (2007d:20052)
4. Marc Burger and Shahar Mozes, Groups acting on trees: from local to global structure, Inst.
Hautes Études Sci. Publ. Math. (2000), no. 92, 113–150 (2001). MR 1839488
5. Pierre-Emmanuel Caprace and Tom De Medts, Simple locally compact groups acting on trees
and their germs of automorphisms, Transform. Groups 16 (2011), no. 2, 375–411. MR 2806497
(2012m:22033)
6. Pierre-Emmanuel Caprace and Nicolas Monod, Decomposing locally compact groups into simple pieces, Math. Proc. Cambridge Philos. Soc. 150 (2011), no. 1, 97–128. MR 2739075
(2012d:22005)
7. Yves Cornulier and Pierre de la Harpe, Metric geometry of locally compact groups,
arXiv:1403.3796 [math.GR].
8. Karl H. Hofmann and Sidney A. Morris, The Lie theory of connected pro-Lie groups, EMS
Tracts in Mathematics, vol. 2, European Mathematical Society (EMS), Zürich, 2007, A
structure theory for pro-Lie algebras, pro-Lie groups, and connected locally compact groups.
MR 2337107
9. Alexander S. Kechris, Classical descriptive set theory, Graduate Texts in Mathematics, vol.
156, Springer-Verlag, New York, 1995. MR 1321597 (96e:03057)
10. Adrien Le Boudec, Groups acting on trees with almost prescribed local action, Comment.
Math. Helv. 91 (2016), no. 2, 253–293. MR 3493371
11. Nicolas Monod, Continuous bounded cohomology of locally compact groups, Lecture Notes in
Mathematics, vol. 1758, Springer-Verlag, Berlin, 2001. MR 1840942
12. Deane Montgomery and Leo Zippin, Topological transformation groups, Interscience Publishers, New York-London, 1955. MR 0073104 (17,383b)
13. V. P. Platonov, Locally projectively nilpotent subgroups and nilelements in topological groups,
Izv. Akad. Nauk SSSR Ser. Mat. 30 (1966), 1257–1274. MR 0202909 (34 #2768)
14. Colin D. Reid and Phillip R. Wesolek, Chief factors in Polish groups, arXiv:1509.00719
[math.GR].
15.
, The essentially chief series of a compactly generated locally compact group,
arXiv:1509.06593 [math.GR].
16.
, Homomorphisms into totally disconnected, locally compact groups with dense image,
arXiv:1509.00156 [math.GR].
17. Luis Ribes and Pavel Zalesskii, Profinite groups, Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and
Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics], vol. 40, SpringerVerlag, Berlin, 2000. MR 1775104
18. Derek J. S. Robinson, A course in the theory of groups, second ed., Graduate Texts in Mathematics, vol. 80, Springer-Verlag, New York, 1996. MR 1357169 (96f:20001)
19. Simon M. Smith, A product for permutation groups and topological groups, arXiv:1407.5697
[math.GR].
20. Phillip Wesolek, Elementary totally disconnected locally compact groups, Proc. Lond. Math.
Soc. (3) 110 (2015), no. 6, 1387–1434. MR 3356810
21. David Zerling, Some theorems on (CA) analytic groups, Trans. Amer. Math. Soc. 205 (1975),
181–192.
University of Newcastle, School of Mathematical and Physical Sciences, University
Drive, Callaghan NSW 2308, Australia
E-mail address: [email protected]
Department of Mathematical Sciences, Binghamton University, PO Box 6000, Binghamton, New York 13902
E-mail address: [email protected]
DENSE NORMAL SUBGROUPS AND CHIEF FACTORS
53
Previous address: Université catholique de Louvain, Institut de Recherche en Mathématiques
et Physique (IRMP), Chemin du Cyclotron 2, box L7.01.02, 1348 Louvain-la-Neuve, Belgium
| 4math.GR
|
1
Online Geographical Load Balancing for
Mobile Edge Computing with Energy
arXiv:1704.00107v3 [cs.IT] 23 Dec 2017
Harvesting
Jie Xu, Hang Wu, Lixing Chen, Cong Shen
Abstract
Mobile Edge Computing (MEC) (a.k.a. fog computing) has recently emerged to enable low-latency
and location-aware data processing at the edge of mobile networks. Since providing grid power supply
in support of MEC can be costly and even infeasible in some scenarios, on-site renewable energy is
mandated as a major or even sole power supply. Nonetheless, the high intermittency and unpredictability
of energy harvesting creates many new challenges of performing effective MEC. In this paper, we develop
an algorithm called GLOBE that performs joint geographical load balancing (GLB) and admission
control for optimizing the system performance of a network of MEC-enabled and energy harvestingpowered base stations. By leveraging and extending the Lyapunov optimization with perturbation technique, GLOBE operates online without requiring future system information and addresses significant
challenges caused by battery state dynamics and energy causality constraints. Moreover, GLOBE works
in a distributed manner, which makes our algorithm scalable to large networks. We prove that GLOBE
achieves a close-to-optimal system performance compared to the offline algorithm that knows full
future information, and present a critical tradeoff between battery capacity and system performance.
Simulation results validate our analysis and demonstrate the superior performance of GLOBE compared
to benchmark algorithms.
I. I NTRODUCTION
Mobile computing and the Internet of Things are driving the development of many new
applications, turning data and information into actions that create new capabilities, richer experiJ. Xu and L. Chen are with the Department of Electrical and Computer Engineering, University of Miami, USA. Email:
[email protected], [email protected].
H. Wu and C. Shen are with the Department of Electric Engineering and Information Science, University of Science and
Technology of China. Email: [email protected], [email protected]
2
ences and unprecedented economic opportunities. Although cloud computing enables convenient
access to a centralized pool of configurable and powerful computing resources, it often cannot
meet the stringent requirements of latency-sensitive or geographically constrained applications,
such as mobile gaming, augmented reality, tactile Internet and connected cars, due to the often
unpredictable network latency and expensive bandwidth [1], [2]. As a remedy to these limitations,
Mobile Edge Computing (MEC) [3] (a.k.a. fog computing [4]) has recently emerged as a new
computing paradigm to enable in-situ data processing at the Internet edge, in close proximity to
mobile devices, sensors, actuators and connected things. In MEC, network edge devices, such
as base stations (BSs) [2], are endowed with cloud-like functionalities to serve users’ requests
as a substitute of clouds, while significantly reducing the transmission latency as they are just
one wireless hop away from end users and data sources.
In increasingly many scenarios, BSs are primarily powered by renewable green energy (e.g.
solar and/or wind), rather than the conventional electric grid, due to various reasons such as
location, reliability, carbon footprint and cost. The high intermittency and unpredictability of
energy harvesting (EH) [5] significantly exacerbate the challenge of the latency requirements of
applications as the computing capacity of an individual MEC-enabled BS is significantly limited
at any moment of time. Geographical load balancing (GLB) [6], [7] is a promising technique
for optimizing MEC performance by exploiting the spatial diversity of the available renewable
energy to re-shape the computation workload distribution among the distributed BSs. However,
energy harvesting leads to extraordinary challenges that existing GLB approaches cannot address:
not only the available energy in the batteries imposes a stringent energy constraint at any time
moment, but also the intrinsic evolution of these constraints couple the GLB decisions across
time, and yet the decisions have to be made without foreseeing the future. Compared to existing
GLB approaches for data center networks that solve time-decoupled problems, GLB for EHpowered MEC networks demands a fundamentally new design that can optimally manage limited
energy, computing and radio access resources in both spatial and temporal domains.
In this paper, we study the joint problem of geographical load balancing, admission control
and energy purchase among a network of EH-powered BSs (See Figure 1 for an illustration)
aimed at improving the mobile edge computing performance. The main contributions of this
paper are as follows.
(1) We develop a novel framework, called GLOBE (Geographical LOad Balancing with Energy
harvesting), for minimizing the long-term system cost (due to violating the computation delay
3
Power grid
(market energy price)
Purchase ?
Load balancing
(communication traffic)
Energy harvesting
Load balancing
(computation tasks)
(intermittent and unpredictable)
Computation offloading
Communication traffic
Fig. 1. Illustration of load balancing in the MEC-enabled BS network with energy harvesting.
constraint and dropping data traffic). In addition to the spatial coupling originated from load
balancing among geographically distributed BSs, the considered problem also exhibits strong
temporal coupling due to the energy harvesting causality constraint. Such spatial-temporal
coupling makes the investigated problem significantly different and more challenging than conventional GLB problems studied in the literature.
(2) We develop an online and distributed algorithm leveraging the Lyapunov stochastic optimization framework to solve the spatial-temporal GLB optimization problem. Instead of the
conventional Lyapunov virtual queue technique, our algorithm is based on the Lyapunov perturbed queue technique to handle the energy causality constraint. The algorithm enables BSs
to make GLB decisions without foreseeing the future system dynamics yet provides provable
performance guarantee. In particular, we prove that GLOBE achieves within a bounded deviation
from the optimal system performance that can be achieved by an oracle algorithm that knows
the complete future information, provided that the battery capacity of the BSs is sufficiently
large. The performance-battery capacity tradeoff is theoretically characterized.
(3) The perturbed battery queue technique decouples the GLB problem across time. In each
time slot, we also develop a distributed algorithm to solve the per-time slot GLB problem.
Lagrangian dual decomposition with quadratic regularization is used to relax the computation
capacity constraint of individual BSs and decouple the spatially-coupled problem into subproblems that can be solved by individual BSs.
(4) We run extensive simulations to evaluate the performance of GLOBE and verify our analytical results for various system configurations and traffic arrival patterns. The results confirm that
our method significantly improves the system performance compared to benchmark algorithms
that do not perform GLB or only optimize the system myopically.
4
The rest of this paper is organized as follows. Section II discusses related works. Section
III presents the system model and formulates the problem. Section IV develops the GLOBE
framework and distributed and online solutions. Section V proves the performance guarantee of
GLOBE. Section VI shows the simulation results. Finally, Section VII concludes the paper.
II. R ELATED WORK
Mobile Edge Computing [2] (a.k.a fog computing computing [4]) has emerged as a new
paradigm to provide service with ultra-low latency and precise location awareness by bringing
the computation resource closer to the end users. However, compared to the conventional mobile
cloud computing, the edge servers are still limited in the computational and storage resource [8]
and hence may fail to offer the expected quality of service (QoS) when facing excessive workload.
To address this challenge, hierarchical edge computing architectures [9], [10] are investigated
to allow overloaded edge servers to further offloading computation tasks to the cloud server or
the higher-tier servers. While this structure effectively solves the potential overloading problem
at the edge servers, the computation tasks further offloaded to the cloud server will suffer from
the large congestion delay due to backhaul transmission and hence degrades the QoS.
Many recent works [11], [12], [13] investigate cooperation among edge entities to improve
uses’ experience by fully utilizing the resource within the edge system and reducing the dependency on the cloud server. Although the cooperation among BSs in the conventional wireless
communication has been well investigated subject to various constraints, e.g., radio resources
[14], energy consumption budgets [15] and backhaul bandwidth capacity [16]), the cooperation
among MEC-enabled BSs is a very different topic since the radio resource and computational
resource need to be considered jointly. Efforts have been made to coordinate the edge entities
to serve the mobile users collaboratively. In [12], [17], computation load distribution among the
network of BSs is investigated by considering both radio and computational resource constraints,
where clustering algorithms are proposed to maximize users’ satisfaction ratio while keeping the
communication power consumption low. In [11], [18], [19], the coalitional game theory is applied
to enable distributed formation of femto-clouds. However, these works study time-decoupled
problems and do not consider energy-harvesting and hence, are very different from our paper.
In [20], cooperation among two BSs powered by renewable energy is investigated. However, it
considers only two BSs and does not study MEC. Our prior work [10] is one of the first works
that consider energy harvesting in the context of MEC. However, it studies the hierarchical
5
offloading problem for a single MEC-enable BS with energy harvesting capability but neglects
the cooperation among multiple BSs. In this paper, we complete the story by exploring the
cooperation among MEC-enabled BSs through geographical load balancing of communication
traffic and computation workload.
Geographical Load Balancing (GLB) has been extensively studied in data center network
(DCN) research [6], [21], [22], [23], [7]. Most of these works study load balancing problems
that are independent across time and hence a myopic optimization problem is often formulated
and solved to derive the GLB policy. Only a few works consider temporally coupled GLB
problems. For example, authors in [6] consider the temporal dependency with respect to the
switching costs (turning on/off) of data center servers, which significantly differs from our
considered problem where the temporal dependency is caused by the energy harvesting and
energy consumption of BSs. Hence, different techniques are required to solve our problem.
A similar temporally coupled GLB problem was studied under the framework of Lyapunov
optimization in [7], where the authors consider a long-term water consumption constraint while
optimizing the load balancing. However, firstly, the constraint in [7] is a long-term average
constraint whereas our paper considers a much more complicated and stringent battery causality
constraint, and secondly, the constraint in [7] is imposed on the entire network whereas in our
paper each individual BS is constrained by its own (time-varying) battery state. To address
these new challenges, we leverage the Lyapunov optimization with perturbation technique [24]
to develop our algorithm with a provable performance guarantee.
Recent works [6], [23] studied the GLB with renewables, which show that GLB creates an
important opportunity by allowing for “follow the renewables” routing. However, renewables
in these works are considered as an instantaneous supplement to grid power, which needs
to be matched by the energy demand instantaneously. By contrast, our considered problem
uses renewables as the major power source and can be stored in the battery co-located with
the BSs. Therefore, the battery state dynamics and the energy causality constraint need to
be carefully considered while performing GLB. More importantly, the GLB algorithms (e.g.
Averaging Fixed Horizon Control [6]) in these works requires future information (albeit few),
whereas our algorithm is able to work without foreseeing the future.
6
III. S YSTEM M ODEL
A. Network Model
We consider a network of N base stations (BS), indexed by N = {1, 2, ..., N}, providing
communication and edge computing services to users. Each BS is mainly powered by renewable
energy harvested from wind and/or solar radiation, and is equipped with a battery for energy
storage. Renewable energy is considered as free. However, BSs can also purchase energy from
the electric grid whenever needed, at a cost depending on the current energy market price and
their energy demand. We consider a dense deployment scenario where BSs have overlapping
coverage areas so that geographical load balancing is possible.
Time is discretized, with each time slot matching the timescale at which load balancing
decisions can be updated. In each time slot t, each BS i has a set of associated users Uit according
to a certain user-cell association rule (e.g., maximizing the received signal strength). The BS can
collect information (e.g. demand, channel conditions etc.) of its associated users and dispatch
their loads to itself or nearby BSs thanks to the dense deployment of BSs. Having the associated
BSs to make load balancing decisions instead of the user themselves can significantly reduce
the algorithm complexity and information exchange overhead, thereby improving the scalability
of the system. Let U t = ∪i∈N Uit be the set of all users in time t. For each user u, let Nu ⊆ N
be the set of BSs that are in its transmission range.
Although our framework and algorithm can handle a time-varying user set across time slots,
for the ease of exposition, we will assume that the user set is static over the considered time
horizon. In an alternative network model, we can also imagine that each user represents a service
area which remains static although users inside it can change. Therefore, in the remainder of
this paper, we drop the time index t in the user set U t and Uit , ∀i ∈ N .
B. Communication Traffic and Cost Model
Users have uplink and downlink communication traffic that has to be served by the BSs. We
assume that uplink and downlink transmissions operate on orthogonal channels and focus on the
downlink traffic since energy consumption associated with downlink transmissions predominates
the total BS energy consumption [25]. Henceforth, communication traffic refers to downlink
communication traffic. In each time slot t, each user u has communication traffic at arrival rate
µti ≤ µmax , which we assume to be upper bounded by µmax . The data size of communication
7
traffic arrival is modeled as an exponential random variable with mean ωu for user u. Using
Shannon’s theorem, the expected energy consumption for each transmission by BS i to its
associated user u can be computed as
pti,u = E
P ωu
,
tx,i H
t
i,u Ptx,i
W log2 1 + σ2
(1)
t
where W is the downlink bandwidth, Hi,u
is a random variable representing the downlink channel
gain between BS i and user u in time slot t, Ptx,i is BS i’s transmitting power, σ 2 is the noise
power, and the expectation is taken over the data size and the channel state.
For each user u, its communication traffic can be transmitted by any of the BSs in Nu that
covers u, e.g., via Coordinated Multi Point (CoMP) with Coordinated Scheduling (CS) [26].
t
t
represents the amount of
Denote αut = {αu,j
}j∈Nu as user u’s transmission strategy where αu,j
traffic that is transmitted by BS j. Since a BS has multiple associated users, the transmission load
balancing strategy of BS i is the collection of individual users’ transmission strategy, represented
by αti = {αtu }u∈Ui . Further, the transmission load balancing strategy of the whole network is
collected in the notation αt = {αti }i∈N . Clearly, a feasible transmission load balancing strategy
must satisfy
X
t
αu,j
≤ µtu , ∀u.
(2)
j∈Nu
Note that the strict inequality in the above constraint means that some communication traffic
of a user may not be fulfilled by any BS. This could happen when energy supply of the BSs
becomes a concern and hence the BSs have to drop some communication traffic to save energy.
An alternative interpretation of constraint (2) is that user requested files, especially video files,
are transmitted at a lower resolution and hence only the incremental high-resolution layers of
the video will be dropped.
Depending on the amount of transmitted traffic and the current channel conditions, the transmission energy consumption of BS i in time slot t is
X
t
Etx,i (αt ) =
pti,u αu,i
.
u∈U
(3)
8
Dropping communication traffic incurs costs to the network operator due to, e.g., user dissatisfaction or reduced Quality of Experience (QoE). The cost is user-specific depending on user
priority. The total communication traffic dropping cost of BS i is thus
X
X
t
αu,j
),
Ctx,i (αt ) =
ctx,u (µtu −
u∈Ui
(4)
j∈Nu
where ctx,u is the unit dropping cost for user u which converts the dropped communication traffic
into a monetary value.
C. Computation Tasks and Cost Model
In each time slot t, each user also generates computationally intensive tasks that have to be
offloaded to the BS/edge server for processing. When the computation is finished, the computation results will be returned to the user. The computation task arrival of user u is denoted by
λtu ≤ λmax , where λmax is the maximum arrival rate. For each computation task, the required
number of CPU cycles is assumed to be an exponential random variable with mean ρ. The
computation capability of BS i is measured by its CPU speed (i.e. CPU cycles per second),
denoted by fi . Given the the constant CPU speed of BS i, and the exponential distribution
of workload of each computation task, the processing time of a computation task follows an
exponential distribution with mean ρ/fi . Therefore, if BS i processes computation tasks with
arrival rate λ, the average computation delay (including the waiting time and the processing
time) for a task, can be computed using a M/M/1 queuing model
dti =
1
.
fi /ρ − λ
(5)
The workload is delay-sensitive which has to satisfy a maximum delay constraint dmax . If the
total computation workload on a BS is too large, then the delay constraint will be easily violated.
Therefore, computation load balancing is also performed to exploit the under-used, otherwise
wasted, computational resources on BSs with light workload to improve the overall system
performance.
In dense networks, computation load balancing is realized by offloading computation tasks
t
}j∈Nu be
to nearby BSs other than the directly associated BS. For each user u, let β tu = {βu,j
its offloading strategy. A computation load balancing strategy of BS i is the collection of its
users’ offloading strategy β ti = {β tu }u∈Ui and a computation load balancing strategy of the whole
9
network is β t = {β ti }i∈N . In a feasible computation load balancing strategy, the total offloaded
workload of user u must not exceed its computation demand, namely
X
t
βu,j
≤ λtu , ∀u.
(6)
j∈Nu
Moreover, the received workload of BS i must not exceed its computation capacity to satisfy
the delay constraint, namely
X
t
βu,i
≤ fi /ρ − 1/dmax , ∀i.
(7)
u∈U
Due to the above constraints, users can also choose to drop some computation workload (or
process locally on user devices, or offload to the remote cloud) if the edge BS network collectively
cannot support it.
The computation energy consumption of BS i is proportional to its received workload (from
P
t
its own users or users of nearby BSs), namely u∈U βu,i
, and the square of its CPU speed (fi )2
[2]. Therefore, the computation energy consumption of BS i in time slot t is
X
t
,
βu,i
Ecom,i (β t ) = κ(fi )2
(8)
u∈U
where the coefficient κ is for unit task energy consumption. Dropping computation workload
incurs a cost, which is linear to the dropped computation workload. The computation dropping
cost for BS i (considering its own associated users) is therefore
X
X
t
βu,j
),
Ccom,i (β t ) =
ccom,u (λtu −
u∈Ui
(9)
j∈Nu
where ccom,i is the unit workload dropping cost for user u which converts the dropped compu-
tation tasks into a monetary value.
D. Energy Harvesting and Purchase
BSs in the considered system are mainly powered by renewable energy harvested from
the environment, such as wind energy and/or solar energy. To capture the intermittent and
unpredictable nature of the energy harvesting process, we model it as successive energy packet
arrivals, i.e. in each time slot t, energy packets with amount Eit ≤ E max arrive at BS i, where
Eit is drawn from some unknown distribution upper bounded by E max . In each time slot t, part
of the arrived energy, denoted by eti , satisfying
0 ≤ eti ≤ Eit , ∀i
(10)
10
will be harvested and stored in the battery of BS i. We start with assuming that the battery
capacity is sufficiently large. Later we will show that by picking the value of eti , the battery
energy levels can be deterministically upper-bounded under the proposed algorithm, thus we
only need finite-capacity batteries in the actual implementation. More importantly, including eti ’s
as decision variables in the optimization facilitates the derivation and performance analysis of
the proposed algorithm. Similar techniques were adopted in existing works [27], [28]. We collect
the energy harvesting decisions of all BSs in the notation et = {et1 , ..., etN }.
When renewable energy falls short, BSs can also purchase energy from the electric grid
depending on its energy demand and the current market energy price, which varies over time.
Let ctgrid be the unit energy price in time slot t. BS i’s energy purchase decision is denoted by
git ∈ [0, g max ] where g max is the maximum grid energy a BS can purchase. The energy purchase
cost is thus
Cgrid,i (git ) = ctgrid git .
(11)
E. Problem Formulation
The objective of the network operator is to minimize the total system cost in (12) due to
dropping communication traffic/computation workload and purchasing grid energy by jointly
optimizing load balancing, energy harvesting and purchase in each time slot.
Ci (αt , βt , git ) , Ctx,i (αt ) + Ccom,i (β t ) + Cgrid,i (git )
(12)
The optimization is subject to an energy availability constraint, namely the consumed energy in
each time slot must not exceed what is available. The total energy consumption of BS i in time
slot t includes communication and computation energy consumption, i.e.
t
t
Eit (αt , βt ) = Etx,i
(αt ) + Ecom,i
(β t )
(13)
Let Bit denote the available battery energy at the beginning of time slot t for BS i. The energy
causality constraint must be satisfied in every time slot
Eit (αt , β t ) ≤ Bit , ∀i ∈ N
(14)
At the end of each time slot, the battery state evolves as follows
Bit+1 = min{Bit − Eit (αt , β t ) + eti + git , B max }
(15)
11
where B max is the battery capacity.
The problem can thus be formulated as follows:
P1
min
αt ,βt ,et ,g t ,∀t
subject to
T
1 XX
E Ci (αt , β t , git )
lim
T →∞ T
t=1 i∈N
(16a)
Constraints (2), (6), (7), (10), (14), ∀t
where (2), (6) are the constraints for feasible transmission load balancing strategy and feasible
computation load balancing strategy, respectively. (7) is the computation capacity constraint for
each BS. (10) indicates that the harvested energy cannot exceed the green energy arrivals. (14)
is the battery causality constraint, indicating that the energy consumption in the current time slot
should not exceed the battery level.
Because of the battery state dynamics and energy causality constraints, the load balancing
decisions are highly coupled across time slots. Let C1∗ be the infimum time average system
cost achievable by any policy that meets the required constraints in every time slot, possibly by
an oracle algorithm that has complete future information of the communication traffic arrival
process, the computation task arrival process, the energy harvesting process, the market energy
price, and the channel conditions. We note that C1∗ represents a performance upper bound for
practical algorithms as they do not possess complete and accurate future information. In the next
sections, we will develop an algorithm that achieves C1∗ within a bounded deviation without
requiring future information.
IV. O NLINE L OAD BALANCING
FOR
EH-P OWERED MEC
A. An Online Algorithm based on Perturbed Lyapunov Optimization
Lyapunov optimization is a powerful framework that enables online stochastic optimization
without requiring future system dynamics yet provides provable performance guarantee. However,
conventional Lyapunov optimization [29] is not directly applicable for solving P1 due to the
presence of energy causality constraints (14). In order to circumvent this obstacle, we take
an alternative approach based on the perturbed Lyapunov technique similar to [30]. First, we
12
formulate a slightly modified version of P1 as follows:
P2
min
αt ,β t ,et ,gt ,∀t
subject to
T
1 XX
E C αt , β t , git
lim
T →∞ T
t=1 i∈N
T
1 XX t t t
E Ei (α , β ) − git − eti = 0
T →∞ T
t=1 i∈N
lim
(17a)
(17b)
Constraints (2), (6), (7), (10), ∀t
where the energy causality constraint (14) in P1 is replaced with a long-term energy demand
and supply clearance constraint (17b). It can be easily shown that P2 is a relaxed version of P1,
as any feasible solution to P1 would also satisfy the constraints in P2. To see this, consider any
policy that satisfies (10) and (14), then summing equation (15) over t ∈ {1, . . . , T }, dividing by
T and taking limits as T → ∞ yields (17b). Let C2∗ denote the optimal value of P2 and then,
we must have C2∗ ≤ C1∗ . Following the framework in [29], it can be shown that the optimal
solution to the relaxed problem P2 can be obtained by the method of stationary randomized
policy, stated in the following lemma.
Lemma 1. There exists a stationary and possibly randomized policy Π that achieves
i
X h
Π,t Π,t
Π,t
E Ci (α , β , gi ) = C2∗
(18)
i∈N
while satisfies the constraints (6), (7), (2), (10) in P2 and
h
i
E Eit (αΠ,t , β Π,t ) − giΠ,t − eΠ,t
=0
i
(19)
Proof. The proof follows the framework in [29] and is omitted here for brevity.
However, the optimal policy Π requires the knowledge of the system dynamics over the entire
time horizon, and hence is very difficult to derive in practice if not impossible. Next, we develop
an online algorithm, called GLOBE, to solve P1 and compare its performance with C2∗ achieved
by Π. We first define the perturbed battery queue for each BS.
Definition 1. The perturbed battery queue B̃it of BS i is defined as
B̃it = Bit − θ, ∀i ∈ N
where θ is the perturbation parameter.
(20)
13
The value of θ will be specified later when we analyze the algorithm performance. In a nutshell,
the proposed GLOBE algorithm aims to minimize the weighted sum of the system cost and the
perturbed energy queue in each time slot, which shall stabilize Bit around the perturbed energy
level θ and meanwhile minimize the system cost, where the weight is adaptively updated over
time. In each time slot t, the load balancing strategies, the amount of harvested and purchased
energy are determined by solving the following optimization problem:
X
t t
t
t
t
t
t
t
t
V · Ci (α , β , gi ) − B̃i · (Ei (α , β ) − gi − ei )
P3
min
β t ,αt ,et ,gt i∈N
subject to
Constraints (2), (6), (7), (10)
(21a)
(21b)
which is parameterized by only the current system state (i.e., communication traffic arrival,
computation task arrival, energy price, channel conditions in time slot t, etc). Therefore, our
algorithm can work online without requiring future information of the system dynamics. Fig.
2 illustrates GLOBE for an arbitrary time slot t. At the beginning of each time slot, GLOBE
observes the system states and battery level. Then, with the observed information, it derives the
optimal energy harvesting, grid power purchase, and load balancing decisions. At the end of the
time slot, the battery states of the BSs are updated based on the acquired energy (via harvesting
and purchase) and the consumed energy (due to transmission and computation), thereby linking
per-time slot problems across time.
Realization
GLOBE
Current System States
Communication traffic: ࣆ௧
Computation task: ࣅ௧
Energy packets: ए௧
௧
Market energy price: ܿௗ
Transmission energy: ௧
Transmission load balancing: ߙ כ௧
System Cost: ܥ௧ ǡ ݅
Computation load balancing: ߚ כ௧
Energy consumption: ܧ௧ ǡ ݅
Grid power purchase: ݃כ௧
Energy harvesting: ݁ כ௧
Battery Dynamics
ܤଵ ൌ ሼܤଵ െ ܧଵ௧ ݁ଵכ௧ ݃ଵכ௧ ǡ ܤଵ௫ ሽ
ڭ
כ௧
ܤே ൌ ሼܤே െ ܧே௧ ݁ேכ௧ ݃ே
ǡ ܤே௫ ሽ
Fig. 2. Illustration of GLOBE for time slot t.
The GLOBE algorithm is formally presented in Algorithm 1. How this algorithm is derived
will be explained and its performance will be theoretically analyzed in Section V. Now, to
complete GLOBE, it remains to solve the per-time slot problem P3.
14
Algorithm 1 The GLOBE algorithm
1: Input: {θi , i ∈ N }, V
2:
Output: Load balancing αt and β t , purchased energy g t , harvested energy et , for every t
3:
for every time slot t do
4:
Observe µt , λt , E t , ctgrid , pt
5:
Solve (P3) to get αt , β t , et and g t
6:
Update the battery state according to (15) for i ∈ N
7:
end for
The objective function in P3 can be decomposed into three parts:
X
V · Ci (αt , β t , git ) − B̃it · (Eit (αt , β t ) − git − eti )
i∈N
X
=
V · ctgrid git + B̃it (git + eti )
i∈N
+
+
X
V ·
X
i∈N
u∈Ui
X
X
i∈N
V ·
ctx,u (µtu −
X
t
αu,j
) − B̃it
ccom,u (λtu −
u∈Ui
t
pti,u αu,i
u
j∈Nu
X
X
t
βu,j
) − B̃it κ(fi )2
X
u∈U
j∈Nu
(22)
!
!
t
βu,i
.
Given the current perturbed battery queues of all BSs, the first term on the right-hand side
depends only on the transmission load balancing strategy αt , the second term depends only on
the computation offloading load balancing strategy β t , and the third term depends only on the
energy harvesting et and purchase strategies g t . Therefore, we can optimize each term separately.
B. Optimal Energy Harvesting and Purchase
We start with solving the per-slot optimal energy harvesting and purchase problem, which is
the first term of (22):
X
t
t
t t
t
V
·
c
g
+
B̃
(g
+
e
)
min
grid i
i i
i
t t
e ,g
subject to
0 ≤ eti ≤ Eit , ∀i.
(23a)
i∈N
This is a simple linear programming (LP) and can be solved by each BS in a fully distributed way
without requiring any message exchange among the BSs. The optimal solution can be derived
as follows
t
t
et∗
i = Ei · 1{B̃i ≤ 0}, ∀i;
git∗ = g max · 1{V · ctgrid + B̃it ≤ 0}, ∀i
(24)
15
where 1{·} is the indicator function. The optimal solution suggests that
1) BS i should harvest all possible energy if the battery queue Bit is less than a threshold θi
and zero energy otherwise;
2) BS i should purchase the maximum amount of grid power if the battery queue is less than
a threshold θi − V · ctgrid and zero grid energy otherwise.
Since θi − V · ctgrid < θi , renewable energy has a high priority than the grid power, i.e.,
renewable energy will be acquired before purchasing any grid energy. This is intuitive because
renewable energy is free. Moreover, because the grid energy price ctgrid varies over time, the
energy purchase strategy also varies depending on ctgrid . The BS is more likely to purchase grid
energy when the price is low. The energy harvesting and purchase strategies are very simple
and both have a “none-or-all” structure. However, we will show later that these simple strategies
indeed ensure that we only need a finite battery capacity for each BS while providing provable
performance guarantee.
C. Optimal Transmission Load Balancing
The optimal transmission load balancing can be obtained by solving the following LP problem:
!
X
X
X X
t
t
(25a)
pti,u αu,i
αu,j
+ B̃it
V ctx,u
max
t
α
u
u∈Ui
j∈Nu
i∈N
X
t
αu,j
≤ µtu , ∀i, ∀u ∈ Ui
(25b)
subject to
j∈Nu
The objective function can be further rearranged as follows
!
X
XX X
X X
X
t
t
t
(V ctx,u + B̃jt ptj,u )αu,j
.
αu,j
+ B̃it
pti,u αu,i
=
V ctx,u
i∈N
u∈Ui
j∈Nu
u
i∈N u∈Ui j∈Nu
This enables each BS i to decide its transmission load balancing strategy for its own associated
users in a distributed way, after exchanging information of the perturbed battery queue lengths
with neighbor BSs. Specifically, BS i solves the following problem for each of its user u ∈ Ui :
X
X
t
t t
t
αu,j
≤ µtu .
(26a)
(V
c
+
B̃
p
)α
subject
to
max
tx,u
j j,u
u,j
αtu j∈N
j∈N
u
u
The optimal solution has a simple structure. If there exist some BS j such that the coefficient
V ctx,u + B̃jt ptj,u ≥ 0, then all user u’s communication traffic will solely be transmitted by the BS
that has the largest positive coefficient. If no BS has positive coefficients, then user u’s downlink
traffic will not be transmitted, thereby causing communication traffic to be dropped. We note
16
that whether traffic dropping occurs or not depends on the cost ctx,u . When ctx,u is sufficiently
large (i.e. dropping communication traffic is very costly), then there will always be BSs with
positive coefficients and hence no traffic will be dropped.
D. Optimal Computation Load Balancing
The optimal computation load balancing strategy can be obtained by solving the following
LP problem
X
max
βt
X
subject to
X
V ccom,u
u∈Ui
i∈N
X
t
βu,j
+ B̃it
X
t
κ(fi )2 βu,i
u
j∈Nu
!
t
βu,j
≤ λtu , ∀i, ∀u ∈ Ui
(27a)
(27b)
j∈Nu
X
t
βu,i
≤ fi /ρ − 1/dmax , ∀i
(27c)
u∈U
Similar to the transmission load balancing problem, the objective function can also be rearranged as follows:
X
i∈N
X
u∈Ui
V ccom,u
X
t
βu,j
+ B̃it κ(fi )2
X
u
j∈Nu
t
αu,i
!
=
XX X
t
(V ccom,u + B̃jt κ(fj )2 )βu,j
.
i∈N u∈Ui j∈Nu
While the computation load balancing problem looks similar to the transmission load balancing
problem in the previous subsection, it indeed is a much more complicated problem because of
the computation capacity constraint (27c) that makes the load balancing decisions among BSs
highly coupled. To enable a distributed solution, we use Lagrangian dual decomposition to relax
the computation capacity constraint and decouple the primal problem into several subproblems.
Unfortunately, the above considered LP problem is sensitive to perturbations and does not work
well with the dual decomposition technique. To overcome this difficulty, we add a quadratic
regularization term to the objective function to smooth the original LP problem1. The regularized
problem is:
max
βt
subject to
XX X
(V ccom,u +
t
B̃jt κ(fj )2 )βu,j
i∈N u∈Ui j∈Nu
X
t
βu,j
≤ λtu , ∀i, ∀u ∈ Ui ;
j∈Nu
1
We note that a similar approach was proposed and adopted in [31].
X
u∈U
1 t 2
)
− (βu,j
2ǫ
t
βu,i
≤ fi /ρ − 1/dmax , ∀i
(28a)
(28b)
17
According to [32], there exists ǫmin > 0 such that the optimal solution to problem (28a) is also
an optimal solution to problem (27a) for any ǫ ≥ ǫmin . For our problem, we have found that
ǫ = 107 works well. Therefore, we will instead solve the smoothed quadratic programming (QP)
problem. We associate a Lagrangian multiplier γi with the computation capacity constraint of
each BS i. The Lagrangian of the primal problem (28a) is thus
XX X
1 t 2 X X t
t
) −
γi (
βu,i − fi /ρ + 1/dmax )
(V ccom,u + B̃jt κ(fj )2 )βu,j
− (βu,j
L(γ, β t ) =
2ǫ
i∈N
u∈U
i∈N u∈U j∈N
u
i
=
XX X
t
(V ccom,u + B̃jt κ(fj )2 − γj )βu,j
−
i∈N u∈Ui j∈Nu
Because the third term
P
i∈N
1 t 2 X
(β ) +
γi (fi /ρ − 1/dmax ).
2ǫ u,j
i∈N
γi (fi /ρ−1/dmax ) is independent of the computation load balancing
strategy β t , the relaxed problem is
XX X
1 t 2
t
(V ccom,u + B̃jt κ(fj )2 − γj )βu,j
− (βu,j
)
max
t
2ǫ
β
i∈N u∈Ui j∈Nu
X
t
βu,j
≤ λtu , ∀i, ∀u ∈ Ui
subject to
(29a)
(29b)
j∈Nu
After this relaxation, the problem can be divided into two levels. At the lower level, each BS
seeks to optimize β t for given γ. The subproblem for BS i is
X X
1 t 2
t
(V ccom,u + B̃jt κ(fj )2 − γj )βu,j
− (βu,j
)
max
t
2ǫ
βi
u∈Ui j∈Nu
X
t
βu,j
≤ λtu , ∀u ∈ Ui
subject to
(30a)
(30b)
j∈Nu
The above problem is a QP. Its optimal value is denoted as Li (γ). Since the problem is strictly
concave, the optimal solution is unique.
At the higher level, each BS i seeks to solve the dual problem:
X
min
(Li (γ) + γi (fi /ρ − 1/dmax ).
γ ≥0
(31)
i∈N
Since the lower level problem has a unique optimal solution, the higher level dual problem is
differentiable. Therefore, we use the gradient descent method to solve this problem in an iterative
manner. In each iteration k, the BSs exchange the current values of γ (k) and each BS i solves
t,(k)
the lower level problem (30a) using the current iteration γ (k) to obtain the optimal β i
t,(k)
the BSs exchange the derived β i
. Then
and use it to update the Lagrangian γ (k+1) for the next
iteration as follows:
(k+1)
γi
(k)
= [γi
− δ (k) (fi /ρ − 1/dmax −
X
u
t,(k)
βu,i (γ (k) ))]+ ,
(32)
18
":DFG
DE
":DFI1G
:H
": F
DE
G
D
":DFG
H1
":DFI1G
:H
H1
D
E
": F
":DFG
HJ
DE
":DFG
DE
":DFG
G
":DFI1G
:H
E
":DFG
HJ
G
": F
DE
G
":DFG
:H
":DFM1G
G
L1
E
D
E
D
":DFG
":DFM1G
HK
":DFM1G
L1
":DFG
HJ
F
":DFG
:H
":DFM1G
LJ
F
": F
DE
":DFG
:H
LJ
G
":DFM1G
LJ
F
":DFM1G
LK
F
":DFM1G
LK
":DFG
HK
each BS updates H( at the lower level
broadcast H( to neighbor BSs
each BS updates L( at the higher level
broadcastNL(N to neighbor BSs
for the next iteration
Fig. 3. Illustration of the iterative algorithm for one iteration.
where δ (k) > 0 is the step size of the subgradient descent method. The dual variable will
converge to the optimal γ ∗ as k → ∞. Fig. 3 illustrates the procedure of the iterative algorithm
for one iteration. Since the Slater’s condition (i.e., the primal problem is convex and there exists
a feasible solution) is satisfied for the primal problem, the strong duality holds. As a result, the
optimal solution β t for the primal problem can also be derived through β t,(k) (γ (k) ) with k → ∞.
We therefore have solved the primal problem. Algorithm 2 summarizes the iterative algorithm
which solves the computation load balancing problem in a distributed manner.
Algorithm 2 The Distributed Algorithm That Iteratively Solves Per-Slot Computation Load
Balancing
1:
Input: B̃it for each BS i at the current time t
2:
Output: Slot t’s computation load balancing β t
3:
Each BS i broadcast B̃it to its neighbor BSs.
4:
for iteration k do
(k)
to its neighbor BSs, ∀i ∈ N
5:
BS i broadcasts its γi
6:
BS i solves (30a) for β i
7:
BS i broadcasts its solution β i
8:
9:
t,(k)
, ∀i ∈ N
t,(k)
BS i updates
(k+1)
γi
to its neighbor BSs, ∀i ∈ N
according to (32), ∀i ∈ N
end for
V. P ERFORMANCE A NALYSIS
In this section, we analyze the performance of GLOBE. To facilitate our exposition, we
min
introduce the following auxiliary notations: cmax
, maxu ctx,u , cmax
,
tx
com , maxu ccom,u , p
min max
mini,u,t pti,u and f min , mini fi . Let cmax , max{cmax
, ccom /(κ(f min )2 )}}. Moreover, we
tx /p
19
max
max
assume that Ei (αt ) is upper bounded by Etx
and Ei (β t ) is upper bounded by Ecom
. This
assumption holds when the communication traffic/computation task arrivals are bounded. Let
max
max
E max = Etx
+ Ecom
.
Lemma 2. For any B max > E max + E max + g max , by choosing
0≤V ≤
B max − E max − E max − g max
cmax
(33)
and θ = V cmax +E max , the battery level Bit is bounded in [0, B max ], ∀i, and the energy causality
constraint is satisfied in every time slot.
Proof. Define θ̃it , θ − V · maxu max{ctx,u /pti,u , ccom,u /(κ(fi )2 )}. Using the definition of cmax ,
we must have θ̃it ≥ E max . Consider the following three cases based on the value of Bit .
Case 1: Bit ∈ [θ, Bmax ]. Our analysis for the optimal energy harvesting and purchase problem
before shows that if Bit > θ, then BS i will not harvest or purchase any energy, namely et∗
i = 0
and git∗ = 0. According to the battery dynamics, Bit+1 ≤ Bit and hence Bit+1 will also be less
than Bmax . On the other hand, because θ > Emax , the energy causality constraint Eit∗ < Bit is
satisfied and the next slot battery will be greater than 0 since Bit+1 = Bit − Eit∗ > θ − E max > 0.
max
t∗
max
Case 2: Bit ∈ [θ̃it , θ]. In this case, we have et∗
, git∗ ≤ g max , Etx,i
≤ Etx
and
i ≤ E
t∗
max
Ecom,i
≤ Ecom
. Therefore,
Bit+1 ≤ θ + E max + g max = V cmax + E max + E max + g max ≤ B max
(34)
where the second equality is obtained by plugging in the definition of θ, and the last inequality
follows the chosen value range of V . On the other hand, because θ̃it ≥ E max , the energy causality
constraint is satisfied and we have Bit+1 ≥ 0.
t∗
t∗
Case 3: Bit ∈ [0, θ̃it ]. We first investigate Etx,i
and Ecom,i
. The objective function in the
transmission load balancing problem (25a) can be rearranged as
!
X
XX
X X
X
t
t
αu,j
+ B̃it
pti,u αu,i
=
(V ctx,u + B̃it pti,u )atu,i .
V ctx,u
i∈N
u∈Ui
j∈Nu
u
(35)
i∈N u∈U
If V ctx,u + B̃it pti,u < 0, ∀u, then the optimal transmission load balancing strategy does not require
BS i to transmit any traffic for any user, namely atu,i = 0, ∀u. This condition is equivalent to
Bit < θ − V ctx,u /pti,u using the definition of B̃it . Further, using the definition of θ̃it , we have
Bit < θ̃it < θ − V ctx,u /pti,u . Therefore, BS i does not incur any transmission energy consumption,
t∗
namely Etx,i
= 0.
20
Similarly, the objective function in the computation load balancing problem (27a) can be
rearranged as
X
i∈N
X
u∈Ui
V ccom,u
X
t
βu,j
+ B̃it κ(fi )2
j∈Nu
X
u
t
αu,i
!
=
XX
t
(V ccom,u + B̃it κ(fi )2 )βu,i
.
(36)
i∈N u∈U
If V ccom,u + B̃it κ(fi )2 < 0, ∀u, then the optimal computation load balancing strategy does not
t
require BS i to perform any computation for any user, namely βu,i
= 0, ∀u. This condition is
equivalent to Bit < θ−V ccom,u /k(fi )2 . Using the definition of θ̃it , we have this condition satisfied.
t∗
Therefore, BS i does not incur any computation energy consumption, namely Ecom,i
= 0.
max
Because we have et∗
and git∗ ≤ g max , clearly the energy causality constraint is
i ≤ E
satisfied, Bit+1 ≥ 0 and Bit+1 ≤ B max .
Lemma 2 is of significant importance because it shows that GLOBE not only yields a feasible
solution to the relaxed problem P2 but also a feasible solution to the original problem P1 since
the energy causality constraint in each time slot is actually satisfied by running GLOBE, provided
that the battery capacity is sufficiently large and the algorithm parameters are properly chosen.
This result helps us to prove the performance guarantee of the proposed GLOBE algorithm.
Next we proceed to show the asymptotic optimality of the GLOBE algorithm, for which we
first define the Lyapunov function as follows:
2
1X t 2 1X t
(B̃i ) =
Bi − θ .
Ψt ,
2 i∈N
2 i∈N
(37)
The Lyapunov drift represents the expected change in the Lyapunov function from one time
slot to another, which is defined as ∆t = E[Ψt+1 − Ψt |B t ], where the expectation is taken with
t
respect to the random process associated the system, given the battery state B t = [B1t , ..., BN
].
Assuming for now that the battery capacity is infinite, the battery state dynamics yields
Bit+1 − θ = Bit − θ − Eit (αt , β t ) + eti + git , ∀i.
(38)
Squaring both sides of the above equation, we obtain,
(Bit+1 − θ)2 = (Bit − θ)2 + (Eit (αt , β t ) − eti − git )2 − 2(Bit − θ)(Eit (αt , β t ) − eti − git ).
(39)
max
max 2
Notice that the term (Eit (αt , β t ) − eti − git )2 ≤ (Eit (αt , β t ))2 + (eti + git )2 ≤ (Etx
+ Ecom
) +
(E max + g max )2 , 2D/N is upper-bounded by a constant 2D/N. Using this bound and rearranging the above equation, we have
(Bit+1 − θ)2 − (Bit − θ)2 ≤ 2D/N − 2(Bit − θ)(Eit (αt , β t ) − eti − git).
(40)
21
Using the above inequality and the definition of ∆t , we have
X
∆t ≤ D − E[ (Bit − θ)(Eit (αt , β t ) − eti − git )|B t ].
(41)
i∈N
P
Adding the system cost multiplied by V , namely V · E[ i∈N Ci (αt , β t , git)|B t ], to both sides
P
and denoting ∆tV = ∆t + V E[ i∈N Ci (αt , β t , g t )|B t ], we have
X
t t
t
t
t
t
t
t
t
t
V Ci (α , β , gi ) − B̃i (Ei (α , β ) − ei − gi ) |B t ].
(42)
∆V ≤ D + E[
i∈N
According to the theory of Lyapunov optimization (drift-plus-penalty method), the control
actions are chosen for each time slot t to minimize the bound on the modified Lyapunov drift
function ∆tV . Therefore, in each time slot t, we solve the per-time slot optimization problem P3
to obtain load balancing strategies and energy harvesting and purchase strategies as in GLOBE.
Theorem 1 provides the theoretical performance guarantee of GLOBE.
Theorem 1. For any V , if the battery capacity satisfies B max ≥ V cmax + E max + E max + g max ,
then the proposed algorithm yields a feasible solution and the achievable time average system
cost satisfies
where D is a constant.
T
1 XX
E Ci (αt , β t , git) ≤ C1∗ + D/V
lim
T →∞ T
t=1 i∈N
(43)
Proof. Consider the bound on the Lyapunov drift function (42). It is clear that the control actions
αt , β t , et , g t by our algorithm minimizes the bound on the Lyapunov function over all possible
control actions. Comparing it with the control actions chosen according to the optimal oracle
policy that achieves C2∗ , we have
X
∆t + V E[
Ci (αt , β t , git)|B t ]
i∈N
≤D + E[
X
i∈N
≤D + E[
X
i∈N
=V E[
X
i∈N
V Ci (αt , β t , git) − B̃it (Eit (αt , β t ) − eti − git ) |B t ]
Π,t
V Ci (αΠ,t , β Π,t , git) − B̃it (Eit (αΠ,t , β Π,t ) − eΠ,t
−
g
)
|B t ]
i
i
Ci (αΠ,t , βΠ,t , giΠ,t )] + D − E[B̃it (Eit (αΠ,t , β Π,t ) − eΠ,t
− giΠt )|B t ].
i
22
Plugging in (18) and (19), taking the expectation on both sides and summing from t = 0, ..., T −1,
normalizing by T and taking the limit T → ∞, we have
T
1 XX
E Ci (αt , βt ) ≤ V C2∗ + D,
V lim
T →∞ T
t=1 i∈N
(44)
where D < ∞ is a constant. The proof is completed with C2∗ ≤ C1∗ .
Theorem 1 proves that GLOBE can achieve the minimum cost achievable by the genieaided offline algorithm within a bounded deviation, without foreseeing the future information.
Moreover, it formalizes a critical tradeoff between the battery capacity and the achievable system
performance: the achievable system performance improves with the increase of the battery
capacity. In particular, the system performance can be made arbitrarily close to optimum if
the battery capacity is large enough. This result provides profound guidelines for EH-powered
MEC network design and deployment, especially on the battery design.
VI. S IMULATION R ESULTS
In this section, we evaluate the performance of GLOBE through simulations. We consider
N = 5 BSs, who are able to perform computation offloading to only some other BSs. To be
specific, we let Mi ∈ N and |Mi | = 3, ∀i. The downlink data traffic arrival at BS i is modeled
as a Poisson process with µti ∈ [0, 10] unit/sec. The expected size of each data traffic E[ω] is
set as 100 Mbits. The transmitting power of BS i is Ptx,i = 1 W, the noise power spectral
density is σ 2 = 0.01 W/Hz, and the bandwidth is W = 20 MHz. The downlink traffic dropping
costs are ctx,i = 10, ∀i. The computation task arrival at BS i is modeled as a Poisson process
with λti ∈ [0, 10] with the mean task size of 1 M unit. The CPU speed is fi = 2.4 GHz
and the expected number of CPU cycles required for each computation task is ρ = 8 × 105 .
The computation constraint is dmax = 1 ms. The energy consumption parameter is chosen as
κ = 2.5 × 10−22 . The cost of dropping one unit of computation task is ccom,i = 0.01. The
harvested energy is modeled as a uniform distribution and satisfies E ∈ [0, 10]. The unit energy
price of the alternative grid power ctgrid is modeled as a uniform distribution with mean 1 if not
specified. The proposed GLOBE is compared with following three benchmark solutions:
•
Non-GLB Stochastic Optimization (SO-NG): SO-NG considers the stochasticity in communication traffic, computation task arrivals and energy harvesting (i.e., battery dynamics
and energy causality) and aims to minimize the long-term system cost. However, in this
23
case, GLB is not enabled in the network. This problem can be solved online by leveraging
the Lyapunov Optimization with Perturbation similar to that in GLOBE.
•
Myopic Optimization with GLB (MO-G): MO-G optimizes GLB and admission control to
minimize the system cost in the current time slot. The decisions are made without concerns
of the future energy harvesting and forthcoming communication/computation workload.
•
Non-GLB Myopic Optimization (MO-NG): MO-NG is the most naive scheme which does
not consider the geographical load balancing or the long-term system performance. Each
BS simply tries to serve all computation workload and communication traffic given the
available energy in the battery and drops whatever cannot be fulfilled.
A. Runtime Performance Comparison
Fig. 4 compares the runtime performance of GLOBE and three benchmarks. We focus on
two metrics: the time-average cost of BSs in Fig. 4(a) and the time-average battery level in
Fig. 4(b). It can be observed from Fig. 4(a) that generally GLOBE achieves the lowest longterm system cost compared to the other three benchmarks. Specifically, GLOBE reduces the
system cost by nearly 50% compared to MO-NG and nearly 30% compared to the second
best scheme, i.e., SO-NG. Moreover, we see that addressing both issues (i.e., loading balancing
of communication/computation and stochastic optimization for long-term system performance)
provides considerable improvements to the system performance. Comparing the time-average
costs of GLOBE and SO-NG (or MO-G and MO-NG), it can be concluded that enabling load
balancing of communication traffic and computation workload among BSs helps to reduce
the system cost. Comparing the time-average costs of GLOBE and MO-G, we see that a
significant system cost reduction can be achieved in the long-run by carefully scheduling the
energy consumption, harvesting and purchasing in each time slot. Another important feature
of GLOBE is that it ensures the performance with bounded battery levels, thereby enabling
practical implement. As can be seen in Fig. 4(b), the time-average battery levels of GLOBE and
SO-NG stabilize at a relatively low value by following the online decisions designed by Lyapunov
optimization with perturbation. This means that the performance of GLOBE is achievable with
a small battery capacity. By contrast, the two myopic benchmarks require a very large battery
capacity to implement these schemes.
20
GLOBE
SO-NG
MO-G
MO-NG
18
16
14
12
10
8
6
0
200
400
600
800
1000
Time-average battery level
Time-average cost of BSs
24
500
GLOBE
SO-NG
MO-G
MO-NG
400
300
200
100
0
0
200
Time slot t
400
600
800
1000
Time slot t
(a) System cost of BSs
Fig. 4. Runtime performance comparison of various solutions.
(b) Battery level of BSs
B. Impact of Control Parameter V
Fig. 5 depicts the time-average system cost and time-average battery level achieved by GLOBE
with various values of V . It shows that the system cost decreases with the increase in V . This
is because a larger V empathizes the cost minimization more in problem P3. However, a lower
system cost is achieved at the price of a higher requirement on battery capacity. As shown in
Fig. 5(a), the time-average battery level stabilizes at a higher value with a larger V , which means
that a larger battery capacity is required to implement the algorithm.
Fig. 6 formally presents the trade-off between the system cost and the battery capacity. It
clearly shows a [O(1/V ), O(V )] trade-off between the stabilized time-average cost and battery
capacity, which is consistent with our analysis in Theorem 1. Moreover, we see that the stabilized
Time-average cost
20
V=0.1
V=1
V=5
15
10
5
0
0
200
400
600
800
1000
Time-average battery level
battery level of BSs closely follows the designed perturbation parameter θ.
200
V=0.1
V=1
V=5
150
100
50
0
Time slot t
(a) Impact on system cost
0
200
400
600
800
1000
Time slot t
(b) Impact on battery level
Fig. 5. Impact of control parameter V to the performance of GLOBE.
C. Impact of Grid Power
We also evaluate the impact of grid power on the algorithm performance. Fig. 7 reports the
performance of GLOBE with or without grid power. Moreover, we vary the expected market
energy price to evaluate the role of energy price in GLOBE. We see from Fig. 7(a) that the
25
500
Time-average cost
Perturbation parameter
Battert level
10
450
400
8
350
300
6
250
4
200
150
2
0
Battery states of BSs
Time-average cost of BSs
12
100
0
5
10
15
50
20
Control parameter V
Fig. 6. Trade-off between system cost and battery capacity.
system cost is reduced with the supplement of grid power. This is due to the fact that the cost
of purchasing grid power is usually lower than that of dropping communication/computation
workloads, and hence the grid power can be used as a supplementary energy source when the
energy harvesting is not sufficient. In the worst case, if the market energy purchasing costs more
than the workload dropping, then GLOBE will simply choose not to purchase. Fig. 7(a) further
indicates that the system cost is reduced along with the decrease in the market energy price
since less fee is charged for the required grid power.
Fig. 7(b) depicts the time-average battery levels of these four cases. It is observed that the
time-average battery levels of cases with grid power stabilize around the same value, i.e., the
perturbation parameter. This is due to the fact that the design of perturbation parameter is
independent of the marker energy price (as shown in Lemma 2). For the case with no grid
power, the stabilized battery level is slightly lower. This is because the BSs cannot maintain the
desired battery level for future use without the grid power, since the energy harvesting may be
Time average cost
30
With grid power (c grid=0.1)
With grid power (c grid=1)
With grid power (c grid=2)
20
No grid power
10
0
0
200
400
600
800
1000
Time slot t
(a) Impact on system cost
Fig. 7. Impact of grid power to the performance of GLOBE.
Time average battery level
insufficient at times.
150
With grid power (c grid=0.1)
With grid power (c grid=1)
With grid power (c grid=2)
100
No grid power
50
0
0
200
400
600
800
Time slot t
(b) Impact on battery level
1000
26
D. Impact of Computation Workload Intensity
Fig. 8 depicts the time-average system costs with different levels of computation workload
intensity. In general, the system cost grows with the increase in computation workload since
larger computation workload incurs higher delay cost and energy consumption. We also see that
if the computation workload intensity becomes too low or too high, the system cost achieved
by GLOBE is almost the same as that achieved by SO-NG. This is reasonable because if the
computation workload is extremely low at every BS, then the BSs can process the workload
locally with their own resources, while incurring low delay cost and energy consumption. Hence,
there is no need for load balancing. On the other hand, if the computation workload intensity
is too high, every BS has already been overloaded, therefore, there is little benefit from loading
balancing among BSs. A large system cost reduction is achieved when computation workload
intensity is at the similar level of the system computation capacity. In this case, there are both
many overloaded BSs and underloaded BSs in the system, and hence the geographical load
balancing can better help to reduce the system cost.
Time-average cost of BSs
35
30
GLOBE
SO-NG
25
20
15
10
5
0
500
1000
1500
2000
2500
3000
Expected computation workload intensity
Fig. 8. Impact of computation workload intensity to the performance of GLOBE and SO-NG.
E. Before versus After GLB
Fig. 9 shows the distribution of computation workload and communication traffic before and
after applying GLB in one particular time slot. It is worth noticing that the GLOBE does not
simply balance the workload evenly among BSs. Instead, GLOBE distributes the workload based
on the battery states of BSs. As can be seen in Fig. 9, BSs with higher battery level (e.g. BS 1
and BS 5) tend to serve more workload and traffic, which helps BSs with low battery level to
avoid battery depletion by conservatively serving less workload, thereby reducing the probability
of purchasing energy from the power grid and decreasing the system cost.
80
70
60
50
BS 1
BS 2
BS 3
BS 4
BS index
BS 5
12
80
Before GLB
After GLB
Battery states
10
70
8
60
6
4
50
2
40
0
(a) GLB for computation workload
BS 1
BS 2
BS 3
BS 4
BS index
Battery states of BSs
Before GLB
After GLB
Battery states
Communication traffic
10000
9000
8000
7000
6000
5000
4000
3000
2000
1000
0
Battery states of BSs
Computation workload
27
40
BS 5
(b) GLB for communication traffic
Fig. 9. Workload distribution before and after GLB.
F. Convergence of Distributed Algorithm
Fig. 10 depicts the achieved objective values by substituting the solutions derived by the
distributed algorithm into the objective functions of primal LP in (27a) and regularized QP in
(28a). There are several observations worth pointing out. First, It can be seen from Fig. 10(a)
that the proposed distributed algorithm ensures the convergence to the optimal solution of the
regularized QP and the optimal objective value of regularized QP is 329.60 in this particular
case. Second, Fig. 10(b) shows that the optimal solution to the regularized QP is exactly the
optimal solution to the primal LP since this solution also achieves the optimal objective value
500
400
300
200
Optimal value of regularized QP
100
0
100
200
300
Objective function value
Objective function value
(i.e., 331.06) of the primal LP problem.
400
iteration k
(a) Convergence of QP during iteration
500
400
300
200
Optimal value of primal LP
100
0
100
200
300
400
iteration k
(b) Convergence of LP during iteration
Fig. 10. Convergence and optimality of the distributed algorithm.
Note that the distributed algorithm may not always obtain the optimal solution to the original
LP problem. However, even when it converges to a suboptimal solution, the gap between the
converged suboptimal value and the optimal value is very small. Among all 1,000 time slots in
our simulation, we only observed 3 time slots that have an error percentage larger than 0.5%
and the maximum error percentage is 3%, which can be reasonably neglected.
28
Fig. 11 compares the performances of distributed and the centralized algorithms. We see that
the performance achieved by the distributed algorithm is almost identical to that achieved by
the centralized algorithm. This further validates that the proposed GLOBE algorithm can be
effectively implemented in a distributed manner at each BS without sacrificing the performance.
500
Time average cost (distributed solution)
Time average cost (centralized solution)
Battery state (distributed solution)
Battery state (centralized solution)
10
400
8
300
6
200
4
Battery states of BSs
Time-average cost of BSs
12
100
2
0
5
10
15
20
Control parameter V
Fig. 11. Performance comparison between distributed and centralized algorithms.
VII. C ONCLUSION
In this paper, we have proposed an online algorithm to perform geographical load balancing
in EH-powered MEC networks. We demonstrated that a fundamentally new design that simultaneously manages the limited energy, computing and radio access resources in both spatial and
temporal domains is key to fully reaping the benefits of EH-power MEC. The proposed GLOBE
algorithm operates online without the need to acquiring future system information. Moreover,
GLOBE can be implemented in a distributed manner, where BSs solve local optimization
problems with very limited information exchange. Our algorithm is simple and easy to implement
in practical deployment scenarios, yet provides provable performance guarantee. Comprehensive
numerical simulations have been carried out to validate the theoretical analysis and illustrate the
performance improvement over other benchmark solutions.
R EFERENCES
[1] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: Vision and challenges,” IEEE Internet Things J., vol. 3,
no. 5, pp. 637–646, 2016.
[2] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey on mobile edge computing: The communication
perspective,” IEEE Commun. Surveys Tuts., 2017.
[3] M. Patel et al., “Mobile-edge computing introductory technical white paper,” White Paper, Mobile-edge Computing (MEC)
industry initiative, 2014.
29
[4] M. Chiang and T. Zhang, “Fog and IoT: An overview of research opportunities,” IEEE Internet Things J., vol. 3, no. 6,
pp. 854–864, 2016.
[5] S. Ulukus, A. Yener, E. Erkip, O. Simeone, M. Zorzi, P. Grover, and K. Huang, “Energy harvesting wireless communications: A review of recent advances,” IEEE J. Select. Areas Commun., vol. 33, no. 3, pp. 360–381, 2015.
[6] M. Lin, Z. Liu, A. Wierman, and L. L. Andrew, “Online algorithms for geographical load balancing,” in 2012 International
Green Computing Conference (IGCC). IEEE, 2012, pp. 1–10.
[7] M. Islam, S. Ren, G. Quan, M. Shakir, and A. Vasilakos, “Water-constrained geographic load balancing in data centers,”
IEEE Trans. Cloud Comput., vol. 5, no. 2, pp. 208–220, 2015.
[8] P. Mach and Z. Becvar, “Mobile edge computing: A survey on architecture and computation offloading,” IEEE Commun.
Surveys Tuts., 2017.
[9] L. Tong, Y. Li, and W. Gao, “A hierarchical edge cloud architecture for mobile computing,” in IEEE INFOCOM 2016.
IEEE, 2016, pp. 1–9.
[10] J. Xu, L. Chen, and S. Ren, “Online learning for offloading and autoscaling in energy harvesting mobile edge computing,”
IEEE Trans. Cogn. Commun. Netw., vol. 3, no. 3, pp. 361–373, Sept 2017.
[11] L. Chen and J. Xu, “Socially trusted collaborative edge computing in ultra dense networks,” in The 2nd ACM/IEEE
Symposium on Edge Computing.
ACM/IEEE, 2017, pp. 1–11.
[12] J. Oueis, E. C. Strinati, and S. Barbarossa, “The fog balancing: Load distribution for small cell cloud computing,” in IEEE
81st Vehicular Technology Conference (VTC Spring), May 2015, pp. 1–6.
[13] Y. Xiao and M. Krunz, “QoE and power efficiency tradeoff for fog computing networks with fog node cooperation,” in
IEEE INFOCOM 2017.
IEEE, 2017, pp. 1–9.
[14] A. Abdelnasser, E. Hossain, and D. I. Kim, “Clustering and resource allocation for dense Femtocells in a two-tier cellular
OFDMA network,” IEEE Trans. Wireless Commun., vol. 13, no. 3, pp. 1628–1641, March 2014.
[15] J. Rubio, A. Pascual-Iserte, J. del Olmo, and J. Vidal, “User association for load balancing in heterogeneous networks
powered with energy harvesting sources,” in 2014 IEEE Globecom Workshops (GC Wkshps), Dec 2014, pp. 1248–1253.
[16] H. H. M. Tam, H. D. Tuan, D. T. Ngo, T. Q. Duong, and H. V. Poor, “Joint load balancing and interference management
for small-cell heterogeneous networks with limited backhaul capacity,” IEEE Trans. Wireless Commun., vol. 16, no. 2, pp.
872–884, Feb 2017.
[17] J. Oueis, E. C. Strinati, S. Sardellitti, and S. Barbarossa, “Small cell clustering for efficient distributed fog computing: A
multi-user case,” in IEEE 82nd Vehicular Technology Conference (VTC2015-Fall), Sept 2015, pp. 1–5.
[18] S. Tanzil, O. Gharehshiran, and V. Krishnamurthy, “A distributed coalition game approach to Femto-Cloud formation,”
IEEE Trans. Cloud Comput., 2016.
[19] S. Guruacharya, D. Niyato, M. Bennis, and D. I. Kim, “Dynamic coalition formation for network MIMO in small cell
networks,” IEEE Trans. Wireless Commun., vol. 12, no. 10, pp. 5360–5372, October 2013.
[20] J. Gong, S. Zhou, and Z. Zhou, “Networked mimo with fractional joint transmission in energy harvesting systems,” IEEE
Transactions on Communications, vol. 64, no. 8, pp. 3323–3336, 2016.
[21] H. Xu, C. Feng, and B. Li, “Temperature aware workload managementin geo-distributed data centers,” IEEE Trans. Parallel
Distrib. Syst., vol. 26, no. 6, pp. 1743–1753, 2015.
[22] J. Luo, L. Rao, and X. Liu, “Spatio-temporal load balancing for energy cost optimization in distributed internet data
centers,” IEEE Trans. Cloud Comput., vol. 3, no. 3, pp. 387–397, July 2015.
[23] Z. Liu, M. Lin, A. Wierman, S. Low, and L. L. Andrew, “Greening geographical load balancing,” IEEE/ACM Trans.
Networking, vol. 23, no. 2, pp. 657–671, 2015.
30
[24] M. J. Neely and L. Huang, “Dynamic product assembly and inventory control for maximum profit,” in The 49th IEEE
Conference on Decision and Control (CDC). IEEE, 2010, pp. 2805–2812.
[25] G. Auer et al., “How much energy is needed to run a wireless network?” IEEE Wireless Commun., vol. 18, no. 5, pp.
40–49, October 2011.
[26] D. Lee, H. Seo, B. Clerckx, E. Hardouin, D. Mazzarese, S. Nagata, and K. Sayana, “Coordinated multipoint transmission
and reception in LTE-advanced: deployment scenarios and operational challenges,” IEEE Commun. Mag., vol. 50, no. 2,
pp. 148–155, February 2012.
[27] S. Lakshminarayana, T. Q. Quek, and H. V. Poor, “Cooperation and storage tradeoffs in power grids with renewable energy
resources,” IEEE J. Select. Areas Commun., vol. 32, no. 7, pp. 1386–1397, 2014.
[28] Y. Mao, J. Zhang, and K. B. Letaief, “A Lyapunov optimization approach for green cellular networks with hybrid energy
supplies,” IEEE J. Select. Areas Commun., vol. 33, no. 12, pp. 2463–2477, 2015.
[29] M. J. Neely, “Stochastic network optimization with application to communication and queueing systems,” Synthesis Lectures
on Communication Networks, vol. 3, no. 1, pp. 1–211, 2010.
[30] R. Urgaonkar, B. Urgaonkar, M. J. Neely, and A. Sivasubramaniam, “Optimal power cost management using stored energy
in data centers,” in Proceedings of the ACM SIGMETRICS joint international conference on Measurement and modeling
of computer systems.
ACM, 2011, pp. 221–232.
[31] K. Yuan, Q. Ling, W. Yin, and A. Ribeiro, “A linearized Bregman algorithm for decentralized basis pursuit,” in Proceedings
of the 21st European Signal Processing Conference (EUSIPCO). IEEE, 2013, pp. 1–5.
[32] M. P. Friedlander and P. Tseng, “Exact regularization of convex programs,” SIAM Journal on Optimization, vol. 18, no. 4,
pp. 1326–1350, 2007.
| 7cs.IT
|
Effective Classification of MicroRNA Precursors Using
Combinatorial Feature Mining and AdaBoost Algorithms
Ling Zhong and Jason T. L. Wang
Bioinformatics Program and Department of Computer Science, New Jersey Institute of
Technology, Newark, New Jersey 07102, USA
Abstract
MicroRNAs (miRNAs) are non-coding RNAs with approximately 22 nucleotides (nt) that
are derived from precursor molecules. These precursor molecules or pre-miRNAs often
fold into stem-loop hairpin structures. However, a large number of sequences with premiRNA-like hairpins can be found in genomes. It is a challenge to distinguish the real
pre-miRNAs from other hairpin sequences with similar stem-loops (referred to as pseudo
pre-miRNAs). Several computational methods have been developed to tackle this
challenge. In this paper we propose a new method, called MirID, for identifying and
classifying microRNA precursors. We collect 74 features from the sequences and
secondary structures of pre-miRNAs; some of these features are taken from our previous
studies on non-coding RNA prediction while others were suggested in the literature. We
develop a combinatorial feature mining algorithm to identify suitable feature sets. These
feature sets are then used to train support vector machines to obtain classification models,
based on which classifier ensemble is constructed. Finally we use an AdaBoost algorithm
to further enhance the accuracy of the classifier ensemble. Experimental results on a
variety of species demonstrate the good performance of the proposed method, and its
superiority over existing tools.
1
Introduction
MicroRNAs (miRNAs) are non-coding RNAs (ncRNAs) of approximately 22 nucleotides
that are known to regulate post-transcriptional expression of protein-coding genes (Bartel,
2004; Bindra et al., 2010). Lee et al. (1993) first reported that in C. elegans, lin-4
regulates the translation of lin-14 mRNA via an antisense RNA-RNA interaction. Since
then, many roles of miRNAs have been discovered, including for example the control of
left/right neuronal asymmetry in C. elegans (Johnston and Hobert, 2003), cell
proliferation in Drosophila (Brennecke et al., 2003), and the regulation of flowering time
in plants (Aukerman and Sakai, 2003). Their role in cancer development has also been
reported (Mack, 2007). However, these roles are only a small portion of total miRNA
functions (Bushati and Cohen, 2007). As a consequence, exploring miRNAs and their
functions continues to be a highly active area of research.
MiRNAs are derived from pre-miRNAs that often fold into stem-loop hairpin
structures. These characteristic stem-loop structures are highly conserved in different
species (Lai et al., 2003). One challenging research problem is to distinguish premiRNAs from other sequences with similar stem-loop structures (referred to as pseudo
pre-miRNAs). Many computational methods have been developed to tackle this
challenge. A common approach is to transform the classification of real and pseudo premiRNAs to a feature selection problem.
Lim et al. (2003) reported some characteristic features in phylogenetically
conserved stem loop pre-miRNAs. Lai et al. (2003) considered hairpin structures
predicted by mfold (Zuker, 2003) as well as the nucleotide divergence of pre-miRNAs.
Xue et al. (2005) decomposed stem-loop hairpin structures into local structure-sequence
2
features, and used these features in combination with a support vector machine to classify
pre-miRNAs. Bentwich et al. (2005) proposed a scoring function for pre-miRNAs with
thermodynamic stability and certain structural features, which capture the global
properties of the hairpin structures in the pre-miRNAs. Ng and Mishra (2007) employed a
Gaussian radial basis function kernel as a similarity measure for 29 global and intrinsic
hairpin folding attributes, and characterized pre-miRNAs based on their dinucleotide
subsequences, hairpin folding, non-linear statistical thermodynamics and topology.
Huang et al. (2007) evaluated features valuable for pre-miRNA classification, such as the
local secondary structure differences of the stem regions of real pre-miRNA and pseudo
pre-miRNA hairpins, and established correlations between different types of mutations
and the secondary structures of real pre-miRNAs. More recently, Zhao et al. (2010)
considered structure-sequence features and minimum free energy of RNA secondary
structure, along with the double helix structure with free nucleotides and base-pairing
features. In general, the quality of selected features directly affects the classification
accuracy achieved by a method.
In this paper, we present a novel combinatorial feature mining method for premiRNA classification. Our method, named MirID, identifies and classifies an input RNA
sequence as a pre-miRNA or not. MirID considers different combinations of features
extracted from pre-miRNAs. For each combination (or each set of features), we create a
support vector machine (SVM) model (Cortes and Vapnik, 1995; Fan et al., 2005) based
on that feature set. SVM models whose accuracies are above a user-determined threshold
are then used to build a classifier ensemble. This classifier ensemble will be refined
through several iterations until its accuracy cannot be enhanced further. We next
3
construct new feature sets based on the best feature sets obtained so far by performing
pairwise merge and split operations on the best feature sets. Then we repeat the above
procedure iteratively by building a SVM model based on each new feature set,
constructing a classifier ensemble from the SVM models whose accuracies are above the
newly computed threshold, and refining the ensemble until it can not be improved further.
Finally we output the best classifier ensemble obtained through this iterative procedure.
To further enhance the accuracy of the classifier ensemble, we apply a boosting algorithm
to the ensemble to obtain a strong classifier, which is used for pre-miRNA classification.
Materials and Methods
Datasets
We collected real pre-miRNAs and pseudo pre-miRNAs from twenty one species,
some of which were studied previously while others have not been explored. These RNA
sequences were evenly divided into training data and test data. Table 1 presents a
summary of the data. The first column of Table 1 shows a species or organism name. The
second column of Table 1 shows the number of training sequences followed by the
number of test sequences with respect to the organism’s real pre-miRNAs. The third
column of Table 1 shows the number of training sequences followed by the number of
test sequences with respect to the organism’s pseudo pre-miRNAs. As an example,
referring to Arabidopsis thaliana in Table 1, its training set contains 66 real pre-miRNAs
and 923 pseudo pre-miRNAs; its test set contains 67 real pre-miRNAs and 924 pseudo
pre-miRNAs.
4
The
real
pre-miRNAs
were
downloaded
from miRBase
available
at
http://www.mirbase.org/ (Kozomara and Griffiths-Jones, 2011). We used RNAfold
(Hofacker, 2003) to predict the secondary structures of all the RNA sequences. The
lengths of the real pre-miRNAs in the final dataset ranged from 60 to 120 nt. The pseudo
pre-miRNAs
used
in
this
study
were
collected
from
GenBank
(http://www.ncbi.nlm.nih.gov/genbank/). As in (Xue et al., 2005), we searched for the
protein-coding regions of the genome sequences of the twenty one species in Table 1, and
divided the regions into short sequences. The lengths of these short sequences were
randomly generated, ranging from 60 to 120 nt. The pseudo pre-miRNAs were chosen
from these short sequences. The criteria used in choosing the pseudo pre-miRNAs are: (i)
they have a stem-loop hairpin structure, (ii) they contain at least 18 base pairs, including
Watson-Crick and wobble base pairs, on the stem region of the stem-loop structure, and
(iii) their secondary structure has a maximum of -15 kcal/mol free energy without
multiple hairpin loops (Kozomara and Griffiths-Jones, 2011). These criteria ensure that
the secondary structures of the pseudo pre-miRNAs are similar to those of the real premiRNAs.
Feature pool
In designing our pre-miRNA classification method, we examined multiple
features extracted from a pre-miRNA sequence and its secondary structure. Some of
these features were taken from our previous studies on ncRNA prediction (Griesmer et al.,
2011; Wang and Wu, 2006) while others were suggested in the literature (Sewer et al.,
2005; Xue et al., 2005; Zheng et al., 2006). These features included the sequence length,
5
the number of base pairs, GC content, the number of nucleotides contained in the hairpin
loop (i.e., the loop size), the free energy of the sequence’s secondary structure obtained
from RNAfold (Hofacker, 2003), the number of bulge loops, and the size of the largest
bulge loop in the secondary structure.
In addition, we considered the features described in (Zheng et al., 2006). These
features included the difference of the lengths of the two tails in the secondary structure
where a tail represented the strand of unpaired bases in the 5’ or 3’ end of the structure,
the number of tails, and the length of the larger tail. Besides, several combined features
were considered. They included the ratio between the number of base pairs and the
sequence length, the length difference of two tails plus the larger tail length, the size of
the hairpin loop plus the larger tail length, the size of the hairpin loop plus the largest
bulge size, the ratio between the larger tail length and the sequence length, the ratio
between the size of the hairpin loop and the sequence length, the ratio between the largest
bulge size and the sequence length, the ratio between the largest bulge size and the
number of base pairs, the normalized free energy (Spirollari et al., 2009), which is the
minimum free energy of the sequence’s secondary structure divided by the sequence
length, and the ratio between the normalized free energy and the GC content.
The next set of features included the triplets of structure-sequence elements
described in (Xue et al., 2005). Here we used the dot-bracket notation (Hofacker, 2003)
to represent an RNA secondary structure. Figure 1 shows the sequence and structure of a
hypothetical pre-miRNA and its dot-bracket notation. A triplet is composed of three
contiguous structure elements (bases or base pairs) (Liu et al., 2005) that correspond to
three contiguous nucleotides along with the middle nucleotide. For example, consider the
6
first three dots (bases) and their corresponding nucleotides AAA in Figure 1. The middle
nucleotide is A. Thus, the structure-sequence elements “A...” constitute a triplet. As
another example, consider the first three brackets (base pairs) and their corresponding
nucleotides UUG in Figure 1. The middle nucleotide is U. Thus, the structure-sequence
elements “U(((” constitute a triplet. There are 32 triplets, and hence 32 such features in
total.
Finally we considered the symmetric and asymmetric loops defined in (Sewer et
al., 2005). We refer to the portion of the sequence from the 5’ end to the hairpin loop as
the left arm, and the portion of the sequence from the hairpin loop to the 3’ end as the
right arm. In a symmetric (internal) loop, the number of nucleotides in the left arm equals
the number of nucleotides in the right arm. In an asymmetric (internal) loop, the number
of nucleotides in the left arm is different from the number of nucleotides in the right arm.
Features related to these loops included the size of each loop, the average size of the
loops, and the average distance between the loops. Other features included the proportion
of A/C/G/U in the stem, and the proportion of A-U/C-G/G-U base pairs in the stem.
Totally, there are 74 features in the feature pool.
Combinatorial feature mining
MirID adopts a novel feature mining algorithm for pre-miRNA classification.
Initially the algorithm randomly generates N feature sets from the feature pool. (The
default value of N used in this study is 100.) Each feature set contains between 1 and 150
features, randomly chosen with replacement from the feature pool. Some features may
repeatedly occur in a feature set; thus a bagging approach is used here (Breiman, 1996).
7
Duplicate features have more weights than the other features in the feature set. The
numbers 1 and 150 are chosen, to ensure that there are enough feature sets containing
duplicate features. We then build a SVM model based on each feature set using training
sequences, and apply the classification model to test sequences to calculate the accuracy
of the model. The SVM used in this study is the LIBSVM package downloaded from
http://www.csie.ntu.edu.tw/~cjlin/libsvm/ (Fan et al., 2005). We use the polynomial
kernel provided in the LIBSVM package. The polynomial kernel achieves the best
performance among all kernel functions included in the package.
Then, we remove the SVM models whose accuracies are less than a userdetermined threshold t. (The default value of t used in this study is 0.8.) The feature sets
used to build those removed SVM models are also eliminated from further consideration.
We construct a classifier ensemble from the remaining SVM models. The ensemble
works by taking the majority vote from the individual SVM models used to build the
classifier ensemble. This ensemble will be refined through several iterations until its
accuracy cannot be enhanced further. In each iteration, the user-determined threshold t is
incremented by a step value, so that more accurate SVM models are used to construct a
(hopefully) better classifier ensemble in the next iteration. (The default value of step used
in this study is 0.005.)
It is likely that different combinations of remaining features may yield an even
better classifier. Our algorithm then performs pairwise merge and split operations on the
set Sb of feature sets used to build the best classifier ensemble obtained so far. In doing so,
MirID takes four steps: (1) picks each pair of feature sets s1 and s2 in Sb; (2) merges s1
and s2 into a single feature set s3 with, say p, features; (3) randomly generates a number q,
8
q < p; (4) randomly assigns q features in s3 into a set s’1 and assigns the remaining p - q
features into another set s’2. Thus, these four steps take two feature sets s1 and s2 in Sb as
input and produce two new feature sets s’1 and s’2 as output. Figure 2 illustrates how the
merge and split operations work on two feature sets.
These pairwise merge and split operations are applied to the feature sets used to
build the best classifier ensemble obtained so far, to generate new feature sets. The new
feature sets are then used to build new SVM models. Accurate new SVM models, whose
accuracies are greater than or equal to the newly computed threshold t, are then used to
build a new classifier ensemble. This procedure is repeated several times to obtain a best
classifier ensemble. Figure 3 summarizes our feature mining algorithm, whose output is
the best classifier ensemble along with the component SVM models (feature sets) used to
build the ensemble. Notice that in the feature mining algorithm in Figure 3, it is possible
that, after removing SVM models/feature sets with accuracy < t, there is no remaining
feature set, and hence Sr becomes empty. Under this circumstance, the classifier ensemble
constructed based on Sr is empty, and the accuracy of the classifier ensemble is 0.
Boosting
The performance of a classification algorithm can be further enhanced through
boosting. We apply AdaBoost (Bindewald and Shapiro, 2006; Freund and Schapire, 1997;
Schapire, 1999) to the classifier ensemble produced by our feature mining algorithm.
Specifically, we treat the classifier ensemble as a weak classifier and continue refining it
into a strong classifier through an iterative procedure. Let X be a set of sequences x1,
x2, . . . , xm where xi, 1 ≤ i ≤ m, is associated with a label yi such that
9
yi
1 if xi is a real pre - miRNA
1 if xi is a pseudo pre - miRNA
The AdaBoost algorithm works with K iterations. (The default value of K used in this
study is 20.) In iteration k, 1 ≤ k ≤ K, the algorithm updates a weight function Wk as
explained below, which will be used in selecting training sequences in iteration k + 1.
Initially, every sequence has an equal weight, i.e. W0(xi) = 1/m, 1 ≤ i ≤ m. In iteration k,
the algorithm samples 1/3 sequences with replacement from X based on the weight
function Wk-1 to form a training set Xk. The set Xk is then used to train a weak classifier Hk,
which classifies each sequence xi as either a real pre-miRNA or a pseudo pre-miRNA.
That is,
H k xi
1 H k classifies xi as a real pre - miRNA
1 H k classifies xi as a pseudo pre - miRNA
Let Ek = {xi|Hk(xi) ≠ yi}. The error rate εk of Hk is:
k
W x
xiEk
k 1
(1)
i
Let
1 1 k
k ln
2 k
(2)
The algorithm updates Wk for each sequence xi, 1 ≤ i ≤ m, as follows:
Wk 1 ( xi ) k
e
Zk
Wk xi
W (x )
k 1 i ek
Zk
if H k ( xi ) yi
if H k ( xi ) yi
Wk 1 ( xi )exp k y i H k xi
Zk
(3)
10
where Zk is a normalization factor chosen such that Wk is normally distributed. Thus, the
sequences causing classification errors in iteration k will have a greater probability of
being selected as training sequences for constructing the weak classifier Hk+1 in iteration
k+1. Using this technique, each weak classifier should have greater accuracy than its
predecessor. The final, strong classifier H combines the vote of each individual weak
classifier Hk, 1 ≤ k ≤ K, where the weight of each weak classifier’s vote is a function of
its accuracy. Specifically, for an unlabeled test sequence x, H(x) is calculated as follows:
K
H x sign k H k x
k 1
(4)
The function sign indicates that if the sum inside the parentheses is greater than or equal
to zero, then H classifies x as positive (i.e. a real pre-miRNA); otherwise H classifies x as
negative (i.e. a pseudo pre-miRNA).
Results
Performance analysis of the MirID method
We carried out a series of experiments to evaluate the proposed MirID method.
All the experiments were performed on a 2 GHz Pentium 4 PC having a memory of 2G
bytes. The operating system was Cygwin on Windows XP and the algorithms were
implemented in Perl. To understand the effect of boosting, we also considered using the
combinatorial feature mining algorithm alone to classify pre-miRNAs, and referred to it
as the CFM method. The performance measure used here is accuracy, defined as follows.
A method is said to classify a test sequence correctly if the sequence is a real pre-miRNA
(pseudo pre-miRNA, respectively) and the method indicates that the sequence is indeed a
real pre-miRNA (pseudo pre-miRNA, respectively). A method is said to classify a test
11
sequence incorrectly if the sequence is a real pre-miRNA (pseudo pre-miRNA,
respectively) but the method mistakenly indicates that the sequence is a pseudo premiRNA (real pre-miRNA, respectively). For each species, the accuracy of a method is
defined as the number of correctly classified test sequences of that species divided by the
total number of test sequences of that species.
We first evaluated how the number of initial feature sets, N, affects the
performance of CFM and MirID. As N increases, more feature sets are generated initially.
This allows the feature mining algorithm to construct a classifier ensemble using more
diverse feature sets, and hence the accuracy of the classifier ensemble increases. On the
other hand, as N increases, the inner loop in Figure 3 is run more times; as a consequence,
the running time increases. MirID requires more time than CFM, due to the extra time
spent in boosting. MirID in general is more accurate than CFM, indicating the benefit of
including the boosting algorithm.
We next evaluated how the threshold, t, used in the feature mining algorithm
affects the performance of CFM and MirID. When t is very large (e.g. t > 0.95), the
accuracies of the methods drop sharply. This happens because the accuracies of most
SVM models are less than 0.95, and hence these SVM models are eliminated from further
consideration early in the feature mining algorithm, cf. Figure 3. When t approaches 1, it
is likely that the set Sb returned by the feature mining algorithm is an empty set, and
therefore the classifier ensemble constructed based on Sb is also empty, yielding an
accuracy of 0. As t increases, fewer feature sets qualify and the set Sr is smaller. As a
result, the inner loop in Figure 3 is executed fewer times, and hence the running time
decreases.
12
Then we evaluated how the value, step, used to increment the threshold t in each
iteration of the inner loop in Figure 3 affects the performance of CFM and MirID. With
the default values of N and t used in this study, the feature mining algorithm is able to
produce a classifier ensemble with high accuracy. The value of step has little impact on
the accuracies of the proposed methods. However, as step increases, fewer iterations of
the inner loop in Figure 3 are executed, and as a consequence, the running time decreases.
We also conducted experiments to test different numbers of iterations, K, in the
boosting algorithm. It was found that when K is sufficiently large (e.g. K ≥ 20), the
behavior of the boosting algorithm becomes stable, with the accuracy approaching 1. On
the other hand, when K is large, more running time will be needed.
Finally we compared CFM and MirID with two closely related methods, PMirP
(Zhao et al., 2010) and TripletSVM (Xue et al., 2005). Like our methods, both PMirP and
TripletSVM were implemented using support vector machines. PMirP adopted a hybrid
coding scheme, combining features such as free bases, base pairs, minimum free energy
of secondary structure, among others. TripletSVM used triplets of structure-sequence
elements, which also were included in our feature pool. Table 2 shows the accuracies of
the four methods on twelve species taken from Table 1. These twelve species were used
to pre-train PMirP and TripletSVM, and available from the tools. For each species, the
highest accuracy yielded by a tool is in bold. It can be seen from Table 2 that MirID is
better than or as good as the existing tools on all the species except Gallus gallus and
Oryza sativa. For Gallus gallus and Oryza sativa, PMirP achieves higher accuracies.
13
Software Tool
We have implemented MirID using Perl into a software tool, available from the
authors upon request. The software tool accepts a test sequence as input and classifies the
test sequence as a pre-miRNA or not. We pre-train our software tool using the training
sequences given in Table 1. In addition to the twelve species available from the PMirP
and TripletSVM web servers (Xue et al., 2005; Zhao et al., 2010), we pre-train our
software tool using nine additional species (shown in Table 1 but not in Table 2). Our
tool achieves high accuracies on these nine species, as shown in Table 3. (The PMirP and
TripletSVM web servers were not pre-trained on these nine species, and hence we only
show the results for CFM and MirID here.) MirID is more accurate than CFM, due to the
boosting algorithm.
Table 4 shows, for each species in Table 1, the number of feature sets produced
by our feature mining algorithm. Table 5 shows the CPU time (in seconds) spent in pretraining the MirID tool. The training time depends on the number of feature sets, the
number of features in each feature set, the number of iterations used by the feature
mining algorithm, and the number of iterations used in the boosting algorithm. Notice
that this training is done once, and no more training is needed on the test data. It takes
less than a second to classify an unlabeled test sequence.
Conclusions
In this paper we present a new method (MirID) for pre-miRNA classification.
Empirical results showed that MirID outperforms two closely related methods, PMirP
and TripletSVM, on the majority of species tested in the experiments. Since all the three
14
methods were implemented using support vector machines with similar features, we
conclude that the superiority of our method is due to its feature mining and boosting
algorithms. In the future, we plan to extend these algorithms for classifying and
predicting other RNA structures; see, e.g. (Laing et al., 2012).
15
References
Aukerman, N.J., and Sakai, H. (2003). Regulation of flowering time and floral organ
identity by a microRNA and its APETALA2-like target genes. Plant Cell 15, 2730-2741.
Bartel, D.P. (2004). MicroRNAs: genomics, biogenesis, mechanism, and function. Cell
116, 281-297.
Bentwich, I., Avniel, A., Karov, Y., et al. (2005). Identification of hundreds of conserved
and nonconserved human microRNAs. Nat Genet 37, 766-770.
Bindewald, E., and Shapiro, B.A. (2006). RNA secondary structure prediction from
sequence alignments using a network of k-nearest neighbor classifiers. RNA 12, 342-352.
Bindra, R.S., Wang, J.T.L., and Bagga, P.S. (2010). Bioinformatics methods for studying
microRNA and ARE-mediated regulation of post-transcriptional gene expression.
International Journal of Knowledge Discovery in Bioinformatics 1, 97-112.
Breiman, L. (1996). Bagging predictors. Machine Learning 24, 123-140.
Brennecke, J., Hipfner, D.R., Stark, A., Russell, R.B., and Cohen, S.M. (2003). bantam
encodes a developmentally regulated microRNA that controls cell proliferation and
regulates the proapoptotic gene hid in Drosophila. Cell 113, 25-36.
Bushati, N., and Cohen, S.M. (2007). MicroRNA functions. Annu Rev Cell Dev Biol 23,
175-205.
Cortes, C., and Vapnik, V. (1995). Support-vector networks. Machine Learning 20, 273297.
Fan, R., Chen, P., and Lin, C. (2005). Working set selection using the second order
information for training SVM. Journal of Machine Learning Research 6, 1889-1918.
Freund, Y., and Schapire, R.E. (1997) A decision-theoretic generalization of on-line
learning and an application to boosting. Journal of Computer and System Sciences 55,
119-139.
Griesmer, S.J., Cervantes-Cervantes, M., Song, Y., and Wang, J.T.L. (2011). In silico
prediction of noncoding RNAs using supervised learning and feature ranking methods.
International Journal of Bioinformatics Research and Applications 7, 355-375.
Hofacker, I.L. (2003). Vienna RNA secondary structure server. Nucleic Acids Res 31,
3429-3431.
16
Huang, T.H., Fan, B., Rothschild, M.F., Hu, Z.L., Li, K., and Zhao, S.H. (2007).
MiRFinder: an improved approach and software implementation for genome-wide fast
microRNA precursor scans. BMC Bioinformatics 8, 341.
Johnston, R.J., and Hobert, O. (2003). A microRNA controlling left/right neuronal
asymmetry in Caenorhabditis elegans. Nature 426, 845-849.
Kozomara, A., and Griffiths-Jones, S. (2011). miRBase: integrating microRNA
annotation and deep sequencing data. Nucleic Acids Res 39, D152-D157.
Lai, E.C., Tomancak, P., Williams, R.W., and Rubin, G.M. (2003). Computational
identification of Drosophila microRNA genes. Genome Biol 4, R42.
Laing, C., Wen, D., Wang, J.T.L., and Schlick, T. (2012). Predicting coaxial helical
stacking in RNA junctions. Nucleic Acids Res 40, 487-498.
Lee, R.C., Feinbaum, R.L., and Ambros, V. (1993). The C. elegans heterochronic gene
lin-4 encodes small RNAs with antisense complementarity to lin-14. Cell 75, 843-854.
Lim, L.P., Glasner, M.E., Yekta, S., Burge, C.B., and Bartel, D.P. (2003). Vertebrate
microRNA genes. Science 299, 1540.
Liu, J., Wang, J.T.L., Hu, J., and Tian, B. (2005). A method for aligning RNA secondary
structures and its application to RNA motif detection. BMC Bioinformatics 6, 89.
Mack, G.S. (2007). MicroRNA gets down to business. Nat Biotechnol 25, 631-638.
Ng, K.L., and Mishra, S.K. (2007). De novo SVM classification of precursor microRNAs
from genomic pseudo hairpins using global and intrinsic folding measures.
Bioinformatics 23, 1321-1330.
Schapire, R.E. (1999). A brief introduction to boosting. In Proceedings of the Sixteenth
International Joint Conference on Artificial Intelligence, 1401-1406.
Sewer, A., Paul, N., Landgraf, P., et al. (2005). Identification of clustered microRNAs
using an ab initio prediction method. BMC Bioinformatics 6, 267.
Spirollari, J., Wang, J.T.L., Zhang, K., Bellofatto, V., Park, Y., and Shapiro, B.A. (2009).
Predicting consensus structures for RNA alignments via pseudo-energy minimization.
Bioinformatics and Biology Insights 3, 51-69.
Wang, J.T.L., and Wu, X. (2006). Kernel design for RNA classification using support
vector machines. International Journal of Data Mining and Bioinformatics 1, 57-76.
17
Xue, C., Li, F., He, T., Liu, G.P., Li, Y., and Zhang, X. (2005). Classification of real and
pseudo microRNA precursors using local structure-sequence features and support vector
machine. BMC Bioinformatics 6, 310.
Zhao, D., Wang, Y., Luo, D., et al. (2010). PMirP: a pre-microRNA prediction method
based on structure-sequence hybrid features. Artificial Intelligence in Medicine 49, 127132.
Zheng, Y., Hsu, W., Lee, M.L., and Wong, L.S. (2006). Exploring essential attributes for
detecting microRNA precursors from background sequences. Lecture Notes in Computer
Science, Springer 4316, 131-145.
Zuker, M. (2003). Mfold web server for nucleic acid folding and hybridization prediction.
Nucleic Acids Res 31, 3406-3415.
18
TABLE 1. Summary of Datasets
Species
Arabidopsis thaliana
Caenorhabditis briggsae
Caenorhabditis elegans
Canis familiaris
Ciona intestinalis
Danio rerio
Drosophila melanogaster
Drosophila pseudoobscura
Epstein barr virus
Gallus gallus
Homo sapiens
Macaca mulatta
Medicago truncatula
Mus musculus
Oryza sativa
Physcomitrella patens
Populus trichocarpa
Pristionchus pacificus
Rattus norvegicus
Schmidtea mediterranea
Taeniopygia guttata
Real pre-miRNA
66, 67
66, 67
84, 85
161, 161
160, 160
170, 170
81, 82
98, 99
12, 13
241, 241
504, 504
222, 223
111, 111
315, 315
172, 172
73, 74
94, 95
60, 61
193, 193
72, 73
94, 95
Pseudo pre-miRNA
923, 924
437, 438
595, 596
904, 905
733, 734
1071, 1072
694, 694
495, 495
119, 119
1186, 1186
1999, 2000
1086, 1086
116, 116
2019, 2019
522, 523
703, 704
809, 810
58, 58
1238, 1238
201, 202
483, 483
19
TABLE 2. Comparison of Four Pre-miRNA Classification Methods
Species
Arabidopsis thaliana
Caenorhabditis briggsae
Caenorhabditis elegans
Danio rerio
Drosophila melanogaster
Drosophila pseudoobscura
Epstein barr virus
Gallus gallus
Homo sapiens
Mus musculus
Oryza sativa
Rattus norvegicus
TripletSVM
0.92
0.96
0.86
0.67
0.92
0.90
1.00
0.85
0.93
0.94
0.95
0.80
PMirP
0.96
0.97
0.86
0.83
0.96
0.92
0.80
1.00
0.95
0.94
1.00
0.92
CFM
0.99
0.98
0.97
0.98
0.97
0.98
0.98
0.96
0.93
0.95
0.97
0.97
MirID
1.00
1.00
0.98
0.99
0.99
1.00
1.00
0.96
0.95
0.97
0.99
0.98
20
TABLE 3. Accuracies of CFM and MirID on Nine Additional Species
Species
Canis familiaris
Ciona intestinalis
Macaca mulatta
Medicago truncatula
Physcomitrella patens
Populus trichocarpa
Pristionchus pacificus
Schmidtea mediterranea
Taeniopygia guttata
CFM
0.97
0.94
0.96
0.95
1.00
0.97
0.96
0.95
0.95
MirID
1.00
1.00
0.96
1.00
1.00
0.99
1.00
0.99
0.99
21
TABLE 4. Number of Feature Sets for Each Species in MirID
Species
Arabidopsis thaliana
Caenorhabditis briggsae
Caenorhabditis elegans
Canis familiaris
Ciona intestinalis
Danio rerio
Drosophila melanogaster
Drosophila pseudoobscura
Epstein barr virus
Gallus gallus
Homo sapiens
Macaca mulatta
Medicago truncatula
Mus musculus
Oryza sativa
Physcomitrella patens
Populus trichocarpa
Pristionchus pacificus
Rattus norvegicus
Schmidtea mediterranea
Taeniopygia guttata
Number of feature sets
1
6
1
1
7
11
3
4
5
3
1
1
1
3
3
1
1
1
10
32
5
22
TABLE 5. Training Time for Each Species in MirID
Species
Arabidopsis thaliana
Caenorhabditis briggsae
Caenorhabditis elegans
Canis familiaris
Ciona intestinalis
Danio rerio
Drosophila melanogaster
Drosophila pseudoobscura
Epstein barr virus
Gallus gallus
Homo sapiens
Macaca mulatta
Medicago truncatula
Mus musculus
Oryza sativa
Physcomitrella patens
Populus trichocarpa
Pristionchus pacificus
Rattus norvegicus
Schmidtea mediterranea
Taeniopygia guttata
Training time
80
348
103
153
269
1272
199
196
113
274
1530
243
104
786
214
90
138
63
349
478
156
23
FIG. 1. Sequence and structure of a hypothetical pre-miRNA and its dot-bracket notation.
24
FIG. 2. Illustration of the merge and split operations on two feature sets.
25
FIG. 3. Algorithm for combinatorial feature mining.
26
| 5cs.CE
|
Learnable Explicit Density for
Continuous Latent Space and Variational Inference
Chin-Wei Huang 1 Ahmed Touati 1 Laurent Dinh 1 Michal Drozdzal 1 2 Mohammad Havaei 1 2
Laurent Charlin 1 3 Aaron Courville 1 4
arXiv:1710.02248v1 [cs.LG] 6 Oct 2017
Abstract
In this paper, we study two aspects of the variational autoencoder (VAE): the prior distribution
over the latent variables and its corresponding
posterior. First, we decompose the learning of
VAEs into layerwise density estimation, and argue that having a flexible prior is beneficial to
both sample generation and inference. Second,
we analyze the family of inverse autoregressive
flows (inverse AF) and show that with further improvement, inverse AF could be used as universal approximation to any complicated posterior.
Our analysis results in a unified approach to parameterizing a VAE, without the need to restrict
ourselves to use factorial Gaussians in the latent
real space.
1. Introduction
Deep Gaussian Latent Models (Rezende et al., 2014),
also known as Variational Autoencoders (VAEs) (Kingma
& Welling, 2014), fall within the paradigm of Maximum
Likelihood Estimate (MLE) and are often applied in computer vision problems. However, training with MLE usually leads to overestimation of the entropy of the data distribution (Minka, 2005). This is an undesirable property, as
natural images are usually assumed to lie within a lower dimensional manifold, and the additional entropy (and other
simplifying modeling assumptions for the purpose of explicit density estimation) often leads to a marginal likelihood with probability mass spread out in the data space
where there is no support in the training data, which causes
the blurriness of samples. These observations motivate the
design of more flexible, complex families of model densities.
Since a continuous latent variable z is introduced to the
model, VAEs can
R be interpreted as an infinite mixture
model p(x) = z p(x|z)p(z)dz where the parameters of
1
MILA, Université de Montréal, Canada 2 Imagia Inc., Canada
HEC Montréal, Canada 4 CIFAR Fellow, Canada. Correspondence to: Chin-Wei Huang <[email protected]>.
3
the class conditional distribution p(x|z) are functions of
the latent variable z (which is thought of as class here),
and there are infinitely many classes. Such models should
theoretically have enough flexibility to capture highly complex distributions such as image manifolds, but in practice
it is found to be overshadowed by tractable density models
such as autoregressive models (Van Den Oord et al., 2016),
or Generative Adversarial Networks (GANs) (Goodfellow
et al., 2014) in terms of sample generation quality.
It is believed that the relative poor performance in sample quality lies in the fact that the introduction of a latent representation requires approximate inference, as the
model distribution is biased by simplifying posterior densities (Buntine & Jakulin, 2004); i.e. training is achieved
by maximizing the variational lower bound on the marginal
log likelihood:
pθ (x|z)pπ (z)
(1)
L(θ, φ, π; x) = Eqφ (z|x) log
qφ (z|x)
where subscripts θ, φ, π denote the parameters of the associated distributions.
We discuss two aspects of training with the bound. First,
maximizing (1) with respect to φ amounts to minimizing
KL(qφ (z|x)||p(z|x)); the variational distribution, q(z|x),
can thus be viewed as an approximate to the true posterior,
p(z|x). Simplifying q(z|x) (e.g. by using a factorial Gaussian as a common practice) is problematic, as the marginal
log likelihood of interest log p(x) can only be optimized to
the extent we are able to approximate the true posterior using the variational distribution. This motivates a direct improvement of variational inference (Rezende & Mohamed,
2015; Ranganath et al., 2015; Kingma et al., 2016).
Second, during training of the VAE, only a part of the latent
space is explored. When marginalizing
R out the input vector
x, we recover the marginal q(z) = x q(z|x)pD (x), where
D indicates the true data distribution. When the marginal
approximate posterior fails to fill up the prior as the priorcontractive term requires, one would risk sampling from
untrained regions in the latent space. A direct and nonparametric treatment of sampling from such regions of the
prior would be to take q(z) as the prior, but the integral is
intractable and the data distribution is only partially speci-
Learnable Explicit Density for Continuous Latent Space and Variational Inference
q(z|x)
p(z|x)
p(z)
Figure 1. Effect of prior on posterior. Matching the prior p(z)
with the marginal approximate posterior Ex [q(z|x)] makes the
true posterior p(z|x) easier to model, since it pushes the true posterior to be closer to the approximate posterior.
fied by a limited training data. Even if we take the empirical distribution of pD (x), we would have a mixture model
of up to n components, where n is the number of training data points, which would be impractical given the scale
of modern machine learning tasks. A workaround of this
problem is to take a random subset of D, or introduce a
learnable set
pseudo-data of size K, and set the prior to
Pof
K
1
be p(z) = j=1 K
q(zj |xj ), which is shown to be promising in the recent work done by Tomczak & Welling (2017).
Another approach is to directly regularize the autoencoder
by matching the aggregated posterior with the prior, as in
Makhzani et al. (2015).
In this paper, we make two main contributions. First, we
analyze the effect of making the prior learnable. We show
that training with the variational lower bound under some
limit conditions matches the marginal approximate posterior with the prior, which is desirable from the generative model point of view. We then decompose the lower
bound, and show that updating the prior alone brings the
prior closer to the marginal approximate posterior, suggesting that having the prior trainable is beneficial to both sample generation and inference. Our second contribution is
to prove that by using the family of inverse AF (Kingma
et al., 2016), one can universally approximate any posterior. This theoretically justifies the use of inverse AF to
improve variational inference. We unified the two aspects
and propose to use invertible functionals Dinh et al. (2016)
and Kingma et al. (2016) to parameterize explicit densities
for both the prior and approximate posterior.
2. Marginal Matching Prior
We claim that maximizing the variational lower bound explicitly matches the marginal q(z) with the prior p(z). By
decomposing the lower bound, we then suggest using a
learnable prior to improve sampling, i.e. to have a prior
that matches the marginal q(z) instead.
Let us define encoding and decoding distributions as q(z|x)
and p(x|z) respectively, a prior as p(z) and a data distribution as pD (x). Our goal is to train an
R auto-encoder as
a generative model by keeping q(z) = x pD (x)q(z|x)dx
close to the prior. This can be achieved at the limits of the
following two conditions (Hoffman & Johnson, 2016):
1. q(z|x) → p(z|x) ∀x ∼ pD (x)
2. p(x) → pD (x)
In words, given a perfect approximate posterior q(z|x) of
p(z|x) and a perfect marginal likelihood p(x) of pD (x), we
have the marginal q(z) converge to the prior, i.e.
Z
q(z) =
pD (x)qφ (z|x)dx
x
Z
1.
→
pD (x)pθ,π (z|x)dx
(2)
Zx
2.
→
pθ,π (x)pθ,π (z|x)dx = pπ (z)
x
That is, to have q(z) → p(z), we need to ensure the two
conditions are satisfied. We can cast it as an optimization
problem by minimizing the KL-divergences:
min EpD (x) [KL(q(z|x)||p(z|x))] + KL(pD (x)||p(x))
= max Ex∼pD (x) [L(θ, φ, π; x)]
(3)
The equality is a direct result of rearrangement of terms.
What (3) implies is that maximizing the variational lower
bound brings us to the limit conditions under which
marginal approximate posterior q(z) should match the prior
given enough flexibility in the assumed form of densities.
Now if we maximize (3) with respect to π while holding θ
and φ fixed like doing coordinate ascent, the samples drawn
from the doubly stochastic process x ∼ pD (x), z ∼ q(z|x)
can be thought of as a projected data distribution that we
want to model using the prior distribution:
max E[L] = min KL(Ex∼pD (x) [q(z|x)]||pπ (z))
π
π
(4)
As a result, having a learnable prior allows us to sample
from the marginal approximate posterior if the above divergence metric goes to zero.
Another advantage of a learnable prior can be visualized
by the cartoon plot in Figure 1. When we fix the approximate posterior and update the prior such that it becomes closer to the marginal approximate posterior, it concentrates the probability mass in such a way that the true
posterior becomes closer to the approximate posterior, as
p(z|x) ∝ p(z). In other words, the region of high posterior density not covered by the approximate posterior will
be reduced, which effectively means our proposal as variational distribution could be improved by having a better
prior which simplifies the true posterior.
Learnable Explicit Density for Continuous Latent Space and Variational Inference
3. Inverse Autoregressive Flows as Universal
Posterior Approximator
In Kingma et al. (2016), a powerful family of invertible
functions called the Inverse Autoregressive Flows (inverse
AF or IAF) were introduced, to improve variational inference. It is thus of practical and fundamental importance
to understand the benefits of using inverse AF and how to
improve them.
In this section, we show that normalizing flows from a base
distribution (such as uniform distribution) under autoregressive assumptions are universal approximators of any
density (as suggested in Goodfellow (2017)), given enough
capacity when a neural network is used to parameterize
non-linear dependencies.
Lemma 1. Existence of solution to a nonlinear independent component analysis problem. Given a random vector
X = (Xi )i=1...m ∈ Rm , there always exists a mapping g
from Rm to Rm such that the components of the random
vector Y = f (X) are statistically independent.
Proof. See Hyvarjnen & Pajunen (1998) for the full proof.
Here we point out that the transformation g used in the
proof falls within the family of autoregressive functions:
f = (fi )i=1...m where yi = fi (xi , y1 , . . . , yi−1 ) = P (x ≤
xi | y1 , ..., yi−1 ), for i = 1 . . . m. fi is the conditional CDF
and Y ∼ U nif ([0, 1]m ). Then any distribution of a random variable x can be warped into an independent distribution via the CDFs, specifically by a kind of Gram-Schmidt
process-like construction.
Proposition 1. Inverse autoregressive transformation as
universal approximator of any density. Let X be a random vector in an open set U ⊂ Rm . We assume that
X has a positive and continuous probability density distribution. There exists a sequence of mappings (Gn )n≥0
from (0, 1)m to Rm parametrized by autoregressive neural networks such that the sequence Xn = Gn (Y ) where
Y ∼ U nif ((0, 1)m ) converges in distribution to X.
Proof. We consider the mapping f defined in the proof of
Lemma 1. As f is autoregressive, the Jacobian of f is an
upper triangular matrix whose diagonal entries are equal
to the conditional densities which are positive by assumption. The determinant of the Jacobian, which is equal to
the product of diagonal entries, is positive. By the inverse
function theorem, f is locally invertible. As f is also injective (as follows from the bijectivity of CDF), f is globally
invertible and let g denotes its inverse. g is an autoregressive function and by the universal approximation theorem
(Cybenkot, 1989), we know that there exists a sequence of
mappings (Gn )n≥0 from (0, 1)m to Rm parametrized by
autoregressive neural networks that converge uniformly to
g. Let Xn = Gn (Y ) where Y ∼ U nif ((0, 1)m ). Let h
be a real-valued bounded continuous function on Rm . The
latter uniform convergence implies that since Gn converge
pointwise to g, then by continuity of h, h ◦ Gn converges
pointwise to h ◦ g. As h is bounded, the dominated convergence theorem gives that E[h(Xn )] = E[h(Gn (Y ))] converges to E[h(g(Y ))] = E[h(X)]. As the latter statement
is valid for all bounded continuous function h, Xn converge
to X in distribution.
Note that G is usually parameterized as an invertible function, at the expense of flexibility, to have a tractable Jacobian. Special designs of such a function, other than affine
transformation (Kingma et al., 2016), could be made to
improve the flow; otherwise one would need to compose
multiple layers of transformations to have a richer distribution family. Our proof shows that, with careful designs of
approximate posteriors, VAEs could have asymptotic consistency.
4. Proposed Method
As suggested in sections 2 and 3, we propose to use oneto-one correspondence to define a learnable explicit density (LED) model for both inference and sample generation. First, inspired by (4), we found that updating the prior
alone is reminiscent of MLE. One can think of data points
projected onto the latent space via Monte Carlo sampling as
a data distribution qD (z) = EpD (x) [q(z|x)] in space z. A
unimodal prior tends to overestimate the entropy of qD (z).
A powerful family of real non-volume preserving (Real
NVP) transformations (Dinh et al., 2016) can be applied
to real variables. It is thus natural to incorporate Real NVP
into VAEs to jointly train an explicit density model as prior.
We define the prior (and also the approximate posterior)
∂h
(z0 )|−1
with change of variable formula: p(z) = p(z0 )| ∂z
0
where h : z0 → z. To compute the density of the projected
data distribution, we inversely (h−1 ) transform the samples
z ∼ qD (z) into the base variable z0 with tractable density
(Dinh et al., 2014). We define the posterior likewise, as in
Rezende & Mohamed (2015), with g : z 0 → z. Objective
(1) can thus be modified as
L = Eq(z0 |x) [log p(x|g(z 0 ))]+
∂h−1
−1
0
0
Eq(z0 |x) log p(h ◦ g(z )) + log
(g(z )) −
∂z
∂g 0
Eq(z0 |x) log q(z 0 |x) − log
(z )
∂z 0
(5)
For permutation invariant latent variables, h is implemented with random masks. For latent variables that preserve the spatial correlation when a convolutional network
is used, we choose to use a checkerboard style mask (Dinh
Learnable Explicit Density for Continuous Latent Space and Variational Inference
pD (x)
4
x1
2
−4
4
p(x)
2
2
0
0
−2
−2
−4
−2
0
z1
2
0
−2
x1
2
4
−4
−3
−3
4
p(z)
−2
−1
0
z0,1
1
2
3
p(z0)
3
2
1
0
−1
−2
−4
ResConv
0
−1
4
z2
x2
0
z0,2
0
−2
4
−4
1
−2
−4
Table 2. Effect of increasing both prior and posterior complexity.
2
z0,2
0
−2
E[q(z0|x)]
3
2
z2
x2
2
−4
E[q(z|x)]
4
−2
−4
−2
0
z1
2
4
−3
−3
−2
−1
0
z0,1
1
2
Table 1. Effect of increasing prior complexity. Lpost : number of
MADE layers used for posterior. Two hidden layers of 512 nodes
were used for each layer of transformation. Lprior : number of NVP
layers used for prior. One hidden layer of 100 nodes was used
for each layer of transformation. For multi-layer perceptron, two
hidden layers with 200 nodes were used and the dimension of the
latent variable is 50. Rectifier is used as non-linear activation. For
Residual ConvNet, we have 3 layers of residual strided convolution (He et al., 2015) with [16,32,32] feature maps, using filter
of size 3×3. Before the stochastic layer a hidden layer of 450
nodes is used. The dimension of the latent variable is 32. We use
exponential linear units (Clevert et al., 2015) as non-linearity.
MLP
ResConv
Lpost
NLL
Lprior
NLL
Lprior
NLL
0
4
8
12
90.78
88.89
88.71
88.70
0
4
8
12
90.78
88.07
87.47
86.59
0
4
8
12
83.11
81.87
81.70
81.44
Lpost
NLL
4 NVP
8 NVP
8 NVP
16 NVP
4 NVP
8 NVP
8 MADE
16 MADE
81.81
81.55
80.81
80.60
3
Figure 2. Fitting a Gaussian mixture distribution. E[·] indicates
marginalization over the data x ∼ pD (x). Clockwise from top
left: projection of data distribution pD (x) onto the prior space
E(q(z|x)), and the base distribution space E(q(z0 |x)); density
maps of the base distribution p(z0 ), the transformed prior p(z)
and and marginal model distribution p(x).
MLP
Lprior
et al., 2016; Agrawal & Dukkipati, 2016). Interestingly,
sampling of such models is similar to block Gibbs sampling
for energy based models (e.g. Ising models) that define the
correlation between adjacent pixels.
Second, for the posterior distribution, we construct g by
inverse AF, which is parallelizable when combined with
MADE (Germain et al., 2015) or PixelCNN (Van Den
Oord et al., 2016). In fact, inverse AF can be thought of as
a generalization of Real NVP, as the Jacobian of the masked
operation used in Real NVP is upper triangular.
5. Experiments
Mixture of Bivariate Gaussians. We experiment on a
Gaussian mixture toy example, and visualize the effect of
having a learnable prior in Figure 2. During training, we
observe that models with flexible prior are easier to train
than models with flexible posterior. Our first conjecture is
that to refine the posterior density, we only draw one sample of z for each data point x, whereas refining the prior
density can be viewed as modeling the projected data distribution and thus depends on as many samples as there are in
the training set. Second, it might be due to the kind of transformation and the distance metrics that are used. To learn
the posterior, we implicitly minimize KL(q(z|x)||p(z|x)),
which is zero forcing since samples in region that has low
target density are heavily penalized. If q begins with a
sharper shape, it pays a high penalty by expansion to move
to another mode. It is thus easy for the distribution to be
stuck in local minima if the true posterior is multimodal,
while learning the prior does not have this mode seeking
problem since the forward KL in (4) is zero avoiding.
MNIST. We also tested our proposed method on binarized
MNIST (Larochelle & Murray, 2011), and report the estimated negative log likelihood as an evaluation metric.
We compare the effects of adding more invertible transformation layers on either the prior or posterior (see Table 1),
or both (Table 2). From Table 1, we see that models having a flexible prior easily outperform models with a flexible posterior. Likelihood of a model with flexible prior can
be further improved by using expressive posterior (Table 2)
such as real NVP (81.70 → 81.55), or with MADE to introduce more autoregressive dependencies (81.55 → 80.81).
6. Discussion and Future Work
In this paper, we first reinterpret training with the variational lower bound as layer-wise density estimation. Treating the Monte Carlo samples from the approximate posterior distributions as projected data distribution suggests
using a flexible prior to avoid overestimate of entropy. We
leave experiments on larger datasets and sample generation
as future work. Second, we showed that parameterizing
inverse AF using neural networks allows us to universally
approximate any posterior, which theoretically justifies the
use of inverse AF. Our proof also implies using affine coupling law to autoregressively warp the distribution is limited. It is thus possible to consider designs of more flexible
invertible functions to improve approximate posterior.
Learnable Explicit Density for Continuous Latent Space and Variational Inference
Acknowledgements
We thank NVIDIA for donating a DGX-1 computer used
in this work.
References
Agrawal, Siddharth and Dukkipati, Ambedkar. Deep Variational Inference Without Pixel-Wise Reconstruction.
ArXiv e-prints, November 2016.
Buntine, Wray and Jakulin, Aleks. Applying discrete pca
in data analysis. In Proceedings of the 20th Conference
on Uncertainty in Artificial Intelligence, UAI ’04, pp.
59–66, Arlington, Virginia, United States, 2004. AUAI
Press. ISBN 0-9749039-0-6.
Clevert, Djork-Arné, Unterthiner, Thomas, and Hochreiter,
Sepp. Fast and accurate deep network learning by exponential linear units (elus). CoRR, abs/1511.07289, 2015.
Cybenkot, G. Mathematics of control, signals, and systems
approximation by superpositions of a sigmoidal function. Math. Control Signals Systems, 2:303–314, 1989.
Dinh, Laurent, Krueger, David, and Bengio, Yoshua.
NICE: Non-linear Independent Components Estimation.
International Conference on Learning Representation, 1
(2):1–12, 2014.
Dinh, Laurent, Sohl-Dickstein, Jascha, and Bengio, Samy.
Density estimation using real NVP. International Conference on Learning Representation, abs/1605.08803,
2016.
Germain, Mathieu, Gregor, Karol, Murray, Iain, and
Larochelle, Hugo. Made: Masked autoencoder for distribution estimation. In Proceedings of the 32nd International Conference on International Conference on Machine Learning, 37, ICML’15, pp. 881–889, 2015.
Goodfellow, Ian, Pouget-Abadie, Jean, Mirza, Mehdi, Xu,
Bing, Warde-Farley, David, Ozair, Sherjil, Courville,
Aaron, and Bengio, Yoshua. Generative adversarial nets.
In Advances in Neural Information Processing Systems
27, pp. 2672–2680. Curran Associates, Inc., 2014.
Goodfellow, Ian J. NIPS 2016 tutorial: Generative adversarial networks. CoRR, abs/1701.00160, 2017.
He, Kaiming, Zhang, Xiangyu, Ren, Shaoqing, and Sun,
Jian. Deep residual learning for image recognition.
CoRR, abs/1512.03385, 2015.
Hoffman, Matthew D. and Johnson, Matthew J. Elbo
surgery: yet another way to carve up the variational evidence lower bound. 2016.
Hyvarjnen, Aapo and Pajunen, Petteri. Nonlinear Independent Component Analysis: Existence and Uniqueness
Results. Neural Networks, 1998.
Kingma, Diederik P and Welling, Max. Auto-Encoding
Variational Bayes. International Conference on Learning Representation, (Ml):1–14, 2014.
Kingma, Diederik P, Salimans, Tim, and Welling, Max. Improving Variational Inference with Inverse Autoregressive Flow. Advances in Neural Information Processing
Systems 27, (2011):1–8, 2016.
Larochelle, Hugo and Murray, Iain. The neural autoregressive distribution estimator. In The Proceedings of the
14th International Conference on Artificial Intelligence
and Statistics, volume 15 of JMLR: W&CP, pp. 29–37,
2011.
Makhzani, Alireza, Shlens, Jonathon, Jaitly, Navdeep, and
Goodfellow, Ian J. Adversarial autoencoders. CoRR,
abs/1511.05644, 2015.
Minka, Tom. Divergence measures and message passing.
Technical report, January 2005.
Ranganath, Rajesh, Tran, Dustin, and Blei, David M.
Hierarchical variational models.
arXiv preprint
arXiv:1511.02386, 2015.
Rezende, Danilo Jimenez and Mohamed, Shakir. Variational Inference with Normalizing Flows. Proceedings
of the 32nd International Conference on Machine Learning, 37:1530–1538, 2015.
Rezende, Danilo Jimenez, Mohamed, Shakir, and Wierstra,
Daan. Stochastic backpropagation and approximate inference in deep generative models. In Proceedings of
the 31st International Conference on International Conference on Machine Learning - Volume 32, ICML’14.
JMLR.org, 2014.
Tomczak, Jakub M. and Welling, Max. VAE with a vampprior. CoRR, abs/1705.07120, 2017.
Van Den Oord, Aäron, Kalchbrenner, Nal, and
Kavukcuoglu, Koray. Pixel Recurrent Neural Networks. 2016.
| 2cs.AI
|
On Low-Risk Heavy Hitter and Sparse Recovery Schemes
Vasileios Nakos∗
Harvard University
[email protected]
Yi Li
Division of Mathematical Sciences
Nanyang Technological University
[email protected]
arXiv:1709.02919v1 [cs.DS] 9 Sep 2017
David P. Woodruff
Carnegie Mellon University
[email protected]
Abstract
We study the heavy hitters and related sparse recovery problems in the low-failure probability
regime. This regime is not well-understood, and has only been studied for non-adaptive schemes.
The main previous work is on sparse recovery by Gilbert et al. (ICALP’13). We recognize an
error in their analysis, improve their results, and contribute new non-adaptive and adaptive
sparse recovery algorithms, as well as provide upper and lower bounds for the heavy hitters
problem with low failure probability. Our results are summarized as follows:
1. (Heavy Hitters) We study three natural variants for finding heavy hitters in the strict
turnstile model, where the variant depends on the quality of the desired output. For
the weakest variant, we give a randomized algorithm improving the failure probability
analysis of the ubiquitous Count-Min data structure. We also give a new lower bound
for deterministic schemes, resolving a question about this variant posed in Question 4 in
the IITK Workshop on Algorithms for Data Streams (2006). Finally, under the strongest
and well-studied `∞ /`2 variant, we provide the first randomized lower bound which is
simultaneously optimal in the approximation factor , the universe size n, and the failure
probability δ, for the full range of such parameters. Our lower bound shows that the
classical Count-Sketch data structure is optimal in all parameters.
2. (Sparse Recovery Algorithms) For non-adaptive sparse-recovery, we give sublinear-time
algorithms with low-failure probability, which improve upon Gilbert et al. (SICOMP’12)
and Gilbert et al. (ICALP’13). In the adaptive case, we improve the failure probability
0.99
from a constant by Indyk et al. (FOCS ’11) to e−k , where k is the sparsity parameter.
3. (Optimal Average-Case Sparse Recovery Bounds) We give matching upper and lower
bounds in all parameters on the measurement complexity of the `2 /`2 sparse recovery
problem in the spiked-covariance model, completely settling its complexity in this model.
1
Introduction
Finding heavy hitters in data streams is one of the most practically and theoretically important
problems in the streaming literature. In this problem, we are asked to report all frequent items
in a very long stream of elements coming from some universe. The main restriction is that the
memory consumption should be very small, much smaller than the universe size and the length of
∗
Supported by ONR grant N00014-15-1-2388
1
the stream. To rephrase the problem, consider a frequency vector x ∈ Rn , where n is the size of
the universe. Each element i in the data stream updates the frequency vector as xi ← xi + 1. At
the end of the data stream, we wish to find the coordinates of x for which |xi | ≥ kxk1 .
Oftentimes the problem is considered under a general streaming model called the turnstile
model, where arbitrary deletions and additions are allowed, that is, the frequency vector x ∈ Rn
receives updates of the form (i, ∆), and each such update causes the change of xi to xi + ∆. The
`p heavy hitters problem, for p ≥ 1, then asks to find all coordinates i for which |xi |p ≥ kxkpp . The
algorithm that treats the `1 case is the Count-min sketch [CM05], and the algorithm that treats
the `2 case is the Count-Sketch [CCFC04]. Both algorithms are randomized and succeed with
probability 1 − 1/ poly(n). In [CM05] the authors also suggest the “dyadic” trick for the strict
turnstile model, in which it is promised that at all times during the stream the vector x satisfies
xi ≥ 0 for all coordinates i. Their “dyadic” trick allows for finding heavy hitters approximatelly
in O( 1 log2 n) time, but with the downside of having a worse update time and a worse space
consumption by an O(log n) factor. The state of the art for heavy hitters is [LNNT16], where the
authors give an algorithm that satisfies the `∞ /lp guarantee, has space O( 1 log n), update time
O(log n), and query time O( 1 log1+γ n) for any constant γ > 0.
Another set of closely related problems occurs in the compressed sensing (CS) literature, which
has seen broad applications to biomedical imaging, sensor networks, hand-held digital cameras, and
monitoring systems, among other places. In this problem we also seek to find the large coordinates
of x ∈ Rn but with a different goal. Instead of finding all coordinates xi for which |xi | ≥ kxk1 , the
CS problem seeks an approximation x̂ to x such that the difference vector x − x̂ has small norm.
In particular, we consider the `2 /`2 error metric, that is, we require that kx − x̂k22 ≤ (1 + )kx−k k22 ,
where x−k is the vector x with the k largest entries (in magnitude) removed. If all `2 heavy hitters
are found, it is clear that the norm of x − x̂ can be made small, but the CS problem allows a small
number of heavy hitters to be missed if their contribution to the approximation error x − x̂ is small.
Gilbert et al. proposed the first sublinear-time algorithm for the `2 /`2 problem that achieves
O((k/) log(n/k)) measurements with constant failure probability [GLPS12]. Earlier sublineartime algorithms all contain several additional log n factors in their number of measurements. The
optimality of O((k/) log(n/k)) measurements was proved by Price and Woodruff [PW11]. Later
Gilbert et al. improved the failure probability to n−k/ poly(log k) [GNP+ 13], while their number of
measurements has a poor dependence on , which is at least 1/11 .
Despite the above works, our understanding of the complexity of heavy hitter and compressed
sensing schemes on the error probability is very limited. The overarching goal of our work is to
give a unified and systematic study of the dependence of the success probability of such schemes,
in terms of both upper and lower bounds. We start with formal definitions of the problems and
then state in detail our improvements over previous work.
1.1
Problem Formulation
Notation. For x ∈ Rn , we define Hk (x) to be the set of its largest k coordinates in magnitude,
breaking ties arbitrarily. For a set S let xS be the vector that occurs after zeroing out every
coordinate i ∈
/ S. We also define x−k = x[n]\Hk (x) and Hk, (x) = {i ∈ Hk (x) : |xi |2 ≥ k kx−k k22 }.
For a sequence of real numbers x1 , . . . , xm , we denote by quanti (xi , α) the α-percentile of x1 , . . . , xm .
A sketch f : Rm → Rn is a function that maps an n-dimensional vector to m dimensions. In
this paper, all sketches will be linear, meaning f (x) = Ax for some A ∈ Rm×n . The sketch length
m will be referred to as the space of our algorithms.
Definition 1 (Heavy hitters). For x ∈ Rn and p ≥ 1, a coordinate xi is called an -heavy hitter
2
in `p norm if |xi |p ≥ kxkpp . We consider the following three variants of the heavy hitters problem
with different guarantees:
1. Return a list containing all -heavy hitters but no /2-heavy hitters.
2. Return a list L of size O(1/) containing all -heavy hitters along with estimates x̂i such that
|xi − x̂i |p ≤ kx−d1/e kpp for all i ∈ L.
3. Return a list of size O(1/) containing all -heavy hitters.
When the algorithm is randomized, it has a parameter δ of failure probability; that is, the algorithm
succeeds with probability at least 1 − δ.
The variant with Guarantee 2 above is also referred to as the `∞ /`p problem. In this paper we
focus on p = 1 and p = 2, with corresponding Count-Min [CM05] and Count-Sketch [CCFC04]
data structures that have been studied extensively.
Definition 2 (`2 /`2 sparse recovery). An `2 /`2 -recovery system A consists of a distribution D on
Rm×n and a recovery algorithm R. We will write A = (D, R). We say that A satisfies the `2 /`2
guarantee with parameters (n, k, , m, δ) if for a signal x ∈ Rn , the recovery algorithm outputs
x̂ = R(Φ, Φx) satisfying
PΦ∼D kx̂ − xk22 ≤ (1 + )kx−k k22 ≥ 1 − δ.
In the above definition, each coordinate of Φx is called a measurement and the vector Φx is
referred to as a measurement vector or just as the measurements. The probability parameter δ is
referred to as the failure probability.
1.2
Our Results
Heavy hitters. Our first result is an improved analysis of the Count-Min sketch [CM05] for
Guarantee 3 under the change of the hash functions from 2-wise to O( 1 )-wise independence. Previous analyses for Guarantees 1 and 2 use O( 1 log nδ ) space; in contrast our analysis shows that this
version of the Count-Min sketch satisfies Guarantee 3 with only O( 1 log(n) + log( 1δ )) space. Notably, the 1 factor does not multiply the log( 1δ ) factor. This result has two important consequences.
First, it gives a uniform scheme for Guarantee 3; second, it implies an improved analysis of the
classic dyadic trick [CM05] for Guarantee 3 using O( 1 log(n)+log n log( logδn )) space. For constant
δ, previous analyses of the dyadic trick needed space O( 1 log n log( log n )) but our analysis shows
that O( 1 log(n) + log(n) log log(n)) space suffices. These results are summarized in Table 1.
Regarding the lower bound, we give the first bound for Guarantee 2 with p = 2, which is
simultaneously optimal in terms of , n and the failure probability δ. We prove that the space has
to be Ω(−1 log(n/δ)), matching the upper bound of Count-Sketch [CCFC04]. We note that the
lower bound of Ω(−1 log n) was previously known [BIPW10, JST11, Pri11], and we note that the
lower bounds given in [BIPW10, Pri11] actually give an Ω(−1 log n) lower bound and are valid for
the full range of parameters of and n. We note that previous analyses cannot be adapted to obtain
non-trivial lower bounds for δ < 1/n, since the lower bound instances used in their arguments have
deterministic upper bounds using O(−1 log n) space.
p
We also show a new
p randomized lower bound of Ω((1/)(log n + log(1/δ))) space for p = 1,
provided that 1/ > log(1/δ). Although not necessarily optimal, this lower bound is the first
to show that
p a term involving log(1/δ) must multiply the 1/ factor for p = 1. The assumption
that 1/ > log(1/δ) is necessary, as there exist deterministic O(1/2 ) space algorithms for p = 1
[GM06, NNW12]. For deterministic algorithms satisfying Guarantee 3 with p = 1, we also show
a lower bound of Ω(1/2 ) measurements, which resolves Question 4 in the IITK Workshop on
Algorithms for Data Streams [McG07].
3
Algorithm
Count-Min [CM05]
This paper
Dyadic Trick [CM05]
This paper
Space
1
1
1
log n + log( δ )
1
1
log(n) + log( δ )
log n
1
log n log( δ· )
log(n)
1
)
log(n) + log(n) log(
δ
Guarantee
1, 2
3
1, 2
3
Query time
Õ(n)
Õ(n)
Õ( 1 )
O(˜1 )
Table 1: Summary of previous heavy hitter algorithms. The notation O(·) for space complexity is
suppressed, Õ(·) hides logarithmic factors in n, 1/ and 1/δ.
Paper
[CCFC04]
[GLPS12]
[GNP+ 13]
[LNNT16]
Measurements
−1 k log n
−1
k log(n/k)
−11 k log(n/k)
−1 k log n
Decoding Time
n log n
−1 k poly(log(n/k))
k 2 · poly(−1 log n)
−1 k · poly(log n)
This paper
−1 k log(n/k)
−1 k log(n/k)
n
)
−1 k log( k
−1 k log3 n
−1 k 2 (log k) log1+γ (n/k)
−1 k 2 poly(log n)
Failure Probability
1/ poly(n)
Ω(1)
13
(n/k)−k/ log k
1/poly(n)
max{e
−
k
log3+γ k
, ( nk )− log k }
3
e−k/ log k
(n/k)−k/ log k
Table 2: Summary of previous sparse recovery results and the results obtained in this paper.
The notation O(·) is suppressed. The paper [GNP+ 13] and the third result of our paper require
k = nΩ(1) .
Sparse Recovery. We summarize previous algorithms in Table 2. Column sparsity denotes how
many 1s there are per column of the measurement matrix, which determines the update time of
the sketch in the turnstile model.
We give algorithms for the `2 /`2 problem with failure probability much less than 1/poly(n)
whenever k = Ω(log n). We present two novel algorithms, one running in O(k poly(log n)) time
and the other running in O(k 2 poly(log n)) time with a tradeoff in failure probability. Namely, the
first algorithm has a larger failure probability than the second one. The algorithms follow a similar
overall framework to each other but are instantiated with different parameters. We also show how
to modify the algorithm of [GNP+ 13] to obtain an optimal dependence on , achieving a smaller
failure probability along the way. All of these results are included in Table 2. Our algorithms,
while constituting a significant improvement over previous work, are still not entirely optimal.
We show, however, that at least in the spiked covariance model, which is a standard average-case
model of input signals, we can obtain optimal upper and lower bounds in terms of the measurement
complexity. Combined with the identification scheme from [GNP+ 13] we also obtain a scheme with
decoding time nearly linear in k, assuming that k = nΩ(1) .
Besides the above non-adaptive schemes, we also make contributions in terms of the failure
probability for adaptive schemes. For adaptive sparse recovery, Indyk et al. gave an algorithm
under the `2 /`2 guarantee [IPW11] using O((k/) log(n/k)) measurements and achieving constant
1−γ
failure probability. We give a scheme that achieves failure probability e−k
for any constant γ,
using the same number of measurements. Moreover, we present an algorithm for the regime when
k/ ≤ poly(log n). Our scheme achieves the stronger `∞ /`2 guarantee and fails with probability
1/ poly(log n). Thus, our algorithms improve upon [IPW11] in both regimes: in the high-sparsity
regime we get an almost exponential improvement in k, and in the low-sparsity regime we get
1/ poly(log n).
4
2
Our Techniques
Heavy hitters. All the schemes we give are for the strict turnstile model. Our first algorithm
is based on a small but important twist to the Count-Min sketch: we change the amount of
independence in the hash functions from 2 to O( 1 ). Since the estimate of any coordinate is an
over-estimate of it, we are able to show that the set of O( 1 ) coordinates with the largest estimates is
a superset of the set of the 1 -heavy hitters. Although changing the amount of independence might
increase both the update and the query time by a multiplicative factor of 1 , we show in our case
we can sufffer only a log2 ( 1 ) factor in the update time, if we are aiming for amortized time, and a
log2 ( 1 ) factor in query time in the worst case. The above observation gives also an improvement
for the well-known dyadic trick. Similary, the deterministc algorithms we give are based on the
important observation that any estimate only over-estimates the value of a coordinate.
Our deterministic lower bounds are based on choosing “bad input vectors” for the sketching
matrix S based on several properties of S itself. Since the algorithm is deterministic, it must
succeed even for these vectors.
Our randomized lower bounds come from designing a pair of distributions which must be distinguished by a heavy hitters algorithm with the appropriate guarantee. They are based on distinguishing a random Gaussian input from a random Gaussian input with a large coordinate planted
in a uniformly random position. By Lipschitz concentration of the 1-norm and 2-norm in Gaussian space, we show for certain input parameters the norms in the two cases are concentrated,
so we have a heavy hitter in one case but not the other. Typically, the planted large coordinate
corresponds to a column in S of small norm, which makes it indistinguishable from the noise on
remaining coordinates. The proof is carried out by a careful analysis of the variation distance of
the distribution of the image of the input under the sketch in the two cases.
Non-Adaptive Sparse Recovery.
1. Differences from [GLPS12] Our result has a similar framework to that of [GLPS12],
though chooses more carefully the main primitives it uses and balances the parameters in a more
effective way. Both schemes consist of O(log k) so-called weak systems: a scheme that takes as
input a vector x and returns a vector x̂ which contains a 2/3 fraction of the heavy hitters of x (the
elements with magnitude larger than √1k kx−k k2 ) along with accurate estimates of (most of) them.
Then it proceeds by considering the vector x − x̂, which contains at most 1/3 of the heavy hitters
of x. We then feed the vector x − x̂ to the next weak-level system to obtain a new vector which
contains at most (2/3)(1/3)k = 2k/9 of the heavy hitters. We proceed in a similar fashion, and
after the i-th stage we will be left with at most k/3i heavy hitters.
Each weak-level system consists of an identification and an estimation part. The identification
part finds a 2/3 fraction of the heavy hitters while the identification part estimates their values.
For the identification part, the algorithm in [GLPS12] hashes n coordinates to Θ(k) buckets using
a 2-wise independent hash function and then uses an error-correcting code in each bucket to find
the heaviest element. Since, with constant probability, a heavy hitter will be isolated and its value
will be larger than the ‘noise’ level in the bucket it is hashed to, it is possible to find a 2/3 fraction
of the heavy hitters with constant probability and O(k · poly(log n)) decoding time. We change
this by hashing again to O(k) buckets but using full independence. Moreover, in each bucket we
use a b-tree, which is a folklore data structure in the data stream literature, the special case of
which (b = 2) first appeared in [CM05]. This modification is the key to obtain smaller column
sparsity and faster decoding time, while it also makes the algorithm simpler. This allows us to
5
achieve better decoding time and much lower failure probability. The estimation part is a different
analysis of the folklore Count-Sketch data structure: we show that the estimation scheme can
be implemented with a better dependence on than the expander-based scheme in [GNP+ 13].
However, to beat constant failure probability, we need another idea on how we use the weak
systems. The intuition for our improvement is the following. We use O(log log k) weak systems with
sparsity parameters k, k3 , . . . , logk k , and after that point we either run the algorithm of [LNNT16]
3
to find the remaining heavy hitters and therefore obtain the result presented in Table 1, or design
the i-th weak system with O( logk k ) rows with sparsity parameter 3ki . The main idea is to choose
parameters such that the failure probability is dominated by the level logk k , while the total number
of measurements does not increase by more than a constant factor.
2. Recovery Time As discussed above, our improvement on failure probability over [GLPS12]
is achieved using more measurements in the i-th iteration for i > log log k. The earlier paper by
Gilbert et. al. [GNP+ 13] also used a similar trick with different choices of parameters. This,
however, raises an issue of recovery time. For instance, in [GNP+ 13], the measurement matrix
contains s = 2i /ic (where c is a constant) repetitions of an expander-based identification matrix
(see Lemma 4.10 and Theorem 4.9 of [GNP+ 13]). The expander-based identification matrix has
at least one non-zero entry in each column and thus the measurement matrix for the i-th iteration
has at least s non-zero entries in each column, which implies that when i = log k − 1, each column has at least Ω(k/ poly(log k)) nonzero entries. Updating measurements y ← y − Φx̂ will then
take Ω(k 2 / poly(log k)) time, where x̂ is the recovered signal and has Ω(k) nonzero coordinates.
Therefore the overall running time of the recovery algorithm will be Ω̃(k 2 ) instead of the claimed
Õ(k 1+α ) in their paper. Our algorithm suffers from the same problem, and we have to choose
different parameters in order to achieve recovery time that is linear in k, at the cost of a larger
failure probability (compare to the results of this paper in Table 2).
Adaptive Compressed Sensing. In [IPW11] the authors suggest a scheme for 1-sparse recovery,
using O(log log n) measurements and having constant failure probability. This scheme serves as the
1
main routine for their `2 /`2 algorithm. We show how to modify it so that we obtain poly(log
n)
1
failure probability for 1-sparse recovery. More generally, when k < log n we obtain poly(log
n) failure
probability with the stronger `∞ /`2 guarantee by using this scheme along with CountSketch.
We then show how the main iterative loop of their `2 /`2 scheme can be modified by changing the
1−γ
way the parameters vary, so that we are able to get e−k
failure probability instead of constant
failure probability.
3
Preliminaries
Count-Sketch. Our estimation will be based on the Count-Sketch data structure. The
Count-Sketch of a vector x is defined as follows. For r = 1, . . . , R, we hash all n coordinates
of x into B buckets with function hr : [n] → [B]. In each bucket we aggregate the elements with
random signs, that is, the value of bucket b ∈ [B] in iteration r is
X
Vb,r =
σi,r xi
i:hr (i)=b
where the σi,r are independent random signs (±1). For i ∈ I the estimate to xi returned by
Count-Sketch is
x̂i = median σi,r Vhr (i),r .
1≤r≤R
6
Weak System. We follow the approach in several previous works [PS12, GLPS17] by first constructing a weak system and then building an overall algorithm upon it.
Definition 3. A probabilistic matrix M with n columns is called a (k, ζ)-weak identification
matrix with the (`, δ, ) guarantee if there is an algorithm that given M x and a subset S ⊆ [n],
with probability at least 1 − δ outputs a subset I ⊆ S such that |I| ≤ ` and at most ζk elements
of Hk, (x) are not present in I. The time to compute I will be called the identification time.
Definition 4. We call an m × n matrix M a (k, ζ, η) weak `2 /`2 system if the following holds
for any vector x = y + z such that | supp(y)| ≤ k: given M x, one can compute x̂ such that there
exist ŷ, ẑ which satisfy the following properties: (i) x − x̂ = ŷ + ẑ, (ii) | supp(x̂)| = O(k), (iii)
| supp(ŷ)| ≤ ζk and (iv) kẑk22 ≤ (1 + η)kzk22 .
3.1
3.1.1
Toolkit
Black-Box Routines
In this subsection we mention three theorems crucial for our sparse recovery scheme. The first one
is from the recent paper by Larsen et al. on finding heavy hitters in data streams [LNNT16].
Theorem 1. Let x ∈ Rn , δ > 0 and K ≥ 1. There exists an (oblivious) randomized construction
of a matrix A such that given y = Ax we can find, with probability 1 − δ, a list L of size O(K) that
contains all K1 -heavy hitters of x with probability 1 − δ. The number of rows of A is O(K log(n/δ)),
the time to find the list L is O(K log3 n), and the column sparsity is O(log(n/δ)).
We also need the folklore b-tree data structure, which is a generalization of the dyadic trick.
The next analysis of a b-tree-based heavy hitter algorithm appears in [LNNT16].
Theorem 2. The b-tree produces a correct output L for (1/K)-heavy hitters with probability at
least 1 − δ. Moreover, when δ > 1/ poly(n), there is a choice of b such that the b-tree uses space
O(K log n), update time O(log n) and query time O((K 1+γ/2 /δ γ ) log1+γ n). The constants in the
big-Oh notations depend on γ.
It is important for our algorithm to have a subroutine that can estimate the values of a set S of
at most k elements, without knowing in advance what this set S is. This task is called a set query.
The following theorem is proven in [Pri11].
Theorem 3. Let c > 0. There is a randomized construction of a O(cK/) × n matrix A such
that for any x ∈ Rn and any set S ⊂ [n] of size at most K, we can find a vector x̂ ∈ Rn such
that kxS − x̂k22 ≤ kxS̄ k22 . The algorithm runs in time O(cK) and succeeds with probability at least
1 − K −c .
We will also need the following theorem from [GNP+ 13]. Although the theorem in the paper is
not stated this way, it is easy to see that by setting the quality of the approximation to be 1 and
the sparsity to be k/ we immediately obtain the desired result.
Theorem 4 ([GNP+ 13]). Suppose that k = nΩ(1) . There exists a randomized construction of a
matrix M of (k/) log(n/k) rows, such that given M x one can find a set S of size O(k/) such
−c
that |Hk/,1 \ S| ≤ ζk with probability 1 − nk , where ζ, c > 0 are absolute constants. The time to
find S is O(k 1+α poly(log n)), where a is any arbitrarily small positive constant.
7
4
Heavy Hitters Problem
4.1
Upper Bounds
In this section we give schemes that enable the detection of heavy hitters in data streams deterministically or with low failure probability. Our first result is Theorem 1, which is slightly more
general than finding heavy hitters but will be useful later when implementing the dyadic trick. This
theorem refers to a variant of the problem which we call the Promise Heavy Hitters problem, where
we are given a subset P of [n] of size m with the guarantee that every heavy hitter is contained in
the subset P .
Theorem 5. Let x ∈ Rn be a vector with non-negative coordinates. Assume a promise that the
heavy hitters of x lie in a set P of size m. There exists a data structure such that with probability
1 − δ, upon query, it finds the `1 heavy hitters (i.e., p = 1) of x under Guarantee 3. The space
usage is O( 1 log(m) + log( 1δ )), the amortized update time is O(log(m) log2 ( 1 ) + log2 ( 1 ) log( 1δ ))
and the worst-case query time is O(m log2 ( 1 )(log m + log( 1δ ))).
As we will see later, the above scheme implies the following corollary.
Corollary 6. There exists a data structure DS which finds the `1 heavy hitters of any x ∈ Rn in
the strict turnstile model under Guarantee 3. The space usage is O( 1 log(n)), the update time is
amortized O(log2 ( 1 ) log(n)) and the query time is O(n log2 ( 1 ) log(n)).
The following theorem follows by an improved analysis of the dyadic trick [CM05].
Theorem 7. There exists a data structure with space O( 1 log(n) + log(n) · log( log(n)
)) that finds
δ
the `1 heavy hitters of x ∈ Rn in the strict turnstile model under Guarantee 3 with probability at
least 1 − δ. The update time is O(log2 ( 1 ) log(n) + log(n) log2 ( 1 ) log( log(n)
)) and the query time
δ
log(n)
2 1
2 1
1
is O( (log ( ) log(n) + log(n) log ( ) log( δ )))).
Theorem 8. There exists a deterministic algorithm that finds the -heavy hitters of any vector
x ∈ Rn using space O(k 1+α (log( 1 ) log n)2+2/α ). The update time is O(poly(log n)) and the query
time is O(n · poly(log n)).
4.1.1
Overview of Techniques
All the schemes we give are in the strict turnstile model. As discussed in the Introduction, our
first algorithm is essentially a small twist of the Count-Min sketch: we change the amount of
independence in the hash functions from 2-wise to O(1/)-wise. Since the estimator of each coordinate never underestimates its value, we show that the set of O(1/) coordinates with the largest
estimates is a superset of the set of the (1/)-heavy hitters. Although changing the amount of
independence might increase both the update and the query time by a multiplicative factor of 1/,
we show in our case we can suffer only a log2 (1/) factor in the update time, if we are aiming
for amortized time, and a log2 (1/) factor in query time in the worst case. This can be achieved
using fast multipoint evaluation of polynomials. For the deterministic case, we show how expanders
that expand only on sets of size Θ(1/) (or equivalently lossless condensers) lead to schemes in the
strict turnstile model under Guarantee 3. Then, we instantiate the Guruswami-Umans-Vadhan
expander [GUV09] properly to get our desired algorithm. The idea to use expanders in the context
of heavy hitters has also been employed by Ganguly [Gan08], although his data structure solves the
`∞ /`1 problem and has Ω(1/2 ) space consumption. The reason why the strict turnstile assumption
is crucial is that we can avoid cancellations.
8
4.2
Low-failure probability algorithms in the strict turnstile model
For the proof of Theorem 5, we choose the constants CR = 5, Cδ = 10(ln 4 − 1), CB = 20, C0 = 30.
The algorithm is presented in Algorithm 1 and the proof is below.
Algorithm 1 Count-Min: Scheme for Heavy Hitters in the Strict Turnstile Model
R ← CR log(m) + d log(1/δ)
e
Cδ
B ← CB k/
Pick hash functions h1 , . . . , hR : [n] → [B], each being (C0 /)-wise independent
Cr,j ← 0 for all r ∈ [R] and j ∈ [B].
function Update(i, ∆)
Cr,hr (i) ← Cr,hr (i) + ∆ for all r ∈ [R]
end function
function Query
for each i ∈ P do
x̂i ← minr Cr,hr (i)
end for
return the (C0 + 1) 1 coordinates in P with the largest x̂ values.
end function
Proof of Theorem 5. We prove first that Algorithm 1 satisfies Guarantee 3 with update time
1
O( 1 log(m) + log 1δ ) and query time O( m
log(m) + log δ ). Later, we will show how to modify
the scheme in order to get the desired result.
Let S be the set of -heavy hitters of x. Let T be any set of at most C0 / coordinates that is
a subset of P \ S. Fix a hash function ha . Let Ba [T ] denote the set of the indices of the buckets
containing elements from T , i.e. Ba [T ] = {b ∈ [B] : ∃j ∈ T, ha (j) = b}. The probability that
|Ba [T ]| < 10/ is at most
C0
10
C0
e 10
10/
B
eCB
10
≤
≤
.
10/
B
10
CB
4
Considering all R hash functions, it follows that with probability ≥ 1 − (e/4)10R/ there exists an
index a∗T ∈ [R] such that |Ba∗T [T ]| ≥ 10/. A union bound over all possible T yields that with
probability at least 1 − δ, there exists such an index a∗T for each T of size C0 /.
Now, let Out be the set of coordinate the algorithm outputs and let T 0 = Out \ S. Clearly,
0
|T | ≥ C0 /. Discard some coordinates of T 0 so that |T 0 | = C0 /. We shall prove that there exists
an element j ∈ T 0 such that its estimate is strictly less than kx−1/ k1 and hence smaller than the
estimate of any element in S. This will imply that every element in S is inside Out. From the
previous paragraph, there exists a∗T 0 such that |BaT 0 [T 0 ] | > 10/. Since we have at most 1/ heavy
hitters and at most 1/ indices b ∈ [B] such that the counter Ca,b ≥ kx−1/ k1 , which implies that
we have at least 10/ indexes (buckets) of B such that the bucket has mass less than kx1/ k1 while
at least one element of T 0 is hashed to that bucket. Therefore, the estimate for this element is less
than kx− 1 k1 , which finishes the correctness of proof.
We note that Algorithm 1 clearly has an update time of O( 1 log(m) + log 1δ ) and a query time
1
O( m log(m) + m log 1δ ). We now show how to improve both runtimes, thus getting the desired
Theorem 5.
9
We split our stream into intervals, which we call epochs, of length C0 /: the l-th epoch starts
from the ((l − 1)C0 / + 1)-st update and ends at the (lC0 /)-th update. Let DS be the data
structure guaranteed by Theorem 7. During an epoch we maintain a list of elements L0 that were
updated in this epoch. This list is initialized to the empty list when an epoch begins. When a
new update (i, ∆) arrives, we store (i, ∆) to L0 . For the query operation, we first query DS to
obtain a set L, and then find the set L0 of indices i such that (i, ∆) ∈ L0 , and return L ∪ L0 as our
answer. Clearly L ∪ L0 has at most (2C0 + 1)/ elements. When an epoch ends, we feed all updates
(i, ∆) ∈ L0 to our data structure. Since we can obtain a (C0 /)-wise independent hash function
from a polynomial of degree C0 /, this means that we can do C0 / evaluations in time O( 1 log2 1 )
using multipoint evaluation of polynomials (see, e.g., Theorem 13 of [KNPW11]). Since we shall
evaluate O(log(n) + log 1δ ) hash functions the amortized update time follows. A similar argument
gives the query time.
Observe that in the previous algorithm the analysis depends only on the set S that contains
the -heavy hitters. By setting P = [n] and taking a union bound over all possible subsets that
the -heavy hitters may lie in, we obtain a uniform algorithm for the strict turnstile model under
Guarantee 3, namely Corollary 6.
We are now ready to prove Theorem 7, which as we said is an improved analysis of the dyadic
trick [CM05], which follows from Theorem 1. We first describe the algorithm of the dyadic trick.
We assume that n is a power of 2. Then for every 1 ≤ l ≤ log n, we partition [n] into 2l equal-sized
and disjoint intervals of length n/2l . Each interval will be called a node. We imagine a complete
binary tree on these nodes, where there is an edge from a node/interval u to a node v if v is an
subinterval of u of exactly the half length. Since in the strict turnstile model the `1 norm of any
interval/node is the sum of its elements, we can set up a Count-min sketch for every level l of the
tree to find out the “heavy” intervals at that level. If a level contains a heavy hitter then it will
always have `1 mass at least kxk1 , while there can be at most 1 intervals with mass more than
kxk1 . Given this observation we traverse the tree in a breadth first search fashion and at every
level keep a list L of all nodes that the Count-Min sketch on that level indicated as heavy. In the
next level there will be at most 2|L| nodes we need to consider: just the children of nodes in L.
At the last level, if all Count-min sketch queries succeed we will be left with at most 1 intervals of
length 1, that is we will have found all heavy hitters. The obstacle for getting suboptimal space by
a factor of log( log n ) stems from the fact that we have to set the parameters of Count-Min sketch
at every level in a way such that we can afford to take a union bound over the at most 1 log n
nodes we will touch while traversing the tree. However, with our improved analysis of Count-Min
sketch, we show that we can essentially avoid this additional log( 1 ) factor. Also, our algorithm
with the data structure guaranteed by Theorem 5 gives the stronger tail guarantee. We continue
with the proof of the theorem.
Proof of Theorem 7. Without loss of generality we can assume that n is a power of 2. We improve
l
the analysis for levels l ≥ 1/. Fix such a level l. Then we consider the vector y ∈ R2 , the ith
entry of which equals
i· n
yi =
2l
X
xj
j=(i−1)· nl +1
2
and run point query on this vector. Every coordinate of y corresponds to a node at level l and
since kx−d 1 e k1 ≥ ky−d 1 e k1 finding the heavy hitters of y corresponds to finding a set of nodes that
contain the heavy hitters of x. At every level, we are solving a version of Promise Heavy Hitters
with m = O( 1 ). We use the data structure guaranteed by Theorem 5 and set the failure probability
10
to be
δ
c log(n) ,
for some large constant c. Our space consumption per level equals O( 1 +log( log(n)
)).
δ
δ
. By a
Hence, at each level we are going to find all heavy nodes with probability at least 1 − log(n)
union bound over all levels the failure probability of our algorithm is at most δ. This means that,
while traversing the tree, we have only O( 1 ) candidates at each level and hence we are solving
a Promise Heavy Hitters Problem with m = O( 1 ). At every level we need O( 1 + log( log(n)
))
δ
log n
1
space and hence the total space of our algorithm is O( log(n) + log(n) log( δ )). Since we
are considering O( 1 log(n)) nodes in total and each function is O( 1 )-wise independent, similar
arguments as in the proof of Theorem 5 give the desired bounds.
4.3
4.3.1
Deterministic Algorithms in the Strict Turnstile Model
Heavy Hitters from Expanders
In this section we show that expanders which can be stored in low space imply schemes for heavy
hitters in the strict turnstile model. Then we show that the Guruswami-Umans-Vadhan expander
is such an (explicit) expander.
Definition 5. Let Γ : [N ] × [D] → [M ] be a bipartite graph with N left vertices, M right
vertices and left degree D. Then, the graph G will be called a (k, ζ) vertex expander if for all sets
S ⊆ [N ], |S| ≤ k it holds that Γ(S) ≥ (1 − ζ)|S|D.
2
Theorem 9. Let ζ, c be absolute constants such that 1−ζ
< c. Suppose there exists an explicit
c
bipartite ( , (1 − ζ)D) expander Γ : [n] × [D] → [m]. Suppose also that Γ can be stored in space S,
and for every i ∈ [n] it is possible to compute the neighbours of i in time t. Then there exists an
algorithm that finds the -heavy hitters of a vector x ∈ Rn with space usage m + S, update time
O(t) and query time O(n · t).
Proof. We maintain counters C1 , . . . , Cm . Whenever an update (i, ∆) arrives, we add ∆ to all
counters Cj for j that is a neighbour of i. The total update time is O(t). The query algorithm is
exactly the same as Count-Min: compute x̂i = mind∈D CΓ(i,d) and output the largest (c + 1)/
coordinates. We now show the correctness of the algorithm.
Fix a vector x ∈ Rn and let S be the set of its -heavy hitters. Let T be any other set of
c/ elements. Observe that for all i ∈ S and all j ∈ [m] such that j, i are neighbours in Γ, it
holds that Cj ≥ kx−d1/e k1 . We claim that there exist adjacent (i, j) for some i ∈ T such that
Cj < kx−d1/e k1 , whence the theorem follows.
Suppose that the claim is false. By the expansion property of G, the neighbourhood of T
has size at least (1 − ζ)(c/)D. Since the claim is false, all of these counters have value at least
kx−d1/e k1 . However, the total number of counters that are at least kx−d1/e k1 , is 2D/. It follows
from our choices of ζ and c that 2D/ < (1 − ζ)(c/)D, which is a contradiction.
Next we review the Guruswami-Umans-Vadhan (GUV) expander. The construction of the GUV
expander is included in Figure 1, and it is known that the construction does give an expander.
Theorem 10 ([V+ 12]). The expander Γ from construction 2 is a (hc , q − ahc) expander.
The following corollary follows with appropriate instantiation of parameters.
Corollary 11. For every constant α > 0 and all positive integers N, K ≤ N and there exists
an explicit (K, (1 − )D) expander with N left vertices, M right vertices and left-degree D =
O(log N log K/)1+1/α and M = k 1+α D2 .
11
Expanders from Parvaresh-Vardy Codes:
1. Define a neighbor function Γ : Faq × Fq → Fq × Fcq by
Γ(f, y) = [y, f0 (y), . . . , fc−1 (y)],
(1)
i
where f (Y ) is a polynomial of degree a − 1 over Fq . We define fi (Y ) = f (Y )h mod E(Y ) where E is an irreducible
polynomial of degree a over Fq .
Figure 1: Guruswami-Umans-Vadhan Expander
The next lemma is immediate, since the time to perform operations between polynomials of
degree a in Fq is O(c · poly(log a, log q)).
Lemma 12. The function Γ can be stored using a words of space and the time to compute Γ(f, y)
given f, y is O(c · poly(log a, log q)).
Our main result, Theorem 8, then follows from combining Corollary 11, Theorem 9 and Lemma 12.
4.3.2
Heavy hitters from Lossless Condensers
Our starting point is an observation that in the proof above we only needed expansion on sets of
size exactly c/ and thus some object weaker than an expander could suffice. The right object to
consider is called lossless condenser, which is essentially an expander that guarantees expansion on
sets of a specific size, but not on sets of smaller size. Here we follow the definitions in [V+ 12].
In order to define a lossless condenser we need the notion of min entropy of a distribution and
the notion of the total variation distance between two distributions.
Definition 6 (Min entropy). Let D be a distribution on a finite sample space Ω. The min entropy
of D is defined as
1
H∞ (D) = min log
.
ω∈Ω
Pω∼D [ω]
Definition 7 (Total variation distance). The total variation distance between two distributions P
and Q on Ω is defined to be
DT V (P, Q) =
1X
|P(ω) − Q(ω)|.
2
ω∈Ω
We also need the definitions of a pseudorandom object called a condenser. In a nutshell, a
condenser takes as input a random variable from a source which has some amount of min-entropy,
and some uniform random bits. It then outputs an element following a distribution that has
sufficiently large min-entropy. Let Un denote the uniform distribution on Fnp .
Definition 8 (Loseless condenser). Let a, b, c be positive integers and let p be a prime number. A
function C : Fap × Fbp → Fcp is called a lossless (κ, ζ) condenser if the following holds.
For every distribution A on Fbp with H∞ (A) ≥ κ, for any random variable A ∼ A and any ‘seed’
variable B ∼ Ub , the distribution (B, C(A, B)) is ζ-close to some distribution (Ub , Z) on Fb+m
with
p
min-entropy at least b + κ.
12
Equipped with these definitions, we are now ready to prove the following theorem relating
lossless condensers and heavy hitters. Although we can repeat the proof of the previous section
and argue that every lossless condenser is equivalent to an expander where only sets of a specific
size expand, we prefer to rewrite the proof in the language of condensers.
Theorem 13. Let p be a prime number and let a be such that n = pa . Let also κ be such that
c/ = 2κ , where c is some absolute constant. Let Con : Fap × Fbp → Fcp be a (κ, ζ) lossless condenser
that can be stored in space S. Let t be the time needed to evaluate Con. If 2−κ+1 / + ζ < 1 then
there exists an algorithm that finds the -heavy hitters of any x ∈ Rn with space S + pc+b , update
time t · pb and query time O(n · t · pb ).
Proof. We consider the following algorithm. We instantiate counters Ci,j for all i ∈ Fbp and j ∈ Fcp .
Upon updating (i, ∆) we perform updates Ci,Con(i,j) ← Ci,Con(i,j) + ∆ for all j ∈ Fbp . Clearly the
update time is t · pb and the space usage is the total number of counters plus the space needed to
store the condenser, in total O(pb · pc + S) words.
The query algorithm is exactly the same as in Count-Min: compute x̂i = minj∈Fbp Ci,Con(i,j)
and output the largest (c + 1)/ coordinates. We now analyze the algorithm.
Fix a vector x ∈ Rn and let S be the set of its -heavy hitters. Let T be any other set of c 1
elements. Observe that for all i ∈ S and all j ∈ Fbp it holds that Ci,Con(i,j) ≥ kx−d1/e k1 . We claim
that there exists i ∈ T, j ∈ Fbp such that Ci,j > kx−d1/e k1 . The theorem then follows.
Suppose that the claim is false, that is, for all i ∈ T and all j ∈ Fbp it holds that Ci,j ≥
kx−d1/e k1 . Consider the uniform distribution A on T and observe that A has min-entropy at
least κ. Let A be a random variable drawn from A. This implies that (UB , Con(A, UB )) is ζ-close
to some distribution D with min-entropy at least κ + b. Since the
number of counters that can
b
have a value at least kx−d1/e k1 is (2/)p , we have that PX∼A CX,Con(X,UB ) > kx−d1/e k1 ≤
(2/)2b 2−κ−b
only on elements in S ∪ T , we have
+ ζ . On the other hand, since A is supported
−κ
that PX∼D CUB ,Con(A,UB ) ≥ kx−d1/e k1 = 1. If (2/)2 + ζ < 1 we reach a contraction.
4.3.3
Heavy hitters from Error-Correcting List-Disjunct Matrices
In this subsection we give another reduction to error-correcting list-disjunct matrices, a combinatorial object that appears in the context of two-stage group testing. Explicit and strongly explicit
constructions of list-disjunct matrices are known [NPR11], although they are very similar to our
expander/condensers proof of the previous section and they do not yield better space complexity. We show that a sufficiently sparse error-correcting list-disjunct matrix that can be stored in
low space, with the appropriate choice of parameters, can induce a scheme for heavy hitters in
the strict turnstile model. Although group testing has been used in finding heavy hitters in data
streams [CH08], to the best of our knowledge this is the first time such a reduction is noticed.
Using such a matrix, one can perform two-stage combinatorial group testing when there also
some false tests, either false positive or false negative. For simplicity, we give the definition only
in the case of false positive tests, as we do not need more complicated ones. For an m × N binary
matrix M and a set S ⊆ [n], let M [S] = {q ∈ [m] : ∃j ∈ S, Mq,j = 1}.
Definition 9. Let k, ` be positive integers and e0 be a non-negative integer. A binary m×n matrix
is called (d, `, e0 )-list-disjunct if for any disjoint sets S, T ⊆ [n] with |S| = k, |T | = `, the following
holds: in any arbitrary subset X of M [T ] \ M [S] of size at most e0 , there exists a column j ∗ ∈ T
such that M [{j}] \ (X ∪ M [S]) 6= ∅.
13
Theorem 14. Let s, n be integers with s ≤ n. Suppose there exists an m × n matrix M that is
(1/, c/, s/) list-disjunct, where s is some integer and C0 some absolute constant. Suppose also
that:
• M can be stored in space S;
• The column sparsity of M is s;
• Each entry of M can be computed in time t.
Then there exists a streaming algorithm which finds the -heavy hitters of any vector x using space
S + m, having update time s · t and query time O(n · st).
Proof. We use M as our sketch, i.e., we have access to y = M x. The inner product of each row of
matrix M with x defines a counter. We describe the query algorithm. For each coordinate i, compute x̂i as the minimum value over all counters it participates in. That is, x̂i = minq:Mq,i =1 yq . The
query algorithm outputs the list L containing the coordinates with the largest (c + 1)/ coordinates
in x̂.
Next we show the correctness of the query algorithm. Let S be the set of -heavy hitters and
let R be the set of coordinates that the algorithm outputs. Define T = R \ S. For any i ∈ S, we
know that x̂i ≥ kx−d1/e k1 . Let X = {q : yq > kx−d1/e k1 and @i ∈ S : Mq,i = 1}, that is, the
set of rows which appear to be ‘heavy’ but contain no -heavy hitter. We claim that |X| ≤ s/.
Indeed, if |X| > s/, the total `1 mass of these counters would be more than skx−d1/e k1 . But since
every coordinate i participates in at most s counters, the total `1 mass of counters with no heavy
hitters is at most skx−d1/e k1 . This gives us the desired contradiction.
Assume now that there exists i ∈ S which is not included in L. This means that for every
j ∈ T , x̂j ≥ kx−d1/e k1 , which means that M [j] ⊆ M [S] ∪ X. Define sets S 0 , T 0 by moving some
coordinates from T to S if needed so that |T 0 | = c/, |S 0 | = 1/. Observe that it still holds that
M [j] ⊆ M [S] ∪ X for all j ∈ T 0 . But this violates the definition of (1/, c/, s/) list-disjunct
matrix. This proves correctness of our algorithm.
Clearly, the space that the streaming algorithms uses is S + m, the space needed to store A
plus the space needed to store y. Moreover, the update time is s · t.
4.4
Lower Bounds
4.5
Preliminaries
We start with some standard facts about the Gaussian distribution.
Fact 15. (One-Dimensional Variation Distance) Let P1 denote the N (0, Ir ) Gaussian distribution,
and P2 the N (τ, Ir ) Gaussian distribution. Then
kτ k2
DT V (P1 , P2 ) =
P
|g| ≤
,
2
g∼N (0,1)
where N (0, 1) denotes a standard one-dimensional normal random variable.
Proof. Let U be an r × r orthogonal matrix which rotates τ to kτ k2 · e1 , where e1 is the first
standard unit vector. Then by rotational invariance, DT V (P1 , P2 ) = DT V (N (0, 1), N (kτ k2 , 1)).
Then by Section 3 of [Pol],
kτ k2
DT V (N (0, 1), N (kτ k2 , 1)) =
P
|g| ≤
.
2
g∼N (0,1)
14
Theorem 16 (Lipschitz concentration, [LT91, p21]). Suppose that f : Rn → R is L-Lipshitz with
respect to the Euclidean norm, i.e. |f (x) − f (y)| ≤ Lkx − yk2 for all x, y ∈ Rn . Let x ∼ N (0, In )
and let f be L-Lipshitz with respect to the Euclidean norm. Then
t2
P {|f (x) − E f (x)| ≥ t} ≤ 2e− 2L2 ,
t ≥ 0.
Fact 17 (Concentration of `1 -Norm). Suppose x ∼ N (0, In ) and n ≥ 32 ln(6/δ). Then
n
δ
3n
P
≥1− .
≤ kxk1 ≤
8
4
3
√
Proof. The function f (x) = kxk1 satisfies |f (x) − f (y)| ≤ kx − yk1 ≤ nkx − yk2 for all x, y in Rn
and so by Theorem 16 we have
t2
P
x∼N (0,In )
{|kxk1 − E kxk1 | > t} ≤ 2e− 2n
p
p
Note that since g ∼ N (0, 1), E |g| = 2/π, and so we have E kxk1 = 2/π. Hence, setting t = n/4,
we have that n/8 ≤ kxk1 ≤ 3n/4 with probability at least 1 − 2e−n/32 , which is at least 1 − δ/3
provided that n ≥ 32 ln(6/δ).
Fact 18 (Concentration of `2 -Norm). Suppose x ∼ N (0, In ) and n ≥ 18 ln(6/δ). Then
√
√
n
δ
3 n
P
≥1− .
≤ kxk2 ≤
2
2
3
Proof. The function f (x) = kxk2 satisfies |f (x) − f (y)| ≤ kx − yk2 for all x, y in Rn and so by
Theorem 16 we have
2 /2
P{kxk2 − E kxk2 | > t} ≤ 2e−t
√
n
if x ∼ N (0, In ). It is well-known (see, e.g., [CRPW10]) that √n+1
≤ E kxk2 ≤ n. Hence, setting
√
√
√
t = n/3, we have that n/2 ≤ kxk2 ≤ 3 n/2 with probability at least 1 − 2e−n/18 , which is at
least 1 − δ/3 provided that n ≥ 18 ln(6/δ).
Fact 19 (Univariatep
Tail Bound). Let g ∼ N (0, 1). There exists δ0 > 0 such that it holds for all
δ < δ0 that P{|g| ≤ 4 log(1/δ)} ≥ 1 − δ/3.
2
Proof. It is well-known (see, e.g.@ [Gor41]) P{g ≥ x} ≤
/2
e−x
√
,
x 2π
and so P{g ≥ 2
p
log(1/δ 2 )} ≤
δ2
2π log(1/δ)
≤ 6δ , for δ less than a sufficiently small absolute constant δ0 > 0. Hence, by symmetry
p
of the normal distribution, |g| ≤ 4 log(1/δ) with probability at least 1 − δ/3.
√
2
In our proofs we are interested in lower bounding the number r of rows of a sketching matrix
S. We can assume throughout w.l.o.g. S has orthonormal rows, since given Sx, one can compute
RSx for any r × r change-of-basis matrix R.
15
4.6
Deterministic Lower Bounds for `1 -Heavy Hitters
Theorem 20 (Strict turnstile deterministic lower bound for Guarantees 1,2). Assume that n =
Ω(−2 ). Any sketching matrix S must have Ω(−2 ) rows if, in the strict turnstile model, it is always
possible to recover from Sx a set which contains all the -heavy hitters of x and contains no items
which are not (/2)-heavy hitters.
Proof. Let r be the number of rows of the sketching
S, which,
Pnw.l.o.g., 2has orthonormal
Pnmatrix
T
2
T
rows. Because the columns of S are orthonormal, i=1 kS Sei k2 = i=1 kSei k2 = r and
n
X
n
n
X
√ X
T
T
kS Sei k1 ≤ n
kS Sei k2 ≤ n
kS T Sei k22
i=1
i=1
!1/2
√
≤ n r,
i=1
where the first inequality uses the relationship between the `1 -norm and the `2 -norm, and the second
inequality is the Cauchy-Schwarz inequality. It follows by an averaging argument that there exists
√
an i∗ ∈ [n] for which kS T Sei∗ k22 ≤ 2r/n and kS T Sei∗ k1 ≤ 2 r. Consider the possibly negative
vector v = ei∗ −S T Sei∗ , which is in the kernel of S, since I −S T S projects onto the space orthogonal
√
to the row space of S. Then vi∗ ≥ 1−2r/n ≥ 1/2, for r ≤ n/4. Also kvk1 ≤ 1+kS T Sei∗ k1 ≤ 2 r+1.
Define w to be the vector with wj = |vj | for all j 6= i∗ , and wi∗ = 0. Note that w is non-negative.
√
Then kw + vk1 ≤ 4 r + 2, while (w + v)i∗ ≥ 1/2, and note that w + v is also a non-negative vector.
Since both w and w + v are non-negative vectors, either can be presented to S in the strict turnstile
model. However, S(w + v) = Sw, and so any algorithm cannot distinguish input w + v from input
w. However, wi∗ = 0 while (w + v)i∗ ≥ 1/2. For the algorithm to be correct, i∗ cannot be an
√
(/2)-heavy hitter for either vector, which implies 4(1/2)
≤ 2 , that is, r = Ω(1/2 ).
r+2
Theorem 21 (Turnstile deterministic lower bound for Guarantee 3). Assume that n = Ω(−2 ).
Any sketching matrix S must have Ω(−2 ) rows if, in the turnstile model, some algorithm never
fails in returning a superset of size O(1/) containing the -heavy hitters. Note that it need not
return approximations to the values of the items in the set which it returns.
Proof. Let r be the number of rows
has orthonormal rows.
P of the sketching matrixPS, which, w.l.o.g., √
As in the proof of Theorem 20, ni=1 kS T Sei k22 = r and ni=1 kS T Sei k1 ≤ n r. It follows that by
an averaging argument there is a set T of 9n/10 indices i ∈ [n] for which both kS T Sei k22 ≤ 20r/n
√
and kS T Sei k1 ≤ 20 r. Let vi = ei − S T Sei , which is in the kernel of S for each i. Furthermore,
as in the proof of Theorem 20, for i ∈ T , the i-th coordinate (vi )i of vi , is at least 1/2 for r ≤ n/4,
√
and also kvi k1 ≤ 2 r + 1.
Now, since Svi = 0 for all vi , a turnstile streaming algorithm cannot distinguish any of the
input vectors vi from the input 0. The output of the algorithm on input 0 can contain at most
O(1/) indices. But |T | ≥ 9n/10 = ω(1/), so for at least one index i ∈ T , the algorithm will be
√
wrong if i is an -heavy hitter for vi , so we require 2(1/2)
≤ , that is, r = Ω(1/2 ).
r+1
4.7
4.7.1
Randomized Lower Bounds
`1 -Heavy Hitters
p
We assume in this
section
that
1/
≥
C
log(1/δ) for a sufficiently large constant C > 0. To
p
obtain an Ω(−1 log(1/δ) lower bound on the number of rows of the sketching matrix, such an
assumption is necessary, otherwise for small enough δ (as a function of n and ) the lower bound
would contradict the −2 poly(log n) deterministic upper bounds of [GM06, NNW12].
16
Theorem 22. (Randomized Turnstile `1 -Heavy
Hitters Lower
l
m Bound for Guarantees 1, 2) Assume
p
p
−1
log(1/δ) . Then for any sketching matrix S, it
that 1/ ≥ C log(1/δ) and suppose n ≥ 64
p
must have Ω(−1 log(1/δ)) rows if, in the turnstile model, it succeeds with probability at least 1 − δ
in returning a set containing all the `1 -heavy hitters and containing no items which are not (/2)
`1 -heavy hitters.
l
m
p
Proof. Let the universe size be n = 64−1 log(1/δ) (note if the actual universe size is larger, we
m
l
p
can set all but the first 64−1 log(1/δ) coordinates of our input to 0). Note that the assumption
in the previous paragraph implies that n ≥ 32 ln(6/δ), which we will need to apply Fact 17.
Let r be the number of rows of thepsketch matrix S, where r ≤ n. Note that if r > n, then we
immediately obtain the claimed Ω(−1 log(1/δ)) lower bound.
Hard Distribution. Let I be a uniformly random element in [n].
Casep1: Let η be the distribution N (0, In ), and suppose xp∼ η. By Fact 17, kxk1 ≥ n/8 ≥
8−1 log(1/δ) with probability 1 − δ/3. By Fact 19, |xI | ≤ 4 log(1/δ) with probability 1 − δ/3.
Let E be the joint occurrence of these events, so that P[E] ≥ 1 − 2δ/3.
By our choice of n, it follows that if E occurs, then |xI | ≤ 2 kxk1 , and therefore I cannot be
output by an `1 -heavy hitters algorithm if the algorithm succeeds.
Case 2: Let γ be the distribution (4n)eI + N (0, In ), where I is drawn uniformly at random
from [n], and eI denotes the standard unit vector in the I-th direction. Suppose p
x ∼ γ, and let
y ∼ N (0, In ). By Fact 17, kyk1 ≤ 3n
with
probability
1
−
δ/3.
By
Fact
19,
|x
|
≤
4
log(1/δ) with
I
4
probability 1 − δ/3.
Let F be the
p joint occurrence of these events, so that P[F] ≥ 1 − 2δ/3.
p
Note that 4 log(1/δ) ≤ n, and consequently if F occurs, then |xI | ≥ (4n)−4 log(1/δ) ≥ 3n,
while kxk1 ≤ 4n + kyk1 ≤ n, provided ≤ 1/16. Consequently, if F occurs, for an `1 -heavy hitters
algorithm to be correct, it must output I.
A Conditioning Argument. We let η 0 be the distribution of η conditioned on E, and let γ 0 be
the distribution of β conditioned on F. For a distribution µ on inputs y, we let µ̄ be the distribution
of Sy.
Note that any `1 -heavy hitters algorithm which succeeds with probability at least 1 − δ can
decide, with probability at least 1 − δ, if x ∼ η 0 or if x ∼ γ 0 . Consequently, DT V (η̄ 0 , γ¯0 ) ≥ 1 − δ.
On the other hand, we have
DT V (η̄, γ̄) ≥ DT V (η̄ 0 , γ¯0 ) − DT V (η̄ 0 , η̄) − DT V (γ¯0 , γ̄)
2δ 2δ
≥ DT V (η̄ 0 , γ¯0 ) −
−
3
3
7δ
≥ 1− ,
3
where the first inequality is the triangle inequality. The second inequality follows from the fact
that for a distribution µ and event W, DT V (µ, µ | W) = P[6 W], together with our bounds on P[E]
and P[F] above.
Therefore, to obtain our lower bound, it suffices to show if the number r of rows of S is too
small, then it cannot hold that DT V (η̄, γ̄) ≥ 1 − 7δ
3 .
17
Bounding the Variation Distance. Since S has orthonormal rows, by rotational invariance of
the Gaussian distribution, the distribution of η̄ is equal to N (0, Ir ). Also by rotational invariance
of the Gaussian distribution, the distribution of γ̄ is (4n)SI +N (0, Ir ), where SI is the I-th column
of S.
Since S has orthonormal rows, by a Markov bound, for 9/10 fraction of values of I, it holds
that kSI k22 ≤ 10r
n . Call this set of columns T .
Let G be the event that I ∈ T . Then P[G] ≥ 9/10. Suppose DT V (η̄, γ̄) ≥ 1 − 7δ
3 . We can write
γ̄ = P[G] · (γ̄ | G) + P[¬G](γ̄ | ¬G), and so if the number r of rows of S is large enough so that
DT V (η̄, γ̄) ≥ 1 − 7δ
3 , then
1−
7δ
3
≤ DT V (η̄, γ̄)
=
=
=
≤
≤
1
kη̄ − γ̄k1
2
1
kη̄ − P[G · (γ̄ | G) − P[¬G](γ̄ | ¬G)k1
2
1
k P[G(η̄ − (γ̄ | G)) + P[¬G](η̄ − (γ̄ | ¬G))k1
2
1
1
P[Gkη̄ − γ̄ | Gk1 + (1 − P[G])kη̄ − γ̄ | ¬Gk1
2
2
1 9
1
kη̄ − γ̄ | Gk1 +
·1 ,
2 10
10
and so DT V (η̄, γ̄ | G) ≥ 1 − 70δ
27 .
The variation distance between N (0, Ir ) and (4n)Si +N (0, Ir ) for a fixed i ∈ T is, by rotational
invariance and by rotating Si to be in the same direction as the first standard unit vector e1 , the
same as the variation distance between N (0, Ir ) and 4nkSi k2 e1 + N (0, Ir ), which is equal
p to the
variation distance between N (0, 1) and N (4nkSi k2 , 1). Using that i ∈ T and so kSi k2 ≤p 10r/n,
we apply Lemma 15 to obtain that the variation distance is at most Pg∼N (0,1) {|g| ≤ 2n 10r/n}.
It follows that
n
o
X 1
p
DT V (η̄, γ̄ | G) ≤
DT V (η̄, γ̄ | (I = i)) ≤
P
|g| ≤ 2n · 10r/n .
|T |
g∼N (0,1)
i∈T
Supposing the number r of rows of S is large enough so that DT V (η̄, γ̄ | G) ≥ 1 −
p
P{|g| ≤ 2n 10r/n} ≥ 1 − 70δ
27 , or equivalently
(
)
r
10r
70δ
P
|g| > 2n
≤
.
n
27
g∼N (0,1)
70δ
27 ,
this implies
(2)
p
p
p
√
Suppose that r ≤ α−1 log(1/δ). Then 2n 10r/n = 2 10rn ≤ C 0 α log(1/δ) for some
absolute constant C 0 . Take α = 1/(2C 02 ). Invoking the well-known bound (see, e.g. [Gor41])
2
e−x /2
1
√
P {g ≥ x} ≥
,
x + 1/x
g∼N (0,1)
2π
p
p
we have that Pg∼N (0,1) {|g| > 2n 10r/n} ≥ C 00 δ/ log(1/δ), when δ is small enough, contradicting
that this probability needs to be at mostp70δ/27 by (2).
It must therefore hold that r > α−1 log(1/δ). This completes the proof.
18
4.7.2
`2 -Heavy Hitters
Our next theorem implies a lower bound for the `∞ /`2 heavy hitters problem as defined in Section 1.1, since it implies such an x̂ satisfying Definition 1 cannot be an acceptable output. It is
based on designing a pair of hard distributions which cannot be distinguished by a small sketch.
We show this by using rotational properties of the Gaussian distribution to reduce our problem to
a univariate Gaussian mean estimation problem, which we show is hard to solve with low failure
probability.
Theorem 23. Suppose that δ < δ0 and < 1/0 for sufficiently small absolute constants δ0 , 0 ∈
(0, 1) and n ≥ 64−1 log(6/δ) . Then for any sketching matrix S, it must have Ω(−1 log(1/δ))
rows if it succeeds with probability at least 1 − δ in returning a set containing all the -heavy hitters
and containing no items which are not (/2)-heavy hitters.
Proof. Let the universe size be n = 64−1 log(6/δ) , which is large enough inorder for us to
apply Fact 18 (note if the actual universe size is larger, we can set all but the first 64−1 log(6/δ)
coordinates of our input to 0).
Let r be the number of rows of the sketch matrix S, where r ≤ n. If r > n, then we immediately
obtain an Ω(−1 log(1/δ)) lower bound. We can assume that S has orthonormal rows, since a change
of basis to the row space of S can always be performed in a post-processing step.
Hard Distribution. Let I be a uniformly random index in [n].
p
Case 1: Let η be the distribution N (0, Ip
n/2 with
n ), and suppose x ∼ η. By Fact 18, kxk2 ≥
probability 1 − δ/3. By Fact 19, |xI | ≤ 4 log(1/δ) with probability 1 − δ/3.
Let E be the joint occurrence of these events, so that P(E) ≥ 1 − 2δ/3.
By our choice of n, it follows that if E occurs, then x2I ≤ 16 log(1/δ) ≤ 2 kxk22 , and therefore I
cannot be output by an `2 -heavy hitters algorithm.
√
Case 2: Let y ∼ N (0, In ) and x√= neI +y, where eI denotes the standard basis vector in the I-th
p
√
direction. By Fact 18, kyk2 ≤ 3 2 n with probability 1−δ/3. By Fact 19, |yI | ≤ 4 log(1/δ) < n/2
with probability 1 − δ/3.
Let F be the joint occurrence of these events, so that P(F)
≥ 1 − 2δ/3.
√
p
√
5 n
If event F occurs, then |xI | ≥ 3 n − 4 log(1/δ) ≥ 2 , and so x2I ≥ 25n
4 . We also have
√
√
√
3 n
2
kxk2 ≤ 3 n + 2 ≤ 2 n, provided ≤ 1/36, and so kxk2 ≤ 4n. Consequently, x2I ≥ kxk22 .
Consequently, if F occurs, for an `2 -heavy hitters algorithm to be correct, it must output I.
Conditioning. Let η 0 be the distribution of η conditioned on E, and let γ 0 be the distribution of
γ conditioned on F. For a distribution µ on inputs y, we let µ̄ be the distribution of Sy.
Note that any `2 -heavy hitters algorithm which succeeds with probability at least 1 − δ can
decide, with probability at least 1 − δ, whether x ∼ η 0 or x ∼ γ 0 . Hence, DT V (η̄ 0 , γ¯0 ) ≥ 1 − δ.
Observe for any measurable set A ⊆ Rm it holds that
η̄(A) − µ̄(A)
2
− (η̄ 0 (A) − µ̄0 (A)) ≤ δ,
2
3
1 − 3δ
and so it then follows that
DT V (η̄, γ̄) ≥
DT V
(η̄ 0 , γ¯0 )
2δ
−
3
2
7δ
1− δ ≥1− .
3
3
Therefore, to obtain our lower bound, it suffices to show if the number r of rows of S is too small,
then it cannot hold that DT V (η̄, γ̄) ≥ 1 − 7δ/3.
19
Bounding the Total Variation Distance. Since S has orthonormal rows, by rotational invariance of the Gaussian distribution, the distribution of η̄ is identical to N (0, Ir ) and the distribution
√
of γ̄ identical to (3 n)SI + N (0, Ir ), where SI is the I-th column of S.
Since S has orthonormal rows, by a Markov bound, for 9/10 fraction of values of I, it holds
that kSI k22 ≤ 10r
n . Call this set of columns T .
Let G be the event that I ∈ T , then P(G) ≥ 9/10. It follows that
DT V (η̄, γ̄) ≤ P(G)DT V (η̄, γ̄|G) + P(¬G)DT V (η̄, γ̄|¬G)
≤ P(G)DT V (η̄, γ̄|G) + 1 − P(G)
= 1 − P(G)(1 − DT V (η̄, γ̄|G))
9
≤ 1 − (1 − DT V (η̄, γ̄|G)).
10
Hence, in order to deduce a contradiction that DT V (η̄, γ̄) < 1 − 7δ/3, it suffices to show that
DT V (η̄, γ̄|G) < 1 − 70δ/27.
√
The total variation distance between N (0, Ir ) and (3 n)Si + N (0, Ir ) for a fixed i ∈ T is,
by rotational invariance and by rotating Si to be in the same direction as the first standard basis
√
vector e1 , the same as the total variation distance between N (0, Ir ) and (3 n)kSi k2 e1 + N (0, Ir ),
√
which is equal to the total variation distance
between N (0, 1) and N (3 nkSi k2 , 1).
p
Using that i ∈ T and so kSi k2 ≤ 10r/n,
p we apply Lemma 15 to obtain that the variation
√
distance is at most P[|N (0, 1)| ≤ (3/2) n · 10r/n]. It follows that
DT V (η̄, γ̄ | G) ≤
n
o
X 1
p
√
DT V (η̄, γ̄ | I = i) ≤
P
|g| ≤ (3/2) n · 10r/n ,
|T |
g∼N (0,1)
i∈T
and thus it suffices to show, when r is small, that
3√
70δ
P
|g| ≥
10r >
.
2
27
g∼N (0,1)
Observe that the left-hand is a decreasing function in r, and so it suffices to show the inequality
above for r = α−1 log(1/δ) for some α > 0.
Invoking the well-known bound that (see, e.g., [Gor41])
1
1 t2
{g ≥ t} ≥ √ · e− 2 ,
g∼N (0,1)
2π 2t
P
we have that
P
g∼N (0,1)
3√
|g| ≥
10r
2
t≥
√
2,
45
3
1
70
≥ √ δ 4 αp
> δ
27
4 5π
α log(1/δ)
when α is small enough. Therefore it must hold that r ≥ α−1 log(1/δ) and the proof is complete.
5
5.1
Non-adaptive Sparse Recovery
Weak System
We start with the following lemma.
20
Lemma 24. Let δ > 0. There exists a randomized construction of a (k, c1 )-weak identification matrix with the (cB k , δ, ) guarantee; the matrix has O(( k + 1 log 1δ ) log n
k ) rows and the identification
time is O( 1 (k + log 1δ ) log1+γ ( n
)),
for
any
constant
γ
>
0,
where
c
,
c
1
B > 0 are absolute constants.
k
Proof. We describe the matrix as a scheme which performs a set of linear measurements on the
vector x. Without loss of generality we can assume that Hk, (x) has size k, otherwise we can expand
it to size k without affecting the guarantee of the weak identification matrix.
Let R = cR d k1 log( 1δ )e and B = cB k/. For each r ∈ [R], we hash every coordinate i ∈ [n] to
B buckets using a B-wise indepedent hash function hr : [n] → [B]. Let Njr be binary variables
defined as follows: Njr = 1 iff |{i : hr (i) = j}| ≥ 4n/B. It is clear that at most B/4 buckets will be
big. Relabel the buckets so that the first 3B/4 buckets are small buckets. Define
3
(r)
B : there exists exactly one j ∈ Hk, (x) such that hr (j) = b ,
Tgood = b ∈
4
X
3
(r)
2
2
B :
kx−k k2 .
Tbad = b ∈
xi ≥
4
5k
i∈H
/ k, (x)
hr (i)=b
(r)
In other words, Tgood is the set of the buckets in iteration r which receive exactly one element of
(r)
Hk, (x), while Tbad is the set of buckets which receive energy more than (/5k)kx−k k22 from elements
outside of Hk, (x).
(r)
(r)
Since every two buckets in Tbad share no coordinates outside Hk, (x), it must hold that |Tbad | ≤
5k/. The probability that there exist more than c3 k coordinates in Hk, (x) that land in a bucket in
(r)
Tbad is at most c3kk ( c5B )c3 k ≤ e−Ω(k) . On the other hand, it is a standard result (see, e.g., [GSTV06,
Lemma 23]) that at least (1−c4 )k elements in Hk, (x) are isolated with probability at least 1−e−Ω(k) ,
(r)
and thus |Tgood | ≥ (1 − c4 )k.
Overall with probability ≥ 1 − e−Ω(k) , at least (1 − c3 − c4 )k elements in Hk, (x) land in buckets
(r)
(r)
(r)
(r)
in Tgood \ Tbad . Each bucket in Tgood \ Tbad is a 1-heavy hitter problem with signal length at most
4n/B and the energy of the heavy hitter is at least 5 times the noise energy. By Theorem 2 we can
1
perform this task in O(log(n/k)) measurements with probability at least 1 − log(n/k)
and decoding
time O(log1+γ (n/k)) for any constant γ > 0. We collect the coordinate found in each bucket, and
return the coordinates with at least R/2 occurrences. Since there are at most BR buckets, the size
of the returned coordinates is at most 2B = 2cB k/.
Since the failure probability of the b-tree is at most a constant, the probability that more than
c5 k buckets fail is at most e−Ω(k) . By considering all R iterations, this means that with probability
1 − exp(−Ω(Rk)) ≥ 1 − δ at most c5 (c3 + c4 )k = c1 k elements of Hk, (x) will not be recognized.
Lemma 25. Let T ⊆ [n] be a set of indices such that |T | ≤ cT k/ for some absolute constant cT
and ζ ≤ 12 be an absolute constant. The Count-Sketch scheme of B = cB (cT + 1)k/ buckets
and R = cR (log 1 + k1 log 1δ ) repetitions yields an estimate x̂ such that
n
o
i ∈ T : |xi − x̂i |2 >
kx−k k22 ≤ ζk,
16k
with probability ≥ 1 − δ.
21
Proof. For the purpose of analysis only we may assume that Hk (x) ⊆ T , otherwise we can include
Hk (x) in T and replace cT with cT + 1. We call elements in T candidates, and the elements not in
T noise elements.
Consider hashing all n elements into B buckets, using B-wise independent hash functions. We
say a bucket b in repetition r is good if
X
x2i ≤
kxT c k22 ,
160k
i6∈T
hr (i)=b
otherwise we say that bucket b in repetition r is bad.
First we show that, with probability at least 1 − exp(−Ω(|T |R)) ≥ 1 − δ 1/ , at least (1 − θ1 )T
candidates are isolated and land in good buckets in at least (1 − θ2 )R repetitions. The isolation
claim is essentially the same to those in [GSTV06, PS12, GLPS17], nevertheless we give a proof
below for completeness. In each repetition r, it follows from a standard result (see [GG11, Section
4.3]) that at least (1 − θ3 )|T | candidates are isolated with probability at least 1 − exp(−c1 θ4 |T |),
by choosing cB large enough. The other claim of landing in good buckets follows from a standard
argument. In each repetition there are at most θ5 B buckets that are bad, since at most θ5 B
buckets contain noise energy greater than kxT c k22 /(θ5 B) ≤ 160k
kxT c k22 . A standard application
of the Chernoff bound shows that with probability at least 1 − exp(−Ω(|T |)), at least (1 − θ6 )|T |
candidates are hashed into good buckets. Call a pair (candidate, repetition) good if the candidate is
isolated and lands in a good bucket in that repetition. Therefore in each repetition with probability
at least 1 − exp(−Ω(|T |)) there are at least (1 − θ7 )|T | good pairs. Taking a Chernoff bound over R
repetitions, with probability ≥ 1 − exp(−Ω(|T |R)), at least (1 − θ8 )R repetitions contain (1 − θ7 )|T |
good pairs. Condition on this event, we know at least (1 − θ8 )(1 − θ7 )R|T | good (candidate,
repetition) pairs. This implies that at least (1 − θ1 )T candidates are isolated and land in good
buckets in at least (1 − θ2 )R repetitions, provided that 1 − θ1 θ2 ≤ (1 − θ8 )(1 − θ7 ).
Condition on the event above. The actual noise in bucket b in repetition r is
X
Wb,r =
σi,r xi ,
i6∈T
hr (i)=b
and by Markov’s inequality,
n
o
1
2
P Wb,r
≥
kxT c k22 ≤ .
16k
10
Let
Xb,r = 1W 2
c 2
b,r ≤ 16k kxT k2
.
We proved above that with high probability at least (1 − θ)T candidates are isolated and land
in good buckets in at least (1 − θ2 )R repetitions; let T 0 be the set of those (1 − θ1 ) candidates and
for each i ∈ T 0 let Ri be the set of repetitions in which i is isolated and land in good buckets. If
1
Xhr (i),r = 1 for at least 2(1−θ
fraction of r ∈ Ri , then
2)
2
|xi − x̂i | ≤
median Wh2r (i),r
r
≤ quant
r∈Ri
Wh2r (i),r ,
1
2(1 − θ2 )
≤
kxT c k22 ≤
kx−k k22
16k
16k
as desired.
The claimed result would follow from a bound on the probability that there exist at most θ11 k
1
elements i ∈ T 0 such that Xhr (i),r = 1 for at most a 2(1−θ
fraction of r ∈ Ri . Observe that Wb,r are
2)
22
independent since the earlier conditioning has fixed the hash functions. It follows from a Chernoff
bound that
X
|Ri | θ6 δ 2/(θ6 k)
Xhr (i),r ≥
≤
,
p := P
2(1 − θ2 )
8cT
r∈Ri
provided that cR is large enough. In expectation there are |T 0 |p ≤ θ6 kδ 2/(θ6 k) /8 elements in T 0 with
bad estimates. Another Chernoff bound over i ∈ T 0 gives the overall failure probability at most δ.
The total number of missed candidates is at most (θ1 + θ6 )|T |.
These two parts together will give us a weak `2 /`2 system. We conclude with:
Theorem 26. There exists a randomized construction of a matrix M with O(( k + 1 log 1δ ) log nk )
rows, which with probability 1−δ is a (k, ζ, ) weak `2 /`2 system. The column sparsity is O d k1 ln 1δ e log nk .
Proof. The matrix M is the concatenation of the matrix of a (k, ζ/2)-weak identification system
with the (O(k/), δ/2, )-guarantee and the estimation matrix in Lemma 25 (where δ is replaced
with δ/2 and ζ with ζ/2). The weak identification system, by Lemma 24, returns a set T of
candidate indices which misses at most ζ2 k elements of Hk, (x), with probability at least 1 − 2δ .
Then by Lemma 25 the estimation process gives ‘bad’ estimates to at most ζ2 k elements in T with
probability at least 1 − 2δ . Then we truncate x̂ to the largest k coordinates. The claim then
follows from the same argument for [PS12, Lemma 4] or [GLPS17, Theorem 14], with the only
change as follows: if some i ∈ Hk, (x) with a good estimate is replaced with some j 6∈ Hk, (x)
with a good estimate, it then
p follows from the good estimate guarantee that µ ≤ xi ≤ 5µ/4 and
3µ/4 ≤ xj ≤ µ, where µ = /kkx−k k2 . Thus |x̂j − xi | ≤ |x̂j − xj | + |xj − xi | ≤ µ/4 + µ/2 = 3µ/4
and there are at most k such replacements, which introduces squared `2 error into ẑ of at most
k(3µ/4)2 ≤ (9/16)kx−k k22 . Finally, the overall success probability is at least 1 − δ.
5.2
Overall algorithms
In this section we show how to combine different weak `2 /`2 systems with the existing algorithms
presented in Section 3.1, to get our desired algorithm.
Theorem 27.
There exists a recovery system A = (D,R) which satisfies the `2 /`2 guarantee with
− k
3
parameters n, k, , O k log nk , e log3+γ k + ( nk )− log k . Moreover, R runs in O k log3 n time,
where γ > 0 is an arbitrary constant and the constant in the O-notations depends on γ.
Proof. Let ` = bC log log kc for some absolute constant C such that 3` = log3+γ k. We shall pick `
weak systems W1 , . . . , W` using Theorem 26. For 1 ≤ i ≤ ` we pick Wi with parameters ( 3ki , 13 , 2i )
ck
k
and failure probability e− 3i . Then we invoke Theorem 1 for K = 1 log3+γ
and δ = n− log k to obtain
k
a matrix A. Lastly, we obtain a matrix B using Theorem 3 by setting K = k and c = log( nk ). Our
randomized matrix is then the vertical concatenaton of all Wi with A and B. Observe that the
total number of rows is O( k log nk ) and the column sparsity is O(log n · log k).
For i = 1, . . . , ` with x(1) = x we run the algorithm of the weak-system on Wi x(i) to find
a vector r(i) . Then we set x(i+1) ← x(i) − r(i) and observe that Wi x(i+1) = Wi (x(i) − r(i) ) =
Wi x(i) − Wi r(i) , which can be computed in O(k log n) time. After ` iterations we run the algorithm
of ExpanderSketch guaranteed by Theorem 1 on vector x(`+1) with matrix A to get candidate set
T . We then set
23
S←T∪
`
[
supp(r(i) ).
i=1
We observe that previous analyses, such as [GLPS12], immediately imply that
kxS k22 ≤ (1 + )kx−k k22 .
We are now exactly in the situation of estimating the vector xS , which be done using the set query
algorithm from [Pri11] on matrix B with set S to obtain x̂. Then x̂ satisfies
kx − x̂k2 ≤ kx − xS k2 + kxS − x̂k2 ≤ (1 + O())kx−k k2 .
This finishes the proof of our theorem.
We move on with the next theorem, which can be compared with the result in [GNP+ 13].
Theorem 28. There
exists an recovery system
A = (D, R) which satisfies the `2 /`2 guarantee
− k3
k
n
with parameters n, k, , O log k , e log k . Moreover, R runs in O k 2 log1+γ ( nk ) time, for
any constant γ.
Proof. Let ` = dlog3 ke + 1 and we shall pick ` weak systems W1 , . . . , W` . Let C be a constant such
that C log log k = log3 k. For 1 ≤ i ≤ C log log k we pick Wi with parameters ( 3ki , 13 , 2i ) and target
ck
failure probability e− 3i . For C log log k + 1 ≤ i ≤ ` we pick Wi with parameters ( 3ki , 13 , log k ) with
3
target failure probability e−k/ log k .
i
Number of Measurements. For i ≤ C log log k, each Wi takes k( 23 )i log 3kn measurements,
k
1
3i n
which sum up to O k log nk . For i > C log log k, each Wi takes O k ( log
+
)
log
=
k
3i
log2 k
i
O logk 2 k log 3kn measurements (by our choice of C), which sum up to O k log nk . The overall
number of measurements is therefore O k log nk .
i
Column Sparsity. The column sparsity of Wi is O(log nk ) for i ≤ C log log k and O( log33 k log nk )
for i > C log log k. Thus the overall column sparsity is O( logk3 k log nk ).
Runtime. For i = 1, 2, . . . , ` with x(1) = x we run the algorithm of the weak-system on Wi x(i) to
find a vector r(i) . Then we set x(i+1) ← x(i) − r(i) and observe that Wi x(i+1) = Wi (x(i) − r(i) ) =
Wi x(i) − Wi r(i) . A standard analysis as in [GLPS12] gives the desired result.
The next theorem is an improvement on the main result in [GNP+ 13]. Our algorithm achieves
the optimal dependence on and better failure probability, using Theorem 4 for identification and
Lemma 25 for estimation. The improvement primarily comes from two changes: (i) an improved
analysis of Count-Sketch, namely Lemma 25, and (ii) better choice of parameters of the weak
system in each iteration, which are set identically to those in the proof of Theorem 28.
Theorem 29. Suppose that k = nΩ(1)
. There exists a recovery
system A = (D, R) which satisfies
k
k
n
n − log k
the `2 /`2 guaranteee with parameters n, k, , log k , ( k )
. Moreover, R runs in O( 1 k 2 poly(log n))
time.
Proof (Sketch). The proof is the same as before but for identification instead of using a weak system
we invoke Theorem 4 to find a set S such that |Hk/,1 (x) \ S| ≤ (1/2)k/. Accurate estimates of
them can be found using Lemma 25. Repeating the same proof as before, but iterating log(k/)
times we get the desired result.
24
6
Adaptive Sparse Recovery
6.1
1-sparse Adaptive Compressed Sensing
Lemma 30 ([IPW11, Lemma 3.2]). Let x ∈ Rn and suppose that there exists a j with |xj | ≥
2
C Bδ2 kx[n]\{j} k2 for some constant C and parameters B and δ. With two non-adaptive measurements, with probability 1 − δ we can find a set S ⊂ [n] such that (i) j ∈ S; (ii)kxS\{j} k2 ≤
1
n
B kx[n]\{j} k2 and (iii) |S| ≤ 1 + B 2 .
The authors in [IPW11] apply the aforementioned lemma O(log log n) times with appropriate
parameters and obtain an algorithm with O(log log n) measurements. However, their approach
gives only constant success probability. We shall show how to boost the success probability, by
first running a preconditioning algorithm and then applying their Lemma with similar parameters
as they do (not exactly the same though).
We first prove the following lemma, which will serve as a preconditioning.
Lemma 31. Let x ∈ Rn and suppose that there exists a j with |xj | ≥ 5kx[n]\{j} k2 . Then there
exists a scheme that uses Ob,c (log log n) measurements and with probability 1− log1c n finds a set S of
size logn n such that (i) j ∈ S and (ii) kxS\{j} k2 ≤ log1b n kx[n]\{j} k2 , where b, c are absolute constants
which can be made arbitrarily large.
Proof. We assume that the coordinates of x are randomly permuted because we can apply a random
permutation π to the vector x, then find a set S satisfying the conditions of the lemma, and at the
end compute π −1 (i).
Let enc : {0, 1}α log log n → {0, 1}C0 log log n be the encoding function of an error-correcting code
E that corrects a 0.45-fraction of errors, where α is a constant. Such codes exist, see, e.g., [Spi96].
Denote by encb (i) the b-th bit of enc(i). Define trunc : [n] → {0, 1}α log log n to be a function such that
trunc(i) equals the first α log log n bits in the binary representation of i. Let also σ : [n] → {+1, −1}
be a 2-wise independent hash function. Then we perform the following 2·C0 log log n measurements:
X
X
Vb,0 =
σi xi ,
Vb,1 =
σi xi ,
i:encb (trunc(i))=0
i:encb (trunc(i))=1
for all b = 1, . . . , C0 log log n.
We form a binary string r of length C0 log log n as follows: for each b = 1, . . . , C0 log log n, rb = 1
if |Vb,1 | > |Vb,0 | and rb = 0 otherwise. At the end we find in the error-correcting code E the closest
codeword to r, say r0 . Define S to be the set of all i such that trunc(i) = r0 .
We now show correctness. Let J = enc(trunc(j)) and for q = 0, 1 let Ib (q) = {i : encb (trunc(i)) =
q}. Observe that
X
2
E[Vb,J
] = x2j +
b
x2i ,
i∈Ib (Jb )\{j}
and
X
2
E[Vb,J
]=
b
x2i .
i∈Ib (Jb )
Observe that by Markov’s inequality,
X
P
x2i ≥ 5 xIb (Jb )\{j}
i∈Ib (Jb )\{j}
25
2
2
≤
1
5
and
P
2
Vb,J
b
≥ 5 xIb (Jb )
It follows that with probability at least 3/5,
√
|Vb,Jb | ≥ |xj | − 5 xIb (Jb )\{j}
2
≥
2
2
√
1
≤ .
5
5 xIb (Jb )
2
≥ Vb,Jb .
This means that, by a Chernoff bound, the string r will agree with Jb at least in a 0.55-fraction
1
of positions with probability at least 1 − logγC
, where γ is an absolute constant. Applying
0 n
the decoding algorithm of E we can recover trunc(j). Observe now that our set S has lognα n
coordinates. Since we assumed that the order of elements in x is random, we have that EkxS\{j} k22 =
Ekxi6=j:trunc(i)=trunc(j) k22 = log1α n kx[n]\{j} k22 . It follows from Markov’s inequality that kxS\{j} k22 ≥
1
1
1
a/2 kx[n]\{j} k2 with probability at most
α/2 . Hence the overall failure probability is logγC0 n +
log
n
1
logα/2 n
log
n
. This concludes the proof of the lemma.
We are now ready to state and prove the main theorem of this section.
Theorem 32. Let x ∈ Rn and suppose that there exists j such that |xj | ≥ Ckx−k k2 . Then, there
exists an algorithm OneSparseRecovery that performs O(log log n) measurements in O(log log n)
rounds, and finds j with probability 1 − log1c n , where c > 0 is some absolute constant.
Proof. We first apply Lemma 31 on vector x and obtain a set S and then follow the approach in
[IPW11] for xS . Consider now the following sequence of parameters:
1
.
logc n
Let r be the first index such that Br ≥ n. For each i = 1, . . . , r, we apply Lemma 30 to xSi−1
with parameters B = Bi and δ = δi and obtain a set Si , where S0 is taken to be the set S. It is easy
B2
to see that r = O(log log n). We will inductively prove that at all steps |xj | ≥ δ2i kxSi k2 . The base
case follows immediately from Lemma 31. By the induction step, at step i we find a set Si such
δi2
B2
1
1
kxSi−1 k2 ≤ Bi−1
|xj | or, equivalently, |xj | ≥ δ2i kxSi k2 . After r iterations,
that kxSi k2 ≤ Bi−1
B2
3
B0 = 2,
2
Bi = Bi−1
,
δ=
i−1
|Sr | ≤ 1 + n/Br2 < 2 and thus we have uniquely identified j.
The overall failure probability is at most log1c n +rδ ≤ 1c0
log
n
for some c0 > 0. The overall number
of measurements is O(log log n), since in every round we use 2 measurements, plus O(log log n)
measurements for the application of Lemma 31 at the very beginning.
6.2
6.2.1
k-sparse recovery
Results
Theorem 33 (`2 /`2 in whole regime of parameters). Let x ∈ Rn and γ > 0 be a constant. There
exists an algorithm that performs O((k/) log log(n/k)) adaptive linear measurements on x in
O(log∗ k · log log(n/k)) rounds, and finds a vector x̂ ∈ Rn such that kx − x̂k22 ≤ (1 + )kx−k k22 . The
algorithm fails with probability at most exp(−k 1−γ ).
Theorem 34 (`∞ /`2 in low sparsity regime). Let x ∈ Rn and parameters k, such that k/ ≤
c poly(log n). There exists an algorithm that performs O((k/) log log n) adaptive linear measurements on x in O(log log n) rounds, and finds a vector x̂ ∈ Rn such that kx − x̂k∞ ≤ k kx−k/ k2 .
The algorithm fails with probability at most 1/ poly(log n).
26
We note that the low sparsity regime result not only improves failure probability but also gives
a stronger guarantee (`∞ /`2 instead of `2 /`2 ) in even fewer rounds.
6.2.2
Whole Regime of Parameters
We first remind the reader the following additive form of the Chernoff bound.
Lemma 35 (Chernoff bound, additive form). Suppose that X1 , . . . , XN are i.i.d. random variables
taking values in {0, 1}. Let p = E[Xi ] and q > p. Then
( n
)
X
P
Xi ≥ N q ≤ exp(−D(qkp)N ),
i=1
where
D(qkp) = −q log
1−q
q
− (1 − q) log
.
p
1−p
Let γ 1 be an absolute constant and T be a constant depending only on γ. Let a sequence
γ0 , γ1 , . . . , γT be a decreasing arithmetic progression such that γ0 = 0.9 and γTγ = γ. Let also C
be an absolute constant.
The algorithm finds a constant fraction of heavy coordinates of x by hashing all coordinates
to a number of buckets. Then it observes their values, subtracts them from x and iterates by
changing (k, , B) parameters. The way these parameters change is crucial in obtaining the desired
low failure probability. We split the algorithm in three different phases. Let φ ∈ {1, 2, 3} be
the variable corresponding to the phase the algorithm is in. The overall algorithm is presented in
(φ)
(φ)
Algorithm 2. In the r-th round of phase φ, we pick for each r ∈ [Rr ] an O(kr )-wise hash functions
(φ)
(φ)
hr,j : [n] → [C 0 kr /(φ) ], where C 0 is a large absolute constant, and run the 1-sparse recovery
algorithm in each bucket induced by the hash functions, as illustrated in HashAndRecover
function below.
function HashAndRecover(x, k, , R)
for j = 1 to R do
0
Pick an
S O(k)-wise independent hash function hj : [n] → [C k/]
Sj ← l∈[C 0 k/] OneSparseRecovery(xh−1 (l) )
j
end forS
return j∈[R] Sj
end function
We need the following lemma.
Lemma 36. In each phase and every round r, it holds with probability at least 1 − exp(−Ω(k 1−γ ))
(φ)
that Hk(φ) ,(φ) (xJ¯) ≤ kr+1 .
r
r
Proof. Define the following events.
(1)
• Bj,r,l : |{i ∈ [n] : hr,j (i) = l}| > 1
(2)
• Bj,r,l : kx{i∈[n]:hr,j (i)=l} k22 ≥
(1)
C 00 r
2
kr kxC 0 tail(kr ) k2
(2)
• Bi,r,j : ∃j ∈ [C 0 Bkrr ] : Bj,r,hr,j (i) or Bj,r,hr,j (i)
27
Algorithm 2 Adaptive sparse recovery algorithm
J ←∅
for r = 0 to log log k − 1 do
(0)
kr ← 2kr
(0)
r ← · ( 34 )r
(0)
Rr ← C
(0) (0)
(0)
J ← J ∪ HashAndRecover(xJ¯, 1, kr , r , Rr )
end for
for r = 0 to log∗ k γ do
(2)
k
kr ← (2↑↑r)·log
k
(1)
r ←
(2)
Rr ← C log k
(1) (2)
(2)
J ← J ∪ HashAndRecover(xJ¯, 2, kr , r , Rr )
end for
for r = 0 to T do
(3)
kr ← k γr
(3)
r ←
(3)
Rr ← Ck 1−γr
(3) (3)
(3)
J ← J ∪ HashAndRecover(xJ¯, 3, kr , r , Rr )
end for
return xJ
We define indicator variables Xi,r,j for the event Bi,r,j . Observe that Xi,r,j are negatively dependent.
We split the analysis in the three phases, depending on the round r. It is essentially an inductive
argument on r. The base case in each phase is easy to verify and we shall show only the inductive
step below.
• Phase 1 (0 ≤ r ≤ log log k − 1): Fix a hash function hr,j . A standard argument shows that
EXi,r,j ≤ C10 . The probability that event Bi,j,r holds equals
kr Rr kr+1
P {∃S ⊆ Hkr ,r (x), |S| ≥ kr+1 : Xi,r,j = 0, ∀i ∈ S, j ∈ Rr } ≤
C0
kr+1
k
≤ exp −c1 r
2
k
≤ exp −c1
log k
• Phase 2 (0 ≤ r ≤ log∗ k γ ): For a coordiante i,
p = P{Xi,r,j = 0 ∀j ∈ [Rr ]} =
Rr
C0
After applying a negative association argument, more than kr+1 heavy hitters are unrecovered
with probability, by the additive form of Chernoff bound, at most (note that kr+1 /kr = 2 ↑↑
28
r/2 ↑↑ (r + 1) ≥ 1/k γ > 3p)
!
!
kr+1
1−p
kr+1
kr+1
p · kr ≤ exp −kr+1 log
·
exp −D
− kr log 1 −
kr
kr p 1 − kr+1
kr
kr
1−p
kr
≤ exp −kr+1 log
+ kr+1
+ kr+1 log
p
kr+1
1
≤ exp −c2 kr+1 log
p
1
0 1−γ
.
≤ exp −c2 k
log
• Phase 3 (0 ≤ r ≤ T ): In this phase kr = k γr . When r < T , similarly to Phase 2, we have
that more than kr+1 heavy hitters are unrecovered with probability at most
γ −γ
1
1
0 1− 0T
exp −c3 kr+1 log
≤ exp −c3 k
log
.
p
In the last step, i.e. r = T , for a coordinate i,
P {Xi,r,j = 0 ∀j ∈ [Rr ]} ≤ exp(−c4 k 1−γ ).
This allows us to take a union bound over all i ∈ HkT , (xr−1 ), so in this step we will recover
all of them with probability 1 − exp(−c04 k 1−γ ).
The proof of the lemma is complete.
6.2.3
Low-Sparsity Regime
We will need the following lemma, which is standard in the sparse recovery and streaming algorithms
literature.
Lemma 37. Let x ∈ Rn and F = {F1 , . . . , FU } be a partition of [n]. For a set S ⊆ [n] define F (S) =
S
i∈Fj Fj . There exists an algorithm that performs O((k/) log |U |) non-adaptive measurements and
with probability 1 − |U |−c finds a set T ⊆ [U ] of size O(k) such that F (Hk, (x)) ⊆ T .
Proof Sketch. For each r = 1, . . . , log |U | we pick a 2-wise independent hash function hr : [U ] →
[Ck], for some absolute constant C and a 2-wise independent hash function σr : [n] → {+1, −1}.
Then, for every r ∈ [log |U |] and j ∈ [Ck] we perform measurement
yj,r =
X
X
σi,r xi .
t∈[U ]:hr (t)=j i∈Ft
For every j ∈ U we compute
ẑj = median yhr (j),r
1≤r≤log k
and find the largest C0 k indices with the biggest ẑ values in magnitude, forming a set T . We then
output T .
We are now ready to prove our main result in the low-sparsity regime, Theorem 34.
29
Proof of Theorem 34. Pick a hash function h : [n] → [logc0 n], where c0 is a large absolute constant.
1
−1
Observe that with probability 1 − poly(log
n) every h (p) contains at most 1 element of Hk, (x). Let
Fgood = {h−1 (p) : |h−1 (p) ∩ Hk, (x)| = 1}. Then we can invoke Lemma 37 and obtain a set S of size
1
O(k) which with probability 1 − poly(log
n) is a superset of Fgood . The number of measurements we
1
need is O( k · log log n). We then run the 1-sparse recovery algorithm guaranteed by Theorem 32
in each Fp ∈ Fgood , for a total of O( 1 k · log log n) measurements in O(log log n) rounds. Since
1
every 1-sparse recovery routine succeeds independently with probability 1 − poly(log
n) we get that
all of them succeed with the desired probability and hence we can obtain a set S containing all
elements of Hk, (x). By observing them directly in another round, we can obtain their values and
thus achieve the `∞ /`2 guarantee.
7
Spiked-Covariance Model
In the spiked covariance model, the signal x is subject to the following distribution: we choose k
coordinates uniformly at random, say,
ik . First construct a vector y ∈ Rn , in which each yki
pi1 , . . . , p
is a uniform Bernoulli variable on {− /k, + /k} and these k coordinate values are independent
of each other. Then let z ∼ N (0, n1 In ) and set x = y + z. We now present a non-adaptive algorithm
(although the runtime is slow) that uses O((k/) log(n/k) + (1/) log(1/δ)) measurements and a
matching lower bound.
Theorem 38. Assume that (k/) log(1/δ) ≤ βn, where β ∈ (0, 1) is a constant.
There exists an
k
n
1
1
`2 /`2 algorithm for the spiked-covariance model that uses O log k + log δ measurements and
succeeds with probability ≥ 1 − δ. Here the randomness is over both the signal and the algorithm.
q
log(1/δ)
Proof. First note that the maximum magnitude in z is O
with probability ≥ 1 − δ,
n
p
which is smaller than /k given the assumption on k. Condition on this event. Furthermore,
n−k
2
with probability ≥ 1 − e−Ω(n) , it holds that (1 − η) n−k
n ≤ kx−k k2 ≤ (1 + η) n . Further condition
on this event.
We essentially repeat Lemma 25 with T = [n]. That is, we estimate
q each coordinate, discard
p
p
the estimates outside the range [(1 − γ) k , (1 + γ) k ], where γ = α (1 + η)(1 − nk ), and retain
only the top k coordinates. To accommodate the larger size of T , the number of repetitions needs
1
1
k
n
1
1
to be R = cR (log n
k + k log δ ), so the total number of measurements is O log k + log δ as
claimed.
Next we show correctness. With probability ≥ 1 − δ, the top k items we retain
p contain at least
(1−θ)k elements of Hk (x), and the estimation error of each of them is at most β /kkx−k k2 . Hence
these elements survive the thresholding. In total these well-recovered heavy hitters contribute at
most β 2 kx−k k22 to the approximation error. For the remaining heavy hitters,pthey could be (i)
unrecovered, (ii) replaced by
p spurious ones with magnitude bounded by (1 + γ) /k, or (iii) have
estimation error at most γ /k. Hence they contribute to the residual energy (squared `2 norm)
at most (1 + γ)2 θ in total. Therefore we conclude that
kx̂ − xk22 ≤ 1 + (β 2 + (1 + γ)2 θ) kx−k k22 .
Remark 39. The algorithm above runs in time Õ(n). Alternatively, when k = nΩ(1) , we can
invoke Theorem 4 to identify a constant fraction
of the heavy hitters and then estimate them as
before. Overall it uses O k log nk + 1 log 1δ measurements and succeeds with probability at least
1 − δ. The error guarantee follows similarly as in the proof above. This alternative algorithm uses
30
the optimal number of measurements when is a constant (and only slightly more measurements
in the general case) but runs significantly faster in time O(k 1+α poly(log n)).
We prove a matching lower bound to conclude this section.
Theorem
40. Suppose
that δ < δ0 for a sufficiently small absolute constant δ0 ∈ (0, 1) and
n ≥ 64−1 log(6/δ) . Then any `2 /`2 -algorithm that solves with probability ≥ 1 − δ the `2 /`2
problem in the spiked-covariance model must use Ω(−1 log(1/δ)) measurements.
Proof. The lower bound proved in Theorem 23 corresponds to the case of k = 1 in the spiked
covariance model. We can follow exactly the same notation and proof for the case of k = 1 for
general k. Doing so, we arrive atPthe point that
p we need to bound the total variation distance
between N (0, Ir ), and N (0, Ir ) + i∈supp(y) (3 n/k)Si σi , where σi are random signs. Note that
q
p
P
≤
S
σ
10(k/n)kSk2F = 10kr/n with constant probability, and thus one must
i∈supp(y) i i
2
√
distinguish N (0, Ir ) and N (0, Ir )+O( r)v for a unit vector v with probability 1−Θ(δ). Now we can
rotate v to the first standard unit vector e1 as before, to conclude that the same Ω((1/) log(1/δ))
lower bound continues to hold for k > 1.
References
[BIPW10] Khanh Do Ba, Piotr Indyk, Eric Price, and David P. Woodruff. Lower bounds for
sparse recovery. In Proceedings of the Twenty-First Annual ACM-SIAM Symposium
on Discrete Algorithms, SODA 2010, Austin, Texas, USA, January 17-19, 2010, pages
1190–1197, 2010.
[CCFC04] Moses Charikar, Kevin Chen, and Martin Farach-Colton. Finding frequent items in
data streams. Theoretical Computer Science, 312(1):3–15, 2004.
[CH08]
Graham Cormode and Marios Hadjieleftheriou. Finding frequent items in data streams.
Proceedings of the VLDB Endowment, 1(2):1530–1541, 2008.
[CM05]
Graham Cormode and Shan Muthukrishnan. An improved data stream summary: the
count-min sketch and its applications. Journal of Algorithms, 55(1):58–75, 2005.
[CRPW10] V. Chandrasekaran, B. Recht, P. A. Parrilo, and A. S. Willsky. The Convex Geometry
of Linear Inverse Problems. ArXiv e-prints, 2010.
[Gan08]
Sumit Ganguly. Data stream algorithms via expander graphs. In International Symposium on Algorithms and Computation, pages 52–63. Springer, 2008.
[GG11]
Subhankar Ghosh and Larry Goldstein. Concentration of measures via size-biased
couplings. Probability Theory and Related Fields, 149(1):271–278, 2011.
[GLPS12]
Anna C Gilbert, Yi Li, Ely Porat, and Martin J Strauss. Approximate sparse recovery:
optimizing time and measurements. SIAM Journal on Computing, 41(2):436–453, 2012.
[GLPS17]
Anna C. Gilbert, Yi Li, Ely Porat, and Martin J. Strauss. For-all sparse recovery in
near-optimal time. ACM Trans. Algorithms, 13(3):32:1–32:26, 2017.
[GM06]
S. Ganguly and A. Majumder. CR-precis: A deterministic summary structure for
update data streams. eprint arXiv:cs/0609032, 2006.
31
[GNP+ 13] Anna C Gilbert, Hung Q Ngo, Ely Porat, Atri Rudra, and Martin J Strauss. `2 /`2 foreach sparse recovery with low risk. In International Colloquium on Automata, Languages, and Programming, pages 461–472. Springer, 2013.
[Gor41]
Robert D. Gordon. Values of mills’ ratio of area to bounding ordinate and of the normal
probability integral for large values of the argument. Ann. Math. Statist., 12(3):364–
366, 09 1941.
[GSTV06] A. C. Gilbert, M. J. Strauss, J. A. Tropp, and R. Vershynin. Algorithmic linear dimension reduction in the `1 norm for sparse vectors, volume 3, pages 1411–1418. University
of Illinois at Urbana-Champaign, Coordinated Science Laboratory and Department of
Computer and Electrical Engineering, 2006.
[GUV09]
Venkatesan Guruswami, Christopher Umans, and Salil Vadhan. Unbalanced expanders
and randomness extractors from parvaresh–vardy codes. Journal of the ACM (JACM),
56(4):20, 2009.
[IPW11]
Piotr Indyk, Eric Price, and David P Woodruff. On the power of adaptivity in sparse
recovery. In Foundations of Computer Science (FOCS), 2011 IEEE 52nd Annual Symposium on, pages 285–294. IEEE, 2011.
[JST11]
Hossein Jowhari, Mert Sağlam, and Gábor Tardos. Tight bounds for lp samplers,
finding duplicates in streams, and related problems. In Proceedings of the thirtieth
ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, pages
49–58. ACM, 2011.
[KNPW11] Daniel M Kane, Jelani Nelson, Ely Porat, and David P Woodruff. Fast moment estimation in data streams in optimal space. In Proceedings of the forty-third annual ACM
symposium on Theory of computing, pages 745–754. ACM, 2011.
[LNNT16] Kasper Green Larsen, Jelani Nelson, Huy L Nguyen, and Mikkel Thorup. Heavy hitters
via cluster-preserving clustering. In Foundations of Computer Science (FOCS), 2016
IEEE 57th Annual Symposium on, pages 61–70. IEEE, 2016.
[LT91]
Michel Ledoux and Michel Talagrand. Probability in Banach spaces. Springer-Verlag,
Berlin, 1991.
[McG07]
Andrew McGregor. Open problems in data streams and related topics: Iitk workshop
on algorithms for data streams, 2006, 2007.
[NNW12]
Jelani Nelson, Huy L. Nguyên, and David P. Woodruff. On deterministic sketching
and streaming for sparse recovery and norm estimation. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques - 15th International Workshop, APPROX 2012, and 16th International Workshop, RANDOM 2012,
Cambridge, MA, USA, August 15-17, 2012. Proceedings, pages 627–638, 2012.
[NPR11]
Hung Q Ngo, Ely Porat, and Atri Rudra. Efficiently decodable error-correcting list disjunct matrices and applications. In International Colloquium on Automata, Languages,
and Programming, pages 557–568. Springer, 2011.
[Pol]
D. Pollard. “Chapter 3: Total Variation Distance Between Measures”. http://www.
stat.yale.edu/~pollard/Courses/607.spring05/handouts/Totalvar.pdf.
32
[Pri11]
Eric Price. Efficient sketches for the set query problem. In Proceedings of the twentysecond annual ACM-SIAM symposium on Discrete Algorithms, pages 41–56. Society
for Industrial and Applied Mathematics, 2011.
[PS12]
Ely Porat and Martin J. Strauss. Sublinear time, measurement-optimal, sparse recovery
for all. In Proceedings of the Twenty-third Annual ACM-SIAM Symposium on Discrete
Algorithms, SODA ’12, pages 1215–1227, 2012.
[PW11]
E. Price and D. P. Woodruff. (1 + )-approximate sparse recovery. In 2011 IEEE 52nd
Annual Symposium on Foundations of Computer Science, pages 295–304, Oct 2011.
[Spi96]
D. A. Spielman. Linear-time encodable and decodable error-correcting codes. IEEE
Transactions on Information Theory, 42(6):1723–1731, 1996.
[V+ 12]
Salil P Vadhan et al. Pseudorandomness. Foundations and Trends R in Theoretical
Computer Science, 7(1–3):1–336, 2012.
33
| 8cs.DS
|
Impact of different time series aggregation methods on optimal energy system design
Leander Kotzura,∗, Peter Markewitza , Martin Robiniusa , Detlef Stoltena,b
a
Institute of Electrochemical Process Engineering (IEK-3), Forschungszentrum Jülich GmbH, Wilhelm-Johnen-Str., 52428 Jülich,
Germany
b Chair for Fuel Cells, RWTH Aachen University, c/o Institute of Electrochemical Process Engineering (IEK-3), Forschungszentrum Jülich
GmbH, Wilhelm-Johnen-Str., 52428 Jülich, Germany
arXiv:1708.00420v1 [math.OC] 1 Aug 2017
Abstract
Modelling renewable energy systems is a computationally-demanding task due to the high fluctuation of supply and
demand time series. To reduce the scale of these, this paper discusses different methods for their aggregation into typical
periods. Each aggregation method is applied to a different type of energy system model, making the methods fairly
incomparable.
To overcome this, the different aggregation methods are first extended so that they can be applied to all types
of multidimensional time series and then compared by applying them to different energy system configurations and
analyzing their impact on the cost optimal design.
It was found that regardless of the method, time series aggregation allows for significantly reduced computational
resources. Nevertheless, averaged values lead to underestimation of the real system cost in comparison to the use of
representative periods from the original time series. The aggregation method itself e.g., k-means clustering plays a
minor role. More significant is the system considered: Energy systems utilizing centralized resources require fewer typical
periods for a feasible system design in comparison to systems with a higher share of renewable feed-in. Furthermore, for
energy systems based on seasonal storage, currently existing models integration of typical periods is not suitable.
Keywords: Energy systems, Renewable energies, Mixed integer linear programming, Typical periods, Cluster analysis,
Extreme periods, Time-series aggregation
1. Introduction
Developing an energy system design that minimizes
costs and environmental impact is a complex task due to
the spatial and temporal gap between energy production
and demand. In consequence, optimization algorithms are
required for solving these design problems [1, 2, 3, 4, 5, 6,
7, 8, 9, 10, 11].
However, the algorithms used hitherto are computationally demanding: The size of the input data directly
influences that of the related optimization problem, and
with it the requirement for processing resources. For this
reason, it is often necessary to simplify the design problem
in advance.
Therefore, different options for complexity reduction
exist and include: Spatial aggregation which reduces the
number of nodes in an energy system network [12]; simplifying the technology models by reducing nonlinearities or
discontinuities so as to avoid non-convexity of the program
[13, 14]; and temporal aggregation, which creates typical
periods representing the original input time series.
The creation of recurring periods is popular because
of the existing patterns in the hourly, daily and seasonal
variation for the majority of design relevant time series.
∗ Corresponding
Therefore, it is reasonable to reduce redundant data until
the minimal representative data set required for the problem is reached. Figure 1 visualizes this redundancy by
showing the result of a Fast Fourier Transformation (FFT)
of different time series that are typically required for an
energy system design. The frequencies with the highest
amplitudes are highlighted and are, as anticipated, the
daily and annual variations.
For this reason, many different methods for the selection of typical periods have been presented. Aside from
custom exact optimization methods [17, 18], and graphical methods [19], the majority use heuristic methods or
greedy clustering algorithms for the aggregation of typical
periods. Creating representative days by averaging time
series, for example over a type of day defined by month
or weekday, has been popular [20, 21, 22, 23]. [24] refers
to it as time-chronological averaging. Recent attempts use
the k-means clustering [25, 26, 27, 28], hierarchical clustering [29, 30], or k-medoids clustering - either based on a
greedy algorithm [31, 32] or an exact solution of a MILP
[33, 34] - for the selection of typical periods. Nevertheless, each method is applied to a different system and it
is difficult to identify which is the most suitable. While
Schuetz et al. [35] compare methods for building energy
systems, they note that future research should focus on
author. Email: [email protected]
Preprint submitted to arXiv.org
August 2, 2017
Nomenclature
Variables
f
Index of the Transformer class
δ
Binary variable determining the existance of a
technology
n
Index of the Collector class
q
Index of the Source/Sink class
µ
Representative values of a typical period
s
Index of the Storage class
D
Scaling of a device
a
Attribute represented by a time series
E
Energy flow between two components
SOC
State of charge
d
Considered device or technology
y
Binary variable determining if the candidate is
a cluster center
g
Time step index inside a period
i
Candidate period index
k
Typical period index
t
Time step index of the full series
z
Binary variable determining candidate cluster
assignment
Parameters
Abbreviations
∆t
Duration of a single time step
CRF
Capital Recovery Factor
η
Efficiency
LB
Lower bound
τ
Lifetime
UB
Upper bound
C
Set of periods inside a cluster
L
Set of component connections
N
Size of an index set
x
Normalized candidate value
CAPEX Specific capital expenditure
Combined Heat and Power plant
GHI
Global Horizontal Irradiance
OPEX Specific fix operational expenditure
RMSE Root Mean Squared Error
Subscripts and sets
CHP
Energy type
WACC Weighted Average Cost of Capital
the appropriateness of clustering algorithms for different
design applications. Moreover, the period length should
be varied so as to assess the impact of storage effects.
A further difficulty is that the system considered determines the minimal required dataset. For renewable energy
systems, a higher resolution of the input times-series is required than with their fossil counterparts [36]. For conventional system design, it could be sufficient to reduce the
dataset to a few time steps [27], while for a storage based
system design different typical weeks are required [37, 30].
In summary, the following open research questions present
themselves:
• Which time-series aggregation method is best suited
for which energy system design application?
• What is the minimum number of aggregated time
steps to model such a system?
tering, k-medoids clustering and hierarchical clustering as
aggregation method are presented in section 2, where the
possibilities of adding extreme periods are also discussed.
In section 3, the aggregation methods are used to select
four typical days of different time series that could be relevant for an energy system design. The aggregated profiles
are then graphically analyzed and through accuracy indicators. In the following, the different methods are applied
in section 4 to three design optimization problems of a
heat and electricity supply system:
1. A cogeneration unit with a heat storage as benchmark system
2. A residential system based, amongst other elements,
on photovoltaics and a heatpump
3. An island system with a high share of renewables
with the support of different storage technologies
To validate the methods, the results for different numbers
of typical periods are compared to the optimal solution of
the original optimization problems with the full time series
and analysed in terms of their accuracy and computational
load. Sections 5, 6 and 7 summarize, draw the principal
• What is an appropriate period length - typical days
or typical weeks?
To answer these, this paper is structured as follows:
First different deterministic methods including k-mean clus2
Amplitude
Wind [m/s]
Amplitude
Temp. [ C]
Amplitude
GHI [W/m2]
0
24 h
8758 h
102
100
10
2
Figure 2: Different steps in the procedure of time series aggregation.
8758 h
101
10
1
10
3
8758 h
100
10
2
10
4
24 h
2.1. Pre-processing the time-series
First, the input data is normalized in terms to evalu0
ate all time series on the same scale. Each time series xa
represents an attribute a ∈ {1, ..., Na } e.g., an electrical
demand profile or the measured solar horizontal irradiance
at a certain location. The time series itself consists of raw
0
data points xa,t where t ∈ {1, ..., Nt } constitutes a single time step. Different possibilities for normalization and
standartization are presented by Rager and Maréchal [31].
In this work, the normalized time series xa are calculated
as follow
584 h
Amplitude
E-Load
House [kW]
0
100
10
2
10
4
24 h
8758 h
xa =
0
xa − min xa
max x0a − min x0a
∀
a ∈ {1, ..., Na },
(1)
which results in time series on the same scale xa,t ∈ [0, 1].
10
4
10
3
10
2
10
1
Figure 1: Fast Fourier Transformation of the Global Horizontal Irradiance (GHI) , the temperature and the wind speed of a test reference
year (Location: Bad Marienberg, Germany) [15] and a representative
electrical load profile of a residential building (Profile 1) [16]
conclusions and give an outlook for further research questions.
All methods introduced are published in the Python
package tsam - Time Series Aggregation Module and can
be easily applied and extended.
Figure 3: Illustration of the nomenclature for a time series a = 1
with two candidate periods i = 1, 2 for a whole period.
2. Time series aggregation methods
For the aggregation of typical periods, the scaled timeThe aim of time series aggregation is to merge a set
series are arranged into the candidate periods considered
of periods into groups such that the group members - the
i ∈ {1, ..., Ni } , each consisting of the same number of time
original periods - are as similar as possible. The group is
steps g ∈ {1, ..., Ng } with Na attributes. The nomenclathen represented by a single period. The grouping of timeture is illustrated in Figure 3 where the normalized timeseries is based in the most of the methods on a distance
series of attribute a = 1 is shown for two periods i = 1, 2,
measure of the attributes between each group member.
which were originally successive.
For an accurate grouping, the raw input data must first
This reordering results in a matrix L in which the numbe pre-processed into the right format (Section 2.1). On
ber of columns is defined by the multiple of the number of
the basis of this, an aggregation method is then applied
period time steps Ng and number of attributes Na . The
to create the groups (Section 2.2). In terms of achieving
number of rows corresponds to the number of periods Ni :
a feasible system design, different variants of integrating
extreme periods can be included afterwards (Section 2.3).
x1,1,1 · · · x1,Ng ,1
x2,1,1 · · · xNa ,Ng ,1
Finally, the aggregated time series must be scaled back
..
..
..
..
..
..
(2)
.
.
such that their average values fit the average values of theL =
.
.
.
.
original time series (Section 2.4).
x1,1,Ni · · · x1,Ng ,Ni x2,1,Ni · · · xNa ,Ng ,Ni
The general procedure for time series aggregation apA single row, described by vector x i represents a canplied in this paper is shown in Figure 2.
didate period, also referred to as an observation point.
3
For example: Pre-processing two hourly time series
over a year (Nt = 8760), like the mentioned electrical demand profile and the measured solar horizontal irradiance
(Na = 2), for the application of a typical day approach
(Ng = 24) would yield 365 days (Ni = 365) and a 365-by48 matrix.
In the case that the raw time series length Nt is no
integer multiple of the period length Ng and the number
of periods Ni , the full time-series xt,a must be cut off or
extended, such that its length becomes an integer multiple
of the number of steps inside a candidate period Ng and
the number of periods Ni .
1. Calculate the integer divisor of the number of candidate periods to the number of representative periods
Ni
bN
c, which represents the size of each.
k
Ni
c candidate periods in their original order
2. Relate b N
k
to the cluster period Ck , except the last cluster CNk
which receives the remaining candidate periods.
3. Calculate the average period profile of each group as:
µa,g,k =
1 X
xa,g,i
|Ck |
∀ a, k, g
(3)
i∈Ck
The advantage of this method is, that it is easy applicable and the resulting typical periods have a clear order.
Nevertheless, the aggregation is only based on the original
sequence of the periods and not on the similarity of the
group members or candidate periods.
2.2. Time-series aggregation methods
Based on the matrix introduced in equation 2, different aggregation methods can be applied to group the Ni
independent candidate periods into clusters defined as Ck .
On the basis of these groups, Nk typical or representative
periods are derived.
2.2.2. k-means clustering
The k-means clustering algorithm calculates also the
average - the mean - profile of a group. The significant
difference to the averaging method is that the k-means
algorithm creates the clusters in order to minimize the
squared error between the empirical mean of a cluster and
all candidates in the cluster [38], and not by their original positional appearance in the year. Different error or
distance functions are possible. In this paper, we consider
the squared error, also known as the Euclidean distance,
which is defined as follows:
min
Ng Na
Nk X
Ni
X
X
X
2
(xa,g,i − µa,g,k ) × zi,k
k=1 i=1
(4)
g=1 a=1
or in the vector form as
min
Nk X
Ni
X
xi − µ k ||2 × zi,k
||x
(5)
k=1 i=1
where zi,k is a binary variable that is equal to 1 if the
candidate i is assigned to cluster k. In order to make sure
that each candidate is assigned to a cluster, the following
constraint is added:
Figure 4: Comparison of cluster methods: k-means with Nk = 2,
k-medoids with Nk = 2, and hierarchical cluster up to Nk = 2.
The representative data point itself of a typical period k in period step g of attribute a is defined as µa,k,g .
The representative candidate is stated as vector µ k . In all
reviewed methods, the number of representative periods
Nk must be defined a priori and is validated following the
aggregation procedure. The applied cluster methods are
visually depicted in Figure 4.
Nk
X
zi,k = 1 ∀ i
(6)
k=1
.
This defines a mixed-integer nonlinear program (MINLP)
which is non-convex and difficult to solve. In consequence,
the k-means method is implemented as a greedy algorithm
that converges to a local minimum. Nevertheless, it is
fairly fast, which is why in practice it is run with many
different starting points in order to determine the global
minimum. The main steps of the algorithm itself are as
follows:
2.2.1. Averaging periods
Perhaps the most obvious option is to take the averaged
values of different periods. Mavrotas et al. [20] do this for
each month in a year for the heating, cooling and electricity
load, resulting in 12 typical days.
A more generic formulation that divides the original
time series into Nk parts based on their order is introduced
as follow:
1. Randomly or heuristically select an initial partition
with Nk clusters and calculate the cluster centers.
4
2. Assign each candidate to its closest cluster center.
3. Compute new cluster centers based on all candidates
belonging to the cluster.
4. If no convergence criterion is met, return to step 2.
Nk clusters are left, which is also shown in Figure 4. The
procedure pf hierarchical clustering is as follows:
1. Set each candidate as an own cluster.
2. Determine the centroid or mean vector of each cluster.
3. Calculate the Euclidian distances to the other cluster
centers
4. Merge the two clusters with the lowest distance. If
the number of cluster is still greater than Nk , return
to step 2.
A more detailed discussion of k-means clustering for
the selection of typical periods can be found in Fazlollahi
et al. [39].
2.2.3. Exact k-medoids clustering
The k-medoids clustering is an adaption of the k-means
algorithm. Instead of calculating the mean as a cluster
center, a representative candidate - the medoid - is chosen
for µk . Figure 4 portrays this difference.
Different algorithms for the determination of the clusters also exist. While Rager and Maréchal [31], Stadler
et al. [32] use a greedy optimization algorithm called Partitioning Around Medoids (PAM) [40], the problem can
also be stated as a Mixed-Integer Linear Program (MILP)
[33, 35], for which good global solving algorithms exist.
First, the distance between each candidate is calculated
as
d(i, j) =
Ng Na
X
X
(xa,g,i − xa,g,j )
2
∀
Based on the resulting clusters, representative periods for
each of these must be derived. In this work, we therefore chose the medoid, respectively the candidate period
in the cluster with the smallest distance to all other cluster candidates. In consequence, the method considered is
essentially a greedy algorithm of the k-medoids algorithm
presented in the previous subchapter.
In comparison to the before presented k-means algorithm, the advantage of this algorithm is that it is independent from initial starting points. For this reason, the
clusters are easily reproducible.
A more detailed introduction of hierarchical clustering
for time series aggregation can be found in Nahmmacher
et al. [30].
i, j ∈ 1, ..., Ni (7)
g=1 a=1
Then, following MILP can be stated
min
Ni
Ni X
X
2.3. Integration of extreme periods
d(i, j) × zi,j
The methods introduced for time series aggregation
have the disadvantage of potentially cut off so called peak
periods, because such periods are not representative for a
whole group or cluster of periods. Rather, they are periods in which the considered time series have design relevant extrema, e.g. peak heat demand. These are important because an accurate energy system design should be
able to meet these demands. Therefore, Stadler et al. [3],
Domı́nguez-Muñoz et al. [33], Fazlollahi et al. [39], Bahl
et al. [27] all manually add so called peak periods to the
aggregated periods.
For a comparison of a good integration of these periods, different variants of integration extreme periods are
presented in Figure 5. The variants are as follows:
(8)
i=1 j=1
subject to
Ni
X
zi,j = 1 ∀
j ∈ 1, ..., Ni
(9)
∀ i, j ∈ 1, ..., Ni
(10)
j=1
zi,j ≤ yi
Ni
X
yi = Nk
(11)
i=1
where both zi,j and yi are binary variables.
This optimization problem can be solved to a global
optimum, but has the disadvantage of requiring a high
computational load. The computational effort is directly
correlated to the number of candidate periods. On the
other hand, the impact of the number of attributes and
number of time steps in each period is negligibly small,
because the distance between the candidates is calculated
a priori.
1. None - no integration of extreme periods at all.
2. Append - add the extreme periods as additional representative periods to the other representative periods.
3. Additional cluster center - the extreme periods are
set as an additional new cluster center. Each candidate period is assigned if it is closer to the new
cluster center than to its original cluster center.
4. Replace representative period - the extreme period is
becoming the new representative period of the cluster it was originally assigned to.
2.2.4. Hierarchical clustering
The hierarchical clustering algorithm also minimizes
the distance between each candidate period x i and the
representative periods µ k , as shown in equation 5. The algorithm starts with each single candidate as its own cluster. Then, pairs of clusters are iteratively merged until
The choice of method will determine the robustness of
the system design.
5
shape of the profiles (section 3.1). In a further step, indicators are used to contrast the suitability of the different
aggregation methods for the different types of profiles (section 3.2).
3.1. Qualitative comparison of the aggregation methods
The number of typical periods is set to Nk = 4 and
the four aggregation methods described in section 2.2 are
applied.
The original profiles are compared to their aggregated
counterparts in Figure 6 for the GHI, Figure 7 for the temperature, Figure 8 for the wind, Figure 9 for the electrical
load of a single household and Figure 10 for the load of a
whole region.
Original
0
800
20
averaging
0
Hour
2.4. Scaling of aggregated time-series
In order for the average values of the aggregated time
series to fit the average values of the original time series,
the time series are scaled a posteriori, similar to the approach of Domı́nguez-Muñoz et al. [33], Nahmmacher et al.
[30]. It is performed for each time series a separately as
follows
PNi PNg
0
g=1 xa,i,g
∀ a, k, g (12)
µa,k,g = µa,k,g PN i=1
PNg
k
g=1 µa,k,g
k=1 |Ck | ×
600
20
k-means
Hour
0
400
k-medoids
Hour
300
200
20
hierarchical
0
Hour
500
20
0
In order to not exceed the extreme values of the original
time series while scaling in accordance to Nahmmacher
et al. [30], all values greater than 1 are set to 1 and the
other values re-scaled again in order to reach the correct
average value.
As a last step, the profiles are scaled back to their
original scale:
00
0
0
0
0
µa,k,g = µa,k,g max xa − min xa + min xa ∀ a, k, g
700
GHI [W/m2]
Hour
Figure 5: Principle illustration of different methods for integrating
the extreme periods, as referred to as peak periods, into the existing
aggregated periods.
20
100
0
0
50
100
150
200
Day
250
300
350
Figure 6: Original annual GHI time series compared to the typical
time series with four typical days for different aggregation methods
(x-axis: day of the year; y-axis: hour of the day)
(13)
3. Exemplary time series aggregation
First, all figures highlight the difference between the
aggregation by the clustering algorithms and the aggregation by averaging profiles. The averaging method is bound
to group a sequence of candidate periods, while the clustering algorithms are able produce aggregated periods which
can represent a nonconsecutive order of candidate periods.
This degree of freedom allows a more accurate representation of the original time series. The differences between
the single groups of the averaging method are high, with
the consequence that by taking the average of the the candidates, much of the fluctuation in the original periods is
smoothed out.
In validating and comparing the introduced methods,
five time series are independently aggregated to four typical days. The time series are hourly values for an entire
year. The Global Horizontal solar Irradiation (GHI), the
temperature and the wind speed of the test reference year
for Bad Marienberg, Germany [15] outline the weather
phenomena relevant to energy system design. Furthermore, two electrical load profiles are aggregated, namely
the load profile of a single household [16], and an electrical
load representing an entire region [41, 42].
The profiles are reduced to four typical days and analyzed by comparing the original shape and aggregated
6
Hour
k-means
k-medoids
hierarchical
Hour
5
50
100
150
200
Day
250
300
6
20
k-medoids
350
4
20
20
2
hierarchical
0
Hour
0
8
k-means
0
0
20
20
Hour
Hour
Hour
5
10
20
0
10
20
0
Hour
15
Hour
Hour
20
0
averaging
0
20
0
12
20
25
averaging
Temp. [ C]
Hour
20
0
Original
0
Wind [m/s]
Original
0
0
0
50
100
150
200
Day
250
300
350
Figure 7: Original annual temperature time series compared to the
typical time series with four typical days for different aggregation
methods (x-axis: day of the year; y-axis: hour of the day)
Figure 8: Original annual wind time series compared to the typical
time series with four typical days for different aggregation methods
(x-axis: Day in year; y-axis: Hour at day)
A comparison of the different time series and their aggregation indicates that a high quality discrepancy exists
between the different types of profile.
Solar radiation has a strong daily pattern as seen in
Figure 6 and is fairly well-represented by all clusteringbased aggregations. The aggregation of daily profiles benefits from the high daily pattern of the irradiance time
series itself, and is able to collect the main daily and annual patterns.
Furthermore, the temperature time series in Figure 7
indicates that the major patterns are represented by the
aggregated profiles. Nevertheless, a comparison with the
original profile indicates that the days with the minimal
temperature are not well integrated, although they are
highly relevant for energy system design. It highlights why
it is important to somehow integrate the extreme periods
into the aggregated time series.
Wind speed profiles are difficult to aggregate, as seen
in Figure 8 due to their negligibly small intra-day pattern.
The variation between the days is fairly well represented,
but the daily profiles themselves seem flattened, as they
do not contain the variance of the original profile. Nevertheless, the medoid based methods seem to embody a
higher degree of variation than the k-mean clustering.
Figure 9 illustrates the high fluctuation and the lag of
a strong pattern of the electrical load profile for a single
household. In consequence, the aggregation methods are
unable to represent it well and cut out many of the peak
loads. Still, a quality difference is once again recognizable
between the mean based aggregation and the medoid based
aggregation, because the chosen centroid of the k-mean
is smoothing out more of the fluctuations. Nevertheless,
for this profile the results also indicate that for all chosen
methods, it is necessary to add the related peak periods
afterwards.
In contrast, Figure 10 shows the electrical load of an entire region and its strong daily pattern. Further differences
between winter and summer, as well as between weekdays
and weekends, are apparent. Due to these strong patterns,
all three cluster methods can aggregate the design-relevant
daily profiles, while no qualitative difference is recognizable between the different aggregations.
For all cases, it is noticeable that the assignment of
candidate periods to typical periods - the grouping - does
not variate much between the three different clustering
algorithms.
3.2. Indicator based quality evaluation
In the literature many different indicators have been
introduced to evaluate which aggregation is the most suitable [18, 39]. The use of indicators has the advantage that
they allow an evaluation of the accuracy of the time series
aggregation before any optimization procedure has even
started. Two indicators have been used in this work:
7
Original
Hour
k-means
Hour
Hour
2.0
k-medoids
1.0
hierarchical
0
50
100
150
200
Day
250
300
350
Figure 9: Original annual electrical load time series compared to the
typical time series with four typical days for different aggregation
methods (x-axis: day of the year; y-axis: hour of the day)
• The RMSE between the duration curve - the sorted
data set - of the original time series and the aggregated time series. The comparison of the duration
curves between the aggregated and original profile
is important because it allows the evaluation of all
potential data points, which are represented independently.
The first indicator (RMSE) for the normalized profiles
for the different profiles and for the different aggregation
methods is shown in Table 1 which displays the same data
as was presented in the previous section.
Table 1: Root-Mean-Squared-Error (RMSE) in percent between the
original profile and aggregated profile of four typical days for different
aggregation methods and different types of time-series
GHI
10.45
6.45
6.83
7.15
Temp.
11.57
6.32
6.55
6.59
Wind
14.44
9.99
10.78
11.00
E-Load House
9.39
7.75
8.37
9.25
1.2
k-means
1.1
1.0
20
k-medoids
0.9
0.8
20
hierarchical
20
0.7
0.6
0
50
100
150
200
Day
250
300
350
Figure 10: Original regional electricity demand time series compared
to the typical time series with four typical days for different aggregation methods (x-axis: day of the year; y-axis: hour of the day)
• The Root-Mean-Squared-Error (RMSE) which is equivalent to the average intra-cluster distance introduced
in equation 4. Essentially, it is the objective functional value of all aggregation methods.
averaging
k-means
k-medoids
hierarchical
1.3
20
0
0.5
Hour
0
averaging
0
1.5
20
1.4
0
Hour
Hour
Hour
0
2.5
20
0
Hour
3.5
3.0
0
Hour
20
averaging
20
20
4.0
E-Load
Region [-]
20
0
Original
0
E-Load
House [kW]
Hour
0
E-Load Region
12.90
6.02
6.32
6.55
The comparison of the RMSE between the profile types
supports the interpretation of the qualitative comparison
8
concerning the quality of the aggregation: The electricity
load of a whole region has the smallest RMSE, at 6.02 %
average for the k-mean clustering, followed by temperature
and GHI. With a big gap, the electrical load of a household
follows with an error of 7.75 %, while the wind profile has
the highest value, with an error of 9.99 %.
The comparison in terms of the aggregation methods
shows the following: The averaging method performs the
worst for all profiles, the k-mean clustering performs best
at the RMSE in comparison to the k-medoids and hierarchical clustering. This is because a centroid generally
features less distance to its candidates in comparison to a
medoid. The hierarchical aggregation performs the worst
in all cases.
Nevertheless, a look at the RMSE of the duration curves
in Table 2 shows a less uniformly distribution between the
profile types and the aggregation methods. The hierarchical aggregation performs the best for the aggregation of
GHI with 1.71 % mean error, and with the electrical load
of a single house featuring a 3.06 % error. k-medoids get
the best results for the wind speed with 4.09 % error and
the electrical load of a region with 1.76 % error. The temperature is best represented in terms of the duration curve
by the k-means aggregation, with an error of 2.65 %.
The hierarchical and k-medoid aggregation perform better in case of the RMSE of the duration curves because the
medoids represent more fluctations and also more peaks in
comparison to the centroids, which partially average them
4.1. CHP-based supply system
Optimizing the configuration and operation of a CHP
system is a common application for typical period aggregation, especially if integer variables are included for
choosing real components or modeling discrete states in
the operation of the system. The CHP system is shown
in Figure 11 and adds to the CHP itself a peak boiler and
heat storage, which are all getting scaled and operated.
The heat demand is simulated by a 5R1C model [34, 45],
with the introduced weather data as an input for a multifamily house. The electrical load is aggregated by six single household profiles that Tjaden et al. [16] represent the
housing units in the multi-family house.
Table 2: Root-Mean-Squared-Error (RMSE) between the duration
curve of the original profile and that of the aggregated profile of four
typical days for different aggregation methods and different profile
types
averaging
k-means
k-medoids
hierarchical
GHI
5.10
1.93
1.75
1.71
Temp.
5.55
2.65
3.21
2.74
Wind
10.43
4.63
4.09
4.28
E-Load House
7.63
3.95
4.05
3.06
E-Load Region
6.11
1.94
1.76
2.57
out. These peaks are also represented in the original duration curve, which is why these aggregations have a higher
overlap.
Still, it would be difficult to favor a certain method
based on the indicators introduced. Their information
value is limited and it is hard to predict how they would
affect optimal system design. To overcome this, we will
apply the typical periods in the following sections to the
design optimization of different energy systems .
4. Application of optimal energy system design
For analyzing the advantages and drawbacks of time series aggregation for energy system optimization, the methods introduced are applied to three different energy supply
systems:
Figure 11: Component network of the considered CHP based energy
supply system
1. A combined heat and power plant (CHP ) in combination with a gas boiler and heat storage for the
supply of the residential electricity and heat demand,
introduced in section 4.1.
2. A residential supply system that is largely based on
a heatpump and photovoltaics, introduced in section
4.2.
3. An island system that supplies the electricity demand of an entire region, as introduced in section
4.3. This is based on photovoltaics and wind as
power sources and also needs storage technologies
because it has a limited access to fossil resources.
4.1.1. Comparing the aggregation methods
For the validation, we optimized the system for the full
original time series, and subsequently for different numbers
of typical days, aggregated by the methods shown in section 2.2. The integration of peak periods was achieved by
setting the days with the highest heat and highest electricity demand as additional cluster centers, which was
discussed in section 2.3.
100
1
80
2
3
averaging
k-means
k-medoids
hierarchical
4
5
0
20
40
60
Number of typical days [d]
System cost [%]
Error [%]
The aggregation methods and number of typical periods are varied and compared for all three systems. Furthermore, in the case of the residential system, we compare
in detail the impact of different methods on integrating extreme periods. The island system is introduced to compare
the impact of the varied length of the typical periods - e.g.
typical days or typical weeks - on the system’s design.
The operation and design of the systems are optimized
to achieve minimal costs of the energy supply. The general modeling approach of the systems is introduced in Appendix A. The models are fairly simple in order to achieve
a rapid solving performance in producing results for many
different aggregation variants. The modeling language is
Pyomo 4.3 [43] and as solver Gurobi 7.0.1 [44] was chosen. The hardware was an Intel i7-4790 CPU with 32 GB
RAM, where 7 of 8 threads were used for solving.
0
GasBoiler
CHP
HeatStorage
GasGrid
EGrid
60
40
20
0
Reference
Figure 12: Relative error of the objective value, the annual energy
cost, of the optimal CHP-system design based on aggregated periods for different types of aggregation methods in comparison to the
annual cost of the cost optimal CHP-system design based on the full
time series as reference.
The results can be seen in Figure 12. It shows the relative error between the objective value, the annual cost,
of the cost-optimal CHP system design based on aggregated periods, in comparison to the annual cost of the
9
Abs. error [%]
cost-optimal CHP system design based on the full time
series for different types of aggregation methods. The bar
plot is the composition of the annual cost of the optimal
system design based on the full time series.
It is noticeable, that all aggregation methods underestimate the cost for a small number of typical periods,
although the extreme periods are included. The reason
for this is in the approach itself: The most representative
periods are in general the days which have a smaller variance and smoother profiles than the original time series.
The consequence is that the typical heat demand and the
typical electricity demand fit better to each other, and the
demand values do not exceed that often the capacity of the
CHP plant. This results in an increased supply rates by
the CHP system, with the consequence that for example
the coverage with the CHP-plant is overestimated and the
resulting system cost underestimated.
Therefore, it is also reasonable that the medoid based
methods (k-medoids and hierarchical ) perform better than
the centroid based method k-mean. The choice of real days
includes higher variations in comparison to an averaged
profile as in the k-mean method. The averaging method
performs, as expected, the worst because it is smoothing
out to many fluctuations. No dominance of the hierarchical
clustering or either the k-medoid clustering exists. For
both methods it is possible to reach errors in the annual
cost smaller than 2% with 8 typical days.
6
5
4
3
2
1
0
10
averaging
k-means
k-medoids
hierarchical
34 5
8 12 18
24
34
45
33 45 8 12
18 24
5 8 12
8 12 18 24
18 24
1
100
72
72
72
101
Solving duration [s]
Reference
102
Figure 13: Trade-off between the error in the objective function and
solving duration for different numbers of typical days, and aggregation methods
4.1.2. Gain in computational solving duration
The gain in the solving performance as a trade-off to
the related accepted error is illustrated in Figure 13. As
expected, the reduction in time steps results in a reduction
of the solving time. The shape relates to the type of system
model, while for our simple CHP-MILP we can consider a
roughly logarithmic reduction of the error related to the
solving performance. The precise shape depends on the
scaling of the optimization problem and is not generalizable. Nevertheless, this pareto-front can be drawn for each
system and it is up to the user to decide which degree of
accuracy should be chosen. For the case of eight typical
days, the solving time could get approximately reduced by
a factor of 50 in solving based on the annual profile.
Figure 14: Component network of the residential energy supply system considered
It is noticeable that the shape of the error functions of
the residential systems are similar to the shape of the CHP
system’s error function. Furthermore, for the residential
system, k-mean clustering and averaging perform poorly
in terms of underestimating the system cost, while the
medoid based aggregation methods converge quickly on
costs which are similar to the cost of the reference system,
but overestimate the cost for higher numbers of typical
days.
The high overestimates of the system cost of the hierarchical aggregation of three and four typical days show
the sensitivity to the actual choice of representative days,
e.g. a medoid day has a small overlay of the photovoltaic
feed-in and electrical load profiles, self-supply rates are
underestimated, with the effect that the energy cost are
overestimated. These effects appear less with the k-mean
aggregation because its profiles are averaged. Therefore,
they converge more smoothly on smaller errors with an
increased number of days.
The scale of the errors is in average twice as high as the
error of the CHP systems for the same number of typical
days. Therefore two factors must be taken into account:
4.2. Residential supply system
The second system analyzed is a supply system for a
residential building based on a central grid supply and
photovoltaics. The heat is supplied by a heat pump, an
immersion heater with heat storage, as is seen in Figure
14. Heat and electricity loads are given for a single family
house, while the photovoltaic feed-in profile is simulated
in advance with the PV-Lib [46].
4.2.1. Comparing the aggregation methods
Like the CHP-system, different aggregation methods
have been used for creating the typical periods, while varying also the number of typical days. The results are shown
in Figure 15.
1. The electricity demand of a single household shows a
higher fluctuation than that by an aggregated number of households, which is more difficult to capture
10
100
0
5
averaging
k-means
k-medoids
hierarchical
10
0
20
40
System cost [%]
Error [%]
5
80
40
20
0
60
Photovoltaic
ElectricHeater
HeatPump
HeatStorage
EGrid
60
Number of typical days [d]
from appending the peak periods. Setting the peak periods manually as a potential cluster center, the underestimation drops to 1%. The resulting design difference
between these two integration methods is so small due to
the fact that the peak periods are mostly at the edge of
a cluster. In consequence, no or not many other observational periods are related to them. The last method, of
replacing a whole cluster by a peak period, results in an
overestimation of 3% and forces a more conservative system design because extreme periods are represented above
average. A consequence of this is that the system design
relies more on electricity supplied by the grid and less on
photovoltaics as seen in Figure 17.
Reference
Capacity [kW or kWh]
Figure 15: Relative error of the objective value, the annual energy
cost, of the optimal residential system design based on aggregated
periods in comparison to the annual cost of the cost optimal residential system design based on the full time series for different types of
aggregation methods.
in typical periods.
2. An additional profile - the photovoltaic feed-in - impacts the optimal system design, which increases the
number of patterns to collect with the aggregation
methods.
ElectricHeater
HeatPump
HeatStorage
Photovoltaic
Reference
Nevertheless, only 12 typical days are enough for reaching
an error, corresponding to less than 2% of the objective
value.
None
append
new
cluster
center
replace
cluster
center
Figure 17: Scaling of the system components for an optimal residential system design for the full time series and system design based
on six typical days with different integrations of extreme days
4.2.2. Comparing the integration of peak periods
In order to find the best method for the integration
of peak periods, we applied all the methods introduced in
section 2.3 to the residential system designed by six typical
days aggregated by the k-medoid clustering. We manually
defined for which time series, what type of peak period
should be integrated:
Also noteworthy is that for all the system designs derived from typical periods, the optimal photovoltaic capacity is overestimated, while the optimal storage capacities
get underestimated. This is because of the already mentioned smoothing effect in the typical periods. The photovoltaic feed-in and electricity demand are predicted to
be more steady, whereby a higher overlap exists. Furthermore, the heat demand is flattened, with less heat storage
capacity required to balance it.
• the day with peak heat load
• the day with peak electricity load
• the day with the smallest total photovoltaic feed-in
98
98
99
103
Annual Cost [%]
100
4.3. Island system
In terms of validating time series aggregation to a highly
renewably based system design, we introduce an island
system that is largely reliant on wind turbines and photovoltaics. The regional electricity demand profile and wind
feed-in are drawn from Robinius et al. [42]. The potential
photovoltaic feed-in is also simulated with PV-Lib [46]. A
power plant gets scaled as backup capacity while its total
feed-in is restricted to 10% of the total electricity demand,
so as to achieve a high share of renewable energy supply.
Furthermore, hydrogen storage based on electrolysers [47],
fuel cells and pressure vessels, along with a battery storage
can balance the fluctuating feed-in of the renewable time
series. The entire system is shown in Figure 18.
Photovoltaic
ElectricHeater
HeatPump
HeatStorage
EGrid
Reference
None
append
new
cluster
center
replace
cluster
center
Figure 16: Composition of the system cost for an optimal residential
system design for the full time series and the optimal system design
based on six typical days with differing integrations of extreme days
4.3.1. Comparing the aggregation methods
As with both prior systems, the different aggregation
methods and the number of typical days are first varied.
The resulting optimal system design and errors are depicted in Figure 19.
The results can be seen in Figure 16. If no extreme periods are integrated, the system costs are underestimated
by roughly 2%. No improvement is recognizable apart
11
4.3.2. Variation of the period length
For this reason, we vary the duration of the typical
periods. Along with the typical days (Nk = 24), typical groups of three days (Nk = 72) and typical weeks
(Nk = 168) are also aggregated. We afix the aggregation
methods to k-medoids and integrate the extreme periods
as an additional cluster center.
Figure 20 shows the resulting error in the objective
function and solving duration for the three different period
lengths and different numbers of typical periods.
As was seen before, for typical days almost no convergence is recognizable.
The aggregation by groups of three days performs slightly
better: With higher numbers of typical periods, the error gets reduced, but this convergence also appears nonmonotonuously.
More predictable results can be achieved with typical weeks, which converge continuously on smaller errors.
Nevertheless, the error also remains over 2% for 24 typical weeks. Furthermore, their solving performance is also
worse due to the higher number of time steps per period.
The results in terms of the aggregation methods are
the same as for both systems introduced previously. The
major difference is the scale of the error: While the CHP
system and the residential system both converged within
less than 12 typical days to an error of less than 2% for
the medoid-based aggregation methods, even for 72 typical
days no robust convergence is identifiable for the island
system.
As in the afore-mentioned systems, the smoothing-effect
of the aggregation also partially contributes to the error,
but therefore an underestimation of the resulting system
cost would be expected. However, k-medoid and hierarchical clustering overestimate the system’s costs especially for
high numbers of typical days.
Objective
Error [%]
20
20
averaging
k-means
k-medoids
hierarchical
40
0
20
40
60
Number of typical days [d]
System cost [%]
Error [%]
0
Photovoltaic
Battery
HydrogenStorage
FuelCell
Electrolyser
Wind
BackupPlant
60
40
20
0
10
0
7
34
10
20 0
10
100
80
12
814
24
21
18
8 1214 212184
12182124
8 14
7 4 7
3
34
101
Length: 24 [h]
Length: 72 [h]
Length: 168 [h]
full system
Figure 18: Component network of the island electricity supply system considered
102
Solving duration [s]
103
Figure 20: Trade-off between the error in the objective function and
solving duration for different numbers of typical periods, and for
different period lengths
Reference
For comparing aggregations with comparable solving
performance, Figure 21 shows the cost composition for
the aggregation of 21 typical days (each 24 time steps),
seven typical periods based on three days (each 72 time
steps) and three typical weeks (each 168 steps), which in
sum have all the same number of time steps. They are
compared to the original system design for validation.
The design based on the aggregation of typical days
highly overestimates the capacities of renewable energies
and battery storage required. In contrast to the original system design, the hydrogen storage is not integrated
at all, as is seen in Figure 22. The same applies for the
aggregation to typical groups of three days, which underestimates the overall system cost by 7%. Only in the case
of the aggregation of typical weeks does a small hydrogen
storage infrastructure gets built. All in all, this system design bears the highest degree of similarity to the original
system, but also underestimates the costs by 9 %.
In summary, a poor performance of the typical period
Figure 19: Relative error of the objective value, the annual energy
cost, of the optimal island system design based on aggregated periods
in comparison to the annual cost of the cost optimal island system
design based on the full time series for different types of aggregation
methods.
Therefore, we expect the cause of the error to lie elsewhere: The cost optimal system design relies heavily on
storage systems that contribute to almost 30% of the system costs, as is seen in Figure 19, including fuel cell, electrolyzer, hydrogen storage and battery. With the applied
approach, each typical period is considered independently
and no energy exchange between the typical periods is possible. For our case, this means that only storage operation
within a single day is possible.
12
111
92
91
7 periods
72 hours
3 periods
168 hours
Annual Cost [%]
100
Reference
21 periods
24 hours
pattern. Furthermore, wind data is challenging: The representative periods smooth out much of the original intraperiod fluctuations. All aggregations were evaluated by
so-called performance indicators, which have limited analytical value.
The methods have been further applied to three different system design optimizations: A CHP-based system,
a residential system based on a heat pump and photovoltaics, and an island system with a high share of renewable energy. The outcome shows that the impact of
the time series aggregation is significantly reliant on the
structure of the energy system itself.
The best results could be achieved for the CHP and the
residential system, which are highly dependent on central
supply infrastructures either gas imports, which would
also hold for other fuel imports, or a connection to the
electricity grid. Therefore, a total of eight to twelve typical
days was sufficient to reach cost prediction errors less than
2% while reducing the solving duration of the optimization
problem by a factor of 50 in comparison to using the time
series of a full year.
For the case of the island system, the current modeling
approach leads to high estimation errors. The optimization based on the aggregated periods caused a modeling
error because the typical periods are independent and cannot exchange energy. Therefore, storage technologies were
not properly taken into account. An extension from typical
days to typical weeks brought about a small improvement
but was, as expected, still unable to account for seasonal
storage.
Photovoltaic
Battery
HydrogenStorage
FuelCell
Electrolyser
Wind
BackupPlant
Capacity [MW or MWh]
Figure 21: Composition of the system cost for an optimal island
system design for the full time series and the optimal system design based on 21 typical days, seven groups of three days, and three
typical weeks
Photovoltaic
Battery
HydrogenStorage
FuelCell
Electrolyser
Wind
BackupPlant
Reference
21 periods
24 hours
7 periods
72 hours
3 periods
168 hours
Figure 22: Scaling of the system components for an optimal island
system design for the full time series and the optimal system design based on 21 typical days, seven groups of three days, and three
typical weeks
aggregation for the highly renewable based energy system
design, including storages, can be identified. The cause of
this is that the sequence of typical periods is not considered
in the system model, with the consequence that long term
storage cannot be properly taken into account.
6. Conclusion
This paper investigated the effect of time series aggregation on optimal energy system design models. We
showed that a time series aggregation based on clustering
algorithms can significantly reduce model complexity and
the required computational time.
Comparing different aggregation methods, the results
indicate that the choice of the aggregation algorithm itself
has only a minor impact on the optimal system design.
Nevertheless, the authors would recommend a hierarchical aggregation due to its small computational load and
reproducibility. The choice of the representative periods
from the aggregated clusters is of more relevance; cluster
medoids, i.e., using real periods, produced more accurate
results than k-mean centroids, which use averaged periods.
A qualitative analysis using aggregated time series samples showed that the series representing entire regions can
be more easily aggregated than those for single locations,
i.e., less typical periods, can be sufficiently representative.
Hence, aggregation methods are expected to be more valuable for spatially-aggregated systems than for models with
finer resolution.
Furthermore, we illustrate that the trade-off between
the depth of time series reduction and the resulting modeling error depends highly on the system configuration.
5. Summary and discussion
Comparing the aggregation methods, the results show
that averaging time series based on their appearance in
the year, for example the average daily profile of a month,
leads to an inaccurate system design.
As alternative aggregation methods, k-mean clustering,
k-medoid clustering and hierarchical clustering have been
compared for the creation of typical periods. In general,
no aggregation method outperforms all the others for every
case, but a trend is recognizable: The medoid, the most
representative candidate period, performs better than the
centroid, respectively an averaged time-series. The choice
of the aggregation method itself, in this case k-medoid or
hierarchical, has a fairly small influence.
The comparison of the aggregations of different time series indicates that solar irradiation, temperature and the
electricity load of a whole region can be easily represented
with a small error by means of the introduced aggregation
methods. An electricity load of a single building is difficult
to aggregate due to high fluctuation and the small daily
13
While energy systems based on centralized supply resources
can be well represented with a few typical days, energy
systems heavily relying on storage technologies cannot be
properly represented by independent typical days at all,
nor by independent typical weeks.
This leads to the conclusion that the impact of applying time series aggregation methods must be evaluated
separately for each energy system model, or at least for a
simplified model version. Merely assessing the aggregated
time series by performance indicators can lead to significant deviation from an optimal design.
The capital expenditure of each component is divided into
the existing related costs [eur], which only appear if the
component is installed, and scale related costs [eur/kW],
specific costs, which are scale dependent [48]. For this reason, each component is modeled by a binary variable δd
that defines whether the component exists, and a continuous variable Dd , which defines the installed capacity of
the component. The resulting device specific annualized
fixed cost can be calculated with the existing related capital expenditure (CAP EXexist ), the scaling-related capital
expenditure (CAP EXspec ) and fixed operational expenditure (OP EXf ix,d ) as follows
7. Outlook
Based on the conclusion, new modeling methods are
required that take the sequence of the aggregated typical
periods into account, with the objective of modeling the
operation of long term storage contingencies.
Furthermore, future research should identify more clusterrelevant parameters like gradients and variance, etc., because the current clustering leads to smoothed typical periods that underestimate the variability of the original time
series.
All methods introduced are published in the Python
package tsam - Time Series Aggregation Module and can
be easily applied and extended.
Acknowledgments
This work was supported by the Helmholtz Association
under the Joint Initiative ”EnergySystem 2050 A Contribution of the Research Field Energy”.
cexist,d = CAP EXexist,d (CRFd + OP EXf ix,d )
cspec,d = CAP EXspec,d (CRFd + OP EXf ix,d )
The costs that variate with the operation of the system
cvar,i,j,t are related to the energy flows Ėi,j,t . Along with
the scaling of the devices Dd , the following objective function can be stated:
X
X X
cexist,d δd + cspec,d Dd +
min
cvar,i,j,t Ėi,j,t 4t
(i,j)∈L t∈T
d
(A.3)
Appendix A.2. Constraints
The device models establish the constraints of the system. They are divided into five classes: Source/Sinks,
Collectors, Transformers and Storages.
The Source/Sink class q represents input and output
flows to the system, like photovoltaic feed-in or electricity
demand. It is essentially defined by a single equation:
X
ηlb,q,t Dq ≤
Ėq,j,t ≤ ηub,q,t Dq ∀ t, q
(A.4)
(q,j)∈L
Appendix A. System modeling
For validation purposes, a comprehensible simple MixedInteger Linear Program has been chosen as the system
model. This is defined by a network of specific technologies that are connected by energy flow variables Ėi,j,t at
time step t. Each connection is therefore defined by an
output component i and input component j and belongs
to a connection set (i, j) ∈ L. These connections are restricted by the introduced component models.
For a typical period, the time steps t are replaced by
the time steps g within a single period k.
Appendix A.1. Objective function
The objective function describes the annualized cost of
the supply system considered. Therefore, for each device d,
the annualized cost are calculated with a capital recovery
factor CRFd , which considers the Weighted Average Cost
of Capital W ACCd and lifetime τd of the device in years:
CRFd =
(1 − W ACCd )τd W ACCd
(1 − W ACCd )τd − 1
(A.2)
(A.1)
where ηlb,q,t Dq could be a certain demand that must at
least be satisfied at timestep t, or ηub,q,t Dq could be the
maximal photovoltaic feed-in per installed capacity.
The Collectors class n can be seen as a hub where all
input energy flows must be equivalent to all output energy
flows:
X
X
Ėi,n,t −
Ėn,j,t = 0 ∀ t, n
(A.5)
(i,n)∈L
(n,j)∈L
The Transformer class f represents devices that transform the energy from one form to another. Examples are
a fuel cells or heat pumps. For their definition, the energy
type (electricity, gas, etc.) must be introduced. Each
energy flow Ėi,j,t has a certain energy type . With the energy type specific transformation efficiency ηf,in ,out , the
following equation can be stated for each energy transformation in the device:
X
X
ηf,in ,out
Ėi,f,t −
Ėf,j,t = 0 ∀ t, f
(i,f )∈L,in
(f,j)∈L,out
(A.6)
14
The Storage class s is defined by an additional variable
the State of Charge SOCs,t at time step t. We can utilize
the Euler method to state for the state of charge in the
next time step SOCs,t+1 :
SOCs,t+1 = SOCs,t (1 − ηsself ∆t)
P
+ηschar
Ėi,s,t ∆t
(i,s)∈L
P
1
− ηdis
Ės,j,t ∆t
s
∀ t, s
(A.7)
(s,j)∈L
char
where Ės,t
describes the charging flow with an efficiency
char
dis
of ηs
and Ės,t
the discharging flow with related effidis
self
ciency ηs . ηs
defines the self-discharge of the storage
and ∆t the step length of a single time step. The state
of charge at the beginning of the considered time frame
SOCs,1 is related to that in the end of the time frame
SOCs,Nt +1 .
The design variable of the storage s which is described
by its capacity Ds limits the state of charge to the following:
SOCs,t ≤ Ds ∀ t, s
(A.8)
The existing related variable δd restricts the scalingdependent device variable Dd by the so called BigM-Method
[49] as follows:
Mδd ≥ Dd
(A.9)
The method is inspired by Stadler et al. [3] and Lindberg
et al. [48].
References
References
[1] R. Baños, F. Manzano-Agugliaro, F. G. Montoya, C. Gil, A. Alcayde, J. Gómez, Optimization methods applied to renewable
and sustainable energy: A review, Renewable and Sustainable
Energy Reviews 15 (4) (2011) 1753–1766, ISSN 13640321, doi:
10.1016/j.rser.2010.12.008.
[2] P. Petruschke, G. Gasparovic, P. Voll, G. Krajačić, N. Duić,
A. Bardow, A hybrid approach for the efficient synthesis of renewable energy systems, Applied Energy 135 (2014) 625–633,
ISSN 03062619, doi:10.1016/j.apenergy.2014.03.051.
[3] M. Stadler, M. Groissböck, G. Cardoso, C. Marnay, Optimizing
Distributed Energy Resources and building retrofits with the
strategic DER-CAModel, Applied Energy 132 (2014) 557–567,
ISSN 03062619, doi:10.1016/j.apenergy.2014.07.041.
[4] S. Kwon, W. Won, J. Kim, A superstructure model of an isolated power supply system using renewable energy: Development and application to Jeju Island, Korea, Renewable Energy
97 (2016) 177–188, ISSN 09601481, doi:10.1016/j.renene.2016.
05.074.
[5] S. Samsatli, N. J. Samsatli, A general spatio-temporal model of
energy systems with a detailed account of transport and storage,
Computers & Chemical Engineering 80 (2015) 155–176, ISSN
00981354, doi:10.1016/j.compchemeng.2015.05.019.
[6] E. Merkel, R. McKenna, W. Fichtner, Optimisation of the capacity and the dispatch of decentralised micro-CHP systems:
A case study for the UK, Applied Energy 140 (2015) 120–134,
ISSN 03062619, doi:10.1016/j.apenergy.2014.11.036.
[7] C. Milan, C. Bojesen, M. P. Nielsen, A cost optimization model
for 100% renewable residential energy supply systems, Energy
48 (1) (2012) 118–127, ISSN 03605442, doi:10.1016/j.energy.
2012.05.034.
15
[8] D. Lauinger, P. Caliandro, J. Van herle, D. Kuhn, A linear
programming approach to the optimization of residential energy systems, Journal of Energy Storage 7 (2016) 24–37, ISSN
2352152X, doi:10.1016/j.est.2016.04.009.
[9] G. Merei, J. Moshövel, D. Magnor, D. U. Sauer, Optimization
of self-consumption and techno-economic analysis of PV-battery
systems in commercial applications, Applied Energy 168 (2016)
171–178, ISSN 03062619, doi:10.1016/j.apenergy.2016.01.083.
[10] H. Wang, W. Yin, E. Abdollahi, R. Lahdelma, W. Jiao, Modelling and optimization of CHP based district heating system
with renewable energy production and energy storage, Applied Energy 159 (2015) 401–421, ISSN 03062619, doi:10.1016/
j.apenergy.2015.09.020.
[11] A. Palzer, H.-M. Henning, A comprehensive model for the German electricity and heat sector in a future energy system with
a dominant contribution from renewable energy technologies
–Part II: Results, Renewable and Sustainable Energy Reviews
30 (2014) 1019–1034, ISSN 13640321, doi:10.1016/j.rser.2013.
11.032.
[12] P. Mancarella, MES (multi-energy systems): An overview of
concepts and evaluation models, Energy 65 (2014) 1–17, ISSN
03605442, doi:10.1016/j.energy.2013.10.041.
[13] M. Geidl, G. Andersson, Optimal Power Flow of Multiple Energy Carriers, IEEE Transactions on Power Systems
22 (1) (2007) 145–155, ISSN 0885-8950, doi:10.1109/tpwrs.2006.
888988.
[14] C. Milan, M. Stadler, G. Cardoso, S. Mashayekh, Modeling of
non-linear CHP efficiency curves in distributed energy systems,
Applied Energy 148 (2015) 334–347, ISSN 03062619, doi:10.
1016/j.apenergy.2015.03.053.
[15] DWD, Testreferenzjahre (TRY) .
[16] T. Tjaden, J. Bergner, J. Weniger, V. Quaschning, Repraesentative elektrische Lastprofile fuer Wohngebaeude in Deutschland
auf 1-sekuendiger Datenbasis doi:10.13140/RG.2.1.5112.0080.
[17] C. Golling, A cost-efficient expansion of renewable energy
sources in the European electricity system - an integrated modelling approach with a particular emphasis on diurnal and seasonal patterns, Thesis, URL http://kups.ub.uni-koeln.de/
4856/, 2012.
[18] K. Poncelet, H. Höschle, E. Delarue, A. Virag, W. D’ haeseleer, Selecting representative days for capturing the implications of integrating intermittend renewables in generation expansion planning problems, IEE Transaction on Power Systems URL http://ieeexplore.ieee.org/stamp/stamp.jsp?
arnumber=7527691.
[19] J. Ortiga, J. C. Bruno, A. Coronas, Selection of typical days
for the characterisation of energy demand in cogeneration and
trigeneration optimisation models for buildings, Energy Conversion and Management 52 (4) (2011) 1934–1942, ISSN 01968904,
doi:10.1016/j.enconman.2010.11.022.
[20] G. Mavrotas, D. Diakoulaki, K. Florios, P. Georgiou, A mathematical programming framework for energy planning in services
sector buildings under uncertainty in load demand: The case of
a hospital in Athens, Energy Policy 36 (7) (2008) 2415–2429,
ISSN 03014215, doi:10.1016/j.enpol.2008.01.011.
[21] E. D. Mehleri, H. Sarimveis, N. C. Markatos, L. G. Papageorgiou, Optimal design and operation of distributed energy systems: Application to Greek residential sector, Renewable Energy 51 (2013) 331–342, ISSN 09601481, doi:10.1016/j.renene.
2012.09.009.
[22] M. Casisi, P. Pinamonti, M. Reini, Optimal lay-out and operation of combined heat & power (CHP) distributed generation
systems, Energy 34 (12) (2009) 2175–2183, ISSN 03605442, doi:
10.1016/j.energy.2008.10.019.
[23] M. A. Lozano, J. C. Ramos, M. Carvalho, L. M. Serra, Structure optimization of energy supply systems in tertiary sector
buildings, Energy and Buildings 41 (10) (2009) 1063–1075, ISSN
03787788, doi:10.1016/j.enbuild.2009.05.008.
[24] C. E. Lythcke-Jørgensen, M. Münster, A. V. Ensinas,
F. Haglind, A method for aggregating external operating conditions in multi-generation system optimization models, Ap-
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
plied Energy 166 (2016) 59–75, ISSN 03062619, doi:10.1016/
j.apenergy.2015.12.050.
S. P. Adhau, R. M. Moharil, P. G. Adhau, K-Means clustering
technique applied to availability of micro hydro power, Sustainable Energy Technologies and Assessments 8 (2014) 191–201,
ISSN 22131388, doi:10.1016/j.seta.2014.09.001.
R. Green, I. Staffell, N. Vasilakos, Divide and Conquer - Means
Clustering of Demand Data Allows Rapid and Accurate Simulations of the British Electricity System, IEEE Transactions
on Engineering Management 61 (2) (2014) 251–260, ISSN 00189391 1558-0040, doi:10.1109/tem.2013.2284386.
B. Bahl, A. Kuempel, M. Lampe, A. Bardow, Time-series aggregation for synthesis of distributed energy supply systems
by bounding error in operational expenditure, in: Z. Kravanja
(Ed.), 26th European Symposium on Computer Aided Process
Engineering, Elsevier B.V, 2016.
S. Fazlollahi, S. L. Bungener, P. Mandel, G. Becker,
F. Maréchal, Multi-objectives, multi-period optimization of district energy systems: I. Selection of typical operating periods, Computers & Chemical Engineering 65 (2014) 54–66, ISSN
00981354, doi:10.1016/j.compchemeng.2014.03.005.
J. H. Merrick, On representation of temporal variability in electricity capacity planningmodels, Energy Economics 59 (2016)
261–274, ISSN 01409883, doi:10.1016/j.eneco.2016.08.001.
P. Nahmmacher, E. Schmid, L. Hirth, B. Knopf, Carpe diem: A
novel approach to select representative days for long-term power
system modeling, Energy 112 (2016) 430–442, ISSN 03605442,
doi:10.1016/j.energy.2016.06.081.
J. M. F. Rager, F. Maréchal, Urban Energy System Design from
the Heat Perspective using mathematical Programming including thermal Storage, Thesis, doi:10.5075/epfl-thesis-6731, 2015.
P. Stadler, A. Ashouri, F. Maréchal, Model-based optimization
of distributed and renewable energy systems in buildings, Energy and Buildings 120 (2016) 103–113, ISSN 03787788, doi:
10.1016/j.enbuild.2016.03.051.
F. Domı́nguez-Muñoz, J. M. Cejudo-López, A. Carrillo-Andrés,
M. Gallardo-Salazar, Selection of typical demand days for CHP
optimization, Energy and Buildings 43 (11) (2011) 3036–3043,
ISSN 03787788, doi:10.1016/j.enbuild.2011.07.024.
T. Schuetz, L. Schiffer, H. Harb, M. Fuchs, D. Müller, Optimal design of energy conversion units and envelopes for residential building retrofits using a comprehensive MILP model,
Applied Energy 185 (2017) 1–15, ISSN 03062619, doi:10.1016/
j.apenergy.2016.10.049.
T. Schuetz, M. H. Schraven, H. Harb, M. Fuchs, D. Müller,
Clustering algorithms for the selection of typical demand days
for the optimal design of building energy systems, in: Efficiency,
Cost, Optimization, Simulation and Environmental Impact of
Energy Systems, 2016.
K. Poncelet, E. Delarue, J. Duerinck, D. Six, W. Dhaeseleer, The importance of integrating the variability of renewables in long-term energy planning models, Report, KU Leuven, URL https://www.mech.kuleuven.be/en/tme/research/
energy_environment/Pdf/wp-importance.pdf, 2014.
H. Harb, C. Schwager, R. Streblow, D. Müller, Optimal Design
of Energy Systems in Residential Districts with Interconnected
Local Heating and Electrical Networks, in: Building Simulation
Conference, doi:10.13140/RG.2.1.2144.6488, 2015.
A. K. Jain, Data clustering: 50 years beyond K-means, Pattern
Recognition Letters 31 (8) (2010) 651–666, ISSN 01678655, doi:
10.1016/j.patrec.2009.09.011.
S. Fazlollahi, G. Becker, F. Maréchal, Multi-objectives, multiperiod optimization of district energy systems: II-Daily thermal
storage, Computers & Chemical Engineering 71 (2014) 648–662,
ISSN 00981354, doi:10.1016/j.compchemeng.2013.10.016.
A. P. Reynolds, G. Richards, B. D. L. Iglesia, V. J. RaywardSmith, Clustering Rules: A Comparison of Partitioning and
Hierarchical Clustering Algorithms, Journal of Mathematical
Modelling and Algorithms 5 (2006) 475–504, URL http://
link.springer.com/article/10.1007/s10852-005-9022-1.
M. Robinius, Strom- und Gasmarktdesign zur Versorgung des
deutschen Straßenverkehrs mit Wasserstoff, Thesis, 2015.
[42] M. Robinius, F. t. Stein, A. Schwane, D. Stolten, A TopDown Spatially Resolved Electrical Load Model, Energies 10 (3)
(2017) 361, ISSN 1996-1073, doi:10.3390/en10030361.
[43] W. E. Hart, C. Laird, D. L. Woodruff, Pyomo: modeling and
solving mathematical programs in Python, Mathematical Programming Computation 3 (3) (2011) 219–260.
[44] I. Gurobi Optimization, Gurobi Optimizer Reference Manual,
URL http://www.gurobi.com, 2016.
[45] D. EN ISO, Thermal performance and energy use in the build
environment (ISO 13790:2008) .
[46] R. W. Andrews, J. S. Stein, C. Hansen, D. Riley, Introduction
to the Open Source PV-LIB for Python Photovoltaic System
Modeling Package .
[47] S. Schiebahn, T. Grube, M. Robinius, V. Tietze, B. Kumar,
D. Stolten, Power to gas: Technological overview, systems analysis and economic assessment for a case study in Germany, International Journal of Hydrogen Energy 40 (12) (2015) 4285–
4294, ISSN 03603199, doi:10.1016/j.ijhydene.2015.01.123.
[48] K. B. Lindberg, G. Doorman, D. Fischer, M. Korpås,
A. Ånestad, I. Sartori, Methodology for optimal energy system design of Zero Energy Buildings using mixed-integer linear
programming, Energy and Buildings 127 (2016) 194–205, ISSN
03787788, doi:10.1016/j.enbuild.2016.05.039.
[49] A. Bemporad, M. Morari, Control of systems integrating logic,
dynamics, and constraints, Automatica 35 (3) (1999) 407–427,
ISSN 0005-1098, doi:Doi10.1016/S0005-1098(98)00178-2, URL
<GotoISI>://WOS:000079457800006.
16
| 5cs.CE
|
1
Distributed Kalman filtering with
minimum-time consensus algorithm
arXiv:1703.05438v1 [cs.SY] 16 Mar 2017
Ye Yuan, Ling Shi, Jun Liu, Zhiyong Chen, Hai-Tao Zhang and Jorge Goncalves
Abstract
Fueled by applications in sensor networks, these years have witnessed a surge of interest in
distributed estimation and filtering. A new approach is hereby proposed for the Distributed Kalman Filter
(DKF) by integrating a local covariance computation scheme. Compared to existing well-established
DKF methods, the virtue of the present approach lies in accelerating the convergence of the state
estimates to those of the Centralized Kalman Filter (CKF). Meanwhile, an algorithm is proposed that
allows each node to compute the averaged measurement noise covariance matrix within a minimal
discrete-time running steps in a distributed way. Both theoretical analysis and extensive numerical
simulations are conducted to show the feasibility and superiority of the proposed method.
I. I NTRODUCTION
Distributed estimation and data fusion have many applications in a variety of fields ranging
from target tracking to control of mobile networks [1]–[12]. In such scenarios, networks of
sensors are employed to take measurements on various kinds of signals of interest, which are
This is an unpublished technical report that summarizes the results in [18], [29].
Y. Yuan and H.-T. Zhang are with School of Automation and State Key Lab of Digital Manufacturing Equipment and
Technology, Huazhong University of Science and Technology, Wuhan 430074, P.R. China.
L.Shi is with Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology,
Hong Kong.
J. Liu is with Department of Applied Mathematics, University of Waterloo, Waterloo, Ontario N2L 3G1, Canada.
Z. Chen is with School of Electrical Engineering and Computer Science, The University of Newcastle, Callaghan, NSW 2308,
Australia.
J. Goncalves is with Department of Engineering, University of Cambridge, CB2 1PZ, UK, and University of Luxembourg,
Faculté des Sciences, de la Technologie et de la Communication, L-4362 BELVAL, Luxembourg.
H.-T. Zhang was supported by the National Natural Science Foundation of China under Grants 61322304, 61673189 and
51120155001. (Corresponding author: H.-T. Zhang, email: [email protected].)
March 16, 2017
DRAFT
2
regarded as the state variables. Since multiple sensors are involved, it is necessary to make full
use of the available measurements to yield an optimal estimate of system states.
Kalman filters [13], which are unbiased, minimum mean-squared error estimators, are commonly used to perform the required integration due to their effective estimation capabilities [14].
Such traditional integration approaches employ a single Kalman filter for the entire network to
estimate system states, which is thus named as the Centralized Kalman Filter (CKF). Another
typical kind of network state estimators is the decentralized Kalman filter proposed by Speyer
[15]. Therein, each node consists of both a sensor and a local Kalman filter. By exchanging
information with every other node in the network, each node will be able to calculate the average
value of the measurements. This value is afterwards used for state estimation, which is as optimal
as the CKF by using a local Kalman filter. However, high communication complexity is still a
main issue of the decentralized Kalman filter. The reason lies in the all-to-all communication
protocol to calculate the optimal state estimate [16].
Afterwards, this issue was alleviated with the introduction of the Distributed Kalman Filter
(DKF) by Olfati-Saber [17], [18]. Therein, each node is equipped with an additional consensus
filter (CF). Thus, it only needs to exchange both its measurement and its measurement noise
covariance with the immediate neighbors. By this means, the CF will be able to compute the
average value of the measurements asymptotically. As a result, each node obtains the optimal
state estimate solely by using local information, which remarkably decreases the communication
complexity. The DKF has similar asymptotic performance to the CKF, but with superior scalability and robustness to uncertainties and external noises in the network [19]. Following this
line, an observer-based distributed Kalman filter was afterwards proposed in [20]. However, so
far, an efficient distributed estimator with minimal convergence time to the DKF is still lacking
for sensor networks.
To fulfill such an urgent yet challenging task, a novel DKF estimator is developed hereby to
enable each node in the sensor network to calculate the consensus value in finite-time. Unlike
conventional asymptotically convergence methods [21], [22], the present algorithm only uses its
own past state estimates to form a Hankel matrix, with which the eventual consensus value can
be calculated in a minimum number of time steps. This algorithm has improved the existing
finite-time consensus algorithms in [23], [24] to a minimal time consensus [18]. The proposed
method shares the same spirit of the well-known Ho-Kalman method [25]. It can substantially
shorten the convergence time of the DKF to the CKF status. Meanwhile, the algorithm is robust
March 16, 2017
DRAFT
3
to the uncertainties in the local measurements. Moreover, a trade-off is considered between
communication cost and computation complexity. Still worth-mentioning is that, each node in
the network calculates the eventual consensus value of CF and therefore guarantees a finite-time
performance. Significantly, the present algorithm supports the convenient “plug-in-and-play”
mode favored by modern industrial filters. Especially for the scenarios where a sensor can be
naturally disabled or dropped when a new one needs to be added.
The remainder of the paper is organized as follows: Section II introduces the preliminaries
of graph theory, the CF and the DKF, and then provides the problems addressed by this paper.
The two DKF estimators are then developed in Section III. Numerical simulations are conducted
in Section IV to show the feasibility and superiority of the proposed DKF estimators. Finally,
conclusion is drawn in Section V.
Throughout the paper, the following symbols will be used. R, Z and N denote the sets of
real numbers, integers and positive integers, respectively. For a matrix A ∈ RM ×N , A[i, j] ∈
R denotes the {i, j}th element, A[i, :] ∈ R1×N and A[:, j] ∈ RM ×1 denotes its ith row and
j th column, respectively. A[i1 : i2 , j1 : j2 ] ∈ R(i2 −i1 +1)×(j2 −j1 +1) denotes the submatrix of A
corresponding to the rows i1 to i2 and the columns j1 to j2 . The symbol ∗[i] represents the ith
element of a column or row vector ∗. The symbol eTr = [0, . . . , 0, 1rth , 0, . . . , 0] ∈ R1×N . IN
denotes the identity matrix of dimension N .
II. P RELIMINARIES AND P ROBLEM D ESCRIPTION
A. Graph Theory
A sensor network with n nodes can be represented by a graph G = (V, E) of order n. The
set of nodes, {v1 , v2 , . . . , vn }, is represented by V and the set of edges, or communication links,
between each node is represented by E ⊆ V × V. A direct communication link from node
vj to node vi is denoted by e[j, i] = (vj , vi ) ∈ E and vi is said to be a neighbor of vj as
a result of this direct connection. All neighbors of a particular node vj constitute to the set
Nj = {vi ∈ V | e[i, j] ∈ E}. For an undirected graph, the degree of a node is the number of
b refers to an n-by-n
edges that are incident on the node. The adjacency matrix of the graph, A,
b j] is the weight of the edge from vj to vi . The degree
matrix where the off-diagonal element A[i,
b is an n-by-n diagonal matrix such that
matrix of the graph, D,
P A[i,
b j]
if i = j;
j
b
D[i, j] =
0
otherwise.
March 16, 2017
DRAFT
4
b is given by L
b := D
b − A.
b
The graph Laplacian matrix, L,
B. Distributed Kalman Filters
The aim of distributed Kalman filtering is to estimate the state of a process with the following
dynamics:
x(k + 1) = Ax(k) + Bw(k),
(1)
with the ith sensor node having a sensing model given by:
zi (k) = Hi (k)x(k) + vi (k).
(2)
Both w(k) and vi (k) are assumed to be zero-mean white Gaussian noise with covariance matrices
given by: E[w(k)w(l)T ] = Q(k)δkl , and E[vi (k)vj (l)T ] = Ri (k)δkl δij , where δ is the Kronecker
delta, i.e., δkl = 1 for k = l, and δkl = 0 otherwise.
Let x
bi (k) and xi (k) be the minimum mean-squared error state estimates of the ith node
based on the available measurement data up to time instants k and k − 1, respectively. The
measurement data, noise, noise covariance matrix, and model of the CKF for time instant k
can be defined in terms of the individual node parameters, respectively, as zc := [z1 , z2 , . . . , zn ],
vc := [v1 , v2 , . . . , vn ], Rc := diag(R1 , R2 , . . . , Rn ), and Hc := [H1 ; H2 ; . . . ; Hn ], where Hc is a
column block matrix [17].
We also assume that (A, Hc ) is observable and (A, Hi ) is observable for every i.
The measurement of the CKF is thus given by zc = Hc x + vc , and the CKF state estimate,
i.e., x
bc , writes
x
bc (k) = xc (k) + Mc (HcT Rc−1 zc − HcT Rc−1 Hc xc (k)),
(3)
where xc is the prior state estimate, Mc = (Pc−1 + HcT Rc−1 Hc )−1 and Pc is the error covariance
matrix of xc .
To propose the main problem, it is necessary to introduce a lemma [17] guaranteeing asymptotic performance of the DKF.
Lemma 1: ( [17]) Suppose every node in the network obtains the average consensus value of
S c and g c , then by performing the following computation
Mi (k) = (Pi (k)−1 + S c (k))−1 ,
x
bi (k) = xi (k) + Mi (k)[g c (k) − S c (k)xi (k)],
Pi (k + 1) = AMi (k)AT + BQi (k)B T ,
(4)
xi (k + 1) = Ab
xi (k),
March 16, 2017
DRAFT
5
where Pi (k) = nPc (k), Mi (k) = nMc (k), Qi (k) = nQ(k) and
P
S c (k) = n1 ni=1 HiT (k)Ri−1 (k)Hi (k)
=
1
H T (k)Rc−1 (k)Hc (k),
n c
g c (k) =
=
1
n
Pn
i=1
HiT (k)Ri−1 (k)zi (k)
1
H T (k)Rc−1 (k)zc (k),
n c
(5)
(6)
with S c (k) and g c (k) being the network-wide average inverse-covariance matrix and average
measurement, respectively, then one has
lim x
bi (k) − x
bc (k) = 0,
k→∞
where x
bi (k) and x
bc (k) are the estimates obtained by the DKF and the CKF, respectively.
C. Consensus filters
By Lemma 1, in order to obtain the same state estimates as the CKF, it is necessary for
each agent to compute the average consensus values g c (k) and S c (k) by exchanging information
only with its neighbors Ni . This can be done through consensus filters. There are three types of
consensus filters that are used in the DKF algorithm in [17]: low-pass, high-pass and band-pass
filters. Note that, Si (k), i.e., the estimate of S c (k) by node i, can be obtained as the output
with HiT (k)Ri−1 (k)Hi (k) as the input. In this set-up, each node exchanges its measurement and
covariance data with its neighbors at each time step. These data will be processed by consensus
filters to obtain estimates of the average consensus values of g c (k) and S c (k). These values
are then used by the local Kalman filter to calculate the state estimate at that time step.Figure
showing the original DKF node setup
By Lemma 1, one has
x
bi (k) = xi (k) + Mi (k)[gi (k) − Si (k)xi (k)].
(7)
If gi and Si are the average consensus values, respectively, i.e., gi = g c , Si = S c , then the states
estimate will asymptotically converge to that the CKF given by substituting Eqs. (5) and (6) to
Eq. (7), or
x
bc (k)
= xc (k) + Mc (HcT Rc−1 zc − HcT Rc−1 Hc xc (k))
(8)
= xc (k) + nMc (g c (k) − S c xc (k)).
March 16, 2017
DRAFT
6
Measurement
Data
Low-Pass
Consensus
Filter
Kalman
Filter
Node i
Covariance
Data
Band-pass
Consensus
Filter
Fig. 1: Architecture for node i employing the algorithm from [17]. The exchange of data from
node i to its neighbors j is not shown in this figure.
Consider the case where consensus has not yet been obtained, the error between the state
estimate x
bi by the ith node and state estimate x
bc of the CKF will thus be given by
x
bc − x
bi
= xc + nMc (g c − S c xc ) − xi − Mi (gi − Si xi )
(9)
= xc − xi + nMc (g c − S c xc ) − Mi (gi − Si xi ).
D. Problem formulation
The DKF algorithm [17] adopts the low-pass and band-pass CFs to compute the average
consensus values g c and S c respectively. Since these CFs only obtain these values asymptotically,
the error between the state estimates given in Eq. (9) only tends towards zero asymptotically as
well, i.e., limk→∞ kb
xc (k) − x
bi (k)k2 = 0. However, in industrial applications, it is unsatisfactory
to asymptotically attain the convergence to the CKF estimates. Accordingly, to guarantee the
convergence of the DKF estimates to the CKF estimates in finite time, rather than asymptotically,
we consider two main problems this paper addresses as below,
Problem 1: DKF estimator with finite steps of convergence: For an n-node connected networked system G = (V, E) governed by Eqs. (1) and (2), design a distributed estimator ui = Ψzi
for each agent i ∈ V such that for all t ≥ M
March 16, 2017
kb
xi (t) − x
bc (t)k2 = 0
DRAFT
7
for arbitrary initial estimates x
bi (0) and x
bc (0). Here, M ∈ N, x
bi (t) and x
bc (t) are estimates gained
by the DKF and the CKF, respectively, and Ψ is a compatible filter matrix to be designed.
Note that, another virtue of the above-mentioned filters lies in that it favors the requirement
of “plug-in-and-play” fashion. That is, when a sensor is disabled and drops out or if a new
one needs to be added– as long as the new graph is connected– the performance of underlying
distributed Kalman filter should be converging to the CKF shortly. To this end, we will propose
a new distributed filter development framework, where each node computes the final consensus
value of CF online in a distributed way.
III. M AIN R ESULTS
We consider a connected DKF network G = (V, E) where the measurement noise covariance
matrix for each node i, Ri , is constant. Since the measurement model, Hi , for each node is fixed
P
as well, the average consensus value S c = n1 ni=1 HiT Ri−1 Hi will therefore be a constant. The
new framework for the DKF is proposed in Algorithm 1 (in abbr. A1), whose technical analysis
is give below.
A. Technical analysis of A1
Consider a discretized system of the band-pass CF in [17]
Si (k + 1)
hP
= Si (k) +
j∈Ni (Sj (k) − Si (k))
i
P
+ j∈Ni ∪{i} (Pj (k) − Si (k)) ,
Pi (k + 1)
= Pi (k) +
P
j∈Ni
[(Pj (k) − Pi (k)) + (Uj (k) − Ui (k))] ,
with Ui (k) = Hi (k)T Ri−1 (k)Hi (k). For arbitrary element in Si (k) and Pi (k), one has
b − D
b
b
ε(k + 1)
I − L
A
ε(k)
=
b
p(k + 1)
0
I − L
p(k)
0
u(k)
+
b
−L
ε(k)
+ Bu(k),
:= A
p(k)
ε(k)
= Si (k)[h, l]
y(k) = eTi ε(k) := C
p(k)
March 16, 2017
(10)
DRAFT
8
with
ε(k) :=
h
iT
S1 (k)[h, l] . . . Sn (k)[h, l] ,
h
iT
p(k) := P1 (k)[h, l] . . . Pn (k)[h, l] ,
h
iT
u(k) := U1 (k)[h, l] . . . Un (k)[h, l] ,
b D,
b A
b being the Laplacian matrix, the degree matrix and the adjacency matrix for the
and L,
b matrix has
underlying sensor network. The sensor network is assumed to be connected so that L
a single eigenvalue at 0 and the sampling time satisfies
0<<
1
.
b i]}
max{L[i,
(11)
which guarantees consensus, as shown in [21], [26].
Proposition 1: Given a consensus filter in Eq. (10), all eigenvalues of A are within the unit
disk apart from one eigenvalue at 1.
Proof: This is straightforward using Gershogrin’s theorem [28] and the block diagonal structure
of A.
Significantly, Proposition 1 guarantees the system asymptotical consensus of the constructed
filter (10).
Next, we shall introduce some definitions and lemmas before giving the main results. For
conciseness, we drop the subscript i, j as the result is valid for any i, j ∈ V.
Definition 1: (Minimal polynomial of a matrix pair) The minimal polynomial associated with
P
the matrix pair (A, C) (A ∈ R2n×2n , C ∈ Rm×2n ) denoted by q(t) := td+1 + di=0 αi ti is the
monic polynomial of smallest degree d + 1 that satisfies Cq(A) = 0.
Since q(t) is the minimal polynomial of the pair (A, C), it then follows from Definition 1
that Cq(A) = 0. Therefore, we obtain
C(Ad+1 + αd Ad + . . . + α1 A + α0 I) = 0,
which immediately leads to that
y(d + 1) = CAd+1 x(0)
= −C(αd Ad + · · · + α1 A + α0 I)x(0)
= −αd y(d) − · · · − α1 y(1) − α0 y(0).
One then has that the dynamics of y(k) satisfies the linear regression equation:
d+1
X
i=0
March 16, 2017
αi y(k + i) = 0,
∀k ∈ N,
(12)
DRAFT
9
with αd+1 = 1 [27].
Denote the z−transform of y(k) as Y (z) := Z(y(k)). From the time-shift property of the
z−transform, it is easy to show that
Pd+1
i=1
Y (z) =
αi
Pi−1
i−`
`=0 y(`)z
q(z)
:=
H(z)
.
q(z)
(13)
It follows from Proposition 1 that the polynomial equation q(t) = 0 does not possess any
unstable roots except one at 1. With the polynomial p(z) :=
p(z) =
d
X
q(z)
,
z−1
one has that
(14)
βi z i ,
i=0
h
iT
for β = β0 . . . βd−1 1 . Then, it can be obtained that
"
#T
d
d
X
X
β = 1+
αi , 1 +
αi , · · · , 1 + αd , 1 .
i=1
(15)
i=2
As a result, one can calculate the consensus value φ by applying the final-value theorem and
some simple algebra
φ = lim y(k) = lim(z − 1)Y (z) =
z→1
k→∞
ydT β
.
1T β
(16)
Remark 1: The final value of y, i.e., y(∞), can be computed once we know β and the historical
sequence of y from y(0) to y(d).
Next, we propose an algorithm that obtains β from the historical observation of y. Let
Y 0,1,...,2k := {y(1) − y(0), y(1), . . . , y(2k + 1) − y(2k)}(k ∈ Z), and consider the following
Hankel matrix
y(1)−y(0)
Γ(Y 0,1,··· ,2k ) =y(2)−y(1)
..
.
y(2)−y(1)
···
..
.
.
..
.
y(2k + 1)−y(2k)
It has been shown in [27] that when Γ(Y 0,1,··· ,2k ) loses rank upon k = K, then β in Eq. (16)
can be calculated by
Γ(Y 0,1,··· ,2K )β = 0.
(17)
Proposition 2: [27] Given a consensus filter in Eq. (10) and the corresponding minimal
polynomial q(A) in Definition 1, then the degree
d + 1 ≤ dim{A}
March 16, 2017
(18)
DRAFT
10
Now we are ready to propose a new DKF algorithm that merely uses local observations
to compute S c in a minimum-number of steps. From Proposition 2, we see that the minimumnumber of steps for every sensor to compute the consensus value is upper-bounded by 2dim{A}.
Accordingly, for an arbitrary node i, given the collection of successive outputs of Si from the
band-pass CF,
(19)
S2k {h, l} = {Si (0)[h, l], . . . , Si (2k)[h, l]}
where the indices of Si refer to the time index. One can calculate a vector of differences given
by1
S2k [h, l]
= {S(1)[h, l] − S(0)[h, l], . . . ,
(20)
S(2k + 1)[h, l] − S(2k)[h, l]}
= {S(0)[h, l], . . . , S(2k)[h, l]},
at every time step k. This vector is then used to build up a Hankel matrix such that
S(0)[h, l] S(1)[h, l]
S(1)[h, l] S(2)[h, l]
Γ(S2k [h, l]) =
S(2)[h, l] S(3)[h, l]
..
..
.
.
S(2)[h, l]
···
S(3)[h, l]
..
.
···
S(2k)[h, l]
.
(21)
The next step is to check whether the Hankel matrix has full rank. If not, the dimension of the
Hankel matrix is increased by including the output of S(·)[h, l] from the next time step k + 1,
and the process is repeated afterwards. Once the Hankel matrix has lost rank or been closed to
losing rank (i.e., its smallest singular value is less than a small positive threshold σ), compute
h,l
the smallest singular vector of this defective Hankel matrix, given by KSh,l = [β0h,l , . . . , βk−1
, 1]T .
This kernel is then used to compute the final consensus value, S c [h, l] according to the following
equation:
S c [h, l] =
h
i
with Sk [h, l] = S(0)[h, l] . . . S(k)[h, l] .
Sk [h, l]T KSh,l
1T KSh,l
,
(22)
Remark 2: Using this approach, the consensus value S c will be obtained in minimum-time as
opposed to asymptotically. Motivated by this result, Algorithm 1 is proposed which incorporates
the minimum-time static consensus algorithm into the DKF.
A robust version of the proposed algorithm is presented in the Appendix.
1
We drop index i for notational simplicity.
March 16, 2017
DRAFT
11
Algorithm 1 Distributed Kalman Filter with minimum-time consensus scheme
Inputs: Measurement of the state zi
Outputs: Estimate of the state of the system x
b
At each time step:
Step 1: Update state of Low-pass Consensus Filter :
uj ← HjT Rj−1 zj , ∀j ∈ Ni ∪ {i},
hP
i
P
gi ← gi +
(g
−
g
)
+
(u
−
g
)
i
i
j∈Ni j
j∈Ni ∪{i} j
where is the sampling period.
Step 2: Update state of Band-pass Consensus Filter:
Uj ← HjT Rj−1 Hj , ∀j ∈ Ni ∪ {i},
P
Pi ← Pi + j∈Ni [(Pj − Pi ) + (Uj − Ui )] ,
i
hP
P
Si ← Si +
j∈Ni ∪{i} (Pj − Si ) .
j∈Ni (Sj − Si ) +
(23)
Step 3: For each node i, compute the differences between successive values of Si (for any
h, l): S2k [h, l] = {Si (1)[h, l] − Si (0)[h, l], . . . , Si (2k + 1)[h, l] − Si (2k)[h, l]} where Si (k) is the
value of Si at the k th time step and form a Hankel matrix Γ(S2k [h, l]). If the smallest singular
value of Hankel matrix is less than a prescribed small real value at some time k h,l , find the
h,l
corresponding singular vector KSh,l = [β0h,l , . . . , βk−1
, 1]T , and compute the final consensus value
for S c [h, l] using eq. (22).
Then for discrete-time k > maxh,l k h,l , Steps 4 and 5 can be computed using S c instead of Si ,
and Steps 2 and 3 itself will no longer be necessary.
Step 4: Estimate the state of the process using the local KF at each node:
Mi ← (Pi−1 + Si )−1 ,
x
b ← x + Mi (gi − Si x).
Step 5: Update the state of the local KF for the next time step
Pi ← AMi AT + BQi B T ,
x ← Ab
x.
March 16, 2017
DRAFT
12
B. Estimator performance analysis
A virtue of the proposed estimator saves communication cost and accelerated convergence
speed to the CKF estimator. Meanwhile, the performance is preserved as that of the CKF. Once
S c is calculated, communication of exchanging Si (k) is no needed. Hence the more quickly
the average consensus matrix is calculated, the more communication savings we have. More
precisely, the communication cost is reduced from O(n4 ) to O(n3 ), while the time taken for
reaching the optimal performance has been reduced from O(n2 ) to O(n) as well.
In terms of performance, by using Algorithm 1, it will take a much shorter time for the error
between x
bi and the CKF estimates x
bc to reduce to Mi (g c − gi ), by noting that nMc = Mi when
Si = S c .
Measurement
Data
Low-Pass
Consensus
Filter
Minimal-time
consensus
algorithm
Node i
Covariance
Data
Kalman
Filter
Band-pass
Consensus
Filter
Fig. 2: Architecture for node i employing Algorithm 1 [18]. In this case, we introduce a new
algorithm to compute S c using S(k). The exchange of data from node i to its neighbors j is not
shown.
A comparison is conducted between the algorithm proposed in [17] (referred to as Algorithm
0 or A0) , and the proposed algorithm A1. More precisely, for A0, each node i exchanges its
estimate on g c (k) and S c (k) with its neighbors j at each time step k. These estimates are then
transferred to the respective CFs to update the estimates, gi and Si . Then a local Kalman filter
is adopted to calculate the state estimate. In A1, the output Si of the band-pass CF is fed to the
minimum-time static consensus algorithm. Once the consensus value for all elements of S c has
March 16, 2017
DRAFT
13
been obtained at a time instant t = K, the local Kalman filter uses S c instead of Si to compute
the state estimates. Therefore, using A1, the difference between the state estimates obtained by
the DKF and the CKF will be minimized in a much shorter time.
Now, we are ready to give the main technical result as below,
Theorem 1: Given a connected sensor network running DKF, for any node i, assume that
a) the prior state estimate xi = xc ;
b) there exists a time constant K and a prescribed small positive value ρ such that kgi (k) −
g c (k)k2 ≤ ρ for any k > K;
then one has
kb
xA1
bc (k)k2 ≤ kb
xA0
bc (k)k2
i (k) − x
i (k) − x
(24)
with x
bA1
bA0
i (k) and x
i (k) being the state estimates of A1 and A0 respectively for any time k > K.
Moreover, the equality holds if and only if Si (k) = S c .
Proof: The state estimate of the ith node is given by: x
bi (k) = xi (k) + Mi (k)[gi (k) − Si (k)xi (k)],
and that of the CKF is given by x
bc (k) = xc (k) + nMc (g c (k) − S c xc (k)).
Thus the error between the state estimates of the ith node and that of the CKF at time k will
be given by:
x
bc − x
bi
=xc + nMc (g c (k) − S c xc ) − xi − Mi (gi (k) − Si (k)xi )
=xc − xi + nMc (g c (k) − S c xc ) − Mi (gi (k) − Si (k)xi )
= (Pi (k) + S c )−1 − (Pi (k) + Si (k))−1 g c +
(−(Pi (k) + S c )−1 S c + (Pi (k) + S i (k))−1 S i (k))xc
+ nMc (g c (k) − gi (k)).
Assume that gi (k) = gc and xc = xi , then we can formulate the following optimization
problem as: minSi kb
xc − x
bi k2 . That is
minSi k ((Pi (k) + S c )−1 − (Pi (k) + Si )−1 ) g c
+ (−(Pi (k) + S c )−1 S c + (Pi (k) + S i )−1 S i ) xc
+nMc (g c (k) − gi (k))k2 ,
which can be approximated as
minSi k ((Pi (k) + S c )−1 − (Pi (k) + Si )−1 ) g c
+ (−(Pi (k) + S c )−1 S c + (Pi (k) + Si )−1 Si ) xc k2
March 16, 2017
(25)
DRAFT
14
for any k > K using the triangle inequality and the assumption that kg c (k) − gi (k)k2 is small.
It follows from Eq. (25) that the minimum value 0 is achieved when Si = S c which is scheme
for A1. Therefore, A1 outperforms A0, which completes the proof.
Remark 3: Theorem 1 shows the theoretical improvement of A1 compared to A0 in terms of
a smaller state estimate error, as seen in Eq. (24). Empirically, the improved performance can
even be gained when assumptions a) and b) are not fulfilled.
IV. N UMERICAL S IMULATIONS
We shall now show how the proposed new local computation can speed up the DKF. In addition
to the theoretical analysis, a comparison by numerical simulations is conducted between the
algorithm proposed in [17] (referred to as Algorithm 0 or A0) , and the proposed algorithm A1.
The control task is to track the position of a target in two-dimensional space, i.e, x = [x1 , x2 ]T .
The target is moving in a noisy circular path with ẋ = F x + Gw where w is a white noise
process with covariance matrix Q and:
0 −3
, G = I2 , Q = 25I2 .
F =
3 0
(26)
x(k + 1) = Ax(k) + Bw(k) with
0.9990 −0.0450
0.0150 −0.0003
,B =
.
A=
0.0450 0.9990
0.0003 0.0150
(27)
half the nodes have sensing model matrix of H1 and the other half have H2 :
1 2
.
H1 = I2 , H2 =
2 1
(28)
This system was then discretized with sampling time = 0.015s to yield a discrete-time system
The sensor network used to track the target, shown in Fig. 3, consists of twenty nodes where
The measurement noise covariance matrix for node i, Ri , is taken to be constant with time and
is given by:
Ri−1
√ −1
0
(0.01 i)
.
=
√
0
(0.01 i)−1
(29)
Fig. 4 shows the evolution comparison of the variation of the first element of S between A0 and
A1. It is observed that A1 calculates the average consensus value S c in about 0.2s, whereas A0
March 16, 2017
DRAFT
15
500
400
300
200
100
0
0
100
200
300
400
500
Fig. 3: A randomly generated twenty-node sensor networks. The network is undirected and
connected. Red dots represent sensors while blue lines their interconnections.
100
50
A0
A1
0
0
1
2
Time(s)
Amplitude
Amplitude
60
40
20
0
3
A0
A1
0
1
2
Time(s)
3
100
40
20
0
A0
A1
0
1
2
Time(s)
3
Amplitude
Amplitude
60
50
A0
A1
0
0
1
2
Time(s)
3
Fig. 4: The simulation results are plotted for node 19, which is randomly picked: A comparison
between A0 and A1 on elements of S, clockwise from top left: S[1, 1], S[1, 2], S[2, 2] and
S[2, 1]. The black curves show that S c [i, j] (i, j = 1, 2) can be calculated in a much shorter time
comparing to A0.
1.6s. Table I compares the shortest, longest and average times taken across the nodes to obtain
the average consensus value for all elements of S c between A0 and A1. It shows that A1 obtains
S c in about 10% of the time taken by A0.
To show the comparison among CFK, A0 and A1 more vividly, we propose a way to assess
the performances in terms of state estimates. To ensure that the estimate difference are only due
to the differences in values of S, it is assumed that all the three algorithms know the consensus
March 16, 2017
DRAFT
x1
16
60
40
20
0
A0
A1
CKF
0
0.2
0.4
0.6
0.8
1
x2
Time (s)
10
0
-10
-20
-30
A0
A1
CKF
0
0.1
0.2
0.3
0.4
0.5
0.6
Time (s)
0.7
0.8
0.9
Fig. 5: Node 19 (a randomly picked node): A comparison of the state estimates by the CKF, A0
and A1. The estimates by A1 is closer to that of the CKF than A0, which indicates an improved
performance.
Time taken
A0
A1
Shortest (s)
1.14
0.18
Longest (s)
1.83
0.21
Average (s)
1.338
0.1845
TABLE I: Times taken for A0 and A1 to calculate average consensus for S. Since different nodes
have different dynamics and estimate, we look at the shortest/longest time that a node takes to
know S c and the average time of the whole network to know S c .
value g c (0). The comparison among the three algorithms for a twenty-node network are shown
in Fig. 5. Therein, in the beginning stage (t < 0.15s), the state estimates by both A0 and A1 are
much larger (beyond what is shown on the graph) than that of the CKF as they have yet to obtain
S c . Afterwards, at t = 0.18s, A1 has obtained S c according to Fig. 4, and the state estimates
drastically improved until it was almost the same as that of the CKF. The slight differences from
this time point onwards are due to numerical inaccuracies in the simulation. As for A0, the state
estimates only converge to that of the CKF at about t = 0.75s and t = 1s for the states x1 and
x2 , respectively. Therefore, A1 outperforms A0 in terms of converge time for the state estimates
to those of the CKF.
March 16, 2017
DRAFT
17
V. C ONCLUSION
This paper develops a DKF algorithm that enables each node in a sensor network to calculate
the global average consensus matrix of measurement noise covariances without access to global
information. Theoretical analyses have shown that state estimates converge to that of the CKF in
a shorter time. A robust DKF is afterwards developed to overcome communication/observation
noises and system uncertainties. Extensive numerical simulations are conducted to show the
feasibility and superiority of the proposed DKF estimators.
VI. ACKNOWLEDGEMENT
We wish to thank Prof. Richard Murray for discussions. The simulation was built on Manuel
Mazo Jr.’s code (TU Delft) on the realization of distributed Kalman filter (Olfati-Saber 2007
[31]) and the simulation was performed by Jerry Thia.
R EFERENCES
[1] A. Esna Ashari, A. Y. Kibangou, and F. Garin. Distributed input and state estimation for linear discrete-time systems. In
Proc. 51st IEEE Conference on Decision and Control, 2012.
[2] P. Torreblanca, P. Giselsson, and A. Rantze. Distributed receding horizon kalman filter. In Proc. 50th IEEE Conference
on Decision and Control, 2011.
[3] R. Olfati-Saber and P. Jalalkamali.
Coupled distributed estimation and control for mobile sensor networks.
IEEE
Transactions on Automatic Control, 57(10):2609–2614, 2012.
[4] A. Petitti, D. Di Paola, A. Rizzo, and G. Cicirelli. Distributed target tracking for sensor networks with only local
communication’. In Proc. 19th Mediterranean Conference on Control & Automation, pages 662–667, 2011.
[5] H. Dong, Z. Wang, and H. Gao. Distributed filtering for a class of time-varying systems over sensor networks with
quantization errors and successive packet dropouts. IEEE Transactions on Signal Processing, 60(6):3164–3173, 2012.
[6] Z. Chen, H.-T. Zhang, M.C. Fan, D. Wang, and D. Li. Algorithms and experiments on flocking of multiagents in a bounded
space. IEEE Transactions on Control Systems Technology, 22(4):1544–1549, 2013.
[7] H.-T. Zhang and Z. Chen. Consensus acceleration in a class of predictive networks. IEEE Transactions on Neural Networks
and Learning Systems, 25(10):1921–1927, 2014.
[8] J. Li and A. Nehorai. Distributed particle filtering via optimal fusion of gaussian mixtures. In International Conference
on Information Fusion, pages 1182–1189, 2015.
[9] Shahin Shahrampour and Ali Jadbabaie. Distributed online optimization in dynamic environments using mirror descent.
2016.
[10] Usman A Khan. On connectivity, observability, and stability in distributed estimation. 58(8):6639–6644, 2010.
[11] Qing Shan Jia, Ling Shi, Yilin Mo, and Bruno Sinopoli. On optimal partial broadcasting of wireless sensor networks for
kalman filtering. IEEE Transactions on Automatic Control, 57(3):715–721, 2012.
[12] S Kar, B Sinopoli, and J. M. F Moura. Kalman filtering with intermittent observations: Weak convergence to a stationary
distribution. Automatic Control IEEE Transactions on, 57(2):405–420, 2012.
March 16, 2017
DRAFT
18
[13] R. Kalman. A new approach to linear filtering and prediction problems. Journal of basic Engineering, 82(2):35–45, 1960.
[14] B. D. O Anderson and J. B. Moore. Optimal Filtering. Prentice-Hall. Englewood Cliffs. NJ., 1979.
[15] J. Speyer. Computation and transmission requirements for a decentralized linear-quadratic-gaussian control problem. IEEE
Transactions on Automatic Control, 24(2):266–269, 1979.
[16] U. A. Khan and J. M. F. Moura. Distributing the kalman filter for large-scale systems. IEEE Transactions on Signal
Processing, 56(10):4919–4935, 2008.
[17] R. Olfati-Saber. Distributed kalman filter with embedded consensus filters. In Proc. 44th IEEE Conference on Decision
and Control, pages 8179–8184, 2005.
[18] J. Thia, Y. Yuan, L. Shi, and J. Goncalves. Distributed kalman filter with minimum-time covariance computation. In Proc.
52st IEEE Conference on Decision and Control, 2013.
[19] M. Nourian, A. S. Leong, S. Dey, and D. E. Quevedo. An optimal transmission strategy for kalman filtering over packet
dropping links with imperfect acknowledgements. IEEE Transactions on Control of Network Systems, 1(1):259 – 271,
2014.
[20] H.-N. Wu and H.-D. Wang. Distributed consensus observers-based h∞ control of dissipative pde systems using sensor
networks. IEEE Transactions on Control of Network Systems, 2(2):112–121, 2014.
[21] R. Olfati-Saber and R. Murray. Consensus problems in networks of agents with switching topology and time-delays. IEEE
Transactions on Automatic Control, 49(9):1520–1533, 2004.
[22] J. Fax R. Olfati-Saber and R.M. Murray. Consensus and cooperation in networked multi-agent systems. Proceedings of
the IEEE, 95(1):215–233, 2007.
[23] S Sundaram and C. N Hadjicostis. Finite-time distributed consensus in graphs with time-invariant topologies. In American
Control Conference, pages 711–716, 2007.
[24] A. Y Kibangou. Finite-time average consensus based protocol for distributed estimation over awgn channels. In Decision
and Control and European Control Conference, pages 5595–5600, 2011.
[25] B. L. Ho and R. Kalman. Effective construction of linear, state-variable models from input/output functions. Regelungstechnik, 14(12):545–548, 1966.
[26] W. Ren and R. W. Beard. Consensus seeking in multiagents systems under dynamically changing interaction topologies.
IEEE Transactions on Automatic Control, 50(5):655–661, 2005.
[27] Y. Yuan, G. Stan, L. Shi, M. Barahona, and J. Goncalves. Decentralised minimum time consensus. Automatica, 49(5):1227–
1235, 2013.
[28] R. A. Horn and C. R. Johnson. Matrix Analysis. Cambridge University Press, 1985.
[29] Y. Yuan, J. Liu, R. M. Murray, and J. Goncalves. Decentralised minimal-time dynamic consensus. In Proc. American
Control Conference, 2012.
[30] S. Rump. Structured perturbations part i: Normwise distances. SIAM Journal of Matrix Analysis, Application, 25(1):1–30,
2003.
[31] R. Olfati-Saber. Distributed kalman filtering for sensor networks. In Prof. 46th IEEE Conference on Decision and Control,
pages 5492–5498, 2007.
March 16, 2017
DRAFT
19
A PPENDIX
A. Robust DKF algorithm, A2
In the application of A1 to the DKF, due to the communication/observation noise and the
numerical error, the Hankel matrix Γ(Y 0,1,··· ,2k ) will not lose rank for any finite dimension k.
It is necessary to propose a new robust algorithm that can incorporate both uncertainties and
internal/external noises. So, we will design a robust DKF estimator to address Problem 2
In this end, the essential issue is to find a Hankel matrix Γ(Yb0,1,··· ,2k ) to approximate Γ(Y 0,1,··· ,2k ).
Since Γ(Yb0,1,··· ,2k ) has finite rank, it can be used to estimate the final consensus value [29] as
follows
Γ(Yb0,1,··· ,2k ) = argminkΓ(Y 0,1,··· ,2k ) − Γ(Yb0,1,··· ,2k )k,
s.t.: det Γ(Yb0,1,··· ,2k ) = 0, Γ(Yb0,1,··· ,2k ) is Hankel
(30)
where k · k can be any norm and here we pick 2-norm without loss of generality.
Accordingly, we present the following procedures (see Algorithm 2 or A2).
To explain how A2 works more clearly, we seek assistance from the following lemma and
then give a proposition.
Lemma 2: [30] Let x ∈ Rn , then there exists a Hankel matrix D ∈ Rn×n , such that
Dx = x and kDk2 ≤ 1.
Proposition 3: Assume that the Hankel matrix Γ(Y 0,1,··· ,2k ) has full rank, then
min kΓ(Y 0,1,··· ,2k ) − H(k, k)k2 = σ(Γ(Y 0,1,··· ,2k ))
s.t.: det H(k, k) = 0, H(k, k) is Hankel.
(31)
where H(k, k) can be obtained by A2.
Proof: We first define the hvec operator as a mapping from a square Hankel matrix Rn×n to a
h
iT
vector R(2n+1)×1 . For example, hvec(Γ(Y 0,1,··· ,2k )) = y0 y1 · · · y2k . We now propose an
algorithm for computing the nearest defective Hankel matrix with respect to Γ(Y 0,1,··· ,2k ).
From A2 , we can see that H(k, k) satisfies the constraints in optimization (30), because
a). H(k, k) is Hankel by construction;
March 16, 2017
DRAFT
20
b). It is verified that Dv(Γ(Y 0,1,··· ,2k ) = v(Γ(Y 0,1,··· ,2k )), and hence
H(k, k)v(Γ(Y 0,1,··· ,2k ))
= Γ(Y 0,1,··· ,2k )v(Γ(Y 0,1,··· ,2k ))
−σ(Γ(Y 0,1,··· ,2k ))Dv(Γ(Y 0,1,··· ,2k ))
= 0.
As a consequence, H(k, k) does not have full rank;
c). Since H(k, k) − Γ(Y 0,1,··· ,2k ) = −σ(Γ(Y 0,1,··· ,2k ))D and kDk2 ≤ 1, then
kH(k, k) − Γ(Y 0,1,··· ,2k )k2 ≤ σ(Γ(Y 0,1,··· ,2k )).
Therefore, Γ(Yb0,1,··· ,2k ) = H(k, k) solves the optimization problem (30), which completes the
proof.
Remark 4: σ(Γ(Y 0,1,··· ,2k )) quantifies the approximation accuracy. More specifically, if it is
large (or a bad approximation), then a greater number of observations are required to increase the
dimension of the Hankel matrix to gain a better approximation. Meanwhile, in A2, the Hankel
matrix H(k, k) is constructed in Step 2. Besides, the c) part of the proof of Proposition 3 has
shown the relationship between D and H(k, k).
Remark 5: We shall replace Step 3 in A1 with the A2 to make the prediction robust to
uncertainty.
March 16, 2017
DRAFT
21
Algorithm 2 Decentralized robust minimum-time consensus value computation
Data: Successive observations of y(i), i = 0, 1, · · · .
Result: Final consensus value: φ.
Step 1: At each time step k starting form 0, we take the singular value decomposition of
Γ(Y 0,1,··· ,2k ) = U ΣV T , where Σ = diag{σ1 , σ2 , . . . , σk+1 } with σ1 ≥ σ2 . . . ≥ σk+1 =
σ(Γ(Y 0,1,··· ,2k ));
Step 2: Conduct a singular value decomposition of Γ(Y 0,1,··· ,2k ) and find the smallest singular
value σ(Γ(Y 0,1,··· ,2k )) and corresponding singular vector v(Γ(Y 0,1,··· ,2k )). If σ(Γ(Y 0,1,··· ,2k )) ≤ ρ,
then go to Step 3, otherwise k = k + 1 and go to Step 1;
Step 3: Compute the Hankel vector
hvec(D) = Cx+ CxT e1 ,
where Cx+ is the Moore-Penrose pseudoinverse of Cx , e1 = [1, 0, . . . , 0]T has length of 2k − 1
and
v[1],
Cx =
v[k]
.
..
v[2]
. . . v[k − 1] v[k]
..
..
.
.
v[1]
..
.
...
v[k]
. . . v[k − 1]
v[k]
.
v[1]
...
v[k − 1]
..
..
.
.
v[1]
..
.
Step 4: Let Γ(Yb0,1,··· ,2k ) = Γ(Y 0,1,··· ,2k ) − σ(Γ(Y 0,1,··· ,2k ))D.
Step 5: Upon obtaining Γ(Yb0,1,··· ,2k ), we adopt Eq. (16) to compute the final consensus value.
March 16, 2017
DRAFT
| 3cs.SY
|
Convex Computation of the Reachable Set for Hybrid Systems with
Parametric Uncertainty
Shankar Mohan∗, Victor Shia†, Ram Vasudevan‡
arXiv:1601.01019v1 [math.OC] 5 Jan 2016
January 7, 2016
Abstract
To verify the correct operation of systems, engineers need to determine the set of configurations of a
dynamical model that are able to safely reach a specified configuration under a control law. Unfortunately,
constructing models for systems interacting in highly dynamic environments is difficult. This paper
addresses this challenge by presenting a convex optimization method to efficiently compute the set of
configurations of a polynomial hybrid dynamical system that are able to safely reach a user defined target
set despite parametric uncertainty in the model. This class of models describes, for example, legged robots
moving over uncertain terrains. The presented approach utilizes the notion of occupation measures to
describe the evolution of trajectories of a nonlinear hybrid dynamical system with parametric uncertainty
as a linear equation over measures whose supports coincide with the trajectories under investigation. This
linear equation with user defined support constraints is approximated with vanishing conservatism using
a hierarchy of semidefinite programs that are each proven to compute an inner/outer approximation to
the set of initial conditions that can reach the user defined target set safely in spite of uncertainty. The
efficacy of this method is illustrated on a collection of six representative examples.
1
Introduction
Computing the set of configurations that are able to safely reach a desired configuration is critical to ensuring
the correct performance of a system in dynamic environments where deviations from planned behavior are to
be expected. Many methods have been proposed to efficiently compute this set that is generally referred to as
the backwards reachable set for deterministic systems. Unfortunately, the effect of intermittent contact with
the world, especially in fluctuating environments, is demanding to model deterministically. A roboticist, for
example, may be tasked with ensuring that a control for a legged robot beginning from an initial configuration
is able to safely reach a desired goal; however, limitations in sensing or environment variability may render
exact modeling of terrain height or friction impossible. The development of numerical tools to tractably
compute the backwards reachable set of dynamical systems undergoing contact, or hybrid dynamical systems,
with parametric uncertainty while providing systematic guarantees has been challenging due to the difficulty
of efficiently accounting for the uncertainty.
Given the potential utility of the set of configurations that are able to reach a target set despite parametric uncertainty, called the uncertain backwards reachable set, many researchers have attempted to develop
numerical tools to approximate this set. For instance, in [9], given a Lyapunov function, the largest ellipsoidal level-set contained in the uncertain backwards reachable set (where the terminal set is the origin) for
polynomial systems is computed by formulating an eigenvalue problem. While the method developed is able
to address systems with either time-varying or constant uncertainties, the computed ellipsoid is dependent
on the provided Lyapunov function and hence is not guaranteed to be the uncertain backwards reachable set.
∗ Shankar Mohan ([email protected]) is with the Department of Electrical Engineering and Computer Science, University
of Michigan, Ann Arbor, MI 48109
† Victor Shia ([email protected]) is with the Department of Electrical Engineering and Computer Sciences, University of California at Berkeley, Berkeley, CA 94720
‡ Ram Vasudevan ([email protected]) is with the Department of Mechanical Engineering, University of Michigan, Ann Arbor,
MI 48109
1
Several other researchers have attempted to utilize this set to construct controllers for legged robots that
are able to walk over terrains of varying heights [8, 11, 15, 27]. Their approach has relied on discretizing
the height of the terrain or selecting specific terrain profiles while constructing a safe controller across these
specified heights, which verifies the performance of the controller only at those specific heights. Moreover,
these approaches are unable to account for uncertainty associated with imperfect knowledge of terrain friction
or parameters affecting the continuous dynamics.
Other researchers have developed tools to outer approximate the uncertain backwards reachable for linear
systems with uncertain parameters using a variety of approaches [2, 13]. These methods can be extended to
nonlinear hybrid systems, but can require the introduction of a large number of discrete states to represent
the nonlinear behavior or require overly conservative estimates of potential uncertainty; this approach of
using zonotopes to generate an envelope for system trajectories has been more recently explored in [20].
More generally, Hamilton-Jacobi Bellman based approaches have also been applied to compute the uncertain
backwards reachable set for nonlinear systems with arbitrary uncertainty affecting the state at any instance
in time [30]. These approaches solve a more general problem, but rely on state space discretization which can
be prohibitive for systems of dimension greater than four without relying upon specific system structure [19].
This paper leverages a method developed in several recent papers [16,21,28,29] that describe the evolution
of trajectories of a deterministic hybrid dynamical system using measures, to describe the evolution of a
hybrid dynamical system with parametric uncertainty as a linear equation over measures. As a result of
this characterization the uncertain backwards reachable set can be computed as the solution to an infinite
dimensional linear program over the space of nonnegative measures. To compute an approximate solution to
this infinite dimensional linear program, a sequence of finite dimensional relaxed semi-definite programs are
constructed that satisfy an important property: each solution to this sequence of semi-definite programs is an
outer approximation to the uncertain backwards reachable set with asymptotically vanishing conservatism.
The approach is most comparable to those that check Lyapunov’s criteria for stability via sums-of-squares
programming to verify the safety of a system [25,26]. In contrast to these approaches, the algorithm described
in this paper does not require solving a bilinear optimization problem that requires feasible initialization and
allows for more general descriptions of the parametric uncertainty in the model.
The remainder of the paper is organized as follows: Section 2 introduces the notation used in the remainder of the paper, the class of systems under consideration, and the backwards reachable set problem under
parametric uncertainty; Section 3 describes how the backwards reachable set under parametric uncertainty
is the solution to an infinite dimensional linear program; Section 4 constructs a sequence of finite dimensional semidefinite programs that outer approximate the infinite dimensional linear program with vanishing
conservatism; Section 5 provides extensions to the contents in Section 3 by presenting amongst others, a
converging inner approximations of the BRS; Section 6 describes the performance of the approach on a set
of examples; and, Section 7 concludes the paper.
2
Preliminaries
This section defines the notation, the class of systems, and the problem considered throughout this paper.
The reader is directed to [5, 6, 12] for an introduction to some of the measure theoretic concepts utilized in
this paper.
2.1
Notation
In the remainder of this text the following notation is adopted: sets are italicized and capitalized (ex. K).
The boundary of a set K is denoted by ∂K. Finite truncations of the set of natural numbers are expressed
as Nn ∶= {1, . . . , n}. The set of continuous functions on a compact set K are denoted by C(K). The ring of
polynomials in x is denoted by R[x], and the degree of a polynomial is equal to degree its largest multinomial;
the degree of the multinomial xα , α ∈ Nn is ∣α∣ = ∥α∥1 ; and Rd [x] is the set of polynomials in x with maximum
degree d.
The dual to C(K) is the set of Radon measures on K, denoted as M(K), and the pairing of µ ∈ M(K)
2
and v ∈ C(K) is:
⟨µ, v⟩ = ∫ v(x) dµ(x).
(1)
K
We denote the nonnegative Radon measures by M+ (K). The space of Radon probability measures on K is
denoted by P(K). The Lebesgue measure is denoted by λ. Finally, supports of measures, µ, are identified
as spt(µ).
2.2
Quasi-Uncertain Hybrid Systems
We next define the class of uncertain hybrid systems considered throughout the remainder of the paper; the
definition is an adaptation of the description in [7].
Definition 1. A ‘quasi-uncertain’ hybrid system is a tuple H = (J , E, D, Γ, F, G, R), where
• J is a finite set of indices of discrete states in of H (discrete states will also be referred to as modes);
• E ⊂ J × J is a set of two-tuples describing directed edges;
• D = {Dj }j∈J is the set of domains where each Dj is a compact nj -dimensional manifold with boundary
where nj ∈ N;
• Γ = {µθj }j∈J where µθj ∈ P(Θj ) describes the uncertainty associated with discrete state j ∈ J with Θj
being a compact set;
• F = {f˜j }j∈J where f˜j ∶ Dj × Θj → Dj is a Lipschitz continuous function (in all variables) describing the
dynamics on Dj ;
• G = {Ge }e∈E is the set of guards where each G(j,j ′ ) ⊂ ∂Dj is a guard in domain j ∈ J that defines a
transition from mode j to mode j ′ ∈ J ;
• R = {Re }e∈E is the set of continuous reset maps, where each map is a continuously differentiable
injection R(j,j ′ ) ∶ G(j,j ′ ) → Dj ′ .
The definition of reset maps needs some clarification. Let A ⊂ G(j,k) be a set such that R(j,k) (A) ⊂ G(k,l) .
In this case, we will not be altering the system behaviour by identifying a new guard (or coalescing with
an existing guard) G(j,l) such that A ⊂ G(j,l) and re-defining G(j,k) ∶= G(j,k) /A. Hence, without loss of
generality, we require that
Re (Ge ) ⋂ Ga = ∅, ∀e ∈ E.
a∈E
a≠e
(2)
To avoid any ambiguity during transitions between discrete states, we assume the following:
Assumption 2. In each discrete state, the guards are mutually exclusive; i.e.
G(i,j) ∩ G(i,k) = ∅, ∀(i, j), (i, k) ∈ E, ∀j ≠ k
(3)
In addition, the systems are not allowed to undergo infinite mode transitions in any finite time-interval.
Assumption 3. H has no zeno execution.
Algorithm 1 describes the finite-time execution, [0, T ], of a hybrid system, H, as in Defn. 1 as follows:
Suppose that the system enters mode j at time t at location x ∈ Dj . Recall that the dynamics in this domain,
f˜j , are a function of a random parameter drawn from the distribution µθj ; let this random variable take
the value θ. The trajectory of the hybrid system beginning at time t at x is then given by any absolutely
continuous function that satisfies the differential equation f˜j with a fixed θ as described in Steps 5 & 6 . This
trajectory evolves until either the time evolution passes T or the trajectory arrives at a guard, whichever
3
Algorithm 1 Execution of H
1: Initialization: t = 0, j ∈ J , x0 ∈ Dj , x(0) = x0
2: While 1 do
3:
Let θ be drawn according to µθj
4:
Let γ∶ [t, T ] → Dj , absolutely continuous st.
5:
γ̇(s) = f˜(γ(s), θ) λat -a.e., s ∈ [t, T ]
6:
γ(t) = x(t)
7:
Λ(j,t) ∶= {r ∈ [t, T ]∣∃(j, k) ∈ E st. (γ(r), θ) ∈ G(j,k) }
8:
If Λ(j,t) ≠ ∅ then
9:
t′ ∶= min Λ(j,t) , k st. γ(t′ ) ∈ G(j,k)
10:
x(s) ← γ(s), ∀s ∈ [t, t′ )
11:
t ← t′ , x(t′ ) ← R(j,k) (γ(t′ )), j ← k
12:
else
13:
x(s) = γ(s), ∀s ∈ [t, T ]
14:
Stop
15:
end
16: end
a
where λt is the Lebesgue measure on [t, T ]
2α
β
δ
θ
γ
Figure 1: Schematic of the rimless wheel with θ describing the effect of unknown terrain.
happens first. Steps 7 – 11 isolates the first hitting-time of a guard of mode j and resets the solution
trajectory to a new mode whereafter the same procedure is repeated until t = T .
Note that the uncertainty does not evolve with time; the uncertainty only changes when the state trajectory is reset. This class of systems is rich as is illustrated by the following two examples: a simple 1D
pedagogical example and a 2D representative of walking models.
Example 4 (1-D Quasi-Uncertain Linear System). Consider a quasi-uncertain linear system evolving with
dynamics:
ẋ = −0.7x + 0.2θ − 0.1, ∀x ∈ D1
(4)
where θ is an unknown parameter affecting the dynamics. The system state evolves on D1 = [−1, 1] and the
value of θ is drawn from a uniform distribution on Θ1 = [0.2, 1]. The uncertain parameter can be thought
of as having arisen due to structural modeling errors, or as a result of reducing a system with time-scale
separation.
Example 5 (Rimless Wheel on an uneven terrain). The rimless wheel—constituted by a massless axle
to which n (angularly) equidistant spokes are connected—is a simple model of legged locomotion [25, 28].
Figure 1 presents a schematic of a rimless wheel—with spokes separated by an angle 2α—rolling down an
inclined plane. The rimless wheel is a hybrid system consisting of one mode; every time the spoke makes
contact with the surface of the inclined plane, the system undergoes a reset. The continuous dynamics of
the rimless wheel are:
β̈ = sin(β)
(5)
where β is the angle between the vertical (which is defined as the line that is perpendicular to the base of
the inclined plane) and the pivoting spoke. Once the swinging spoke makes contact with the terrain, the
4
states are reset as:
R(1,1) (β − , β̇ − ) = [2γ − β −
T
cos(2α) β̇ − ] .
(6)
For a rimless wheel rolling down an inclined plane with flat terrain, at the instance when the swinging spoke
makes contact with the ground, β = γ + α. To encode the uncertainty due to terrain height, suppose the
δ
), the guard is defined as:
rimless wheel encounters a step of size δ, then if we let θ = arcsin ( 2l sin
α
G(1,1) = {(β, β̇, θ) ∣ β = γ + α + θ}.
(7)
Observe that as the rimless wheel continues to roll, the terrain is allowed to change since the random variable
θ is allowed to take a distinct value after each contact with the ground.
2.3
Problem Description
The objective of this work is to estimate the largest set of initial conditions from which all state trajectories
of H, regardless of any encountered uncertainty, reach a terminal set by a pre-specified time, T . To formalize
the definition of this uncertain backwards reachable set, we denote the terminal set as XT and its projection
into each mode by X(T,j) , which we assume is compact. For convenience, we let T = [0, T ]. We define the
uncertain backwards reachable set mode-wise:
Alg. 1
X(0,j) = {x0 ∈ Dj ∣ ∀x ∶ [0, T ] ÐÐÐ→ D, with x(0) = x0 , x(T ) ∈ XT }
(8)
The uncertain backwards reachable set is then defined as X0 = {X(0,j) }j∈J . Observe that by definition, all
initial conditions originating in any member of X0 must reach XT at time T regardless of mode transitions
and uncertainty encountered along the way.
Lastly, we assume that mode transitions take place instantaneously. As a consequence, XT and guards
must have empty intersections.
Assumption 6. The guards and terminal set are mutually exclusive.
Remark 7. In describing the execution of the hybrid system, H, it is assumed that guard is a subset of ∂D.
There is no loss of generality in making this assumption, since it is always possible to define new modes (by
partitioning existing modes) such that the guard lies at the boundary.
3
Problem Formulation
In this section, we present a pair of dual infinite dimensional linear programs that compute the uncertain
backwards reachable set. Critically, note that despite the uncertainty being drawn from a distribution at
the arrival into each mode, it remains constant throughout that mode. As a result, this unknown parameter
can be appended to the dynamics of every mode j and treated as a portion of the state-space:
T
fj = [f˜jT
0Tnθ ] .
j
(9)
To address the problem of estimating the uncertain backwards reachable set (BRS), we rely on the notion
of occupation measures, first introduced in [24], to transform the hybrid nonlinear dynamics of the system
into a set linear dynamics over measures that can more readily be solved. For instance, suppose the system
enters mode j at τk with the states being initialized as x(τk ) = x0 and θ(τk ) = θ, the occupation measure,
µj (⋅ ∣ τk , x0 , θ) ∈ M+ (T × Dj × Θj ), is defined as:
T
µj (A × B × C∣τk , x0 , θ) =∫ IA×B×C (t, x(t∣τk , x0 , θ), θ)dt.
0
5
(10)
Note that the following relation between the Lebesgue measure on T , λ, and µj (⋅ ∣ τk , x0 , θ) holds for all
v ∈ C(T × Dj × Θj ):
⟨µj (⋅ ∣ τk , x0 , θ), v⟩ = ⟨λ, v(t, x(t ∣ τk , x0 , θ), θ)⟩ .
(11)
The occupation measure, as defined, is a conditional measure – conditioned on the arrival-time and initial
values of the states in that mode. To consider a set of possible arrival-times and initial conditions, we define
the average occupation measure by integrating the conditional occupation measure against a measure on the
set of possible initial conditions of the mode, µsj ∈ M+ (T × Dj × Θj ):
µj (A × B × C) =∫ µj (A × B × C ∣ τk , x0 , θ) dµsj .
(12)
T ×Dj ×Θj
The occupation measure is to be interpreted as the total time that a particular solution trajectory (with a
given initial conditions of states, uncertainty and initial time) spends in a set of interest. In contrast, the
average occupation measure is the total time spent by all solution trajectories whose initial conditions are in
a given set, in a set of interest.
Observe that by definition, the uncertain variables are independent of the states’ initial conditions; hence
µsi ∈ M+ (T × Dj × Θj ) is expressible as a product measure:
µsj = µ̄0j ⊗ µθj ,
(13)
where µ̄0j ∈ M+ (T × Dj ) is a measure describing the set of initial conditions of the states, and µθj ∈ M+ (Θj )
is as in the definition of H. Lemma 9 completes the characterization of the initial measure, µsj .
Similarly, measures on terminals sets, µTj ∈ M+ (X(T,j) × Θj ):
µTj (A × B) = ∫ IA×B (x(T ∣ τk , x0 , θ), θ) dµsj ,
(14)
T ×Dj ×Θj
and guards, µGe ∈ M+ (T × G(j,k) × Θj ):
µG(j,k) (A × B × C) =∫ µj (A × B × C ∣ τk , x0 , θ) dµsj ,
(15)
T ×Dj ×Θj
for all (j, k) ∈ E are defined. The measures µG(j,k) are supported on the guards of mode j and should be
interpreted as the hitting times of the guard. Note that µG(i,j) has been defined with a slight abuse of
notation. Guards were introduced in Defn. 1 as subsets of ∂D. In the uncertain description, if the guard
description does not depend on θ, then spt(µGe ) = G(j,k) × Θj ; otherwise spt(µGe ) ⊂ G(j,k) × Θj .
Remark 8. Note the similarity between the definitions of µJ and µGE (Eqns. (12) and (15)); µG(j,k) is the
restriction of µj to G(j,k) (denoted by µj ∣G(j,k) ).
The final measure in each mode j is then defined as:
µfj = δT ⊗ µTj +
∑
µG(j,k) .
(16)
k∈{l∣(j,l)∈E}
To compute X0 , we relate {µsj }j∈J with {µfj }j∈J using the dynamics of the system. As a first step, define
linear operators Lfj ∶ C 1 (T × Dj × Θj ) → C(T × Dj × Θj ) as:
Lfj v =
∂v
+ ⟨∇x v, f˜j ⟩
∂t
(17)
where v ∈ C 1 (T × Dj × Θj ; R) is an arbitrary test function and ∇x v computes the gradient of v in the Dj
coordinates. Suppose the system transitioned to mode j at t = τk−1 with the state vector taking value upon
6
reset x(τk−1 ) and θ. The value of v, evaluated along the flow of the system and at t = τk is computed using
the Fundamental Theorem of Calculus:
v(τk , x(τk ∣ x(τk−1 ), θk−1 )) = v(τk−1 , x(τk−1 ), θk−1 )
τk
+∫
Lf v(t, x(t ∣ τk−1 , x(τk−1 ), θk−1 )) dt.
(18)
τk−1
Using Eqn. (11), Eqn. (18) can be re-written as:
v(τk ,x(τk ∣ τk−1 , x(τk−1 ), θk−1 )) = v(τk−1 , x(τk−1 ), θk−1 )
+ ⟨µj (⋅ ∣ τk−1 , x(τk−1 , θk−1 ), Lf v⟩ ,
(19)
which can be simplified further by using Eqns. (12)–(16):
⟨µfj , v⟩ = ⟨µsj , v⟩ + ⟨µj , Lf v⟩ .
(20)
Alternatively, using the standard definition of adjoint operators1 , Eqn. (20) is re-written as:
⟨µfj , v⟩ = ⟨µsj , v⟩ + ⟨L′f µj , v⟩ .
(21)
Eqn. (21) defines a linear relation that initial and final measures evolving according to the hybrid dynamics
must satisfy. Lemma 19 formalizes this relation between trajectories of the system and Eqn. (21).
With the average occupation and final measures defined, Lemma 9 completes the characterization of the
Liouville eqn. by providing an explicit expression for the starting measure.
Lemma 9. The initial measure in each mode j, µsj , is expressible as the following
µsj = δ0 ⊗ µ0j ⊗ µθj +
∗
∗
R(i,j)
(π(t,x)
µG(i,j) ) ⊗ µθj ,
∑
(22)
i∈{k∣(k,j)∈E}
∗
where π(t,x)
denotes the pushforward constructed by lifting the (t, x)-projection operator, π(t,x) ∶ T ×Dj ×Θj →
∗
T × Dj , to measures; and R(i,j)
is the lifting of the reset map between modes i and j.
Proof. During the execution of a hybrid system, any mode can be entered either at t = 0 or due to a reset.
By definition of quasi-uncertain systems (Defn. 1 and Alg. 1), the value of the uncertain parameter in each
mode is independent of the state’s initial condition. Hence, the measure on initial states and uncertainty
values is a product measure.
Let δ0 ⊗ µ0j × µθj ∈ M+ ({0} × Dj × Θj ) be the measure on the set of initial conditions of system states at
t = 0; and σ0j ∈ M+ (T × Dj × Θj ) be the measure on initial conditions because of trajectory resets. Then,
the initial measure in the (t, x)-coordinate, using notations from Eqn. (13), can be decomposed as:
∗
µ̄0j = δ0 ⊗ µ0j + π(t,x)
σ0j
(23)
∗
where π(t,x)
is the pushforward measure under the projection operator (refer to Chapter 11 in [18] for an
introduction to pushforwards).
Trajectory resets occur if and only if trajectories reach any of the guards; hence σ0j and µGe , ∀e ∈ E must
be related. To formalize this relationship notice that σ0j can be decomposed into measures corresponding
to the source of each reset:
σ0j =
σ(i,j) ⊗ µθj ,
∑
i∈{k∣(k,j)∈E}
where σ(i,j) is the measure describing initial conditions that are reset into mode j from guard G(i,j) .
1A
linear operator L and its adjoint, L′ , satisfy the following relation:
⟨L′ µ, v⟩ = ⟨µ, Lv⟩ .
7
(24)
Upon reaching guard G(i,j) , the trajectories transition according to the reset map, R(i,j) . Reset R(i,j)
is independent of the value of the uncertainty (by Defn. 1); and the uncertainty in the dynamics upon reset
(if any), in mode j, is independent of the uncertainty (if any) in mode i. Thus, only the (t, x)-marginal of
∗
µG(i,j) affects σ(i,j) . Finally, since µθi is a probability measure, the (t, x)-marginal is equal to π(t,x)
µG(i,j) .
Applying a change of variables formula we have for all w ∈ C(T × Dj ) [5, Theorem 3.6.1]:
∗
⟨σ(i,j) , w⟩ = ⟨π(t,x)
µG(i,j) , w ○ R(i,j) ⟩ .
(25)
∗
Essentially, σ(i,j) is the pushforward measure of π(t,x)
(µG(i,j) ) under R(i,j) , and we have the expression in
the statement of the Lemma.
3.1
The primal
The problem of computing the uncertain backwards reachable set of H can be formulated as an infinitedimensional linear program that supremizes the volume of the set of initial condition:
∑ ⟨µ0j , 1⟩
sup
(P )
Λ
j∈J
st.
µsj + L′f µj = µfj
∀j ∈ J (26)
µ0j + µ̂0j = λj
∀j ∈ J (27)
∑ ⟨µTj , 1⟩ = ∑ ⟨µ0j , 1⟩
(28)
j∈J
j∈J
where λj is the Lebesgue measure supported on Dj ,
Λ = {(µJ , µ0J , µTJ , µ̂0J , µGE ) ∈ ⨉ M+ (T × Dj × Θj ) ⨉ M+ (Dj ) ⨉ M+ (X(T,j) × Θj ) ⨉ M+ (Dj )
j∈J
j∈J
j∈J
j∈J
⨉ M+ (T × Ge × Θj )}, and 1 denotes the function that takes value 1 everywhere. µ̂0j ∈ M(Dj ) are slack
e∈E
variables introduced to ensure that the mass of the µ0j are identical to the volume (under the Lebesgue
measure) of the uncertain backwards reachable set, as proven in Thm. 10. Eqn. (28) ensures that all
trajectories that emanate ∪j∈J spt(µ0j ) reach XT at t = T .
Theorem 10. If ({µsj }j∈J , {µj }j∈J , {µfj }j∈J ) is a solution to (P ), then X(0,j) = spt(µ0j ) for each j ∈ J .
In addition, the optimal value of (P ) is equal to the sum of volumes of the uncertain backwards reachable
set in each mode, i.e. ∑j∈J λj (X0j ).
Proof. We prove this Theorem by showing that the BRS contains the supports of µ0j , ∀j ∈ J and that the
union of supports of µ0j contains the BRS.
Suppose that ⋃j∈J spt(µ0j )/X0 ≠ ∅; this implies that by Lemma 19 (in the appendix), there exist
trajectories that begin in ⋃j∈J (spt(µ0j )/X(0,j) ) and terminate in XT . This is a contradiction since the BRS
is the largest set of initial conditions from which trajectories can reach XT at time t = T . Thus,
⋃ spt(µ0j ) ⊂ ⋃ X(0,j) ,
j∈J
(29)
j∈J
∑ λj (spt(µ0j )) ≤ ∑ λj (X(0,j) ).
j∈J
(30)
j∈J
By definition of the BRS, all state trajectories that emanate from a subset of X0 end in XT . That is, for each
j ∈ J and initial measure µ0j , if spt(µ0j ) ⊂ X(0,j) , there exist measures µj and µfj that satisfy Eqn. (26).
Thus the following inequality is true:
∑ λj (spt(µ0j )) ≥ ∑ λj (X(0,j) )
j∈J
(31)
j∈J
From Eqns. (30) and (31), ⋃j∈J spt(µ0j ) is the BRS of the system. That the optimal value of (P ) is the
volume of the uncertain backward reachable set follows by noting that the slack variables ensure absolute
continuity of each µ0j with respect to the Lebesgue measure and the observation that λj ∣X(0,j) , ∀j ∈ J is
feasible in (P ).
8
3.2
The dual
The dual to (P ) for a quasi-uncertain hybrid system H can be written as:
∑ ⟨λj , wj ⟩
inf
(D)
F
j∈J
st.
wj ≥ 0
∀x ∈ Dj
∀j ∈ J
(32)
vj (T, x, θ) + q ≥ 0,
∀(x, θ) ∈ Φj
∀j ∈ J
(33)
− Lf vj (t, x, θ) ≥ 0,
∀(t, x, θ) ∈ Ωj
∀j ∈ J
(34)
wj − ⟨µθj , vj (0, x, θ)⟩ − q ≥ 1,
∀x ∈ Dj
∀j ∈ J
(35)
vj − ⟨µθk , vk ⟩ ○ R(j,k) ≥ 0,
∀(x, θ) ∈ G(j,k) × Θj
∀j ∈ J
(36)
(j, k) ∈ E
where F = {(vJ , wJ , q) ∈ ⨉ C 1 (T × Dj × Θj ) ⨉ C(Dj ) × R}, Φj = X(T,j) × Θj and Ωj = T × Dj × Θj .
j∈J
j∈J
Remark 11. The dual to (P ) as stated in (D), is a result of a slight abuse of notations. By definition,
∀(i, j) ∈ E, R(i,j) ∶ Di → Dj ; however, for notational convenience, in (D), R(i,j) is assumed to have been
defined as R(i,j) ∶ T × Di → T × Dj with the mapping being identity in the t-component and the standard
R(i,j) in the x-components.
The solution to D can be used to determine the uncertain backwards reachable set:
Lemma 12. If ({v}j∈J , {wj }j∈J , q) is a feasible point to D, then the super-level set:
⋃ {x ∈ Dj ∣ wj (x) ≥ 1}
(37)
j∈J
is an outer approximation of the uncertain backwards reachable set of H. Furthermore there is a sequence
of feasible solutions to (D) such that for each j ∈ J , the 1-super-level set of the feasible wj converges from
above to the indicator function on X(0,j) in the L1 norm and almost uniformly.
Proof. To prove this lemma we project the uncertain backwards reachable set into each mode and show
that it is part of the 1-level set of w. Assume that the state trajectory terminates in X(T,jk ) for some jk
The state trajectory must have arrived in mode jk through a finite sequence of mode-transitions (according
to Assumption 3). Let this sequences of mode-transitions be of length k. Suppose the trajectory entered
mode jk at time τk , then from the Fundamental Theorem of Calculus (FTC) and the constraints in (D),
the following inequalities hold:
−q ≤ vjk (T, x(T ∣ x(τk+ ), θ), θ)
≤ vjk (τk , x(τk+ ), θ)
(Eqn. (33))
(38)
(FTC & Eqn. (34)) (39)
Now, integrating both sides of Eqn. (39) wrt. µθjk , and noting that µθjk is a probability measure, we get
−q ≤ ⟨µθjk , vjk (τk , x(τk+ ), θ)⟩ .
9
(40)
By iterative application of the constraints in (D), it follows that:
−q ≤ ⟨µθjk , vjk ⟩ ○ R(jk−1 ,jk ) (τk , x(τk− ))
(Reset)
≤ vjk−1 (τk , x(τk−
≤
+
∣ x(τk−1
), θ), θ)
+
⟨µθjk−1 , vjk (τk , x(τk−1 ), θ)⟩
(41)
(Eqn. (36)) (42)
(43)
⋮
≤ vj0 (τ1 , x(τ1− ∣ x0 , θ), θ)
(44)
≤ vj0 (0, x0 , θ)
(45)
≤ ⟨µθj0 , vj0 (0, x0 , θ)⟩
(46)
≤ wj0 (x0 ) − q − 1.
(Eqn. (35)) (47)
The final inequality implies that the initial condition of every trajectory that ends in the terminal set belongs
to the 1-superlevel set of wj for some j ∈ J . The remainder of the proof follows from a straightforward
extension to [16, Theorem 2].
Finally, note that the value computed by either optimization problem is equal which follows from [4,
Theorem 3.10] and is similar to [16, Theorem 2]:
Lemma 13. Formulations (P ) and (D) are equivalent and have the same optimal value.
In the next section, we present a method to numerically solve the dual problem.
Remark 14. There are two key aspects of the presentation that deserve re-iteration: First, the uncertainties
that influence the dynamics are drawn from the distribution each time a trajectory enters a new mode;
Second, the uncertain backwards reachable set corresponds to the set of initial conditions for all trajectories
that are able to reach the terminal set in spite of all possible sequences of uncertainty that each have non-zero
probability. Notice that the uncertain backwards reachable set is the intersection of the backwards reachable
set for every possible discrete uncertainty with non-zero probability.
4
Numerical Implementation
In this section, a sequence of Semidefinite Programs (SDP)s that approximate the solution to the infinite
dimensional primal and dual defined in §3.1 and §3.2 are introduced. This sequence of relaxations is constructed by characterizing each measure using a sequences of moments2 and assuming the following:
Assumption 15. The vector field in each mode and reset map between modes is a polynomial. Moreover
the domain, the value of uncertainties, the guard, and the target set in each mode is a semi-algebraic set.
Recall that polynomials are dense in the set of continuous functions by the Stone-Weierstrass Theorem
[12]; so this assumption is made without much loss of generality.
Under this assumption, given any finite d-degree truncation of the moment sequence of all measures in
the primal (P ), a primal relaxation, (Pd ), can be formulated over the moments of measures to construct an
SDP. The dual to (Pd ), (Dd ), can be expressed as a sums-of-squares (SOS) program by considering d-degree
polynomials in place of the continuous variables in D.
To formalize this dual program, first note that a polynomial p ∈ R[x] is SOS or p ∈ SOS if it can be written
2
m
as p(x) = ∑m
i=1 qi (x) for a set of polynomials {qi }i=1 ⊂ R[x]. Note efficient tools exist to check whether a
finite dimensional polynomial is SOS using SDPs [23]. Next, suppose we are given a semi-algebraic set
A = {x ∈ Rn ∣ hi (x) ≥ 0, hi ∈ R[x], ∀i ∈ Nm }. We denote the d-degree quadratic module of A as:
Qd (A) = {q ∈ Rd [x] ∣ ∃{sk }k∈Nm ∪{0} ⊂ SOS s.t. q = s0 + ∑ hk sk }
k∈Nm
2 The
nth moment of a measure (µ) is obtained by evaluating the following expression
yµ,n = ⟨µ, xn ⟩ .
10
(48)
The d-degree relaxation of the dual, Dd , can be written as:
inf ∑ ∫
Ξ
d
j∈J
Dj
wj (x) dλj (x)
st. wjd ∈ Qd (X(T,j) )
vjd (T, ⋅) + q ∈ Qd (Dj × Θj )
− Lfj vjd ∈ Qd (T × Dj × Θj )
wjd − ⟨µθj , vjd (0, ⋅)⟩ − q − 1 ∈ Qd (Dj )
vjd − ⟨µθk , vkd ⟩ ○ R(j,k) ∈ Qd (T × Dj × Θj )
(Dd )
(49)
∀j ∈ J
(50)
∀j ∈ J
(51)
∀j ∈ J
(52)
∀j ∈ J
(53)
∀(j, k) ∈ E (54)
d
d
where Ξd = {(vJ
, wJ
, q) ∈ ⨉ Rd [t, x, θ] ⨉ Rd [x] × R}. A primal can similarly be constructed, but the
j∈J
j∈J
solution to the dual can be used directly generate a sequence of outer approximations to the uncertain
backwards reachable set:
Theorem 16. For each d ∈ N and j ∈ J , let wjd denote the j-slice of the w-component of the solution to Dd .
Then X(0,jd ) = {x ∈ Dj ∣ wjd (x) ≥ 1} is an outer approximation to X(0,j) and limd→∞ λnj (X(0,jd ) /X(0,j) ) = 0.
Proof. The proof to this lemma is an extension of Theorems 5–7 in [28] given Lemma 12.
5
Extensions
In this section, we present extensions to the problem formulation presented in §3. All results presented in
§3 and 4 hold for the formulations contained herein; proofs are largely identical and are omitted.
5.1
Free terminal time
The formulation in §3 aimed at identifying the BRS of quasi-uncertain hybrid systems given a fixed terminal
time T and a terminal set XT . Suppose it is of interest to estimate the BRS which includes all initial
conditions from which trajectories reach XT at some time t ≤ T . This problem of computing the uncertain
backwards reachable set of H can be formulated as the following infinite-dimensional linear program that
supremizes the volume of the set of initial conditions:
sup
(P T )
∑ ⟨µ0j , 1⟩
Λ
j∈J
st.
µsj + L′f µj = µfj
∀j ∈ J (55)
µ0j + µ̂0,j = λj
∀j ∈ J (56)
∑ ⟨µTj , 1⟩ = ∑ ⟨µ0j , 1⟩
j∈J
(57)
j∈J
where λj is the Lebesgue measure supported on Dj ,
Λ = {(µJ , µ0J , µTJ , µ̂0J , µGE ) ∈ ⨉ M+ (T × Dj × Θj ) ⨉ M+ (Dj ) ⨉ M+ (X(T,j) × Θj ) ⨉ M+ (Dj )
j∈J
j∈J
j∈J
j∈J
M+ (T × Ge × Θj )}.
⨉
e∶=(j,k)∈E
This formulation differs from (P ) only in that µTj , is now supported on T × X(T,j) × Θj as opposed
to X(T,j) × Θj . This change in support translates into admitting initial conditions from which solution
trajectories reach XT at some time before t = T + for all sequences of uncertainties, as a part of the BRS.
Note that this does not mean that solution trajectories remain in X(T,j) × Θj for all time, upon first entry.
5.2
Inner approximations
As as review, according to the execution of quasi-uncertain hybrid systems (Alg. 1), the BRS of a set
XT is the set of initial conditions that reach XT for all possible sequences of uncertainties (refer to §2 for
11
definition). Given a terminal set XT , the problem formulation presented in §3 and the subsequent relaxations
in §4 provide outer approximations of the BRS. In some situations, inner approximations of the BRS are
more informative since they guarantee that points in the interior of the set obtained through relaxations will
satisfy the problem objectives. In this section, we present a problem formulation whose relaxations provide
convergent inner approximations of the BRS by adapting the technique presented in [17].
For quasi-uncertain hybrid systems with the BRS as defined in §2, constructing a sequence of convergent
inner approximations of the BRS of a set XT is analogous to constructing a sequence of convergent outer
approximations of the set of initial conditions for which there exists a set (of non-zero measure) of sequences
of uncertainties, (θ)n , for which the resulting trajectories fail to arrive at the terminal set at t = T .
The infinite-dimensional problem of interest in this case, employing the nomenclature adopted in §3, is
the following:
(P p )
∑ ⟨µ0j , 1⟩
sup
Λ
j∈J
st.
µsj + L′f µj = µfj
∀j ∈ J (58)
µ0j + µ̂0,j = λj
∀j ∈ J (59)
∑ ⟨µTj , 1⟩ = ∑ ⟨µ0j , 1⟩
j∈J
(60)
j∈J
where λj is the Lebesgue measure supported on Dj ,
c
Λ = {(µJ , µ0J , µTJ , µ̂0J , µGE ) ∈ ⨉ M+ (T × Dj × Θj ) ⨉ M+ (Dj ) ⨉ M+ ((X(T,j)
⋂ Dj )) ⨉ M+ (Dj )
j∈J
j∈J
j∈J
j∈J
M+ (T × Ge × Θj )}.
⨉
e∶=(j,k)∈E
Problem (P p ) differs from (P ) on two main counts – (1) how uncertainty in incorporated; (2) definition
of the final measure in each mode. In problem (P ), uncertainty was augmented as a state with a given initial
distribution; in (P p ), the uncertainty is considered as a bounded input to the system. This difference harks
back to the objective of the inner approximation problem – find all initial conditions that fail to reach XT
for some sequence of uncertainty. By assuming that θ is a bounded control whose value can be arbitrarily
chosen, we are able to search over all possible uncertainty sequences (which includes, specifically, the class
of uncertainties that is constant in each mode and changes value upon trajectory reset).
For solution trajectories to fail to arrive ar XT at t = T , one of the following must be true: solution
trajectories either leave the space at some time3 t ∈ [0, T ] or they arrive at D/XT at t = T for some sequence
of uncertainty. That is, employing the same nomenclature as in §3, the final measure in each mode is give
by
µfj = δT ⊗ µTj + µ∂j +
∑
µG(j,k)
(61)
k∈{l∣(j,l)∈E}
c
where µTj ∈ M+ ((X(T,j)
⋂ Dj )×Θj ), µ∂j ∈ M+ (T ×(∂Dj − ⋃k∈{l∣(j,l)∈E} G(j,k) )) and µG(j,k) ∈ M+ (T ×G(j,k) ×
Θj ), ∀(j, k) ∈ E.
In the above definition, µ∂j is a measure that traps trajectories that leave the domain, Dj , and not pass
through any of the guards. Note that µTj in (P p ), unlike in (P ), is supported on the complement of X(T,j) ;
that is, mirroring out desire to characterize the set of initial conditions from which solution trajectories fail
to reach X(T,j) .
It should be remarked that to be able to derive guarantees/results for (P p ) as we have done for (P ), we
need the following technical assumption.
Assumption 17. XT is an open subset of D; hence XTc is closed; and XTc ⋂(⋃e∈E Ge ) = ∅.
6
Examples
In this section, the proposed method is applied to six examples of varying complexities. The examples are
chosen to highlight different variants of the problem description:
3 For
a hybrid system, given that guards are subsets of the boundary (Remark 7), leaving the space is equivalent to arriving
at ⋃j (∂Dj − ⋃k∈{l∣(j,l)∈E} G(j,k) ) at t ∈ [0, T ].
12
βst
βsw
Figure 2: Schematic of the compass gait walker.
1. The compass-gait without uncertainty shows the applicability to deterministic systems
2. The 1D example in Ex. 4 and another example adopted from [9] are used to as benchmark examples
for computing the uncertain BRS for linear and nonlinear systems
3. The rimless wheel is an example in which guards are uncertain
4. The bean-bag toss serves as an example for the free terminal-time formulation in §5.1 and frictional
cone uncertainties
5. A variant of logistic resource growth dynamics is used to demonstrate inner approximations
The relaxed dual problems are constructed using the SPOTLESS toolbox [1] and solved with MOSEK on
a computer equipped with a Intel Xeon W3540 processor and 12GB of RAM. Note that, considering the
definition of BRS in §2, for all probability distributions with identical support, the problem formulation will
generate the same uncertain backwards reachable set. As a result it is assumed that the disturbance, θ, is
uniformly distributed (denoted as θ ∼ U([a, b]) for θ uniformly distributed in the interval [a, b]). Additionally
for numerical stability, the domains of modes of the hybrid system are scaled to a box of the appropriate
dimension.
6.1
Compass Gait Walker without Uncertainty
The compass gait (CG) walker is a simple model of legged locomotion consisting of two legs: one leg fixed
to the ground called the stance leg, and one leg that swings called the swing leg and shown in Fig. 2. The
CG is a one mode hybrid system in which the guard is reached when the swing leg makes contact with the
inclined plane, upon which the swing and stance leg switch. With different slopes and parameters, the CG
has been found to reach a limit cycle consisting of 1, 2+ steps [14]. In this example, we consider a passive
CG walker with no actuation. Let β = [βsw , βst ] and l = a + b, the dynamics of the passive CG are given by:
M (β, β̇)β̈ + C(β, β̇)β̇ + N (β) = 0
where
M (β, β̇) = [
C(β, β̇) = [
mb2
−mlb cos(βst − βsw )
0
mlb sin(βst − βsw )β̇sw
N (β) = [
(62)
−mlb cos(βst − βsw )
]
(mh + m)l2 + ma2
(63)
mlb sin(βst − βsw )β̇st
]
0
(64)
mbg sin(βsw )
]
−(mh l + ma + ml)g sin(βst )
(65)
The guard is defined as when the swing leg hits the inclined slope and mathematically defined as:
G(1,1) = {(β, β̇) ∣ βsw + βst + 2γ = 0}.
13
(66)
3
2
β̇
1
0
-1
-2
-3
-0.4
-0.2
0
0.2
0.4
0.6
β
Figure 3: The finite time region of attraction, shaded in gray, for the compass gait’s limit cycle. The swing
and stance leg limit cycles are projected down to the (β, β̇) domain and denoted by the blue and red line,
respectively. The time horizon is 1s.
The reset map is given by:
R(1,1) (β − , β̇ − ) = [βst
where
Q−α = [
Q+α = [
−mab
0
βsw
(Q+α )
−1
Q−α β̇ − ]
′
−mab + (mh l2 + 2mal) cos(2α)
]
−mab
mb(b − l cos(2α))
mb2
ml(l − b cos(2α)) + ma2 + mh l2
]
−mbl cos(2α)
(67)
(68)
(69)
and α = βsw2−βst . The reset dynamics are derived using conservation of momentum resulting in a loss of
kinetic energy. The loss of kinetic energy is recovered via change in potential energy as the CG walks down
the slope.
Prior to [22] which presents an inner-approximation to the BRS, the BRS was limited to exhaustive
simulation. However, [22] is limited to a small region and misses much of the BRS. For computation, we
consider the 5th order Taylor approximation of the dynamics about the origin and a linearized reset map
about the point where the limit cycle encounters the guard.
Figure 3 presents the polynomial degree 10 approximation to the backwards reachable set for the compass
gait (with γ = 0.05) which is tasked with reaching within 0.1 of the limit cycle (in black) in T = 1 second
with mh = 10kg, m = 5kg, a = b = 1. Through simulation of 10,000 points in the BRS, we find that 70% of
the BRS reaches within 0.1 of the limit cycle.
6.2
1-D Quasi-Uncertain Linear System
Recall the 1-D linear dynamical system from Ex. 4 whose dynamics are:
ẋ1 = −0.7x1 + 0.2θ − 0.1,
14
(70)
1.6
1.4
w20
1.2
@BRS
1
@BRS
0.8
@X
0.6
0.2
0
0
0.4
0.2
0
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
x
Figure 4: Comparison between indicator functions on analytically computed backwards reachable sets of the
1D linear system in Eqn. (70) and w20 obtained as a part of the solution to D20 . The 1-level superset of w20
encompasses the intersection of the backward reachable sets when the uncertainty takes extreme constant
values.
where θ ∈ U([0.2, 1]). Setting T = 1, the target set is chosen as XT = [0.2, 0.4]. If θ was a fixed constant then
the BRS for the system evolving with this known constant is analytically computed to be
2θ − 1 0.7 2θ − 1
)e +
,
7
7
2θ − 1 0.7 2θ − 1
(0.4 −
)e +
].
7
7
BRSθ = [(0.2 −
(71)
Note that the expression for the BRSθ is linear in θ and that the width of BRSθ is constant for all values
of θ. As the value of θ changes, BRSθ slides along R; thus, the intersection of BRS0.2 and BRS1 is the
uncertain backwards reachable set of the system in Eqn. (70) system.
Figure 4 plots the degree 20 approximation of the indicator function of X0 , w20 , that solves D20 and
the analytically computed indicator functions supported on BRS0.2 and BRS1 , and χX0 . Observe that the
1-level superset of w20 contains X0 and hence is an outer approximation of X0 .
6.3
Rimless Wheel on Uneven Terrain
The rimless wheel, introduced in Ex. 5, is a one mode hybrid system in which the guard is reached when the
swinging spoke makes contact with the inclined plane. For a rimless wheel rolling along an inclined plane with
no terrain height variation (apart from the deterministic incline), an analytically computable stable limit
cycle exits [10]; however, for the case considered in this example—with the inclined plane having variations
in terrain height—the definition of a limit cycle is less clear. In this example, we define the terminal set as
an band around the stable limit-cycle of the disturbance-free system.
Figure 5 presents the polynomial degree 14 approximation to the uncertain backwards reachable set
(black dashed) for the rimless wheel (with α = 0.4) which is tasked with arriving within the red band in T = 4
seconds, as it is rolling down an inclined plane with slope γ = 0.2. The uncertain parameter, θ, which affects
the terrain height as described earlier in Ex. 5, is drawn from a uniform distribution, θ ∼ U([−0.1, 0.1]). The
maximum terrain variation is about 25% of the length of each spoke. With this setup, by the terminal time,
somewhere between four and six spokes will have made contact with the wedge.
The outer approximation of the uncertain BRS is validated by performing Monte Carlo simulations; the
unit box is discretized into 51 points in either direction and 100 independent trajectories are simulated
(using MATLAB’s ode45 function) from each initial condition. The blue dots depict the initial conditions
that arrived within the terminal set at the desired time without violating any constraint. Note that the set
15
1
0.9
0.8
0.7
-_
0.6
0.5
0.4
0.3
0.2
0.1
0
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
Figure 5: Comparing outer approximations with Monte Carlo simulations. The dotted curve depicts the
boundary of the outer approximation of the backward reachable set of a PRW tasked with reaching the red
band by t = 4 seconds. The blue dots represent a sampling of the true backwards reachable set obtained
by performing Monte Carlo simulation; for each initial condition, 100 trials were performed and only those
initial conditions from which no trajectory violated constraints and reached XT at t = 4 were included.
of points that succeeded in the Monte Carlo simulation is contained entirely in the uncertain BRS computed
using our formulation. In fact, as a result of our method, we know that for points outside of the black region
there exist a sequence of terrain heights that produces a trajectory that does not arrive at the target set at
the designated time.
Monte Carlo simulations for this example were performed on a computer with two Intel Xeon E5-2660
processors and 128 GB of RAM. The program was written and executed in parallel in MATLAB 8.4.0 using
the parallel processing toolbox and took 21019 seconds to compute the result. In comparison, the proposed
method, when solved on an arguably lesser computational resource, took 4487 seconds to solve D14 .
6.4
An example from literature
We next consider the following benchmark example from [9, 31]
−x1
−6x2 + x22 + x31
4x2 − x22
ẋ = [
]+[
]δ + [
] δ2
3x1 − 2x2
−10x1 + 6x2 + x1 x2
12x2 − 4x2
(72)
with δ ∼ U([0, 1]). We solve the outer approximation dual formulation with the terminal set described
by a ball of radius 0.05 centered at the origin, for different values of terminal time, T . Figure 6 presents a
comparison between the obtained 16-degree outer approximations and the true BRS as estimated by sampling
the space for when T = 5 s and T = 10 s; in both cases, the 1-superlevel set is an outer approximation of the
BRS.
6.5
Bean-bag toss
In this example, we consider a simplified version of the party game involving tossing a bean-bag onto an
inclined plane which has a designated target-zone. If the motion of the bean-bag (of mass m) is considered
16
5
0
x1
1
1
T =5 s
T = 10 s
0.5
x2
0.5
x2
0
0
-0.5
-0.5
-0.5
0
x1
-1
-1
0.5
0
x1
1
Figure 6: Estimated degree-16 BRSs with terminal set B0 (0.05) (dashed) and corresponding0.5
result of MC
simulation (dots); (a) terminal time T = 5 s, (b) terminal time T = 10 s.
as it leaves the hand of the player, its equation of motion can be modeled as a hybrid system0 with three
states: (mode 1) flight, (mode 2) sliding, and (mode 3) rest. The dynamic equations in each mode is given
by the following
x2
1
-1
Mode 1
Mode 2
Mode 3
T
T
[ẋ1 ẋ2 ẏ1 ẏ2 ] = [x2 0 y2 −g]
d˙
d2
[ ˙1 ] = [
]
mg
sin(γ)
−
sgn(d
d2
2 ) µk mg cos(γ)
[d˙1 ] = [0] .
-0.5
-0.5
(73)
(74)
(75)
In the above description, in mode 1, the states correspond to position and velocities in cartesian coordinates;
when the bean-bag makes contact with the wedge (of slope γ), it is assumed that the impact is inelastic and
the system transitions to mode 2. Upon transitioning to mode 2, a coordinate transformation is applied and
the states of the dynamic model correspond to the position of the bean-bag relative to the end of the wedge
˙ If the bean-bag slides along the wedge, it may reach the
(d), and the sliding velocity along the wedge (d).
target-zone located at d = 0.55 and shaped as a ball of radius 0.05 m; whereupon the bean-bag falls and
comes to rest (dynamics transitions to mode 3). It should be noted that as the bean-bag slides, its motion
is impeded by friction with an uncertain kinetic frictional coefficient µk ∈ [0.37, 0.97].
The various guards and reset maps of this hybrid system are mathematically represented as
G(1,2) = {(x1 , x2 , y1 , y2 ) ∈ R4 ∣ y1 = x1 tan(γ)},
2
(76)
G(2,3) = {(d1 , d2 ) ∈ R ∣ d2 = 0},
(77)
R(1,2) ↦ [y1 csc(γ); x2 cos(γ) + y2 sin(γ)],
(78)
R(2,3) ↦ d1 .
(79)
Suppose the mass of the bean-bag is m = 0.5 Kg, γ = π/6, and the length of the inclined face of the wedge is 1
m, and additionally that the velocity of the bean-bag as it slides d˙ ∈ [−1, 1]. We are interested in knowing the
positions, (x1 , y1 ), and velocities, (x2 , y2 ), at which one should toss the bag to ensure that it falls into the
hole on the wedge. In Fig. 7 we present the projection of the BRS onto mode 2. The blue dots correspond
to the initial values of states on mode 2 that reach XT at some time for all admissible values of µk (the true
BRS); the dashed red line is the outer approximation obtained by utilizing the formulation in §5.1.
6.6
An example of inner approximation
Consider the 1D system whose dynamics is described by a variant of the logistic resource growth equation
ẋ = 0.2x2 + θx, ∀(x, θ) ∈ [−1, 1] × [0, 0.3].
17
(80)
0
x1
1
d_
0.5
0
-0.5
-1
0
0.2
0.4
0.6
0.8
1
d
Figure 7: Projection of the BRS of a bean-bag of mass 0.5 Kg thrown onto a 1 m long π/6-wedge with hole
located at XT = [0.5, 0.6]. The kinetic frictional coefficient of the wedge’s slope, µf ∈ [0.37, 0.97], is not
known with certainty. The blue dots are the landing points that hit the target; the interior of the red region
in the outer approximation of the BRS.
iBRS3
1.5
oBRS3
w
1
0.5
0
-1
-0.5
0
x
0.5
1
Figure 8: Outer and inner approximations of the BRS of a 1D system exhibiting transcritical bifurcation.
The terminal set, XT , T = 1 is a ball of radius 0.3 about the origin. The dashed black boxes are characteristic
functions of the true BRS of the uncertain system (Eqn. (80)). The indicator functions of degree 12 for outer
and inner approximations are presented in solid red and dot-dashed blue respectively.
This system exhibits transcritical bifurcation as the value of θ changes. That is, when θ = 0, the origin
is attractive for all x < 0, and for other values of θ, the origin is unstable and a new stable equilibrium is
created in the left half plane; regardless of the value of θ, the origin is repulsive for all x > 0.
Let the space be divided two modes D1 = [−1, 0] and D2 = [0, 1], with a guard, G(1,2) at x = 1, and
associated reset map R(1,2) (x) = −x/6. This reset maps traps trajectories leaving the space from the right
and maps them into the left half plane, and closer to the equilibria. Suppose it is of interest to determine
the BRS associated with the terminal set XT = [−0.3, 0.3] with T = 1. Figure 8 presents the degree 12 outer
(oBRSθ ) and inner approximations (iBRSθ ) of the indicator function, w, on the BRS, and the true BRS of
this uncertain system. Observe that the inner and outer approximations of the BRS are as expected.
18
7
Conclusions
In this paper, a convex optimization approach is presented to compute the backwards reachable set of
quasi-uncertain hybrid systems. The presented method optimizes over the set of unsigned measures using
converging moment relaxations that can be solved using SDPs. A commentary on the accuracy and the
adequacy of the proposed method is provided using examples. Future work will extend the work herein by
incorporating control laws while computing the backwards reachable set.
A
Existence of solutions
In Lemma 19 we concern ourselves with establishing that if measures µsJ , µJ , and µfJ satisfy Eqn. (20),
then there exist solutions to the hybrid system H that originate in the support of µsJ and terminate at t = T
in the support of µTJ .
The critical ingredient of the proof is the relation between the existence of solutions to a Conservative
Continuity Equations (CCE) and an Ordinary Differential Equation (ODE) – if a solution to the CCE exists,
solutions to the ODE exist [3]. As a review, a continuity equation is a PDE of the following form
∂ρ
+ ∇ ⋅ (ρv) = 0,
∂t
(81)
i where v is the ‘flux’ and ρ is the conserved quantity mass.
We first show that solutions to CCEs are related to the average occupation measure defined in Eqn. (12).
Subsequently, we show that the Liouville eqn. in Eqn. (20) is a CCE with the associated ODE given by
Eqn. (9), and conclude that µJ which solves the Liouville eqn. is an average occupation measure associated
with H. Since µJ solves a CCE, solutions to the ODE in Eqn. (9) and hence to H, exist.
As as review, we present, below, the problem of finding solutions to CCE and an associated theorem
about the existence of a representing measure.
Lemma 18. Consider the homogeneous conservative continuity problem in any one mode of the system; say
mode j:
d
ζt + ∇ ⋅ (f˜j ζt ) = 0
dt
(82)
where ζt are stochastic kernels supported on Dj , conditioned on time, t; and f˜j is the Lipschitz vector field.
If ζt satisfies the continuity equation, then there exists η ∈ M+ (Dj × Ξ) such that ∀(φ, t) ∈ Cb (Dj ) × T
∫
Dj
φ(x) dζt = ∫
Dj ×Ξ
φ(γ(t)) dη(x, γ),
(83)
where Ξ is the space of all absolutely continuous functions supported on {T } and map to Dj ; i.e. Ξ ∶=
C(T ; Dj ); and Cb is the space of bounded functions. The measure η can be interpreted as a measure on the
space of absolutely continuous solutions to the differential equation
γ̇(t) = f˜(γ(t))
γ(0) = x0 ,
(84a)
x0 ∈ Dj (84b)
such that the conditionals ηx are dirac masses.
In addition, if λ is the Lebesgue measure on T , the measure ζtη ⊗ λ is the average occupation measure
defined in Eqn. (12). Here ζtη satisfies the following equality
⟨ζtη , φ⟩ = ∫
Dj ×Ξ
φ(γ(t)) dη(x, γ)
∀φ ∈ Cb (Dj ) (85)
Proof. The first part of this Lemma is an immediate consequence of Theorem 3.1 in [3]; the remainder of
this proof will establish the relation between the average occupation measure and solutions to the continuity
equation.
19
By definition, the following equality holds
(ζtη ⊗ λ)(A × B) = ∫ ∫
T
Dj ×Ξ
IA×B (t, x(t)) d(ζtη ⊗ λ)
(86)
Since η ∈ M+ (Dj ×Ξ), and this is a Polish space, we can decompose η to satisfy the following equality [6, Corr.
10.4.13]
dη = dηx dµ0
where µ0 is the distribution of initial conditions for the ODE in Eqn. (84) and ηx is a regular conditional
measure, conditioned on the initial value of x. Thus,
(ζtη ⊗ λt )(A × B) = ∫ ∫ ∫ IA×B (t, x(t ∣ x0 )) dηx dµ0 dt
(87)
T Dj Ξ
= ∫ IA (t) ∫ IB (x(t ∣ x0 )) dµ0 dt
(88)
Dj
T
= ∫ ∫ IA×B (t, x(t ∣ x0 )) dt dµ0
(89)
Dj T
= ∫ µj (A × B ∣ x0 ) dµ0
(90)
Dj
= µj (A × B);
(91)
where the transition from Eqn. (87) to Eqn. (88) uses the fact that ηx is a dirac measure; Fubini’s theorem
is used to get to Eqn. (89); and the definition of occupation measure and average occupation measure are
used to arrive at Eqns. (90) and (91) respectively.
Lemma 19. Let (µ0J , µTJ , µGE , µJ ) satisfy Eqn. (26). Then, there exists a family of absolutely continuous
state trajectories starting emanating from each µ0j , j ∈ J such the occupation and terminal measures (at
t = T ) in each mode generated by this family of trajectories coincide with µj and µTj , j ∈ J . In addition,
µGE s coincide with the restrictions of µJ on the respective guards.
Proof. This lemma is a generalization of [16, Lemma 3] and is proven using the same technique adopted
by the authors of [16]. In presenting this proof, it is assumed that uncertain parameters are augmented to
the state-space description of the system, similar to §3. In addition, given the description of hybrid systems
considered in this paper (refer §2.2), it is sufficient to restrict our attention to one particular mode of the
system, say mode j. For notational convenience, let σ0j denote the measure on the image of reset maps
originating from the guards of in H; and let the starting measure µsj be defined as
µsj = δ0 ⊗ µ0j ⊗ µθj + σ0j
(92)
In accordance with the hybrid system definition, trajectories of the state can arrive in mode j at time
τ ∈ [0, T ] via spt(µsj ); and once in the mode, can arrive at spt(µfj ) at any time t ∈ [0, T ], and hence
possibly leave the mode. Equation (20) has a conserved quantity – mass. To see this, consider the test
function φ(t) = tk with k = 0; then it follows that
µsj (A × B × C) = µfj (A × B × C).
(93)
That is, along the flow of solutions in each mode, mass is conserved and the Liouville eqn. is a CCE in terms
of the measure µj . The measures given by the statement of this Lemma satisfy the Liouville eqn. and hence
a CCE; thus solutions to the associated family of ODEs exists in each mode. To see that solutions to the
hybrid description of ODEs exists, refer to the results in [7]. Also, using Lemma 18, since µj (x, θ ∣ t) solves
the CCE in each mode, µj is the average occupation measure.
Having established that solutions to the ODE exists in every mode, we have to demonstrate that the
solution trajectories begin in ⋃j∈J spt(µ0j ) and terminate in ⋃j∈J spt(µTj ); this, again, can be demonstrated
20
in a per-mode basis. Since µj (⋅ ∣ t) solves a CCE, conditioned at t = 0, this measure coincides with the initial
distribution (ν0j ) of states and at t = T , it coincides with the final distribution of states; we have to show
that spt(ν0j ) = spt(µ0j ) and that spt(µj (⋅ ∣ T )) = X(T,j) . We do the same in the ensuing presentation.
Since we are concerned with solutions evolving on Polish spaces, we can decompose µj into the following
form [6, Corr. 10.4.13]
dµj (t, x, θ) = dµ̃j (x, θ ∣ t) dξµj (t)
(94)
where µ̃j (⋅ ∣ t) is the regular conditional measure and ξµj is the normalized projection of µj on T . We claim
that if λ is the Lebesgue measure on T , ξµj is a scaled version of λ. That is, we have to demonstrate that
the moments of ξµ are a scaled version of that of λ. Since each µj is an average occupation measure, by
definition (Eqn. (12)), we have that
T
µj (A × B × C) = ∫ ∫ IA×B×C (t, x(t∣τk , x0 , θ), θ) dt dν0j ,
0
T ×Dj ×Θj
(95)
where ν0j ∈ M+ (T × Dj × Θj ) is the initial distribution of states (not yet known to be related to µsj ). Hence,
the t-moments of µj are
T
∫
tk dµj =
∫
k
∫ t dt dν0j ,
(96)
T ×Dj ×Θj 0
T ×Dj ×Θj
=
T k+1
ν0 (T × Dj × Θj ),
k+1 j
(97)
which are scaled moments of the λ. Thus it follows that Eqn. (94) can be written as
dµj = dµj (x, θ ∣ t) dt.
(98)
Recall that in each mode, the supports of measures in Liouville eqn. have the following properties: spt(µsj ) ⊂
(T × Dj○ × Θj ); spt(µj ) ⊂ (T × Dj × Θj ), spt(µTj ) ⊂ X(T,j) / ⋃e∶=(j,k),e∈E Ge × Θj ; and spt(µGe ) ⊂ (T × Ge × Θj ).
Observe that only µj and µGe are supported on T × Ge × Θj ; it thus follows from Eqn. (26)that for all
e ∶= (j, k), ∀k ∈ {l ∣ (j, l) ∈ E}
µGe (T × Ge × Θj ) = L′fj µj (T × Ge × Θj ).
(99)
Since µj is an unsigned measure whose t-marginal is a scaled version of λ, it follows that for all A ⊂
T st. λ(A) = 0
0 = µj (A × Dj × Θj ) ≥ µj (A × Ge × Θj ).
(100)
From Eqns. (99) and (100), it follows that the t-marginal of µGe is absolutely continuous wrt. λ. Thus, the
measures on the guards can be decomposed as follows:
dµGe = we (t) dµGe (x, θ ∣ t) dt,
∀e ∈ E (101)
where we (t) is the density of the t-marginal of µGe wrt. to λ.
In each mode j ∈ J , σ0j is a sum of the push-forward of some measures on guards through reset maps that
are identity maps in the t-component; that is,
σ0j =
∗
Re∗ (π(t,x)
µGe ) ⊗ µθj .
∑
(102)
e∈{(k,j)∣E}
Now, by considering test functions of the form v(t, x, θ) = ϕ(t)φ(x, θ) in Eqn. (26), we get (by employing
integration by parts):
T
ϕ(T )αT − ϕ(0)α0 = ∫
0
ϕ̇αt + ϕ ⋅ (αf + ασ − αG ) dt
21
(103)
where
αT ∶=
[φ(x, θ)] dµTj ,
∫
(104)
X(T ,j) ×Θj
α0 ∶=
[φ(x, θ)] dµ0j dµθj ,
∫
(105)
X(0,j) ×Θj
αt (t) ∶=
∫
[φ(x, θ)] dµj (x, θ ∣ t),
(106)
[∇φ(x, θ) ⋅ fj ] dµj (x, θ ∣ t),
(107)
Dj ×Θj
αf (t) ∶=
∫
Dj ×Θj
ασ (t) ∶= ∑
∫ [we ∫
e∈{(k,j)∈E}G ×Θ
e
k
αG (t) ∶=
∑
e∈{(j,k)∈E}
∫
Θj
φ(Re (x), θ) dµθj ] dµGe (x, θ ∣ t),
[we φ(x, θ)] dµGe (x, θ ∣ t).
Ge ×Θj
(108)
(109)
Given functions αT , α0 , αf , ασ and αG , Eqn. (103) admits an dt-a.e. unique solution for αt . To see this,
suppose c1 (t) and c2 (t) are any two admissible solutions. Then
T
ϕ(T )αT − ϕ(0)α0 =∫ [ϕ̇c1 + ϕ(αf + ασ − αG )] dt
(110a)
0
T
ϕ(T )αT − ϕ(0)α0 =∫ [ϕ̇c2 + ϕ(αf + ασ − αG )] dt
(110b)
0
Taking the difference between Eqns. (110a) and (110b) and noting that ϕ(t) ∈ C(T ) is any arbitrary function
and that C(T ) is dense in L1 (T ), it follows that c1 (t) = c2 (t), dt-a.e..
By construction, with αT , α0 , αf , ασ and αG as defined, we know that αt (t) as defined in Eqn. (106)
satisfies Eqn. (103). We claim that α̃t (t) defined as
t
α̃t (t) = α0 + ∫ [αf + ασ − αG ] dτ
(111)
0
also solves Eqn. (103). Observe from Eqn. (111) that:
0
α̃t (0) = α0 + ∫ [αf + ασ − αG ] dτ = α0 ,
(112)
0
T
α̃t (T ) = α0 + ∫ [αf + ασ − αG ] dτ = αT
(113)
0
dα̃t
= αf + ασ − αG ,
dt
(114)
where the equality in Eqn. (113) follows from considering v = ϕ(x, θ) as the test function in Eqn. (26), and
the last equality follows from Leibniz rule.
Now consider evaluating ∫T (ϕ̇α̃t ) dt.
T
dα̃t
dt
dt
(115)
= ϕ(T )αT − ϕ(0)α0 − ∫ ϕ [αf + ασ − αG ] dt.
(116)
∫ (ϕ̇α̃t ) dt = ϕ(T )α̃t (T ) − ϕ(0)α̃t (0) − ∫ ϕ
0
T
T
0
22
This is of the form of Eqn. (103) and hence α̃t (t) solves Eqn. (103).
Since the solution to Eqn. (103) is dt-a.e. unique, it follows that, ∀φ ∈ C 1 (Dj × Θj ):
t
∫
φ(x, θ) dµj (x, θ ∣ t) = α0 + ∫ [αf + ασ − αG ] dτ
(117)
0
Dj ×Θj
From Eqn. (117) it follows that the stochastic kernel of µj at t = 0 coincides with µ0j , and that it coincides
with µTj at t = T . This implies that the solutions to the associated ODE, at time t = T terminate in
spt(µTj ) = X(T,j) .
To complete the proof, we now show that the restriction of µj onto the guards is, dt-a.e., identical to the
measure on the guards; further that ν0 (from Eqn. (95)) is identical to µsj (from Eqn. (93)).
Recall the definition of the average occupation measure (Eqn. (12)), its relation to the initial state
distribution µsj , and the form of the Liouville eqn. (Eqn. (26)); it is clear that (in mode j) the following
equality holds
ν0 + L′fj µj = δT ⊗ µT + µG ,
(118)
where µG is the sum of measures on all guards in mode j.
From Remark 8, the following equality holds
ν0 + L′fj µj = δT ⊗ µT +
∑
µj ∣G(j,k) .
(119)
k∈{l∣(j,l)∈E}
Taking the difference between Eqns. (118) and (119), noting that since µj is decomposable in the form of
Eqn. (98), so is its restriction, recalling that µG is also decomposable (Eqn. (101)), and considering the
product of arbitrary test functions in C 1 (T ) and C 1 (Dj × Θj ), it is concluded that the t-conditionals of µG
and µj are equal dt-a.e.. Hence, µG is the restriction of µj to the guards. That that ν0 is equal to µ0j + µsj
follows as a natural consequence of Eqn. (92) and our conclusion about the t-conditional of µj at t = 0.
References
[1] SPOTLESS. Available at https://goo.gl/n6CS7f.
[2] Matthias Althoff, Olaf Stursberg, and Martin Buss. Reachability analysis of nonlinear systems with uncertain parameters using conservative linearization. In 47th IEEE Conference on Decision and Control,
pages 4042–4048. IEEE, 2008.
[3] Luigi Ambrosio. Transport equation and cauchy problem for non-smooth vector fields. In Bernard
Dacorogna Paolo Marcellini, editor, Calculus of Variations and Nonlinear Partial Differential Equations,
volume 1927 of C.I.M.E. Foundation Subseries, page 206. Springer-Verlag Berlin Heidelberg, 1 edition,
2008.
[4] E.J. Anderson and P. Nash. Linear programming in infinite-dimensional spaces: theory and applications.
Wiley-Interscience series in discrete mathematics and optimization. Wiley, 1987.
[5] V.I. Bogachev. Measure Theory. Number v. 1 in Measure Theory. Springer Berlin Heidelberg, 2007.
[6] V.I. Bogachev. Measure Theory. Number v. 2 in Measure Theory. Springer Berlin Heidelberg, 2007.
[7] S. Burden, H. Gonzalez, R. Vasudevan, R. Bajcsy, and S. Shankar Sastry. Metrization and Simulation
of Controlled Hybrid Systems. IEEE Trans. Autom. Control, 60(9):2307–2320, 2015.
[8] Katie Byl. Metastable legged-robot locomotion. PhD thesis, Massachusetts Institute of Technology, 2008.
[9] G. Chesi. On the estimation of the domain of attraction for uncertain polynomial systems via lmis. In
Decision and Control, 2004. CDC. 43rd IEEE Conference on, volume 1, pages 881–886 Vol.1, Dec 2004.
23
[10] Michael Jon Coleman. A stability study of a three-dimensional passive-dynamic model of human gait.
Cornell University, May, 1998.
[11] Hongkai Dai and Russ Tedrake. Optimizing robust limit cycles for legged locomotion on unknown
terrain. In 2012 IEEE 51st Annual Conference on Decision and Control, pages 1207–1213. IEEE, 2012.
[12] Gerald B Folland. Real analysis: modern techniques and their applications. John Wiley & Sons, 2013.
[13] Antoine Girard. Reachability of uncertain linear systems using zonotopes. In Hybrid Systems: Computation and Control, pages 291–305. Springer, 2005.
[14] Ambarish Goswami, Thuilot Benoit, and Bernard Espiau. A study of the passive gait of a compass-like
biped robot: Symmetry and chaos. The International Journal of Robotics Research, 17:1282–1301, 1998.
[15] Brent Griffin and Jessy Grizzle. Walking gait optimization for accomodation of unknown terrain height
variations. In American Control Conference 2015, 2015.
[16] Didier Henrion and Milan Korda. Convex computation of the region of attraction of polynomial control
systems. IEEE Trans. Autom. Control, 59(2):297–312, 2014.
[17] Milan Korda, Didier Henrion, and Colin N. Jones. Inner approximations of the region of attraction for
polynomial dynamical systems. In 9th IFAC Symposium on Nonlinear Control Systems (NOLCOS),
Toulouse, France, September 4-6, 2013, 2013.
[18] John M Lee. Smooth manifolds. Springer, 2003.
[19] John N Maidens, Shahab Kaynama, Ian M Mitchell, Meeko MK Oishi, and Guy A Dumont. Lagrangian
methods for approximating the viability kernel in high-dimensional systems. Automatica, 49(7):2017–
2029, 2013.
[20] M. Maiga, N. Ramdani, L. Trave-Massuyes, and C. Combastel. A comprehensive method for reachability
analysis of uncertain nonlinear hybrid systems. IEEE Trans. Autom. Control, PP(99):1–1, 2015.
[21] Anirudha Majumdar, Ram Vasudevan, Mark M Tobenkin, and Russ Tedrake. Convex optimization of
nonlinear feedback controllers via occupation measures. The International Journal of Robotics Research,
page 0278364914528059, 2014.
[22] Ian Manchester, Mark Tobenkin, Michael Levashov, and Russ Tedrake. Regions of attraction for hybrid
limit cycles of walking robots. Proceedings of the 18th IFAC World Congress, 2011.
[23] Pablo A Parrilo. Structured semidefinite programs and semialgebraic geometry methods in robustness
and optimization. PhD thesis, Citeseer, 2000.
[24] J. W. Pitman. Occupation measures for markov chains. Advances in Applied Probability, 9(1):pp. 69–86,
1977.
[25] M. Posa, M. Tobenkin, and R. Tedrake. Stability analysis and control of rigid-body systems with
impacts and friction. IEEE Trans. Autom. Control, PP(99):1–1, 2015.
[26] Stephen Prajna and Ali Jadbabaie. Safety verification of hybrid systems using barrier certificates. In
Hybrid Systems: Computation and Control, pages 477–492. Springer, 2004.
[27] Cenk Oguz Saglam and Katie Byl. Switching policies for metastable walking. In 2013 IEEE 52nd
Annual Conference on Decision and Control (CDC), pages 977–983. IEEE, 2013.
[28] Victor Shia, Ram Vasudevan, Ruzena Bajcsy, and Russ Tedrake. Convex computation of the reachable
set for controlled polynomial hybrid systems. In 2014 IEEE 53rd Annual Conference on Decision and
Control (CDC), pages 1499–1506. IEEE, 2014.
24
[29] Christoffer Sloth and Rafael Wisniewski. Safety analysis of stochastic dynamical systems. IFACPapersOnLine, 48(27):62 – 67, 2015. Analysis and Design of Hybrid Systems ADHSAtlanta, GA, USA,
Oct. 14-16, 2015.
[30] Claire J Tomlin, Ian Mitchell, Alexandre M Bayen, and Meeko Oishi. Computational techniques for the
verification of hybrid systems. Proceedings of the IEEE, 91(7):986–1001, 2003.
[31] Ufuk Topcu, Andrew K. Packard, P. Seiler, and G.J. Balas. Robust region-of-attraction estimation.
IEEE Trans. Autom. Control, 55(1):137–142, Jan 2010.
25
| 3cs.SY
|
Output-sensitive Complexity of Multiobjective Combinatorial
Optimization
Fritz Bökler*1 , Matthias Ehrgott2 , Christopher Morris†1 , and Petra Mutzel1
1
arXiv:1610.07204v1 [math.OC] 23 Oct 2016
Department of Computer Science, TU Dortmund University, Dortmund, Germany,
{fritz.boekler, christopher.morris, petra.mutzel}@tu-dortmund.de
2
Department of Management Science, Lancester University, Lancester, United Kingdom,
[email protected]
Abstract
We study output-sensitive algorithms and complexity for multiobjective combinatorial optimization problems. In this computational complexity framework, an algorithm for a general enumeration problem is regarded efficient if it is output-sensitive, i.e., its running time is
bounded by a polynomial in the input and the output size. We provide both practical examples
of MOCO problems for which such an efficient algorithm exists as well as problems for which
no efficient algorithm exists under mild complexity theoretic assumptions.
Keywords Multiobjective Optimization, Combinatorial Optimization, Output-sensitive Complexity, Linear Programming
1 Introduction
Computational complexity theory in multiobjective optimization has been considered in different
shapes for quite a while. In this paper, we argue that in contrast to the traditional way of investigating the complexity of multiobjective optimization problems, especially multiobjective combinatorial
optimization (MOCO) problems, output-sensitive complexity theory yields deeper insights into the
complexity of these problems.
We stress here, that our discussions and results apply not only to MOCO problems, but also
to every multiobjective optimization problem with a well-defined and finite set to output. For
example, they also apply to multiobjective integer optimization problems with finite Pareto-fronts.
The problem on which we will exemplify our considerations is the following.
Definition 1.1 (Multiobjective Combinatorial Optimization Problem). An instance of a multiobjective combinatorial optimization problem is a pair (𝒮, 𝐶), where
∙ 𝒮 ⊆ {0, 1}𝑛 is the set of solutions, and
∙ 𝐶 ∈ Q𝑑×𝑛 is the objective function matrix.
The multiobjective combinatorial optimization problem consists of all its instances. The goal is to
enumerate for a given instance (𝒮, 𝐶) all minimal elements of 𝒴 = {𝐶𝑥 | 𝑥 ∈ 𝒮} with respect to
the canonical partial order on vectors.
*
Correspondence to: Department of Computer Science, TU Dortmund, Otto-Hahn-Str. 14, 44225 Dortmund,
Germany. E-mail: [email protected]. The author has been supported by the Bundesministerium
für Wirtschaft und Energie (BMWi) within the research project “Bewertung und Planung von Stromnetzen”
(promotional reference 03ET7505) and by DFG GRK 1855 (DOTS).
†
The author is funded by the German Science Foundation (DFG) within the Collaborative Research Center SFB 876
“Providing Information by Resource-Constrained Data Analysis”, project A6 “Resource-efficient Graph Mining”.
1
The canonical partial order on vectors is the componentwise less-or-equal partial order, denoted
as ≤. We call the minima of 𝒴 the Pareto-front, which we denote by 𝒴N . Points in 𝒴N are
called nondominated points. The preimage of the Pareto-front is called Pareto-set; solutions in the
Pareto-set are called Pareto-optimal or efficient solutions. Note that the MOCO problem defined
as above is different from the problem of enumerating the Pareto-set.
In the definition of the MOCO problem, we only allow rational numbers in the input as real
numbers cannot be encoded in our model of computation. But of course, if we show that these
problems are already hard on rational numbers, they do not become easier when extending the set
of numbers allowed.
When investigating the complexity of these problems, it is important to define the input size
precisely. As 𝒮 can be very large, it is supposed to be implicitly given. This accounts for the fact
that searching 𝒮 exhaustively is explicitly not what we want to do. Thus, the input size is supposed
to be 𝑛 plus the encoding length of the matrix 𝐶, denoted as ⟨𝐶⟩, which will be defined in the
preliminaries in Section 2.
In practice we also want to find a solution 𝑥 ∈ 𝒮 which maps to a point 𝑦 ∈ 𝒴N . Regarding the
formal definition above, this renders the problem to not be well-defined. Thus, when describing an
algorithm, we usually expect that also a solution is produced for each point of the Pareto-front.
When we prove a hardness result, Definition 1.1 gives us problems which are not harder than a
variant of the problem where we also want to find solutions. Hence, proving hardness of the problem
as defined in Definition 1.1 will lead to a hardness result for the problem including finding of a
representative solution.
1.1 Traditional Complexity Results for MOCO
In the past, complexity of MOCO problems has been investigated in the sense of NP-hardness.
The canonical decision problem of a MOCO problem is as follows.
Definition 1.2 (Canonical Decision Problem). Given a MOCO problem 𝑃 , the canonical decision
problem 𝑃 Dec is defined as follows: We are given an instance (𝒮, 𝐶) of 𝑃 and a vector 𝑘 ∈ Q𝑑 and
the goal is to decide whether there exists an 𝑥 ∈ 𝒮, such that 𝐶𝑥 ≤ 𝑘. The input size is supposed
to be 𝑛 + ⟨𝐶⟩ + ⟨𝑘⟩.
Obviously, any MOCO problem that is NP-hard for 𝑑 = 1 is also NP-hard for 𝑑 ≥ 2. Moreover,
known results in the literature show that even for MOCO problems with 𝑑 = 2 objectives 𝑃 Dec
is NP-hard, even for problems that are in P when 𝑑 = 1 such as the biobjective shortest path
problem (Serafini, 1986), biobjective minimum spanning tree problem (Camerini, Galbiati, and
Maffioli, 1984), biobjective assignment problem (Serafini, 1986), and biobjective uniform matroid
problem (Ehrgott, 1996). The proof of Proposition 3.2 also shows that 𝑃 Dec for a MOCO problem
with two objectives and no constraints is NP-complete.
One remark is in order here. It is not clear if the NP-hardness of 𝑃 Dec implies the NP-hardness
of the MOCO problem 𝑃 . In fact, Serafini (1986) proved the polynomial-time equivalence of 𝑃
and 𝑃 Dec only for the case where ||𝐶𝑥 − 𝐶𝑥′ ||∞ < 𝛿𝑛𝑙 for every two solutions 𝑥, 𝑥′ ∈ 𝒮 and some
𝛿, 𝑙 ≥ 1.
Furthermore, we note that 𝑃 Dec for 𝑑 = 2 is the same as the decision problem of the so-called
resource-constrained combinatorial optimization problem
(𝒮, 𝑐, 𝑟, 𝑟^),
where 𝒮 is as in Definition 1.1, 𝑐 ∈ Q𝑛 , 𝑟 ∈ Q𝑛 , 𝑟^ ∈ Q, and the goal is to find a minimizer
of {𝑐𝑇 𝑥 | 𝑥 ∈ 𝒮, 𝑟𝑇 𝑥 ≤ 𝑟^}. This observation reveals that the canonical decision problem for a
MOCO problem with 𝑑 = 2 and that for a related resource-constrained combinatorial optimization
2
problem are identical, despite the fundamental difference in the goals of these two problems. We
can therefore question, whether studying the NP-hardness of 𝑃 Dec can provide much insight in
the hardness of MOCO problems on top of insights into the hardness of resource-constrained single
objective combinatorial optimization problems.
Another important consideration is the size of the Pareto-front. Many researchers have provided
instances of MOCO problems for which the size of 𝒴N is exponential, we refer to (Hansen, 1979)
for the biobjective shortest path problem, (Hamacher and Ruhe, 1994) for the biobjective minimum spanning tree problem, (Ruhe, 1988) for the biobjective integer minimum-cost flow problem,
and (Ehrgott, 2005) for the biobjective unconstrained combinatorial optimization problem. This
consideration also raises the question, if the NP-hardness of a MOCO problem 𝑃 which has a
Pareto-front of exponential size, does add information to the picture. Because NP-hardness is still
only an indication that 𝑃 cannot be solved in polynomial time, but 𝑃 having a Pareto-front of
exponential size is an actual proof of this fact. Despite these negative results, a few results on
MOCO problems with a Pareto-front of polynomial size or which belong to P are known. We refer
to (Figueira et al., 2016) in this issue for more details.
This brief overview on complexity results for MOCO problems reveals what one could consider
a bleak picture: Multiobjective versions of even the simplest combinatorial optimization problems
have Pareto-fronts of exponential size, and thus there is no prospect for finding polynomial time
algorithms to determine the Pareto-front. A closer look at some of the pathological biobjective
instances does, however, reveal, that despite the exponential size of the Pareto-front, all elements
of the Pareto-front lie on a single line in Q2 . Hence, we might hypothesize that traditional worst
case complexity analysis, which measures complexity in relation to input size, is not the right tool to
investigate the complexity of MOCO problems. Instead, we propose to investigate output-sensitive
complexity.
1.2 The Pareto-front Size within the Smoothed Analysis Framework
Before we move forward to output-sensitive complexity, we need to adress one practical objection
regarding MOCO problems. It is often argued that computing the entire Pareto-front of a MOCO
problem is too costly to pursue, because it can have exponential size in the worst-case. But in
practice, it was observed that usually the situation is not so bad when the number of objectives is
small. This observed discrepancy between practice and the traditional worst case analysis motivates
stochastic running time analysis, where the inputs are drawn from a certain distribution. One
prominent stochastic running time analysis framework, so called Smoothed Analysis, can be used
to bound the expected worst-case size of the Pareto-front of a MOCO problem.
In classical running time analysis, we play against an adversary who gives us ill posed instances
in the sense that the running time is very high or the Pareto-front is very large. Smoothed Analysis
aims at weakening this adversary. In the context of multiobjective optimization, a model by Ackermann et al. (2007) was used to show the best expected bound on the smoothed worst-case size
of the Pareto-front of general MOCO problems.
Instead of all the objective function matrix entries, the adversary is only allowed to choose the
first row of it deterministically and for all remaining rows 𝑖 ∈ {2, . . . , 𝑑} and columns 𝑗 ∈ [𝑛] it may
choose a probability density function 𝑓𝑖,𝑗 : [0, 1] → R, describing how potential entries are drawn.
The adversary is not allowed to choose these densities arbitrarily, because otherwise it could again
choose the coefficients deterministically. Rather, the 𝑓𝑖,𝑗 are bounded by a model parameter 𝜑.
Consequently, the adversary gives us a set of solutions 𝒮 ⊆ {0, 1}𝑛 , the first row of the objective
function matrix 𝑐 ∈ Q𝑛 and the probability densities 𝑓𝑖,𝑗 . Brunsch and Röglin (2012) showed that
the expected size of the Pareto-front of these instances is at most 𝒪(𝑛2𝑑 𝜑𝑑 ).
Thus, when we fix the number of objectives, we can expect to have only polynomial Pareto-front
sizes of any MOCO problem in practice. This emphasizes the need for output-sensitive algorithms,
3
because when the Pareto-front is small, we want our algorithms to be fast and when the Pareto-front
is large, we want them to be not too slow.
1.3 Organization
In the remainder of the paper, we will first give some definitions in Section 2. In particular, we will
give necessary definitions from theoretical computer science, including output-sensitive complexity.
Afterwards, we will be concerned with a sufficient condition for a MOCO problem to be efficiently
solvable in the sense of output-sensitive complexity in Section 3. This sufficient condition can be
applied to the multiobjective global minimum-cut problem which will be our first example of an
efficiently solvable MOCO problem.
We will show that the multiobjective shortest path problem is an example of a problem which is
not efficiently solvable under weak complexity theoretic assumptions in Section 4. Moreover, we will
demonstrate a general method from output-sensitive complexity for showing that an enumeration
problem is not efficiently solvable under the assumption that P ̸= NP.
In Section 5, we will again be concerned with efficiently solvable multiobjective optimization
problems. We will discuss some connections between MOCO problems and multiobjective linear
programming. We will show that the impression that computing extreme points of the Pareto-front
is easier than computing the whole Pareto-front is indeed true in the world of output-sensitive
complexity, at least for the multiobjective shortest path problem. Moreover, we review some recent
results from the literature about output-sensitive algorithms for computing supported solutions
and extreme nondominated points and also provide some new results for the special case of 𝑑 = 2.
2 Preliminaries
In the following section we fix notation. Let 𝑣 be a vector in R𝑛 , then we denote its 𝑖-th component
by 𝑣𝑖 . We assume that all vectors are column vectors. The 𝑖-th unit vector is denoted by 𝑒𝑖 . Let
𝑀 be a matrix in R𝑚×𝑛 , then we denote its 𝑖-th row by 𝑀𝑖 and the scalar in the 𝑖-th row and the
𝑗-th column by 𝑚𝑖𝑗 . The transpose of 𝑀 is denoted by 𝑀 𝑇 . For 𝑟 in R, the ceiling function is
⌈𝑟⌉ := min {𝑧 ∈ Z | 𝑧 ≥ 𝑟} and let [𝑛] := {0, . . . , 𝑛} ⊂ N and [𝑛 : 𝑚] := [𝑛, 𝑚] ∩ N.
Let (𝑆, ⪯) be a partially ordered set. Then 𝑠 <lex 𝑡 for 𝑠 and 𝑡 in 𝑆 𝑝 , 𝑝 ∈ N, if there is 𝑗 in [1 : 𝑝]
such that
1. 𝑠𝑖 = 𝑡𝑖 for 𝑖 < 𝑗, and
2. 𝑠𝑗 < 𝑡𝑗 .
2.1 Theoretical Computer Science
In theoretical computer science, it is important to define the model of computation used for running
time analyses. In the complexity of enumeration problems, the model of computation usually is
the Random Access Machine (RAM). It can be concieved as a formal model of a computer, with a
fixed and small set of instructions and an infinite number of memory cells. These memory cells are
only allowed to hold integer numbers with up to 𝒪(𝑛𝑙 ) bits on inputs of size 𝐿 for some constant
𝑙 ≥ 1 to account for the precision needed in optimization. For more details, we refer the reader to
the book by Cormen et al. (2001, 23 sqq.).
Moreover, we also need to discuss encoding lengths of numbers. The following is largely based
on Grötschel, Lovász, and Schrijver (1988, 29 sqq.). For 𝑧 ∈ Z ∖ {0}, the encoding length of 𝑧, i.e.,
an upper bound for the minimum number of bits to encode 𝑧 in binary representation, is
⟨𝑧⟩ := 1 + ⌈log2 (|𝑧| + 1)⌉ .
4
That is, one bit to represent the sign and ⌈log2 (|𝑧| + 1)⌉ bits to encode the binary representation
of the absolute value of 𝑧. For 𝑧 = 0 one bit is sufficient, i.e., ⟨0⟩ := 1. Let 𝑞 ∈ Q and let 𝑞 = 𝑎/𝑏
for 𝑎 and 𝑏 ∈ Z such that 𝑎 and 𝑏 are coprime, and 𝑏 > 0, then the encoding length of 𝑞 is
⟨𝑞⟩ := ⟨𝑎⟩ + ⟨𝑏⟩ .
Let 𝑣 be a vector in Q𝑛 , then ⟨𝑣⟩ := 𝑛 + 𝑛𝑖=1 ⟨𝑣𝑖 ⟩. Moreover, let 𝑀 be a matrix in Q𝑚×𝑛 , then
∑︀
⟨𝑀 ⟩ := 𝑚𝑛 + 𝑚
𝑖=1 ⟨𝑀𝑖 ⟩.
Due to the special importance of polynomial running time, we define the set of functions not
growing faster than a polynomial function as follows: Let 𝑓 : R𝑑 → R be a function, then 𝑓 is in
poly(𝑛1 , . . . , 𝑛𝑑 ) if there is a polynomial function 𝑝 : R𝑑 → R such that 𝑓 in 𝒪(𝑝(𝑛1 , . . . , 𝑛𝑑 )). For
a more concise introduction to formal languages and complexity theory, e.g., NP and co-NP, we
refer to (Arora and Barak, 2009).
∑︀
2.2 Introduction to Output-sensitive Complexity
Classically, the running time of an algorithm is measured with respect to the input size. An
algorithm which has a running time which is polynomially-bounded in the input size is widely
regarded as efficient. For certain problems, such as MOCO problems, the size of the output varies
widely or even becomes exponentially large in the input size.
Here, the notion of output-sensitive algorithms comes into play, where running time is not only
measured in the input size but also in the size of the output, cf. (Johnson, Yannakakis, and Papadimitriou, 1988) for examples of output-sensitive algorithms. The point here is that we can view
the MOCO problem of Definition 1.1 as an enumeration problem, i.e., enumerating the elements of
the (finite) Pareto-front, see Section 2.3. This task is solved by an enumeration algorithm, i.e., an
algorithm that outputs every element of the the Pareto-front exactly once.
In the following, we give a formal definition of the terms enumeration problem and enumeration
algorithm. Moreover, we define complexity classes for enumeration problems. This section is largely
based on (Schmidt, 2009).
Definition 2.1 (cf. (Schmidt, 2009, 7 sq.)). An enumeration problem is a pair (𝐼, 𝐶), such that
1. 𝐼 ⊆ 𝛴 * is a language for some fixed alphabet 𝛴,
2. 𝐶 : 𝐼 → 𝛴 * maps each instance 𝑥 in 𝐼 to its configurations 𝐶(𝑥), and
3. the encoding length ⟨𝑠⟩ for 𝑠 in 𝐶(𝑥) for 𝑥 in 𝐼 is in poly(𝑥).
We assume that 𝐼 is decidable in polynomial time and that 𝐶 is computable.
If the reader is not familiar with the term language, the set 𝛴 * can be interpreted as the set of
all finite strings over {0, 1}. This is important, because all instances and all configurations need to
be encoded by finite strings, while the exact encoding is not relevant here. The third requirement
means, that the configurations need to be compact, i.e., the encoding length of a configuration of
an instance 𝑥 should be polynomially bounded in the size of 𝑥. The reason for this is to limit the
number of configurations to be at most exponentially many.
Definition 2.2 (cf. (Schmidt, 2009, p. 8)). Let 𝐸 = (𝐼, 𝐶) be an enumeration problem. An
enumeration algorithm for 𝐸 is a RAM that
1. on input 𝑥 in 𝐼 outputs 𝑐 in 𝐶(𝑥) exactly once, and
2. on every input terminates after a finite number of steps.
Let 𝑥 in 𝐼 and let |𝐶(𝑥)| = 𝑘, then the 0-th delay is the time before the first solution is output, the
𝑖-th delay for 𝑖 in [1 : 𝑘 − 1] is the time between the output of the 𝑖-th and the (𝑖 + 1)-th solution,
and the 𝑘-th delay is the time between the last output and the termination of the algorithm.
5
The following complexity classes can be used to classify enumeration problems, cf. (Johnson,
Yannakakis, and Papadimitriou, 1988).
Definition 2.3 (cf. (Schmidt, 2009, p. 12)). Let 𝐸 = (𝐼, 𝐶) be an enumeration problem. Then 𝐸
is in
1.
2.
3.
4.
TotalP (Output-Polynomial Time/Polynomial Total Time)1 ,
IncP (Incremental Polynomial Time),
DelayP (Polynomial Time Delay),
PSDelayP (Polynomial Time Delay with Polynomial Space)
if there is an enumeration algorithm such that
1.
2.
3.
4.
its running time is in poly(|𝑥|, |𝐶(𝑥)|) for 𝑥 in 𝐼,
its 𝑖-th delay for 𝑖 in [𝑛] is in poly(|𝑥|, |𝐶 𝑖 (𝑥)|) for 𝑥 in 𝐼,
its 𝑖-th delay for 𝑖 in [𝑛] is in poly(|𝑥|, |𝐶 𝑖 (𝑥)|) for 𝑥 in 𝐼, and
the same as in 3. holds and the algorithm requires at most polynomial space in the input size,
respectively, where 𝐶 𝑖 (𝑥) denotes the set of solution that have been output before the 𝑖-th solution
has been output.
We say that an algorithm is output-sensitive if it fulfills condition 1 of the above definition.
Moreover, we say that an algorithm has incremental delay (polynomial delay) if it fulfills the second
(third) condition of the above definition. The following hierarchy result holds.
Theorem 2.4 (Schmidt, 2009, 14 sqq.).
PSDelayP ⊆ DelayP ⊆ IncP ⊂ TotalP .
An enumeration algorithm will be regarded efficient if it is output-sensitive.
2.3 MOCO Problems are Enumeration Problems
We will now explain, how MOCO problems as defined in Definition 1.1 relate to enumeration
problems as defined in Definition 2.1. More formally, we will show that the general MOCO problem
is an enumeration problem. Given a MOCO instance (𝒮, 𝐶) we need to define the set 𝐼 of instances
of the enumeration problem and the configuration mapping.
To define 𝐼, we fix a binary representation of (𝒮, 𝐶) with encoding length of at most 𝒪(𝑛 + ⟨𝐶⟩).
Then 𝐼 is the set of all such encodings of instances of our MOCO problem. The configuration
mapping now maps an instance to its Pareto-front 𝒴N , by fixing an arbitrary encoding of rational
numbers with at most 𝒪(⟨𝑟⟩) bits for a rational number 𝑟. We observe that the value of each com∑︀
ponent 𝑗 of each point of the Pareto-front is at most 𝑛𝑖=1 |𝐶𝑗𝑖 | ∈ 𝒪(⟨𝐶⟩) and thus the encoding of
each point in the Pareto-front is polynomial in the instance encoding size. Thus, the configurations
of a MOCO problem are the points of the Pareto-front and not, for example, the solutions of the
MOCO problem.
Solving the so defined enumeration problem will give us the set of configurations, i.e., the Paretofront in the given encoding. This corresponds exactly to our definition of solving a MOCO problem.
3 A Sufficient Condition for Output-sensitivity from the MOCO Literature
In this section, we will show a sufficient condition for a MOCO problem being solvable in outputpolynomial time, which follows directly from the multiobjective optimization literature. A broad
1
Although we use the term output-polynomial time in the remaining text, we abbreviate it to TotalP for historical
and notational reasons.
6
subject of investigation in the multiobjective optimization literature is the 𝜀-constraint scalarization:
min{1𝑇 𝐶𝑥 | 𝑥 ∈ 𝒮, 𝐶1𝑇 𝑥 ≤ 𝜀1 , . . . , 𝐶𝑑𝑇 𝑥 ≤ 𝜀𝑑 }, for 𝜀 ∈ Q𝑑 .
It is well known that the complete Pareto-front can be found using this scalarization, cf. (Chankong
and Haimes, 1983), but the scalarization itself can be hard to solve, cf. (Ehrgott, 2006). In the past,
authors have proven several bounds on the number of 𝜀-constraint scalarizations needed to find the
Pareto-front of a general MOCO problem. It is possible to construct a sufficient condition for
enumerating the Pareto-front of a MOCO problem in output-polynomial time from these results.
One of the first such bounds is due to Laumanns, Thiele, and Zitzler (2006), who prove that
at most 𝒪(|𝒴N |𝑑−1 ) single-objective 𝜀-constraint scalarization problems need to be solved. The
currently best known bound is due to Klamroth, Lacour, and Vanderpooten (2015), who prove a
𝑑
bound of 𝒪(|𝒴N |⌊ 2 ⌋ ).
From these results we can formulate the following corollary.
Corollary 3.1. If the 𝜀-constraint scalarization of a given MOCO problem 𝑃 can be solved in
polynomial time for a fixed number of objectives, 𝒴N can be enumerated in output-polynomial time
for a fixed number of objectives.
One example for such a problem is the multiobjective global minimum-cut problem. Armon
and Zwick (2006) showed, that the 𝜀-constraint version of the problem can be solved in time
𝒪(𝑚𝑛2𝑑 log 𝑛). Thus, the multiobjective global minimum-cut problem is an example of a multiobjective optimization problem that can be solved in output-polynomial time for each fixed number
of objectives.
Now, the question arises if this condition is also necessary, i.e., can we show that a MOCO
problem cannot be solved in output-polynomial time by showing that the 𝜀-constraint scalarization
cannot be solved in polynomial time. But as the following proposition shows, the condition of
Corollary 3.1 is not necessary, unless P = NP.
Proposition 3.2. There is a MOCO problem 𝑃 ∈ IncP with 𝑃 Dec being NP-complete.
Proof. We consider the following biobjective problem
(𝑃 ) min{(𝑐𝑇 𝑥, −𝑐𝑇 𝑥) | 𝑥 ∈ {0, 1}𝑛 }
with 𝑐 ∈ N𝑛 . We will first show NP-hardness and membership in NP of 𝑃 Dec and then show
membership in IncP of 𝑃 .
Recall that 𝑃 Dec is defined as the decision problem in which we are given an instance (𝒮, 𝐶) of
𝑃 and a vector 𝑘 ∈ Q𝑑 and the goal is to decide whether there exists an 𝑥 ∈ 𝒮, such that 𝐶𝑥 ≤ 𝑘.
Setting
(︃
)︃
1
1
𝑘 := 1𝑇 𝑐
,
−1
2
we can reformulate 𝑃 Dec in the following way: Decide for a given vector 𝑐 ∈ N𝑛 if there is an
∑︀
𝑥 ∈ {0, 1}𝑛 , such that 𝑐𝑇 𝑥 = 21 𝑛𝑖=1 𝑐𝑖 . We observe that this decision problem is the well-known
partition problem shown by Garey and Johnson (1979) to be NP-hard. It follows from Definition 1.2
that 𝑃 Dec is in NP and thus 𝑃 Dec is NP-complete.
Enumerating the Pareto-front of 𝑃 can be done in incremental polynomial time by employing a
recursive enumeration scheme on the first 𝑖 variables (fixing an arbitrary variable ordering). Setting
all variables to 0 yields point (0, 0)𝑇 . Now we fix all variables after the 𝑖-th to 0 and allow the first
𝑖 variables to vary. Let 𝐹𝑖 be the Pareto-front for this restricted problem. If we know 𝐹𝑖 , we can
compute 𝐹𝑖+1 by computing 𝐹𝑖 ∪ (𝐹𝑖 + {𝑐𝑖+1 }). Each such step yields time 𝒪(|𝐹𝑖 | log |𝐹𝑖 |) and we
have 𝐹0 ( 𝐹1 ( . . . ( 𝐹𝑛 , i.e., we find at least one new point in each iteration.
7
4 Example of a Hard Problem: Multiobjective Shortest Path
In this section, we will classify a problem as not efficiently solvable under mild complexity theoretic
assumptions in the sense of output-sensitive complexity, namely the multiobjective shortest path
problem, or more precisely, the multiobjective 𝑠-𝑡-path problem.
Definition 4.1 (Multiobjective 𝑠-𝑡-path problem (MOSP)). Given a directed graph 𝐺 = (𝑉, 𝐸)
where 𝐸 ⊆ 𝑉 × 𝑉 , two nodes 𝑠, 𝑡 ∈ 𝑉 , and a multiobjective arc-cost function 𝑐 : 𝐸 → Q𝑑≥ . A
∑︀
feasible solution is a path 𝑃 from node 𝑠 to node 𝑡. We set 𝑐(𝑃 ) := 𝑒∈𝑃 𝑐(𝑒). The set of all
𝑠-𝑡-paths is 𝒫𝑠,𝑡 . The goal is to enumerate the minima of 𝑐(𝒫𝑠,𝑡 ) with respect to the canonical
partial order on vectors.
It has long been investigated in the multiobjective optimization community. One of the first
systematical studies was by Hansen (1979). In practice, MOSP can be solved relatively well. But
usually algorithms which solve MOSP solve a more general problem which we call the multiobjective
single-source shortest-path (MO-SSSP) problem. In the MO-SSSP problem, instead of enumerating
one Pareto-front of the Pareto-optimal 𝑠-𝑡-paths, we enumerate for every vertex 𝑣 ∈ 𝑉 ∖{𝑠} the
Pareto-front of all Pareto-optimal 𝑠-𝑣-paths. One example of such an algorithm is the well known
label setting algorithm by Martins (1984), which solves the MO-SSSP problem in incremental
polynomial time.
We stress here that MOSP and MO-SSSP are indeed different problems. The insight that the
MO-SSSP problem is solvable in incremental polynomial time does not immediately transfer to the
MOSP problem. This can be seen, when using Martins’ algorithm for solving MOSP on a given
instance, where we cannot even guarantee output-polynomial running time.
More modern label setting and label correcting algorithms usually start from there and try to
avoid enumerating too many unneccessary 𝑠-𝑣-paths. The following considerations show that a large
number of these 𝑠-𝑣-paths are needed in these algorithms and that—in contrast to the MO-SSSP
problem—there is no ouput-sensitive algorithm solving MOSP in general, if we assume P ̸= NP.
Moreover, we will exemplify a general methodology in output-sensitive complexity for showing
that there is no output-sensitive algorithm unless P = NP. Similar to single objective optimization,
we show that an enumeration problem is hard by showing the hardness of a decision problem, which
is defined as follows.
Definition 4.2 (Finished Decision Problem). Given an enumeration problem 𝐸 = (𝐼, 𝐶) the
finished decision problem 𝐸 Fin is the following problem: We are given an instance 𝑥 ∈ 𝐼 of the
enumeration problem and a subset 𝑀 ⊆ 𝐶(𝑥) and the goal is to decide if 𝑀 = 𝐶(𝑥).
That is, when we are given an instance 𝑥 ∈ 𝐼 of the enumeration problem (𝐼, 𝐶) and a subset
𝑀 ⊆ 𝐶(𝑥) of the configuration set, we ask the question: Do we already have all configurations? As
commonly done in theoretical computer science, we will identify a decision problem 𝑃 with its set
of “Yes”-instances. So for an instance 𝑥 of 𝑃 , we write 𝑥 ∈ 𝑃 if 𝑥 is a “Yes”-instance and 𝑥 ∈
/ 𝑃 if
𝑥 is a “No”-instance of 𝑃 .
Lawler, Lenstra, and Rinnooy Kan (1980) showed that if there is an output-polynomial algorithm for an enumeration problem, then we can solve the associated finished decision problem in
polynomial time. Or equivalently: If the finished decision problem cannot be solved in polynomial
time, then we cannot solve the enumeration problem in output-polynomial time.
Lemma 4.3 (Lawler, Lenstra, and Rinnooy Kan, 1980). If 𝐸 ∈ TotalP then 𝐸 Fin ∈ P.
Proof. Let 𝐸 = (𝐼, 𝐶) be given. Since 𝐸 ∈ TotalP, there exists a polynomial function 𝑝 and a
RAM 𝐴 which enumerates 𝐶(𝑥) for a given 𝑥 ∈ 𝐼 in time at most 𝑝(|𝑥|, |𝐶(𝑥)|).
We will now construct an algorithm which will decide for a given instance (𝑥, 𝑀 ) of 𝐸 Fin whether
it is a “Yes”- or “No”-instance. We simulate 𝐴 on 𝑥 for time 𝑝(|𝑥|, |𝑀 |). If 𝐴 does not halt on 𝑥,
8
s
1
c1
0
1
c2
0
...
0
c21
t
0
c22
k1 + 1
0
0
c2 1 − k2 + 1
T
v
Figure 1: Showing the reduction in the proof for Theorem 4.4. Arcs with no label have cost 0.
then |𝑀 | < |𝐶(𝑥)| and we can safely answer “No”. If 𝐴 does halt on 𝑥, we are almost done. It
can still be the case that the input 𝑀 was invalid, i.e., 𝑀 ∖𝐶(𝑥) ̸= ∅ or that 𝑀 ( 𝐶(𝑥) and the
algorithm was by chance faster than expected. Both conditions can be tested by checking whether
𝑀 is equal to the output of 𝐴. If it is, we return “Yes”, otherwise “No”.
Simulation takes time poly(|𝑥|, |𝑀 |). The output of 𝐴 has at most size poly(|𝑥|, |𝑀 |) and the
final check can thus be done in poly(|𝑥|, |𝑀 |).
We can now use this method to finally show that the MOSP problem is indeed a hard enumeration
problem. For a definition of co-NP, we refer the reader to the book by Arora and Barak (2009).
Theorem 4.4. There is no output-sensitive algorithm for the MO 𝑠-𝑡-path problem unless P = NP,
even if the input graph is outerplanar.
Proof. We will show that MOSPFin is co-NP-hard. By Lemma 4.3 this shows that a TotalP
algorithm for MOSP implies P = co-NP and thus P = NP.
We reduce instances of the complement of the Knapsack problem:
𝑇
𝑇
(KP) {(𝑐1 , 𝑐2 , 𝑘1 , 𝑘2 ) | 𝑐1 𝑥 ≤ 𝑘1 , 𝑐2 𝑥 ≥ 𝑘2 , 𝑥 ∈ {0, 1}𝑛 } .
𝑇
𝑇
Without loss of generality, we can assume that 𝑐1 , 𝑐2 ∈ N𝑛 , 𝑘1 , 𝑘2 ∈ N, 𝑐1 1 > 𝑘1 and 𝑐2 1 >
𝑘2 . The problem above is still NP-complete under these restrictions, cf. (Kellerer, Pferschy, and
Pisinger, 2004).
We now construct an instance 𝐼^ of the MOSPFin problem from an instance 𝐼 of the KP problem.
1 . It has one arc
The instance has nodes {𝑣𝑖1 , 𝑣𝑖2 } for every variable 𝑥𝑖 and one additional node 𝑣𝑛+1
1 ) with
for every 𝑖 ∈ [1 : 𝑛] : (𝑣𝑖1 , 𝑣𝑖2 ) with cost (𝑐1𝑖 , 0)𝑇 and for every 𝑖 ∈ [1 : 𝑛] it has one arc (𝑣𝑖1 , 𝑣𝑖+1
2
𝑇
2
1
1
1
cost (0, 𝑐𝑖 ) and one arc (𝑣𝑖 , 𝑣𝑖+1 ) with cost 0. The node 𝑣1 is identified with 𝑠, the node 𝑣𝑛+1
is
𝑇
identified with 𝑡. There is one additional arc (𝑠, 𝑡) with cost (𝑘1 + 1, 0) and one additional path
𝑇
𝑇
(𝑠, 𝑣, 𝑡) with cost (0, 𝑐2 1 − 𝑘2 + 1). To complete the reduction, we set 𝑀 := {(𝑘1 + 1, 0)𝑇 , (0, 𝑐2 1 −
𝑘2 + 1)𝑇 }. An example of this reduction can be seen in Figure 1.
We observe that the instance is valid, and that there are at least two Pareto-optimal paths,
𝑇
namely (𝑠, 𝑡) and (𝑠, 𝑣, 𝑡) having cost (𝑘1 + 1, 0)𝑇 and (0, 𝑐2 1 − 𝑘2 + 1)𝑇 , respectively. Both paths
are Pareto-optimal because 𝑐1𝑖 , 𝑐2𝑖 > 0 for all 𝑖 ∈ [1 : 𝑛] and thus all other paths have either non-zero
components in their objective function values or the sum of all values in one objective function and
0 in the other. All steps can be performed in polynomial time in the input instance 𝐼.
𝑇
Now, we take an instances 𝐼 ∈ KP. Accordingly, there exists 𝑥 ∈ {0, 1}𝑛 with 𝑐1 𝑥 ≤ 𝑘1 and
𝑇
𝑇
𝑇
𝑐2 𝑥 ≥ 𝑘2 ⇔ 𝑐2 (1 − 𝑥) ≤ 𝑐2 1 − 𝑘2 . Using this solution, we construct a path 𝑃 in the MOSP
^ For every 𝑖 with 𝑥𝑖 = 1, we take the route through node 𝑣 2 , inducing cost of (𝑐1 , 0)𝑇 .
instance 𝐼:
𝑖
𝑖
9
1 ), inducing cost (0, 𝑐2 )𝑇 .
For every 𝑖 with 𝑥𝑖 = 0, we take the route directly through arc (𝑣𝑖1 , 𝑣𝑖+1
𝑖
𝑇
𝑇
𝑇
We observe that 𝑐1 (𝑃 ) = 𝑐1 𝑥 ≤ 𝑘1 and 𝑐2 (𝑃 ) = 𝑐2 (1 − 𝑥) ≤ 𝑐2 1 − 𝑘2 . But then, 𝐼^ ∈
/ MOSPFin ,
since 𝑃 is neither dominated by (𝑠, 𝑡) nor (𝑠, 𝑣, 𝑡).
Now, suppose for some instance 𝐼, the constructed instance 𝐼^ ∈
/ MOSPFin , i.e., there is an
additional nondominated path 𝑃 apart from (𝑠, 𝑡) and (𝑠, 𝑣, 𝑡). Since it is not dominated by (𝑠, 𝑡)
𝑇
and (𝑠, 𝑣, 𝑡), it must hold that 𝑐1 (𝑃 ) ≤ 𝑘1 and 𝑐2 (𝑃 ) ≤ 𝑐2 1 − 𝑘2 . But then, we can construct a
solution to KP in 𝐼 as follows: The path 𝑃 cannot take arcs from (𝑠, 𝑡) or (𝑠, 𝑣, 𝑡), so it needs to
take the route through the 𝑣𝑖1 and 𝑣𝑖2 nodes. For every 𝑖 ∈ [1 : 𝑛] it can only either take arc (𝑣𝑖1 , 𝑣𝑖2 )
1 ). If it takes the first arc, we set 𝑥 := 1; if it takes the second arc, we set 𝑥 := 0. This
or (𝑣𝑖1 , 𝑣𝑖+1
𝑖
𝑖
𝑇
𝑇
𝑇
𝑇
𝑇
solution then has cost 𝑐1 𝑥 = 𝑐1 (𝑃 ) ≤ 𝑘1 and 𝑐2 𝑥 = 𝑐2 1 − 𝑐2 (𝑃 ) ≥ 𝑐2 1 − 𝑐2 1 + 𝑘2 = 𝑘2 .
Hence, 𝐼 ∈ KP.
This proves that the reduction is a polynomial time reduction from the complement of KP to
the finished decision variant of MOSP and thus the theorem.
The proof also shows that deciding whether the Pareto-front of a MO 𝑠-𝑡-path instance is larger
than 2 is NP-hard. Moreover, it gives us a lower bound on the quality we can approximate the
size of the Pareto-front.
Corollary 4.5. It is co-NP-hard to approximate the size of the Pareto-front of the MO 𝑠-𝑡-path
problem within a factor better than 23 .
Since MOSP can be seen as a special case of many important MOCO problems, e.g., the multiobjective minimum perfect matching problem and the multiobjective minimum-cost flow problem,
these problems also turn out to be hard.
5 Enumerating Supported Efficient Solutions and Nondominated Extreme
Points
In the history of multiobjective optimization, it has been repeatedly observed that in the case of
two objectives a certain subset of the Pareto-front can be enumerated efficiently. This subset is
usually defined as the set of Pareto-optimal solutions that can be obtained by using weighted-sum
scalarizations of a MOCO problem 𝑃 , i.e.,
(𝑃 WS (ℓ)) min{ℓ𝑇 𝐶𝑥 | 𝑥 ∈ 𝒮}, for ℓ ≥ 0 .
One of the pioneering works in this respect was the discovery of the Dichotomic Approach (DA)
independently by Aneja and Nair (1979); Cohen (1978); Dial (1979). The algorithm finds the
extreme supported points of the Pareto-front of a biobjective combinatorial optimization problem,
i.e., the nondominated extreme points of conv 𝑐(𝒮). We will give different definitions for these terms
later in this section. Especially, Aneja and Nair showed that if we have access to an algorithm 𝐴
which solves a lexicographic version of the MOCO problem, we can find the set of nondominated
extreme points of size 𝑘 ≥ 2 by needing 2𝑘 −1 calls to 𝐴. The theory of output-sensitive complexity
is able to add more ground to the intuition that the set of nondominated extreme points can be
computed efficiently for every fixed 𝑑 ≥ 2.
A more general view on supported Pareto-optimal solutions can be obtained by extending the
view to multiobjective linear programming.
Definition 5.1 (Multiobjective Linear Programming (MOLP)). Given matrices 𝐴 ∈ Q𝑚×𝑛 , 𝐶 ∈
Q𝑑×𝑛 and a vector 𝑏 ∈ Q𝑚 , enumerate the extreme points 𝒴X of the polyhedron 𝒫 := {𝐶𝑥 | 𝐴𝑥 ≥
𝑏} + R𝑑≥ . We also write an MOLP in the form
min 𝐶𝑥
𝐴𝑥 ≥ 𝑏 .
10
We call the polyhedron 𝒫 the upper image of a given MOLP instance, cf. also (Hamel, Löhne,
and Rudloff, 2013)), and set 𝒴 := {𝐶𝑥 | 𝐴𝑥 ≥ 𝑏}. In the definition of the MOLP problem, we again
only allow rational numbers, as real numbers cannot be encoded in our model of computation.
We stress here that this is not the only possible definition of MOLP. One could also be interested
in finding the nondominated facets of the upper image and this would result in a different computational problem. But for the study of extreme points of MOCO problems, the above problem will
serve us better.
We observe that there is a difference between the above version of the MOLP problem and the
problem of enumerating all Pareto-optimal basic feasible solutions: The latter problem cannot be
solved in output-polynomial time unless P = NP, because it subsumes the vertex enumeration
problem even when allowing only two objectives. Khachiyan et al. (2008) proved that the vertex
enumeration problem for general polyhedra cannot be solved in output-polynomial time unless
P = NP. But we conjecture that there exists an output-sensitive algorithm for the MOLP problem
as in Definition 5.1.
The central connection between supported efficient solutions and supported nondominated points
and MOLP lies in the concept of convex relaxations of MOCO problems, cf. (Ehrgott and
Gandibleux, 2007; Cerqueus, Przybylski, and Gandibleux, 2015).
Definition 5.2 (Convex Relaxation). Given a MOCO problem 𝑃 = (𝒮, 𝐶), the convex relaxation
of 𝑃 is the MOLP
min 𝐶𝑥
𝑥 ∈ conv 𝒮 .
Definition 5.2 hides the fact, that a convex relaxation is not in the correct form of Definition 5.1.
Although the convex relaxation of a MOCO problem exists in general, it might be computationally
hard to obtain a representation of the MOLP as in Definition 5.1 (i.e., an inequality representation),
as it might have a large number of constraints or it might have large numbers in the matrix 𝐴. But
the convex relaxation will serve as a theoretical vehicle to define the extreme points of a MOCO
problem and show some connections between MOCO problems and MOLP.
Accordingly, we define the nondominated extreme points 𝒴X of a MOCO problem 𝑃 to be the
extreme points of the upper image of the associated convex relaxation. An efficient solution 𝑥 to a
MOCO problem 𝑃 is called supported, if 𝑥 is an efficient solution to the convex relaxation of 𝑃 .
In the following sections, we will first review some literature to shed some lights on solving
MOLP in the output-sensitivity framework in Section 5.1. Then in Section 5.2 we are concerned
with finding nondominated extreme points to MOCO problems and show how the results from
Section 5.1 can be applied to MOCO problems using the relation between a MOCO problem and
its convex relaxation.
5.1 Multiobjective Linear Programming
There are many ways of solving MOLP problems in the multiobjective optimization literature.
But performance guarantees are seldomly given. Recently, Bökler and Mutzel proved that MOLP,
under some restrictions, is solvable efficiently in the theory of output-sensitive complexity (Bökler
and Mutzel, 2015). We will first review this result and later show that in some special cases better
running times can be achieved.
5.1.1 The General Case with an Arbitrary Number of Objectives
In (Bökler and Mutzel, 2015), the authors prove the following theorem: If the ideal point of an
MOLP exists, we can enumerate the extreme points of the upper image in output-polynomial time.
11
Theorem 5.3 (Enumeration of Extreme Points of MOLP having an ideal point (Bökler and Mutzel,
2015)). If an ideal point of an MOLP (𝐶, 𝐴, 𝑏) exists, the extreme points 𝒴X of 𝒫 can be enumerated
𝑑
in time 𝒪(|𝒴X |⌊ 2 ⌋ (poly(⟨𝐴⟩, ⟨𝐶⟩) + |𝒴X | log |𝒴X |)) for each fixed number 𝑑 of objectives.
The ideal point of an MOLP is the point (min{𝐶𝑖 𝑥 | 𝑥 ∈ 𝒮})𝑖∈[𝑑] , where 𝐶𝑖 is the 𝑖-th row of 𝐶.
We remark that an ideal point does not exist for every MOLP, even though the MOLP is feasible,
e.g., if one of the linear programming problems min{𝐶𝑖 𝑥 | 𝑥 ∈ 𝒮} does not have a finite optimal
value.
To achieve this result, the authors conducted a running time analysis of a variant of Benson’s
algorithm. Benson’s algorithm was proposed by Benson (1998) and the mentioned variant, called
the dual Benson algorithm, was proposed by Ehrgott, Löhne, and Shao (2012).
From a highlevel point of view, the algorithm computes a representation of a polyhedron 𝒟, called
lower image, which is geometrically dual to the upper image 𝒫. The theory of geometric duality of
MOLP was introduced by Heyde and Löhne (2008), and in some sense discovered independently by
Przybylski, Gandibleux, and Ehrgott (2010). The representation consists of the extreme points and
facet inequalities which define 𝒟. Because of geometric duality of MOLP, the extreme points and
facet inequalities of 𝒟 correspond to the facet inequalities and extreme points of 𝒫, respectively,
and can be efficiently computed from them.
To compute the extreme points and facet inequalities of 𝒟, the algorithm first computes a
polyhedron 𝑃0 which contains 𝒟 entirely. Then, in each iteration 𝑖, it discovers one extreme point
𝑣 of the polyhedron 𝑃𝑖 and checks if 𝑣 lies on the boundary of 𝒟. If this is the case, the point is
saved as a candidate extreme point of 𝒟. If this is not the case, the algorithm finds an inequality
which separates 𝑣 from 𝒟 and supports 𝒟 in a face. Then, 𝑃𝑖+1 is computed by intersecting 𝑃𝑖
with the halfspace defined by the inequality found. In the end, redundant points are removed from
the set of candidates and the remaining set is output.
To check if 𝑣 lies on the boundary of 𝒟 and to compute the separating hyperplane, we only have
to solve a weighted-sum LP of the original MOLP P:
(𝑃 WS (ℓ)) min {ℓ𝑇 𝐶𝑥 | 𝐴𝑥 ≥ 𝑏},
where ℓ ∈ Q𝑑 and ||ℓ||1 = 1. Another important subproblem to solve is a variant of the well-known
vertex enumeration problem.
To arrive at the running time of Theorem 5.3, Bökler and Mutzel also provide an improvement
to the algorithm by guaranteeing to find facet supporting inequalities exclusively. This is possible
at the expense of solving a lexicographic linear programming problem
(𝑃 Lex-WS (ℓ)) lexmin {(ℓ𝑇 𝐶𝑥, 𝑐1 𝑥, . . . , 𝑐𝑑 𝑥) | 𝐴𝑥 ≥ 𝑏},
which turns out to be polynomial-time solvable in general.
5.1.2 The Special Case with Two Objectives
Now let 𝑃 be a biobjective linear program. In the following we present a polynomial delay algorithm,
which only needs polynomial space in the size of the input, for enumerating nondominated extreme
points. We assume that we have access to a solver for the lexicographic linear programming
problem. The idea of the algorithm is the following: We start by computing a nondominated
extreme point 𝑦 1 that minimizes the second objective, i.e., we solve 𝑃 Lex-WS ((0, 1)). Then, by
solving a lexicographic linear program, a facet 𝐹 of 𝒴 is computed that supports 𝑦 1 such that 𝑦 1 is
the lexicographic maximum of 𝐹 . By solving another lexicographic linear program, the algorithm
finds the other nondominated extreme point 𝑦 * that is also supported by 𝐹 , see Figure 2 for a
geometric illustration. We then restart the procedure with 𝑦 * . The algorithm terminates when the
nondominated extreme point is found that minimizes the first objective.
12
C2
Y
Y
y∗
F
y1
C1
Figure 2: Illustration of the idea of the algorithm with polynomial delay and polynomial space.
In order to compute a facet that supports a nondominated extreme point 𝑦 of 𝑃 , we need the
LP 𝐷2 (𝑦), cf. (Isermann, 1974):
maximize
𝑏𝑇 𝑢 − 𝑦 𝑇 𝜆
subject to (𝑢, 𝜆) ≥ 0𝑚+𝑝
𝐴𝑇 𝑢 = 𝐶 𝑇 𝜆
(𝐷2 (𝑦))
𝑝 𝑇
(1 ) 𝜆 = 1
(𝑢, 𝜆) ∈ Q𝑚+𝑝 .
We get the following result.
Lemma 5.4. Let 𝑃 be a biobjective linear program, let 𝑦 be a nondominated extreme point for 𝑃
¯ for 𝐷2 (𝑦) so
and let (𝑢, 𝜆) be a maximum
𝑢, 𝜆)
2 (𝑦) such that
{︁ for 𝐷
}︁ there is no other maximum (¯
⃒ 𝑇
2
𝑇
¯
⃒
that 𝜆 <lex 𝜆. Then 𝐹 := 𝑦 ∈ R 𝜆 𝑦 = 𝑏 𝑢 is a facet of 𝒴 that supports 𝒴 in 𝑦 such that 𝑦 is
the lexicographic maximum of 𝐹 .
Proof. The result easily follows from Proposition 4.2 in Ehrgott, Löhne, and Shao (2012).
Moreover, in order to compute the other extreme point of 𝐹 , we can simply add the constraint
𝜆𝑇 𝐶𝑥 = 𝑏𝑇 𝑢
to 𝑃 and solve the lexicographic linear programm for 𝑃 Lex-WS (𝜆), where (𝑢, 𝜆) is a maximum of
𝐷2 (𝑦) as stated in Lemma 5.4. We get the following result.
Theorem 5.5. Let 𝑃 be a biobjective linear program, then the above described algorithm outputs
the nondominated extreme points of 𝑃 with polynomial delay and polynomial space.
Proof. Let 𝑦 * in 𝒴X and let (𝑢, 𝜆) be a lexicographic maximum{︁ for the lexicographic
}︁ linear program
2
𝑇
𝑇
𝐷2 (𝑦) as stated in Lemma 5.4. Then, by Lemma 5.4, 𝐹 := 𝑦 ∈ R | 𝜆 𝑦 = 𝑏 𝑢 is a facet of 𝒴
with 𝑦 in 𝐹 such that 𝑦 is a lexicographic maximum of 𝐹 . At some point, the algorithm will find
the point that minimizes the first objective as it is the lexicographic minimum of a facet of 𝒴.
Hence, the algorithm is finite.
13
By Bökler and Mutzel (2015), each iteration can be computed in polynomial time. In each
iteration one nondominated extreme point is output, therefore the 𝑖-th delay of the algorithm is
polynomial in the input size. Moreover, observe that at most three nondominated extreme points
have to be kept in memory during each delay. Hence, the space is bounded polynomially in the
input size.
5.2 Enumerating Supported Nondominated Points
Now, we will come back to MOCO problems and show how we can find supported nondominated
and extreme nondominated points. In Section 5.2.1 we show how we can make use of the results
from Section 5.1.1, even when we do not have a compact LP formulation of our MOCO problem. We
show that in the case of two objectives, we can again achieve better running times in Section 5.2.2.
In the end, we discuss a result by Okamoto and Uno (2011), where the authors are concerned with
finding supported spanning trees.
5.2.1 Enumerating Extreme Points of MOCO with an Arbitrary Number of
Objectives
To find nondominated extreme points we can apply the algorithms from Section 5.1.1 to the convex
relaxation of the MOCO we want to solve. On one hand, convex relaxations of MOCO problems
have an ideal point if and only if the MOCO instance is feasible.
On the other hand, because of the aforementioned problems, we cannot easily compute the convex
relaxation of a MOCO problem to apply the Dual Benson algorithm. But the only interaction
between the algorithm and the MOLP at hand lies in solving weighted-sum linear programs. Hence,
instead of solving the weighted-sum objective over the feasible set of the convex relaxation, we can
optimize the weighted-sum objective over the feasible set of the MOCO problem.
These considerations lead us to the following theorem.
Theorem 5.6 (Enumeration of Extreme Points of MOCO Problems I (Bökler and Mutzel, 2015)).
For every MOCO problem 𝑃 with a fixed number of objectives, the set of extreme nondominated
points of 𝑃 can be enumerated in output-polynomial time if we can solve the weighted-sum scalarization of 𝑃 in polynomial time.
Moreover, if we are able to solve lexicographic objective functions over the feasible set of the
MOCO problem, we can even bound the delay of the algorithm.
Theorem 5.7 (Enumeration of Extreme Points of MOCO Problems II (Bökler and Mutzel, 2015)).
For every MOCO problem 𝑃 with a fixed number of objectives, the set of extreme nondominated
points of 𝑃 can be enumerated in incremental polynomial time if we can solve the lexicographic
version of 𝑃 in polynomial time.
One example of a problem where it is hard to compute the whole Pareto-front, but enumerating
the nondominated extreme points can be done in incremental polynomial time is thus the MOSP
problem although both the size of the Pareto-front and the number of nondominated extreme points
can be superpolynomial in the input size, cf. (Gusfield, 1980; Carstensen, 1983).
There are two further methods which are concerned with finding extreme nondominated points of
the Pareto-front of general multiobjective integer problems: The methods by Przybylski, Gandibleux,
and Ehrgott, 2010 and Özpeynirci and Köksalan, 2010. Both do not give any running time guarantees, but the investigation of these can be subject to further research.
14
y0
y0
Y
Y
y3
Y
Y
y3
ȳ
ȳ
y2
y2
y1
y1
(a) Case 1
(b) Case 2
Figure 3: Illustration of the idea of the DA
5.2.2 Enumerating Extreme Points of Bicriterial Combinatorial Optimization
Problems
In the following, we investigate different variants of the DA, i.e., with and without access to an
algorithm to solve the lexicographic version of a biobjective combinatorial optimization (BOCO)
problem 𝑃 . Especially, we show that the nondominated extreme points of 𝑃 can be enumerated
with polynomial delay if we have access to an algorithm to solve the lexicographic version of 𝑃 .
5.2.3 The Dichotomic Approach
Let 𝑃 be a BOCO problem. We begin by describing the DA without access to an algorithm to
solve the lexicographic version of 𝑃 . Notice that Aneja and Nair (1979) assume access to such an
algorithm.
In a first step, the DA computes optimal solutions 𝑦 0 and 𝑦 1 to the weighted-sum scalarizations
WS
𝑃 ((1, 0)) and 𝑃 WS ((0, 1)), respectively. If 𝑦 0 equals 𝑦 1 , the algorithm terminates since the ideal
point is a member of the Pareto-front. Otherwise, the pair (𝑦 0 , 𝑦 1 ) is added to the queue 𝒬.
In each iteration, the algorithm retrieves a pair (𝑦 2 , 𝑦 3 ) from 𝒬 and 𝜆 is computed such that
𝜆𝑇 𝑦 2 = 𝜆𝑇 𝑦 3 .
(1)
Hence, the algorithm computes 𝜆 = (|𝑦22 − 𝑦23 |, |𝑦12 − 𝑦13 |). The algorithm proceeds by computing
an optimal solution 𝑦¯ to the weighted-sum scalarization 𝑃 WS (𝜆). If 𝑦¯ is located in the line segment
spanned by 𝑦 2 and 𝑦 3 , i.e., 𝜆𝑇 𝑦 2 = 𝜆𝑇 𝑦¯ = 𝜆𝑇 𝑦 3 , the algorithm neglects 𝑦¯ as it is either not a
nondominated extreme point or was already discovered by the algorithm in a previous iteration,
see Figure 3(a). Otherwise, 𝑦¯ is added to the set 𝒜 and the pairs (¯
𝑦 , 𝑦 2 ) and (¯
𝑦 , 𝑦 3 ) are added
to 𝒬, see Figure 3(b). The algorithm terminates when 𝒬 is empty. Moreover, the nondominated
non-extreme points in 𝒜 are removed and the remaining points are output. We get the following
result.
Theorem 5.8. Let 𝑃 be a BOCO problem, then the DA without access to an algorithm to solve
the lexicographic version of 𝑃 outputs the nondominated extreme points of 𝑃 . Moreover, if the
weighted-sum scalarization of 𝑃 is solvable in polynomial time, the algorithm is output-sensitive.
15
Proof. First, observe that there is no nondominated extreme point 𝑦 such that 𝑦1 < 𝑦10 and 𝑦2 < 𝑦21 ,
where 𝑦 0 and 𝑦 1 are the points found in the first step of the algorithm. Secondly, observe that there
is at most one nondominated non-extreme point found by the algorithm per facet of 𝒴. Hence,
the algorithm is finite. Moreover, by Isermann (1974) all points computed by the algorithm except
possibly 𝑦 0 and 𝑦 1 are nondominated.
Now assume there is a nondominated extreme point 𝑦¯ that is not found by the algorithm. By
the first observation, 𝑦 0 <lex 𝑦¯ <lex 𝑦 1 . Hence, there must be a pair of points (𝑦 2 , 𝑦 3 ) such that
𝑦 2 <lex 𝑦¯ <lex 𝑦 3 and there is no other pair (𝑦 4 , 𝑦 5 ) that is “closer” to 𝑦¯, i.e., 𝑦 2 <lex 𝑦 4 <lex 𝑦¯ <lex
𝑦 5 <lex 𝑦 3 . Since the algorithm investigates each such pair, the algorithm will eventually discover
𝑦¯ and we arrive at a contradiction. Hence, the algorithm finds all nondominated extreme points of
𝑃.
Because of the second observation and the fact that the number of nondominated facets of 𝒴 is
linearly bounded in the cardinality of 𝒴X , the number of weighted-sum scalarizations the algorithm
has to solve is linearly bounded in the cardinality of 𝒴X . Moreover, the encoding lengths of the
points discovered during the computation of the algorithm are polynomially bounded in the size of
the input, cf. (Bökler and Mutzel, 2015). Hence, the algorithm is output-sensitive if the weightedsum scalarization of 𝑃 is solvable in polynomial time.
The above algorithm does not lead directly to an algorithm with polynomial delay since there
may exist several solution to the weighted-sum scalarization 𝑃 WS (𝜆), which are not necessarily
extreme. However, if we assume access to an algorithm to solve the lexicographic version of 𝑃 , we
get the following result.
Corollary 5.9. Let 𝑃 be a BOCO problem, then the DA with access to an algorithm to solve
the lexicographic version of 𝑃 enumerates the nondominated extreme points of 𝑃 . Moreover, if the
lexicographic version of 𝑃 is solvable in polynomial time, the points are enumerated with incremental
delay.
Proof. The correctness follows from the proof of Theorem 5.8. Since we have access to an algorithm
to solve the lexicographic version of 𝑃 , each point found by the algorithm is nondominated extreme.
Hence, the running time to discover a new nondominated extreme point or to conclude that no
such point exists is linearly bounded in the number of the nondominated extreme points found so
far. Therefore, the algorithm enumerates the nondominated extreme points with incremental delay
if the lexicographic version of 𝑃 is solvable in polynomial time.
Note that a lexicographic version of many BOCO problems can be solved as fast as the single
objective problem. For example the biobjective shortest path, matroid optimization and assignment
problems. Also if a compact formulation is available, the lexicographic weighted sum problem can
be solved as a lexicographic linear programming problem, which can also be solved as fast as a
single objective linear programming problem (cf. Bökler and Mutzel, 2015).
A Variant of the Dichotomic Approach with Polynomial Delay We present a simple
modification of the DA with access to an algorithm to solve the lexicographic version of 𝑃 . The
idea is the following: When a new nondominated extreme point is found, the DA puts two pairs
of nondominated extreme points into the queue 𝒬. At some point, the orginal algorithm checks if
there is any nondominated extreme point 𝑦¯ “between” such a pair (𝑦 2 , 𝑦 3 ) of points, i.e., 𝑦 2 <lex
𝑦¯ <lex 𝑦 3 or 𝑦 3 <lex 𝑦¯ <lex 𝑦 2 . Now instead the modified algorithm computes 𝜆 as in (1) and solves
𝑃 Lex-WS (𝜆) resulting in a nondominated extreme point 𝑦. If 𝑦 equals 𝑦 2 or 𝑦 3 , we have verified
that there is no point between the two points. Otherwise we add the triple (𝑦, 𝑦 2 , 𝑦 3 ) to the queue
̃︀ If such a triple is retrieved from 𝒬,
̃︀ we can output the nondominated extreme point 𝑦. The rest
𝒬.
of the algorithm is analogous to original Dichotomic Approach. Therefore, in each iteration exactly
16
one nondominated extreme point is output. Moreover, in each iteration at most two lexicographic
weighted sum-scalarization problems have to be solved. Hence, we can easily derive the following
result.
Corollary 5.10. Let 𝑃 be a BOCO problem, then the above described variant of the DA outputs
the nondominated extreme points of 𝑃 with polynomial delay if the lexicographic version of 𝑃 is
solvable in polynomial time.
We remark here that we obtain a polynomial delay algorithm by keeping back points which we
would have presented to the user already in the original DA. This is not what we usually want to do
in practice. To overcome this issue, one could investigate the amortized delay, i.e., the amortized
time needed per output. Moreover, we get the following result which follows from Theorem 5.5.
Corollary 5.11. Let 𝑃 be a BOCO problem. If there is a compact formulation of 𝑃 , then the
above described algorithm outputs the nondominated extreme points of 𝑃 with polynomial delay and
polynomial space.
5.2.4 Enumerating Supported Efficient Spanning Trees
The first paper discussing techniques from enumeration algorithmics in the context of multiobjective
optimization is by Okamoto and Uno (2011). They propose a solution to the multiobjective spanning
tree problem, which is defined as follows.
Definition 5.12 (Multiobjective Spanning Tree Problem (MOST)). Given an undirected Graph
𝐺 = (𝑉, 𝐸), where 𝐸 ⊆ {𝑒 ⊂ 𝑉 | |𝑒| = 2}, and a multiobjective edge-cost function 𝑐 : 𝐸 → Q𝑑≥ .
A feasible solution is a spanning tree 𝑇 of 𝐺, i.e., an acyclic, connected subgraph of 𝐺 containing
∑︀
all nodes 𝑉 . We set 𝑐(𝑇 ) := 𝑒∈𝑇 𝑐(𝑒). The set of all such trees is denoted by 𝒯 . The goal is to
enumerate the minima of 𝑐(𝒯 ) with respect to the canonical partial order on vectors.
However, the model considered by Okamoto and Uno (2011) is different from the one considered
in this paper. They consider enumerating all supported weakly efficient spanning trees, i.e., all
𝑇 ∈ 𝒯 , such that 𝑇 is an optimal solution to MOSTWS (ℓ) for some ℓ ≥ 0. This includes nonPareto-optimal spanning trees for ℓ ∈ {𝑒𝑖 | 𝑖 ∈ [𝑑]} and also equivalent trees 𝑇 and 𝑇 ′ with 𝑇 ̸= 𝑇 ′
and 𝑐(𝑇 ) = 𝑐(𝑇 ′ ).
For this problem, the authors propose an algorithm which runs with polynomial delay by employing a variant of the reverse search method by Avis and Fukuda (1992). The authors also generalize
this algorithm to the case of enumerating Pareto-optimal basic feasible solutions of an MOLP if it
is nondegenerate, i.e., every basic solution has exactly 𝑛 active inequalities.
6 Conclusion
In this paper, we showed that output-sensitive complexity is a useful tool to investigate the complexity of multiobjective combinatorial optimization problems. We showed that in contrast to
traditional computational complexity in multiobjective optimization, output-sensitive complexity
is able to separate efficiently solvable from presumably not efficiently solvable problems.
On one hand, we provided examples of problems for which output-sensitive algorithms exist, e.g.,
the multiobjective global minimum-cut problem, and also computing the nondominated extreme
points for many MOCO problems can be done efficiently.
On the other hand, we demonstrated that it is also possible to show that a MOCO problem does
not admit an output-sensitive algorithm under weak complexity theoretic assumptions as P ̸= NP.
One example is the multiobjective shortest path problem, which in turn also rules out the existence
17
of output-sensitive algorithms for the multiobjective versions of the minimum perfect matching and
the minimum-cost flow problem.
To conclude, we will raise some questions which can be investigated in further work.
Multiobjective Linear Programming While in Section 5.2.1 it was shown that we can enumerate the extreme points of the upper image of an MOLP in incremental polynomial time, we
needed the assumption that an ideal point exists. We are looking forward to being able to lift this
assumption to the general case of MOLP and obtain at least an output-sensitive algorithm.
Another question is, if it is necessary to fix the number of objectives 𝑑 in the running time
to obtain an output-sensitive algorithm. We conjecture that this is actually the case when using
standard assumptions from complexity theory and will investigate this in the future.
A longer reaching question is, if we can find an algorithm with polynomial delay for the case of
general MOLP and a fixed number of objectives.
Multiobjective Combinatorial Optimization Not much is known in this regard. A very
intriguing problem is the following:
Definition 6.1 (Unconstrained Biobjective Combinatorial Optimization Problem). Given two vectors 𝑐1 , 𝑐2 ∈ Q𝑛 . The set of feasible solutions is {0, 1}𝑛 and the objective function is
(︃ 𝑇 )︃
1
𝑐 : 𝑥 ↦→
𝑐 𝑥
.
𝑇
𝑐2 𝑥
The goal is to enumerate the minima of 𝑐({0, 1}𝑛 ) with respect to the canonical partial order on
vectors.
Even for this problem, which is an unconstrained version of the biobjective knapsack problem, it
is unkown if an output-sensitive algorithm exists. Moreover, if there is no output-sensitive algorithm
for this problem we can rule out output-sensitive algorithms for many MOCO problems, e.g., the
multiobjective assignment and the multiobjective spanning tree problem.
Additional Complexity Classes There has been some research on new complexity classes,
combining fixed parameter tractability theory and output-sensitive complexity theory, which is
Parameterized Enumeration by Creignou et al. (2013). It is very interesting to see if parameters
can be found which make MOCO problems hard, apart from the fact that multiple objectives are
present.
References
[1]
H. Ackermann, A. Newman, H. Röglin, and B. Vöcking. “Decision-making based on approximate and smoothed Pareto curves”. In: Theoretical Computer Science 378.3 (2007), pp. 253–
270.
[2]
Y. P. Aneja and K. P. K. Nair. “Bicriteria transportation problem”. In: Management Science
25.1 (1979), pp. 73–78.
[3]
A. Armon and U. Zwick. “Multicriteria Global Minimum Cuts”. In: Algorithmica 46.1 (2006),
pp. 15–26.
[4]
S. Arora and B. Barak. Computational Complexity: A Modern Approach. 1st. New York:
Cambridge University Press, 2009.
18
[5]
D. Avis and K. Fukuda. “A pivoting algorithm for convex hulls and vertex enumeration
of arrangements of polyhedra”. In: Discrete and Computational Geometry 8.1 (Dec. 1992),
pp. 295–313.
[6]
H. P. Benson. “An outer approximation algorithm for generating all efficient extreme points
in the outcome set of a multiple objective linear programming problem”. In: Journal of Global
Optimization 13.1 (1998), pp. 1–24.
[7]
F. Bökler and P. Mutzel. “Output-Sensitive Algorithms for Enumerating the Extreme Nondominated Points of Multiobjective Combinatorial Optimization Problems”. In: Algorithms
– ESA 2015. LNCS 9294. Berlin: Springer, 2015, pp. 288–299.
[8]
T. Brunsch and H. Röglin. “Improved Smoothed Analysis of Multiobjective Optimization”.
In: STOC 2012. New York: ACM, 2012, pp. 407–426.
[9]
P.M. Camerini, G. Galbiati, and F. Maffioli. “The complexity of multi-constrained spanning
tree problems”. In: Theory of Algorithms. Ed. by L. Lovasz. Amsterdam: North-Holland, 1984,
pp. 53 –101.
[10]
P. J. Carstensen. “The Complexity of Some Problems in Parametric Linear and Combinatorial
Programming”. PhD thesis. University of Michigan, 1983.
[11]
A. Cerqueus, A. Przybylski, and X. Gandibleux. “Surrogate upper bound sets for bi-objective
bi-dimensional binary knapsack problems”. In: European Journal of Operational Research
244.2 (2015), pp. 417–433.
[12]
V. Chankong and Y.Y. Haimes. Multiobjective Decision Making Theory and Methodology.
Amsterdam: Elsevier, 1983.
[13]
J. L. Cohen. Multiobjective Programming and Planning. Vol. 140. Mathematics in Science
and Engineering. New York: Academic Press, 1978.
[14]
T. H. Cormen, C. Stein, R. L. Rivest, and C. E. Leiserson. Introduction to Algorithms. 2nd ed.
Cambridge, MA: MIT Press, 2001.
[15]
N. Creignou, A. Meier, J.-S. Müller, J. Schmidt, and H. Vollmer. “Paradigms for Parameterized Enumeration”. In: Mathematical Foundations in Computer Science. LNCS 8087. Berlin:
Springer, 2013, pp. 290–301.
[16]
R. B. Dial. “A Model and Algorithm for Multicriteria Route-Mode Choice”. In: Transportation
Research 13B (1979), pp. 311–316.
[17]
M. Ehrgott. “A Discussion of Scalarization Techniques for Multiple Objective Integer Programming”. In: Annals of Operations Research 147.1 (2006), pp. 343–360.
[18]
M. Ehrgott. Multicriteria Optimization. 2nd. Vol. 491. Lecture Notes in Economics and Mathematical Systems. Berlin: Springer, 2005.
[19]
M. Ehrgott. “On matroids with multiple objectives”. In: Optimization 38.1 (1996), pp. 73–84.
[20]
M. Ehrgott and X. Gandibleux. “Bound Sets for Biobjective Combinatorial Optimization
Problems”. In: Computers & Operations Research 34.9 (2007), pp. 2674–2694.
[21]
M. Ehrgott, A. Löhne, and L. Shao. “A dual variant of Benson’s “outer approximation algorithm” for multiple objective linear programming”. In: Journal of Global Optimization 52.4
(2012), pp. 757–778.
[22]
J. R. Figueira, C. Fonseca, P. Halffmann, K. Klamroth, L. Paquete, S. Ruzika, B. Schulze,
M. Stigelmayr, and D. Willems. “Easy to say they’re hard, but hard to say they are easy –
Towards a categorization of tractable multiobjective combinatorial optimization problems”.
In: JMCDA (to appear) (2016).
19
[23]
M. R. Garey and D. S. Johnson. Computers and Intractability; A Guide to the Theory of
NP-Completeness. 1st ed. Series of Books in the Mathematical Sciences. New York: W. H.
Freeman & Co., 1979.
[24]
M. Grötschel, L. Lovász, and A. Schrijver. Geometric Algorithms and Combinatorial Optimization. 1st. Vol. 2. Algorithms and Combinatorics. Berlin: Springer, 1988.
[25]
D. Gusfield. “Sensitivity Analysis for Combinatorial Optimization”. PhD thesis. UC Berkeley,
1980.
[26]
H.W. Hamacher and G. Ruhe. “On spanning tree problems with multiple objectives”. In:
Annals of Operations Research 52 (1994), pp. 209–230.
[27]
A. H. Hamel, A. Löhne, and B. Rudloff. “Benson type algorithms for linear vector optimization
and applications”. In: Journal of Global Optimization (2013).
[28]
P. Hansen. “Bicriterion Path Problems”. In: Multiple Criteria Decision Making Theory and
Application. Ed. by G. Fandel and T. Gal. Vol. 177. Lecture Notes in Economics and Mathematical Systems. Berlin: Springer, 1979, pp. 109–127.
[29]
F. Heyde and A. Löhne. “Geometric duality in multiple objective linear programming”. In:
SIAM Journal of Optimization 19.2 (2008), pp. 836–845.
[30]
H. Isermann. “Technical note—Proper efficiency and the linear vector maximum problem.”
In: Operations Research 22.1 (1974), pp. 189–191.
[31]
D. S. Johnson, M. Yannakakis, and C. H. Papadimitriou. “On generating all maximal independent sets”. In: Information Processing Letters 27.3 (1988), pp. 119–123.
[32]
H. Kellerer, U. Pferschy, and D. Pisinger. Knapsack Problems. Berlin: Springer, 2004.
[33]
L. Khachiyan, E. Boros, K. Borys, K. Elbassioni, and V. Gurvich. “Generating All Vertices
of a Polyhedron Is Hard”. In: Discrete and Computational Geometry 39 (2008), pp. 174–190.
[34]
K. Klamroth, R. Lacour, and D. Vanderpooten. “On the representation of the search region
in multi-objective optimization”. In: European Journal of Operational Research 245.3 (2015).
[35]
M. Laumanns, L. Thiele, and E. Zitzler. “An efficient, adaptive parameter variation scheme for
metaheuristics based on the epsilon-constraint method”. In: European Journal of Operational
Research 169 (2006), pp. 932–942.
[36]
E. L. Lawler, J. K. Lenstra, and A. H. G. Rinnooy Kan. “Generating All Maximal Independent
Sets: NP-Hardness and Polynomial-Time Algorithms”. In: SIAM Journal on Computing 9.3
(Aug. 1980), pp. 558–565.
[37]
E. Q. V. Martins. “On a Multicriteria Shortest Path Problem”. In: European Journal of
Operational Research 16 (1984), pp. 236–245.
[38]
Y. Okamoto and T. Uno. “A polynomial-time-delay and polynomial-space algorithm for enumeration problems in multi-criteria optimization”. In: European Journal of Operational Research 210.1 (2011), pp. 48–56.
[39]
Ö. Özpeynirci and M. Köksalan. “An exact algorithm for finding extreme supported nondominated points of multiobjective mixed integer programs”. In: Management Science 56.12
(2010), pp. 2302–2315.
[40]
A. Przybylski, X. Gandibleux, and M. Ehrgott. “A recursive algorithm for finding all nondominated extreme points in the outcome set of a multiobjective integer programme”. In:
INFORMS Journal on Computing 22.3 (2010), pp. 371–386.
[41]
G. Ruhe. “Complexity results for multicriteria and parametric network flows using a pathological graph of Zadeh”. In: Zeitschrift für Operations Research 32 (1988), pp. 9–27.
20
[42]
J. Schmidt. Enumeration: Algorithms and Complexity. Diplomarbeit. Mar. 2009.
[43]
P. Serafini. “Some considerations about computational complexity for multi objective combinatorial problems”. In: Recent Advances and Historical Development of Vector Optimization.
Vol. 294. Lecture Notes in Economics and Mathematical Systems. Berlin: Springer, 1986,
pp. 222–232.
21
| 8cs.DS
|
arXiv:1612.00198v2 [math.GR] 7 Jul 2017
Pro-p subgroups of profinite completions of
3-manifold groups
Henry Wilton∗ and Pavel Zalesskii†
July 10, 2017
Abstract
We completely describe the finitely generated pro-p subgroups of
the profinite completion of the fundamental group of an arbitrary 3manifold. We also prove a pro-p analogue of the main theorem of
Bass–Serre theory for finitely generated pro-p groups.
1
Introduction
In recent years there has been a great deal of interest in detecting properties
of the fundamental group π1 M of a 3-manifold via its finite quotients, or
more conceptually by its profinite completion. This motivates the study of
the profinite completion π[
1 M of the fundamental group of a 3-manifold. As
in the case of finite groups, there is a Sylow theory for profinite groups, and
pro-p subgroups play the same central role in profinite group theory that
p-subgroups play in the theory of finite groups. Note that pro-p subgroups
of π[
1 M are infinitely generated in general, so it is natural to begin the study
from the finitely generated case. Amazingly enough, it is possible to give a
complete description of the finitely generated pro-p subgroups of the profinite
completions of all 3-manifold groups, and this is one of the objectives of this
paper (see Theorem 1.3 below).
∗
Department of Pure Mathematics and Mathematical Statistics, Centre for Mathematical Sciences, Wilberforce Road, Cambridge, CB3 0WB, United Kingdom
†
Department of Mathematics, University of Brası́lia, 70910-9000 Brası́lia, Brazil
1
The work of this paper was partially motivated by [WZ17, Theorem F],
where it was proved that if M is hyperbolic then any finitely generated prop subgroup of π[
1 M is free. In fact the proof of Theorem 1.3 follows the
same strategy as in [WZ17]. By applying first the (profinite) Kneser–Milnor
decomposition and then the (profinite) JSJ decomposition, the theorem is
reduced to the cases of Seifert-fibred and hyperbolic manifolds, possibly with
cusps. The main difficulty at this point is provided by hyperbolic manifolds
with cusps. In the closed case the result was proved in [WZ17]. In this
case (after passing to a finite-sheeted cover) there is a suitable hierarchy in
which the corresponding actions on profinite trees are relatively profinitely
acylindrical. However, we have to deal with cusps.
At this point we make heavy use of the profinite analogue of Bass–Serre
theory for groups acting on trees. Note that this theory does not have the
full strength of its classical original. The main theorem of Bass–Serre theory
asserts that a group G acting on a tree T is the fundamental group of a graph
of groups (G, T /G, D), where D is a maximal subtree of T /G and G consists
of edge and vertex stabilizers of a connected transversal of T /G. This does
not hold in the profinite case: firstly D does not always exist and secondly
even if it exists it might not lift to T .
In this paper we prove a pro-p analogue of the main theorem of Bass–
Serre theory for finitely generated pro-p groups, that stands independently
as a contribution to the combinatorial theory of pro-p groups.
Theorem 1.1. Let G be a finitely generated pro-p group acting on a pro-p
tree T . Then G is the fundamental pro-p group of a profinite graph of pro-p
groups (G, Γ). Moreover, the vertex and edge groups of (G, Γ) are stabilizers
of certain vertices and edges of T respectively, and stabilizers of vertices and
edges of T in G are conjugate to subgroups of vertex and edge groups of (G, Γ)
respectively.
We observe that (G, Γ) in this theorem is not (G, T /G) as in classical
Bass–Serre theory. To construct the graph of pro-p groups (G, Γ) we use
a decomposition of G as an inverse limit G = limU GU of finitely generated
←−
virtually free pro-p groups GU , and their splittings as the fundamental groups
of finite graphs of finite p-groups obtained in [HZ13].
We then apply this theorem to describe the finitely generated pro-p subgroups of the profinite completion of the fundamental group of a cusped
hyperbolic 3-manifold.
2
Theorem 1.2. Let π1 M be the fundamental group of an orientable hyperbolic
3-manifold with cusps and H a finitely generated pro-p subgroup of π
b1 M.
Then H is a free pro-p product of free abelian pro-p groups of rank ≤ 2.
[
Analysing π
1 M for Thurston’s geometries case-by-case, we obtain the
[
classification of finitely generated pro-p subgroups of π
1 M for an arbitrary
compact, orientable 3-manifold.
Theorem 1.3. A finitely generated pro-p subgroup of the profinite completion π[
1 M of the fundamental group π1 M of a compact, orientable 3-manifold
M is a free pro-p product of the pro-p groups from the following lists of isomorphism types.
(i) For p > 3: Cp ; Zp ; Zp ×Zp ; the pro-p completion of (Z×Z)⋊Z; and the
pro-p completion of a residually-p fundamental group of a non-compact
Seifert fibred manifold with hyperbolic base orbifold;
(ii) For p = 3, in addition to the list of (i) of possible free factors we have
a torsion-free extension of Z3 × Z3 × Z3 by C3 .
(iii) For p = 2, in addition to the list in (i) we have the following free
factors:
(1) the cyclic groups C2m ;
(2) the dihedral groups D2k ;
(3) the generalized quaternion groups Q2n ;
(4) the infinite dihedral pro-2 group Z2 ⋊ C2 ;
(5) the torsion-free extensions of Z2 × Z2 × Z2 by one of the following
finite 2-groups: C2 , C4 , C8 , D2 , D4 , D8 , Q16 ;
(6) the pro-2 extension of the Klein-bottle group Z ⋊ Z;
(7) the pro-2 completion of all torsion free extensions of a soluble
group (Z ⋊ Z) ⋊ Z with a group of order at most 2.
We note that, because the finitely generated pro-p subgroups of profinite
completions of 3-manifold groups are so restricted, they are not very useful
for distinguishing 3-manifolds from each other. However, for the same reason,
they are potentially useful for distinguishing the profinite completions of 3manifold groups from the profinite completions of other discrete groups.
3
Acknowledgements
The first author was supported by the EPSRC and the second author by
Capes. Much of the work for this paper was completed while the second
author was a visiting fellow at Trinity College, Cambridge, and he acknowledges their hospitality. Thanks are also due to the anonymous referee for
pointing out a mistake in the first version of this paper.
2
Preliminaries
In this section, we recall the necessary elements of the theory of profinite
trees.
A graph Γ is a disjoint union E(Γ)∪V (Γ) with two maps d0 , d1 : Γ → V (Γ)
that are the identity on the set of vertices V (Γ). For an element e of the set
of edges E(Γ), d0 (e) is called the initial and d1 (e) the terminal vertex of e.
Definition 2.1. A profinite graph Γ is a graph such that:
1. Γ is a profinite space (i.e. an inverse limit of finite discrete spaces);
2. V (Γ) is closed; and
3. the maps d0 and d1 are continuous.
Note that E(Γ) is not necessary closed.
A morphism α : Γ −→ ∆ of profinite graphs is a continuous map with
αdi = di α for i = 0, 1.
By [ZM88, Prop. 1.7] every profinite graph Γ is an inverse limit of finite
quotient graphs of Γ.
For a profinite space X that is the inverse limit of finite discrete spaces
b
b j ], where [ZX
b j ] is the free Z-module
b
Xj , [[ZX]]
is the inverse limit of [ZX
with basis Xj . For a pointed profinite space (X, ∗) that is the inverse limit
b
of pointed finite discrete spaces (Xj , ∗), [[Z(X,
∗)]] is the inverse limit of
b
b
b
[Z(Xj , ∗)], where [Z(Xj , ∗)] is the free Z-module with basis Xj \ {∗} [RZ10,
Chapter 5.2].
For a profinite graph Γ, define the pointed space (E ∗ (Γ), ∗) as Γ/V (Γ),
with the image of V (Γ) as a distinguished point ∗, and denote the image of
4
e ∈ E(Γ) by ē. By definition a profinite tree Γ is a profinite graph with a
short exact sequence
δ
ǫ b
b ∗ (Γ), ∗)]] →
b (Γ)]] →
0 → [[Z(E
[[ZV
Z→0
where δ(ē) = d1 (e)−d0 (e) for every e ∈ E(Γ) and ǫ(v) = 1 for every v ∈ V (Γ).
If v and w are elements of a profinite tree T , we denote by [v, w] the smallest
profinite subtree of T containing v and w and call it a geodesic.
A profinite graph is called a pro-p tree if one has the following exact
sequence:
δ
ǫ
0 → [[Fp (E ∗ (Γ), ∗)]] → [[Fp V (Γ)]] → Fp → 0
where δ(ē) = d1 (e)−d0 (e) for every e ∈ E(Γ) and ǫ(v) = 1 for every v ∈ V (Γ).
Note that any profinite tree is a pro-p tree, but the converse is not true. In
fact, a connected profinite graph is a profinite tree if and only if it is a pro-p
tree for every prime p.
By definition, a profinite group G acts on a profinite graph Γ if we have
a continuous action of G on the profinite space Γ that commutes with the
maps d0 and d1 .
When we say that G is a finite graph of profinite groups, we mean that it
contains the data of the underlying finite graph, the edge profinite groups,
the vertex profinite groups and the attaching continuous maps. More precisely, let ∆ be a connected finite graph. The data of a graph of profinite
groups (G, ∆) over ∆ consists of a profinite group G(m) for each m ∈ ∆, and
continuous monomorphisms ∂i : G(e) −→ G(di (e)) for each edge e ∈ E(∆).
A finite graph of groups (G, Γ) is said to be reduced, if for every edge e
which is not a loop neither ∂1 : G(e) → G(d1 (e)) nor ∂0 : G(e) → G(d0 (e)) is
an isomorphism; we call an edge e which is not a loop and such that one of
the edge maps ∂i is an isomorphism fictitious. Any finite graph of groups can
be transformed into a reduced finite graph of groups by collapsing fictitious
edges using the following procedure. If e is a fictitious edge, we can remove
{e} from the edge set of Γ, and identify d0 (e) and d1 (e) to a new vertex y.
Let Γ′ be the finite graph given by V (Γ′ ) = {y} ⊔ V (Γ) \ {d0 (e), d1 (e)} and
E(Γ′ ) = E(Γ)\{e}, and let (G′ , Γ′ ) denote the finite graph of groups based on
Γ′ given by G′ (y) = G(d1 (e)) if ∂0 (e) is an isomorphism, and G′ (y) = G(d0 (e))
if ∂0 is not an isomorphism. This procedure can be continued until there are
no fictitious edges. The resulting finite graph of groups (G, Γ) is reduced.
5
Remark 2.2. The reduction procedure just described does not change the
fundamental group (as a group given by presentation), i.e. choosing a maximal subtree to contain the collapsing edge the morphism (G, Γ) −→ (G′ , Γ′ )
induces the identity map on the fundamental group given by presentation by
eliminating redundant relations associated with fictitious edges that are just
collapsed by reduction.
The definition of the profinite fundamental group of a connected profinite graph of profinite groups is quite involved (see [ZM89]). However, the
profinite fundamental group Π1 (G, Γ) of a finite graph of finitely generated
profinite groups (G, Γ) can be defined as the profinite completion of the abstract (usual) fundamental group Πabs
1 (G, Γ) (we use here that every subgroup
of finite index in a finitely generated profinite group is open [NS07, Theorem
1.1 ]). The fundamental profinite group Π1 (G, Γ) has the following presentation:
Π1 (G, Γ) = hG(v), te | rel(G(v)), ∂1 (g) = ∂0 (g)te , g ∈ G(e),
te = 1 for e ∈ T i;
(1)
here T is a maximal subtree of Γ and ∂0 : G(e) −→ G(d0 (e)), ∂1 : G(e) −→
G(d1 (e)) are monomorphisms.
In contrast to the abstract case, the vertex groups of (G, Γ) do not always
embed in Π1 (G, Γ), i.e., Π1 (G, Γ) is not always proper. However, the edge
and vertex groups can be replaced by their images in Π1 (G, Γ) and after this
replacement Π1 (G, Γ) becomes proper. Thus from now on we shall assume
that Π1 (G, Γ) is always proper, unless otherwise stated.
To obtain the definition of the fundamental pro-p group of a finite graph
of finitely generated pro-p groups one simply replaces ‘profinite’ by ‘pro-p’ in
the above definition.
The profinite (resp. pro-p) fundamental group Π1 (G, Γ) acts on the standard profinite (resp. pro-p) tree S (defined analogously to the abstract one)
associated to it, with vertex and edge stabilizers being conjugates of vertex
and edge groups, and such that S/Π1 (G, Γ) = Γ [ZM88, Proposition 3.8].
Example 2.3. If G = π1 (G, Γ) is the fundamental group of a finite graph of
(abstract) groups then one has the induced graph of profinite completions
b Γ) (not necessarily proper) and a natural hoof edge and vertex groups (G,
b Γ). It is an embedding if π1 (G, Γ) is
momorphism G = π1 (G, Γ) −→ Π1 (G,
b is simply the profinite completion.
residually finite. In this case Π1 (G, Γ) = G
6
b if and only if the
Moreover, the standard tree S(G) naturally embeds in S(G)
edge and vertex groups G(e), G(v) are separable in π1 (G, Γ), or equivalently
G(e) are closed in G(d0 (e)), G(d1 (v)) with respect to the topology induced
by the profinite topology on G [CB13, Proposition 2.5]. In particular, this is
the case if vertex and edge groups are finitely generated and G is subgroup
separable.
In what follows, the notion of the fundamental pro-p group of an infinite
profinite graph of pro-p groups is used only in Theorem 3.1, and in fact only
in the form of the inverse limit of fundamental groups of finite graphs of
finite p-groups. We state this as a proposition that the reader can use as a
definition. We need to define first a profinite graph of profinite groups. The
definition is different from, but equivalent to, the one given in [ZM89]1 .
Definition 2.4. An inverse limit of finite graphs of finite groups is called a
profinite graph of profinite groups.
Proposition 2.5. An inverse limit (G, Γ) = limi (Gi , Γi ) of finite graphs of
←−
profinite (resp. pro-p) groups induces an inverse limit of the fundamental
profinite (resp. pro-p) groups limi Π1 (Gi , Γi ) and Π1 (G, Γ) = limi Π1 (Gi , Γi ).
←−
←−
Proof. One uses Definition 9 (a) in [Ser77] with a base vertex (rather than a
maximal subtree). Then by [Bas93, Proposition 2.4 ] a morphism of graphs of
groups induces a homomorphism of fundamental groups, that in turn induces
a homomorphism of profinite completions. So choosing a base vertex v0 in
Γ and considering its image voi in Γi to be the base point in Γi , we have an
inverse system {Π1 (Gi , Γi )}. The proof of Π1 (G, Γ) = limi Π1 (Gi , Γi ) is just a
←−
straightforward verification of the universal property from [ZM89, 2.1].
3
Finitely generated pro-p groups acting on
profinite trees
We now prove Theorem 1.1.
Theorem 3.1. Let G be a finitely generated pro-p group acting on a pro-p
tree T . Then G is the fundamental pro-p group of a profinite graph of pro-p
groups (G, Γ). Moreover, the vertex and edge groups of (G, Γ) are stabilizers
1
The equivalence is an easy argument which we leave as an exercise to the reader.
7
of certain vertices and edges of T respectively, and stabilizers of vertices and
edges of T in G are conjugate to subgroups of vertex and edge groups of (G, Γ)
respectively.
Proof. For every open normal subgroup U of G consider Ũ , a subgroup generated by all intersections with vertex stabilizers. Then by Proposition 3.5 and
Corollary 3.6 of [RZ00], the quotient group U/Ũ acts freely on the pro-p tree
T /Ũ and therefore it is free pro-p. Thus GU := G/Ũ is virtually free pro-p.
By [HZ13, Theorem 1.1] it follows that GU is the fundamental pro-p group
Π1 (GU , ΓU ) of a finite graph of finite p-groups with cyclic edge stabilizers.
By Remark 2.2 we have a morphism of graphs of groups η : (GU , ΓU ) −→
(GU , ΓU ) to a reduced graph of groups. By [WZar, Corollary 3.3] the number
of reduced graphs of groups (GU , ΓU ) that can be obtained from (GU , ΓU )
is finite. Let ΩU be the set of such reduced graphs of groups. Since G is
finitely generated we can choose a linearly ordered system B of open normal
subgroups U that form a basis of neighbourhoods of 1.
By [RZ14, Proposition 1.10] (see also its proof), for V ⊆ U both in B
with G/Ṽ written as the fundamental pro-p group G/Ṽ = Π1 (GV , ΓV ) of a
reduced graph of p-groups, one has a natural decomposition of G/Ũ as the
pro-p fundamental group G/Ũ = Π1 (GV,U , ΓV ) of a finite graph of finite pgroups (GV,U , ΓV ), where the vertex and edge groups are GV,U (x) = G(x)Ũ /Ũ ,
x ∈ V (ΓV )⊔E(ΓV ). Thus we have a morphism νV,U : (GV , ΓV ) −→ (GV,U , ΓV )
of graphs of groups such that the induced homomorphism on the pro-p fundamental groups coincides with the canonical projection ϕV,U : G/Ṽ −→ G/Ũ .
Choose a reduction morphism ηU : (GV,U , ΓV ) −→ (GV,U , ΓU ) to a reduced
graph of groups (GV,U , ΓU ) (it is not unique); it induces the identity map on
the fundamental group G/Ũ (see Remark 2.2) and so ηU νV,U induces the homomorphism Π1 (GV , ΓV ) −→ Π1 (GV,U , ΓU ) on the pro-p fundamental groups
that coincides with the canonical projection ϕU V : G/Ṽ −→ G/Ũ. Observe
that the morphisms νV,U and therefore ηU νV,U restricted on the second variable is just the collapse of fictitious edges. Thus, for V ⊆ U, one has a map
ωV,U : ΩV → ΩU . Since B is linearly ordered this yields an inverse system
(otherwise it would have not, since the reduction morphism is not unique).
Hence Ω = limU ΩU is non-empty. Let (GU , ΓU ) ∈ Ω. Then (G, Γ) given
←−
by (G, Γ) = lim(GU , ΓU ), is a profinite graph of pro-p groups satisfying G ≃
←−
Π1 (G, Γ) (cf. Proposition 2.5).
8
Corollary 3.2. Suppose the graphs ΓU in the proof of Theorem 3.1 are
bounded independently of U. Then Γ is finite and so G is the fundamental group of a finite graph of pro-p groups (G, Γ).
It is not known whether Γ can be made finite in Theorem 3.1. In the
abstract case it comes automatically. One could achieve it if one could bound
the size of a reduced finite graph of finite p-groups in terms of the minimal
number of generators of its pro-p fundamental group. In the case of cyclic
edge groups this is possible.
Corollary 3.3. Let G be a finitely generated pro-p group acting on a pro-p
tree T with cyclic edge stabilizers. Then G is the fundamental pro-p group of
a finite graph of pro-p groups (G, Γ). Moreover, the vertex and edge groups
of (G, Γ) are stabilizers of certain vertices and edges of T respectively, and
stabilizers of vertices and edges of T in G are conjugate to subgroups of vertex
and edge groups of (G, Γ) respectively.
Proof. By [SZ14, Lemma 2.2] in the case of cyclic edge stabilizers the size
of ΓU from the proof of Theorem 3.1 is bounded independently of U, so the
result follows from Corollary 3.2.
b on a profinite (or pro-p)
Definition 3.4. The action of a profinite group Γ
tree T is said to be k-acylindrical, for k a constant, if the set of fixed points
of γ has diameter at most k whenever γ 6= 1.
Proposition 3.5. Let G be a finitely generated pro-p group acting k-acylindrically
and virtually freely on a pro-p tree. Then G is a free pro-p product of finite
p-groups and a free pro-p group.
Proof. Let F be an open subgroup of G acting freely on T . By [RZ00,
Corollary 3.6] F is free pro-p. By [HZ08, Theorem 1] a virtually free pro-p
group whose torsion elements have finite centralizers satisfies the conclusion
of the proposition. So we just need to show that torsion elements have finite
centralizers. Let g be a torsion element of G and let T g be the subtree of
fixed points (see [ZM88, Theorem 2.8]. Then CG (g) acts on T g . Since the
action is virtually free |CG (g)| = ∞ would imply |T g | = ∞ contradicting the
hypothesis. Hence CG (g) is finite. This finishes the proof.
Remark 3.6. The proposition is valid for second countable pro-p groups. One
just needs to replace the reference [HZ08] in the proof with [MZ17].
9
For a normal subgroup U of a profinite group G acting on a profinite
tree T we denote by Ũ the subgroup of U generated by all vertex stabilizers.
Clearly Ũ is normal in G and by [ZM88, Prop. 2.5] T /Ũ is a profinite tree.
Lemma 3.7. Let G be a pro-p group acting k-acylindrically on a profinite
tree T and U a closed normal subgroup of G. Then the action of GU = G/Ũ
on TU = T /Ũ is k-acylindrical.
Proof. First note that a nontrivial stabilizer of any edge e coincides with
one of its vertex stabilizers Gv or Gw , say Gw , since otherwise by [WZ17,
Lemma 11.2] the stabilizers of two vertices of e do not generate a pro-p
group. It follows that a maximal connected subgraph D of T having nontrivial stabilizers in G of all its edges has diameter at most k.
We shall prove that the same is true in GU ; we use the subscript U for
images in T /Ũ . Consider 1 6= gU ∈ GU that stabilizes a vertex in TU . It
suffices to show that a maximal connected subgraph containing vU such that
all edge stabilizers of it are non-trivial is contained in DU . Let eU be an edge
not in DU having a vertex wU ∈ DU . Let g ∈ G, stabilizing a vertex v ∈ T ,
be such that g Ũ/Ũ = gU , and choose D with v ∈ D. Let w be a vertex
of D whose image in TU is wU . Then Gw 6= {1} and there exists an edge e
incident to w whose image in TU is eU . But Ge is trivial and so (GU )eU is
trivial. Thus all edges not in DU with vertices in DU have trivial stabilizers
in GU , as required.
Proposition 3.5 and Lemma 3.7 enable us to generalize [WZ17, Theorem
11.1] from the 1-acylindrical case to the k-acylindrical case, for any natural
number k.
Theorem 3.8. Let G be a finitely generated pro-p group acting k-acylindrically
on a profinite tree T . Then G is a free pro-p product of vertex stabilizers and
a free pro-p group.
Proof. Since G is finitely generated, its Frattini series Φn (G) is a fundamen^
n (G) be the subgroup of Φn (G)
tal system of neighbourhoods of 1. Let Φ
^
n (G)
generated be all vertex stabilizers of Φn (G). By Lemma 3.7, Gn = G/Φ
^
n (G) and so by Proposition
acts k-acylindrically on a profinite tree Tn = T /Φ
3.5 is a free pro-p product
!
a
^
n (G) =
Gn = G/Φ
Gnv ∐ F0n ,
v
10
of finite p groups Gnv and a free group F0n in Gn .
Since G = lim Gn and G is finitely generated, by choosing n large enough
←−
we may assume that the number of free factors is the same for every m >
n, i.e. v ranges over a finite set V . By [RZ00, Theorem 4.2] every finite
subgroup of a free pro-p product is conjugate to a subgroup of a free factor.
Therefore, the free factors of Gm+1 are mapped onto the free factors of Gm
up to conjugation. But in a free pro-p product decomposition replacing any
free factor by its conjugate does not change the group. So, starting from
n, we can inductively choose Gm+1v in such a way that its image in Gm is
Gmv .`Let Gv be the inverse limit of Gmv . Then by [RZ10, Lemma 9.1.5]
G = v∈V Gv ∐ F0 .
It remains to observe that Gmv is a stabilizer of a vertex in Tm so the set of
fixed points T Gmv is non-empty and closed [RZ00, Theorem 3.7]. Therefore,
Gv is the stabilizer of a non-empty set of vertices lim T Gvm . This finishes the
←−
proof.
We shall need these auxiliary lemmas in Section 4.
Lemma 3.9. Let H be a pro-p group acting on a profinite tree T with compact
non-empty set of edges having only finitely many maximal vertex stabilizers
up to conjugation. Then H acts on a profinite tree D with non-empty set of
edges such that, for every edge e of D, the edge stabilizer He fixes at least
two edges in the original tree T .
Proof. Let Hv1 , . . . Hvn be maximal stabilizers of vertices in T up to conjugation. Then the stabilizer Hm of any edge or vertex m ∈ T is contained in
some of Hvi up to H-translation of m. So assuming Hm ≤ Hvi we see that
Hm fixes the geodesic [m, vi ]. Since E(T ) is compact, Hm stabilizes an edge
incident to vi whenever m 6= vi by [ZM88, Lemma 2.15]. So
Snif m 6∈ star(vi ),
the stabilizer Hm fixes at least two distinct edges. Now i=1 Hstar(vi ) is
a profinite subgraph of T and, collapsing all its connected components in
T , by the Proposition on page 486 of [Zal92], we get a profinite tree D on
which H acts with edge stabilizers each stabilizing at least two distinct edges
of S
T (as was just explained). Finally note that the connected components
of ni=1 Hstar(vi ) are stars, since the geodesic [gvi , hvj ] is infinite whenever
gvi 6= hvj and g, h ∈ H. Indeed, otherwise choose a geodesic [gvi , hvj ] of
minimal length with gvi 6= hvj for g, h ∈ H. To arrive at a contradiction
we show that the group K generated by the stabilizers of gvi and hvj is not
11
pro-p. Note that B = K[gvi , hvj ] is a profinite tree with B/K finite. Since
[gvi , hvj ] is minimal, the stabilizers of all internal vertices of the geodesic are
equal to the stabilizers of their incident edges. Thus collapsing all but one
edges of this geodesic and their K translates, we may assume that B/K has
just one edge. Now we can apply [WZ17, Lemma 11.2 ] to deduce that K is
not pro-p as required.
Lemma 3.10. Let G be a profinite group acting on a profinite tree T with
compact set of edges such that any non-trivial subgroup K of G stabilizing
two edges e1 , e2 is at most procyclic. Let H be a finitely generated pro-p
subgroup of G. Then H is the fundamental pro-p group Π1 (H, ∆) of a finite
graph of finitely generated pro-p groups with edge groups that are stabilizers
of some pairs of distinct edges of T . In particular, if the action
of G on T is
`
1-acylindrical then H splits as a free pro-p product H = v∈V (∆) H(v) ∐ F ,
where F is a free pro-p group intersecting trivially all conjugates of vertex
groups in H.
Proof. First note that a nontrivial stabilizer of any edge e coincides with one
of its vertex stabilizers Hv or Hw , say Hw , since otherwise by [WZ17, Lemma
11.2] the stabilizers of two vertices of e do not generate a pro-p group. This
means that connected components of the abstract subgraph whose edges have
non-cyclic H-stabilizers are at most edges.
Now since H is finitely generated, its Frattini series Φn (H) is a fundan (H) be the subgroup of Φn (H)
mental system of neighbourhoods of 1. Let Φ^
n (H) acts on a
generated be all vertex stabilizers of Φn (H). Then Hn = H/Φ^
n (H) (cf. [ZM88, Proposition 2.5]) and Φn (H)/Φ
^
n (H)
profinite tree Tn = T /Φ^
acts freely on Tn and therefore is free pro-p (see Theorem 2.6 [ZM88]). Note
that the vertex and edge stabilizers of Hn acting on Tn are finite epimorphic images of the corresponding vertex and edge stabilizers of H acting on
T and so the images in Tn of edges of T with trivial edge stabilizers have
trivial stabilizers. Therefore, every abstract connected component Ωn of the
subgraph of points of Tn with non-cyclic edge stabilizers in Hn is exactly the
image of some abstract connected components Ω of the subgraph of points
of T with non-procyclic H-stabilizers in T . Indeed, suppose on the contrary
there exists an edge en ∈ Ωn not in the image of Ω but having a vertex vn in
it. Let v be a vertex of Ω whose image in Tn is vn . Then Hv is not procyclic
and there exists an edge e incident to v whose image in Tn is en . But e 6∈ Ω,
so He is procyclic and therefore Hen is cyclic, contradicting the choice of en .
12
By [HZ08, Lemma 8], a virtually free pro-p group has only finitely many
finite subgroups up to conjugation. This means that Tn has only finitely
many edges en up to translation with (Hn )en that are not cyclic. Hence the
union of them is a profinite subgraph of S and collapsing all its connected
components in Tn , by the Proposition on page 486 in [Zal92] we get a profinite
tree Dn on which Hn acts with cyclic edge stabilizers. Hence, by Corollary
3.3, Hn is the fundamental pro-p group of a finite graph of finite p-groups
(Hn , Γn ) with cyclic edge groups, where the vertex and edge groups of (Hn , Γ)
are stabilizers of certain vertices and edges of Dn respectively. Moreover, by
collapsing edges that are not loops with a vertex group equal to the edge
group we may assume that (Hn , Γ) is reduced (see Remark 2.2). Then by
[SZ14, Lemma 2.2], the size of Γ is limited in terms of the minimal number
of generators d(H) of H, i.e. is bounded independently of n. Therefore,
by Corollary 3.2, H is the fundamental group Π1 (H, Γ) of a finite graph of
pro-p groups with cyclic edge stabilizers whose vertex and edge groups are
stabilizers of certain vertices and edges of T . In particular, it has only finitely
many maximal stabilizers of vertices in T up to conjugation. Then we can
apply Lemma 3.9 to deduce that H acts on a profinite tree D such that
every edge stabilizer He , e ∈ D fixes at least two edges in T . Now applying
Corollary 3.2 again we deduce that H is the fundamental group Π1 (H, ∆) of
a finite graph of pro-p groups satisfying the statement of the lemma.
Finite generation of the vertex groups H(v) follow easily from the presentation of Π1 (H, ∆), the finiteness of ∆, and the fact that the edge groups
H(e) are procyclic.
To show the last statement, observe that in the case of a 1-acylindrical
action the edge groups H(e) of (H, ∆) are trivial (since each of them stabilizes
a pair of distinct edges of T ), so the result follows from the presentation (1)
of Π1 (H, ∆).
We finish the section with two simple lemmas.
Lemma 3.11. Let G = HNN(F (Y ∪ X), Ci , ti ), i = 1, . . . n be a pro-p
HNN-extension of a free pro-p group F = F (Y ∪ X) on a finite set Y ∪ X
ti
such that
` each Ci and Ci are a conjugate of hyi in F for some y ∈ Y . Then
G = y∈Y \Z (hyi × F (Ty )) ∐ F (Z ∪ X) is a free pro-p product, where Z is
a subset of Y and Ty is a (possibly empty) subset of the set of stable letters
t1 , . . . tn .
Proof. We use induction on the cardinality of Y . If there exists ti such
13
that y1f1ti = y2f2 for some y1 6= y2 and f1 , f2 ∈ F then we can rewrite G as
G = HNN(F (Y \ y2 ∪ (X ∪ f1 ti f2−1 ), Cj , tj ), j = 1, . . . i − 1, i + 1 . . . , n and
the result follows from the inductive hypothesis.
Thus we may assume that Cifi = Citi for some fi ∈ F for all i. Then
replacing ti with ti fi−1 we may assume that ti centralizes Ci for every i.
Moreover, conjugating Ci in F , we may assume that every Ci is generated
by an element of Y . Now the result follows from the presentation G =
HNN(F (Y ∪ X), Ci , ti ), i = 1, . . . n.
Lemma 3.12. (i) Let H = H1 ∐H0 H2 be a pro-p free amalgamated product
over a procyclic group H0 . Suppose CH (H0 ) is abelian of rank at most
2 and Hi = CH (H0 ) ∐ Li . Then H = CH (H0 ) ∐ L1 ∐ L2 .
(ii) Let H = HNN(H1 , H0 , t) be a pro-p HNN-extension over a procyclic
subgroup H0 . Suppose CH (H0 ) is abelian of rank at most 2, H1 =
CH (H0 ) ∐ L1 and H0t is conjugate in H1 either to H0 or to a subgroup
of L1 . Then either H = (Zp × H0 ) ∐ L1 or H = Zp ∐ L1 .
Proof. (i) Since CH (H0 ) is abelian of rank at most 2, either H0 is selfcentralized in H1 or it is self-centralized in H2 ; say H0 is self-centralized
in H1 . Then H = L1 ∐ H2 and the result follows.
(ii) If H0 is conjugate to a subgroup
of L1 then H = L1 ∐ hti. If H0t = H0h1
`
−1
for some h1 ∈ H1 then H = L1 (hth1 i × H0 ) so we are done as well.
4
4.1
Pro-p subgroups of profinite completions of
3-manifold groups
Reduction to the geometric cases
Let M be a compact, orientable, non-geometric 3-manifold. Since every
compact manifold is a retract of a closed manifold, it follows easily that
we may reduce to the case in which M is closed. If M is reducible then
its Kneser–Milnor decomposition is non-trivial. In particular, π1 M acts on
a tree with trivial edge stabilizers (i.e. 0-acylindrically), and with vertex
stabilizers exactly the conjugates of the reducible factors of M; such an
action is 0-acylindrical. Since 3-manifold groups are residually finite, π[
1M
acts 0-acylindrically on a profinite tree, with vertex stabilizers the conjugates
14
of the completions of the irreducible factors. Therefore, by Theorem 3.8, we
may reduce to the case where M is irreducible.
If M is non-geometric, we now argue similarly using the JSJ decomposition. The action of an irreducible, non-geometric 3-manifold on its JSJ
tree is 4-acylindrical [WZ10, Lemma 2.4]. The same is true for the action of
the profinite completion on the profinite tree, as follows essentially from the
results of [WZ10] and [HWZ13]. For completeness and ease of reference, we
state the result here.
Proposition 4.1. Let M be a closed, irreducible 3-manifold. Then the action
[
of π
1 M on the profinite tree associated to the JSJ decomposition of M is 4acylindrical.
This was stated in [WZ17, p. 376], and a careful proof was written down in
[Wil17, Proposition 6.8]. Proposition 4.1 follows in a straightforward manner
from the following statement.
Lemma 4.2. Let N be geometric 3-manifold with incompressible toral boundary, not homeomorphic to an interval bundle over a torus or Klein bottle, and
let P1 , P2 be two Z2 subgroups conjugate to the fundamental groups of boundγ̂
ary components. Then, for any γ̂ ∈ πd
1 N, P 1 ∩ P 2 is equal to the maximal
normal procyclic subgroup of πd
1N .
Gareth Wilkes has pointed out that the results of [WZ10] and [HWZ13]
in fact prove a weaker statement than Lemma 4.2, since they only handle
the case in which γ̂ ∈ π1 N. However, similar techniques do indeed establish
Lemma 4.2, and hence Proposition 4.1. For the Seifert fibred case, one argues
as in [WZ10], appealing to [CZ07, Proposition 3.4 and Lemma 3.5] instead
of [RZ96, Lemma 3.6]. (Alternatively, a proof using more recent technology
might appeal to [WZ17, Theorem 3.3].) For the hyperbolic case, one argues
as in [HWZ13], but replaces [HWZ13, Lemma 4.7] by [WZ17, Lemma 4.5].
With Proposition 4.1 in hand, by Theorem 3.8, a finitely generated pro-p
[
subgroup of π
1 M is a free pro-p product of pro-p subgroups of the vertex
groups, so it suffices to consider the geometric cases.
4.2
Hyperbolic manifolds
The closed hyperbolic case is the subject of [WZ17, Theorem F]; in this case
our subgroup is free pro-p.
15
Thus we consider cusped hyperbolic 3-manifolds in this subsection. Recall that a subgroup of π1 M is called peripheral if it is conjugate to the
fundamental group of a cusp and so is isomorphic to Z × Z. A conjugate of
[
the closure of a peripheral group in π
1 M will also be called peripheral; it is
b
b
isomorphic to Z × Z.
Definition 4.3. Suppose that a group G is hyperbolic relative to a collection
of parabolic subgroups {P1 , . . . , Pn }. A subgroup H of G is called relatively
malnormal if, whenever an intersection of conjugates H γ ∩H is not conjugate
into some Pi , we have γ ∈ H.
It is well known that π1 M is hyperbolic relative to its peripheral subgroups. Following Wise, we analyse the cusped hyperbolic 3-manifold M by
cutting along an embedded family of surface Σi so every cusp is cut by some
Σi . The following theorem is [WZ17, Theorem 9.1].
Theorem 4.4. A hyperbolic 3-manifold M with cusps has a finite-sheeted
covering space N → M that contains a disjoint family of connected, geometrically finite, incompressible subsurfaces {Σ1 , . . . , Σn } such that:
1. each cusp of N contains a boundary component of some Σi ;
2. each π1 Σi ⊆ π1 N is relatively malnormal.
Cutting along the family of surfaces given by Theorem 4.4 produces a
graph-of-groups decomposition (G, ∆) for G = π1 N, with the property that
every vertex and edge group is hyperbolic and virtually special. Any peripheral subgroup Pi leaves invariant an infinite (Tits’) axis and splits into
direct product of infinite cyclic groups Pi = Li × Ri such that Li fixes every edge of the axis and Ri acts by translations. The stabilizer of any pair
of distinct edges of the Bass–Serre tree S is conjugate to such Li for some
peripheral subgroup Pi . Moreover, π1 N is subgroup separable (see [AFW15,
Corollary 5.5]) and so by Example 2.3 we can pass to the corresponding
b ∆) such that G
b ∆) with G
b = Π1 (G,
b acting on
graph of profinite groups (G,
b By [WZ17, Theorem 4.2], π\
the standard profinite tree S(G).
1 (Σi ) are relatively malnormal in the profinite completion of peripheral subgroups and so
b is conjugate to
every non-trivial stabilizer of a pair of distinct edges of S(G)
bi ≤ Pbi .
a certain L
We shall need however a more specific hierarchy on vertex groups G(v) of
G given by the next proposition.
16
Theorem 4.5. Let M be a cusped hyperbolic 3-manifold. The fundamental
group π1 M possesses a finite index subgroup G that admits a a graph-ofgroups decomposition (G, Γ) such that:
1. every vertex and edge group is hyperbolic and virtually special, and the
stabilizer of any pair of two distinct edges of the Bass–Serre tree S is
cyclic and is conjugate to a cyclic subgroup Li of a peripheral subgroup;
2. moreover, each vertex group G(v) has a malnormal hierarchy, which
can be chosen so that, at each step, every Li belongs to a vertex group
of the corresponding step-decomposition (so that at the last step of the
vertex group being a free product of cyclic groups each Li is contained
in some cyclic free factor).
Proof. We consider the finite-sheeted cover N of M provided by Theorem
4.4. The fundamental group of N comes equipped with a decomposition that
satisfies item (1) above. For item (2), we appeal to [AGM16, Theorem 2.11].
Indeed, let Nv be a component of the given decomposition of N. Then the
Li that are conjugate into Nv from a malnormal family Lv , and so the pair
(π1 Nv , Lv ) is relatively hyperbolic by [Bow12, Theorem 7.11]. Since π1 Nv
is also hyperbolic and virtually special, we may apply [AGM16, Theorem
2.11] to obtain a normal subgroup Gv of finite index in π1 Nv so that the
induced pair (Gv , L′v ) ⊳ (π1 Nv , Lv ) admits, in the terminology of [AGM16],
a malnormal, quasiconvex, fully L′v -elliptic hierarchy. This means precisely
that there is a finite hierarchy in which every vertex group is hyperbolic and
virtually special, and every edge group is malnormal and quasiconvex, and
in which the intersection with every conjugate of Li ∈ Lv is always elliptic,
terminating in free products of groups from L′v and free groups.
Since the subgroups π1 Nv are separable in π1 N, we may pass to a finiteindex normal subgroup G of π1 N so that every vertex group of the induced
decomposition is contained in some π1 Nv . The subgroup G then has the
properties claimed.
We deduce from Lemma 3.10 the following.
Corollary 4.6. Let H be a finitely generated pro-p subgroup of the profinite
b Then H = (H, ∆) is the fundamental pro-p group of a finite
completion G.
graph of finitely generated pro-p groups with cyclic edge groups conjugate to
subgroups of Li , and where the vertex groups of (H, ∆) are conjugate to some
vertex groups of (G, Γ).
17
Corollary 4.7. Let H be a finitely generated pro-p subgroup of the profinite
b
b
completion of a vertex group G(v).
Let ṽ be a vertex in the standard tree S(G)
mapping to v; consider the action of H on the set of incident edges at ṽ, and
let e1 , . . . , en be a set of orbit representatives, with corresponding
stabilizers
`
H(ei ) in H. Then H splits as a free pro-p product H = i H(ei ) ∐ F , where
F is a free pro-p group.
Proof. We use induction on the hierarchy of Theorem 4.5 for G(v). So let K
be a subgroup that belongs to this hierarchy.
The base of the induction is the case in which K is a free product of cyclic
vertex groups K(v). Since Pb is malnormal in π[
1 M (see [WZ17, Lemma 4.5])
g
b
b
and H(e) = Li for some g ∈ G, in this case each non-trivial intersection
bg = K(v) so each H(e) = L
bg generates a free factor.
K(v) ∩ H(e) = K(v) ∩ L
i
i
Now,
to
perform
the
inductive
step,
note
that
in
this case by Lemma 3.10
`
H = v∈V (∆K ) H(v) ∐ F is a free pro-p product of vertex groups and a free
pro-p group F intersecting trivially all conjugates of vertex groups. Applying
the induction hypothesis to each vertex group HK (v) and using that a pro-p
free factor of a finitely generated pro-p group is finitely generated, we deduce
the statement.
We are now ready to deduce the classification of pro-p subgroups in the
cusped hyperbolic case.
Theorem 4.8. Let π1 M be the fundamental group of an orientable hyperbolic
3-manifold with k cusps, and H a finitely generated pro-p subgroup of π[
1M .
Then H is a free pro-p product of free abelian pro-p groups of rank ≤ 2.
Proof. Let G be the finite-index subgroup of π1 M given by Theorem 4.5. We
b By Corollary 4.6, H is the fundamental
prove first the theorem for H ≤ G.
pro-p group of a finite graph of finitely generated pro-p groups (H, ∆), with
bi , where the vertex groups of
cyclic edge groups conjugate to subgroups of L
b Moreover, by
(H, ∆) are stabilizers of certain vertices and edges of`
S(G).
Corollary 4.7, for each vertex v ∈ ∆ we have H(v) = ( He ≤H(v) H(e)) ∐ Lv .
We use induction on the size of ∆ to prove that H is a free pro-p product
bg g ∈ G
b being
of free abelian pro-p groups of rank ≤ 2, with every H ∩ L
i
conjugate to a free factor. If`∆ is a vertex then the result just follows from
the decomposition H(v) = ( He ≤H(v) H(e)) ∐ Lv mentioned above.
Now let e be an edge of ∆. Then H = H1 ∐H(e) H2 or H = HNN(H1 , H(e), t),
where H1 and H2 are the fundamental groups of the graph of groups (H, ∆)
18
restricted to the connected components ∆i of ∆ \ {e} (i = 1, 2 o i = 1). If
H(e) = 1 then the amalgamated free products or HNN-extensions are free
products and the result follows directly from the induction hypothesis.
Suppose H(e) 6= 1. By the induction hypothesis we may assume that
Hi = (H(e) × H0i ) ∐ Li . Since the profinite completion Pb of a parabolic
subgroup P is malnormal in π[
1 M the centralizer of H(e) in H is abelian of
at most rank 2 so the result follows from Lemma 3.12.
Now, to prove the general case, we observe that by what we just proved
b satisfies the statement of the theorem. Then the statement for H
H∩G
follows from [Zal16, Theorem 1.2].
4.3
Seifert fibred manifolds over hyperbolic bases
In this section we deal with the cases in which N admits a geometric structure
^
modelled on either H2 × R or SL
2 (R). Recall that, when N is Seifert fibred,
the fundamental group π1 N is torsion-free of the form
1 → Z → π1 N → π1 O → 1
(not necessarily central) where π1 O is a Fuchsian group. (We refer the reader
to [Sco83] for details.) Our classification theorem in this case shows that
the pro-p subgroups of the profinite completion of π1 N are either free pro-p
products, or pro-p completions of fundamental groups π1 M, where M is a
non-compact Seifert fibred manifold with residually p fundamental group.
Theorem 4.9. Let G be the fundamental group a Seifert fibred manifold with
^
either H2 × R or SL
2 (R) geometry and H a finitely generated pro-p subgroup
b Then H is a cyclic extension of a free pro-p-product of cyclic pro-p
of G.
groups.
Proof. Consider H0 , the image of H in πd
1 O, and note that H0 is a finitely
generated, pro-p subgroup. It suffices to prove that H0 is a free pro-p-product
of cyclic pro-p groups. Indeed, let S be a surface subgroup of finite index of
π1 O. Then H ∩ Sb is free pro-p by [Zal05, Proposition 1.2].
We now prove that the centralizer in πd
1 O of a torsion element of H0 is
d
finite. Since no open subgroup of π1 O is pro-p, we may replace O by a nontrivial finite-sheeted cover of degree n > 2, where n is coprime to p, whose
completion contains H0 . We may therefore assume that π1 O is not a triangle
19
group. If π1 O is torsion-free there is nothing to prove; otherwise, π1 O splits
as an amalgamation over an infinite cyclic group π1 O = K ∗C F , where
F is free and K = ∗i Ci is a free product of cyclic groups. It follows that
b b Fb is a free profinite product with cyclic amalgamation. Applying
πd
1 O = K ∐C
b
[ZM88, Theorem 3.10] twice, any finite subgroup of πd
1 O is conjugate into K,
b ∩K
b g is contained in
and thence to some Ci . By [ZM88, Theorem 3.12], K
b
b It follows that the centralizer C d (ci ) of any element
a K-conjugate
of C.
π1 O
b and so by [RZ10, Theorem 9.1.12], the centralizer
ci of Ci is contained in K,
coincides with Ci , as required.
By [HZ08, Theorem 1], virtually free pro-p groups in which torsion elements have finite centralizers split as a free pro-p product of finite subgroups
and a free pro-p group. Applying this to H0 now completes the proof.
4.4
Sol and Nil manifolds
Suppose that N is a compact 3-manifold with solvable fundamental group.
The groups that arise in this way were classified by Evans and Moser [EM72],
and are all virtually polycyclic. It follows from this classification that a
compact, orientable manifold N with Nil or Sol geometry has a cover N0 of
index at most two such that π1 N0 ∼
= Z2 ⋊ Z, and we deduce the following.
Theorem 4.10. Let N be a compact, orientable, Nil- or Sol-manifold. If p is
odd then the Sylow p-subgroup of πd
1 N is isomorphic to the pro-p completion
2
of a semidirect product Z ⋊ Z, and for p = 2, the Sylow p-subgroup has a
subgroup of index at most 2 with this structure.
Proof. Let H = π1 N0 as above. Denote by Op (G) the kernel of the epimorb −→
phism of a profinite group G to its maximal pro-p quotient. Let f : H
2
b
b
H/Op (Z ) be the canonical map. Clearly f restricts to an injection on H, so
b 2 ) is Z2 on which H/O
b p (H)
b acts
we identify H with f (H). The image f (Z
p
b p (H)
b −→ GL2 (Zp ).
by conjugation. Thus we have a homomorphism η : H/O
Note that GL2 (Zp ) is virtually pro-p. Hence the preimage U of a p-Sylow subgroup of η(H) is open in H/Op (H). Consider the intersection K = U ∩ π1 N,
a finite-index subgroup of π1 N. The pro-p completion Kp̂ has a normal
subgroup Z2p which is the closure of Z2 and the kernel of the natural map
Kp̂ −→ U intersects it trivially. One deduces then that this kernel is trivial,
and the result follows.
20
4.5
The Euclidean case
This case consists of all 3d Bieberbach groups – i.e. torsion-free Z3 -by-finite
groups with the finite groups of the following 32 types [Hah87, Chapter 10.1,
p. 794]: 27 cyclic and dihedral groups, 3 tetrahedral groups T, Td , Th and 2
octahedral groups O and Oh . The orders of the these groups are divisors of
48, so only the primes p = 2, 3 can lead to non-trivial Sylow subgroups. For
p = 3, all Sylow subgroups are C3 . See also [EM72, Lemma 4.9].
Suppose p = 2. Then we have the following list.
1. Cyclic groups: C2 , C4 , C8 .
2. Dihedral groups: D2 , D4 , D8 of orders 4, 8, 16 respectively.
3. Tetrahedral groups T, Td , Th have orders 12, 24, 24 and their Sylow 2subgroups are D4 and D8 , so we do not get anything new from the
previous item.
4. The two octohedral groups O ∼
= S4 and Oh of order 24 and 48 respectively. The Sylow 2-subgroup of the first one is D4 , and of the second
is the generalized quaternion group Q16 .
Thus we can deduce the following:
Proposition 4.11. The Sylow p-subgroup of the profinite completion of the
fundamental group of a Euclidean 3-manifold is a pro-p group of one of the
following types:
(i) Z3p for any p.
(ii) if p = 3, Z33 -by-C3 .
(iii) p = 2, Z32 -by-C2 , -C4 , -C8 , -D2 , -D4 , -D8 , -Q16 .
Remark 4.12. Each extension may yield several isomorphism classes, so this
does not give the number of groups.
There are the cases of Z and D∞ . We do not get anything new, just the
pro-p and pro-2 completions respectively.
21
4.6
Interval bundles over Klein bottles
The JSJ decomposition of a non-geometric 3-manifold M may include pieces
N which are homeomorphic to twisted interval bundles over the Klein bottle,
with fundamental group the Klein-bottle group Z⋊Z. These can be removed
by passing to a double cover, but when p = 2, the pro-2 completion of Z ⋊ Z
\
arises as the Sylow 2-subgroup of Z
⋊ Z. We record this observation for later
use.
Proposition 4.13. The Sylow 2-subgroup of the profinite completion of the
fundamental group of the twisted interval bundle over the Klein bottle is the
pro-2 completion of Z ⋊ Z.
4.7
Spherical manifolds
We shall use here [Wol11, Theorem 2.4 ].
Theorem 4.14. (Theorem 2.4 [Wol11] ) Finite fixed-point free subgroup of
SO(4) belongs to the following list:
1. Cyclic group Cn of order n.
2. Dihedral groups D2k (2n+1) , for k ≥ 2 and n ≥ 1, with presentation
k
hx, y | x2 = y 2n+1 = 1, xyx−1 = y −1i of order 2k (2n + 1).
3. Generalised quaternion groups Q4k of order 4k.
4. The binary icosahedral group I = SL2 (F5 ) of order 120.
′
5. The groups P8·3
k = Q8 ⋊ C3k defined by the following presentation
k
2
hx, y, z|x = (xy)2 = y 2, zxz −1 = y, zyz −1 = xy, z 3 = 1i of order
8 · 3k .
6. The direct product of any of the above group with a cyclic group of
relatively prime order.
We shall list p-subgroups of all these groups. Since Case 6 does not
produce anything new we analyse cases 1–5 only.
Looking at the orders we see that for p odd p-Sylow subgroups are cyclic.
So we concentrate on the case p = 2. Then we have only 3 types of 2groups: cyclic, dihedrals D2k , generalized quaternions Q2n . Thus we deduce
the following
22
Theorem 4.15. Let G be the profinite completion of the fundamental group
of a spherical 3-manifold.
(a) If p is odd, its Sylow p-subgroups are cyclic.
(b) If p = 2, we have the following list of possible Sylow 2-subgroups:
1. cyclic;
2. dihedrals D2k ;
3. generalized quaternions Q2n .
Putting the above results together we obtain Theorem 1.3.
References
[AFW15] Matthias Aschenbrenner, Stefan Friedl, and Henry Wilton. 3manifold groups. Series of Lecture Notes in Mathematics. European Mathematical Society (EMS), Zürich, 2015.
[AGM16] Ian Agol, Daniel Groves, and Jason Fox Manning. An alternate
proof of Wise’s malnormal special quotient theorem. Forum Math.
Pi, 4:e1, 54, 2016.
[Bas93]
Hyman Bass. Covering theory for graphs of groups. J. Pure Appl.
Algebra, 89(1-2):3–47, 1993.
[Bow12]
B. H. Bowditch. Relatively hyperbolic groups. Internat. J. Algebra
Comput., 22(3):1250016, 66, 2012.
[CB13]
Owen Cotton-Barratt. Detecting ends of residually finite groups
in profinite completions. Math. Proc. Cambridge Philos. Soc.,
155(3):379–389, 2013.
[CZ07]
S. C. Chagas and P. A. Zalesskii. Limit groups are conjugacy
separable. Internat. J. Algebra Comput., 17(4):851–857, 2007.
[EM72]
Benny Evans and Louise Moser. Solvable fundamental groups of
compact 3-manifolds. Trans. Amer. Math. Soc., 168:189–210, 1972.
23
[Hah87]
Theo Hahn, editor. International tables for crystallography. Vol. A.
Published for the International Union of Crystallography, Chester;
by D. Reidel Publishing Co., Dordrecht, 5-th edition, 2002. Spacegroup symmetry.
[HWZ13] Emily Hamilton, Henry Wilton, and Pavel A. Zalesskii. Separability of double cosets and conjugacy classes in 3-manifold groups.
J. Lond. Math. Soc. (2), 87(1):269–288, 2013.
[HZ08]
W. Herfort and P. A. Zalesskii. Virtually free pro-p groups whose
torsion elements have finite centralizer. Bull. Lond. Math. Soc.,
40(6):929–936, 2008.
[HZ13]
Wolfgang Herfort and Pavel Zalesskii. Virtually free pro-p groups.
Publ. Math. Inst. Hautes Études Sci., 118:193–211, 2013.
[MZ17]
J. W. MacQuarrie and P. A. Zalesskii. Second countable virtually
free pro-p groups whose torsion elements have finite centralizer.
Selecta Math. (N.S.), 23(1):101–115, 2017.
[NS07]
Nikolay Nikolov and Dan Segal. On finitely generated profinite
groups. I. Strong completeness and uniform bounds. Ann. of Math.
(2), 165(1):171–238, 2007.
[RZ96]
Luis Ribes and Pavel A. Zalesskii. Conjugacy separability of amalgamated free products of groups. Journal of Algebra, 179(3):751—
774, 1996.
[RZ00]
Luis Ribes and Pavel Zalesskii. Pro-p trees and applications. In
New horizons in pro-p groups, volume 184 of Progr. Math., pages
75–119. Birkhäuser Boston, Boston, MA, 2000.
[RZ10]
Luis Ribes and Pavel Zalesskii. Profinite groups, volume 40 of
Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A
Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in
Mathematics]. Springer-Verlag, Berlin, second edition, 2010.
[RZ14]
Luis Ribes and Pavel A. Zalesskiı̆. Normalizers in groups and in
their profinite completions. Rev. Mat. Iberoam., 30(1):165–190,
2014.
24
[Sco83]
Peter Scott. The geometries of 3-manifolds. The Bulletin of the
London Mathematical Society, 15(5):401–487, 1983.
[Ser77]
Jean-Pierre Serre.
Arbres, amalgames, SL2 , volume 46 of
Astérisque. Société Mathématique de France, Paris, 1977.
[SZ14]
Ilir Snopce and Pavel A. Zalesskii. Subgroup properties of prop extensions of centralizers. Selecta Math. (N.S.), 20(2):465–489,
2014.
[Wil17]
Gareth Wilkes. Profinite rigidity of graph manifolds and JSJ decompositions of 3-manifolds. arXiv:1605.08244v2, 2017.
[Wol11]
Joseph A. Wolf. Spaces of constant curvature. AMS Chelsea Publishing, Providence, RI, sixth edition, 2011.
[WZ10]
Henry Wilton and Pavel Zalesskii. Profinite properties of graph
manifolds. Geometriae Dedicata, 147(1):29—45, 2010.
[WZ17]
Henry Wilton and Pavel Zalesskii. Distinguishing geometries using
finite quotients. Geom. Topol., 21(1):345–384, 2017.
[WZar]
Thomas S. Weigel and Pavel A. Zalesskii. Virtually free pro-p
products. Israel J. Math., to appear.
[Zal92]
P. A. Zalesskiı̆. Open subgroups of free profinite products. In
Proceedings of the International Conference on Algebra, Part 1
(Novosibirsk, 1989), volume 131 of Contemp. Math., pages 473–
491. Amer. Math. Soc., Providence, RI, 1992.
[Zal05]
P. A. Zalesskii. Profinite surface groups and the congruence kernel
of arithmetic lattices in SL2 (R). Israel J. Math., 146:111–123,
2005.
[Zal16]
Pavel A. Zalesskii. Virtually free pro-p products. J. Group Theory,
19(3):543–560, 2016.
[ZM88]
P. A. Zalesskii and O. V. Mel’nikov. Subgroups of profinite groups
acting on trees. Matematicheskii Sbornik, 135(177)(4):419—439,
1988.
25
[ZM89]
P. A. Zalesskii and O. V. Mel’nikov. Fundamental groups of graphs
of profinite groups. Algebra i Analiz, 1(4):117—135, 1989.
26
| 4math.GR
|
TESTING FOR PRINCIPAL COMPONENT DIRECTIONS
UNDER WEAK IDENTIFIABILITY
arXiv:1710.05291v1 [math.ST] 15 Oct 2017
By Davy Paindaveine∗ , Julien Remy and Thomas Verdebout
Université libre de Bruxelles
We consider the problem of testing, on the basis of a p-variate
Gaussian random sample, the null hypothesis H0 : θ 1 = θ 01 against
the alternative H1 : θ 1 6= θ 01 , where θ 1 is the “first” eigenvector of
the underlying covariance matrix and θ 01 is a fixed unit p-vector. In
the classical setup where eigenvalues λ1 > λ2 ≥ . . . ≥ λp are fixed,
the Anderson (1963) likelihood ratio test (LRT) and the Hallin, Paindaveine and Verdebout (2010) Le Cam optimal test for this problem
are asymptotically equivalent under the null, hence also under sequences of contiguous alternatives. We show that this equivalence
does not survive asymptotic scenarios where λn1 − λn2 = o(rn )
√
with rn = O(1/ n). For such scenarios, the Le Cam optimal test
still asymptotically meets the nominal level constraint, whereas the
LRT becomes extremely liberal. Consequently, the former test should
be favored over the latter one whenever the two largest sample eigenvalues are close to each other. By relying on the Le Cam theory
of asymptotic experiments, we study in the aforementioned asymptotic scenarios the non-null and optimality properties of the Le Cam
optimal test and show that the null robustness of this test is not obtained at the expense of efficiency. Our asymptotic investigation is
extensive in the sense that it allows rn to converge to zero at an arbitrary rate. To make our results as striking as possible, we not only
restrict to the multinormal case but also to single-spiked spectra of
the form λn1 > λn2 = . . . = λnp .
1. Introduction. Principal Component Analysis (PCA) is one of the
most classical tools in dimension reduction. For a random p-vector X with
mean zeroPand a covariance matrix Σ admitting the spectral decomposition Σ = pj=1 λj θ j θ 0j (λ1 ≥ . . . ≥ λp ), the jth principal component is θ 0j X,
that is, the projection of X onto the jth unit eigenvector θ j of Σ . In practice, Σ is usually unknown, so that one of the key issues in PCA is to perform
∗
Research is supported by the IAP research network grant nr. P7/06 of the Belgian
government (Belgian Science Policy), the Crédit de Recherche J.0113.16 of the FNRS
(Fonds National pour la Recherche Scientifique), Communauté Française de Belgique,
and a grant from the National Bank of Belgium.
MSC 2010 subject classifications: Primary 62F05, 62H25; secondary 62E20
Keywords and phrases: Le Cam’s theory of asymptotic experiments, Local asymptotic
normality, Principal component analysis, Spiked covariance matrices, Weak identifiability
1
2
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
inference on eigenvectors. The seminal paper Anderson (1963) focused on
the multinormal case and derived asymptotic results for the maximum likelihood estimators of the θ j ’s and λj ’s. Later, Tyler (1981, 1983a) extended
those results to the elliptical case, where, to avoid moment assumptions, Σ is
then the corresponding “scatter” matrix rather than the covariance matrix.
Still under ellipticity assumptions, Hallin, Paindaveine and Verdebout (2010)
obtained Le Cam optimal tests on eigenvectors and eigenvalues, whereas
Hallin, Paindaveine and Verdebout (2014) developed efficient R-estimators
for eigenvectors. Croux and Haesbroeck (2000), Hubert, Rousseeuw and Vanden Branden (2005) and He et al. (2011) proposed various robust methods
for PCA. Recently, Johnstone and Lu (2009), Berthet and Rigollet (2013)
and Han and Liu (2014) considered inference on eigenvectors of Σ in sparse
high-dimensional situations. PCA has also been extensively considered in
the functional case; see, e.g., Boente and Fraiman (2000), Bali et al. (2011)
or the review paper Cuevas (2014).
In this work, we focus on the problem of testing, on the basis of a
random sample X1 , . . . , Xn from the p-variate normal distribution with
mean µ and covariance matrix Σ , the null hypothesis H0 : θ 1 = θ 01 against
the alternative H1 : θ 1 6= θ 01 , where θ 01 is a fixed unit vector. Denoting
as λ̂1 ≥ λ̂2 P
≥ . . . ≥ λ̂p the eigenvalues of the sample covariance
maP
trix S := n1 ni=1 (Xi − X̄)(Xi − X̄)0 (as usual, X̄ := n1 ni=1 Xi here),
the classical test for this problem is the Anderson (1963) likelihood ratio
test, φA say, rejecting the null at asymptotic level α when
0
−1 00
2
−1 0
QA := n λ̂1θ 00
1 S θ 1 + λ̂1 θ 1 S θ 1 − 2 > χp−1,1−α ,
where χ2`,1−α stands for the α-upper quantile of the chi-square distribution with ` degrees of freedom. Hallin, Paindaveine and Verdebout (2010)
obtained the Le Cam optimal for the same problem. This test, φHPV say,
rejects the null at asymptotic level α when
QHPV :=
p
n X
λ̂1
0
λ̂−1
θ̃θ j Sθθ 01
j
2
> χ2p−1,1−α ,
j=2
where θ̃θ j , j = 2, . . . , p, defined recursively through
(1.1)
θ̃θ j :=
(Ip − θ 01θ 00
1 −
k(Ip − θ 01θ 00
1 −
0
θ kθ̃θ k )θ̂θ j
k=2 θ̃
Pj−1
0
θ kθ̃θ k )θ̂θ j k
k=2 θ̃
Pj−1
(with summation over empty collection of indices being equal to zero), result
from a Gram-Schmidt orthogonalization of θ 01 , θ̂θ 2 , . . . , θ̂θ p , where θ̂θ j is a unit
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
3
eigenvector of S associated with the eigenvalue λ̂j , j = 2, . . . , p. When the
eigenvalues of Σ are fixed and satisfy λ1 > λ2 ≥ λ3 ≥ . . . ≥ λp (the minimal
condition under which θ 1 is identified — up to an unimportant sign, as
usual), both tests above are asymptotically equivalent under the null, hence
also under sequences of contiguous alternatives, which implies that φA is
also Le Cam optimal; see Hallin, Paindaveine and Verdebout (2010). The
tests φA and φHPV can therefore be considered perfectly equivalent, at least
asymptotically so.
In the present paper, we compare the asymptotic behaviors of these tests
in a non-standard asymptotic framework where eigenvalues may depend on n
and where λn1 − λn2 is o(1) as n goes to infinity. Such asymptotic scenarios
provide weak identifiability since the first eigenvector θ 1 is not properly identified in the limit. To make our results as striking as possible, we will stick
to the Gaussian framework considered above and restrict to single-spiked
spectra of the form λn1 > λn2 = . . . = λnp . In other words, we will consider triangular arrays of observations Xni , i = 1, . . . , n, n = 1, 2, . . ., where
Xn1 , . . . , Xnn form a random sample from the p-variate normal distribution
with mean µ n and covariance matrix
(1.2)
Σ n := σn2 (Ip + rn v θ 1θ 01 )
= σn2 (1 + rn v)θθ 1θ 01 + σn2 (Ip − θ 1θ 01 ),
where v is a positive real number, (σn2 ) is a positive real sequence, (rn )
is a bounded nonnegative real sequence, and I` denotes the `-dimensional
identity matrix. The eigenvalues of the covariance matrix Σ n are then λn1 =
σn2 (1 + rn v) (with corresponding eigenvector θ 1 ) and λn2 = . . . = λnp =
σn2 (with corresponding eigenspace being the orthogonal complement of θ 1
in Rp ). If rn ≡ 1 or if rn converges to 1 as n → ∞, then this setup is
similar to the classical one where the first eigenvector θ 1 remains identified
in the limit. In contrast, if rn = o(1), then the resulting weak identifiability
(n)
(n)
intuitively makes the problem of testing H0 : θ 1 = θ 01 against H1 : θ 1 6= θ 01
increasingly hard as n goes to infinity.
Our results show that, while they are, as mentioned above, equivalent in
the standard asymptotic scenario associated with rn ≡ 1, the tests φHPV
and φA actually exhibit very different behaviors under weak identifiability.
More precisely, we show that this asymptotic equivalence survives scenarios
√
√
where rn = o(1) with nrn → ∞, but not scenarios where rn = O(1/ n).
Irrespective of the asymptotic scenario considered, the test φHPV asymptotically meets the nominal level constraint, hence may be considered robust to
√
weak identifiability. On the contrary, in scenarios where rn = O(1/ n), the
test φA becomes dramatically liberal. Consequently, despite the asymptotic
4
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
equivalence of these tests in standard asymptotic scenarios, the test φHPV
should be favored over φA .
Of course, this nice robustness property of φHPV only refers to the null
asymptotic behavior of this test, and it is of interest to investigate whether
or not this null robustness is obtained at the expense of efficiency. In order
to do so, we study the non-null and optimality properties of φHPV under
suitable local alternatives. This is done by exploiting the Le Cam theory of
asymptotic experiments. In every asymptotic scenario considered, we show
that the corresponding sequence of experiments converges to a limiting experiment in the Le Cam sense. Interestingly, (i) the corresponding contiguity
rate crucially depends on the underlying scenario and (ii) the resulting limiting experiment is not always a Gaussian shift experiment, such as in the
standard local asymptotic normality (LAN) setup. In all cases, however, we
can derive the asymptotic non-null distribution of QHPV under contiguous
alternatives by resorting to the Le Cam third lemma, and we can establish
that this test enjoys excellent optimality properties.
The problem we consider in this paper is characterized by the fact that
the parameter of interest (here, the first eigenvector) is unidentified when a
nuisance parameter (here, the difference between both largest eigenvalues)
is equal to some given value. Such situations have been much considered in
the statistics and econometrics literatures; we refer, e.g., to Dufour (1997),
Pötscher (2002), Forchini and Hillier (2003), Dufour (2006), or Paindaveine
and Verdebout (2017). To the best of our knowledge, however, no relevant
results have been obtained in PCA. We think that, far from being of academic interest only, our results should also be crucial for practitioners: they
indeed provide a clear warning that, when the underlying distribution is
close to spherical (more generally, when both largest sample eigenvalues are
nearly equal), the daily-practice Gaussian test φA is extremely liberal, hence
may lead to wrong conclusions (false positives) with very high probability,
whereas the test φHPV remains a reliable procedure in such cases.
The paper is organized as follows. In Section 2, we introduce the distributional setup and notation to be used throughout and we derive preliminary
results on the asymptotic behavior of sample eigenvalues/eigenvectors. In
Section 3, we show that the null asymptotic distribution of QHPV is χ2p−1
under all asymptotic scenarios, whereas that of QA is guaranteed to be χ2p−1
√
only if nrn → ∞. We also explicitly provide a case where the null asymptotic distribution of QA indeed fails to be χ2p−1 . In Section 4, we show that,
in all asymptotic scenarios, the sequence of experiments considered converges to a limiting experiment. Then, this is used to study the non-null
and optimality properties of φHPV . Both in Sections 3 and 4, our results
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
5
are illustrated through Monte Carlo exercises. We wrap up in Section 5 and
present all proofs in an appendix.
2. Preliminary results. As mentioned above, we will consider throughout triangular arrays of observations Xni , i = 1, . . . , n, n = 1, 2, . . ., where
Xn1 , . . . , Xnn form a random sample from the p-variate normal distribution
with mean µ n and covariance matrix Σ n = σn2 (Ip +rn v θ 1θ 01 ), where σn2 , v > 0,
rn ≥ 0 and θ 1 is a unit p-vector. The resulting hypothesis will be denoted
(n)
as Pµ n ,σn2 ,θθ 1 ,rn ,v = Pµ ,σ2 ,θθ 1 ,rn ,v (the superscript (n) will be dropped in the sen n
P
P
quel). Throughout, X̄n := n1 ni=1 Xni and Sn := n1 ni=1 (Xni − X̄n )(Xni −
X̄n )0 will denote the sample average and sample covariance matrix based
on Xn1 , . . . , Xnn . For any j = 1, . . . , p, the jth largest eigenvalue of Sn and
the corresponding unit eigenvector will be denoted as λ̂nj and θ̂θ nj , respec(n)
(n)
tively. With this notation, the tests φA = φA and φHPV = φHPV from the
introduction reject the null at asymptotic level α when
(n)
0
−1 00
−1 0
QA = QA = n λ̂n1θ 00
1 Sn θ 1 + λ̂n1 θ 1 Snθ 1 − 2
(2.1)
=
p
n X
λ̂n1
0
2
θ nj θ 01
λ̂−1
nj (λ̂n1 − λ̂nj ) θ̂
2
> χ2p−1,1−α
j=2
and
(2.2)
QHPV =
(n)
QHPV
=
p
n X
λ̂n1
0
θ nj Snθ 01
λ̂−1
nj θ̃
2
> χ2p−1,1−α ,
j=2
respectively, where θ̃θ n2 , . . . , θ̃θ np result from the Gram-Schmidt orthogonalization in (1) applied to θ 01 , θ̂θ n2 , . . . , θ̂θ np . Under Pµ n ,σn2 ,θθ 1 ,rn ,v , the sample
eigenvalue λ̂nj is uniquely defined with probability one, but θ̂θ nj is, still with
probability one, defined up to a sign only. Clearly, this sign does not play
any role in (2.1)-(2.2), hence will be fixed arbitrarily. At a few places below,
this sign, however, will need to be fixed in an appropriate way.
For obvious reasons, the asymptotic behavior of Sn will play a crucial
role when investigating the asymptotic properties of the tests above. To
describe this behavior, we need to introduce the following notation. For
an ` × ` matrix A, denote as vec(A) the vector obtained by stacking the
columns of A on top of each other. We will let A⊗2 := A ⊗ A, where
A ⊗ B is the Kronecker product of A and B. The commutation matrix Kk,` ,
that is such that Kk,` (vec A) = vec(A0 ) for any k × ` matrix A, satisfies
Kp,k (A ⊗ B) = (B ⊗ A)Kq,` for any k × ` matrix A and p × q matrix B; see,
e.g., Magnus and Neudecker (2007). If X is p-variate standard normal, then
6
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
the covariance matrix of vec(XX0 ) is Ip2 + Kp , with Kp := Kp,p ; the LevyLindeberg central limit theorem then easily provides the following result.
Lemma 2.1. Fix a unit p-vector θ 1 , v > 0 and a bounded nonnegative
√
Σ−1/2
real sequence (rn ). Then, under P0,1,θθ 1 ,rn ,v , n(Σ
)⊗2 vec(Sn − Σ n ) is
n
asymptotically normal with mean zero and covariance matrix Ip2 + Kp . In
√
particular, (i) if rn ≡ 1, then n vec(Sn − Σ n ) is asymptotically normal
Σ(v))⊗2 , with Σ (v) :=
with mean zero and covariance matrix (Ip2 + Kp )(Σ
√
0
Σn ) is asymptotically normal
Ip + v θ 1θ 1 ; (ii) if rn is o(1), then n vec(Sn −Σ
with mean zero and covariance matrix Ip2 + Kp .
Clearly, the tests φA and φHPV above are invariant under translations
and scale transformations, that is, respectively, under transformations of
the form (Xn1 , . . . , Xnn ) 7→ (Xn1 + t, . . . , Xnn + t), with t ∈ Rp , and
(Xn1 , . . . , Xnn ) 7→ (sXn1 , . . . , sXnn ), with s > 0. This implies that, when
investigating the behavior of these tests, we may assume without loss of generality that µ n ≡ 0 and σn2 ≡ 1, that is, we may restrict to hypotheses of the
form Pθ 1 ,rn ,v := P0,1,θθ 1 ,rn ,v , as we already did in Lemma 2.1. We therefore
restrict to such hypotheses in the rest of the paper.
The tests φA and φHPV are based on statistics that not only involve the
sample covariance matrix Sn , but also the sample eigenvalues and eigenvectors. It is therefore no surprise that investigating the asymptotic behavior of
these tests under weak identifiability will require controlling the corresponding asymptotic behaviors of sample eigenvalues and eigenvectors. For eigenvalues, we have the following result (throughout, diag(A1 , . . . , Am ) stands
for the block-diagonal matrix with diagonal blocks A1 , . . . , Am ).
Lemma 2.2. Fix a unit p-vector θ 1 , v > 0 and a bounded nonnegative
real sequence (rn ). Let Z(v) be a p × p random matrix such that vec(Z(v)) ∼
Λ(v))⊗2 ), with Λ (v) := diag(1 + v, 1, . . . , 1), and let Z22 (v)
N (0, (Ip2 + Kp )(Λ
be the matrix obtained from Z(v) by deleting its first row and first column.
Write Z := Z(0) and Z22 := Z22 (0). Then, under Pθ 1 ,rn ,v ,
`n :=
√
n(λ̂n1 −(1+rn v)),
√
0 D
√
n(λ̂n2 −1), . . . , n(λ̂np −1) → ` = (`1 , . . . , `p )0 ,
D
where → denotes weak convergence and where ` is as follows:
(i) if rn ≡ 1, then `1 and (`2 , . . . , `p )0 are mutually independent, `1 is
normal with mean zero and variance 2(1 + v)2 , and `2 ≥ . . . ≥ `p are
the eigenvalues of Z22 (v);
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
7
√
(ii) if rn is o(1) with nrn → ∞, then `1 and (`2 , . . . , `p )0 are mutually
independent, `1 is normal with mean zero and variance 2, and `2 ≥
. . . ≥ `p are the eigenvalues of Z22 ;
√
(iii) if rn = 1/ n, then `1 is the largest eigenvalue of Z − diag(0, v, . . . , v)
and `2 ≥ . . . ≥ `p are the p−1 smallest eigenvalues of Z+diag(v, 0, . . . , 0);
√
(iv) If rn = o(1/ n), then ` is the vector of eigenvalues of Z (in decreasing
order), hence has density
(2.3)
Y
p
1X 2
0
(`1 , . . . , `p ) 7→ cp exp −
`j
(`k − `j ) I[`1 ≥ . . . ≥ `p ]
4
j=1
1≤k<j≤p
where cp is a normalizing constant and where I[A] is the indicator
function of A.
Lemma 2.2 shows that, unlike the sample covariance matrix Sn , sample
eigenvalues exhibit an asymptotic behavior that crucially depends on (rn ).
√
The important threshold, associated with rn = 1/ n, provides sequences of
hypotheses Pθ 1 ,rn ,v that are contiguous to the spherical hypotheses Pθ 1 ,0,v
under which the first eigenvector θ 1 is unidentified (contiguity follows, e.g.,
from Proposition 2.1 in Hallin and Paindaveine, 2006). Lemma 2.2 then
identifies four regimes that will be present throughout our double asymptotic
investigation below, namely away from contiguity (rn ≡ 1, case (i)), above
√
contiguity (rn is o(1) with nrn → ∞, case (ii)), under contiguity (rn =
√
√
1/ n, case (iii)), and under strict contiguity (rn = o(1/ n), case (iv)).
While Lemmas 2.1-2.2 will be sufficient to study the asymptotic behavior
of φHPV , the test φA , as hinted by the expression in (2.1), will further require
0
0
investigating the joint asymptotic behavior of θ̂θ n2θ 01 , . . . , θ̂θ npθ 01 . To do so, fix
arbitrary p-vectors θ 2 , . . . , θ p such that Γ := (θθ 01 , θ 2 , . . . , θ p ) is orthogonal.
Γn := (θ̂θ n1 , . . . , θ̂θ np ), where the “signs” of θ̂θ nj , j = 1, . . . , p, are
Let further Γ̂
fixed by the constraint that, with probability one,
0
En,11 En,12
Γ nΓ =
(2.4)
En := Γ̂
En,21 En,22
has positive diagonal elements only (note that the diagonal entries of En are
non-zero with probability one). With this notation, En,21 collects the ran0
0
dom variables θ̂θ n2θ 01 , . . . , θ̂θ npθ 01 of interest above. We then have the following
result.
8
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
Lemma 2.3. Fix a unit p-vector θ 1 , v > 0 and a bounded nonnegative real
sequence (rn ). Let Z be a p×p random matrix such that vec(Z) ∼ N (0, Ip2 +
Kp ). For any v > 0, let E(v) := (w1 (v), . . . , wp (v))0 , where wj (v) is the unit
eigenvector associated with the jth largest eigenvalue of Z + diag(v, 0, . . . , 0)
and such that (w1 (v))j > 0 almost surely. Then, we have the following
under Pθ 1 ,rn ,v :
√
(i) if rn ≡ 1, then En,11 = 1+oP (1), En,22 E0n,22 = Ip−1 +oP (1), nEn,21 =
√ 0
√ 0
OP (1), and both nEn,22 En,21 and nEn,12 are asymptotically normal
with mean zero and covariance matrix v −2 (1 + v)Ip−1 ;
√
(ii) if rn is o(1) with nrn → ∞, then En,11 = 1 + oP (1), En,22 E0n,22 =
√
√
Ip−1 + oP (1), nrn En,21 = OP (1), and both nrn E0n,22 En,21 and
√
nrn E0n,12 are asymptotically normal with mean zero and covariance
matrix v −2 Ip−1 ;
√
(iii) if rn = 1/ n, then En converges weakly to E(v);
√
(iv) if rn = o(1/ n), then En converges weakly to E(0).
This result shows that the asymptotic behavior of En,21 , which, as mentioned above, is the only part of En involved in the Anderson test statistic QA , depends on the regimes identified in Lemma 2.2. (i) Away from
contiguity, En,21 converges to the zero vector in probability at the standard
root-n rate. (ii) Above contiguity, En,21 is still oP (1), but the rate of convergence is slower. (iii) Under contiguity, consistency is lost and En,21 converges
weakly to a distribution that still depends on v. (iv) Under strict contiguity,
on the contrary, the asymptotic distribution of En does not depend on v
and is the same as for v = 0. As a direct corollary, the asymptotic distribution of En is then the same as in the spherical Gaussian case, so that, up
to the fact that E has almost surely positive diagonal entries (a constraint
inherited from the corresponding one on En ), this asymptotic distribution
is the invariant Haar distribution on the group of p × p orthogonal matrices;
see Anderson (1963), page 126.
3. Null results. In this section, we study the null asymptotic behaviors
of φA and φHPV under weak identifiability, that is, we do so under the
sequences of (null) hypotheses Pθ 01 ,rn ,v introduced in the previous section.
This will obviously heavily rely on Lemmas 2.2-2.3. Regarding φHPV , we
have the following result.
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
9
Theorem 3.1. Fix a unit p-vector θ 01 , v > 0 and a bounded nonnegative
real sequence (rn ). Then, under Pθ 01 ,rn ,v ,
D
QHPV → χ2p−1 ,
so that, in all regimes (i)-(iv) from the previous section, the test φHPV has
asymptotic size α under the null.
This result shows that the test φHPV is validity-robust to weak identifiability in the sense that it will asymptotically meet the nominal level constraint
in scenarios that are arbitrarily close to the spherical case. The following
result shows that, if one restricts to regimes (i)-(ii), then this extends to
the Anderson test φA .
Theorem 3.2. Fix a unit p-vector θ 01 , v > 0 and a nonnegative real
√
sequence (rn ) satisfying (i) rn ≡ 1 or (ii) rn = o(1) with nrn → ∞. Then,
under Pθ 01 ,rn ,v ,
D
QA → χ2p−1 ,
so that, in regimes (i)-(ii), the test φA has asymptotic size α under the null.
Note that (2.1) allows to write
QA =
p
X
√
j=2
(3.1)
=:
p
X
n(λ̂n1 − (1 + rn v)) −
q
λ̂n1 λ̂nj
√
n(λ̂nj − 1)
0
θ̂θ nj θ 01
!2
√
nrn v 0 0
+q
θ̂θ nj θ 1
λ̂n1 λ̂nj
2
Rnj + Snj .
j=2
In regimes (i)-(ii), Parts (i)-(ii) of Lemmas 2.2-2.3 imply P
that, for any j =
2 + o (1)
2, . . . , p, Rnj = oP (1) and Snj = OP (1), so that QA = pj=2 Snj
P
in these regimes. In regimes (iii)-(iv), on the contrary, the Rnj ’s will play
a role in the null asymptotic behavior of the Anderson test statistic QA ,
which suggests that the corresponding asymptotic distribution may differ
from the χ2p−1 one obtained in regimes (i)-(ii). To explore this and to illustrate our theoretical results above, we performed the following Monte
Carlo exercise. For any ` = 0, 1, . . . , 5, we generated M = 10, 000 mutually
(`)
independent random samples Xi , i = 1, . . . , n, from the (p = 10)-variate
normal distribution with mean zero and covariance matrix
(3.2)
−`/6 0 00
Σ (`)
θ 1θ 1 ,
n := Ip + n
10
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
where θ 01 is the first vector of the canonical basis of Rp . For each sample, we
(n)
performed the tests φHPV and φA for H0 : θ 1 = θ 01 at nominal level 5%.
The present simulation therefore focuses on the null behavior of these tests.
The value of ` allows to consider the various regimes above, namely (i) away
from contiguity (` = 0), (ii) beyond contiguity (` = 1, 2), (iii) under contiguity (` = 3), and (iv) under strict contiguity (` = 4, 5). Increasing values
of ` therefore provide harder and harder inference problems. In order to investigate the finite-sample relevance of our asymptotic results and to check
their correctness, this exercise was performed for n = 200 and n = 500, 000,
respectively. Figure 1, that reports the resulting empirical rejection frequencies, clearly support Theorems 3.1-3.2 in the sense that φHPV shows the
target Type 1 risk in all regimes and that φA does so (asymptotically) in
regimes (i)-(ii). More importantly, these results reveal that φA is dramatically liberal in regimes (iii)-(iv), even asymptotically. We conclude that
only φHPV is validity-robust to weak identifiability.
(i)
ℓ=0
(ii)
ℓ=1
n=500,000
(iii)
ℓ=2
ℓ=3
(iv)
ℓ=4
0.00 0.15 0.30 0.45 0.60 0.75 0.90
0.00 0.15 0.30 0.45 0.60 0.75 0.90
Null rejection frequencies
n=200
ℓ=5
(i)
(ii)
(iii)
(iv)
HPV
A
ℓ=0
ℓ=1
ℓ=2
ℓ=3
ℓ=4
ℓ=5
Fig 1. Empirical rejection frequencies, under the null hypothesis, of the tests φHPV and φA
performed at nominal level 5%. Results are based on M = 10, 000 independent tendimensional Gaussian random samples of size n = 200 and size n = 500, 000. Increasing
values of ` bring the underlying spiked covariance matrix closer and closer to a multiple
of the identity matrix; see Section 3 for details. The links between the values of ` and the
asymptotic regimes (i)-(iv) from Section 2 are provided in each barplot.
The previous simulation showed empirically that, in regimes (iii)-(iv), the
null asymptotic distribution of φA is not χ2p−1 . From an academic point of
view, of course, it would be more satisfactory to derive theoretically the corresponding asymptotic distributions. Unfortunately, the complicated asymptotic behavior of sample eigenvalues and eigenvectors in Lemmas 2.2-2.3
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
11
makes this extremely challenging. For regime (iv), though, we have the following result in the bivariate case p = 2.
Theorem 3.3. Fix p = 2, a unit p-vector θ 01 , v > 0 and a nonnegative
√
real sequence (rn ) such that nrn → 0. Then, under Pθ 01 ,rn ,v ,
D
QA → 4χ21 ,
so that, irrespective of α ∈ (0, 1), the test φA has an asymptotic size under
the null that is strictly larger than α.
This result shows in a striking way the impact weak identifiability may
have, in the bivariate case, on the null asymptotic distribution of the Anderson test statistic QA : away from contiguity and beyond contiguity (regimes
(i)-(ii)), QA is asymptotically χ21 under the null (Theorem 3.2), whereas under strict contiguity (regime (iv)), this statistic is asymptotically 4χ21 under
the null. The result also allows to quantify, for any nominal level α, how
liberal the bivariate Anderson test will asymptotically be in regime (iv).
More precisely, the Type 1 risk, in this regime, of the level-α bivariate Anderson test converges to P[4Z 2 > χ21,1−α ], where Z is standard normal.
For α = 0.1%, 1% and 5%, this provides in regime (iv) an asymptotic
Type 1 risk of about 10%, 19.8% and 32.7%, respectively (which exceeds
the nominal level by about a factor 100, 20 and 6.5, respectively!)
We conducted another simulation to illustrate Theorem 3.3 and its consequences. To do so, we generated, for any ` = 0, 1, . . . , 5, a collection
of M = 10, 000 mutually independent random samples of size n = 500, 000
from the bivariate normal distribution with mean zero and covariance ma−`/6θ 0θ 00 , with θ 0 = (1, 0)0 . This is therefore essentially
trix Σ (`)
n := I2 + n
1 1
1
the bivariate version of the ten-dimensional numerical exercise above. For
each value of `, Figure 2 provides histograms of the resulting M values of
the Anderson test statistic QA , along with plots of the densities of the χ21
and 4χ21 distributions, that is, densities of the null asymptotic distribution
of QA in regimes (i)-(ii) and in regime (iv), respectively. In these three
regimes, the histograms are perfectly fitted by the corresponding density.
The figure also provides the empirical Type 1 risks of the level-α Anderson test for α = 0.1%, 1% and 5%. Clearly, these Type 1 risks are close to
the theoretical asymptotic ones both in regimes (i)-(ii) (namely, α) and in
regime (iv) (namely, the Type 1 risks provided at the end of the previous
paragraph).
12
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
0.04 (α=0.1%)
0.89 (α=1%)
1.0
1.12 (α=1%)
4.72 (α=5%)
0.0
0.0
0.5
5.03 (α=5%)
0.5
1.0
2.0
0.12 (α=0.1%)
χ21 (pdf)
4χ21 (pdf)
ℓ=1 (ii)
1.5
1.5
2.0
ℓ=0 (i)
0.0
0.5
1.0
1.5
2.0
2.5
0.5
1.0
1.5
2.0
0.23 (α=0.1%)
2.5
4.6 (α=0.1%)
10.86 (α=1%)
1.0
1.32 (α=1%)
1.0
2.0
ℓ=3 (iii)
1.5
1.5
2.0
ℓ=2 (ii)
0.0
0.5
21.27 (α=5%)
0.0
0.0
0.5
5.71 (α=5%)
0.0
0.5
1.0
1.5
2.0
2.5
0.5
1.0
1.5
2.0
9.3 (α=0.1%)
2.5
9.83 (α=0.1%)
19.18 (α=1%)
1.0
17.97 (α=1%)
1.0
2.0
ℓ=5 (iv)
1.5
1.5
2.0
ℓ=4 (iv)
0.0
0.5
32.38 (α=5%)
0.0
0.0
0.5
30.8 (α=5%)
0.0
0.5
1.0
1.5
2.0
2.5
0.0
0.5
1.0
1.5
2.0
2.5
Fig 2. For each ` = 0, 1, . . . , 5, histograms of values of QA from M = 10, 000 independent
(null) Gaussian random samples of size n = 500, 000 and dimension p = 2. Increasing
values of ` bring the underlying spiked covariance matrix closer and closer to a multiple
of the identity matrix; see Section 3 for details. The links between the values of ` and the
asymptotic regimes (i)-(iv) from Section 2 are provided in each case. For any value of `, the
density of the null asymptotic distribution of QA in regimes (i)-(ii) (resp., in regime (iv))
is plotted in orange (resp., in blue) and the empirical Type 1 risk of the level-α Anderson
test is provided for α = 0.1%, 1% and 5%.
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
13
4. Non-null and optimality results. The previous section shows
that, unlike φA , the test φHPV is validity-robust to weak identifiability.
However, the trivial level-α test, that randomly rejects the null with probability α, of course enjoys the same robustness property. This motivates
investigating whether or not the validity-robustness of φHPV is obtained at
the expense of efficiency. In this section, we therefore study the asymptotic
non-null behavior of this test and show that it actually still enjoys strong
optimality properties under weak identifiability.
Throughout, optimality is to be understood in the Le Cam sense. In the
present hypothesis testing context, Le Cam optimality requires studying
local log-likelihood ratios of the form
(4.1)
Λn := log
dPθ 01 +νnτ n ,rn ,v
dPθ 01 ,rn ,v
,
where the bounded sequence (ττ n ) in Rp and the positive real sequence (νn )
are such that, for any n, θ 01 + νnτ n is a unit p-vector, hence, is an admissible
perturbation of θ 01 . This imposes that (ττ n ) is bounded and satisfies
θ 00
1 τn = −
(4.2)
νn
kττ n k2
2
for any n. The following result then describes, in each of the four regimes (i)(iv) considered in the previous sections, the asymptotic behavior of the loglikelihood ratio Λn .
Theorem 4.1. Fix a unit p-vector θ 01 , v > 0 and a bounded nonnegative
real sequence (rn ). Then, we have the following under Pθ 01 ,rn ,v :
√
(i) if rn ≡ 1, then, with νn = 1/ n,
∆n =
v √
v2
θ 01 and Γ =
n(Ip − θ 01θ 00
(Ip − θ 01θ 00
1 )(Sn − Σ n )θ
1 ),
1+v
1+v
we have that Λn = τ 0n ∆ n − 12 τ 0nΓτ n + oP (1) and that ∆ n is asymptotically normal with mean zero and covariance matrix Γ ;
√
√
(ii) if rn is o(1) with nrn → ∞, then, with νn = 1/( nrn ),
√
θ 01 and Γ = v 2 (Ip − θ 01θ 00
∆ n = v n(Ip − θ 01θ 00
1 )(Sn − Σ n )θ
1 ),
we similarly have that Λn = τ 0n ∆ n − 21 τ 0nΓτ n + oP (1) and that ∆ n is
still asymptotically normal with mean zero and covariance matrix Γ ;
14
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
√
(iii) if rn = 1/ n, then, letting νn ≡ 1,
h √
1 i
Λn = τ 0n v n(Sn − Σ n ) θ 01 + τ n
(4.3)
2
−
v2
v2
kττ n k2 + kττ n k4 + oP (1),
2
8
√
where, if (ττ n ) → τ , then τ 0n n(Sn − Σ n )(θθ 01 + 12 τ n ) is asymptotically
normal with mean zero and variance kττ k2 − 41 kττ k4 ;
√
(iv) if rn = o(1/ n), then, even with νn ≡ 1, we have that Λn is oP (1).
This result shows that, for any fixed v > 0 and for any fixed sequence rn
associated with regime (i) (away from contiguity) or regime (ii) (beyond
contiguity), the sequence of models {Pθ 1 ,rn ,v : θ 1 ∈ S p−1 } is LAN, with
central sequence
√
nv
θ 01
(4.4)
∆ n,δ :=
Ip − θ 01θ 00
1 (Sn − Σ n )θ
1 + δv
and Fisher information matrix
(4.5)
Γ δ :=
v2
Ip − θ 01θ 00
1 ,
1 + δv
where δ := 1 if regime (i) is considered and δ := 0 otherwise. Denoting as A−
the Moore-Penrose inverse of A, it follows that the locally asymptotically
(n)
(n)
maximin test for H0 : θ 1 = θ 01 against H1 : θ 1 6= θ 01 rejects the null at
asymptotic level α when
n
0
00
0 00
2
(4.6)
Qδ = ∆ 0n,δΓ −
δ ∆ n,δ = 1 + δv θ 1 Sn Ip − θ 1θ 1 Snθ 1 > χp−1,1−α .
In view of (A.11) in the proof of Theorem 3.1, we have that QHPV = Qδ +
oP (1) under Pθ 01 ,rn ,v , for any v > 0 and any bounded sequence (rn ), hence
also, from contiguity, under local alternatives of the form Pθ 01 +ττ n /(√nrn ),rn ,v .
We may therefore conclude that, away from contiguity and beyond contiguity, the test φHPV is Le Cam optimal for the problem at hand. We have the
following result.
Theorem 4.2. Fix a unit p-vector θ 01 , v > 0 and a nonnegative real
√
sequence (rn ) satisfying (i) rn ≡ 1 or (ii) rn = o(1) with nrn → ∞. Then,
(n)
the test φHPV is locally asymptotically maximin at level α when testing H0 :
(n)
θ 1 = θ 01 against H1 : θ 1 6= θ 01 . Moreover, under Pθ 01 +ττ n /(√nrn ),rn ,v , with (ττ n )
→ τ , the statistic QHPV is asymptotically non-central chi-square with p − 1
degrees of freedom and with non-centrality parameter (v 2 /(1 + δv))kττ k2 .
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
15
Under strict contiguity (Theorem 4.1(iv)), no asymptotic level-α test can
show non-trivial asymptotic powers against the most severe alternatives of
the form θ 01 + τ . Therefore, the test φHPV is also optimal in regime (iv),
even though this optimality is degenerate since the trivial level-α test is
also optimal in this regime. We then turn to Theorem 4.1(iii), where the
situation is much less standard, as the sequence of experiments {Pθ 1 ,1/√n,v :
θ 1 ∈ S p−1 } there is not LAN. However, by using the Le Cam first lemma (see,
e.g., Lemma 6.4 in van der Vaart, 1998), Theorem 4.1(iii) readily entails that,
for any v > 0, the hypotheses Pθ 01 ,1/√n,v and Pθ 01 +ττ n ,1/√n,v , with (ττ n ) → τ ,
are mutually contiguous. Consequently, the asymptotic non-null distribution
of the test statistic QHPV under contiguous alternatives may still be obtained
from the Le Cam third lemma. We have the following result.
Theorem 4.3. Fix a unit p-vector θ 01 , v > 0 and a nonnegative real
√
√
sequence (rn ) satisfying (iii) rn = 1/ n or (iv) rn = o(1/ n). Then, in
case (iii), the test statistic QHPV , under Pθ 01 +ττ n ,rn ,v , with (ττ n ) → τ , is
asymptotically non-central chi-square with p − 1 degrees of freedom and with
non-centrality parameter
(4.7)
2
v2
kττ k2 4 − kττ k2 2 − kττ k2 ,
16
so that the test φHPV is rate-consistent. In case (iv), this test is locally
(n)
(n)
asymptotically maximin at level α when testing H0 : θ 1 = θ 01 against H1 :
θ 1 6= θ 01 , but trivially so since its asymptotic power against any sequence of
alternatives Pθ 01 +ττ n ,rn ,v , with (ττ n ) → τ , is then equal to α.
Figure
3 plots the non-centrality parameter in (4.7) as a function of τ ∈
√
[0, 2] (since θ 1 is defined up to a sign, one may restrict to alternatives θ 01 +ττ
in the hemisphere centered at θ 01 ), as well as the asymptotic powers of the
test φHPV in dimensions p = 2, 3. This test
√ shows no power when θ 1 is or0
thogonal to θ 1 (equivalently, when kττ k = 2), hence clearly does not enjoy
global-in-ττ optimality properties in regime (iii). As we now explain, however, φHPV exhibits excellent local-in-ττ optimality properties in this regime.
0 00
In order to see this, note that decomposing Ip into Ip − θ 01θ 00
1 + θ 1θ 1 and
using repeatedly (4.2) allows to rewrite (4.3) as
(4.8)
1
Λn = τ 0n∆ n,0 − τ 0nΓ 0τ n
2
√
√
nv
nv 0
0
2 00
−
kττ n k θ 1 (Sn − Σ n )θθ 1 +
τ (Sn − Σ n )ττ n + oP (1),
2
2 n
16
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
where ∆ n,0 and Γ 0 were defined in (4.4) and (4.5), respectively. For small
perturbations τ n , the righthand side of (4.8), after neglecting the secondorder random terms in τ n , becomes
1
τ 0n∆ n,0 − τ 0nΓ 0τ n + oP (1),
2
so that the sequence of experiments is then LAN again, with central sequence ∆ n,0 and Fisher information matrix Γ 0 . This implies that the test
in (4.6) and (in view of the asymptotic equivalence stated below (4.6)) the
test φHPV are locally(-in-ττ ) asymptotically maximin.
Now, if the objective is to construct a test that will perform well also for
large perturbations τ n , it may be tempting to consider as a test statistic the
˜ n := v √n(Sn −Σ
Σn )θθ 01 .
linear-in-ττ part of the random term in (4.3), namely ∆
˜ n , under P 0 √ , is asymptotically normal with mean zero and
Since ∆
θ 1 ,1/ n,v
Γ := v 2 (Ip + θ 01θ 100 ), the resulting test, φoracle say, rejects the null
covariance Γ̃
at asymptotic level α when
−1
0 00 −1
˜ 0 Γ̃
˜
θ 00
Q̃ := ∆
(Sn − Σ n )θθ 01
n Γ ∆ n = nθ
1 (Sn − Σ n ) Ip + θ 1θ 1
0
= n Snθ 01 − 1 + √vn θ 01 Ip − 12 θ 01θ 00
Snθ 01 − 1 + √vn θ 01 > χ2p,1−α ;
1
(4.9)
the terminology “oracle” stresses that this test requires knowing the true
˜ n , under P 0
√
value of v. The Le Cam third lemma entails that ∆
θ 1 +ττ n ,1/ n,v ,
1
2
2
with (ττ n ) → τ , is asymptotically normal with mean v (1 − 2 kττ k )ττ −
Γ. Therefore, under the same sequence
(v 2 /2)kττ k2θ 01 and covariance matrix Γ̃
of hypotheses, Q̃ is asymptotically non-central chi-square with p degrees of
freedom and with non-centrality parameter
(4.10)
v2
1
kττ k2 4 − kττ k2 4 − 2kττ k2 + kττ k4 .
16
2
Note that the difference between this non-centrality parameter and the one
in (4.7) is O(kττ k4 ) as kττ k goes to zero. Since φHPV is based on a smaller
number of degrees of freedom (p − 1, versus p for the oracle test), it will
therefore exhibit larger asymptotic powers than the oracle test for small
values of τ , which reflects the aforementioned local-in-ττ optimality of φHPV .
Figure 3 also plots the non-centrality parameter in (4.10) as a function
of kττ k, as well as the asymptotic powers of the oracle test in dimensions p =
2, 3. As predicted above, φHPV dominates φoracle for small values of kττ k,
that is, for small perturbations. The opposite happens for large values of
the perturbation and it is seen that φoracle overall is quite efficient. It is
17
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
important to recall, however, that this test cannot be used in practice since
it requires knowing the value of v. The figure further reports the results of
a Monte Carlo simulation we conducted to check correctness of the highly
non-standard asymptotic results obtained in this regime. In this simulation,
(k)
we generated M = 200, 000 mutually independent random samples Xi ,
i = 1, . . . , n = 10, 000, k = 0, 1, . . . , 20, of p-variate (p = 2, 3) Gaussian
random vectors with mean zero and covariance matrix
−1/2 0
Σ (k)
(θθ 1 + τ k )(θθ 01 + τ k )0
n := Ip + n
(4.11)
0.2
0.4
0.09
0.6
0.8
Norm of τ
1.0
1.2
1.4
0.08
p=2
p=3
0.07
p=2
p=3
0.06
Asymptotic powers, rejection frequencies
0.4
0.3
0.2
0.1
0.0
Asymptotic powers
Rejection frequencies
0.05
Oracle
HPV
0.0
Non-centrality parameters
0.5
where θ 01 = (1, 0, . . . , 0)0 ∈ Rp and θ 01 +ττ k = (cos(kπ/40), sin(kπ/40), 0, . . . , 0)0
∈ Rp . In each replication, we performed the tests φHPV and φoracle for
(n)
H0 : θ 1 = θ 01 at nominal level 5%. The value k = 0 is associated with
the null hypothesis, whereas the values k = 1, . . . , 20 provide increasingly
severe alternatives, the most severe of which involves a first eigenvector that
is orthogonal to θ 01 . The resulting rejection frequencies are plotted in the
right panel of Figure 3. Clearly, they are in perfect agreement with the
asymptotic powers, which supports our theoretical results.
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Norm of τ
√
Fig 3. (Left:) Non-centrality parameters (4.7) and (4.10), as a function of kττ k(∈ [0, 2]),
in the asymptotic non-central chi-square distributions of the test statistics of φHPV
and φoracle , respectively, under alternatives of the form Pθ 0 +ττ ,1/√n,1 . (Right:) The corre1
sponding asymptotic power curves in dimensions p = 2 and p = 3, as well as the empirical
power curves resulting from the Monte Carlo exercise described at the end of Section 4.
We conclude this section by stressing that, as announced in the introduction, the contiguity rate in Theorem 4.1 depends on the regime considered.
Clearly, the weaker the identifiability (that is, the closer the underlying distribution to the spherical Gaussian one), the slower the contiguity rate, that
18
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
is, the hardest the inference problem on θ 1 .
5. Wrap up. In this paper, we tackled the problem of testing the
(n)
(n)
null hypothesis H0 : θ 1 = θ 01 against the alternative H1 : θ 1 6= θ 01 ,
where θ 1 is the eigenvector associated with the largest eigenvalue of the
underlying covariance matrix and where θ 01 is some fixed unit vector. We
analyzed the asymptotic behavior of the classical Anderson (1963) test φA
and of the Hallin, Paindaveine and Verdebout (2010) test φHPV under sequences of p-variate Gaussian models with spiked covariance matrices of the
form Σ n = σn2 (Ip + rn vθθ 1θ 01 ), where (σn2 ) is a positive sequence, v > 0 is
fixed, and (rn ) is a nonnegative sequence that converges to zero. We showed
that in these situations where θ 1 is increasingly closer to being unidentified,
φHPV performs better than φA : (i) φHPV , unlike φA , meets asymptotically
the nominal level constraint without any condition on the rate at which rn
converges to zero, and (ii) φHPV remains locally and asymptotically max√
imin in all regimes, but in the contiguity regime rn = 1/ n where φHPV
still enjoys the same optimality property locally in τ . These considerations,
along with the asymptotic equivalence of φHPV and φA in the standard
case rn ≡ 1, clearly imply that the test φHPV , for all practical purposes,
should be favored over φA .
APPENDIX: PROOFS
Throughout the appendix, we will write vn := rn v and δ will take value
one if regime (i) is considered and zero otherwise. All convergences are
as n → ∞.
Proof of Lemma 2.2. Fix arbitrarily θ 2 , . . . , θ p such that the p × p matrix Γ := (θθ 01 , θ 2 , . . . , θ p ) is orthogonal. Letting Λ n := diag(1 + vn , 1, . . . , 1),
we have ΣnΓ = ΓΛn , so that Γ is an eigenvectors matrix for Σn . Clearly,
√
`n1 := n(λ̂n1 − (1 + vn )) is the largest root of the polynomial Pn1 (h) :=
√
√
√
det( n(Sn −(1+vn )Ip )−hIp ) and `n2 := n(λ̂n2 −1), . . . , `np := n(λ̂np −1)
√
are the p − 1 smallest roots of the polynomial Pn2 (h) := det( n(Sn − Ip ) −
hIp ). Letting
√
√
Γ0 SnΓ − Λ n ) = Γ 0 n(Sn − Σ n )Γ
Γ,
(A.1)
Zn := n(Γ
a key ingredient in this proof is to rewrite these polynomials as
√ 0
√
Γ SnΓ − n(1 + vn )Ip − hIp )
Pn1 (h) = det( nΓ
√
Λn − (1 + vn )Ip ) − hIp )
(A.2)
= det(Zn + n(Λ
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
19
and
(A.3)
√ 0
√
Γ SnΓ − nIp − hIp )
Pn2 (h) = det( nΓ
√
Λn − Ip ) − hIp ).
= det(Zn + n(Λ
Note that Lemma 2.1 readily implies that Zn converges weakly to Z(v) in
case (i) and to Z = Z(0) in cases (ii)-(iv), where Z(v) is the random matrix
defined in the statement of the theorem. In all cases, thus, Zn converges
weakly to Z(δv), where δ was introduced at the beginning of this appendix.
We start with the proofs of (i)-(ii). Partition Zn and Z(δv) into
Zn,11 Z0n,21
Z11 (δv) Z021 (δv)
Zn =
and Z(δv) =
,
Z21 (δv) Z22 (δv)
Zn,21 Zn,22
where Zn,11 and Z11 (δv) are random variables, whereas Zn,22 and Z22 (δv)
are (p − 1) × (p − 1) random matrices. It follows from the discussion above
that `n1 is the largest root of the polynomial
Pn1 (h)
Qn1 (h) := √
( nvn )p−1
1
Zn,11 − h
Z0n,21
√
det
= √
Zn,21
Zn,22 − ( nvn + h)Ip−1
( nvn )p−1
√
Zn,11 − h
( nvn )−1 Z0n,21
√
√
= det
.
Zn,21
( nvn )−1 (Zn,22 − ( nvn + h)Ip−1 )
√
In cases (i)-(ii), ( nvn )−1 Zn,21 is oP (1), so that the largest root `n1 converges
weakly to the root of the weak limit of Zn,11 − h, namely to Z11 (δv) (note
√
that, still since ( nvn )−1 Zn,21 is oP (1), the p − 1 smallest roots of Qn1 (h)
converge to −∞ in probability). Similarly, `n2 , . . . , `np are the p − 1 smallest
roots (in decreasing order) of the polynomial
Pn2 (h)
Qn2 (h) := √
nvn
√
1
Zn,11 + nvn − h
Z0n,21
= √
det
Zn,21
Zn,22 − hIp−1
nvn
√
√
( nvn )−1 (Zn,11 + nvn − h)
Z0n,12
√
= det
.
( nvn )−1 Zn,21
Zn,22 − hIp−1
√
Using again the fact that ( nvn )−1 Zn,21 is oP (1) yields that (`n2 , . . . , `np )0
converges weakly to the vector of roots (still in decreasing order) of the weak
20
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
limit of the polynomial det(Zn,22 (δv) − hIp−1 ) = 0, namely to the vector
of (ordered) eigenvalues of Z22 (δv); note that the largest root of Qn2 (h)
converges to ∞ in probability. The result then follows from the fact that,
in both cases (i)-(ii), Z11 (δv) ∼ N (0, 2(1 + δv 2 )) and Z22 (δv) are mutually
independent.
We turn to the proof of (iii). In this case, we have Pn1 (h) = det(Zn −
diag(0, v, . . . , v) − hIp ) and Pn2 (h) = det(Zn + diag(v, 0, . . . , 0) − hIp ). It
readily follows that `n1 converges weakly to the largest root of the polynomial det(Z − diag(0, v, . . . , v) − hIp ), that is, to the largest eigenvalue
of Z − diag(0, v, . . . , v), while (`n2 , . . . , `np )0 converges weakly to the vector
of the p − 1 smallest roots (in decreasing order) of the polynomial det(Z +
diag(v, 0, . . . , 0) − hIp ), namely to the vector of the p − 1 smallest (ordered)
eigenvalues of Z + diag(v, 0, . . . , 0).
Finally, we prove the result in (iv). In that case, we have
√
Pn1 (h) = det Zn − nvn diag(0, 1, . . . , 1) − hIp
and
Pn2 (h) = det Zn +
√
nvn diag(1, 0, . . . , 0) − hIp ,
√
where nvn is o(1). Parallel as above, it follows that `n1 (resp., (`n2 , . . . , `np )0 )
converges weakly to the largest root (resp., the p − 1 smallest roots, in
decreasing order) of the polynomial det(Z − hIp ). In other words, ` =
(`1 , . . . , `p )0 is equal in distribution to the vector collecting the eigenvalues (in decreasing order) of Z. It only remains to establish that the density
of ` is the one given in (2.3). To do so, let Dp be the p-dimensional duplication matrix, that is such that Dp (vech A) = vec A for any p × p symmetric matrix A, where vech A is the p(p + 1)/2-vector obtained from vec A
by restricting to the entries from the upper-triangular part of A. Con−1 0
0
−
sider then W := vech Z = D−
p (vec Z), where Dp = (Dp Dp ) Dp is the
Moore-Penrose inverse of Dp ; see page 57 in Magnus and Neudecker (2007).
Lemma 2.1 implies that W has density
1 p(p+1)/2
1
− 0 −1
w 7→ g(w) =
exp − w0 (D−
(I
+
K
)(D
)
)
w
2
p
p p
p
2π
2
1
1 p(p+1)/2
=
exp − w0 D0p Dp w ,
2π
4
where we used the identity Kp Dp = Dp ; see again page 57 in Magnus and
Neudecker (2007). The resulting density for Z is therefore
z 7→ f (z) = g(vech z) =
1 p(p+1)/2
1
exp − (tr z2 ) .
2π
4
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
21
The explicit density for ` in (2.3) then follows from Lemma 2.3 of Tyler
(1983b).
The proof of Lemma 2.3 requires the following linear algebra result.
Lemma A.1. Let A be a p × p matrix and λ be an eigenvalue of A.
Assume that v = (C11 , . . . , C1p )0 6= 0, where C = (Cij ) stands for the cofactor matrix of A − λIp . Then v is an eigenvector of A associated with
eigenvalue λ.
Proof of Lemma A.1. For any j = 1, . . . , p, denote as (A − λIp )j the jth
row of A − λIp (left as a row vector). Since λ is an eigenvalue of A, we then
have (A − λIp )1 v = det(A − λIp ) = 0. Now, for j = 2, . . . , p,
(A − λIp )j
(A − λIp )2
(A − λIp )j v = det
= 0,
..
.
(A − λIp )p
since this is the determinant of a matrix with (at least) twice the same row.
We conclude that (A − λIp )v = 0. Since v 6= 0, this shows that v is an
eigenvector of A associated with eigenvalue λ.
Proof of Lemma 2.3. First note that since
0
En,11 En,12
Γ nΓ =
En = Γ̂
En,21 En,22
is an orthogonal matrix, we easily obtain that
(A.4)
(A.5)
En,21 = −
1
En,11
En,22 E0n,12 ,
En,22 E0n,22 = Ip−1 − En,21 E0n,21
and
(A.6)
En,11 E0n,12 = −E0n,22 En,21 .
We start with the proof of (i)-(ii). As in the proof of Lemma 2.2, we have
from (A.2) that, still with the random vector Zn defined in (A.1), `n1 =
√
n(λ̂n1 − (1 + vn )) is the largest eigenvalue of
√ 0
√
√
Γ SnΓ − n(1 + vn )Ip = Zn + n(Λ
Λn − (1 + vn )Ip )
nΓ
√
√
(A.7)
= Zn − diag(0, nvn , . . . , nvn ).
22
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
One readily checks that wn1 := Γ 0θ̂θ n1 = (En,11 , En,12 )0 is a corresponding
unit eigenvector. Using the same notation as in the proof of Lemma 2.2,
Lemma A.1 yields that wn1 is proportional to the vector of cofactors associated with the first row of
Zn,11 − `n1
Z0n,21
√
,
(A.8)
Mn,1 :=
Zn,21
Zn,22 − nvn Ip−1 − `n1 Ip−1
or equivalently, of
1
√
Mn,1 =
( nvn )p−1
Z0n,21
Zn,11 − `n1
√ 1 Zn,21
nvn
√ 1 Zn,22
nvn
− Ip−1 −
!
√ 1 `n1 Ip−1
nvn
.
Since `n1 is OP (1) (Lemma 2.2) and so are Zn,21 and Zn,22 , we obtain
that (En,11 , En,12 )0 = (1, 0, . . . , 0)0 + oP (1) (recall that En,11 > 0 almost
√
surely) and that nvn E0n,12 = OP (1). Since En,22 is bounded, it directly fol√
lows from (A.4) that nvn En,21 = OP (1). In view of (A.5), we then obtain
that En,22 E0n,22 − Ip−1 is oP (1). Now, by using (A.6), we have
Zn,21 = `n1 En,11 E0n,12 + E0n,22 diag(`n2 , . . . , `np )En,21
√
√
+ n(1 + vn )En,11 E0n,12 + nE0n,22 En,21
√
(A.9)
= E0n,22 diag(`n2 − `n1 , . . . , `np − `n1 )En,21 − nvn E0n,22 En,21 ,
√
which yields nvn E0n,22 En,21 = −Zn,21 + oP (1). Since Lemma 2.1 entails
that Zn,21 is asymptotically N (0, Ip−1 ) in case (i) and N (0, (1 + v)Ip−1 ) in
√
case (ii), the asymptotic normality result for nrn E0n,22 En,21 follows, which,
√
in view of (A.6), also establishes the one for nrn E0n,12 .
We turn to the proof of (iii)-(iv). As above, wn1 = Γ 0θ̂θ n1 = (En,11 , En,12 )0
√
is the unit eigenvector associated with the largest eigenvalue `n1 = n(λ̂n1 −
(1 + vn )) of (A.7), or equivalently, with the largest eigenvalue `˜n1 = `n1 +
√
√
nvn = n(λ̂n1 − 1) of
√
(A.10)
Zn + diag( nvn , 0, . . . , 0).
Similarly, wnj := Γ 0θ̂θ nj = E0n ej , j = 2, . . . , p, where ej stands for the jth
vector of the canonical basis of Rp , are the unit eigenvectors associated with
√
√
the p − 1 smallest eigenvalues `n2 = n(λ̂n2 − 1), . . . , `np = n(λ̂np − 1)
of (A.10). Consequently, the joint distribution of wnj , j = 1, . . . , p — that
is, the joint distribution of the columns of E0n — converges weakly to the
joint distribution of the unit eigenvectors (associated with eigenvalues in
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
23
decreasing order, and with the signs fixed as in the statement of the theorem)
of
√
Z + lim diag( nvn , 0, . . . , 0)
n→∞
(recall that, in cases (iii)-(iv), Zn converges weakly to the random matrix Z
defined in Lemma 2.2). This establishes the result.
Proof of Theorem 3.1. In this proof, all stochastic convergences are as n →
0
∞ under Pθ 01 ,rn ,v . Since θ̃θ j Σ nθ 01 = 0 for j = 2, . . . , p, the test statistic in (2.2)
rewrites
p
2
n X −1 0
QHPV =
λ̂nj θ̃θ nj Snθ 01 = N0n Nn ,
λ̂n1 j=2
with
0
1 −1/2 0 √
0 √
0
θ
θ
θ̃
n(S
−
Σ
)θ
Nn := p
λ̂n2 θ̃θ n2 n(Sn − Σ n )θθ 01 , . . . , λ̂−1/2
n
n 1 .
np
np
λ̂n1
√
Since the θ̃θ nj ’s are unit vectors and since n(Sn − Σ n ) is OP (1) as soon
0 √
as rn = O(1) (Lemma 2.1), we have that θ̃θ nj n(Sn − Σn )θθ 01 is OP (1) for
any j = 2, . . . , p. Therefore, Lemma 2.2 entails that, still with δ being equal
to one if rn ≡ 1 and to zero if rn = o(1),
Nn = √
0 √
1
0 √
θ̃θ n2 n(Sn − Σ n )θθ 01 , . . . , θ̃θ np n(Sn − Σ n )θθ 01 + oP (1).
1 + δv
Using the fact that
QHPV
0
θ nj θ̃θ nj
j=2 θ̃
Pp
= Ip − θ 01θ 00
1 is idempotent, it follows that
p
n X 00
0
=
θ 1 (Sn − Σ n )θ̃θ nj θ̃θ nj (Sn − Σ n )θθ 01 + oP (1)
1 + δv
j=2
(A.11)
=
n
θ 01 + oP (1)
θ 00 (Sn − Σ n )(Ip − θ 01θ 00
1 )(Sn − Σ n )θ
1 + δv 1
= Vn0 Vn + oP (1),
√
Σn )θθ 01 . Since Lemma 2.1
where we let Vn := (1+δv)−1/2 (Ip −θθ 01θ 00
1 ) n(Sn −Σ
entails that Vn is asymptotically normal with mean zero and covariance
0 00
matrix Ip − θ 01θ 00
1 and since Ip − θ 1θ 1 is idempotent with trace p − 1, the
result then follows from, e.g., Theorem 9.2.1 in Rao and Mitra (1971).
24
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
Proof of Theorem 3.2. In this proof, all stochastic convergences are as n →
∞ under Pθ 01 ,rn ,v . Recall that we showed below (3.1) that, in regimes (i)-(ii),
QA =
p
X
j=2
!2
√
nrn v 0 0
q
θ̂θ nj θ 1 + oP (1).
λ̂n1 λ̂nj
Consequently,
0 √
√
v2
nrn En,21
nrn En,21 + oP (1),
1 + δv
√
where we used the fact that nrn En,21 = OP (1). Therefore, we may write
QA =
QA =
0 √
√
v2
nrn E0n,22 En,21
nrn E0n,22 En,21
1 + δv
+
0
√
√
v2
nrn En,21 (Ip−1 − En,22 E0n,22 ) nrn En,21 + oP (1),
1 + δv
and the result follows from Lemmas 2.3(i)-(ii).
Proof of Theorem 3.3. First note that, for p = 2, (2.1) yields
!2
!2
√
√
n(λ̂n1 − λ̂n2 ) 0 0
`n1 − `n2 + nrn 0 0
p
p
(θ̂θ n2θ 1 )
=
(θ̂θ n2θ 1 ) .
QA =
λ̂n1 λ̂n2
λ̂n1 λ̂n2
In view of Lemmas 2.2-2.3, the continuous mapping theorem therefore implies that
2
D
(A.12)
QA → `1 − `2 (E21 (0))2 ,
where ` = (`1 , `2 )0 and E21 (0) are mutually independent and have the distributions described in Part (iv) of Lemma 2.2 and Lemma 2.3, respectively
(mutual independence follows from Anderson (1963), once it is seen that
0
the asymptotic joint distribution of ` n = (`n1 , `n2 )0 and En,21 = θ̂θ n2θ 01 is
the same in regime (iv) as in the spherical case associated with rn ≡ 0).
2 is the squared of the lower left entry of a random
In particular, Z := E21
matrix E that has an invariant Haar distribution over the collection of 2 × 2
orthogonal matrices; see the remark below Lemma p
2.3. It follows that Z is
1 p−1
Beta( 2 , 2 ), hence has density z 7→ fZ (z) = 1/(π z(1 − z)). As for ` , it
has density
1
2
2
(`1 , `2 )0 7→ f(`1 ,`2 ) (`1 , `2 ) = √ (`1 − `2 )e−(`1 +`2 )/4 I[`1 ≥ `2 ],
4 2π
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
25
√
where I[A]√is the indicator function of A. Letting U1 = (`1 −`2 )/ 2 and U2 =
(`1 + `2 )/ 2, we thus have that (U1 , U2 )0 has density
√
√
(u1 , u2 )0 7→ f(U1 ,U2 ) (u1 , u2 ) = f(`1 ,`2 ) (u1 + u2 )/ 2, (−u1 + u2 )/ 2
1 √
1
2
2
2
2
= √
2u1 e−(u1 +u2 )/4 I[u1 > 0] = √ u1 e−(u1 +u2 )/4 I[u1 > 0].
4
π
4 2π
√
Therefore, U1 = (`1 − `2 )/ 2 has density
Z ∞
1
2
2
√ u1 e−(u1 +u2 )/4 du2
fU1 (u1 ) = I[u1 > 0]
4
π
−∞
Z ∞
1
1
1
2
2
2
√
√ e−u2 /4 du2 = u1 e−u1 /4 I[u1 > 0].
= u1 e−u1 /4 I[u1 > 0]
2
2
2π × 2
−∞
We conclude that Y = (`1 − `2 )2 = 2U12 has density
1
1
√ √
y 7→ fY (y) = √ fU1 ( y/ 2) = e−y/8 I[y > 0].
8
2 2y
Finally, we obtain that the asymptotic distribution of QA in the regime
considered, that coincides with the distribution of Y Z (see (A.12)), has
density
Z 1
Z 1
1
1
e−q/(8z)
q 7→ fQ (q) = I[q > 0]
fY (q/z)fZ (z) dz =
I[q > 0]
dz
3/2 (1 − z)1/2
8π
0 z
0 z
√
r
1
1
4 2π −q/8
q(1 − z) 1
(A.13) =
I[q > 0] = √
e−q/8 I[q > 0],
− √ e
Φ
8π
q
4z
8πq
0
which, as was to be proved, is the pdf of the 4χ21 distribution. To justify the
integral computation in (A.13), note that, denoting respectively as Φ and φ
the cumulative distribution function and probability density function of the
standard normal distribution,
r
r
s
d
q(1 − z)
q(1 − z) 1
4z
−4qz − 4q(1 − z)
Φ
=φ
dz
4z
4z
2 q(1 − z)
16z 2
r
r
r
1
q(1 − z)
q
1
q(1 − z)
q
= − √ exp −
,
=− φ
3
3
4
4z
z (1 − z)
8z
z (1 − z)
4 2π
so that
√
r
s
4 2π
q d
q(1 − z)
q
1
− √ exp −
Φ
= exp −
,
3
q
8 dz
4z
8z
z (1 − z)
26
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
that is,
Z
√
r
4 2π
e−q/(8z)
q
q(1 − z)
dz = − √ exp −
Φ
.
q
8
4z
z 3/2 (1 − z)1/2
Proof of Theorem 4.1. Clearly, Σn = (1 + rn v)θθ 1θ 01 + (Ip − θ 1θ 01 ) has
−1θ θ 0 + (I −
determinant 1 + rn v and inverse matrix Σ −1
1 1
p
n = (1 + rn v) P
0
0
1
θ 1θ 1 ) = Ip − (rn v/(1 + rn v))θθ 1θ 1 . Thus, letting Sn0 := n ni=1 Xni X0ni ,
the hypothesis Pθ 1 ,rn ,v is associated with the density (with respect to the
Lebesgue measure µ over (Rp )n )
n
1 X
dPθ 1 ,rn ,v
(2π)−np/2
0
0
θ
=
exp
−
X
I
−
(r
v/(1
+
r
v))θ
θ
X
n
n
1 1
ni
ni p
dµ
2
(1 + rn v)n/2
i=1
n
1 X
(2π)−np/2
nrn v
0
2
θ
S
θ
=
exp
−
kX
k
+
,
n0
1
ni
1
2
2(1 + rn v)
(1 + rn v)n/2
i=1
which yields (see (4.1))
Λn =
=
nrn v 0
0
(θθ 1 + νnτ n )0 Sn0 (θθ 01 + νnτ n ) − θ 00
S
θ
1 n0 1
2(1 + rn v)
nrn vνn2
nrn vνn 0
τ n Sn0θ 01 +
τ 0 Sn0τ n .
1 + rn v
2(1 + rn v) n
Throughout the rest of the proof, we work under Pθ 01 ,rn ,v , and, accordingly,
0
Σ n = Ip + rn v θ 01θ 00
1 . Since n(Sn0 − Sn ) = nX̄n X̄n is OP (1), we obtain that,
in cases (i)-(iv),
Λn =
nrn vνn 0
nrn vνn2
τ n Snθ 01 +
τ 0 Snτ n + oP (1).
1 + rn v
2(1 + rn v) n
Now, τ 0nΣ nθ 01 = (1 + rn v)ττ 0nθ 01 = − 12 (1 + rn v)νn kττ n k2 (see (4.2)), so that
Λn =
nrn vνn 0
nrn vνn2
nrn vνn2
τ n (Sn − Σ n )θθ 01 +
τ 0n Snτ n −
kττ n k2 + oP (1)
1 + rn v
2(1 + rn v)
2
=
nrn vνn 0
nrn vνn2
τ n (Sn − Σ n )θθ 01 +
τ 0 (Sn − (1 + rn v)Ip )ττ n + oP (1).
1 + rn v
2(1 + rn v) n
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
27
The identity (1 + rn v)Ip = Σ n + rn v(Ip − θ 01θ 00
1 ) then provides
Λn =
nrn vνn2
nrn vνn 0
τ n (Sn − Σ n )θθ 01 +
τ 0 (Sn − Σ n )ττ n
1 + vn
2(1 + rn v) n
−
=
nrn vνn 0
nrn vνn2
τ n (Sn − Σ n )θθ 01 +
τ 0 (Sn − Σ n )ττ n
1 + rn v
2(1 + rn v) n
−
(A.14)
nrn2 v 2 νn2 0
τ n + oP (1)
τ (Ip − θ 01θ 00
1 )τ
2(1 + rn v) n
nrn2 v 2 νn2
nrn2 v 2 νn4
kττ n k2 +
kττ n k4 + oP (1),
2(1 + rn v)
8(1 + rn v)
where we used (4.2) again. We can now consider the cases (i)-(iv).
√
Let us start with cases (i)-(ii). Take then νn = 1/( nrn ) (recall that rn ≡ 1
in case (i)) and let δ = 1 (resp., δ = 0) if case (i) (resp., case (ii)) is
√
considered. The facts that, in both cases (i)-(ii), n(Sn − Σ n ) = OP (1)
√
(Lemma 2.1), τ 0nθ 01 = − 21 νn kττ n k2 = o(1) and nrn νn2 = o(1) yield
√
v
v2
τ 0n n(Sn − Σ n )θθ 01 −
kττ n k2 + oP (1)
1 + δv
2(1 + δv)
√
√
v
v
θ 01 +
θ 01
=
τ 0n (Ip − θ 01θ 00
(ττ 0nθ 01 )θθ 00
1 ) n(Sn − Σ n )θ
1 n(Sn − Σ n )θ
1 + δv
1 + δv
Λn =
−
=
v2
v2
τ
τ 0n (Ip − θ 01θ 00
)τ
−
(ττ 0 θ 0 )2 + oP (1)
n
1
2(1 + δv)
2(1 + δv) n 1
√
v
v2
τ n + oP (1).
θ 01 −
τ 0n (Ip − θ 01θ 00
τ 0 (Ip − θ 01θ 00
1 ) n(Sn − Σ n )θ
1 )τ
1 + δv
2(1 + δv) n
Now, Lemma 2.1 implies that
√
√
0 00
θ 01 = (θθ 00
(Ip − θ 01θ 00
1 ) n(Sn − Σ n )θ
1 ⊗ (Ip − θ 1θ 1 ))vec( n(Sn − Σ n ))
is asymptotically normal with mean zero and covariance matrix (1+δv)(Ip −
θ 01θ 00
1 ), which establishes the result in cases (i)-(ii).
Let us turn to the proof in cases (iii)-(iv). The facts that, in both cases,
√
√
we have n(Sn − Σ n ) = OP (1) and nrn = O(1) implies that, with ν ≡ 1,
(A.14) becomes
nrn v 0
τ (Sn − Σ n )ττ n
2 n
nr2 v 2
nr2 v 2
− n kττ n k2 + n kττ n k4 + oP (1).
2
8
Λn = nrn vττ 0n (Sn − Σ n )θθ 01 +
(A.15)
28
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
√
In case (iv), where nrn = o(1), it directly follows that Λn = oP (1).
√
In case (iii), where rn = 1/ n, (A.15) yields the announced stochastic
quadratic expansion of Λn . Finally, still in case (iii), Lemma 2.1 implies
√
that, if (ττ n ) → τ , then τ 0n n(Sn − Σ n ) θ 01 + 21 τ n is asymptotically normal
with mean zero and variance kττ k2 − 41 kττ k4 , which establishes the result.
Proof of Theorem 4.2. Since optimality of the test φHPV was established
above the statement of the theorem, we only have to derive the asymptotic non-null distribution of QHPV under Pθ 01 +ττ n /(√nrn ),rn ,v . Recall that,
√
θ 01 is asymptotically normal with
under Pθ 01 ,rn ,v , n(Ip − θ 01θ 00
1 )(Sn − Σ n )θ
mean zero and covariance matrix (1 + δv)(Ip − θ 01θ 00
1 ); see the proof of Theorem 4.1(i)-(ii). A routine application of the Le Cam third lemma thus
√
provides that, under Pθ 01 +ττ n /(√nrn ),rn ,v , with (ττ n ) → τ , n(Ip − θ 01θ 00
1 )(Sn −
0
0 00
Σ n )θθ 1 is asymptotically normal with mean v (Ip − θ 1θ 1 )ττ and covariance
matrix (1 + δv)(Ip − θ 01θ 00
1 ). Since contiguity implies that the asymptotic
equivalence (see (A.11))
QHPV =
n
θ 01 + oP (1)
θ 00 (Sn − Σ n )(Ip − θ 01θ 00
1 )(Sn − Σ n )θ
1 + δv 1
also holds under Pθ 01 +ττ n /(√nrn ),rn ,v , we obtain that, under the same sequence
of hypotheses, QHPV is asymptotically non-central chi-square with p − 1
τ )0 ((1 +
degrees of freedom and with non-centrality parameter (v (Ip − θ 01θ 00
1 )τ
−
θ 01θ 100 )ττ ), which establishes the result (note that (4.2)
δv)(Ip −θθ 01θ 00
1 )) (v (Ip −θ
00
here implies that θ 1 τ = 0).
Proof of Theorem 4.3. Since the claims in regime (iv) are trivial, we
only show the result in regime (iii). The sequences of hypotheses Pθ 01 ,1/√n,v
and Pθ 01 ,1/√n,v are mutually contiguous (which follows, as already mentioned,
from the Le Cam first lemma). The Le Cam third lemma then yields that,
√
Σn )θθ 01 is asymptotically
under Pθ 01 +ττ n ,rn ,v , with (ττ n ) → τ , n(Ip −θθ 01θ 00
1 )(Sn −Σ
normal with mean
1
1
0 00
2
2 0
τ
τ
τ
kτ
k
τ
+
kτ
k
θ
v(1 + θ 00
τ
)(I
−
θ
θ
)τ
=
v
1
−
p
1
1 1
1
2
2
and covariance matrix Ip −θθ 01θ 00
1 , where we used (4.2). As in the proof of the
previous theorem, contiguity yields that the asymptotic equivalence (which,
in (A.11), also holds in regime (iii))
QHPV =
n
θ 01 + oP (1)
θ 00 (Sn − Σ n )(Ip − θ 01θ 00
1 )(Sn − Σ n )θ
1 + δv 1
TESTING FOR PC DIRECTIONS UNDER WEAK IDENTIFIABILITY
29
extends to Pθ 01 +ττ n ,rn ,v , with (ττ n ) → τ , and we may therefore conclude that,
under this sequence of hypotheses, QHPV is asymptotically non-central chisquare with p − 1 degrees of freedom and with non-centrality parameter
h
h
i0
i
1
1
1
1
−
v 1 − kττ k2 τ + kττ k2θ 01 .
v 1 − kττ k2 τ + kττ k2θ 01 (Ip − θ 01θ 00
1)
2
2
2
2
This establishes the result since a direct computation shows that this noncentrality parameter coincides with the one in (4.7).
REFERENCES
Anderson, T. W. (1963). Asymptotic theory for principal component analysis. Ann.
Mathem. Statist. 34 122–148.
Bali, J. L., Boente, G., Tyler, D. E., Wang, J.-L. et al. (2011). Robust functional
principal components: A projection-pursuit approach. Ann. Statist. 39 2852–2882.
Berthet, Q. and Rigollet, P. (2013). Optimal detection of sparse principal components
in high dimension. Ann. Statist. 41 1780–1815.
Boente, G. and Fraiman, R. (2000). Kernel-based functional principal components.
Statist. Probab. Lett. 48 335–345.
Croux, C. and Haesbroeck, G. (2000). Principal component analysis based on robust
estimators of the covariance or correlation matrix: influence functions and efficiencies.
Biometrika 87 603–618.
Cuevas, A. (2014). A partial overview of the theory of statistics with functional data. J.
Statist. Plann. Inference 147 1–23.
Dufour, J. M. (1997). Some impossibility theorems in econometrics with applications to
instrumental variables and dynamic models. Econometrica 65 1365–1388.
Dufour, J. M. (2006). Monte Carlo tests with nuisance parameters: a general approach
to finite-sample inference and nonstandard asymptotics. J. Econometrics 133 443–477.
Forchini, G. and Hillier, G. (2003). Conditional inference for possibly unidentified
structural equations. Econometric Theory 19 707–743.
Hallin, M. and Paindaveine, D. (2006). Semiparametrically efficient rank-based inference for shape. I. Optimal rank-based tests for sphericity. Ann. Statist. 34 2707–2756.
Hallin, M., Paindaveine, D. and Verdebout, T. (2010). Optimal rank-based testing
for principal components. Ann. Statist. 38 3245–3299.
Hallin, M., Paindaveine, D. and Verdebout, T. (2014). Efficient R-estimation of
principal and common principal components. J. Amer. Statist. Assoc. 109 1071–1083.
Han, F. and Liu, H. (2014). Scale-invariant sparse PCA on high-dimensional metaelliptical data. J. Amer. Statist. Assoc. 109 275–287.
He, R., Hu, B.-G., Zheng, W.-S. and Kong, X.-W. (2011). Robust principal component
analysis based on maximum correntropy criterion. IEEE Trans. Image Process. 20
1485–1494.
Hubert, M., Rousseeuw, P. J. and Vanden Branden, K. (2005). ROBPCA: a new
approach to robust principal component analysis. Technometrics 47 64–79.
Johnstone, I. M. and Lu, A. Y. (2009). On consistency and sparsity for principal components analysis in high dimensions. J. Amer. Statist. Assoc. 104 682–693.
Magnus, J. R. and Neudecker, H. (2007). Matrix Differential Calculus with Applications in Statistics and Econometrics, 3rd edition ed. John Wiley & Sons, Chichester.
30
D. PAINDAVEINE, J. REMY AND TH. VERDEBOUT
Paindaveine, D. and Verdebout, T. (2017). Inference on the mode of weak directional
signals: a Le Cam perspective on hypothesis testing near singularities. Ann. Statist. 45
800–832.
Pötscher, B. M. (2002). Lower risk bounds and properties of confidence sets for ill-posed
estimation problems with applications to spectral density and persistence estimation,
unit roots, and estimation of long memory parameters. Econometrica 70 1035–1065.
Rao, C. R. and Mitra, S. K. (1971). Generalized Inverses of Matrices and its Applications. J. Wiley, New York.
Tyler, D. E. (1981). Asymptotic inference for eigenvectors. Ann. Statist. 725–736.
Tyler, D. E. (1983a). A class of asymptotic tests for principal component vectors. Ann.
Statist. 1243–1250.
Tyler, D. E. (1983b). The asymptotic distribution of principal component roots under
local alternatives to multiple roots. Ann. Statist. 1232–1242.
van der Vaart, A. W. (1998). Asymptotic Statistics. Cambridge Univ. Press, Cambridge.
Université libre de Bruxelles
ECARES and Département de Mathématique
Avenue F.D. Roosevelt, 50
ECARES, CP114/04
B-1050, Brussels
Belgium
E-mail: [email protected]
[email protected]
[email protected]
URL: http://homepages.ulb.ac.be/˜dpaindav
http://tverdebo.ulb.ac.be
| 10math.ST
|
arXiv:1306.0514v4 [cs.NE] 3 Feb 2015
Riemannian metrics for neural networks II:
recurrent networks and learning symbolic data
sequences
Yann Ollivier
February 4, 2015
Abstract
Recurrent neural networks are powerful models for sequential data,
able to represent complex dependencies in the sequence that simpler
models such as hidden Markov models cannot handle. Yet they are
notoriously hard to train.
Here we introduce a training procedure using a gradient ascent in
a Riemannian metric: this produces an algorithm independent from
design choices such as the encoding of parameters and unit activities.
This metric gradient ascent is designed to have an algorithmic cost close
to backpropagation through time for sparsely connected networks.
We use this procedure on gated leaky neural networks (GLNNs), a
variant of recurrent neural networks with an architecture inspired by
finite automata and an evolution equation inspired by continuous-time
networks.
GLNNs trained with a Riemannian gradient are demonstrated to
effectively capture a variety of structures in synthetic problems: basic block nesting as in context-free grammars (an important feature
of natural languages, but difficult to learn), intersections of multiple
independent Markov-type relations, or long-distance relationships such
as the distant-XOR problem.
This method does not require adjusting the network structure or
initial parameters: the network used is a sparse random graph and the
initialization is identical for all problems considered.
The problem considered here is to learn a probabilistic model for an
observed sequence of symbols (x0 , . . . , xt , . . .) over a finite alphabet A. Such
a model can be used for prediction, compression, or generalization.
Hidden Markov models (HMMs) are frequently used in such a setting.
However, the kind of algorithmic structures HMMs can represent is limited
because of the underlying finite automaton structure. Examples of simple
sequential data that cannot be, or cannot conveniently be, represented by
HMMs are discussed below; for instance, subsequence insertions, or intersections of multiple independent constraints.
1
Recurrent neural networks (RNNs) are an alternative with higher modelling power. However, their training comes with its own limitations; in
particular, picking long-distance dependencies remains problematic [BSF94,
HS97, Jae02]. Techniques to deal with this problem include long short-term
memory (LSTM) networks [HS97] or echo state networks (ESN) [Jae02].
Here we use a new training procedure which realizes a gradient ascent
using a suitable Riemannian metric, instead of backpropagation, at a small
computational cost. Moreover, we use gated leaky neural networks (GLNNs),
a variation on the RNN architecture. More precisely:
• Rather than standard backpropagation through time, for training the
model we use a gradient inspired by Riemannian geometry, using metrics for neural networks as introduced in [Oll13], adapted to a recurrent context. This makes learning less sensitive to arbitrary design
choices, and provides a substantial improvement in learning speed and
quality. An important point is doing so while keeping a scalable algorithm. Here the asymptotic algorithmic complexity is identical to
backpropagation through time for sparsely connected networks.
• In GLNNs, at each time in the production of a sequence of symbols, the
neural network weights depend on the symbol last produced (“gated”
units). This is inspired by finite automata in which the next state
depends both on the current state and the currently produced symbol,
and allows for an easy representation of automaton-like structures.
Such “gated” models have already been used, e.g., in [SMH11], and
arguably the LSTM architecture.
• The dynamics of GLNNs is modified in a way inspired by continuoustime (or “leaky”) neural networks: the connection weights between
the units control the variation of the activation levels, rather than
directly setting the activation levels at the next step. This provides
an integrating effect and is efficient, for instance, at modelling some
hierarchical, context-free-grammar–like structures in which an internal
state must be held constant while something else is happening.
Much of this text is devoted to the derivation of Riemannian metrics
for recurrent networks. Indeed, we believe the use of a proper gradient is a
major ingredient for an effective learning procedure. The standard gradient
ascent update over a parameter θ can be seen as a way to increase the value
of a function f (θ) while changing as least as possible the numerical value θ:
∂f
θ =θ+η
∂θ
′
⇒
1
θ ≈ arg max f (θ ) −
θ − θ′
2η
θ′
′
′
2
(1)
for small enough learning rates η (where ≈ means “up to O(η 2 ) when η →
0”). The norm kθ − θ ′ k depends on how the parameters are cast as a set of
2
real numbers. If, instead, one uses a measure of distance between θ and θ ′
depending on what the network does, rather than how the numbers in θ and
θ ′ differ, the penalty for moving θ in different directions becomes different
and hopefully yields better learning. One possible benefit, for instance,
is self-adaptation of the cost of moving θ in certain directions, depending
on the current behavior of the network. Another benefit is invariance of
the learning procedure from a number of designing choices, such as using
a logistic or tanh activation function, or scaling the values of parameters
(choices which affect the conventional gradient ascent).
The primary example of an invariant gradient ascent is Amari’s natural
gradient, which amounts to replacing kθ − θ ′ k2 with the Kullback–Leibler
divergence KL(Prθ || Prθ′ ) between the distributions defined by the network
(seen as a probabilistic model of the data). However, the natural gradient
comes at a great algorithmic cost. “Hessian-free” techniques [Mar10, MS11,
MS12] allow to approximate it to some extent and have yielded good results,
but are still quite computationally expensive.
Here we build two metrics for recurrent neural networks having some
of the key properties of the natural gradient, but at a computational cost
closer to that of backpropagation through time. The resulting algorithm
is first presented in Section 2 in its final form. The algorithm might look
arbitrary at first sight, but is theoretically well-grounded; in Sections 3.1–3.5
we derive it step by step from the principles in [Oll13] adapted to a recurrent
setting.
This construction builds on the Riemannian geometry framework for
neural networks from [Oll13]. The activities of units in the network are
assumed to belong to a manifold: intuitively, they represent “abstract quantities” representable by numbers, but no preferred correspondence with R is
fixed. This forces us to write only invariant algorithms which do not depend
on the chosen numerical representation of the activities. Such algorithms
are more impervious to design choices (e.g., changing the activation function
from logistic to tanh has no effect); as a consequence, if they work well on
one problem, they will tend to work well on rewritings of the same problem
using different numerical representations. Thus, such algorithms are more
“agnostic” as to physical meaning of the activities of the units (activation
levels, activation frequencies, log-frequencies, ...).
Remark 1. The three changes introduced above with respect to standard
RNNs are independent and can be used separately. For instance, the metrics
can be used for any network architecture.
Remark 2. The approach is not specific to symbolic sequences: instead
of transition parameters τijxt depending on the latest symbol xt , one can
use transition weights which depend on the components of the latest input
vector xt .
3
Remark 3. The gradient update proposed is independent of the training example management scheme (batch, online, small batches, stochastic
gradient. . . ).
Remark 4. The algorithm presented here is quadratic in network connectivity (number of connections per unit), and we have used it with very sparse
networks (as few as 3 connections per unit), which apparently perform well.
For non-sparse networks, a version with complexity linear in the number of
connections, but with fewer invariance properties, is presented at the end of
Section 2.
Examples. Let us present a few examples of data that we have found can
be efficiently learned by GLNNs. Other techniques that have been used to
deal with such sequences include long short-term memory (LSTM) networks
[HS97] (see for instance [Gra13] for a recent application using stacked LSTMs
for text modelling) and echo state networks (ESN) [Jae02]. Here we do not
have to engineer a particular network structure or to have prior knowledge
of the scale of time correlations for initialization: in our experiments the
network is a sparse random graph and parameter initialization is the same
for all problems.
Example 1 illustrates a type of operation frequent in natural languages
(and artificial programming languages): in the course of a sequence, a subsequence is inserted, then the main sequence resumes back exactly where it
was interrupted. This kind of structure is impossible to represent within a
Markovian model, and is usually modelled with context-free grammars (the
learning of which is still problematic).
In this example, the main sequence is the Latin alphabet. Sometimes a
subsequence is inserted which spells out the digits from 0 to 9. In this subsequence, sometimes a subsubsequence is inserted containing nine random
(to prevent rote learning) capital letters (Example 1).
abcdefghijklmnopqrs(01[HSATXUEUZ]2[OYNFIWWOR]345[ZYMBOMYBZ]6789)tuvwxyz
abcde(01234567[FFRLCMKVI]89)fghijklmnopqrstuvwxyz
...
Example 1: Inserting subsequences, a simple context-free grammar.
Here the difficulty, both for HMMs and recurrent neural networks trained
by ordinary backpropagation through time, is in starting again at the right
point after the interruption caused by the subsequence.
Example 2 is a pathological synthetic problem traditionally considered
among the hardest for recurrent neural networks (although it can be represented by a simple finite automaton): the distant XOR problem. In a
random binary sequence, two positions are marked at random (here with
the symbol X), and the binary symbol at the end of each line is the logical
4
XOR of the two random bits following the X marks. Use of the XOR function prevents detecting a correlation between the XOR result and any one
of the two arguments.
1 1 1 0 0X1 1X1 1 1 1 1 0 0 1 0 1 0 0 0 1 0 1 0 0 1=0
0X1X0 1 1 0 1 0 1 1 1 0 1 0 1 0 0 1 0 0 1 0 0 1 0 0=1
...
Example 2: Long-distance XOR.
On this example, apparently the best performance for RNNs is obtained
in [MS11]: with 100 random bits on each line, the failure rate is about 75%,
where “failure” means that a run examines more than 50 million examples
before reaching an error rate below 1% [MS11, legend of Figure 3].
Example 3 is synthetic music notation (here in LilyPond format1 ), meant
to illustrate the intersection of several independent constraints. Successive
musical bars are separated by a | symbol. Each bar is a succession of notes
separated by spaces, where each note is made of a pitch (a, b, c, ...) and value
(4 for a quarter note, 2 for a half note, 4. for a dotted quarter note, etc.). In
each bar, a hidden variable with three possible values determines a harmony
which restricts the possible pitches used in this bar. Harmonies in successive
bars follow a specific deterministic pattern. Additionally, in each bar, the
successive durations are taken from a finite set of possibilities (rhythms
commonly encountered in waltzes). Rhythm is chosen independently from
pitch and harmony. The resulting probability distribution is the intersection
of all these constraints.
c2 c4 | f4. a8 c4 | g4 b4 g8 d8 | g4. g8 g4 | e4 c4 c4 | ...
Example 3: Synthetic music.
This example can be represented as a Markov chain, but only using a
huge state space. The “correct” representation of the constraints is more
compact, which allows for efficient learning, whereas a Markov representation would essentially need to see every possible combination of rhythm and
pitches to learn the underlying structure.
Example 4 is the textbook example of sequences that cannot be represented by a finite automaton (thus also excluding an HMM): sequences of
the form an bn . The sequence alternates blocks of random length containing
only a’s and only b’s, with the constraint that the length of a b-block is
equal to the length of the a-block preceding it. The blocks are separated
with newlines.
Seen as a temporal sequence, this exhibits long-term dependencies, especially if the block lengths used in the training sequence are long. GLNNs
1
http://lilypond.org/
5
aaaaaaa
bbbbbbb
aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa
bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb
aaaaaaaaaaaaaaaaaaaaaaaaaaa
bbbbbbbbbbbbbbbbbbbbbbbbbbb
...
Example 4: an bn
are found to be able to learn this model within minutes with a training set
of as few as 10 examples with the block lengths ranging in the thousands.
Experiments for each of these examples are given in Section 4, both for
GLNNs and more traditional RNNs: a GLNN or RNN network is trained on
a single (long) training sequence2 and evaluated on an independent validation sequence, for a given computation time. More experiments attempt to
isolate the respective contributions of the three changes introduced (leakiness, gatedness, and Riemannian training). Hidden Markov models, LSTMs,
and classical text compression methods are included as a baseline.
The code for these experiments can be downloaded at http://www.yann-ollivier.org/rech/cod
1
1.1
Definition of the models
Generative models for sequential data
A generative model for symbolic sequences is a model which produces an
infinite random sequence of symbols (x0 , . . . , xt , . . .) over a finite alphabet
A. The model depends on a set of internal parameters θ: each θ defines
a probability distribution Prθ ((xt )t=0,1,... ) over the set of infinite sequences.
Given an actual training sequence (xt ), the goal of learning is to find the
value of θ that maximizes the probability of the training sequence (xt ):
θ = arg max Prθ ((xt )t=0,1,... ) = arg max log Prθ ((xt )t=0,1,... )
θ
= arg max
θ
(2)
θ
X
t
log Prθ (xt |x0 x1 . . . xt−1 )
(3)
where the latter sum can usually be computed step by step. This value of θ
is then used for prediction of future observations, generation of new similar
sequences, or compression of the training sequence.
2
We chose a single long training sequence rather than several short sequences, first, to
avoid giving the algorithms a hint about the time scales at play; second, because in some
of the problems presented here, there are no marked cuts (music example), or finding the
relevant cuts is part of the game (an bn example); third, because having several training
sequences is not always relevant, e.g., if there is a single temporal stream of data.
6
The generative models considered here work in an iterative way. At each
time step t, the system has an internal state. This internal state is used to
compute a probability distribution πt over the alphabet. The symbol xt
printed at time t is drawn from this distribution πt . Then the new internal
state as time t + 1 is a deterministic or random function of the internal state
at time t together with the symbol xt just printed.
Computing the probability of an actual training sequence (xt ) can be
done iteratively, by computing the probability π0 assigned by the model
to the first symbol x0 , then revealing the actual value of x0 , using this
x0 to compute the internal state at time 1, which is used to compute the
probabilistic distribution of x1 , etc. (forward pass).
In a variant of the problem, only some of the symbols in the sequence (xt )
have to be predicted, while the others are given “for free”. For instance, in
a classification task the sequence (xt ) might be of the form y0 z0 y1 z1 y2 z2 . . .
where for each instance yi we have to predict the corresponding label zi . In
this case the problem is to find the θ maximizing the probability of those
symbols to be predicted:
θ = arg max
θ
where
χt =
1.2
X
t
(
χt log Prθ (xt |x0 x1 . . . xt−1 )
1 if xt is to be predicted,
0 otherwise.
(4)
(5)
Recurrent neural network models
We now present the recurrent network models discussed in this work. These
include ordinary recurrent neural networks (RNNs), gated neural networks
(GNNs), and leaky GNNs (GLNNs).
Neural network–based models use a finite oriented graph N , the network,
over a set of units. The internal state is a real-valued function over N (the
activities), and edges in the graph indicate which units of the network at
time t contribute to the computation of the state of units at time t + 1.
At each time step t, each unit i in the network N has an activation
level ati ∈ R. As is usual for neural networks, we include a special, alwaysactivated unit i = 0 with at0 ≡ 1, used to represent the so-called “biases”.
The activation levels at time t are used to compute the output of the network
at time t and the activation levels at time t + 1. This transition function
is different for RNNs, GNNs, and GLNNs, as defined below (Sections 1.2.1–
1.2.3).
For the output of the network we always use the softmax function: each
unit i ∈ N (including i = 0) has time-independent writing weights wix for
each symbol x in the alphabet A. At each time, the network outputs a
7
random symbol x ∈ A with probabilities given by the exponential of the
writing weights weighted by the activation levels at that time:
πt (x) := P
P
e
ati wix
i
P
y∈A e
i
ati wiy
(6)
where πt (x) is the probability to print x ∈ A. This allows any active unit
to sway the result by using a large enough weight. One effect of this “nonlinear voting” is to easily represent intersections of constraints: If an active
unit puts high weight on a subset of the alphabet, and another active unit
puts high weight on another subset of the alphabet, only the symbols in the
intersection of these subsets will have high probability.
Thus, given the activities (ati )i∈N at time t, the network prints a random
symbol xt drawn from πt . Then the network uses its current state and its
own output xt to compute the activities at time t + 1: the (at+1
i ) are a
t
deterministic function of both the (ai )i∈N and xt .
Given the writing weights wix , the model-specific transition function (depending on model-specific transition parameters τ ), and the initial activation
levels a0i , the model produces a random sequence of symbols x0 , x1 , . . . , xt , . . ..
Given a training sequence (xt ), the goal of training is to find parameters wix ,
τ and a0i maximizing the probability to print (xt ):
Pr((xt )t=0,...,T −1 ) =
TY
−1
πt (xt )
(7)
t=0
The parameters θ = (w, τ, a0 ) can be trained by gradient ascent θ ← θ +
The gradient of the (log-)probability to print (xt ) with respect to
the parameters can be computed by the standard backpropagation through
time technique [RHW87, Jae02]. Appendix B describes backpropagation
through time for the GLNN model (Proposition 11).
However, here we will use gradient ascents in suitable, non-trivial metrics kθ − θ ′ k given by a symmetric, positive-definite matrix M (θ). The corresponding gradient ascent will take the form θ ← θ + ηM (θ)−1 ∂ log∂θPr(x)
(see Section 3.1). These metrics are built in Sections 3.3–3.5 to achieve
reparametrization invariance at a reasonable computational cost, based on
ideas from [Oll13].
We first give the full specification for the three neural network models
used.
η ∂ log∂θPr(x) .
8
1.2.1
Recurrent Neural Networks
In this article we use the following transition function to compute the RNN
activation levels at step t + 1 (see for instance [Jae02]):
Vjt+1 := ρjxt +
X
τij ati
(8)
i
:= s(Vjt+1 ),
at+1
j
(9)
where s is a fixed activation function, xt ∈ A is the symbol printed at time
t, and the sum runs over all edges ij in the network. The sum also includes
the always-activated unit i = 0 to represent “biases”3 .
The parameters to be trained are the input parameters ρjxt and the
transition parameters τij . The parameter ρixt can equivalently be thought
of as a connection weight from an input unit activated when reading xt .
Two standard choices for the activation function are the logistic function
s(V ) := eV /(1 + eV ) = 1/(1 + e−V ) and the hyperbolic tangent s(V ) :=
tanh(V ). Actually the two are related: one is obtained from the other by
an affine transform of V and a. Traditional learning procedures would yield
different results for these two choices. With the training procedures below
using an invariant metric, using the tanh function instead of the logistic
function would result in the same learning trajectory so that this choice is
indifferent. To fix ideas, the experiments were implemented using tanh.
1.2.2
Gated Neural Networks
GNNs are an extension of recurrent neural networks, in which the neural
network transition function governing the new activations depends on the
last symbol written. This is inspired by finite automata. Such models have
also been used in [SMH11], the main difference being the non-linear softmax
function (6) we use for the output.
In GNNs the activation levels at step t + 1 are given by
Vjt+1 :=
X
ati τijxt
(10)
i
:= s(Vjt+1 ),
at+1
j
(11)
where s is the same activation function as in RNNs. The sum includes the
always-activated unit i = 0.
In the above, xt ∈ A is the symbol printed at step t, and the parameters
τijx are the transition weights from unit i to unit j given context x ∈ A:
contrary to RNNs, τijx depends on the current signal x. This amounts
3
Biases are actually redundant in this case: the bias τ0i at unit i has the same effect as
adding τ0i to all the input weights ρix for all symbols x, since at any time, one and exactly
one symbol is active. Still, since backpropagation is not parametrization-invariant, using
or not using these biases has an effect on learning.
9
to having an RNN with different parameters τ for each symbol x in the
alphabet. (This is not specific to discrete-valued sequences here: a continuous vector-valued signal xt with components xkt could trigger the use of
P k
k xt τijk as transition coefficients at time t.)
Hidden Markov models are GNNs with linear activation function [Bri90]:
if we set s(V ) := V and if τijx is set to (HMM probability that unit i prints symbol x)×
(HMM transition probability from i to j), then the GNN transition (10) yields
the update equation for the HMM forward probabilities4 . If, in
Paddition, we
replace the softmax output (6) with a linear output πt (x) :=
t
a w
i i ix
P
where
at
i
i
wix is the HMM probability to write x in state i, then the GNN model
exactly reduces to the HMM model.5
GNNs have more parameters than standard recurrent networks, because
each edge carries a parameter for each letter in the alphabet. This can be
a problem for very large alphabets (e.g., when each symbol represents a
word of a natural language): even storing the parameters can become costly.
This is discussed in [SMH11], where a factorization technique is applied to
alleviate this problem.
1.2.3
Gated Leaky Neural Networks
Gated leaky neural networks are a variation over GNNs which allow for
better handling of some distant temporal dependencies. They are better
understood by a detour through continuous-time models. In GNNs we have
P
Vjt+1 = i τijxt ati . One possible way to define a continuous-time analogue
is to set
dVjt X
τijxt ati
(12)
=
dt
i
and set atj = s(Vjt ) as before. See [Jae02] for “continuous-time” or “leaky”
neural networks.
This produces an “integration effect”: units become activated when a
certain signal xt occurs, and stay activated until another event occurs. Importantly, the transition coefficient τiixt from i to i itself provides a feedback
control. For this reason, in our applications, loops i → i are always included
in the graph of the network.
Here, contrary to the models in [Jae02], the differential equation is written over V which results in a slightly different equation for the activity a.6
4
More precisely ati becomes the probability to have emitted y0 , . . . , yt−1 and be in state
i at time t, i.e., the HMM probabilities right before emitting xt but after the t − 1 → t
state transition.
5
Conversely, any system of the form at+1 = F (at , xt ) and law(xt+1 ) = G(at+1 ), can be
viewed as a Markov process on the infinite continuous space in which (at , xt ) take values.
6
Making V rather than a the leaky variable comes from looking for the simplest possible
nonlinear dynamics in the context of differential geometry for neural networks [Oll13]. In
full generality, if the activity unit j is a point aj in a manifold Aj , the continuous-time
10
Gated leaky neural networks are obtained by the obvious time discretization of this evolution equation. This is summed up in the following definition.
Definition 5. A gated leaky neural network (GLNN) is a network as
above, subjected to the evolution equation
Vjt+1 := Vjt +
X
atj := s(Vjt )
τijxt ati ,
(13)
i
(where the sum includes the always-activated unit i = 0). The probability
to output symbol x at time t is given by
πt (x) := P
P
e
ati wix
i
P
y∈A e
i
ati wiy
(14)
Appendix A provides a further discussion of the integrating effect by
studying the linearized regime. This is useful to gain an intuition into GLNN
behavior and to obtain a sensible parameter initialization.
2
An algorithm for GLNN training
In Section 3 we expose theoretical principles along which to build Riemannian algorithms for RNN, GNN and GLNN training. For convenience, we
first collect here the explicit form of the final algorithm obtained for GLNNs,
and discuss its algorithmic cost.
The derivatives of the log-likelihood of the training data with respect to
the writing and transition weights, can be computed using backpropagation
through time adapted to GLNNs (Appendix B). These derivatives are turned
into a parameter update
θ ← θ + ηM (θ)−1
∂ log Pr(x)
∂θ
(15)
through a suitable metric M (θ). We present two algorithmically efficient
choices for M : the recurrent backpropagated metric (RBPM) and the recurrent unitwise outer product metric (RUOP metric).
dynamics will be daj /dt = Fj ((ai )i→j , xt ) where Fj is a vector field on Aj depending on
the activitives of units connected to j and on the current signal xt . Looking for dynamics
with a simple form, it makes sense to assume that the vector-field–valued function Fj
is the product of a fixed vector field Fj0 times a real-valued function of the of the ai ,
and that the latterPdecomposes as a sum of the influences of individual units i, namely:
Fj ((ai )i→j , xt ) = ( i fi (ai , xt ))Fj0 . For one-dimensional activitives, if Fj0 does not vanish,
there always exists a particular chart of the manifold Aj , unique up to an affine transform,
in which Fj0 is constant: we call this chart Vj . Further assuming that fi (ai , xt ) decomposes
as a product of a function of xt and a function of ai , namely fi (ai , xt ) = τi (xt )gi (ai ), we
can set ai := gi (ai ), and we obtain the dynamics (12). Both variables V and a are thus
recovered uniquely up to affine transform, respectively, as the variable that makes the
time evolution uniform and the variable that makes the contribution of incoming units
additive.
11
For the update of the writing weights wix , we use the quasi-diagonal
reduction [Oll13, Sect. 2.3] of the Hessian or Fisher information matrix (the
two coincide in this case) as the metric. Quasi-diagonal reduction is a process
producing an update with algorithmic cost close to using only the diagonal
of the matrix, yet has some of the reparametrization invariance properties
of the full matrix. The expression for this metric on wix is worked out in
Section 3.2.
The metric M used for updating the transition weights τijx is built in Sections 3.3–3.5. First, in Section 3.3 we build a metric on recurrent networks
from any metric on feedforward networks. This involves “time-unfolding”
[RHW87, Jae02] the recurrent network to view it as a feedforward network with T times as many units (T being the length of the training data),
and then summing the feedforward metric over time (Definition 6). In Sections 3.4 and 3.5 we carry out this procedure explicitly for two feedforward
metrics described in [Oll13]: this yields the RUOP metric and the RBPM,
respectively.
Before starting the gradient ascent, the parameters of the network are
initialized so that at startup, the activation of each unit over time is a
random linear combination of the symbols xt observed in the recent past.
As this latter point provides interesting insight into the behavior of GLNNs,
we discuss it in Appendix A.
Algorithm description. Training consists in adjusting the writing weights
wix , transition weights τijx , and starting values Vi0 (used by the network at
t = 0), to increase the log-likelihood of the training sequence (xt )t under the
model.
The variable χt encodes which symbols in the sequence have to be predicted: it is set to 1 if the symbol xt has to be predicted, and to 0 if xt is
given. Namely, the problem to be solved is
arg max
w,τ,V 0
X
χt log πt (xt )
(16)
t
where πt is the probability attributed by the network to the next symbol
knowing x0 , . . . , xt−1 .
For simplicity we work with a single (long) training sequence (xt )t=0,...,T −1 ;
the algorithm can be extended in a straightforward manner to cover the case
of several training examples, or mini-batches of training sequences (as in a
stochastic gradient algorithm), simply by summing the gradients W , G and
the metrics h̃, M̃ below over the training examples.
The procedure alternates between a gradient step with respect to the
wix , and a gradient step with respect to the τijx and Vi0 , with two distinct
learning rates ηw and ητ . We describe these two steps in turn. It is important
to start with an update of wix , otherwise the metric at startup may be
singular.
12
In the following expressions, all sums over units i in the network N
include the always-activated unit i = 0 with at0 ≡ 1.
Gradient update for the writing weights wix .
to the following steps.
This is done according
1. Forward pass: Compute the activations of the network over the training sequence (xt )t=0,...,T −1 , using the GLNN evolution equations in
Definition 5.
2. Compute the partial derivatives with respect to the writing weights:
Wiy =
X
t
χt ati (1xt =y − πt (y))
(17)
3. Compute the following terms of the Hessian (or Fisher information
matrix) of the log-likelihood with respect to w, using
hyii = εy +
TX
−1
t=0
hy0i =
TX
−1
t=0
χt (ati )2 πt (y)(1 − πt (y)),
χt ati πt (y)(1 − πt (y)),
i ∈ N, y ∈ A
i 6= 0, y ∈ A
(18)
(19)
where εy is a dampening term to avoid divisions by 0. We set εy to
the frequency of y in the training sequence plus the machine epsilon.
4. Update the weights using the quasi-diagonal reduction of the inverse
Hessian:
wiy ←wiy + ηw
w0y
Wiy − W0y hy0i /hy00
hyii − (hy0i )2 /hy00
i 6= 0
(20)
W0y X hy0i Wiy − W0y hy0i /hy00
←w0y + ηw y −
h00
hy hyii − (hy0i )2 /hy00
i6=0 00
(21)
(These formulas may look surprising, but they amount to using weighted
covariances over time between desired output and activity of unit i,
rather than just sums over time [Oll13, Sect. 1.1]; the constant terms
are transferred to the always-activated unit.)
Gradient update for the transition weights τijx .
This goes as follows.
1. Forward pass: Compute the activations of the network over the training sequence (xt )t=0,...,T −1 , using the GLNN evolution equations in
Definition 5.
13
2. Backward pass: Compute the backpropagated values Bit for each unit
i 6= 0 using
Bit = Bit+1 + s′ (Vit ) χt wixt −
P
y πt (y)wiy
+
X
j
τijxt Bjt+1 (22)
initialized with BjT = 0. This is the derivative of data log-likelihood
with respect to Vit . Here s′ is the derivative of the activation function.
3. Compute the following “modulus” m̃ti for each unit i 6= 0 at each time
t. In the RUOP variant, simply set
m̃ti = (Bit )2
(23)
In the RBPM variant, set by induction from t + 1 to t:
m̃ti = s′ (Vit )2 χt
P
2
y πt (y)wiy
2
+ 1 + τiixt s′ (Vit )
−(
P
2
y πt (y)wiy )
m̃t+1
i
+
X
j6=i
2
(τijxt ) m̃t+1
j
(24)
initialized with m̃Ti = 0.
4. For each unit j 6= 0, for each symbol y ∈ A, compute the following
(jy)
(jy)
vector Gi
and matrix M̃ii′ indexed by the units i with i → j in
the network N , including i, i′ = 0.
(jy)
Gi
=
TX
−1
1xt=y ati Bjt+1
(25)
t=0
(this is the derivative of the log-likelihood with respect to τijx ) and
(jy)
M̃ii′
=
TX
−1
1xt=y ati ati′ m̃t+1
j
(26)
t=0
(jy)
Dampen the matrix M̃ii′ by adding ε to the diagonal (we use ε = 1
which is small compared to the T terms in the sum making up M̃ ).
5. Set
G(jy) ← (M̃ (jy) )−1 G(jy)
(27)
and update the transition weights with
(jy)
τijy ← τijy + ητ Gi
for each j 6= 0 and y ∈ A.
14
(28)
6. Update the starting values Vj0 with
Vj0 ← Vj0 + ητ Bj0 /(m̃0j + ε)
(29)
(this is obtained by analogy: this would be the update of τijy with
i = 0 and y a special dummy symbol read at startup—consistently
with the fact that m̃0j and Bj0 have not been used to update τ ).
Initialization of the parameters. At startup, the network N is chosen
as an oriented random graph with d distinct edges from each unit i, always
including a loop i → i. For the tanh activation function, the parameters are
set up as follows (see the derivation in Appendix A):
w0y ← log νy ,
where νy =
P
wiy ← 0
(i 6= 0),
(30)
χt 1xt =y
P
is the frequency of symbol y among symbols to be
χ
t
t
t
predicted in the training data (this way the initial model is an i.i.d. sequence
with the correct frequencies). The transition parameters are set so that each
unit’s activation reflects a random linear combination of the signal in some
time range, as computed in Appendix A from the linearization of the network
dynamics, namely
τiiy ← −α,
τijy ← 0
(i 6= j, i 6= 0)
(31)
and
P
µj
(32)
(ujy − y′ ν̃y′ ujy′ )
4
where the uP
jy are independent random variables uniformly distributed in
1x =y
[0; 1], ν̃y = t T t is the frequency of symbol y in the data, and where
τ0jy ← βj +
µj = 1/(j + 1),
α = 1/2,
q
βj = − α(α − µj )
(33)
for unit j (j > 1) are adjusted to control the effective memory7 of the
integrating effect at unit j (see Appendix A). These values apply to the tanh
activation function. The initial activation values are set to Vj0 = s−1 (βj /α)
with s−1 the inverse of the activation function.
7
In particular, any foreknowledge of the time scale(s) of dependencies in the sequence
may be used to choose relevant values for µj . With our choice, from Appendix A the time
scale for unit j is O(j) at startup, though it may change freely during learning. Multiple
time scales for recurrent networks can be found in several places, e.g., [HB95, KGGS14].
15
Learning rate control. Gradient ascents come with a guarantee of improvement at each step if the learning rate is small enough. Here we test at
each step whether this is the case: If an update of the parameter decreases
data log-likelihood, the update is cancelled, the corresponding learning rate
(ηw or ητ ) is divided by 2, and the update is tried again. On the contrary,
if the update improves data log-likelihood, the corresponding learning rate
is multiplied by 1.1. This is done separately for the writing weights and
transition weights. This is a primitive, less costly form of line search8 .
At startup the value ηw = ητ = 1/N (with N the number of units)
seems to work well in practice (based on the idea that if each unit adapts its
writing weights by O(1/N ) then the total writing probabilities will change
by O(1)).
Computational complexity. If the network connectivity d (number of
edges i → j per unit j) is not too large, the cost of the steps above is
comparable to that of ordinary backpropagation through time.
Let N be the network size (number of units), A the alphabet size, T the
length of the training data, and d the maximum number of edges i → j per
unit j in the network.
The cost of one forward pass is O(N T d) for computing the activities and
O(N T A) for computing the output probabilities. The cost of computing the
quantities Wiy is O(N T A) as well, as is the cost of computing the Hessian
values hy . Applying the update of w costs O(N A). Thus the cost of the w
update is O(N T (d + A)).
Computing the backpropagated values Bjt costs O(N T (d+ A)). The cost
of computing the backpropagated modulus m̃ti is identical.
(jy)
The cost of computing the gradients Gi
is O(N T d) (note that each
time t contributes for only one value of y, namely y = xt , so that there is
no A factor).
(jy)
The costliest operation is filling the matrices M̃ii′ . For a fixed j and
y this matrix is of size d × d. Computing the entries takes time O(T d2 ) for
each j, hence a total cost of O(N T d2 ). (Once more, each time t contributes
for only one value of y so that there is no A factor.) Inverting the matrix has
a cost of O(N d3 ): as this requires no sum over t, this is generally negligible
if T ≫ d.
Thus, the overall cost√(if T ≫ d) of one gradient √
step is O(N T (d2 + A)).
This suggests using d ≈ A. In particular if d = O( A) the overall cost is
the same as backpropagation through time.
If network connectivity is large, there is the possibility to use the quasidiagonal reduction of the matrices M̃ , as described in [Oll13, Sect. 2.3]. This
8
Experimentally, this leads to some slight oscillating behavior when the learning rate
gets past the optimal value (as is clear for a quadratic minimum). This might be overcome
by averaging consecutive gradient steps.
16
(jy)
requires computing only the terms M̃ii′ with i = i′ or i = 0. This removes
the d2 factor and also allows for O(d) inversion, as follows.
Non-sparse networks: quasi-diagonal reduction. The algorithm above
must maintain a matrix of size d×d for each unit i, where d is the number of
units j pointing to i in the network. When d is large this is obviously costly.
The quasi-diagonal reduction process [Oll13, Sect. 2.3] provides a procedure
linear in d while keeping most invariance properties of the algorithm. This is
the procedure already used for the writing weights wiy in (20)–(21). Essentially, at each unit j, the signals received from units i → j are considered to
be mutually orthogonal, except for those coming from the always-activated
unit i = 0. Thus only the terms M̃ii and M̃0i of the matrix are used. The
update of the transition parameters τijy becomes as follows.
1. For each unit j ∈ N and each symbol y ∈ A, compute the vector G(jy)
(jy)
(jy)
(jy)
as before. Compute only the terms M̃00 , M̃ii , and M̃0i of the
(jy)
(jy)
matrix M̃ (jy) in (26). Dampen the diagonal terms M̃00 and M̃ii as
before.
2. Update the transition weights τijy with
(jy)
τijy ←τijy + ητ
τ0jy ←τ0jy +
3
Gi
(jy)
M̃ii
(jy)
− G0
(jy)
(jy)
M̃0i /M̃00
(jy)
i 6= 0
(jy)
− (M̃0i )2 /M̃00
(jy)
G
ητ 0(jy)
M̃00
−
(34)
X M̃ (jy) G(jy) − G(jy) M̃ (jy) /M̃ (jy)
0
00
i
0i
0i
i6=0
(jy)
M̃00
(jy)
M̃ii
(jy)
(jy)
− (M̃0i )2 /M̃00
(35)
Constructing invariant algorithms for recurrent
networks
We now give the main ideas behind the construction of the algorithm above.
The approach is not specific to GLNNs and is also valid for classical recurrent
networks.
3.1
Gradients and metrics
Backpropagation performs a simple gradient ascent over parameter space
to train a network. However, for GLNNs (at least), this does not work
well. One reason is that gradient ascent trajectories depend on the chosen
numerical representation of the parameters. For instance, a non-orthogonal
change of basis in parameter space will yield different learning trajectories;
yet such changes can result from simple changes in data representation (see
the introduction of [Oll13]).
17
This is clear from the following viewpoint. Given a real-valued function
f to be maximized depending on a vector-valued parameter θ, the gradient
ascent update
∂f
θ′ = θ + η
(36)
∂θ
with learning rate η, can be viewed, for small η, as a maximization of f
penalized by the change in θ, namely
1
θ − θ′
θ ≈ arg max f (θ ) −
′
2η
θ
′
′
2
(37)
where the equality holds up to an approximation O(η 2 ) for small η. The
term kθ − θ ′ k2 defines a “cost” of changing θ.
Clearly, different ways to represent the parameter θ as a vector will yield
different costs kθ − θ ′ k2 . For instance, a linear change of basis for θ amounts
to replacing kθ − θ ′ k2 with (θ − θ ′ )⊤M (θ − θ ′ ) with M a symmetric, positivedefinite matrix. The associated gradient update will then be
θ ′ = θ + ηM −1
∂f
∂θ
(38)
which is the general form of a gradient ascent when no privileged norm or
basis is chosen for the parameter vector θ. Moreover, in general the matrix
M may depend on the current value of θ, defining a (Riemannian) metric in
which the norm of an infinitesimal change θ → θ + δθ of the parameter θ is
kδθk2 = δθ⊤M (θ)δθ
(39)
The gradient ascent update defined by such a metric is thus
θ ′ = θ + ηM (θ)−1
∂f
∂θ
(40)
A suitable choice of M can greatly improve learning, by changing the cost
of moving into various directions. Amari, in particular, advocated the use
of the “natural gradient” for learning of probabilistic models: this is a norm
kθ − θ ′ k2nat which depends on the behavior of the probability distribution
represented by θ, i.e., the output probabilities of the network, rather than
on the way θ is decomposed as a set of numbers. Thus the natural gradient
provides invariance with respect to some arbitrary design choices. (As a
consequence, learning does not depend on whether a logistic or tanh is used
as the activation function, for instance, since one can be changed into the
other by a change of variables.)
In [Oll13] we introduced several metrics for feedforward neural networks
sharing this key feature of the natural gradient, at a lesser computational
cost. The main idea is to define the metric according to what the network
does, rather than the numerical values of the parameters. We now show how
these can be used to build invariant metrics for recurrent networks.
18
3.2
The Fisher metric on the output units and writing weights
Whole-sequence Fisher metric and conditional Fisher metric. Metrics for neural networks first rely on choosing a metric on the output of the
network [Oll13]. Here the network’s output is interpreted as a probability
distribution on the sequence (xt ) printed by the network. Amari’s natural gradient and the metrics we use are both based on the Fisher metric
[AN00] on the space of probability distributions. One way to define the
Fisher metric is as an infinitesimal Kullback–Leibler divergence between
two infinitesimally close probability distributions on the same set.
For recurrent neural networks, there is a choice as to which probability
distribution should be considered. One can either view the network as defining a probability distribution Pr over all output sequences (x0 , . . . , xt , . . .),
or equivalently as defining a sequence of conditional probability distributions
πt for the next symbol xt knowing the previous symbols. Thus there are two
ways to define a divergence on the parameter θ based on Kullback–Leibler
divergences for finite-length output sequences (xt )06t6T . One is
D1 (θ, θ ′ ) := KL(Prθ (x0 , . . . , xt , . . .) || Prθ′ (x0 , . . . , xt , . . .))
(41)
where Prθ is the probability distribution over the set of all sequences (x0 , . . . , xt , . . .)
defined by the network with parameter θ. The other depends on the actual
training sequence x and is
D2 (θ, θ ′ ) :=
X
t
=
X
t
KL(Prθ (xt |x0 , . . . , xt−1 ) || Prθ′ (xt |x0 , . . . , xt−1 ))
(42)
KL πt || πt′
(43)
where πt (resp. πt′ ) is the probability distribution on the next symbol xt
defined by the network with parameter θ (resp. θ ′ ) knowing past observations
x0 , . . . , xt−1 .
Arguably, D2 is more adapted to prediction or (online) compression,
while D1 is better suited for generalization and learning. For instance, if
the actual training sequence starts with the letter a, a gradient ascent based
on D2 will not care how a change of θ affects the probability of sequences
starting with a b.
Assuming that the training sequence (xt ) has actually been sampled
from Prθ and is long enough, and under reasonable stationarity and ergodicity assumptions for Prθ , D2 should be a reasonable approximation of D1 .9
9
Indeed one has D1
=
Ex∼Prθ ln
Prθ (x)
Prθ′ (x)
=
Q
Ex∼Prθ ln Q t
Prθ (xt |x0 ...xt−1 )
Prθ′ (xt |x0 ...xt−1 )
=
t
P
P
Prθ (xt |x0 ...xt−1 )
Prθ (xt |x0 ...xt−1 )
E
E
E
ln
=
=
x∼Pr
x
∼Pr
(x
|x
...x
) ln Pr ′ (xt |x0 ...x
(x
...x
)∼Pr
t
t
0
0
θ
t−1
t−1
θ
θ
Prθ′ (xt |x0 ...xt−1 )
t
t−1 )
θ
Pt
KL(Prθ (xt |x0 . . . xt−1 ) || Prθ′ (xt |x0 . . . xt−1 )) so that if averaging over
E
t (x0 ...xt−1 )∼Prθ
t in the actual training sequence is a good approximation of averaging over t for a
Prθ -random sequence, then D1 and D2 are close.
19
However, in a learning situation, it may not be reasonable to assume that
(xt ) is a sample from Prθ until the training of θ is finished. So we do not
assume that D1 ≈ D2 .
Algorithmically, when an actual training sequence x is given, the conditional divergence D2 is much easier to work with, because it can be computed
in linear time, whereas computing D1 would require summing over all possible sequences, or using a Monte Carlo approximation and sampling a large
number of sequences.
For these reasons, we will define a metric based on D2 , i.e., on the Fisher
metric on the successive individual distributions πt .
Fisher metric on the output units. At each time step, the output of the
network is a probability distribution over the alphabet A. The set of these
probability distributions is naturally endowed with the Fisher metric: the
square distance between two infinitesimally close probability distributions π
and π + δπ is
kδπk2nat := 2KL(π || π + δπ)
=
X (δπ(x))2
x∈A
π(x)
(44)
= Ex∼π (δ log π(x))2
(45)
at second order, where δ log π(x) = δπ(x)/π(x) is the resulting variation of
log π(x).
In the networks we consider, at each step the distribution πt for the next
symbol is given by a softmax output
πt (x) = P
P
e
ati wix
i
P
y∈A e
i
for each x in the alphabet A. Let us set Eyt :=
t
Eyt ′
(46)
ati wiy
P
i
ati wiy , so that πt (y) =
eEy / e .
Then the norm kδπt knat of a change δπt resulting from a change δE t in
the values of E t is, by standard arguments for exponential families, found
to be
X
X
(47)
πt (y)(δEyt )2 −
πt (y)πt (y ′ )δEyt δEyt ′
kδπt k2nat =
P
y
y,y ′
(see Appendix C). By a property of exponential families, this is also, for any
y ′′ , the Hessian of − log πt (y ′′ ) with respect to the variables E t . In particular,
in this situation, for the parameters w, the natural gradient with learning
rate 1 coincides with the Newton method.
Metric over the writing coefficients. We can now compute the natural
metric over the writing coefficients wix . Let δwix be an infinitesimal change
20
in the parameters wix : this creates a change δπt in the distribution πt , for
all t. By the discussion above, we are interested in the quantity
X
t
kδπt k2nat
(48)
Changing the writing weights wix does not change the activities of units
P
in the network. Consequently, we have δEyt = i ati δwiy . Thus the above
yields
X
t
kδπt k2nat =
XXX
t y,y ′ i,i′
πt (y ′ )(1y′ =y′′ − πt (y ′′ ))ati ati′ δwiy δwi′ y′
(49)
so that the metric t kδπt k2nat over the parameters wiy is given by the Fisher
matrix
X
(50)
πt (y ′ )(1y′ =y′′ − πt (y ′′ ))ati ati′
Iwiy wi′ y′ =
P
t
which is also, up to sign, the Hessian of the log-likelihood of the training
sequence with respect to the parameters w.
This is a full matrix whose inversion can be costly. The update of the
parameters wiy given in Section 2 corresponds to the quasi-diagonal inverse
of this metric, whose only non-zero terms correspond to y = y ′ and i = i′ or
i = 0. By [Oll13, Sect. 2.3], the quasi-diagonal inverse respects invariance
under affine reparametrization of the activities of each unit.
3.3
Invariant metrics for recurrent networks
The natural gradient arising from the whole-network Fisher metric is algorithmically costly to compute for neural networks (though the “Hessianfree” conjugate gradient method introduced in [Mar10, MS11, MS12] approximates it). We now introduce metrics for recurrent networks that enjoy
some of the main properties of the Fisher metric (in particular, invariance
with respect to a number of transformations of the parameters or of the activities), at a computational cost close to that of backpropagation through
time.
Any invariant metric for feedforward networks can be used to build an invariant metric for recurrent networks, by first “time-unfolding” the network
as in backpropagation through time [RHW87, Jae02], and then by defining
the norm of a change of parameters of the recurrent network as a sum over
time of the norms of corresponding changes of parameters at each time in
the time-unfolded network, as follows.
A recurrent neural network with n units, working on an input of length T ,
can be considered as an ordinary feedforward neural network with nT units
with shared parameters [RHW87, Jae02]. We will refer to it as the timeunfolded network. In the time-unfolded network, a unit is a pair (i, t) with i
a unit in the original network and t a time. The unit (i, t) directly influences
21
the units (j, t + 1) where i → j is an edge of the recurrent network. We also
consider the output distribution πt at time t as a (probability distribution–
valued) output unit of the time-unfolded network, directly influenced by all
time-unfolded units (i, t).
If all time-unfolded units (i, t) use the same parameters θi as the corresponding unit i in the recurrent network, then the behaviors of the timeunfolded and recurrent networks coincide. Thus, let us introduce dummy
time-dependent parameters θit for unit (i, t) of the time-unfolded network,
and decide that the original parameter θi for unit i in the recurrent network
is a “meta-parameter” of the time-unfolded network, which sets all dummy
parameters to θit = θi .
We are now ready to build a metric on recurrent networks from a metric
k·k on feedforward networks. A variation δθ of the parameters of the recurrent network determines a variation δθ t of the (dummy) parameters of the
time-unfolded network, which is an ordinary feedforward network. Thus we
can simply set
kδθk2 :=
X
δθ t
2
(51)
t
where for each t, δθ t is a variation of the parameters of an ordinary feedforward network, for which we can use the norm δθ t .
If the metric used on the time-unfolded network is reparametrizationinvariant, then so will be the metric on the recurrent network (since its
definition does not use any choice of coordinates).
Using this definition for kδθk is actually the same as making independent
gradient updates δθ t for each θ t based on the metric δθ t , then projecting
the resulting update onto the subspace where the value of θ t does not depend
on t (where the projection is orthogonal in the metric k·k). Equivalently, this
amounts to making independent updates for each θ t and then finding the
P
2
time-independent update δθ minimizing t δθ t − δθ .10
Thus, we can use any of the metrics mentioned in [Oll13] for feedforward
networks. Two will be of particular interest here, but other choices are
possible; in particular, in case network connectivity is high, quasi-diagonal
reduction [Oll13, Sect. 2.3] should be used.
Definition 6. Let k·kff be a metric for feedforward networks. The recurrent metric associated with k·kff is the metric for recurrent network parameters defined by
kδθk2rff :=
X
10
t
δθ t
2
ff
(52)
For these equivalent interpretations, one has to assume that there is more than one
training sequence. Indeed, for typical choices of the feedforward network metric δθt ,
with only one training sequence the metric δθt on each individual θt is only of rank one,
hence the corresponding update of θt is ill-defined. On the other hand, even with only
one training sequence, the metric on δθ is generally full-rank, being a sum over time of
rank-one metrics.
22
namely, by summing over time the metric k·kff on the time-unfolded network.
The recurrent backpropagated metric (RBPM) is the norm k·krbp on a
recurrent network associated with the backpropagated metric k·kbp on the
time-unfolded network.
The recurrent unitwise outer product metric (RUOP metric) is the norm
k·kruop associated with the unitwise outer product metric k·kuop on the timeunfolded network.
The latter two metrics are described in more detail below. Both of
them are “unitwise” in the sense that the incoming parameters to a unit are
orthogonal to the incoming parameters to other units, so that the incoming
parameters to different units can be treated independently. (Given a unit
k in the network, we call incoming parameters to k the parameters directly
influencing the activity of unit k, namely, the weights of edges leading to k
and the bias of k.)
Remark 7. We shall use these metrics only for the transition parameters
τ of recurrent networks and GLNNs. For the writing parameters w, the
Hessian, or equivalently the Fisher metric, is easily computed (Section 3.2)
and there is no reason not to use it.
Remark 8 (Multiple training sequences). Definition 6 is given
for a single training sequence (xt ). In the case of multiple training sequences,
one has to first compute the metric for each sequence separately (since the
time-unfolded networks are different if the training sequences have different
lengths) and then define a metric by averaging the square norm kδθk2rff
over the training dataset, as in [Oll13]. There is a choice to be made as
to whether training sequences of different lengths should be given equal
weights or weights proportional to their lengths; the relevant choice arguably
depends on the situation at hand.
Remark 9 (Natural metric and recurrent natural metric). The natural metric of a recurrent network is defined in its own right
and should not be confused with the recurrent-natural metric obtained by applying Definition 6 to the natural metric of the time-unfolded network. For
the natural metric of the recurrent network, the norm of a change of parameter δθ is the norm of the change it induces on the network outputs πt . For
the recurrent-natural metric, the square norm of δθ is the sum over time t, of
the square norm of the change induced on the output by a change δθ t = δθ
of the dummy parameter θ t, so that the influence of δθ is decomposed as
the sum of its influences on each dummy parameter δθ t considered independently. (Still, the influence of θ t on the output at times t′ > t is taken into
account.) Explicitly, if πt is the network output at time t, then the natural
norm is kδθk2nat =
composing
∂πt
∂θ
=
P
P
t
t′ 6t
∂πt
∂θ δθ
∂πt
∂θ t′
2
where k·k is the norm on the outputs πt . De-
this is
P
t
P
23
t′ 6t
∂πt
∂θ t′
2
δθ . On the other hand
the recurrent-natural norm is kδθk2rnat =
P
′
t′
δθ t
2
nat
=
P P
t′
t>t′
∂πt
∂θ t
2
δθ
′
which is generally different and accounts for fewer cross-time dependencies.
We now turn to obtaining more explicit forms of these metrics for the
case of GLNNs. We describe, in turn, the RUOP metric and the RBPM.
For simplicity we will assume that all symbols in the sequence have to be
predicted (χt ≡ 1). Section 2 includes the final formulas for the general case.
3.4
The recurrent unitwise outer product metric
Let us now describe the recurrent unitwise outer product metric (RUOP
metric) in more detail.
We briefly recall the definition of the (non-recurrent) unitwise outer product metric. Suppose we have a loss function L depending on a parameter
θ, and moreover that L decomposes as a sum or average L = Ex∈D ℓ(x) of
a loss function ℓ over individual data samples x in a dataset D. The outer
products of the differentials ∂ℓ(x)
∂θ , averaged over x, provide a metric on θ,
∂ℓ(x)
∂ℓ(x)
namely, Ex∈D ∂θ ⊗ ∂θ given by the matrix
Cij = Ex∈D
∂ℓ(x) ∂ℓ(x)
∂θi ∂θj
(53)
This is the outer product (OP) metric on θ.
The associated gradient ascent for L, with step size θ, is thus θ ←
θ + ηC −1 ∂L
∂θ , and this gradient direction is parametrization-invariant. (One
must be careful that scaling L by a factor λ will result in scaling this gradient
step by 1/λ, which is counter-intuitive, thus step-size for this gradient must
be carefully adjusted.)
When the loss function ℓ is the logarithmic loss ℓ(x) = log Prθ (y|x) of a
probabilistic model Prθ (y|x), as is the case for feedforward networks with y
θ (y|x)
θ (y|x)
⊗ ∂ log Pr
is
the desired output for x, then the OP metric Ex∈D ∂ log Pr
∂θ
∂θ
a well-known approximation to the Fisher metric (for the latter, y would be
sampled from the output of the network seen as a probability distribution,
instead of using only the target value of y). In this context it has been
used for a long time [APF00, RMB07]—sometimes under the name “natural
gradient”, though it is in fact distinct from the Fisher metric, see discussion
in [PB13] and [Oll13].
The OP metric has the following unique property: For a given increment
δL in the value of L, the OP gradient step is the one for which the increment
is most uniformly spread over all samples x ∈ D, in the sense that the
variance Varx∈D δℓ(x) is minimal [Oll13, Prop. 15].
For feedforward networks, the OP metric is given by a full matrix on
parameter space. This is computationally unacceptable for large networks; a
more manageable version is the unitwise OP metric (UOP metric), in which
24
the incoming parameters for each unit are made orthogonal [Oll13]. The
unitwise OP metric is still invariant under reparametrization of the activities
of each unit. This decomposition is also used in [RMB07] (together with a
further low-rank approximation in each block which breaks invariance).
The recurrent UOP metric is obtained from the UOP metric by Definition 6, through summing over time in the time-unfolded network. Let i be a
unit in the recurrent network, and let θi be the set of incoming parameters
to i. A change δθi in θi results in a change δθit of all the dummy parameters
θit of units (i, t) in the time-unfolded network. The square norm of δθi in
the RUOP metric is, by definition (52), the sum over t of the square norms
of δθit in the UOP metric of the time-unfolded network.
For each t and each unit i, the unitwise OP metric on the dummy parameter θit is given by the outer product square of the associated change of the
Prθ (x)
.
objective function log Prθ (x), namely, the outer product square of ∂ log∂θ
t
i
Now θit is a dummy parameter of the time-unfolded network, and is used
exactly once during computation of the network activities, namely, only at
time t to compute the activity Vit and ati = s(Vit ) of unit i. Thus we have
t
∂ log Prθ (x) ∂Vit
∂ log Prθ (x)
t ∂Vi
=
=
B
i
∂θit
∂Vit
∂θit
∂θit
(54)
where the derivatives Bit := ∂ log∂VPrtθ (x) are computed in the usual way by
i
backpropagation through time (Appendix B).
∂V t
The partial derivative ∂θit is readily computed from the evolution equai
tion defining the network: for instance, for GLNNs, the evolution equation
of the time-unfolded network (using dummy parameters) is Vit = Vit−1 +
P t
t−1
t
t
j τjixt−1 aj , so that the derivative of Vi w.r.t. the parameter τjiy is
1y=xt−1 at−1
j .
The unitwise OP metric for the dummy parameter θit is given by the
∂V t
∂V t
Prθ (x)
, which by the above is (Bit )2 ∂θit ⊗ ∂θit .
outer product square of ∂ log∂θ
t
i
i
i
This has to be summed over time to find the recurrent UOP metric for
the true parameter θi . So in the end, the RUOP metric for the incoming
parameters θi at unit i is given for each i by the matrix
(i)
M̃kk′ =
(Bit )2
X
t
∂Vit ∂Vit
∂(θit )k ∂(θit )k′
(55)
where (θit )k denotes the k-th component of the parameter θit , and where the
derivative is with respect to the dummy parameter θit used only at time t.
For GLNNs, this results in the expression given in the algorithm of Section 2: In the end, for the GLNN transition parameter θ = (τjiy )j,i,y , using
t−1
t =1
that ∂Vit /∂τjiy
y=xt−1 aj , the recurrent UOP metric is
kδθk2ruop =
XXX
i
j,j ′
t
(Bit+1 )2 atj atj ′ δτjix
δτjt′ ixt
t
t
25
(56)
The same expression holds for GNNs (but B has a different expression).
The form of the metric. Thus, we find that the RUOP metric on τ is
given by a symmetric matrix with the following properties. These remarks
also hold for the other metric we use, the RBPM below.
First, different units i are orthogonal (there are no cross-terms between
δτjix and δτj ′ i′ x′ for i 6= i′ ).
Second, for GNNs and GLNNs, different symbols x are independent: the
transition parameters τijx and τijx′ with x 6= x′ are mutually orthogonal in
the RUOP metric, i.e., there are no cross-terms for x 6= x′ . This is because,
at any given time t, only the parameters τjixt using the currently read symbol
xt contribute to the evolution equation. This results in a separate matrix
M̃ (ix) for each pair ix in the final algorithm, reducing computational burden.
On the other hand, for RNNs with the evolution equation at+1
= s(ρixt +
i
P
t
j τji aj ), there is no such block decomposition because the transition parameters τij have non-trivial scalar product with all the input parameters
ρix for all x; thus, handling this metric would be quadratic in alphabet size.
If alphabet size is large, one solution is to restrict input to a subset of units.
Another is to use quasi-diagonal reduction [Oll13, Sect. 2.3] to obtain a more
lightweight but still invariant algorithm; this was tested in Section 4.
Third, different units j and j ′ connected to the same unit i are not
independent. (In particular, the “biases” τ0ix corresponding to the alwaysactivated unit j = 0, aj ≡ 1 are not orthogonal to the other transition
weights.) The cross-term between δτjix and δτj ′ ix is
X
1xt=x atj atj ′ (Bit+1 )2
(57)
t
Besides, the derivative of log-likelihood with respect to τjix is t 1xt =x atj Bit+1
(Proposition 11), and the gradient step is obtained by applying the inverse
of the matrix above to this derivative. This problem has an interesting
structure. Indeed, vectors obtained as M −1 G where M is a matrix of the
P
P
form Mjk = t atj atk ct , and G of the form Gj = t atj Y t , are weighted leastsquare regression problems: M −1 G gives the best way to write the vector
Y t /ct , seen as a function of t, as a linear combination of the family atj , seen
as functions of t. This is the “best-fit” interpretation [Oll13, Section 3.3].
Thus, using metrics of this form, each unit i in the network combines
the signals from its incoming units j in an optimal way to match a desired
change in activity (given by Bit ) over time. The two metrics presented here,
RUOP and RBPM, differ by the choice of the weighting ct .
P
Remark 10 (UOP metric and recurrent UOP metric). The
recurrent unitwise OP metric should not be confused with the unitwise OP
metric applied to the recurrent network, which is defined in its own right but
unsuitable for several reasons. For instance, with only one training sequence
26
x, the OP metric for the recurrent network is simply ∂ log∂θPr(x) ⊗ ∂ log∂θPr(x) ,
which is a rank-1 matrix and thus not invertible. On the other hand, on a
single training sequence of length T , the recurrent UOP metric is a sum of T
matrices of rank 1. Thus for a recurrent network, k·kruop 6= k·kuop in general:
one is a time sum of outer product squares, the other is the outer product
square of a time sum. (Compare Remark 9.) So the recurrent UOP metric
performs an averaging of the metric over time rather than over samples, as
is expected in a recurrent setting.
Another similar-looking metric would be the OP metric associated with
P
P
the decomposition log Pr(x) =
t log πt (xt )
t log Pr(xt |x0 , . . . , xt−1 ) =
of the objective function. Such a decomposition gives rise to a metric
P ∂ log πx (xt ) ⊗2
) . This metric is generally full-rank even for a single training
t(
∂θ
P
sequence. The recurrent OP metric, on the other hand, is t ( ∂ log∂θPr(x)
)⊗2 .
t
So while the recurrent OP is the sum over time of the effect of the dummy
time-t parameter θ t on the objective function, the metric just introduced is
the sum over time of the effect of the parameter θ on the t-th component of
the objective function. These are generally different. Computing all partial
πx (xt )
for all t and θ is algorithmically costlier, which is why
derivatives ∂ log∂θ
we did not use this metric.
3.5
The recurrent backpropagated metric
We now work out an explicit form for the recurrent backpropagated metric.
For a feedforward network, the backpropagated metric (BPM), introduced in [Oll13], is defined as follows. Given a metric on the output units of
a network (here the Fisher metric on πt ), one can inductively define a metric
on every unit by defining the square norm kδai k2bp of a change of activity
P
δai at unit i, as the sum j, i→j kδaj k2bp of the square norms of the resulting
changes in activity at units j directly influenced by i, thus “backpropagating”
the definition of the metric from output units to inner units. The metric
kδaj k2bp at unit j is then turned into a metric on the incoming parameters
to j, by setting kδθj k2bp := kδaj k2bp with δaj the change of aj resulting from
the change δθj .
The recurrent BPM is obtained from the BPM by Definition 6, through
summing over time in the time-unfolded network. Let i be a unit in the
recurrent network, and let θi be the set of incoming parameters to i. A
change δθi in θi results in a change δθit of all the dummy parameters θit of
units (i, t) in the time-unfolded network. The square norm of δθi in the
RBPM is, by definition (52), the sum over t of the square norms of δθit in
the backpropagated metric metric of the time-unfolded network.
So let us work out the backpropagated metric in the time-unfolded network. The time-unfolded unit (i, t) directly influences the time-unfolded
units (j, t + 1) for all edges i → j in the graph of the original network, and
27
it also directly influences the distribution πt at time t.
Thus, let δati be an infinitesimal change in the activity of time-unfolded
unit (i, t). Let δπt be the resulting change in the probability distribution πt ,
∂at+1
and δat+1
= ∂aj t δati the resulting change in the activity of time-unfolded
j
i
unit (j, t + 1). The BPM is obtained by backwards induction over t
δati
2
bp
:= kδπt k2nat +
δat+1
j
X
j
2
(58)
bp
The term kδπt k2nat is readily computed from Section 3.2: in the notation
P
above, the change in Eyt = j wjy atj from a change of activity in ati is
δEyt = wiy δati , so that (47) yields
kδπt k2nat = (δati )2
P
2
y πt (y)wiy
−(
P
2
y πt (y)wiy )
(59)
i.e., proportional to the πt -variance of wiy (in line with the fact that translating weights does not change output).
Since activities are
one-dimensional, the backpropagated metric is simply
2
proportional to δati , so that we have
2
δati
bp
2
X
=: mti δati
(60)
for some positive number mti , the backpropagated modulus [Oll13]. The definition (58) of the backpropagated metric thus translates as
mti
=
P
2
y πt (y)wiy
−(
P
2
y πt (y)wiy )
+
∂at+1
j
∂ati
j
!2
mt+1
j
(61)
(initialized with mTi = 0), in which one recognizes a source term from the
output at time t, and a term transmitted from t + 1 to t.
It is advisable to express the backpropagated metric using the variable Vit
∂V t+1
rather than ati (because the expression for ∂Vj t is simpler). The variables
i
V and a correspond bijectively to each other, and
their variations
are related
2
2
by δati = s′ (Vit )δVit so that δati bp = mti δati = m̃ti δVit with
m̃ti := mit s′ (Vit )2
(62)
from which we derive the induction equation for m̃, namely
m̃ti
=s
′
(Vit )2
P
2
y πt (y)wiy
−(
P
2
y πt (y)wiy )
in which we can now easily compute the
equation defining the recurrent network.
28
+
∂Vjt+1
∂Vit
X
j
∂Vjt+1
∂Vit
!2
m̃t+1
(63)
j
term from the evolution
For instance, for GLNNs we have Vjt+1 = Vjt +
P
t
i τijxt s(Vi )
so we find
∂Vjt+1
= 1i=j + τijxt s′ (Vit )
∂Vit
(64)
which, plugged into the above, yields the explicit equation (24) given in the
algorithm description.
Once the backpropagated modulus is known, the backpropagated metric
on the dummy parameters θit at each unit (i, t) of the time-unfolded network
∂at
is obtained by δθit bp := δati bp where δati = ∂θti .δθit is the variation of ati
i
resulting from a variation δθit . Thus
2
δθit
bp
=
mti
∂ati t
.δθ
∂θit i
!2
=
m̃ti
∂Vit t
.δθi
∂θit
!2
(65)
∂V t
where, as in the case of the RUOP metric above, the derivative ∂θit can be
i
obtained from the evolution equation defining the network. In components,
2
δθit bp is thus given by a matrix whose kk′ entry is
m̃ti
∂Vit ∂Vit
∂(θit )k ∂(θit )k′
(66)
where (θit )k denotes the k-th component of the incoming parameter θit to
unit i.
A parameter θi of the recurrent network influences all dummy parameters
t
θi for all t. The recurrent backpropagated metric is obtained by summing
the backpropagated metric over time as in (52). So in the end the recurrent
backpropagated metric for the incoming parameter θi to unit i is given by
the matrix
X
∂Vit ∂Vit
(i)
(67)
m̃ti
M̃kk′ =
∂(θit )k ∂(θit )k′
t
with (θit )k the k-th component of θit , and where the derivative is with respect
to the dummy parameter θit used only at time t.
For instance, in GLNNs, the incoming parameter to unit i is θi = (τjiy )j,y .
P t
at−1 using the dummy paramThe evolution equation Vit = Vit−1 + j τjix
t−1 j
eters yields
∂Vjt
t
∂τjiy
= 1xt−1=y at−1
j . This results in the expression given in the
algorithm of Section 2. In the end, for the GLNN parameter θ = (τjiy )j,i,y ,
the recurrent backpropagated metric is
kδθk2rbp =
XXX
i
j,j ′
t t
m̃t+1
i aj aj ′ δτjixt δτj ′ ixt
(68)
t
The structure of this metric is the same as for the RUOP metric above,
and the same remarks apply (see Section 3.4): incoming parameters to distinct units i are independent; parameters corresponding to distinct symbols
29
y 6= y ′ are independent for GNNs and GLNNs but not for RNNs; finally, the
transition parameters from different units j and j ′ incoming to the same unit
i are not independent, and the gradient ascent in this metric realizes, at each
unit i, a weighted least-square regression on the incoming signals from units
j to best match a desired activation profile given by the backpropagation
values.
3.6
Invariance of the algorithms
Amari [Ama98, AN00] pioneered the use of “invariant” algorithms for statistical learning that do not depend on a chosen numerical representation
(parametrization) of the parameter space of the model. Invariance can often
improve performance; for instance, in the standard RNNs in the experiments
below, replacing the standard inverse diagonal Hessian with the (invariant)
quasi-diagonal inverse brings performance of RNNs closer to that of GLNNs,
at very little computational cost.
The gradient ascent presented above is invariant by reparametrization of
the activities and by reparametrization of the incoming parameters to each
unit (but not by reparametrizations mixing incoming parameters to different
units, as the natural gradient is).
This stems from its construction using a metric which depends only on
the behavior of the network. For instance, using tanh instead of sigmoid
activation function and following the same procedure would result in an
algorithm with identical learning trajectories.
However, in practice three factors limit this invariance.
1. The invariance holds, in theory, only for the continuous-time gradient
trajectories. The actual gradient steps with non-zero learning rate are
only approximately invariant when the learning rate is small. Still, the
actual gradient steps are exactly invariant under affine reparametrizations of the parameters and activity (such as changing sigmoid into
tanh).
2. Parameter initialization is done by setting numerical values for the parameters in an explicit numerical representation. Changing parametrization obviously means changing the initial values in the same way. If
initialization is based on an intended parametrization-independent behavior at startup, as in Section 2, this is not a problem.
3. The dampening procedure for matrix inversion (the various ε terms
in Section 2) formally breaks invariance. Using a Moore-Penrose pseudoinverse (which is simply the limit ε → 0) does not solve this problem.
It would be nice to have a dampening scheme preserving invariance11 .
11
Here is a possibility for defining a matrix for the incoming parameters to a unit i,
which could be used as dampening the metric at i in an invariant way: Compute a copy
30
4
Preliminary experiments
Here we report a comparison of the performance of GLNNs and more traditional RNNs on some synthetic data examples: the “alphabet with insertion”
(Example 1 from the Introduction), synthetic music (Example 3), the distant XOR problem (Example 2), and finally the an bn problem (Example 4).
LSTMs are used as an additional benchmark.
GLNNs were trained with either the recurrent backpropagated metric or
the recurrent unitwise outer product metric, as described in Section 2.
The reference RNN was trained using traditional (but not naive) techniques as described below. For the distant XOR example, RNN performance
is known to be poor unless the “Hessian-free” technique is used [MS11], so we
did not test RNN on this example and instead directly compare performance
to [MS11].
Reference RNN training. The RNN used as a baseline is described in
Section 1.2.1. In particular, both this RNN and GLNNs use a softmax (6)
for the probability to produce a symbol x given the internal state of the
network.
RNN training is done via backpropagation through time. As plain backpropagation was too slow, for the parameters wiy the inverse diagonal Hessian (obtained from (18)) is applied to the gradient update, and the learning
rate for each ρix is inversely proportional to the frequency of symbol x in
the data (thus compensating for the number of terms making up the corresponding gradient, so that rare symbols learn as fast as frequent symbols12 ).
A method similar to RMSprop or Adagrad [DHS11], in which the learning
rate for each transition parameter is divided by the root mean square (RMS)
average over time of the gradient magnitude, is also reported in Table 1.
Initialization of the RNN parameters has been set along the same principles as for GLNNs, namely
w0y ← log νy ,
τii ← 1 − 1/i,
ρjy ←
P
1
(ujy − y′ ν̃y′ ujy′ )
2
(69)
with u and ν̃ as in Section 2, and with all other weights set to 0, where the
symbol frequencies νy and ν̃y are as in Section 2, and the ujy are independent
random variables uniformly distributed in [0; 1]. This way, at startup the
of the metric (RUOP or RBPM) but replacing the actual training sequence (xt ) with a
randomly generated sequence (e.g., uniform, or a perturbation of (xt )). More copies with
more random sequences can be used until one gets a non-degenerate metric. The resulting
metric can be multiplied by a small number and used as a dampening term. But this does
not solve all problems: for instance, if a unit i has no effect whatsoever on the output
given the current parameters, the corresponding metric will vanish. It seems difficult to
define a non-zero invariant metric in the latter situation.
12
If ρix is seen as the weight from an input unit activated when symbol x occurs, then
this is equivalent to scaling the input unit signals to a given L2 norm over time.
31
activation of each unit is given by a random linear combination of past
symbols with weights exponentially decreasing with time, with unit i having
a decay time of order i thanks to τii .
More combinations of models (RNN, GNN, GLNN) and training methods are reported in Table 1.
LSTMs. LSTMs [HS97] are included as an additional benchmark. For this
we have kept the same overall procedure and simply replaced each RNN cell
with an LSTM cell following Eqs. (7)–(11) in [Gra13], and modified the gradient accordingly. We kept the softmax output from the other models (also
as used in [Gra13]). The weights were initialized to uniform random values in [−0.1, 0.1] [Gra12, GSS03]. Network construction, network sizes, and
CPU time budget were identical to the other models, as described below.
Since plain gradient resulted in slow training, we have also included a variant described above for RNNs: using the diagonal Hessian for the writing
parameters w, and frequency-adjusted learning rates for the input symbols
(equivalent to rescaling the inputs). Still, training is quite slow and from
Table 1 it appears that LSTMs are not competitive in this setup13 , at least
for the computational time budget used here.
Regularization. When working with discrete alphabets, the problem arises
of having probability 0 for certain symbols in certain situations after training; if the trained model is used on a validation set, validation log-likelihood
can thus be very low. In our situation this is especially the case near the
beginning of the sequence: since the model is trained on only one training
sequence and has parameters for the activities at startup, it can frequently
learn to start in a specific configuration to reproduce the first few letters of
the training sequence. For this reason, a crude regularization procedure was
used: before computing log-likelihood of the validation sequence, the predic1
1
)πt + t+2
unif A
tion πt for the next symbol at time t was replaced with (1− t+2
with unif A the uniform distribution over the alphabet. (This kind of regularization has some backing from information theory.)
Experimental setup. The same overall procedure (construction of a random graph, learning rate control) has been used for both GLNNs and RNNs
as described in Section 2, following nearly identical implementations.
In each case, a single14 training sequence (xt ) is generated using the exact
synthetic model. Another, independent sequence (x′t ) is used for validation:
we report the log-likelihood (in base 2) of the validation sequence (x′t ) using
13
Good performance of LSTMs has been reported for one of the problems we use, the
a b problem [GSS03]. However this involved more samples and small values of n in the
training set. With these settings we were able to obtain similar results.
14
see footnote 2
n n
32
the GLNN or RNN trained on (xt ). The baseline for performance is the
number of random bits used by the exact synthetic model to generate (x′t ).
As a sanity check, we also report the performance of a well-known, efficient online text prediction method, context tree weighting (CTW): the
algorithm is presented with the concatenation of the training and validation
sequence, and we report the number of bits used to predict the validation
sequence after having read the training sequence.
The comparison between GLNNs and RNNs is made for identical computation time on the same machine, for a series of hyper-parameter settings
(network size and connectivity). Indeed, as RNNs and GLNNs have different parameter sets, direct comparison for the same network size is difficult.
Spanning different network sizes shows the performance each model can attain for a given time budget if the right hyper-parameters are used.
In each case, the size of the network was chosen to increase√from 4 units
to a maximum of 256 or 512 units by increments of a factor 2. For each
network size, we tested both a sparse network with connectivity d = 3 edges
per unit (including
√ a loop at each unit), and a “semi-sparse” network with
connectivity d = 2#A for GLNNs and d = #A for RNNs, where #A is
the alphabet size; this latter choice balances the various contributions to
algorithmic complexity (see Section 2). This way, RNNs can take advantage
of their lesser computational sensitivity to connectivity d.
For each hyper-parameter setting, the corresponding model was allowed
to learn for the same time (10 or 30 minutes depending on the example).
The experiments were run on a standard laptop computer with an Intel
Core i7-3720QM CPU at 2.60GHz 15 , using a straightforward implementation in C++.
The code for these experiments can be downloaded at http://www.yann-ollivier.org/rech/cod
Let us now discuss each example in turn.
Alphabet with insertions. The synthetic generative model is as follows.
The training sequence is the concatenation of 1000 lines, separated by a
newline symbol. Each line is obtained by writing the 26 lowercase letters of
the Latin alphabet, in the standard order, and then inserting (independently)
a sub-block after each letter with probability 1/26 for each letter. A subblock starts with an opening parenthesis, followed by the 10 digits from 0 to
9 (in that order), and ends with a closing parenthesis. After each digit in the
sub-block, with probability 1/5 a sub-sub-block is inserted, which consists
of an opening square bracket, nine random uppercase letters chosen from
A–Z, and a closing bracket. Thus a typical line might look like
ab(0123[WZPYCPEEH]456789[HYDVTWATR])cdefghijklmnopqrstuvwxyz
15
For technical reasons the experiments for LSTMs and the RMS variant of RNNs were
done on a slightly faster machine; an empirically adjusted scaling factor was applied to
the corresponding CPU time.
33
-80000
True model entropy
Sanity check (CTW)
RNN BP
GLNN RUOP
GLNN RBPM
Validation log-likelihood (bits)
-85000
-90000
-95000
-100000
-105000
-110000
0
200
400
600
800
1000
1200
Computation time (seconds)
1400
1600
1800
Figure 1: Validation log-likelihood on the alphabet with insertions example.
The validation sequence has the same law: the concatenation of 1000
independent such lines. Randomization of the innermost blocks prevents
rote learning.
GLNNs and RNNs with a variety of network sizes ranging from 4 to
512 units, as described above, were run for 30 minutes each on the training
sequence. The validation sequence log-likelihood is reported in Figure 1 and
Table 1.
GLNNs come more than ten times closer to the true model log-likelihood
than RNNs: the best validation log-likelihood for GLNNs is -89,126 bits
while that for RNNs is -96,099 bits, compared to -88,482 bits for the true
model. Such a difference of roughly 7,000 bits represents roughly 7 bits per
line of the training sequence. Note that the cost of representing a letter in
the alphabet is log(26)/ log(2) ≈ 4.7 bits: this would be the log-likelihood
difference, for each line of the training sequence, between a model that
resumes at the correct place in the alphabet after a sub-block insertion,
and one that resumes at a random letter.
This is confirmed by visual inspection of the models obtained after training. Indeed, since we train generative models, the trained network can be
used to generate new sequences, hopefully similar to the training sequence.
Doing so with RNNs and GLNNs reveals qualitative differences in the models learned, in line with the difference in performance: After a sub-block has
been inserted, GLNNs resume at the correct letter or sometimes one letter
off the correct position in the alphabet; on the other hand, RNNs seldom
34
resume at the correct position.
The remaining small difference in log-likelihood between GLNNs and
the true model can, from visual inspection, be attributed to various factors:
residual errors like occasional duplicated or omitted letters, or resuming
one letter off after an insertion, as well as arguably good generalizations of
the training sequence such as having more than one sub-block between two
letters or starting a new line with a sub-block.
There is no obvious pattern of dissimilar performance between sparse
and semi-sparse networks.
However, GLNNs are apparently quite sensitive to overfitting over time:
validation log-likelihood increases at first, then steadily decreases as parameter optimization progresses. This phenomenon is also present to a lesser
extent for RNNs, but only after much longer training times. Note that for
a given network size, GLNNs have more parameters (because each edge has
as many parameters as symbols in the alphabet A).
This illustrates the importance of using a validation sequence to stop
training of GLNNs.
One GLNN run exhibits wild variations of validation log-likelihood, for
unknown reasons (perhaps a badly invertible matrix M̃ ).
On the other hand, surprisingly, GLNNs are less sensitive to overfitting
due to a too large network size: while increasing network size past some
value results in worse performance for RNNs (lower curves on Figure 1), for
GLNNs it seems that the best validation log-likelihood over an optimization
trajectory stays the same for a wide range of network sizes.
Running RNNs for longer times only partially bridges the gap in performance: RNNs after 4 hours are still seven times farther from the true model
than GLNNs are after 30 minutes (with a gain of 2,810 bits of log-likelihood
for RNNs). After some time, RNNs slow down considerably or sometimes
exhibit the same overfitting phenomenon as GLNNs and their validation
performance decreases.
Overall, the “resume-after-insertion” phenomenon illustrated by this example is well captured by GLNNs.
Synthetic music. The next example is synthetic music notation, meant
to illustrate the intersection of several independent constraints. The training sequence is a succession of musical bars. Successive musical bars are
separated by a | symbol and a newline symbol. Each bar is a succession
of notes separated by spaces, where each note is made of a pitch (a,b,c,...)
and value (4 for a quarter note, 2 for a half note, 4. for a dotted quarter
note, etc.). In each bar, a hidden variable determines a harmony with three
possible values I, IV, or V. If the harmony is I, every pitch in the bar is taken
uniformly at random from the set (“chord”) {c,e,g}; pitches are taken from
{c,f,a} if harmony is IV, and from {g,b,d} if harmony is V. Harmonies in
35
True model entropy
Sanity check (CTW)
RNN BP
GLNN RUOP
GLNN RBPM
Validation log-likelihood (bits)
-16000
-18000
-20000
-22000
-24000
-26000
-28000
0
100
200
300
400
Computation time (seconds)
500
600
Figure 2: Validation log-likelihood on the synthetic music example.
successive bars follow a specific deterministic pattern: an 8-bar-long cycle
I-IV-I-V-I-IV-V-I as encountered in simple tunes. Finally, in each bar, the
successive durations are taken from a finite set of 5 rhythmic possibilities
(commonly encountered in waltzes), namely: 4-4-4; 2-4; 4.-8-4; 2.; 4-4-8-8.
Rhythm is chosen independently from pitch and harmony. See Example 3.
The training sequence is made of 2,700 musical bars. The validation
sequence is taken independently with the same law.
GLNNs and RNNs with a variety of network sizes ranging from 4 to
256 units, as described above, were run for 10 minutes each on the training
sequence. The validation sequence log-likelihood is reported in Figure 2 and
Table 1.
Only one RNN run beats the sanity check (CTW). There is a difference
of roughly 2,000 bits between the best RNN and best GLNN performance;
GLNNs come roughly three times closer to the true model.
Visual inspection of the output of the networks seen as generative models
confirms that this difference is semantically significant: GLNNs correctly
learn the rhythmic and harmonic constraints inside each bar, whereas RNNs
still display “mistakes”.
On the other hand, even GLNNs were not able to learn the underlying
8-bar-long harmonic progression, which was apparently approximated by
probabilistic transitions. This is reflected in the remaining gap between the
true model and GLNNs.
Running an RNN with backpropagation for a longer time (3 hours in-
36
stead of 10 minutes) only partially bridged the gap, only bringing RNN
an additional 604 bits in log-likelihood. Once more, visual inspection of
RNN output revealed a correct learning of the possible set of rhythms, but
imperfect learning of the harmonic constraints even inside each musical bar.
The pattern of decrease in validation log-likelihood because of overfitting
is present but less pronounced than for the alphabet-with-insertions example.
Still, on Figure 2 one can notice one GLNN run exhibiting a wild variation
of validation log-likelihood at some point. Once more this points out the
importance of using validation sets during GLNN training, although using
only one training sequence of relatively small size may also play a role here.
Distant XOR. The setting is taken from [MS11], after [HS97]; here we
recast it in a symbolic sequence setting. A parameter T is fixed (T =
100 below), which determines the length of the instances. The training
sequence is a concatenation of lines separated by newline symbols. Each
line is made of T ′ random bits preceded by whitespaces, where T ′ is taken
at random between T and 1.1T . Two of these random bits are preceded by a
special symbol X instead of a whitespace. The positions of these two special
symbols are taken at random from the intervals J0; T ′ /10J and JT ′ /10; T ′ /2J
respectively. At the end of each line, a symbol = is inserted and is followed
by a bit giving the XOR result of the two bits following the two X symbols.
Example 2 gives a typical training sequence.
The goal is to correctly predict the value of the final bit of each line. So
in the gradient computation an error term is included only for the bits to be
predicted, as in [HS97]. Namely, in the notation of Section 2, we set χt = 1
if and only if xt−1 is the symbol =.
For this problem, we did not run the reference RNN and directly compared to the best performance we found in the literature, in [MS11], using
“Hessian-free” second-order RNN training. The success rate reported in the
latter, for T = 100, is about 25% (proportion of runs achieving a classification error below 1% using at most 50,000 minibatches of 1,000 instances
each).
We ran eight distinct instances of the problem, each with a different
random training and validation sequence. Each such sequence was the concatenation of 10,000 lines as above with T = 100. We used a fully connected
network with 10 units. Optimization was run for 1,500 gradient passes over
the training sequence (amounting to roughly 12 hours of computation and
750 gradient steps for each of the writing and transition parameters, since
we alternate those). We discuss the results for training using the recurrent
BPM; the results using the recurrent UOP metric are extremely similar.
Figure 3 reports two measures of performance on the validation sequence:
the log-likelihood score for prediction of the final bit of each line (following
the score (4)), and the classification error (equal to 0 if the correct bit value
37
2000
60
0
50
-2000
40
-4000
30
-6000
20
-8000
10
-10000
0
0
100
200
300
400
500
Computation time (minutes)
600
700
Classification error (%)
Validation log-likelihood (bits)
log-likelihood
classification error
800
Figure 3: Validation log-likelihood and classification error on the distant
XOR problem using GLNNs.
is given a probability > 1/2 and to 1 otherwise—this is always bounded by
the log-likelihood error) expressed as a percentage.
The results are binary: each run either successfully achieves low error
rates after enough time, or does not perform better than random prediction.
4 out of 8 independent runs reached error rates below 1% within less
than 1,500 gradient passes over the training set, and 6 out of 8 within 2,000
gradient passes. The sample size is too small to tell for sure that this is
better than the success rate in [MS11]. Still, the algorithm is simpler and
uses fewer training examples.
Direct comparison of the algorithmic cost with the approach in [MS11]
is difficult, because for each gradient pass the latter can perform up to 300
passes of the conjugate gradient algorithm used in the implicit Hessian computation. For reference, in our approach, each run of the experiment above
(1,500 gradient passes on a training sequence of 10,000 lines) takes slightly
above 4h of CPU time on an Intel Core i7-3720QM CPU at 2.60GHz using
a straightforward C++ implementation (no parallelism, no use of GPUs).
an bn problem. In this problem, the training sequence is made of lines
separated by newlines. The first line is a block of n1 symbols a; the second
line is a block of n1 symbols b; the third line contains n2 a, the fourth line
contains n2 b, etc. See Example 4.
In this experiment, the block lengths n were taken at random in J1024; 2048J
38
True model entropy
Sanity check (CTW)
RNN BP
GLNN RUOP
GLNN RBPM
-50
Validation log-likelihood (bits)
-100
-150
-200
-250
-300
-350
0
100
200
300
400
Computation time (seconds)
500
600
Figure 4: Validation log-likelihood on the an bn example.
to build the training and validation sequences.
We used training and validation sequences made of only ten an bn blocks.
RNNs and GLNNs with sizes ranging from 4 to 64, as described above,
were run for 10 minutes each. For each independent run, a new random
training sequence and validation sequence was generated. The results are
reported in Figure 4 and Table 1.
The log-likelihood of a validation sequence under the true model is 10
bits for each block of a (choosing an integer n between 1024 and 2048),
after which the length next block of b is known and comes for free. Thus
the reference log-likelihood of the whole validation sequence (which contains
10 blocks of each) is 100 bits. However, from only 10 training samples as used
here, the exact distribution of the length n cannot reasonably be inferred;
a reasonable inference would be, for instance, a geometric law with mean
= 1536
somewhere in this interval. The geometric law with mean 1024+2048
2
has an entropy of about 12 bits instead of 10.
Thus, at best, one can expect a reasonable model to attain an entropy
of about 120 bits on the 10-instance-long validation set. On the other hand,
a model which would not catch the equality of the sizes of consecutive a
and b blocks would require twice as much entropy, i.e., about 240 bits for
the validation set. Indeed, the sanity check (CTW) has a log-likelihood of
-243.5 bits.
The best GLNN log-likelihood value obtained is -129.7 bits, while the
best RNN log-likelihood value is -222.4 bits.
39
Surprisingly, the best GLNN value was obtained with a network of size
4; a size-23 network came close second at -129.98 bits.
Not all GLNN runs find the optimum: there is a cluster of runs around
-230 bits, presumably corresponding to the model with independent lengths
for a and b blocks, and one run (with 64 units) provided aberrant validation
log-likelihood after some point because of overfitting.
Visual inspection of the output of the best trained GLNN runs, used as
generative models, shows that consecutive blocks of a and b indeed have the
same or very close lengths, with sometimes an error of ±1 on the length. This
imperfection would likely disappear with more than ten training sequences.
The kind of internal representation used by the GLNN to reach this
result is unclear, especially given the small network size: does it build a
kind of base-2 counter, does it take advantage of the analog nature of the
units’ activities, or something in between?
Influence of the various choices. The difference in performance between GLNNs and RNNs above results from various factors: choices in
model design (leakiness and gatedness) and in the training method (backpropagation or a Riemannian gradient). We now try to isolate these factors,
by testing various combinations of models (RNNs, GNNs and GLNNs) and
training methods.
In particular, it is possible to apply invariant training methods to RNNs.
The recurrent BPM and recurrent UOP metric are well-defined for RNNs.
However, contrary to GNNs and GLNNs, the parameters corresponding to
different symbols in the alphabet are not mutually orthogonal, and thus,
using them directly would result in a complexity quadratic in the alphabet
size, which we deem unacceptable. Therefore, we used the quasi-diagonal
reductions of these metrics, as defined in [Oll13]. This still provides training
methods that are invariant under reparametrization of the activity of each
unit.
Each model and training method was tested as described above, spanning various values of the hyperparameters (network size and connectivity).
For each method we report the best performance found over the hyperparameters.
The performance reported is the cumulative regret with respect to the
true generating model, a standard measure used in sequential prediction
contexts. It is defined as the difference between the log-likelihood of the
validation data sequence under the true model used to generate the data,
and the log-likelihood of the validation data sequence under the trained
model.
We also included three sanity checks for reference. Two are text compressors known for their performance (CTW as mentioned above, and the
file compressor bzip2), for which, to incorporate the effect of training, we
40
Cumulative regret (bits)
Alphabet Music an bn
Method
Sanity checks:
bzip2
Context tree weighting
Hidden Markov model
Non-invariant methods:
RNN with DH and FB
RNN with DH and RMS
GNN with DH and FB
GLNN with DH and FB
LSTM with plain gradient
LSTM with DH and FB
Invariant methods:
RNN with QDH and QDRUOP
RNN with QDH and QDRBPM
GNN with QDH and RUOP
GNN with QDH and RBPM
GLNN with QDH and RUOP
GLNN with QDH and RBPM
27206.1
18272.1
10212.0
6035.7
3611.0
4818.1
244.0
143.5
125.7
7616.8
6721.5
4059.0
5883.9
20073.1
11843.1
2898.9
2295.0
2688.7
3616.4
8796.1
5590.9
122.4
129.9
132.0
74.4
139.8
101.7
3372.8
3623.9
1759.0
2166.0
1011.9
644.2
1349.5
1224.3
1148.3
1596.0
1055.9
1055.9
89.5
106.0
71.6
115.4
29.7
30.0
Table 1: Cumulative regret (bits) of the learned model with respect to the
true generative model, over the validation data sequence. Best value found
over the hyperparameters in the allocated CPU time budget is reported.
Models are: RNN, GNN, GLNN, LSTM, and HMM. Training methods for
the writing parameters: inverse diagonal Hessian (DH), or quasi-diagonal
inverse Hessian (QDH). Training methods for the transition parameters:
frequency-adjusted backpropagation (FB), root mean square gradient rescaling (RMS), recurrent backpropagated metric (RBPM), recurrent unitwise
outer product metric (RUOP), and the quasi-diagonal reduction of the latter (QDRBPM and QDRUOP).
41
report the number of bits used to compress the concatenation of the training and validation sequences minus the number of bits used to compress the
training sequence alone.
The third sanity check is a hidden Markov model (HMM), trained using
a variety of network sizes as for the neural networks.16 The comparison with
HMMs is especially interesting, since these are a classical tool for modelling
sequential data.
The “classical” training method is as described above for RNNs: diagonal
inverse Hessian for the writing parameters w, and backpropagation for the
transition parameters; for the latter, parameters like ρiy (for RNNs) or τijy
(for GNNs and GLNNs) related to a given symbol y have a learning rate
divided by the frequency of y in the training sequence (“frequency-adjusted”
backpropagation, which compensates for the number of terms making up
the corresponding gradient, and, for RNNs, is equivalent to scaling the input
signals). Pure backpropagation was tested but is simply too slow.
The results are collected in Table 1.
From this table it is clear that an invariant method is the first step to
improve performance: RNNs trained with an invariant method beat GNNs
and GLNNs trained with a non-invariant method.
Still, the leaky aspect of GLNNs seems to be necessary to bring the
best performance in problems with very long dependencies (the alphabet
with insertions and the an bn example). On the other hand, on the problem
where dependencies are most local (synthetic music), all network models
achieve quite comparable results if trained with an invariant method.
Conclusions and perspectives
The viability of GLNNs with Riemannian training to capture complex algorithmic dependencies in symbolic data sequences has been established.
Metrics inspired by a Riemannian geometric viewpoint, allow us to write
invariant algorithms at an algorithmic cost comparable to backpropagation
through time for sparsely connected networks.
These metrics bring down the necessary number of gradient steps to a
few hundreds in the various examples studied. This approach seems to work
with small training samples. Better than state-of-the-art performance has
been obtained on difficult synthetic problems.
16
Details as follows. Training is done by the expectation-maximization algorithm. The
network is obtained, as for the neural networks, by taking an oriented random graph with
a given number of edges per node (including loops); this number of edges per node is set
to the alphabet size, because this gives an algorithmic complexity similar to that of the
neural networks. Initialization of the transition probabilities is by a Dirichlet(1/2, . . . , 1/2)
(i.e., Jeffreys) prior on the edges from a node. Initialization of the production probabilities
is done by multiplying the actual frequency of each symbol in the sequence to be modelled,
by a random uniform([0; 1]) number.
42
In the experiments, the importance of invariance seems to supercede that
of model choice: in our tests, any model with an invariant training algorithm
did better than any model with a non-invariant one.
More experiments are needed to investigate the isolated effect of each feature of this training procedure (memory effect in the definition of GLNNs,
gatedness, and the choice of metric). Other issues in need of investigation are the influence of parameter initialization (especially if some expert
knowledge on the time scale of dependencies in the data is available) and
a better, invariant dampening procedure. It would also be interesting to
acquire insight into the dynamical behavior of GLNNs (ergodicity, multiple
equilibrium regimes, etc.) and how it is affected by training. Furthermore,
the Riemannian approach can in principle be extended to more complex
architectures: testing Riemannian methods for LSTMs seems promising.
Finally, scalable Riemannian training algorithms should be developed
for a fully online “lifelong learning” setting where there is a single training
sequence which grows with time and where it is not possible to fully store the
past states and signal, so that backpropagation through time is excluded.
A
Parameter initialization, the linearized regime,
and integrating effects in GLNNs
Let us examine the dynamics of a GLNN, and in particular the linearized
regime (the regime in which the connection weights are small). This will
provide some insight into the time-integrating effect of the model, and also
suggest relevant initializations of the parameter values before launching the
gradient ascent, as presented in the algorithm above.
P
In the GLNN evolution equation Vjt+1 = Vjt + i τijxt ati let us isolate the
contributions of i = j and of the always-activated unit i = 0. Substituting
atj = s(Vjt ) and at0 ≡ 1 we get
Vjt+1 = Vjt + τjjxt s(Vjt ) + τ0jxt +
X
τijxt ati
(70)
i6=0, i6=j
Since s(Vjt ) is an increasing function of Vjt , the contribution i = j provides a feedback loop: if τjjx is negative for all x, then the feedback will be
negative, whereas positive τjjx would result in perpetual increase of Vjt if
the other contributions are ignored. Meanwhile, τ0jxt provides the reaction
of unit j to the signal xt .
For instance, if we set τjjx = −α for all x with α > 0, τ0jx = β for all x,
and all other weights τijx to 0, the dynamics is
Vjt+1 = Vjt − αs(Vjt ) + β
(71)
which has a fixed point at Vjt = V̄ := s−1 (β/α), i.e., atj = β/α (assuming
β/α lies in the range of the activation function s). The linearized dynamics
43
around this fixed point (Vjt close to V̄ ) is
Vjt+1 − V̄ ≈ (1 − αs′ (V̄ )) Vjt − V̄
(72)
so that if 1 − αs′ (V̄ ) < 1 this fixed point is attractive.
A more interesting choice is to let
τjjx = −α,
τ0jx = β + ερx
(73)
with small ε, where ρx is chosen to that the average of ρ over the data xt
is 0. Then, the value of V t as a function of t and the data can be found by
induction using the linearized dynamics:
Vjt ≈ V̄ + ε
X
t′ <t
′
(1 − µ)t−t ρxt′
(74)
where
µ := αs′ (V̄ )
(75)
namely, the activation level Vjt is a linear combination of the past values
of the signal xt , with weights exponentially decreasing with time at rate
(1 − µ).
This provides insights into reasonable values of the parameter leading
to interesting internal dynamics, to be used at the start of the learning
procedure. Indeed, negative values of α would lead to unstability, whereas
positive values of α presumably stabilize the network. However, values of
α above 1/ sup s′ (= 1 for tanh activation) will provide too much feedback,
resulting in non-monotonous V t+1 as a function of V t and an oscillating
behavior. Indeed we have found that setting α = 1/(2 sup s′ ), i.e.,
τjjx = −α = −1/2
(76)
(for tanh) for all j and x at startup, yields a good behavior of the network.
With τjjx and τ0jx as above, the value of (1−µ) controls the effective time
window of the integrating effect: data much older than t − t′ ≫ µ1 has little
weight. Thus µ1 presumably gives the order of magnitude of the distances
t−t′ for which the model can reasonably be expected to capture correlations
in the data (at least at startup, since µ will change during learning).
The value of µ can be directly controlled through β via µ = αs′ (V̄ ) =
′
αs (s−1 (β/α)): for the tanh activation function, this yields
q
β = − α(α − µ)
(77)
which is used to set β from an arbitrary choice for µ. We have found that
using different values of µ for different units yields good results. We have
used
µj = 1/(j + 1)
(78)
44
for unit number j (starting at j = 1); this yields a characteristic time of
order j and seems to perform well.
Finally, the “reading rates” ρx are taken at random independently for
each unit j in the following way. The value of ε must be small enough to
ensure that Vjt stays close to Vj (otherwise the linear regime assumption is
P
′
unjustified), namely, that the sum ε t′ <t (1 − µ)t−t ρxt′ stays small. If each
ρ is roughly of size 1, the sum is ε/µ so taking ε somewhat smaller than µ
is reasonable. We have used
µ
(79)
ε=
4
which apparently yields good performance. Finally, ρx is taken at random
uniformly in [0; 1] for each symbol x (independently for each unit j), and
then shifted by a constant so that the average of ρxt over the training data
P
xt is 0 (namely, the constant
νx ρx is removed from each ρx where νx is
the frequency of symbol x in the training data)17 .
The other transition weights τijx , with i 6= 0, i 6= j, were set to 0 at
startup.
The explicit initialization values described here are specific to the tanh
activation function; however, the reasoning extends to any activation function.
B
Derivative of the log-likelihood: Backpropagation through time for GLNNs
Let (xt )t=0,...,T −1 be an observed sequence of T symbols in the alphabet A.
Here we compute the derivatives of the log-probability that a GLNN prints
(xt ) with respect to the GLNN parameters, via the standard backpropagation through time technique.
Given a training sequence x = (xt ), let Pr(x) be the probability that
the model prints (x0 , . . . , xT −1 ). Here, for simplicity we assume that all
symbols in the sequence have to be predicted (i.e., χt ≡ 1). The algorithm
in Section 2 gives the formulas for the general case.
Proposition 11 (log-likelihood derivative for GLNNs). The
derivative of the log-probability of a sequence x = (xt )t=0,...,T −1 with respect
to the parameters of a gated leaky neural network is given as follows.
Setting
∂ log Pr(x)
(80)
Bjt :=
∂Vjt
17
The choice to use a uniform random variable in [0; 1] rather than, e.g., Gaussian
random variables, is justified by the feedback mechanism. Indeed since the activation
function s ranges in [0; 1], the feedback term −αs(Vjt ) is bounded. If an unbounded signal
ρxt can occur at each step, it may take a long time to stabilize. Empirically, using Gaussian
rather than bounded random variables seems to decrease performance, confirming this
viewpoint.
45
we have the backpropagation relation
Bit =
Bit+1
+ s′ (Vit ) wixt −
X
πt (y)wiy +
y
X
j
τijxt Bjt+1
(81)
(initialized with BjT := 0). In particular Bj0 gives the derivative with respect
to the initial values Vj0 at time 0.
The derivatives with respect to the writing weights are
∂ log Pr(x) X t
ai (1xt =y − πt (y))
=
∂wiy
t
(82)
and the derivatives with respect to the transition weights are
∂ log Pr(x) X
1xt=y ati Bjt+1
=
∂τijy
t
(83)
These relations include the always-activated unit i = 0, ai ≡ 1.
The meaning of the partial derivative with respect to Vjt is the following:
P
if, in the equation Vjt+1 = Vjt + j τijxt ati defining GLNNs, we artificially
P
introduce a term ε ≪ 1 at unit j at time t, namely, Vjt+1 = Vjt + j τijxt ati +ε
for a single unit at a single time, and let the network evolve normally except
for this change, then the value of log PT changes by εBjt + O(ε2 ).
Proof. Given a training sequence (xt )t=0,...,T −1 of length T , let P0 := 1 and
Pt+1 := πt (xt )Pt
(84)
so that PT is the probability of printing (x0 , . . . , xT −1 ).
By definition of πt we have
log Pt+1 = log Pi +
X
ati wixt
i
− log
X P
e
at w
i i iy
y
!
(85)
Let us compute the infinitesimal variations of these quantities under an
infinitesimal variation δw, δτ of the parameters. Ultimately we are interested in the variation of log PT , to perform gradient ascent on the parameters.
By a first-order Taylor expansion, the variation of log Pt+1 satisfies
δ log Pt+1 = δ log Pt +
X
ati δwixt +
−
y
πt (y)
wixt δati
i
i
X
X
X
i
46
ati δwiy
+
X
i
wiy δati
!
(86)
and rearranging and substituting
δati = s′ (Vit )δVit
where
s′
(87)
is the derivative of the activation function, this yields
δ log Pt+1 = δ log Pt +
X
ati
δwixt −
i
+
X
wixt −
i
X
X
πt (y)δwiy
y
!
πt (y)wiy s
y
′
!
(88)
(Vit )δVit
Consequently, the variation δ log Pt of log Pt can be expressed in terms
of the variation of log Pt−1 , the variations of the parameters w and τ , and
the variations of the values Vjt−1 at time t − 1.
Let us assume, by backward induction, that we can write the differential
of log PT with respect to the parameters, as
δ log PT =: δ log Pt +
X
Bit δVit +
X
t
Ciy
δwiy +
t
Dijy
δτijy
(89)
i,j,y
i,y
i
X
For t = T this is satisfied with B T = C T = D T = 0.
Thus Bit represents the backpropagated value at unit i at time t, and C
and D will cumulatively compute the gradient of log PT with respect to the
parameters w and τ , namely:
and
∂ log Pr(x)
0
= Ciy
∂wiy
(90)
∂ log Pr(x)
0
= Dijy
∂τijy
(91)
and moreover Bj0 will contain the derivatives with respect to the initial levels
Vj0 .
P
Using the evolution equation Vjt+1 = Vjt + i τijxt ati we find
δVjt+1 = δVjt +
X
δτijxt ati +
X
τijxt s′ (Vit )δVit
(92)
i
i
Using these relations to go from time t + 1 to time t in (89), namely,
expressing δ log Pt+1 in terms of δ log Pt and expanding V t+1 in terms of V t ,
we find
t+1
t
Ciy
= Ciy
+ 1xt =y ati − πt (y)ati
(93)
t+1
t
Dijy
= Dijy
+ 1xt =y ati Bjt+1
and
Bit = Bit+1 + s′ (Vit ) wixt −
X
πt (y)wiy +
y
0 and D 0 follow.
from which the expressions for Ciy
ijy
47
(94)
X
j
τijxt Bjt+1
(95)
C
Fisher metric for the output distribution πt
Let us compute the Fisher norm of the variation δπ of π resulting from a
P
change δEyt in the values of Eyt = i ati wiy . (Such a change in E can result
from a change in the writing weights w or the activities a; this will be used
to compute the metric on the writing weights and the transition weights,
respectively.) The effect of a change δE t on log πt is
δ log πt (y) =
X ∂ log πt (y)
y′
∂Eyt ′
δEyt ′
(96)
and the norm of this δπt in Fisher metric is
kδπt knat = Ey∼πt (δ log πt (y))2
= Ey∼πt
(97)
X ∂ log πt (y) ∂ log πt (y)
y ′ ,y ′′
∂Ey′
∂Ey′′
δEyt ′ δEyt ′′
(98)
By a standard formula for exponential families of probability distributions we find:
∂ log πt (y)
(99)
= 1y=y′ − πt (y ′ )
∂Ey′
so that
Ey∼πt
"
#
∂ log πt (y) ∂ log πt (y)
= Ey∼πt (1y=y′ − πt (y ′ ))(1y=y′′ − πt (y ′′ ))
′
′′
∂Ey
∂Ey
(100)
= πt (y ′ )(1y′ =y′′ − πt (y ′′ ))
(101)
(this is also18 the Hessian of − log πt (y) with respect to the values E t ). Consequently, the Fisher metric for πt , expressed in terms of the variations δEyt ,
is
X
X
kδπt k2nat =
πt (y)(δEyt )2 −
πt (y)πt (y ′ )δEyt δEyt ′
(102)
y
y,y ′
References
[Ama98]
Shun-ichi Amari. Natural gradient works efficiently in learning.
Neural Comput., 10:251–276, February 1998.
[AN00]
Shun-ichi Amari and Hiroshi Nagaoka. Methods of information
geometry, volume 191 of Translations of Mathematical Monographs. American Mathematical Society, Providence, RI, 2000.
Translated from the 1993 Japanese original by Daishi Harada.
18
because for exponential families, the Hessian of log π(y) does not depend on y
48
[APF00]
Shun-ichi Amari, Hyeyoung Park, and Kenji Fukumizu. Adaptive method of realizing natural gradient learning for multilayer
perceptrons. Neural Computation, 12(6):1399–1409, 2000.
[Bri90]
John S. Bridle. Alpha-nets: A recurrent ’neural’ network architecture with a hidden Markov model interpretation. Speech
Communication, 9(1):83–92, 1990.
[BSF94]
Y. Bengio, P. Simard, and P. Frasconi. Learning long-term dependencies with gradient descent is difficult. Neural Networks,
IEEE Transactions on, 5(2):157–166, 1994.
[DHS11]
John C. Duchi, Elad Hazan, and Yoram Singer. Adaptive subgradient methods for online learning and stochastic optimization.
Journal of Machine Learning Research, 12:2121–2159, 2011.
[Gra12]
Alex Graves. Sequence transduction with recurrent neural networks. Preprint, http://arxiv.org/abs/1211.3711 , 2012.
[Gra13]
Alex Graves. Generating sequences with recurrent neural networks, 2013. Preprint, http://arxiv.org/abs/1308.0850v5 .
[GS11]
Lise Getoor and Tobias Scheffer, editors. Proceedings of the
28th International Conference on Machine Learning, ICML 2011,
Bellevue, Washington, USA, June 28 – July 2, 2011. Omnipress,
2011.
[GSS03]
Felix A. Gers, Nicol N. Schraudolph, and Jürgen Schmidhuber.
Learning precise timing with lstm recurrent networks. The Journal of Machine Learning Research, 3:115–143, 2003.
[HB95]
Salah El Hihi and Yoshua Bengio. Hierarchical recurrent neural
networks for long-term dependencies. In David S. Touretzky,
Michael Mozer, and Michael E. Hasselmo, editors, Advances in
Neural Information Processing Systems 8, NIPS, Denver, CO,
November 27-30, 1995, pages 493–499. MIT Press, 1995.
[HS97]
Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735–1780, 1997.
[Jae02]
Herbert Jaeger. Tutorial on training recurrent neural networks,
covering BPTT, RTRL, EKF and the ‘‘echo state network’’ approach. Technical Report 159, German National Research Center
for Information Technology, 2002.
49
[KGGS14] Jan Koutník, Klaus Greff, Faustino J. Gomez, and Jürgen
Schmidhuber. A clockwork RNN. In Proceedings of the 31th International Conference on Machine Learning, ICML 2014, Beijing, China, 21-26 June 2014, volume 32 of JMLR Proceedings,
pages 1863–1871. JMLR.org, 2014.
[Mar10]
James Martens. Deep learning via hessian-free optimization. In
Johannes Fürnkranz and Thorsten Joachims, editors, Proceedings of the 27th International Conference on Machine Learning
(ICML-10), June 21-24, 2010, Haifa, Israel, pages 735–742. Omnipress, 2010.
[MS11]
James Martens and Ilya Sutskever. Learning recurrent neural
networks with Hessian-free optimization. In Getoor and Scheffer
[GS11], pages 1033–1040.
[MS12]
James Martens and Ilya Sutskever. Training deep and recurrent neural networks with Hessian-free optimization. In Grégoire
Montavon, Geneviève B. Orr, and Klaus-Robert Müller, editors,
Neural Networks: Tricks of the Trade, volume 7700 of Lecture
Notes in Computer Science, pages 479–535. Springer, 2012.
[Oll13]
Yann Ollivier. Riemannian metrics for neural networks I: feedforward networks. Preprint, http://arxiv.org/abs/1303.0818 ,
2013.
[PB13]
Razvan
Pascanu
and
Yoshua
Bengio.
ing natural gradient for deep networks.
http://arxiv.org/abs/1301.3584 , 2013.
RevisitPreprint,
[RHW87] D. E. Rumelhart, G. E. Hinton, and R. J. Williams. Learning internal representations by error propagation. In D. E. Rumelhart
and J. L. McClelland, editors, Explorations in the Microstructure
of Cognition, volume 1 of Parallel Distributed Processing, pages
318–362. MIT Press, Cambridge, MA, 1987.
[RMB07]
Nicolas Le Roux, Pierre-Antoine Manzagol, and Yoshua Bengio.
Topmoumoute online natural gradient algorithm. In John C.
Platt, Daphne Koller, Yoram Singer, and Sam T. Roweis, editors, NIPS. Curran Associates, Inc., 2007.
[SMH11]
Ilya Sutskever, James Martens, and Geoffrey E. Hinton. Generating text with recurrent neural networks. In Getoor and Scheffer
[GS11], pages 1017–1024.
50
| 9cs.NE
|
arXiv:1802.02183v1 [cs.CV] 5 Jan 2018
Enhanced Image Classification With Data
Augmentation Using Position Coordinates
Avinash Kori
Ganapathy Krishnamurthi
Balaji Srinivasan
[email protected]
[email protected]
[email protected]
Abstract—In this paper we propose the use of image pixel
position coordinate system to improve image classification
accuracy in various applications. Specifically, we hypothesize
that the use of pixel coordinates will lead to (a) Resolution
invariant performance. Here, by resolution we mean the spacing
between the pixels rather than the size of the image matrix. (b)
Overall improvement in classification accuracy in comparison
with network models trained without local pixel coordinates.
This is due to position coordinates enabling the network to
learn relationship between parts of objects, mimicking the
human vision system. We demonstrate our hypothesis using
empirical results and intuitive explanations of the feature maps
learnt by deep neural networks. Specifically, our approach
showed improvements in MNIST digit classification and beats
state of the results on the SVHN database. We also show that
the performance of our networks is unaffected despite training
the same using blurred images of the MNIST database and
predicting on the high resolution database.
Index Terms — Position Digit classification, MNIST,SVHN,
Deep Learning, CNN
I. I NTRODUCTION
Image classification using images acquired with different
pixel spacing or scales is a difficult problem and is in
general solved using very deep neural networks trained with
a large amount of data. However, in many problems of
interest such large data is not readily available and alternate
approaches are desirable. Typically, CNNs are trained using
a multi-resolution approach wherein images down sampled
or blurred to various resolutions, these whole lot of images
with different resolutions are used to train network and their
predictions are used for final classification. However, this
leads to large networks and a cumbersome training process.
The capability to classify images at various resolutions is,
however, achieved almost seamlessly by the human vision
system. One model of human perception is based on sets of
neurons representing hierarchical representations of objects
starting with basic shapes. For instance, certain cells in the
human retina are sensitive to horizontal and vertical lines
which can be used to compose basic shapes[1]. CNNs mimic
this model by using layers of filter to construct hiearchical
representations and the presence of a local coordinate system
in the input would definitely help learn these representations
better by learning relationships between parts of objects.
One key difference between typical CNNs and the human
vision system is that position information is more immediately present in the latter due to the geometric arrangement
of neurons; In the case of usual neural networks position
information can only be learnt (and that too, implicitly) via
training as, typically, any image is unrolled into a single
column vector before training. Similarly, the level of explicit
coordinate position encoded into even CNNs is limited by
the size of the filter.
In this article, we explore the advantages of mimicing the
geometric feature of the human vision system by explicitly
encoding a normalized coordinate system into the CNN. We
hypothesize that introducing a local pixel coordinate system
on the images, i.e. a XY cartesian coordinate system and
using it as one of the input channels would enable the
network to learn the relationships between parts of objects.
We expect, in turn, that this recognition would enhance
performance in both image classification as well in multiresolution problems. In a simple sense, one would expect
the network to learn Euclidean distances between significant
features in an object. As a result, an improvement in classification accuracy can be expected. One would also expect an
invariance to pixel spacing, as recognition should be based
on relative distances rather than on the actual resolution of
the images.
In our work we test the use of local pixel coordinates using
commonly used digit datasets like MNIST[2], SVHN[3](
MNIST is a benchmark dataset for an images of segmented
handwritten digits, each with 28x28 pixels. It includes 60,000
training examples and 10,000 testing examples, Figure 3. The
SVHN dataset is a real life application oriented dataset for
digits which includes images with 32x32 pixels, with 73257
digits for training and 26032 digits for testing, Figure 4),
and show empirical evidence in terms of improvement in
classification accuracy over regular approaches. In the case
of SVHN, we demonstrate the capability to beat state of the
art results using position coordinates as input to a deep neural
network. We also examine feature maps for less complex
networks like the one used for digit recognition and we
demonstrate how these networks combine coordinate data and
shape information.
II. M ETHODOLOGY
We discuss below our approach to encoding position
information and creating tests for multi-resolution efficacy.
a) Encoding Position Information: The CNN is fed with a 3
channel input, the first channel comprises of the actual input
image (gray-scale image). The remaining two channels are
for encoding spatial position along X and Y dimension with
respect to Cartesian coordinate frame of reference(auxillary
frames). These position encoded image frames are spatial
coordinate grids normalized between 0 and 1. Figure: 1 .
illustrates this.
conventional network without using position information[4].
In other words, the error rate was reduced by approximately
a factor of 2 due to the use of position information.
Figure 1. a. Standard input image used — b. Shows the layers of position
information with an input image — c. Our input image encoded with position
information
b) Validating multi-resolution behavior : In this analysis
we trained the network with images of one specific resolution
and in testing phase, we used images of entirely different
resolution. In this study of multi-resolution behavior, the
previously shown method is followed however the coordinate
grid is also resized to the same resolution as that of an
image. In many applications involving images, we train the
model using images of a fixed resolution; but for most of
the practical and real life applications there will always be
varying resolutions, frame of reference which need to be
detected, due to which most of the models fails. Since the
downsampled coordinate information is obviously image
independent, position encoding is a potentially promising
approach to tackle such scenarios. In the next subsection
we discuss how the results were improved just by encoding
position information with input image.
Figure 2. MNIST network — architecture details: convolutional layer :
5x5x32— pooling1: 2x2—convolutional layer : 5x5x64— pooling2: 2x2
III. E XPERIMENTS AND R ESULTS
Our hypothesis was that encoding position information
explicitly would improve both image classification as well
as multi-resolution performance. We tested our hypothesis
through the relative performance of our network in comparison to conventional networks on the following tasks:
1) Classification accuracy on the MNIST dataset
2) Classification accuracy on the SVHN dataset
3) Digit translation test
4) Multi-resolution test
A. Test On MNIST Dataset
A schematic of the network architecture followed for
MNIST digit classification is shown in Figure 2. In this
network we use two sets convolutional layers with maxpooling layer followed by two fully connected layers with
softmax layer to obtain the posterior probability of an input
digit. On MNIST handwritten digits dataset, with 45,000
images for training, 5,000 for validation and 10,000 images
for testing, with our auxiliary input (XY-Position matrix)
channels, We were able to achieve 99.84% accuracy. This was
about 0.2% greater than maximum accuracy achieved using a
Figure 3. examples of MNIST digits
B. Test On SVHN Dataset
In the SVHN classification task, we set an upper limit
for number of digits in the image as 5, and remove all
other images with more than 5 digits from dataset and didn’t
use them either in training or for testing. We use VGG16
(16 layered deep convolutional network which is used for
image classification tasks) convolutional architecture [5] for
feature extraction and we use attention based modeling [6]
for predicting existing digits in an image. The input image is
appended with XY position information and all the digits are
predicted simultaneously. The entire model was trained endto-end without data augmentation. In attention mechanism
is basically weighted superposition of higher level features
extracted by VGG16 convolutional neural network.
Figure 4. examples of SVHN digits
Training images were obtained by down sampling the
original 28x28 image to 14x14 and again up sampling it
to the original resolution. Basically, the network is trained
on low resolution images and tested on high resolution
images (original data). In this case, adding XY position
information, in addition to gray channel, increases digit
classification accuracy by 2-3%. With this kind of data our
network with encoded position information gives an accuracy
about 98.26% while the standard network gives an accuracy
of about 96.32%. Once again, the error rate was reduced
by approximately a factor of 2 due to the use of position
information.
D. Digit Translation Test
Figure 5. Network architecture followed for SVHN classification
Typically, in CNNs[7] used for digit recognition tasks,
features learnt by network after a few convolutional layers
would correspond to entire digit itself. However, the learnt
features could also be combination of multiple digits. The
attention model used in our network has learnable parameters
which helps in providing specific weightage to feature maps
in predicting final output. Our results indicate that the use
of attention model mechanism has a positive impact on the
classification accuracy.
We were able to reduce the the digit recognition MSE by
1.2%, by including the XY position channel information. On
a real life dataset like SVHN, with auxiliary (XY-position
matrix) channels, we achieved around 95.44%(average of all
5 digits) accuracy with 1.83% classification error. Whereas
the current state of the art result is 94.6%(average of all 5
digits) accuracy with 1.64% classification error[8].
In this test we show how translation invariance can be
obtained by using by employing XY position information
with an image. (which is equivalent to labeling pixels in an
image) our model could predict digits correctly, regardless of
finite translations. This would not be possible to accomplish
merely with a standard input or even with max pooling
(without position information) Some of these results are
shown in the figure below.
C. Resolution Change Test
Our next endeavor was to show that using XY position
information makes the network resolution invariant. Networks were trained with MNIST digits whose resolution was
artificially degraded by blurring and tested with digits at
the original resolution. The classification results indicate that
implementing position matrix aided in achieving exceedingly
good outputs (improvement by 2-3%) despite mismatch in
training and testing resolution.
Figure 7. Example images which are used in translation test
The results obtained so far on MNIST, SVHN databases
and the translational invariance experiment indicate that the
XY coordinate information improves prediction accuracy. We
assert that using XY inputs could also help to reduce the
number of parameters required for achieving a certain level of
accuracy. The following numerical experiments demonstrate
that using XY position input can help reduce the number of
network weights.
IV. D ISCUSSION AND R ESULTS
A. What does this position matrix learn?
Figure 6. Variation in resolution among training and testing images
To see what exactly could be learnt from the position
matrix, we implemented convolutional variational autoencoders([9] & [10]). These auxillary layer (XY position
matrix) helps in removing useless information(noise) for
any given data. Figure 8 & figure 9 shows input and
corresponding output for convolutional variational autoencoders experiment. The reconstructed outputs are revealing
as the reconstruction of the position matrix contains the
digits superimposed on it.
B. What is the significance of using XY Position Information?
Figure 8. Normalized input given to Variational Auto Encoder
Figure 11. The layer wise loss of edge information can be clearly seen in
the above image.
Figure 9. Variational Auto Encoder Output (After Reconstruction)
Figure 10. Joint position matrix(Sum of reconstructed x and y position
matrix)
As the input and output images are normalized to unity,
the black pixel values correspond to lower probability values
and white pixel correspond to higher probability information.
As the XY position matrix was auxiliary input to the network
there are no specific features which helps in reconstruction
of these matrices. But at the same time, as mentioned earlier
we can see that the reconstructed position matrices are
related to the corresponding input digit image. If we consider
the sum of the reconstructed position matrices as the joint
probabilities of X and Y matrix (As all image pixels are
bounded between 0 and 1, and pixel probability distribution
is linear(uniform to certain extent)) then the input image
is clearly reconstructed. As a result of which, we conclude
that the position matrix helps in learning joint probability
distribution of the pixels constituting the digits.
In standard convolution operation of any image, only the
local (bound to kernel size) relative information is captured.
It takes many number of layers to capture the global relative
pixel information with some loss of edge information at each
convolutional step.
But by adding an XY position layer as a channel information,
capturing global pixel relative information becomes easier,
and the information is procured at a very early stage. As
the position information is encoded within the input image
channel, the network learns some sort of features or patterns
and their relative positions which makes it possible to capture
global information at the very initial layers.
Mathematically, for any image, it will take about
n−s
[
]
(1)
k
steps of convolution to capture whole global information(without pooling). [where, n is the image dimension(assuming image is square), s is the stride size,and k
is the kernel size]
Mathematically Convolution is defined as:
XXXX
Y (i, j) =
I(i + ki, j + kj) ∗ K(ki, kj), (2)
i
j
ki
kj
where I is the image and K is the kernel Convolution on an
image with XY position stacked is
XXXX
Y (i, j) =
I(i + ki, j + kj, x, y) ∗ K(ki, kj),
i
⇒
j
ki
kj
XXXX
i
j
ki
I(i + ki, j + kj) ∗ K(ki, kj)+
kj
δ(x, y, i + ki, j + kj)
The term
δ(x, y, i + ki, j + kj)
(3)
saves the encoded XY position information, technically saving pixel-to-pixel relative information which is preserved
after successive convolutions, and can enable the network
to learn the relative position between features. Another observation was that the feature maps learnt by networks using
XY position information was less redundant as compared to
feature maps of standard image input.
maps shown in the bottom row(with position info.). Using
position information seems to help the network learn features
that are normally learnt at deeper layers. Consequently using
position helps to reduce the depth and consequently the
number of weights making the network computationally less
expensive.
V. C ONCLUSIONS AND F UTURE WORK
We have demonstrated that implementing position
coordinates as an auxillary input results in enhanced
accuracy as well improved mutliresolution performance. Our
results consistently provide an improvement over state of the
art results on publicly available classification datasets. We
aim to extend this study from gray images to colored images
and test the result on CIFAR-10 and STL-10 datasets. We
also expect that improved performance could be obtained by
integrating the position information into a complex number
i.e. by replacing XY position information with Z (i.e X+iY).
We expect that implementing position information via the
complex domain will helps in reducing the number of
parameters as well as number of layers.
R EFERENCES
Figure 12. Without position information
Figure 13. With position information
The figure 12 and 13 shows feature maps obtained after the
first convolutional layer. It can be observed that quite a few of
the feature maps on the top image(without position info.) are
fully blank (corresponding to redundant information), but our
network using position information was still able to extract
some unique features, as is evident from some of the feature
[1] “Foundationsofvision.stanford.edu. (2017). foundations of vision
chapter 5: The retinal representation. [online] available
at:
https://foundationsofvision.stanford.edu/chapter-5-the-retinalrepresentation/ [accessed 27 oct. 2017].,”
[2] “Mnist handwritten digit database, yann lecun, corinna cortes and
chris burges [online] available at: http://yann.lecun.com/exdb/mnist/
[accessed 27 oct. 2017].,”
[3] “Ufldl.stanford.edu. (2017). the street view house numbers (svhn)
dataset. [online] available at: http://ufldl.stanford.edu/housenumbers/
[accessed 27 oct. 2017].,”
[4] “Benenson,
r.
(2017).
classification
datasets
results.
[online]
rodrigob.github.io.
available
at:
http://rodrigob.github.io/are we there yet/build/
classification datasets results.html/ [accessed 31 oct. 2017].,”
[5] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556, 2014.
[6] Y. Kim, C. Denton, L. Hoang, and A. M. Rush, “Structured attention
networks,” arXiv preprint arXiv:1702.00887, 2017.
[7] “[ebook] iccv2015 tutorial on convolutional feature maps [online]
available at: http://kaiminghe.com/iccv15tutorial/iccv2015 tutorial
convolutional feature maps kaiminghe.pdf [accessed 27 oct. 2017].,”
[8] P. Sermanet, S. Chintala, and Y. LeCun, “Convolutional neural networks applied to house numbers digit classification,” in Pattern Recognition (ICPR), 2012 21st International Conference on, pp. 3288–3291,
IEEE, 2012.
[9] A. Makhzani, J. Shlens, N. Jaitly, I. Goodfellow, and B. Frey, “Adversarial autoencoders,” arXiv preprint arXiv:1511.05644, 2015.
[10] C. Doersch, “Tutorial on variational autoencoders,” arXiv preprint
arXiv:1606.05908, 2016.
| 1cs.CV
|
Typical Approximation Performance for Maximum Coverage Problem
Satoshi Takabe,1, 2, ∗ Takanori Maehara,2 and Koji Hukushima3
1
arXiv:1706.08721v2 [cond-mat.dis-nn] 7 Feb 2018
Department of Computer Science, Nagoya Institute of Technology,
Gokiso-cho, Showa-ku, Nagoya, Aichi, 466-8555, Japan
2
RIKEN Center for Advanced Intelligence Project 15F,
1-4-1, Nihonbashi, Chuo-ku, Tokyo, 103-0027, Japan
3
Graduate School of Arts and Sciences, The University of Tokyo,
3-8-1 Komaba, Meguro-ku, Tokyo 153-8902, Japan
(Dated: February 8, 2018)
This study investigated typical performance of approximation algorithms known as belief propagation, greedy algorithm, and linear-programming relaxation for maximum coverage problems on
sparse biregular random graphs. After using the cavity method for a corresponding hard-core
lattice–gas model, results show that two distinct thresholds of replica-symmetry and its breaking
exist in the typical performance threshold of belief propagation. In the low-density region, the
superiority of three algorithms in terms of a typical performance threshold is obtained by some
theoretical analyses. Although the greedy algorithm and linear-programming relaxation have the
same approximation ratio in worst-case performance, their typical performance thresholds are mutually different, indicating the importance of typical performance. Results of numerical simulations
validate the theoretical analyses and imply further mutual relations of approximation algorithms
PACS numbers: 75.10.Nr, 89.70.Eg, 89.20.Ff, 02.60.Pn
I.
INTRODUCTION
Approximation algorithms, which are important for
hard optimization problems, have attracted researchers’
interest because, for NP-hard optimization problems, one
encounters difficulty when solving optimal solutions exactly in polynomial time of the problem size. Since the P
versus NP problem arose, development and performance
analyses of approximation algorithms have persisted as a
central issue of computer science and operations research.
Two performance evaluations of approximation algorithms exist: worst-case performance and typical (or
average-case) performance. As described in this paper, we specifically examine the latter mainly using
statistical–mechanical methods, although the former has
been investigated mainly in the literature of theoretical
computer science [1]. Worst-case performance is defined
by the pair of an optimization problem and its approximation algorithm. An approximation ratio is then defined by the maximal ratio of an optimal value to an
approximation value over all instances if the problem is
a maximization problem (and vice versa, otherwise). It
is important to provide strict performance guarantee of
the algorithm, although the worst-case instance is sometimes pathological. However, the typical performance is
defined as the average performance of approximation algorithm for a given optimization problem over randomized instances. It is sometimes useful for practical use.
It sheds light on properties of optimization problems and
approximation algorithms in a perspective that is different from worst-case analysis.
∗
E-mail: s [email protected]
An interesting point of the typical property of optimization problems is found in a close relation to the
spin-glass theory in statistical physics. Extensive studies
of various computational problems have revealed that the
concept of replica symmetry (RS) and its breaking (RSB)
in spin-glass theory reflects average computational complexity [2] and structure of the solution space [3] of the
problems.
Typical performance of approximation algorithms often exhibits a phase transition on typical goodness or
accuracy of approximation, which is called threshold phenomena in the literature of theoretical computer science [4]. Some approximation algorithms for minimum
vertex covers (min-VC) are good examples of threshold phenomena. The problem is defined on a graph.
The randomized problem is characterized using a random graph ensemble with c being the average degree.
From a statistical–mechanical perspective, the problem
on Erdös-Rényi random graphs has the RS-RSB threshold at c = e = 2.71 . . . [5]. Moreover, three approximation algorithms have been investigated: belief
propagation (BP) (or message passing) [6], greedy leafremoval algorithm [7], and linear-programming (LP) relaxation [8, 9]. Approximation algorithms other than BP
naively have no direct connection to the spin-glass theory. Nevertheless, for Erdös-Rényi random graphs, these
algorithms have the same performance threshold as the
RS-RSB threshold. However, subsequent studies of general random graphs [10] indicate that their thresholds
are not equivalent for some random graphs because of
their graph structure. Using statistical–mechanical techniques, typical performance has been studied for variants
of leaf removal [11–15] and relaxation technique such as
LP relaxation [16, 17] and semidefinite-programming relaxation [18]. These studies have revealed not only typi-
2
cal approximate performance itself but also a suggestive
connection to typical properties of optimization problems, random graph structure, and the spin-glass theory.
As described in this paper, we examine the unweighted
maximum coverage (max-COV) problem defined in the
next section. Although the max-COV belongs to the class
of NP-hard, it has several practical applications such as
pan and scan problems [19], multi-topic blog watch [20],
and text summarization [21]. Our main purpose is to
examine the typical performance of approximation algorithms for the problem and to compare their performance
thresholds. We specifically examine three approximation
algorithms: belief propagation, greedy algorithm, and LP
relaxation. Some statistical–mechanical methods applied
to both analytical and numerical analyses together with
mathematical rigorous discussions clarify typical performance of those algorithms and a suggestive mutual relation among approximation algorithms.
This paper is organized as follows. In the next section,
we define details of the max-COV and its approximation
algorithms. As a random graph ensemble, biregular random graphs are also defined. In Section III, a hard-core
lattice–gas model for the problem is introduced. Its BP
equations are obtained based on Bethe–Peierls approximation. In Section IV, we present a study of the typical
performance of BP using the RS cavity method. Calculation of the spin-glass susceptibility provides the threshold
below which BP typically approximate max-COV with
high accuracy. In Section V, the greedy algorithm is analyzed based on a mean-field rate equation of its deletion
process. Using the obtained solution, the typical performance threshold of the greedy algorithm is evaluated. In
Section VI, LP-relaxed approximate values are evaluated
rigorously. The theorem is proved using the weak duality theorem. In Section VII, we describe some numerical
results which support the validity of theoretical analyses presented in the previous sections. We also execute
some additional simulations to consider the typical performance of a modified greedy algorithm and randomized
rounding of LP relaxation. These results provide suggestive relation between approximation algorithms. The
last section is devoted to a summary and discussion of
the results.
II. MAX-COVER PROBLEM AND
APPROXIMATION ALGORITHMS
A.
Maximum coverage problem
The max-COV is defined as follows: Let S be a set
of M elements and S = {S1 , · · · , SN } be a collection of
subsets of S. For a given positive integer K(≤ N ), the
problem is to choose at most K subsets to maximize the
total number of elements in the union of chosen subsets.
An example of an instance with M = 4 and N = 3 is
shown at the left-hand side of Fig. 1. Given that K = 2,
one can select subsets S1 and S3 such that all elements
FIG. 1. Example of the maximum coverage problem. (Left)
An example with M = 4 and N = 3. S1 , S2 , and S3 represent subsets of S = {1, 2, 3, 4}. (Right) A bipartite graph
corresponding to the left example. Each vertex in V1 stands
for a subset in the left example.
are included in the union of the subsets.
Converting an instance to a bipartite graph, one obtains an integer-programming (IP) representation of the
max-COV. Let G = (V1 , V2 , E) be a bipartite graph
where each edge in E connects to a vertex in V1 and
a vertex in V2 . Vertices in V1 and V2 are labeled respectively by i ∈ {1, . . . , N } and a ∈ {1, . . . , M }. V1
corresponds to a collection of subsets S, whereas V2 represents S. Each edge (i, a) is set if subset Si includes
element a. For Fig. 1, the left example is converted to
a bipartite graph in the right-hand side. We then introduce binary variables {xi } and {ya }, respectively, to
V1 and V2 . Also, xi is set to one if vertex i (or subset
Si ) is selected and zero otherwise. Similarly, ya is set to
one if vertex a is connected to a selected vertex (or element a belongs to the union of chosen subsets) and zero
otherwise. The problem is therefore represented by the
following IP problem:
Max.
M
X
ya ,
a=1
subject to ya ≤
X
xi
∀a ∈ V2 ,
i; (i,a)∈E
N
X
i=1
xi ≤ K,
xi ∈ {0, 1} ∀i ∈ V1 , ya ∈ {0, 1} ∀a ∈ V2 .
(1)
The inequality in the second constraint can be replaced
with an equality sign because the problem is a maximization problem.
As described in this paper, typical-case property of the
max-COV is analyzed by randomizing its instance. Then,
we introduce ρx = K/N and assume that ρx ∈ [0, 1] is
constant. For random bipartite graphs, we specifically
examine (L, R)-biregular random graphs where degrees of
each vertex in V1 and V2 respectively denote L and R. Using this simple random graph ensemble, our statistical–
mechanical analyses reveal interesting properties of the
problem.
To take the large graph limit, the number of vertices
in V2 is rescaled as M = αN with a constant factor α.
We assume that a random graph is sparse, i.e., L and
R are constant with respect to N . For the randomized
3
max-COV, we define an average optimal cover ratio over
random graphs with cardinality N + M and a given ρx
by
M
X
Max.
ya ,
a=1
M
1 X opt
ρy (ρx ; N ) =
ya (G, ρx ),
M a=1
(2)
Approximation algorithms
We introduce three approximation methods for maxCOVs. The first algorithm is belief propagation (BP).
The recursive equations called BP equations are derived
from Bethe–Peierls approximation for a spin system corresponding to a given optimization problem. For systems on trees, graphs with no cycles, the Bethe–Peierls
approximation and BP are exact. In general, however,
there exist cycles in a graph yielding correlations between
variables or spins, which results in inexact estimation of
solutions (or configurations). BP on graphs with some
cycles is regarded as an approximation method. It is
called loopy BP [22].
The second one is a simple greedy algorithm. At
each step, this algorithm has the following procedure:
(i) choose one vertex named i with the maximum degree in V1 , (ii) delete vertices neighboring to vertex i
from V2 , and (iii) update V1 to V1 \i and return to (i) if
|V1 | > N − K. This simple algorithm gives an approximation ratio of 1 − 1/e [23]. The problem cannot be
approximated within this ratio unless P=NP [24].
The last is linear programming (LP) relaxation. An
integer programming problem including the max-COV is
relaxed to LP problems by replacing integral constraints
with real constraints. The LP-relaxed max-COV is given
III.
X
subject to ya ≤
xi
∀a ∈ V2 ,
i; (i,a)∈E
N
X
where {yaopt (G, ρx )} represents optimal solutions in V2
and (· · · ) represents an average over random bipartite
graphs with size N + M . Its limiting value to N → ∞
is denoted by ρy . We simply call it the average optimal
cover ratio.
B.
as
(3)
xi ≤ K,
i=1
xi ∈ [0, 1] ∀i ∈ V1 , ya ∈ [0, 1] ∀a ∈ V2 .
Actually, LP approximation value gives the upper bound
of the problem. Because LP problems are solvable exactly in polynomial time, LP relaxation is a widely used
approximation technique. The approximate solution obtained by LP relaxation usually involves non-integers.
One must round those numbers appropriately to obtain
an approximate integral solution for the IP problem.
Here, we consider randomized rounding [25]. Using the
obtained approximation solution {xLP
i }, one selects vertex i ∈ V1 to set I ⊂ V1 with probability xLP
i /K up
to K vertices. Then, the rounded solution {xLPr
} is set
i
to xLPr
= 1 for i ∈ I and xLPr
= 0 otherwise. The
i
i
rounded approximation value is readily calculated from
{xLPr
}. In terms of the worst-case performance, LP rei
laxation and its randomized rounding in expectation has
the same approximation ratio as the greedy algorithm.
As described in this paper, we study the typical performance of these approximation algorithms. It is evaluated by approximate values averaged over randomized
max-COVs defined in the last subsection. Similar to the
average optimal cover ratio, the average cover ratio is
defined as the average ratio of the approximate value to
the cardinality M of vertex set V2 . It is regarded as exhibiting good typical performance if the average cover
ratio obtained by an approximation algorithm is equal
to the average optimal cover ratio in the large-N limit.
The main aim of this paper is to evaluate the typical
performance threshold of ρx below which an approximate algorithm exhibits good typical performance with L
and R fixed. Evaluation of the approximation algorithms
is accomplished by comparing their typical performance
thresholds.
BP EQUATIONS FOR MAX-COV
As explained in this section, BP equations for max-COV are derived based on the statistical–mechanical model for
the problem. We set particles on vertices in V1 and V2 , which respectively occupy vertex i and a if xi = 1 and ya = 1.
The hard-core lattice-gas model for max-COVs on bipartite graph G is then naively given as the following partition
function.
!
! M
!
M
N
X
X
X
X
Y
X
Ξ0 (µ; G) =
exp µ
ya H K −
xi
H
xi − ya ,
(4)
x∈{0,1}N y∈{0,1}M
a=1
i=1
a=1
i∈∂a
Therein, H(x) = 1 (x ≥ 0), 0 (x < 0) is the Heaviside step function and ∂a = {i ∈ V1 | (i, a) ∈ E} stands for a set
of neighbors of vertex a ∈ V2 . In the partition function, µ represents a chemical potential for particles on V2 . One
4
can construct BP equations for this model. However, it is inconvenient for practical use because of the constraint
PN
i=1 xi ≤ K.
Therefore, we use the following alternative partition function of
! M
!
N
M
X
X
X
X
Y
X
0
Ξ(µ; G) =
exp µ
xi + µ
ya
H
xi − ya ,
(5)
a=1
i=1
x∈{0,1}N y∈{0,1}M
a=1
i∈∂a
where µ0 = µ0 (µ, ρx ; G) is a chemical potential for particles on V1 . The ratio of µ0 with respect to µ is defined as
κ=−
µ0
.
µ
(6)
This parameter is regarded as a Lagrange multiplier for the constraint on the number of selected vertices. The
appropriate value of κ must satisfy the condition given by
*
+
N
1 X
ρx (µ, κ) = ρx , ρx (µ, κ) ≡
xi
,
(7)
N i=1
µ
where h. . . iµ is a grand-canonical average with a given µ. To consider ground states, we take the large-µ limit with
parameter κ fixed.
First, we construct BP equations for Eq. (5). Using the
Bethe–Peierls approximation, the single spin probability
Pi (x) which xi takes x is
Y
Pi (x) ' Zi−1 e−µκx
Pa→i (x), ,
(8)
ya under the condition xi = x. Similarly, single spin
probability Pa (y) that ya takes y reads
!
Pa (y) '
Za−1 eµy
Y
−1 −µκx
Pi→a (x) ' Zi→a
e
H
x∂a
a∈∂i
where ∂i = {a ∈ V2 | (i, a) ∈ E} is a set of neighbors of
vertex i ∈ V1 . Z∗ is a normalization factor hereinafter.
Pa→i (x) represents the marginal probability of x∂a\i and
X
X
xi − y
i∈∂a
Y
Pi→a (x), (9)
i∈∂a
where Pi→a (x) is the probability of xi taking x on the
cavity graph G\a. These probabilities satisfy the following recursive relations under the same approximation.
Pa→i (x),
(10)
b∈∂i\a
−1
Pa→i (x) ' Za→i
X
eµy
y
X
H x +
x∂a\i
To take the large-µ limit later, it is convenient to introduce cavity fields {hia } and {ĥai } defined respectively
by Pi→a (x) ∝ exp(µhia x) and Pa→i (y) ∝ exp(µĥai y).
Then, BP equations for cavity fields are given as
X
hia = −κ +
ĥbi ,
(12)
b∈∂i\a
ĥai = −
Y
1
1
1
.
ln 1 −
−µ
µ
1+e
1 + eµhja
X
xj − y
j∈∂a\i
Y
Pj→a (x).
(11)
j∈∂a\i
and {ξˆa }, Eqs. (8) and (9) then read
ξi = −κ +
X
ĥai ,
(14)
a∈∂i
"
#
Y
1
1
1
ξˆa = − ln 1 −
.
µ
1 + e−µ
1 + eµhia
(15)
i∈∂a
(13)
j∈∂a\i
By rescaling the single spin probability as Pi (x) ∝
exp(µξi x) and Pa (y) ∝ exp(µξˆa y) using local fields {ξi }
One can estimate the single spin probability by solving
BP equations (12) and (13) as the loopy belief propagation.
5
IV.
Taking the large-µ limit, the average cover ratio for ρx <
1/R is obtained as
REPLICA-SYMMETRIC SOLUTION
In this section, typical performance of BP is studied
using the RS cavity method based on the simplest RS
ansatz. The RS ansatz assumes that cavity fields {h}
and {ĥ} are independent random variables respectively
following probability distributions P (h) and P̂ (ĥ). For
biregular random graphs, it is apparent that these distributions have no variance because of the absence of
fluctuation of degree in V1 and V2 . We therefore introduce the cavity fields h and ĥ on (L, R)-biregular random
graphs. Using BP equations (12) and (13), they satisfy
the following RS cavity equations in the large-N limit as
h = −κ + (L − 1)ĥ,
1
1
ĥ = − ln 1 −
.
µ
(1 + e−µ )(1 + eµh )R−1
(16)
(17)
(18)
(19)
According to Eq. (7), for a given ρx , the local field ξ is
determined as
eµξ
= ρx .
1 + eµξ
(20)
Then, using Eqs. (18) and (20), the appropriate parameter κ is represented as
κ = Lĥ −
1
ln r,
µ
(26)
This relation indicates that each vertex in V2 is connected
to at most one chosen vertex in V1 . However, if ρx is
larger than 1/R (R ≥ 2), then the solution x remains
constant for sufficiently large µ. In this case, a simple
solution ρRS
= 1 is obtained by Eq. (24) because the
y
average cover ratio in Eq. (26) touches to one at ρx =
1/R.
Before closing this section, we discuss the stability of
the RS solutions using spin-glass susceptibility defined as
X X
1
(hxi ya iµ − hxi iµ hya iµ )2 .
χSG (µ) =
N +M
i∈V1 a∈V2
Similarly, the local fields on (L, R)-biregular random
ˆ which satisfy
graphs are set to ξ and ξ,
ξ = −κ + Lĥ,
1
1
ˆ
.
ξ = − ln 1 −
µ
(1 + e−µ )(1 + eµh )R
ρRS
y = Rρx .
(21)
where r = ρx /(1 − ρx ).
By substituting Eqs. (16) and (21) to Eq. (17) and
by introducing x = e−µĥ , one obtains the self-consistent
equation as
(27)
Another representation of χSG using cavity fields is
known based on the fluctuation–dissipation theorem [26].
As shown in [27], in the case of biregular random graphs,
it reads
!2
∞
X
X
∂
ĥ
ai
d
,
χSG (µ) '
λ , λ = E
∂
ĥ
bj
d=0
j∈∂a\i; b∈∂j\a
(28)
where E[· · · ] represents an average over vertices in V1 and
random graphs. The cavity fields are mutually correlated
in general. Actually, BP cannot converge any more [28] if
the susceptibility diverges. In this sense, the divergence
of χSG not only means the instability of the RS solution,
but also poor typical performance of BP as an approximation algorithm in the static sense. In the low-density
region where ρx < 1/R holds, it is evaluated as
2
r(1 − x)
,
λ = (L − 1)(R − 1)
1 + rx
→ (L − 1)(R − 1)r2
(µ → ∞).
(29)
(22)
The RS-RSB threshold in the large-µ limit therefore
reads
1
p
ρRS
.
(30)
x =
1 + (L − 1)(R − 1)
The order of this solution changes depending on the value
of ρx . The solution x vanishes as µ becomes large if
ρx < 1/R holds. Using Taylor expansion of the left-hand
side of Eq. (22), it is estimated as
Otherwise, the other threshold ρRS
x (> 1/R) satisfies
"
#2
0
∗
ρRS
x (1 − x )
(L − 1)(R − 1)
= 1,
(31)
0
∗
(1 − ρRS
x )(1 + x )
(1 + rx)
R−1
(1 − x) =
1
.
1 + e−µ
e−µ
x=
+ O(e−2µ ).
1 − (R − 1)r
0
where x∗ is a solution of
(23)
Using Eq. (19), an average density of particles on V2 reads
0
ρRS
x
∗
1+
0 x
1 − ρRS
x
!R−1
(1 − x∗ ) = 1.
0
*
lim
M →∞
M
1 X
ya
M i=1
+
=
µ
−R
1 − (1 + rx)
1 − (1 + rx)−R + e−µ
= Rρx + O(e−µ )
(24)
(25)
(32)
RS
In Fig. 2, RS-RSB thresholds ρRS
on (3R, R)x and ρx
regular random graphs are shown as a function of R. Except for the case R = 1, the RS-RSB thresholds separate
RS and RSB regions. The RSB region remains for a finite
ρx while they converge to zero in the large-R limit.
6
the algorithm, i.e., ρgy = Rρx . However, the algorithm
typically mistakes selections of vertices in V1 and underestimates the average cover ratio. Results show that
ρgx ≤ ρRS
x where the equality holds for R = 1, indicating
that BP is better than the greedy algorithm in terms of
the typical performance threshold.
1
ρRS
x
′
ρRS
x
0.8
RS
ρx
0.6
0.4
RSB
0.2
0
VI.
RS
1
2
3
4
5
R
6
7
8
9
10
0
FIG. 2. RS-RSB thresholds ρRS
(squares) and ρRS
(cycles)
x
x
on (3R, R)-regular random graphs.
V.
TYPICAL ANALYSIS OF GREEDY
ALGORITHM
In this section, typical behavior of the greedy algorithm on (L, R)-biregular random graphs is investigated
in the low-density region of ρx < 1/R. The corresponding RS solution indicates that one can choose vertices in
V1 without overlapping their neighbors. It is therefore
sufficient to analyze the fraction of vertices in V1 with
maximum degree during the deletion process. Here we
analyze a rate equation that is used frequently for analyses of similar greedy algorithms [4, 29]. Let V (T ) be the
expected number of vertices with maximum degree in V1
at T -th step of the algorithm. By the definition of the
algorithm in sec. II B, we find that
N V (T )
N V (T + 1) = N V (T ) − 1 − L(R − 1)
+ O(N −1 ),
N − RT
(33)
where the assumption L = O(1) is used. By introducing
v(t) = V (T )/N and t = T /N , we obtain the following
differential equation in the large graph limit.
dv(t)
L(R − 1)
= −1 −
v(t).
dt
1 − Rt
(34)
Under the initial condition v(0) = 1, the solution reads
L(R−1)
R
(L − 1)(R − 1)(1 − Rt)
v(t) =
LR − L − R
An LP-relaxed value of the max-COV on any biregular
graph is evaluated exactly using the LP duality [30] as
follows.
Theorem 1 An LP-relaxed value of the max-COV on
any (L, R)-biregular graph is LK if K ≤ 1/R holds.
Proof. xi = K/N (1 ≤ i ≤ N ), ya = RK/N (1 ≤ a ≤ M )
is a feasible solution of LP-relaxed max-COV (3). The
value of the cost function is then RKM/N = LK.
The Lagrangian function of Eq. (3) is written as
!
M
N
X
X
L(x, y, p, q) =
ya + p K −
xi
a=1
+
i=1
M
X
!
qa
a=1
X
xi − ya
,
(37)
i∈∂a
where p ∈ R and q ∈ RM . Using the weak duality theorem, one finds the following inequalities.
max min L(x, y, p, q) ≤ min max L(x, y, p, q)
x,y
p,q
p,q
x,y
≤ max L(x, y, p, q).
x,y
(38)
Consequently, maxx,y L(x, y, p, q) is an upper bound of
the LP-relaxed value. Because Eq. (37) is represented by
!
N
M
X
X
X
L(x, y, p, q) =
−p +
qa xi + (1−qa )ya +pK,
i=1
a∈∂i
a=1
(39)
where a solution xi = ya = 1 (1 ≤ i ≤ N, 1 ≤ a ≤ M )
realizes
P the maximum of the function with p, q satisfying
−p+ a∈∂i qa ≥ 0 and 1−qa ≥ 0(1 ≤ a ≤ M ). Assuming
that qa = q holds for 1 ≤ a ≤ M , one finds that
max L (x, y, p, q) = −pN + LN q + M − M q + pK. (40)
x,y
− (1 − Rt)
. (35)
Let ρgx be a threshold below which vertices with the maximum degree are left at the end of the algorithm. Considering that t represents a fraction of chosen vertices in
V1 , we find v(ρgx ) = 0, that is,
ρgx =
ANALYSIS OF LP RELAXATION
o
R
1 n
1 − [(L − 1)(R − 1)]− LR−L−R .
R
(36)
If ρx < ρgx , then no chosen vertex in V1 has overlapped
neighbors in V2 resulting in good typical performance of
On the right-hand side, setting (p, q) = (M/N, K/N )
results in LK, which indicates that the upper bound of
the LP-relaxed max-COV on (L, R)-biregular graph is
LK. Because the approximate solution denoted above
achieves this bound, the LP-relaxed value is equivalent
to the bound.
This theorem claims that ρy = Rρx (ρx ≤ 1/R), suggesting that LP relaxation typically finds good approximate values in the RS regime. The LP-relaxed value
is equivalent to ρRS
in the high-density regime where
y
ρx > 1/R.
7
NUMERICAL RESULTS
0.9
0.7
0.6
Average cover ratio
To examine the validity of theoretical analyses on the
typical performance of approximation algorithms, the average cover ratio ρy is evaluated using several methods.
We employ a BP-guided decimation (BPD) algorithm [22] as a variant of loopy BP to obtain a feasible solution satisfying all the constraints. The algorithm
fixes a value of a variable based on a solution of BP equations (11) until all variables are fixed to either zero or one.
When a variable is fixed, BP equations are updated by
applying the fixed value. If the number of variables in
V1 fixed to one reaches to K, the remainders of x are
immediately fixed to zero to satisfy the constraint. An
approximate value of the algorithm is thus a function of
the parameter κ, which depends on inputs, i.e., µ, K and
a graph. Here, the BP equations with µ = 20 are solved
iteratively up to 150 steps, which enables the algorithm
to fix a variable practically even if the iterations cannot
reach to the RS fixed point. The parameter κ is tuned
to maximize an approximate value while the detail will
be reported elsewhere. In addition, the RS estimation of
RS
average cover ratio ρRS
y in the RS regime ρx < ρx is used
for comparison to BPD though its typical performance is
possibly evaluated directly as in [32].
The theoretical result presented in section VI is used
for LP relaxation, which is valid for arbitrary biregular
graphs. Average approximate values of the greedy algorithm are estimated numerically with N = 103 .
We also evaluate the average optimal value using the
replica exchange Monte Carlo (EMC) method [33]. Results show that single-spin flip updates of the model (4)
takes a long relaxation time for equilibration even using the exchange MC, indicating the existence of deep
valleys of the free energy. Consequently, it is necessary
to accelerate equilibration for a system with sufficiently
small chemical potential µ. To avoid such slow relaxation, we consider an alternative lattice–gas model on
bipartite graph G represented as
Ξ1 (µ; G)
=
X
x∈{0,1}N
exp µ
M
X
a=1
!!
θ
X
i∈∂a
xi
δ K,
N
X
!
xi
,
i=1
(41)
0.04
0.5
0.03
0.4
0.02
0.01
0.3
0.2
A.
BPD (N = 400)
EMC (N = 200)
EMC (N = 400)
Greedy (N = 1000)
RS/LP
0.8
∆ρy
This section presents description of some numerical
results for validation of the theoretical analyses. Here,
we set (L, R) = (9, 3) as an example. Then, the RSRSB threshold and the typical performance threshold of
the greedy algorithm are evaluated as ρRS
= 0.2 and
x
ρgx = (1 − 2−4/5 )/3 = 0.1418 . . . . Biregular random
graphs are generated based on implementation of the
configuration model [31]. At least 400 random graphs
are used to take a random graph average.
ρy
VII.
ρgx
0.1
0.12
0.14
ρRS
x
0.16
0.18
0.2
ρx
0
0.22
0.1
0.15
0.24
ρx
0.2
0.26
0.25
0.28
0.3
FIG. 3. Maximum cover ratio ρy obtained by some approximation algorithms as a function of ρx . Average optimal cover
ratios are estimated for the cardinality N = 200 (squares) and
400 (triangles) by the replica exchange Monte Carlo (EMC)
method, for N = 1000 by the greedy algorithm (circles), and
for N = 400 by the BP-guided decimation (BPD) algorithm
(cross marks). The solid line is the RS solution, equivalent
to the LP relaxed value. Two vertical lines respectively represent the analytical performance threshold ρgx of the greedy
algorithm and the RS-RSB threshold ρRS
x . (Inset) The difference ∆ρy between the RS optimal cover ratio and numerical
estimations by corresponding approximation algorithms as a
function of ρx .
where θ(x) takes one if x > 0 and zero otherwise, and
δ(x, y) is Kronecker’s delta. In this
P model, variables y
are eliminated because ya = θ( i∈∂a xi ) holds in the
large-µ limit. The ground states of the alternative model
therefore correspond to the optimal solutions of the maxCOV on the same graph. Moreover, single-spin flip updates in the model are substantially equivalent to multispin updates in the original model (4), which makes the
relaxation time to equilibrium states markedly short. Because optimal values are invariant by replacing inequalPN
ity constraint i=1 xi ≤ K to equality one, the equality constraint is adopted and Kawasaki dynamics for a
density-conserved system [34] is applied.
The results are presented in Fig. 3. Numerical results
are compatible to the RS estimation nearly up to the RSRSB threshold ρRS
x = 0.2 for the BPD and EMC, and up
to the typical performance threshold ρgx predicted analytically for the greedy algorithm. As shown in the inset
of Fig. 3, the average cover ratio by the greedy algorithm deviates from the average optimal ratio above ρgx ,
while other estimates split above ρRS
x , indicating that the
Bethe–Peierls approximation and LP relaxation are no
longer appropriate. In fact, as shown in Fig. 4, the loopy
BP for N = 400 fails its convergence above ρx ' 0.195,
which may imply the existence of dynamical one-step
RSB phase. It is expected that BPD finds approximate
solutions with good accuracy if a corresponding loopy
BP converges to a fixed point. Otherwise, decimation in
BPD is performed based on a wrong estimation by the
8
1
0.3
Greedy (N = 1000)
0.8
0.25
Greedy (N = 5000)
0.2
Modified (N = 1000)
Theory
Modified (N = 5000)
rg
pconv
0.6
0.15
0.4
0.1
0.2
0
0.1
0.2
0.3
0.4
ρx
0.5
0.6
0.7
0.8
FIG. 4. Probability of successfully convergent samples in the
loopy BP for (9, 3)-biregular random graphs as a function of
ρx . The dotted vertical lines represents RS-RSB thresholds
shown in Fig. 2.
Bethe–Peierls approximation. We emphasize, however,
that BPD finds relatively good approximate solutions
even in the RSB phase as shown in Fig. 3. From these observations, we confirm that statistical–mechanical analyses successfully predict typical performance thresholds
by approximation algorithms.
B.
Greedy algorithm and its variant
Here we examine the greedy algorithm more closely.
To validate the correctness of our analysis, we specifically examine a fraction of selected vertices without the
maximum degree. The value rg is evaluated by
rg = 1 −
ρgx
,
ρx
ρgx
0.05
N = 100
N = 400
0
L = 9, R = 3
0
0.1
0.12
0.14
ρx
0.16
0.18
0.2
FIG. 5. Fraction rg of selected vertices without maximum
degree until the greedy algorithm stops as a function of ρx .
Symbols represent numerical results of the greedy algorithm
with cardinality N = 1000 (squares) and 5000 (circles) and
modified one with N = 1000 (cross marks) and 5000 (stars).
The solid line exhibits the analytical estimation of the fraction
rg .
W1 . The modified greedy algorithm is given as follows:
at each step, (i) choose vertex named i ∈ W1 so that
|∂ 2 i ∩ W1 | is minimized, (ii) delete vertices neighboring
to vertex i from V2 , and (iii) update W1 and V1 and return to (i) if |V1 | > N − K. In Fig. 5, the fraction rg
of the modified algorithm is also shown with that of the
original greedy algorithm. As shown in Fig. 5, the typical performance threshold of the proposed algorithm is
improved, although it is still below the RS-RSB threshold. This fact suggests that the typical performance of
greedy algorithms depends of approximate choices and
BP selects vertices better than those greedy selections.
(42)
where ρgx is given as Eq. (36). The algorithm finds an
optimal solution of the problem if rg = 0. As depicted
in Fig. 5, the analytical estimations of rg averaged over
biregular random graphs agree well with the numerical
results. The fraction arises at ρgx corresponding to the
typical performance threshold.
Additionally, we modify the greedy algorithm to reduce the gap of its typical performance to that of BP.
As suggested above, the reason lies in the point that the
greedy algorithm often chooses wrong vertices leading to
shortage of vertices with maximum degree. To avoid the
situation, one must select a vertex to preserve as many
vertices to a maximum degree as possible, meaning that
the optimal selection requires an exhaustive search. We
therefore propose a modified algorithm in consideration
of the influence of the selection on other vertices in V1 ,
the simplest improved algorithm toward the optimal selection. Let ∂ 2 i = {j ∈ V1 \i | j ∈ ∂a (∀a ∈ ∂i)} be a set
of the second neighbors of vertex i ∈ V1 . We also define
the subset of vertices with the maximum degree in V1 by
C.
Randomized rounding of LP relaxation
In Sec. VII A, we examine the typical performance of
LP relaxation in terms of its approximate value. In this
subsection, randomized rounding, a practical means of
constructing a feasible integer solution from LP relaxation, is applied to LP-relaxed solutions. It is noteworthy that the typical performance of randomized rounding
probably depends on the selection of an LP solver and its
setting. Here, we use IBM ILOG CPLEX with a default
setting. The approximate value of the rounded solution
is compared to that of the RS solution in the RS phase,
which is LK. One considers that the rounding finds an
optimal integer solution of the problem if two estimates
mutually coincide. We define a success ratio pr by the
fraction of random graphs on which the rounding finds an
optimal solution. Fig. 6 presents the average approximation ratio ρry and the success ratio pr as a function of ρx ,
which strongly suggests that the randomized rounding
exhibits a phase transition in terms of its typical performance. The threshold ρrx is less than 0.1, which is much
9
0.6
1
ρry (N = 1000)
ρry (N = 5000)
0.5
ρLP
y
pr (N = 1000)
0.4
pr (N = 5000)
0.6
0.3
pr
ρy
0.8
0.4
L = 9, R = 3
0.2
0.2
0.1
0
0
0.02
0.04
0.06
ρx
0.08
0.1
0
0.12
FIG. 6. Typical performance of randomized rounding of LP
relaxation as a function of ρx (left vertical axis). Symbols
represent approximation values ρry of rounded LP-relaxed solutions with cardinality N = 1000 (squares) and 5000 (cross
marks). The solid line represents the average cover ratio
equivalent to that of LP relaxation in the region shown.
Squares and stars respectively represent the success ratio pr
(right vertical axis) of randomized rounding with N = 1000
and 5000.
lower than ρgx , whereas the LP-relaxed value is regarded
as nearly optimal in the RS region ρx < ρRS
x = 0.2. We
therefore conclude that the typical performance of the
randomized rounding of LP-relaxed solution is inferior
to that of the greedy algorithm.
VIII.
SUMMARY AND DISCUSSION
As described in this paper, we investigate the typical performance of approximation algorithms called belief
propagation, greedy algorithm, and linear-programming
relaxation for maximum coverage problem on sparse
biregular random graphs. The typical performance of
BP is studied by application of the RS cavity method
to a correspondent hard-core lattice–gas model. Results show that, in the large-µ limit, there exist two distinct RS-RSB thresholds regarded as typical performance
thresholds of BP. In addition, the greedy algorithm performance and LP relaxation were studied especially in
the low-density region. Results show that the typical
performance threshold of the greedy algorithm is lower
than that of BP and that LP-relaxed values are always
equivalent to the RS solutions leading to the threshold
equivalent to that of BP. Those analytical results were
validated by executing some numerical simulations. Results of additional numerical studies suggest that BP typ-
[1] V. V. Vazirani, Approximation Algorithms (Springer
Berlin Heidelberg, Berlin, Heidelberg, 2003).
ically works better than the modified greedy algorithm
and that randomized rounding of LP-relaxed solutions
has a lower threshold than the greedy algorithm.
To assess the typical performance of BP as an approximation algorithm, we concentrated on statistical–
mechanical analysis of max-COV up to the RS level.
Further analyses based on the one step RSB will be necessary to reveal statistical–mechanical properties of the
problem and typical performance analysis of another algorithm called survey propagation. Another possible avenue of future work is the extension of our analyses to
other random bipartite graphs. As with min-VCs, it is
an attractive question whether the magnitude relation of
typical performance thresholds changes depending on the
random graph ensembles.
Our results provide not only respective typical performance of approximation algorithms but also their suggestive mutual relations. Specifically examining LP relaxation, it is worth emphasizing that the LP relaxation
finds good approximate values compared to optimal values but the typical performance of its randomized rounding has the smallest threshold among approximation algorithms studied here. To fill the gap of thresholds, it
is important to examine modifications of LP relaxation
such as the cutting-plane approach [35]. As for the greedy
algorithms and their modification, numerical results suggest that evaluation of the influence of their deletion process affects the marked improvement of the typical performance threshold. The fact that BP is better than the
greedy algorithm and its modification indicates that BP
incorporates the influence more efficiently. These suggestions are expected to be of great help to understand
properties and relations of approximation algorithms in
terms of typical performance. In addition, our analyses of the greedy algorithm and randomized rounding of
LP relaxation illustrate that typical case and worst case
evaluations capture different notions of approximate performance in optimization problems. This fact indicates
the importance of the typical-case analysis of approximation algorithms. We hope that the arguments and results presented herein stimulate further studies and that
the typical performance analyses of approximation algorithms will attract the interest of researchers in many
diverse fields.
ACKNOWLEDGMENTS
ST warmly thanks T. Takaguchi for stimulating this
study. The use of IBM ILOG CPLEX has been supported by the IBM Academic Initiative. This research
was supported by JSPS KAKENHI Grant Nos. 25120010
(KH), 16K16011 (TM), and 15J09001 (ST).
[2] R. Monasson, R. Zecchina, S. Kirkpatrick, B. Selman,
10
and L. Troyanskyk, Nature 400, 133 (1999).
[3] M. Mézard, F. Ricci-Tersenghi, and R. Zecchina, J. Stat.
Phys. 111 (2003).
[4] R. M. Karp and M. Sipser, in 22nd Annu. Symp. Found.
Comput. Sci. (sfcs 1981) (IEEE, 1981) pp. 364–375.
[5] M. Weigt and a. K. Hartmann, Phys. Rev. Lett. 84, 6118
(2000).
[6] M. Weigt and H. Zhou, Phys. Rev. E 74, 046110 (2006).
[7] M. Bauer and O. Golinelli, Eur. Phys. J. B - Condens.
Matter 352, 339 (2001).
[8] T. Dewenter and A. K. Hartmann, Phys. Rev. E 86,
041128 (2012).
[9] S. Takabe and K. Hukushima, Journal of the Physical
Society of Japan 83, 043801 (2014).
[10] S. Takabe and K. Hukushima, J. Stat. Mech. Theory Exp.
2016, 113401 (2016).
[11] Y.-Y. Liu, E. Csóka, H. Zhou, and M. Pósfai, Phys. Rev.
Lett. 109, 205703 (2012).
[12] S. Takabe and K. Hukushima, Phys. Rev. E 89, 062139
(2014).
[13] C. Lucibello and F. Ricci-Tersenghi, Int. J. Stat. Mech.
2014, 1 (2014).
[14] Y. Habibulla, J.-H. Zhao, and H.-J. Zhou, in Int. Work.
Front. Algorithmics. (Springer International Publishing.,
2015) pp. 78–88.
[15] J. Takahashi, S. Takabe, and K. Hukushima, Journal of
the Physical Society of Japan 86, 073001 (2017).
[16] H. Schawe and A. K. Hartmann, Eur. Phys. Lett. 113,
30004 (2016).
[17] S. Takabe and K. Hukushima, Phys. Rev. E 93, 053308
(2016).
[18] A. Javanmard, A. Montanari, and F. Ricci-Tersenghi,
Proc. Natl. Acad. Sci. 113, E2218 (2016).
[19] M. P. Johnson and A. Bar-Noy, Proc. - IEEE INFOCOM
, 1071 (2011).
[20] B. Saha and L. Getoor, in Ninth SIAM Int. Conf. Data
Min. (Society for Industrial and Applied Mathematics,
Philadelphia, PA, 2009) pp. 697–708.
[21] H. Takamura and M. Okumura, in Proc. Conf. Eur.
Chapter Assoc. Comput. Linguist. (2009) pp. 781–789.
[22] M. Mézard and A. Montanari, Information, Physics, and
Computation (Oxford University Press, 2009) p. 560,
arXiv:1412.0611.
[23] G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, Math.
Program. 14, 265 (1978).
[24] U. Feige, J. ACM 45, 634 (1998).
[25] P. Raghavan and C. D. Tompson, Combinatorica 7, 365
(1987).
[26] O. Rivoire, G. Biroli, O. C. Martin, and M. Mézard,
Eur. Phys. J. B - Condens. Matter 37, 55 (2003).
[27] M. Mézard and M. Tarzia, Phys. Rev. E 76, 041124
(2007).
[28] L. Zdeborová, Acta Phys. Slovaca. Rev. Tutorials 59, 169
(2009).
[29] M. Weigt, Eur. Phys. J. B - Condens. Matter 28, 369
(2002).
[30] R. T. Rockafellar, Convex Analysis (Princeton University
Press, 1972).
[31] M. E. J. Newman, S. H. Strogatz, and D. J. Watts, Phys.
Rev. E 64, 026118 (2001).
[32] F. Ricci-Tersenghi and G. Semerjian, J. Stat. Mech. Theory Exp. 2009, P09001 (2009).
[33] K. Hukushima and K. Nemoto, J. Phys. Soc. Japan 65,
1604 (1996).
[34] K. Kawasaki, Phys. Rev. 145, 224 (1966).
[35] A. Schrijver, Theory of linear and integer programming
(John Wiley & Sons Ltd,, West Sussex, 1998).
| 8cs.DS
|
F -SINGULARITIES UNDER GENERIC LINKAGE
arXiv:1607.03972v2 [math.AC] 18 Mar 2018
LINQUAN MA, JANET PAGE, REBECCA R.G., WILLIAM TAYLOR, AND WENLIANG ZHANG
Dedicated to Prof. Craig Huneke on the occasion of his 65th birthday
Abstract. Let R = k[x1 , . . . , xn ] be a polynomial ring over a prefect field of positive characteristic. Let
I be an equi-dimensional ideal in R and let J be a generic link of I in S = R[uij ]c×r . We describe the
parameter test submodule of S/J in terms of the test ideal of the pair (R, I) when a reduction of I is a
complete intersection or almost complete intersection. As an application, we deduce a criterion for when
S/J has F -rational singularities in these cases. We also compare the F -pure threshold of (R, I) and (S, J).
1. Introduction
Let R = k[x1 , . . . , xn ] be a polynomial ring over a field of positive characteristic. Let I = (f1 , . . . , fr ) be an
equi-dimensional ideal in R of height c, where equi-dimensional means that all associated primes of I have the
same height [Mat86]. We can define a regular sequence g1 , . . . , gc in S = R[uij ]c×r via gi := ui1 f1 +· · ·+uir fr ,
where the uij are variables over S. Then J = (g1 , . . . , gc ) : I is called a generic link of I in S = R[uij ].
The study of generic linkage has attracted considerable attention and has been developed widely from both
algebraic and geometric points of view [HU87], [HU88], [CU02], [EHU04], [Niu14].
In contrast to the quick and deep development of singularity theories in the past decades, much less
has been known about the behaviors of singularities under generic linkage. A special case is a result of
Chardin and Ulrich [CU02] which says that if R/I is a complete intersection and has rational (resp. F rational) singularities, then a generic link S/J also has rational (resp. F -rational singularities). This result
in characteristic zero has been vastly extended in recent work of Niu [Niu14], which is our main motivation
for this research.
Theorem 1.1 (Theorem 1.1 in [Niu14]). Let J be a generic link of a reduced and equidimensional ideal I
in S = R[uij ] and assume that the characteristic of k is 0. We have
GR
GR ∼
denotes the Grauert-Riemenschneider canonical sheaf of S/J
(1) ωS/J
= J (R, I c ) · (S/J), where ωS/J
c
and J (R, I ) denotes the multiplier ideal of the pair (R, I c ),
(2) lct(S, J) ≥ lct(R, I). In particular, if the pair (R, I c ) is log canonical, then the pair (S, J c ) is also
log canonical.
This result gives a nice criterion for a generic link to have rational singularities in characteristic 0. It also
has applications to bounding the Castelnuovo-Mumford regularity of projective varieties [Niu14, Corollary
1.2]. Since test ideals and F -pure thresholds are characteristic p analogues of multiplier ideals and log
canonical thresholds (c.f. [BST15] and [HY03]), it is natural to ask whether analogues of Theorem 1.1 hold
for test ideals and F -pure thresholds. Our main result is the following, which partially extends Theorem 1.1
to characteristic p and generalizes [CU02, Theorem 3.13] in characteristic p.
Theorem 1.2 (Theorem 3.3, Corollary 4.4). Let J be a generic link of an equi-dimensional ideal I in
S = R[uij ] and assume the characteristic of k is p > 0.
(1) Suppose I is reduced and that a reduction of I is a complete intersection or an almost complete
intersection. Then τ (ωS/J ) ∼
= τ (R, I c ) · (S/J), where τ (ωS/J ) denotes the parameter test submodule
c
and τ (R, I ) denotes the test ideal of the pair (R, I c ).
(2) Suppose that a reduction of I is a complete intersection. Then fptS (J) ≥ fptR (I). In particular, if
the pair (R, I c ) is F -pure, then the pair (S, J c ) is also F -pure.
This paper is organized as follows: in Section 2 we recall and prove some basic result for F -singularities
and test ideals; in Section 3 we give a description of the parameter test submodule of S/J in terms of the test
1
ideal of the pair (R, I), when a reduction of I is a complete intersection or an almost complete intersection.
This generalizes earlier results in [CU02]. In Section 4 we compare the F -pure threshold of the pairs (S, J)
and (R, I) when a reduction of I is a complete intersection.
Acknowledgements. Part of this work was done at Mathematics Research Community (MRC) in Commutative Algebra in June 2015. The authors would like to thank the staff and organizers of the MRC and
the American Mathematical Society for their support. The first author would like to thank Karl Schwede,
Shunsuke Takagi and Bernd Ulrich for fruitful discussions. The first author was partially supported by NSF
Grant DMS #1600198, NSF CAREER Grant DMS #1252860/1501102 and a Simons Travel Grant when
preparing this article. The second author was partially supported by NSF RTG grant DMS #1246844. The
third author was partially supported by the NSF grant DGE #1256260. The fifth author was partially
supported by the NSF grant DMS #1606414. The authors thank the referee for some comments which lead
to improvement of the presentation of the paper.
2. F -singularities and test ideals
In this section we collect some basic definitions of F -singularities and test ideals and prove a characteristic
p > 0 analogue of Ein’s Lemma in [Niu14], which will be used in later sections.
Let R be a Noetherian commutative ring of characteristic p > 0. We will use F∗e R to denote the target of
r7→r p
e
the e-th Frobenius endomorphism F e : R −−−−→ R, i.e. F∗e R is an R-bimodule, which is the same as R as
an abelian group and as a right R-module, that acquires its left R-module structure via the e-th Frobenius
r7→r p
e
e
endomorphism F e : R −−−−→ R. When R is reduced, we will use R1/p to denote the ring whose elements
e
are pe -th roots of elements of R. Note that these notations (when R is reduced) F∗e R and R1/p are used
interchangeably in the literature; we will do so in this paper as well assuming no confusion will arise.
Remark 2.1. If R is a commutative ring essentially of finite type over a perfect field of characteristic p > 0,
then R admits a canonical module denoted by ωR . Applying HomR (−, ωR ) to the e-th Frobenius R → F∗e R
produces an R-linear map
HomR (F∗e R, ωR ) → HomR (R, ωR ) = ωR .
e
∼
Moreover, we have F∗ ωR = HomR (F∗e R, ωR ) (see [BST15, Example 2.4] for more details). Hence we have a
natural R-linear map:
ΦeR : F∗e ωR ∼
= HomR (F∗e R, ωR ) → HomR (R, ωR ) = ωR
called the trace map of the e-th Frobenius.
Example 2.2. When R = k[x1 , . . . , xn ] is a polynomial ring over a perfect field k of characteristic p > 0,
we can identify ωR with R, and ΦeR can be identified with the usual trace TreR , that is:
e −1)
i2 −(pe −1)
in −(pe −1)
e
i1 −(p
−1)
pe
pe
pe
x
∈ Z for each t
x
·
·
·
x
, if it −(p
e
in
e i1 i2
n
1
2
pe
TrR (F∗ (x1 x2 · · · xn )) =
0,
otherwise
In this case HomR (F∗e R, R) is a cyclic F∗e R-module generated by TreR . Furthermore, if f1 , . . . , fc is a regular
sequence in R and T = R/(f1 , . . . , fc ), then we have ([Fed83, Corollary on page 465]1)
e
ΦeT (F∗e (−)) = image of TreR (F∗e (f1p
−1
e
· · · fcp
−1
· −)) in T .
Remark 2.3. Let R = k[x1 , . . . , xn ] be a polynomial ring over a field k of characteristic p > 0 and A =
R[y1 , . . . , ym ] be a polynomial ring over R. For each ideal I in R, it is well known and straightforward to
check that
TreR (F∗e (I))A = TreA (F∗e (IA)).
Lemma 2.4. Let S → R be a surjection of Noetherian commutative rings of characteristic p. Assume that
both S and R admit canonical module ωS and ωR respectively and dim S = dim R. Then
ΦeR = ΦeS |ωR .
1Fedder’s result [Fed83, Corollary on page 465] assumes that the ring R is a Gorenstein local ring only to ensure that
HomR (F∗ R, R) ∼
= F∗ R. In our case, R = k[x1 , . . . , xn ] is a polynomial ring, so HomR (F∗ R, R) is clearly isomorphic to
F∗ R. Hence Fedder’s result applies in our case.
2
Proof. Under our assumptions, we have ωR = HomS (R, ωS ) and the surjection S → R induces an inclusion
ωR = HomS (R, ωS ) ֒→ ωS . Consider the following diagram
HomR (F∗e R, HomS (R, ωS ))
// HomR (R, HomS (R, ωS ))
∼
// HomS (R, ωS )
HomS (F∗e S, ωS )
// HomS (S, ωS )
∼
// ωS
Note that the top row (resp. the bottom row) induces the trace map ΦeR (resp. ΦeS ). To prove our lemma,
it suffices to prove
(a) the vertical map on the left is an inclusion, and
(b) the diagram commutes
To prove (a), note that the vertical map on the left can be refined further as
HomR (F∗e R, HomS (R, ωS )) = HomS (F∗e R, HomS (R, ωS ))
֒→ HomS (F∗e S, HomS (R, ωS )) since F∗e S ։ F∗e R
֒→ HomS (F∗e S, ωS ) since HomS (R, ωS ) ֒→ ωS
To prove (b), note that the commutativity follows directly from the commutativity of
S
// F∗e S
R
// F e R
∗
Definition 2.5 (cf. Definition 3.1 in [Har01] and Definition 2.33 in [BST15]). Let R be an F -finite Noetherian integral domain of characteristic p. The parameter test submodule τ (ωR ) is the unique smallest
nonzero submodule M of ωR such that ΦR (F∗ M ) ⊆ M . R is called F -rational if R is Cohen-Macaulay and
τ (ωR ) = ωR . Note that this is not the original definition of F -rationality, but is known to be equivalent
[Smi97].
Definition 2.6 (cf. Definition 3.16 and Theorem 3.18 in [Sch11]). Let R be an F -finite Noetherian integral
domain of characteristic p. Let I ⊆ R be a nonzero ideal and t ∈ Q≥0 . We define the test ideal τ (R, I t ),
e
abbreviated τ (I t ), to be the unique smallest nonzero ideal J ⊆ R such that φ(F∗e (I ⌈t(p −1)⌉ J)) ⊆ J for all
e > 0 and all φ ∈ HomR (F∗e R, R).
Definition 2.7 (cf. Definitions 1.3 and 2.1 and Proposition 2.2 in [TW04]). Let R be an F -finite, local,
Noetherian, integral domain of characteristic p. Let I ⊂ R be an ideal and t ≥ 0 be a real number.
e
(1) The pair (R, I t ) is F -pure if for all large e ≫ 0, there exists an element d ∈ I ⌊t(p −1)⌋ such that
(F∗e d)R ֒→ F∗e R splits as an R-module homomorphism.
e
(2) The pair (R, I t ) is strongly F -regular if for every c 6= 0 there exists e ≥ 0 and d ∈ I ⌈tp ⌉ such that
F∗e (cd)R ֒→ F∗e R splits as an R-module homomorphism.
(3) The F -pure threshold fptR (I) of (R, I) is sup{s ∈ R≥0 | the pair (R, I s ) is F -pure}, and when R is
strongly F -regular, we also have fptR (I) = sup{s ∈ R≥0 | the pair (R, I s ) is strongly F -regular}.
Remark 2.8. Note that when R is local, (R, I t ) is strongly F -regular if and only if τ (I t ) = R. Indeed,
e
suppose (R, I t ) is strongly F -regular. Pick a nonzero element c ∈ J and take e ≫ 0 and d ∈ I ⌈tp ⌉ satisfying
the conditions of strong F -regularity for c, and let φ : F∗e R → R be a map such that φ(F∗e (cd)) = 1. Then
e
φ(F∗e (I ⌈t(p
−1)⌉
e
J)) ⊇ φ(F∗e (I ⌈tp ⌉ J)) = R,
and so τ (I t ) = R.
On the other hand, if τ (I t ) = R, 0 6= c ∈ R, and a ∈ I ⌈t⌉ , then there exists e ≥ 0 and φ : F∗e R → R such
e
e
that φ(F∗e (I ⌈t(p −1)⌉ acR)) = R. Let b ∈ I ⌈t(p −1)⌉ and f ∈ R such that φ(F∗e (c(abf ))) = 1. Then we are
e
e
done once we note that abf ∈ I ⌈t⌉ I ⌈t(p −1)⌉ ⊆ I ⌈tp ⌉ .
3
We will need the following important description of test ideals:
Theorem 2.9 (cf. Proof of Theorem 3.18 in [Sch11]). With the notations as in Definition 2.6, for any
nonzero a ∈ τ (I t ), we have:
XX
e
τ (I t ) =
φ(F∗e (aI ⌈t(p −1)⌉ ))
e≥0 φ
where the inner sum runs over all φ ∈ HomR (F∗e R, R).
Remark 2.10. With the notations as in Definition 2.6, the following holds ([BSTZ10, 3.3])
X
X
e
φ(F∗e (dI ⌈tp ⌉ ))
(2.10.1)
τ (I t ) =
e≥0 φ∈HomR (F∗e R,R)
where d is a big test element (which is just a nonzero element in τ (R) = τ (R, I 0 )).
If R = k[x1 , . . . , xn ] is a polynomial ring over a perfect field k of characteristic p > 0, then one can set
d = 1 in (2.10.1) and HomR (F∗e R, R) is a cyclic F∗e R-module generated by TreR as discussed earlier. Hence
by (2.10.1),
X
X
X
e
e
φ(F∗e (aI ⌈t(p −1)⌉ )) =
TreR (F∗e (aI ⌈t(p −1)⌉ )), for any a ∈ τ (I t )
τ (I t ) =
e≥0 φ∈HomR (F∗e R,R)
=
X
X
e≥0 φ∈HomR (F∗e R,R)
e≥0
e
φ(F∗e (I ⌈tp ⌉ ))
=
X
e
TreR (F∗e (I ⌈tp ⌉ ))
e≥0
Remark 2.11. With the notations as in Definition
2.5, one can show that if Ra′ is regular, then for every
P
sufficiently large power a of a′ , τ (ωR ) = e ΦeR (F∗e (a · ωR )). This can be proved by a similar argument as
[ST12, Lemma 3.6, Lemma 3.8] so we omit the details.
The following result from [ST12] will also be used. These results were originally proved in [HY03] and
[HT04], and they hold as long as R is F -finite. We will only state the version of these results that we need.
Lemma 2.12 (cf. Theorem 6.9 in [ST12]). Let R be an integral domain essentially of finite type over a
perfect field of characteristic p > 0 and let I, J ⊆ R be nonzero ideals and t ∈ R≥0 .
(1) If J is a reduction of I, then τ (I t ) = τ (J t ).
(2) If J is generated by r elements, then τ (J r ) = Jτ (J r−1 ).
We are ready to prove the characteristic p > 0 analogue of Ein’s Lemma in [Niu14]:
Lemma 2.13 (Ein’s Lemma in characteristic p > 0). Let R be an integral domain essentially of finite
type over a perfect field of characteristic p > 0 and let I ⊆ R be an equi-dimensional and unmixed ideal
of codimension c. If τ (I c−1 ) = R, then τ (I c ) = I. In particular, if R is strongly F -regular and (R, I c ) is
F -pure, then τ (I c ) = I.
Proof. The lemma will follow from the following two inclusions:
(2.13.1)
τ (I c ) ⊆ I.
(2.13.2)
Iτ (I t−1 ) ⊆ τ (I t ) for all t ≥ 1.
Indeed, if τ (I c−1 ) = R, then I = Iτ (I c−1 ) ⊆ τ (I c ) ⊆ I, and so we have equality throughout.
Proof of (2.13.1). Since inclusion is a local condition, we may assume that R is local with maximal ideal m.
By replacing R by R[x]mR[x] , we may assume that R has infinite residue field: it is straightforward to check
that τ (I c )R[x]mR[x] = τ ((IR[x]mR[x] )c ). Now let p be a minimal prime of I. Since I is equi-dimensional,
dim Rp = c. Hence IRp has a reduction J ⊆ IRp generated by c elements. Therefore, since test ideals
localize,
τ (I c )Rp = τ ((IRp )c ) = τ (J c ) = Jτ (J c−1 ) ⊆ J ⊆ IRp .
Since every associated prime of I is minimal, this inclusion holds for all associated primes of I, hence it holds
globally, i.e. τ (I c ) ⊆ I.
4
Proof of (2.13.2). This should be well known to experts in the field; we opt to provide a proof here since we
could not locate a proper reference. Let t ∈ R≥1 , and pick 0 6= a ∈ τ (I t ). Then
XX
e
Iτ (I t−1 ) = I
φ F∗e aI ⌈(t−1)(p −1)⌉
e≥0 φ
XX
e
e
=
φ F∗e aI [p ] I ⌈(t−1)(p −1)⌉
e≥0 φ
XX e
e
⊆
φ F∗e aI p I ⌈(t−1)(p −1)⌉
e≥0 φ
XX e
e
=
φ F∗e aI p +⌈(t−1)(p −1)⌉
e≥0 φ
XX
e
⊆
φ F∗e aI ⌈t(p −1)⌉
e≥0 φ
= τ (I t ),
where the inner sum runs over all φ ∈ HomR (F∗e R, R) and the last inclusion following from the fact that
pe + ⌈(t − 1)(pe − 1)⌉ = ⌈pe + (t − 1)(pe − 1)⌉ = ⌈t(pe − 1) + 1⌉ > ⌈t(pe − 1)⌉.
For the last statement, if (R, I c ) is F -pure, then the F -pure threshold of I is at least c. Since the F -pure
threshold is the supremum of those values t for which (R, I t ) is strongly F -regular when R is strongly F regular [TW04, Proposition 2.2], we have that (R, I c−1 ) is strongly F -regular. This means that τ (I c−1 ) = R
by Remark 2.8, and hence the first statement of the lemma tells us τ (I c ) = I.
3. F -rationality under generic linkage
In this section, we investigate how F -singularities (e.g. F -purity, F -rationality, etc) behave under a
generic linkage. To this end, we will also consider the behaviors of test ideals under a generic linkage. We
begin with recalling the definition of a generic link.
Definition 3.1. Let R = k[x1 , . . . , xn ] be a polynomial ring over a perfect field of positive characteristic.
Let I be an equi-dimensional and unmixed ideal of R of height c. Fix a generating set {f1 , . . . , fr } of I. Let
(uij ), 1 ≤ i ≤ c, 1 ≤ j ≤ r, be a c × r matrix of variables. Consider c elements g1 , . . . , gc in S = R[uij ]
defined by
gi := ui1 f1 + ui2 f2 + · · · + uir fr
for 1 ≤ i ≤ c. Then J = (g1 , . . . , gc ) : (IS) is called the first generic link of I with respect to {f1 , . . . , fr }
(we also call S/J the generic link of R/I with respect to {f1 , . . . , fr }).
Remark 3.2. It is well known that under the above assumptions, if I is reduced, then IS and J are geometrically linked, i.e., IS = (g1 , . . . , gc ) : J and IS ∩ J = (g1 , . . . , gc ). Moreover, J is actually a prime ideal of
height c [HU85, 2.6].
The following theorem is our main technical result in this section.
Theorem 3.3. With the notation as in Definition 3.1, assuming I is reduced, we have
(1) τ (ωS/J ) ⊆ τ (I c ) · (S/J);
(2) If I has a minimal reduction generated by at most c + 1 elements, then τ (ωS/J ) ⊇ τ (I c ) · (S/J);
hence τ (ωS/J ) = τ (I c ) · (S/J) in this case.
Our proof of Theorem 3.3(2) requires considering different sets of generators of I. A priori, a generic link
(S, J) depends on the choice of generators. The following lemma guarantees that the statement in Theorem
3.3(2) is independent of the choice of generators of I. Its proof follows the same line as the proof of [HU87,
Proposition 2.11].
Lemma 3.4. Let Λ1 and Λ2 be two sets of generators of I and let (S1 , J1 ) and (S2 , J2 ) be generic links of
I with respect to Λ1 and Λ2 respectively. Then τ (ωS1 /J1 ) ⊇ τ (I c ) · (S1 /J1 ) iff τ (ωS2 /J2 ) ⊇ τ (I c ) · (S2 /J2 ).
5
Proof. By considering Λ1 ∪ Λ2 , we can assume that Λ1 ⊆ Λ2 . By induction on the difference between the
cardinality of Λ1 and Λ2 , we may assume that Λ2 has one more element than Λ1 , i.e. we may assume that
Λ1 = {f1 , . . . , fr } and Λ2 = Λ1 ∪ {fr+1 }.
Denote the height of I by c. Let {uij | 1 ≤ i ≤ c, 1 ≤ j ≤ r + 1} be indeterminates over R. Set
S1 = R[uij ]1≤i≤c,1≤j≤r and S2 = R[uij ]1≤i≤c,1≤j≤r+1 . For i = 1, . . . , c, set
gi := ui1 f1 + · · · + uir fr
and
hi := ui1 f1 + · · · ui,r+1 fr+1 .
Then J1 = ((g1 , . . . , gc ) :S IS) is the first generic link of I with respect to Λ1 and J2 = ((h1 , . . . , hc ) :S2 IS2 )
is the first generic link of I with respect to Λ2 .
Pr
It is clear that S2 = S1 [u1,r+1 , . . . , uc,r+1]. Since fr+1 ∈ I, we must have that fr+1 = j=1 aj fj for some
aj ∈ R. Let ϕ : S2 → S2 be the automorphism given by the linear change of variables
uij 7→ uij + ui,r+1 aj
for 1 ≤ i ≤ c and 1 ≤ j ≤ r and
ui,r+1 7→ ui,r+1
for 1 ≤ i ≤ c.
We claim that ϕ(J1 S2 ) = J2 and we reason as follows. For i = 1, . . . , c, we have that
ϕ(gi ) =
r
X
j=1
(uij + ui,r+1 aj )fj =
r
X
uij fj + ui,r+1
r
X
aj f j =
j=1
j=1
r
X
uij fj + ui,r+1 fr+1 = hi .
j=1
Now since S1 ֒→ S2 is a faithfully flat extension, we have that
J1 S2 = ((g1 , . . . , gc ) :S1 IS1 )S2 = ((g1 , . . . , gc )S2 :S2 IS2 ),
and hence
ϕ(J1 S2 ) = ϕ((g1 , . . . , gc )S2 :S2 IS2 ) = (ϕ(g1 , . . . , gc )S2 :S2 ϕ(IS2 )) = ((h1 , . . . , hc ) :S2 IS2 ) = J2 .
S2ϕ
denote the S1 -algebra that is the same as S2 as a ring and whose S1 -module structure is induced by
Let
ϕ
→ S2 . Then we have shown that J1 ⊗S1 S2ϕ = J2 and hence S1 /J1 ⊗S1 S2ϕ = S2 /J2 . Combining
S1 ֒→ S2 −
Remarks 2.3 and 2.11, one can check that
τ (ωS1 /J1 ) ⊗S1 S2ϕ = τ (ωS1 /J1 ⊗S1 S2ϕ )
where the right hand side is precisely τ (ωS2 /J2 ). Our lemma follows immediately since S2ϕ is faithfully flat
over S1 .
The following lemma is also needed in the proof of Theorem 3.3.
Lemma 3.5. Let c, r be positive integers such that c = r orP
c = r − 1. Let β = (β1 , . . . , βr ) be an element
of Nr , where N is the set of non-negative integers. Assume i βi = c(pe − 1). Then there exist c elements
α1 , ..., αc in Nr such that:
(1) each αi has at most two nonzero entries;
e
(2) the sum
P of the entries of each αi is p − 1;
(3) βj = i αij , where αi = (αi1 , . . . , αir ).
Proof. We will induce on r. If c = r = 1, then β = (pe − 1) and we let α1 = β. If c = 1, r = 2, we have
β = (β1 , β2 ) where β1 + β2 = pe − 1 and we can let α1 = (β1 , β2 ) and again (1)-(3) hold.
If c = r and β1 = · · · = βc = pe − 1, then we can set αi to be the vector with pe − 1 at i-th spot and 0
elsewhere. Otherwise, there must be a βi < pe − 1. Without loss of generality, we assume that βr < pe − 1.
We claim that βj ≥ pe − 1 − βr for some j between 1 and r − 1, and we reason as follows. If c = r, then
there must be a j such that βj > pe − 1, and hence βj ≥ pe − 1 − βr . Now assume that c = r − 1. Suppose
βi < pe − 1 − βr for all i ≤ r − 1, as then we would have:
r
X
βi < (r − 1)(pe − 1 − βr ) + βr ≤ (r − 2)(pe − 1 − βr ) + (pe − 1) ≤ (r − 1)(pe − 1) = c(pe − 1)
i=1
6
Pr
which contradicts the assumption that i=1 βi = c(pe − 1). So, there is a j between 1 and r − 1 such that
βj ≥ p e − 1 − βr .
Set αc := (0, . . . , 0, pe − 1 − βr , 0, . . . , βr ) where pe − 1 − βr appears in the j-th spot. Consider
(β1 , . . . , βj−1 , βj − (pe − 1 − βr ), βj+1 , . . . , βr−1 ).
This is an element of Nr−1 such that the sum of its entries is (c − 1)(pe − 1). By our induction hypotheses,
there are γ1 , . . . , γc−1 ∈ Nr−1 that satisfy (1), (2), and (3). For 1 ≤ i ≤ c − 1, setting αi be γi with a 0 added
to the end completes the proof of our lemma.
Proof of Theorem 3.3. By Remark 3.2, J is a minimal prime of (g1 , . . . , gc ). Hence once we identify
ωS/J = HomS/(g1 ,...,gc ) (S/J, S/(g1 , . . . , gc )) = ((g1 , . . . , gc ) : J) · (S/J) = I · (S/J),
we know from Lemma 2.4 that
e
ΦeS/J (F∗e (−)) = ΦeS/(g1 ,...,gc ) (F∗e (−))|ωS/J = TreS (F∗e (g1p
−1
e
· · · gcp
−1
· −))|I·(S/J) .
Next we notice that for every 1 ≤ k ≤ r, (S/J)fk ∼
= Rfk [uij |j 6= k] is regular. Hence for N ≫ 0, fkN is a test
element for S/J. Thus by Remark 2.11, we have:
X
X
e
e
(3.5.1)
τ (ωS/J ) =
ΦeS/J (F∗e (fkN · ωS/J )) =
TreS (F∗e (g1p −1 · · · gcp −1 · fkN · IS)) · (S/J)
e≥0
e≥0
Since fk ∈ I and R is regular, by Remark 2.10, for N ≫ 0 we also have:
X
e
(3.5.2)
τ (I c ) · (S/J) =
TreR (F∗e ((f1 , . . . , fr )c(p −1) · fkN · R)) · (S/J)
e≥0
e
e
When we expand g1p −1 · · · gcp −1 , it is easy to see from (3.5.1) that τ (ωS/J ) can be generated by elements of
the form
e
e
Y
Y
p −1
p −1
γ
α
uijij
uijij · fkN · s ·
···
f β1 f β2 · · · frβr
(3.5.3)
TreS F∗e
α11 , . . . , α1r
αc1 , . . . , αcr 1 2
1≤i≤c
1≤i≤c
1≤j≤r
1≤j≤r
where 0 ≤ αij ≤ pe − 1, βj =
is equal to
Y
αij +γij −(pe −1)
pe
uij
·
Pc
i=1
TreR
1≤i≤c
αij ,
Pr
j=1
βj = c(pe − 1) and s ∈ I. By definition of the trace map, this
pe − 1
pe − 1
β1 β2
e
N
βr
···
F∗
f f · · · fr · fk · s
α11 , . . . , α1r
αc1 , . . . , αcr 1 2
1≤j≤r
α +γ −(pe −1)
denotes 0 if αij + γij 6≡ −1 mod pe . But it is clear that this element is in τ (I c ) · S by
where ij ijpe
expression (3.5.2). This proves (1).
Next we prove (2). By Lemma 3.4 we can assume that I˜ = (f1 , . . . , fc+1 ) is a reduction of I (the case
that I has a reduction generated by c elements is similar). Hence by the arguments above, we have that, for
1 ≤ k ≤ c and N ≫ 0,
X
e
τ (I c ) · (S/J) = τ (I˜c ) · (S/J) =
TreR (F∗e ((f1 , . . . , fc+1 )c(p −1) · fkN +1 · R)) · (S/J)
e≥0
f1β1
βc+1
· · · fc+1
Given a generator
sion of Lemma 3.5. Then
e
of (f1 , . . . , fc+1 )c(p
Y
, we can find α1 , . . . , αc ∈ Nc+1 satisfying the concluY
(uij fj )αij =
1≤i≤c
1≤i≤c
1≤j≤c+1
1≤j≤c+1
pe −1
pe −1
α11 ,...,α1,c+1 · · · αc1 ,...,αc,c+1
pe −1
pe −1
for some ji by
αi1 ,...,αi,c+1 = αiji
appears with coefficient
mial coefficient
−1)
α
uijij
β
Y
fj j
1≤j≤c+1
e
in the product g1p
−1
e
· · · gcp
−1
. Because each multino-
Lemma 3.5 (1)-(2), they are nonzero in k.
7
Each αij is less than pe , so let
Y
s′ =
pe −1−αij
uij
1≤i≤c
1≤j≤c+1
Then TreS (F∗e (− · s′ )) sends
e
Q
e
1≤l≤n
xlp
e
−1
Q
Y
1≤i≤c
c+2≤j≤r
pe −1
uij
.
α
1≤i≤c
1≤j≤c+1
uijij to 1 and all other basis elements to 0. Hence,
TreS (F∗e (g1p −1 · · · gcp −1 · fkN +1 · s′ · R))
Y
c+1
e
e
Y
e
p −1
p −1
β
p −1
fj j · fkN +1 R
uij
···
·
= TreS F∗e
α11 , . . . , α1,c+1
αc1 , . . . , αc,c+1
j=1
1≤i≤c
1≤j≤r
= TreR F∗e
c+1
Y
j=1
In particular,
TreR F∗e
c+1
Y
j=1
β
fj j
β
fj j · fkN +1 · R .
e
· fkN +1 · R · (S/J) = TreS (F∗e (g1p
−1
e
· · · gcp
−1
· fkN · fk s′ R)) · (S/J) ⊆ τ (ωS/J )
Q
e
βj
of (f1 , . . . , fc+1 )c(p −1) , where the second inclusion follows from expression
for every generator c+1
j=1 fj
(3.5.1). Therefore we have
τ (I c ) · (S/J) = τ (I˜c ) · (S/J)
X
e
=
TreR (F∗e ((f1 , . . . , fc+1 )c(p −1) · fkN +1 · R)) · (S/J)
e≥0
⊆ τ (ωS/J ).
Remark 3.6. The proof of Theorem 3.3 (2) requires the minimal reduction be generated by at most c + 1
elements. If not, then we are not in the case of Lemma 3.5 and it may be the case that there are always at least
three nonzero entries in some αi . Consequently, multinomial coefficients must be taken into consideration.
Corollary 3.7. With the notation as in Definition 3.1 and the assumptions as in Theorem 3.3 (2), τ (ωS/J ) =
ωS/J if and only if τ (I c ) = I. In particular, S/J has F -rational singularities if and only if S/J is CohenMacaulay and τ (I c ) = I.
Proof. If τ (I c ) = I, then Theorem 3.3 immediately implies τ (ωS/J ) = ωS/J .
Conversely, assume that τ (I c ) 6= I and τ (ωS/J ) = ωS/J . Since τ (I c ) is always contained in I by (2.13.1), at
least one of the generators of I is not in τ (I c ), say f1 . From Theorem 3.3, we can see that τ (I c )S+J = IS+J;
hence f1 ∈ τ (I c )S + J. Thus, there are elements a ∈ τ (I c )S and b ∈ J such that f1 = a + b. (Note that
b 6= 0.) Then we have f1 − a ∈ J which implies that (f1 − a)f1 ∈ (g1 , . . . , gc ). However, this is impossible
because of the degrees in the uij . This is a contradiction.
The last assertion is clear because S/J is F -rational if and only if S/J is Cohen-Macaulay and τ (ωS/J ) =
ωS/J .
Corollary 3.8. With the notation as in Definition 3.1 and the assumptions as in Theorem 3.3 (2), if the
pair (R, I c ) is F -pure and R/I is Cohen-Macaulay, then S/J is F -rational. In particular, if R/I is an
F -pure complete intersection, then S/J is F -rational.
Proof. By Lemma 2.13, (R, I c ) is F -pure implies τ (I c ) = I. The first statement thus follows from Corollary
3.7. Finally, it is well known that when R/I is an F -pure complete intersection, the pair (R, I c ) is F -pure.
This follows from a Fedder type criterion ([Tak04, Lemma 3.9] and others).
We can recover [Niu14, Corollary 3.4] in the complete intersection and almost complete intersection cases.
8
Corollary 3.9. Let I = (f1 , . . . , fr ) be an ideal of C[x1 , . . . , xn ] and let c be the codimension of I. Let S
and J be in Definition 3.1. Assume that r ≤ c + 1. Then S/J has rational singularities if and only if S/J
is Cohen-Macaulay and I (I c ) = I, where I (I c ) is the multiplier ideal of I c .
Proof. By [Smi97] and [Har98], S/J has rational singularities if and only if its reduction (S/J)p is F -rational
for all p ≫ 0. It is easy to see that, for p ≫ 0, the reduction Jp of J is a generic link of the reduction Ip of
I. Hence, S/J has rational singularities if and only if (S/J)p is Cohen-Macaulay and τ (Ipc ) = Ip for p ≫ 0
by Corollary 3.7. On the other hand, it was proved in [HY03] that I (I c )p = τ (Ipc ) for all p ≫ 0. Therefore,
we have S/J has rational singularities if and only if (S/J)p is Cohen-Macaulay and I (I c )p = Ip for p ≫ 0.
This completes the proof.
4. Behavior of F -pure threshold under generic linkage
In this section we investigate behaviors of F -pure thresholds under generic linkages. We begin with an
easy lemma.
Lemma 4.1. Let R = k[x1 , . . . , xn ] be a polynomial ring over a perfect field of characteristic p and I be an
equi-dimensional and unmixed ideal of R. Let Λ1 and Λ2 be 2 sets of generators of I and let (Si , Ji ) be the
generic link with respect to Λi (i=1,2). Then
fptS1 (J1 ) = fptS2 (J2 ).
Proof. As in the proof of Lemma 3.4, we can assume that Λ1 = {f1 , . . . , fr } and Λ2 = {f1 , . . . , fr , fr+1 }.
Let ϕ and S2ϕ be the same as in the proof of Lemma 3.4. It is straightforward to check that
τ (J1t ) ⊗S1 S2ϕ = τ (J2t )
for each nonnegative real number t. Our lemma follows immediately.
Remark 4.2. Let k ⊆ K be an extension of perfect fields and let R = k[x1 , . . . , xn ] and T = K[x1 , . . . , xn ].
e
e
Since HomR (R1/p , R) and HomT (T 1/p , T ) are generated by the same projection, we have τR (I t ) = τT ((IT )t )
(c.f. [BMS08, Remark 2.18]).
Theorem 4.3. Let R = k[x1 , . . . , xn ] be a polynomial ring over a perfect field of characteristic p and I be
an equi-dimensional and unmixed ideal of height c in R. Assume that I = (f1 , . . . , fs ) and that I has a
reduction I˜ generated by r elements. Let S = R[uij ]1≤i≤c,1≤j≤s be a polynomial ring over R. For 1 ≤ i ≤ c,
let
gi = ui1 f1 + ui2 f2 + . . . + uis fs .
c
Then fptS (g1 , . . . , gc ) ≥ r fptR (I).
Proof. By Lemma 4.1, we can add the generators of I˜ to those of I and then assume that I˜ = (f1 , . . . , fr ).
˜ Hence it
Since I˜ is a reduction of I, it follows from [TW04, Proposition 2.2(6)] that fptR (I) = fptR (I).
ct
t
˜
suffices to show that τR (I ) = R implies τS ((g1 , . . . , gc ) r ) = S for any positive real number t < c. To this
end, assume that τR (I˜t ) = R. By Remark 4.2, we may assume that k is algebraically closed.
ct
We wish to show that τS ((g1 , . . . , gc ) r ) = S. Suppose otherwise and we seek a contradiction. There
ct
is a maximal ideal m of S such that τS ((g1 , . . . , gc ) r ) ⊆ m. Since k is algebraically closed, we can write
m = (x1 − a1 , . . . , xn − an , u11 − b11 , . . . , ucr − bcr ) for some ai , bij ∈ k. Set n = (x1 − a1 , . . . , xn − an ).
e
Since τR (I˜t ) = R, there exist an integer e, an R-linear map φ ∈ HomR (R1/p , R), and nonnegative integers
e
e
P
α /p
α /p
/ n.
α1 , . . . , αr with i αi = ⌈tpe ⌉ such that φ(f1 1 · · · fr r ) ∈
At this point we show that each fj ∈ n, and therefore αj ≤ pe − 1 for all j. Indeed, let e ≥ 1 such
e
(pe −1)/pe
(pe −1)/pe (pe −1)/pe
to 1
· · · ucj
u2j
that pe ≥ c/(c − t) and let ψ : S 1/p → S send the basis element u1j
a /pe b /pe
and all other basis elements xℓ ℓ uijij
(ct/r)pe ≤ tpe ≤ c(pe − 1). Therefore
fjc
−1 pe −1
g2
· · · ucj
c/pe (pe −1)/pe (pe −1)/pe
g1
g2
· · · gc(p
ψ(fj
=
ψ(fj
⌈(ct/r)pe ⌉ 1/pe
ψ(((g1 , · · · , gc )
)
9
e
· · · gcp
−1
e
)⊆m
−1)/pe
)
e
∈ (g1 , · · · , gc )⌈(ct/r)p ⌉ , because
(pe −1)/pe c(pe −1)/pe
)
fj
c/pe (pe −1)/pe (pe −1)/pe
u2j
u1j
=
⊆
e
to 0. Now fjc g1p
ct
by our choice of ψ and the assumption that τS ((g1 , . . . , gc ) r ) ⊆ m. It follows that fj ∈ m ∩ R = n.
Without loss of generality, we may assume that pe − 1 ≥ α1 ≥ α2 ≥ · · · ≥ αr . Consequently,
lc
m lc
m lc m
α1 + · · · + αc ≥
(α1 + · · · + αr ) =
⌈tpe ⌉ ≥
tpe
r
r
r
α
/pe
α /pe
α
/pe
α /pe
c+1
c+1
· · · fr r
followed by the application
· · · fr r ·−), i.e. pre-multiplication by fc+1
Let φα = φ(fc+1
e
α1 /pe
αc /pe
1/p
→ R is an R-linear map and that φα (f1
· · · fc
)∈
/ n. We can extend
of φ. It is clear that φα : R
e
e
φα to an S-linear map ψα : R1/p [uij ] → S = R[uij ] that sends each uij to itself and restricts to φα on R1/p .
e
e
Q
e
e
e
b /p
1/p
It is clear that S 1/p = R1/p [uij ] is a free R1/p [uij ]-module with a basis { 0≤bij ≤pe −1 uijij }. Let
e
1/pe
α /pe
α /pe
e
to 1 and all other basis elements
πα : R1/p [uij ] → R1/p [uij ] be the projection that sends u111 · · · uccc
to 0.
ψz a
πα
e
e
Let θα be the composition of S 1/p −−→ R1/p [uij ] −−→ S. It is clear that θα is S-linear. By the
construction of πα , it is straightforward to check that
α /pe
e
α /pe
Since φ(f1 1
α /pe
· · · fr r
e
α /p
φ(f1 1
e
e
α /pe
· · · gcαc /p ) = θα ((u11 f1 )α1 /p · · · (ucc fc )αc /p ) = φ(f1 1
θa (g1 1
e
· · · frαr /p ).
) in R but not in n = (x1 − a1 , . . . , xn − an ), we must have
e
· · · frαr /p ) ∈
/ m = (x1 − a1 , . . . , xn − an , u11 − b11 , . . . , ucr − bcr ),
ct
ct
e
a contradiction to the assumption that τS ((g1 , . . . , gc ) r ) ⊆ m (note that g1α1 · · · gcαc ∈ (g1 , . . . , gc )⌈ r p ⌉ ).
We have some immediate corollaries.
Corollary 4.4. Let R = k[x1 , . . . , xn ] be a polynomial ring over a perfect field of characteristic p and I be
an equi-dimensional and unmixed ideal of height c in R. Let J be a generic link of I in S = R[uij ]. The
following hold:
(1) If I has a reduction generated by r elements, then fptS (J) ≥ rc fptR (I).
(2) If I has a reduction generated by c elements, in particular if I is a complete intersection, then
fptS (J) ≥ fptR (I).
(3) fptS (J) ≥ nc fptR (I) (note n = dim(R)).
Proof. To prove (1), note that since (g1 , . . . , gc ) ⊆ J, we have fptS (J) ≥ fptS (g1 , . . . , gc ). Theorem 4.3 then
completes the proof.
(2) is an immediate consequence of (1).
(3) By Remark 4.2, passing to the algebraic closure of k doesn’t affect fptR (I) and fptS (J). Hence we
can assume that k is algebraically closed and hence is infinite. [Lyu86, Theorem] asserts that each ideal I
admits a reduction generated by n elements. We are done by (1).
References
[BMS08] Manuel Blickle, Mircea Mustaţǎ, and Karen E. Smith. Discreteness and rationality of F -thresholds. Michigan Math.
J., 57:43–61, 2008. Special volume in honor of Melvin Hochster.
[BST15] Manuel Blickle, Karl Schwede, and Kevin Tucker. F -singularities via alterations. Amer. J. Math., 137(1):61–109,
2015.
[BSTZ10] Manuel Blickle, Karl Schwede, Shunsuke Takagi, and Wenliang Zhang. Discreteness and rationality of F -jumping
numbers on singular varieties. Math. Ann., 347(4):917–949, 2010.
[CU02]
Marc Chardin and Bernd Ulrich. Liaison and Castelnuovo-Mumford regularity. Amer. J. Math., 124(6):1103–1124,
2002.
[EHU04] David Eisenbud, Craig Huneke, and Bernd Ulrich. Heights of ideals of minors. Amer. J. Math., 126(2):417–438, 2004.
[Fed83]
Richard Fedder. F -purity and rational singularity. Trans. Amer. Math. Soc., 278(2):461–480, 1983.
[Har98]
Nobuo Hara. A characterization of rational singularities in terms of injectivity of Frobenius maps. Amer. J. Math.,
120(5):981–996, 1998.
[Har01]
Nobuo Hara. Geometric interpretation of tight closure and test ideals. Trans. Amer. Math. Soc., 353(5):1885–1906,
2001.
[HT04]
Nobuo Hara and Shunsuke Takagi. On a generalization of test ideals. Nagoya Math. J., 175:59–74, 2004.
[HU85]
Craig Huneke and Bernd Ulrich. Divisor class groups and deformations. Amer. J. Math., 107(6):1265–1303 (1986),
1985.
[HU87]
Craig Huneke and Bernd Ulrich. The structure of linkage. Ann. of Math. (2), 126(2):277–334, 1987.
10
[HU88]
[HY03]
[Lyu86]
[Mat86]
[Niu14]
[Sch11]
[Smi97]
[ST12]
[Tak04]
[TW04]
Craig Huneke and Bernd Ulrich. Algebraic linkage. Duke Math. J., 56(3):415–429, 1988.
Nobuo Hara and Ken-Ichi Yoshida. A generalization of tight closure and multiplier ideals. Trans. Amer. Math. Soc.,
355(8):3143–3174 (electronic), 2003.
Gennady Lyubeznik. A property of ideals in polynomial rings. Proc. Amer. Math. Soc., 98(3):399–400, 1986.
Hideyuki Matsumura. Commutative ring theory, volume 8 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1986. Translated from the Japanese by M. Reid.
Wenbo Niu. Singularities of generic linkage of algebraic varieties. Amer. J. Math., 136(6):1665–1691, 2014.
Karl Schwede. Test ideals in non-Q-Gorenstein rings. Trans. Amer. Math. Soc., 363(11):5925–5941, 2011.
Karen E. Smith. F -rational rings have rational singularities. Amer. J. Math., 119(1):159–180, 1997.
Karl Schwede and Kevin Tucker. A survey of test ideals. In Progress in commutative algebra 2, pages 39–99. Walter
de Gruyter, Berlin, 2012.
Shunsuke Takagi. F-singularities of pairs and inversion of adjunction of arbitrary codimension. Invent. Math.,
157(1):123–146, 2004.
Shunsuke Takagi and Kei-ichi Watanabe. On F-pure thresholds. J. Algebra, 282(1):278–297, 2004.
Department of Mathematics, University of Utah, Salt Lake City, UT 84112
E-mail address: [email protected]
Department of Mathematics, University of Illinois at Chicago, Chicago, Illinois 60607
E-mail address: [email protected]
Department of Mathematics, Syracuse University, Syracuse, NY 13244
E-mail address: [email protected]
Department of Mathematical Sciences, University of Arkansas, Fayetteville, AR 72701
E-mail address: [email protected]
Department of Mathematics, University of Illinois at Chicago, Chicago, Illinois 60607
E-mail address: [email protected]
11
| 0math.AC
|
1
A Stochastic Geometry-based Demand Response
Management Framework for Cellular Networks
Powered by Smart Grid
arXiv:1612.01636v1 [cs.SY] 6 Dec 2016
Muhammad Junaid Farooq, Hakim Ghazzai, and Abdullah Kadri
Qatar Mobility Innovations Center (QMIC), Qatar Science & Technology Park, Doha, Qatar,
Email: {junaidf, hakimg, abdullahk}@qmic.com.
Abstract—In this paper, the production decisions across multiple energy suppliers in smart grid, powering cellular networks
are investigated. The suppliers are characterized by different
offered prices and pollutant emissions levels. The challenge is
to decide the amount of energy provided by each supplier to
each of the operators such that their profitability is maximized
while respecting the maximum tolerated level of CO2 emissions.
The cellular operators are characterized by their offered quality
of service (QoS) to the subscribers and the number of users
that determines their energy requirements. Stochastic geometry
is used to determine the average power needed to achieve the
target probability of coverage for each operator. The total average
power requirements of all networks are fed to an optimization
framework to find the optimal amount of energy to be provided
from each supplier to the operators. The generalized alphafair utility function is used to avoid production bias among the
suppliers based on profitability of generation. Results illustrate
the production behavior of the energy suppliers versus QoS level,
cost of energy, capacity of generation, and level of fairness.
Index Terms—Cellular networks, demand response, smart
grid, stochastic geometry.
I. I NTRODUCTION
Cellular networks have seen an enormous growth in the
number of connected subscribers, particularly after the advent
of smart phones and 3G/4G mobile broadband technologies.
To cater for the existing, as well as future communication
requirements of emerging wireless technologies such as the 5G
and internet of things (IoT), there is a need for massive network deployment, i.e., base stations (BSs) to increase network
capacity of cellular operators [1]. This will lead to a surge
in the energy consumption of cellular networks and hence,
negative environmental impact in the form of pollutant gases
emissions mainly carbon dioxide (CO2 ). It is estimated that
the carbon footprint of wireless communications could almost
triple in 2020 if no actions are taken [2]. In order to counter the
increasing carbon footprint of cellular networks, steps must be
taken to limit the pollutant emissions by incorporating green
renewable sources of energy alongside conventional fossil fuel
energy generation. Moreover, effective energy management
techniques are required in order to achieve the goal of reducing
CO2 emissions.
The development of the smart grid, a modern electricity
grid that allows integration of energy generated from different
sources, has enabled intelligent energy management both at the
demand and supply side. The smart grid comprises of multiple
energy suppliers, some of which provide renewable energy.
This work was made possible by NPRP grant # 6-001-2-001 from the Qatar
National Research Fund (A member of The Qatar Foundation). The statements
made herein are solely the responsibility of the authors.
The demand side management (DSM) in smart grid powered
cellular networks allows operators to make profitable and
environment-friendly energy procurement decisions according
to the changes in price or carbon footprint of the energy [3].
On the other hand, an increase in energy demand from the
side of cellular operators requires a decision at the smart grid
level about the quantity of energy provided by the suppliers
to the cellular operators and the price of energy. This is also
known as demand response management (DRM) [4].
The objective of DRM is to determine the average amount of
energy to be produced by each supplier in smart grid such that
the profit of the suppliers is maximized while respecting the
CO2 emissions threshold set by the regulatory authority. For
this, an accurate estimate of the total energy requirements of
the cellular operators is required, that depends on the quality
of their transmissions and the number of subscribers using
their services. The average radiated power of BSs of each
operator is the most crucial component in determining the
total power consumption. Stochastic Geometry (SG) [5] is
employed to estimate the transmit power required by operators
to provide their subscribers with a certain quality of service
(QoS) level. SG allows the computation of important network
statistics such as the probability of coverage (or outage) as
spatial averages [6]. Since cellular networks involve BSs and
users which can be modeled as points from a particular
point process, we can use SG for performance analysis of
such networks. The Poisson point process (PPP), which is
well known for its analytical tractability, is used to model
the location of operator BSs and subscribers. The resulting
analysis enables us to ensure that, on average, any user in the
network is served with a predefined QoS and to determine the
corresponding power requirements.
Existing works on DRM in smart grid deal with capacity
planning by a single energy supplier in response to the
changes in the energy demand of generic consumers [4]. Realtime pricing in smart grid with demand of the consumers is
investigated in [7]. Most of the previous works are based on
DRM for general electricity consumers. Few works have dealt
with smart grid DRM for cellular networks. However, they
are based on instantaneous network statistics [8], [9] without
considering fairness in the supply decision. In this paper,
a generalized α-fairness-based DRM framework for cellular
networks is proposed using average network statistics provided
by SG. The system model comprises a set of cellular operators,
characterized by the offered QoS to their subscribers, powered
by a common pool of energy suppliers. SG is exploited to
evaluate the total power requirements of the cellular networks.
Then, closed-form expressions for the amount of energy
provided by each supplier to the operators are derived while
respecting the maximum tolerated CO2 emissions level set
by the environmental regulator. A generalized α-fair utility
function based on the profit of energy suppliers is used to avoid
production bias towards any single supplier. The optimization
is performed while taking into account the production capacity
of each supplier, their cost of generation, the dynamic pricing
strategy, and the pollutant emissions levels. Simulation results
investigate the energy suppliers’ behaviors according to different system parameters. Also, the obtained results show that
suppliers’ production decisions are adapted to the changing
demand of cellular networks to achieve maximum profitability
in an environment-friendly manner.
II. S YSTEM M ODEL
A. Network Geometry
The network consists of Nop mobile operators serving
their subscribers in an area of A km2 . The BSs of each
mobile operator l = 1, . . . , Nop are randomly deployed in R2
according to a homogeneous PPP denoted by Φ(l) with density
(l)
λBS BS/km2 . The total number of BSs that belong to each
(l)
(l)
operator is NBS = λBS A. Each mobile operator uses frequency
reuse within its network to avoid adjacent cell interference.
The set of BSs using the same frequency are, therefore,
separated by a certain exclusion distance denoted by R. The
location of the interfering BSs of each operator are modeled
by the Matérn hard core point process (MHCPP) obtained via
a dependent thinning of the original PPP. No two selected
BSs in the MHCPP are closer than the exclusion distance.
This new set is denoted by Φ̃(l) ⊆ Φ(l) (see [10] for details).
For tractability of analysis, the MHCPP is approximated by
an equi-dense PPP. The intensity of the MHCPP is evaluated
(l)
−λ
(l)
πR2
as ΛBS = 1−e πRBS2
, [11]. The subscribers of operator l are
(l)
uniformly distributed in R2 with density λu users/km2 . Each
user is associated to its nearest BS and, hence, the coverage
regions of BSs form a Voronoi tessellation [6]. The number
of users served by each BS of operator l are evaluated as
(l)
(l)
(l)
Nu = λu A/NBS .
B. Operator Characteristics
In this framework, each mobile operator is characterized
by the offered QoS, determined by the downlink signalto-interference-plus-noise ratio (SINR) thresholds T (l) and
the probability of coverage P (l) for l = 1, . . . , Nop . The
assumption is that a user connected to the BS of operator l, is
served only if the downlink SINR at the user exceeds the predefined threshold T (l) . To ensure that each operator meets the
set transmission quality standards, the transmit power levels of
its BSs need to be adjusted. The total transmit power from all
the BSs determines the energy requirements of the operator.
C. Channel Model
It is assumed that the BSs of operator l transmit with
(l)
an average power denoted by Pt for l = 1, . . . , Nop to
serve a single user. The power decays with the distance
according to the power law and, hence, the received power
at a test user located at the origin from a BS located at xj is
(l)
(l) (l)
(l)
(l)
given as, Pr = Pt hj kxj k−η , where hj represents the
channel fast fading effect and is assumed to be independent
and identically distributed (i.i.d.) exponential random variables
with mean µ−1 for each j, l, k.k represents the vector norm,
and η is the path-loss exponent. Without loss of generality, it
is assumed that the effects of shadowing are catered for by
the randomness of the BS locations. The downlink SINR at a
typical user located at the origin served by operator l can be
expressed as:
−η
(l) (l)
SINR(l) =
Pt h0 r(l)
,
P
(l) (l)
(l)
Pt hj kxj k−η
σ2 +
(1)
j∈Φ̃(l) \0
(l)
where h0 and r(l) are the channel fading and the distance
between the test user and the nearest BS of operator l, respectively, and σ 2 is the Gaussian noise power. The interference
P
(l)
(l) (l)
(l)
power denoted by Iagg is
Pt hj kxj k−η , where
j∈Φ̃(l) \0
Φ̃(l) \0 denotes the set of BSs of operator l excluding the BS
closest to the considered user.
D. Power Consumption Model for Mobile Operators
Each BS of operator l is considered to be equipped with a
single omni-directional antenna. Its average power consump(l)
tion, denoted by PBS , is expressed as follows [12]:
(l)
(l)
PBS = a Ptx + b,
(2)
(l)
Ptx
where
is the average power radiated by each BS of operator l. The coefficient a corresponds to the power consumption
that scales with the radiated power due to amplifier and feeder
losses. The term b models an offset of BS site power which
is consumed independently of the average transmit power and
(l)
is due to signal processing, battery backup, and cooling. Ptx
(l)
(l) (l)
can be evaluated as Ptx = Pt Nu .
E. Energy Pricing and Environmental Impact
In this study, it is assumed that the cellular networks are
powered by a smart grid where NR different sources exist to
supply energy with different prices and have different pollutant
levels depending on the nature of the generated energy. The
suppliers use a dynamic pricing strategy, whereby, the unit
charge of supplied energy depends directly on the energy
demand. In order to increase production, an energy supplier
needs to employ additional infrastructure and resources, thus
bearing additional expenses. Therefore, the price of energy
from supplier n charged to operator l, denoted by π (n,l) , is
expressed in monetary units (MU) as follows:
π
(n,l)
=w
(n)
q (n,l)
Q̄(n)
γ (n)
,
(3)
where w(n) is the benchmark price charged per unit by supplier
n, q (n,l) is the average energy provided by supplier n to
operator l, Q̄(n) is the maximum generation capacity of each
supplier, and γ (n) is a non-negative integer representing the
price sensitivity of supplier n. If γ (n) = 0, the dynamic pricing
is ignored and the suppliers charge a fixed price for energy
independently of the quantity purchased. Otherwise, the price
increases with the demand and vice versa.
The suppliers have different environmental impacts in terms
of CO2 emissions depending on the quantity and type of
energy generated. This can be quantified using the following
pollutant emissions function F (see [12]):
F(q (n,l) ) = ψn
Nop
Nop
2
X
X
q (n,l) + φn
q (n,l) ,
l=1
(4)
l=1
where ψn and φn are the pollutant emissions coefficients of
supplier n.
III. M ETHODOLOGY OF A NALYSIS
In this section, the probability of coverage for a typical user
in the network of operator l is determined using analytical
tools from SG. This helps us in identifying the transmit
power requirements of each operator to meet the transmission
quality standards and, hence, determining the average power
consumption of the operators. Consequently, an optimization
framework is developed to supply the required energy from the
smart grid in an environment-friendly yet profitable manner.
A. Operator Coverage Analysis
SG is employed to evaluate the probability that a typical
user of operator l is covered by the cellular network. The
coverage is defined as the probability that the received SINR
per user exceeds a target service threshold T (l) . In other words,
it corresponds to the complementary cumulative distribution
function (ccdf) of the SINR, denoted by P (l) , evaluated as
P (l) = P[SINR(l) > T (l) ] and it can be expressed as [6]:
(l) σ 2 r η
Z∞
− µT (l)
(l)
(l) −λ(l)
(l)
(l)
πr 2
Pt
BS
P (Pt , T ) = 2πrλBS e
×
e
0
e
(l)
−ΛBS πT (l)
∞
R
2/η 2
r
T (l)
−2/η
1
1+uη/2
du
dr.
(5)
Note that the probability of coverage depends on the transmitted power and the target SINR threshold of each operator.
Therefore, for a fixed P (l) , the transmit power can be compromised to achieve a higher SINR threshold and vice versa.
B. Problem Formulation and Solution
The decision on production quantity by the suppliers is
based on their profitability as well as the CO2 emissions.
The profitability of the suppliers depends on the quantity of
energy supplied to mobile operators Q, where Q is a matrix
of size NR × Nop with elements q (n,l) ∀ n = 1, . . . , NR , l =
1, . . . , Nop , the per unit price of energy set by suppliers π (n,l) ,
and the per unit cost of generation c(n) . The profit of supplier
n can be expressed as follows:
Πn (Q) =
Nop
X
q (n,l) π (n,l) − c(n) , ∀ n = 1, . . . , NR .
(6)
l=1
The environment friendliness of the suppliers is determined
by the pollutant emissions levels that depend on the amount
of energy supplied. It is quantified by the cost function
NR
P
C(Q) =
F(q (n,l) ). In order to reduce pollutant emissions,
n=1
the smart grid needs to produce more green energy from renewable sources while curtailing the use of fossil fuel sources.
However, the renewable energy generation might be expensive
and/or limited in amount and therefore, may lead to lower
profitability for the suppliers. Hence, the suppliers use multiple
energy sources depending on their cost, available amount,
and environmental impact in order to increase their revenues
while respecting the maximum tolerated CO2 emissions level
imposed by the regulator.
The objective now is to solve a constrained optimization
problem which results in an optimal matrix Q∗ that maximizes
a utility U based on the vector Π = [Π1 , Π2 , . . . , ΠNR ]
containing the individual profits of the energy suppliers. The
optimization problem is expressed as follows:
maximize U(Π)
(7)
Q
subject to
Nop
X
(l)
NBS q (n,l) ≤ Q̄n , ∀ n = 1, . . . , NR ,
(8)
l=1
NR
X
(l)
(l)
q (n,l) = NBS (a Pt Nu(l) + b)τ, ∀ l = 1, . . . , Nop ,
(9)
n=1
C(Q) ≤ Cth .
(10)
The constraint (8) enforces the total amount of energy produced by a supplier n not to exceed its maximum energy
production capacity Q̄n . Assuming that there are no losses,
the average energy provided by the suppliers should be equal
to the total power consumption of all BSs of operator l during
its operation time τ , as indicated in constraint (9). The total
(l)
transmit power for each operator Pt can be obtained by
solving (5) for a fixed QoS requirement P (l) . As the expression
of P (l) involves computation of a complicated integral which
does not have a closed form, it is solved numerically using
(l)
Newton’s method to obtain Pt , ∀l. In constraint (10), a
restriction is imposed on the pollutant emissions due to energy
consumption. This forces the suppliers to produce more green
energy to remain within regulated limits, denoted by Cth .
C. Utility Function
Since the objective is to determine the optimal amount
of energy provided by the suppliers in the smart grid to
multiple mobile operators, a fair power sharing framework
needs to be introduced. Therefore, a unified mathematical
formulation for power allocation known as α-fairness is used,
where the parameter α determines the degree of fairness of
profit distribution among the operators [13]. The generalized
α-fairness utility function of the supplier profits is defined as
follows:
NR 1−α
P Πn
,
if α ≥ 0, α 6= 1,
1−α
n=1
U(Π) =
(11)
N
R
P log(Πn ), if α = 1.
n=1
The parameter α controls the tradeoff between the efficiency
as well as the fairness in energy provided by the suppliers to
the operators. In other words, the framework tends to avoid
unfair allocation of energy depending on the value of α. For
example, situations where the suppliers only supply the most
expensive energy to meet the demand in order to achieve
maximum profit. In the special cases, when α = 0, 1, ∞, the
utility reduces to the sum, proportional fair and max-min fair
utility, respectively. These cases are described as follows:
Sum Utility (α = 0): The utility of this metric is equivalent
to the sum of the profits of the energy suppliers (U(Π) =
NR
P
Πn ). This approach promotes suppliers with higher profit
n=1
margins by offering them the possibility to supply the most
expensive energy to the operators at the least cost via smart
grid to maximize their profits. On the other hand, suppliers
with lower energy price will be deprived from the possibility
to sell the cheapest energy as their profits are already low.
Proportional fair utility (α = 1): The proportional fair metric
maximizes the sum of the log of the profit of suppliers. It is
equivalent to the maximization
geometric mean of the
QNR of the
Πn )1/NR ), which is equivalent
profits Πn i.e. (U(Π) = ( n=1
PNR
ln(Πn )). The proportional fair (PF) metric
to (U(Π) = n=1
is fairer since an energy supplier with a profit close to zero will
make the entire utility go to zero. Hence, this metric avoids
having a very low profit for any of the suppliers. In addition, it
reasonably promotes suppliers with good profit margins since
a high profit will contribute in increasing the product.
Max-Min Utility (α → ∞): Max-Min utilities are a family of
utility functions attempting to maximize the minimum profit
among all energy suppliers: U(Π) = min(Πn ). By increasing
n
the priority of suppliers having lower profit margins, MaxMin utilities lead to more fairness in the system. In order to
simplify the problem for this approach, a new decision variable
Πmin = min(Πn ) is introduced. Therefore, the optimization
n
problem becomes:
maximize
Πmin
(12)
Q,Πmin
subject to
(8), (9) and (10),
Πn ≥ Πmin , ∀n = 1, . . . , NR .
(13)
(14)
D. Solution
The optimization problem presented in eqs. (7) to (10) can
be solved for special cases using the Lagrangian method [14]
exploiting its strong duality property as follows:
minimize maximize
δ≥0
Q≥0
L(Q, δ, ξ, ζ),
(15)
where L(Q, δ, ξ, ζ) is the Lagrangian, δ = [δ1 , . . . , δNR ],
ξ = [ξ1 , . . . , ξNop ], and ζ are the Lagrange multipliers corresponding to the constraints in (8), (9), and (10) respectively.
Deriving the optimal amount of energy for a general value of
α is highly elaborate and is left for future work. The closed
form expressions for the special cases, i.e., α = 0, 1, ∞ are
provided here. However, the investigation of the performance
for different values of α is done via simulations in Section IV.
For the case of Sum utility, i.e., α = 0, the optimal amount
of energy supplied by supplier n to operator l, q (n,l)∗ is:
1
(l)
(l)
w(n) − c(n) − δn NBS − ζφn NBS + ξl ,
(l)
2ζψn NBSif γ = 0,
n
(n,l)∗
q
=
(l)
(l)
1
−c(n) − δn N
BS − ζφn NBS + ξl ,
(l) w(n)
2 ζψn NBS − Q̄n
if γn = 1.
We can deduce that q (n,l)∗
(16)
is inversely proportional to the
coefficient related to CO2 emissions of energy supplier n i.e.,
ζ. The Sum utility depends on the prices charged by the
suppliers where the higher the price, the higher the amount
of energy supplied from that source.
For the case of Proportional fair utility, i.e., α = 1, the
optimal amount of energy supplied by supplier n to operator
l is given as follows:
N
QR
(l)
1
w(n) − c(n)
Πm − δn NBS −
(l)
2ζψn NBS
m=1
m6=n
(l)
ζφn NBS + ξl , if γn = 0,
(n,l)∗
N
q
=
QR
(l)
1
c(n)
Πm + δn NBS +
N
R
Q
(n)
(l)
2w
m=1
Q̄n
Πm −2ζψn NBS
m6=n
m=1
m6=n
ζφn N (l) − ξl , if γn = 1.
BS
(17)
The supplied energy q (n,l)∗ depends directly on the product
of other supplier profits i.e., Πm . Hence, this approach avoids
having any supplier with very low profit and maximizes the
product of all suppliers profits simultaneously.
For the case of Max-Min fair utility, i.e., α → ∞, the
optimal amount of energy supplied by supplier n to operator
l is given as follows:
1
(l)
(l)
(n)
(n)
θ
(w
−
c
)
−
δ
N
−
ζφ
N
+
ξ
,
n
n
n
l
(l)
BS
BS
2ζψn NBS
if
γ
=
0,
n
q (n,l)∗ =
(l)
(l)
1
c(n) θn + δn N
−
ξ
,
+
ζφ
N
n
l
BS
BS
(n)
(l) w
2 ζψn NBS − Q̄n θn
if γn = 1.
(18)
θn are the additional Lagrange multipliers corresponding to
(14). By taking the derivative of
the Lagrangian with respect
PN
R
to Πmin , we can deduce that
n=1 θn = 1. By comparing
with (16), we can notice that θn control the priority of the
energy supply according to the profitability of the suppliers.
Note that if a fixed pricing strategy is used, i.e., γn = 0, the
amount of energy supplied is directly proportional to the price
of energy, whereas for dynamic pricing strategy, the amount
of energy supplied is inversely related to the price of energy.
This is because the energy price increases with demand under
dynamic pricing policy. Hence, the amount supplied should
decrease with increasing price.
The optimal Lagrange multipliers δ ∗ , ξ ∗ , ζ ∗ and θn∗ can
be obtained using the subgradient method (see [15]) or other
heuristic approaches. In order to achieve the optimal solution,
we can start with random initial values of the Lagrange
multipliers and evaluate the corresponding amount of energy.
Then, we update the Lagrange multipliers at the next iteration
with a step size updated according to the non-summable
diminishing step length policy. The values of the optimal
amount of energy supplied and the Lagrange multipliers are
updated until convergence.
IV. S IMULATION R ESULTS
A cellular network with Nop = 3, denoted by Op. 1, Op.
2 and Op. 3, is considered. The BSs of each operator are
distributed uniformly in R2 according to PPPs with BS density
(3)
Energy consumed (kJ)
100
90
80
70
60
50
Op. 1
Op. 2
Op. 3
40
30
10
11
12
13
14
15
SINR Threshold of operator 1, T1 (dB)
(a) Energy consumed.
Price of supplied energy π(n,l) (MU)
(2)
2
1.5
1
n=1, l=1
n=2, l=1
n=1, l=2
n=2, l=2
n=1, l=3
n=2, l=3
0.5
0
10
11
12
13
14
15
SINR Threshold of operator 1, T1 (dB)
(b) Price of energy.
Fig. 1: Price of energy charged by suppliers to operators and the total
energy consumed by operators versus SINR threshold of Op. 1. The
simulation parameters are: α = 0, γ1 = γ2 = 1, Q̄1 = Q̄2 = 170
kJ, Cth = 4.5 × 107 kg/h.
140
120
Supplier Profit (MU)
(1)
λBS = λBS = λBS = (π(200)2 )−1 BS/m2 respectively. The
simulation area A is chosen to be a square region of side 10
km2 in R2 . The subscribers of the operators are also uniformly
(1)
distributed in the simulation area with intensity λu = 15
(2)
(3)
2
2
users/km , λu = 30 users/km , and λu = 40 users/km2 .
The channel is assumed to experience exponential fading with
a mean channel gain µ = 1. The operators are differentiated
by the offered SINR levels and the QoS. The SINR levels
offered by the three operators are T (1) = 15 dB, T (2) = 10
dB, and T (3) = 5 dB while the QoS requirements are specified
as P (1) = 0.7, P (2) = 0.8, and P (3) = 0.9. Without loss of
generality, the BS operation time τ is set to be 1 s.
All BSs use the same power model [12] with a = 7.84 and
b = 71.5. The path loss exponent η = 4 and the noise power
is σ 2 = −115 dB. The BSs procure energy from a smart grid
in which there are several available energy suppliers offering
different prices and CO2 emissions levels. It is assumed that
there are three available energy suppliers (i.e., NR = 3),
namely; Sup. 1, Sup. 2, and Sup. 3. They are characterized as
follows: Sup. 1 provides the cheapest energy at a benchmark
price of w(1) = 1 MU but with the highest carbon footprint,
i.e., pollutant coefficients ψ1 = 0.004 and φ1 = 0.001, Sup.
2 provides energy with an intermediate benchmark price and
lower carbon footprint (w(2) = 2 MU, ψ2 = 0.002, φ2 =
0.0005), and Sup. 3 offers clean but expensive energy (i.e.,
w(3) = 3 MU, ψ3 = 0, φ3 = 0.0001). The maximum CO2
emissions level is taken as Cth = 107 kg/h. The costs of energy
production for the suppliers is c(1) = 0.1 MU, c(2) = 0.5 MU,
and c(3) = 2.5 MU. The base profit margin of the suppliers
is, therefore, 0.9 MU, 1.5 MU and 0.5 MU, respectively. Each
supplier has a maximum energy production capacity which is
assumed to be equal (i.e., Q̄1 = Q̄2 = Q̄3 = 150 kJ), unless
otherwise stated.
The objective is to study the behavior of the system while
changing system parameters and thresholds. Fig. 1 plots the
price of energy charged to operators and the total energy
consumed by the operators versus the SINR threshold of Op.
1. For ease of exposition, NR = 2, i.e., Sup. 1 and Sup.
3, is used in Fig. 1. We set γn = 1, which means that the
dynamic pricing is in effect, and α = 0, which means that
the Sum utility is used. It can be observed from Fig. 1(a),
that an increase in transmission quality standards of a mobile
operator increases its energy consumption due to additional
transmit power required. This increased demand triggers a
response from the suppliers as shown in Fig. 1(b). Since Op. 3
is the most power hungry operator, it is provided with the most
expensive energy followed by Op. 2 and Op. 1, respectively.
Op. 1 and Op. 2 are completely powered by the cheapest
energy supplier with high carbon footprint while Op. 3 is
powered by both suppliers. As the power requirements of
Op. 1 increase, its price also increases due to the dynamic
pricing policy. To supply more energy to Op. 1 from Sup. 1,
the framework progressively cuts down the energy supplied
by Sup. 1 to Op. 3 and increases the energy supplied by Sup.
2 to Op. 3. As soon as CO2 emissions threshold is reached,
Op. 3 suffers further by completely being supplied by Op.
2 with very expensive energy. Further increase in demand is
100
Sup. 1 (α = 0)
Sup. 2 (α = 0)
Sup. 3 (α = 0)
Sup. 1 (α = ∞)
Sup. 2 (α = ∞)
Sup. 3 (α = ∞)
80
60
40
20
5
10
15
SINR Threshold of operator 1, T1 (dB)
Fig. 2: Profit of energy suppliers versus SINR threshold of Op. 1.
compensated by providing Op. 2 with more green expensive
energy. Finally, the framework shows an abrupt change in
the production decision after T1 = 14.8 dB. This is because
Op. 1 becomes the most power hungry operator as evident
from Fig. 1(a). Therefore, it starts to receive green expensive
energy and less cheap energy. The results in Fig. 1 motivate
the introduction of fairness to the framework to avoid price
exploitation of any single operator.
In the sequel, the price sensitivity of suppliers γn = 0, ∀n =
1, . . . , NR and α = 0.5 is used unless otherwise stated. Fig. 2
plots the profit of the suppliers against the SINR threshold of
Op. 1 for extreme values of α. It can be observed that for
α = 0, i.e., Sum utility case, all the additional share of profit
is received by Sup. 1 since it has the lowest profit among
the suppliers. However, for α = ∞, i.e, Max-Min fair case,
operator 2 already receives a higher profit and the additional
share of the profit is given to Sup. 2 and Sup. 3.
Energy Production (kJ)
140
the distribution of profit is biased. The supplier with the
highest profit margin, i.e., Sup. 2 receives the highest share
and the supplier with the lowest profit margin, i.e., Sup. 1
receives the lowest share. For α = 1, i.e., Proportional fair,
the distribution of profits is according to the profit margins,
however, the discrepancy is smaller. As α is increased, the
share of the profits equalizes further to increase fairness.
Sup. 1
Sup. 2
Sup. 3
120
100
80
60
40
0.5
V. C ONCLUSION
1
1.5
2
(3)
Cost of energy production of supplier 3, c
2.5
(MU)
(a) Energy production by suppliers.
350
Sup. 1
Sup. 2
Sup. 3
Supplier Profit (MU)
300
250
200
150
100
50
0
0.5
1
1.5
2
(3)
Cost of energy production of supplier 3, c
2.5
(MU)
(b) Profit of energy suppliers.
Fig. 3: Energy production by suppliers and profit of suppliers versus
production cost of Sup. 3 for α = 0.5.
In this paper, a fair and optimized framework is proposed
for energy production decisions made by suppliers in the smart
grid powering cellular networks. Multiple energy suppliers
are available in smart grid offering different energy prices
and pollutant emissions. The goal is to first assess the power
requirements of the mobile operators, for a given quality
of transmission and the number of subscribers. Then, the
framework needs to meet these requirements by intelligently
providing energy from different suppliers in order to maximize
the profitability of the suppliers. Stochastic geometry is used
to quantify the network performance from the perspective of a
typical user in terms of the average transmission power of the
networks. It is observed that using the proposed framework,
suppliers adapts their energy production decisions under different situations such that the CO2 emissions levels of the
network do not exceed the set threshold and generate the
optimal amount of energy for maximum profitability.
180
Supplier Profit (MU)
R EFERENCES
Sup. 1
Sup. 2
Sup. 3
160
140
120
100
80
60
40
20
0
0.5
1
1.5
2
Level of fairness (α)
Fig. 4: Profit of energy suppliers versus the fairness level α.
In Fig. 3(a), the impact of increasing the cost of energy
production of Sup. 3 on the amount of energy produced by
the suppliers is studied for T1 = 15 dB. It can be observed that
increasing the cost of production reduces the profit margin of
Sup. 3, thus leading to a reduction in its energy production.
In order to meet the same energy demand, the production of
Sup. 2 increases since Sup. 2 has the highest profit margin.
An increase in the production of Sup. 1 is also observed
since the fairness level α = 0.5 does not allow any single
supplier to exploit the demand. The corresponding profit of
the suppliers is shown in Fig. 3(b). It is shown that the Sup.
3 profit is linearly falling as the cost increases while there
is a gentle increase in Sup. 1 and Sup. 2 profits due to
increased production.
Finally, in Fig. 4, the effect of the fairness level α on the
the profits of the suppliers is demonstrated. It is apparent
that Sup. 2 achieves the highest profit due to its high profit
margin followed by Sup. 3 and Sup. 1 in order of their profit
margins. As α is increased, the distribution of profits among
the three suppliers becomes fairer. Particularly, when α = 0,
[1] “5G radio access technology and capabilities,” Ericsson white paper.
Available at: http://www.ericsson.com/res/docs/whitepapers/wp-5g.pdf.
[2] A. Fehske, G. Fettweis, J. Malmodin, and G. Biczok, “The global
footprint of mobile communications: The ecological and economic
perspective,” IEEE Commun. Mag., vol. 49, no. 8, pp. 55–62, Aug. 2011.
[3] M. J. Farooq, H. Ghazzai, and A. Kadri, “Optimized energy procurement
for cellular networks powered by sart grid based on stochastic geometry,”
in Proc. of IEEE Global Commun. Conf. (Globecom workshops), San
Diego, USA, Dec. 2015.
[4] L. Jiang and S. Low, “Real-time demand response with uncertain
renewable energy in smart grid,” in Proc. of 49th Annual Allerton Conf.
on Commun., Control, and Comput. (Allerton), IL, USA, Sept. 2011.
[5] M. Haenggi, “Stochastic geometry for wireless networks,” Cambridge
University Press, 2012.
[6] J. Andrews, F. Baccelli, and R. Ganti, “A tractable approach to coverage
and rate in cellular networks,” IEEE Trans. Commun., vol. 59, no. 11,
pp. 3122–3134, Nov. 2011.
[7] P. Tarasak, “Optimal real-time pricing under load uncertainty based on
utility maximization for smart grid,” in Proc. of IEEE Intl. Conf. on
Smart Grid Commun. (SmartGridComm), Brussels, Belgium, Oct. 2011.
[8] H. Ghazzai, E. Yaacoub, and M.-S. Alouini, “A game theoretical
approach for cooperative environmentally friendly cellular networks
powered by the smart grid,” in IEEE Online Conf. Green Commun.
(OnlineGreencomm), pp. 1 – 6, Nov. 2014.
[9] S. Bu, F. Yu, Y. Cai, and X. Liu, “When the smart grid meets energyefficient communications: Green wireless cellular networks powered by
the smart grid,” IEEE Trans. Wireless Commun., vol. 11, pp. 3014–3024,
Aug. 2012.
[10] H. ElSawy, E. Hossain, and M. Haenggi, “Stochastic geometry for
modeling, analysis, and design of multi-tier and cognitive cellular
wireless networks: A survey,” IEEE Commun. Surveys Tuts., vol. 15,
no. 3, pp. 996–1019, Third 2013.
[11] D. Stoyan, W. S. Kendall, and J. Mecke, “Stochastic geometry and its
applications,” Wiley series in probability and mathematical statisitics,
Wiley, 1987.
[12] H. Ghazzai, E. Yaacoub, M.-S. Alouini, and A. Abu-Dayya, “Optimized
smart grid energy procurement for LTE networks using evolutionary
algorithms,” IEEE Trans. Veh. Technol., vol. 63, no. 9, pp. 4508–4519,
Nov. 2014.
[13] E. Altman, K. Avrachenkov, and A. Garnaev, “Generalized alpha-fair
resource allocation in wireless networks,” in Proc. of 47th IEEE Conf.
on Decision and Control (CDC), Dec. 2008.
[14] S. Boyd and L. Vandenberghe, “Convex optimization,” Cambridge
University Press, NY, USA, 2004.
[15] S. Boyd and A. Mutapcic, “Subgradient methods,” Notes for EE364b,
Stanford University, Winter 2006-07, 2008.
| 3cs.SY
|
ROBUST STABILIZATION OF LINEAR PLANTS UNDER UNCERTAINTIES
AND HIGH-FREQUENCY MEASUREMENT NOISES1
Igor B. Furtat*,**. Artem N. Nekhoroshikh, *,**
*Institute for Problems of Mechanical Engineering Russian Academy of Sciences, 61 Bolshoy ave V.O., St.-Petersburg,
199178, Russia (Tel: +7-812-321-47-66; e-mail: [email protected]).
**ITMO University, 49 Kronverkskiy ave, Saint Petersburg, 197101, Russia.
Abstract: The paper describes the robust algorithm for linear time-invariant plants under parametric
uncertainties, external disturbances and high-frequency noises in measurements. The proposed algorithm
allows one to reduce the noise impact on the output variable of the plant and to compensate parametric
uncertainties and external disturbances independently. The modeling results illustrate the effectiveness of
the algorithm.
Keywords: Robust control, linear plants, high-frequency noise, time delay, Lyapunov-Krasovskii
functional.
1. INTRODUCTION
Design of simple control systems under parametric
uncertainties and external disturbances when only plant
output is available for measurement is important problem of
control theory and practice. To construct such control
schemes many solutions have been proposed in this regard.
One of the most effective tools is to synthesize control
structure using high-gain observer. At first high-gain
observer was proposed in (Esfandiary and Khalil, 1992;
Gauthier et al., 1992) for minimum phase plants. Later other
kind of high-gain observer were considered in (Gauthier et
al., 1992; Teel and Praly, 1994; Bobtsov, 2002; Tsykunov,
2008; Furtat, 2015). In (Esfandiary and Khalil, 1992;
Gauthier et al., 1992; Teel and Praly, 1994; Bobtsov, 2002;
Furtat, 2015) the dimension of the high-gain observer is equal
to γ – 1, where γ is the relative degree of plant model.
However, using high-gain observer can be unsatisfactory in
case high frequency noise measurement application. The
investigations of the high-gain observers under noises were
considered in (Vasiljevic and Khalil, 2008; Boizot et al.,
2010; Sanfelice and Praly, 2011). The problem is that the
value of estimate derivative could be sufficiently greater than
the real one. Moreover, the error accumulates in further
estimation.
In (Ahrens and Khalil, 2009; Sanfelice and Praly, 2011;
Prasov and Khalil, 2013) adaptive high-gain observer was
proposed to partially overcome this problem. Thus, initially
1
high-gain parameter of the observer has a large value, while
in steady state mode high-gain parameter is decreased.
In (Astolfi and Marconi, 2015; Wang et al., 2015) an
extended high-gain observer was considered. The dimension
of the modified observer is 2γ – 2. Doubling dimension is
caused by the introduction of additional differential equations
reducing the impact of high frequency measurement noises.
The simulation results showed the effectiveness of the
modified algorithm as compared with the standard high-gain
observer. However, in (Astolfi and Marconi, 2015; Wang et
al., 2015) quality of estimate derivatives and quality of
filtering simultaneously depend on the solution of the
observer equation.
In the present paper we consider two independent algorithms:
filtering and control ones. Differently from (Astolfi and
Marconi, 2015; Wang et al., 2015), the proposed algorithm
allows one
1) to improve the quality of the estimation derivatives;
2) to calculate independently the filter parameters and the
parameters of the observer.
The paper is organized as follows. The problem statement is
presented in Section 2. In Section 3 we design the highfrequency filtering algorithm. In Section 4 we propose the
control algorithm for linear plants. In Section 5 we consider
simulation results and discuss an efficiency of the proposed
control structure. Concluding remarks are given in Section 6.
The results of Section 3 was developed under support of RSF (grant 14-29-00142) in IPME RAS. The results of Section 4
wsa supported solely by the Russian Federation President Grant (No. 14.W01.16.6325-MD (MD-6325.2016.8)). The other
research were partially supported by grants of Russian Foundation for Basic Research No. 16-08-00282, № 16-08-00686,
Ministry of Education and Science of Russian Federation (Project 14.Z50.31.0031) and Government of Russian Federation,
Grant 074-U01.
hx (t ) = f (x (t ), u(t ) ),
2. PROBLEM STATEMENT
(7)
where x ∈ R , f(x, u, h) is Lipschitz function in x and u, u is a
bounded function, h > 0 is a small coefficient. Let system (7)
be asymptotically stable when u = 0. Consider the set
Ω = {x : f ( x, u ) = 0}. Then there exists h0 > 0 for any ε > 0
such that for any h < h0 the following condition holds
s
Consider a plant model in the form
Q( p) z (t ) = kR( p )u (t ) + f (t ), y (t ) = z (t ) + w(t ),
(1)
where y(t) ∈ R is an output, u(t) ∈ R is an input, f(t) ∈ R is a
unmeasured bounded disturbance, w(t) ∈ R is a high
frequency bounded noise, Q(p), R(p) are linear differential
operators with unknown coefficients, deg Q(p) = n,
deg R(p) = m, k > 0, p = d / dt.
lim dist (x(t ), Ω ) < ε .
t →∞
The problem is to design the control system such that the
following condition holds
Let us verify conditions of Lemma for system (6). System (6)
is asymptotically stable for z = 0, because the matrix G is
Hurwitz. Substituting µ = 0 into (6), we get Gξ = –Bz or
ξ1 = z and ξi = ξi + 1, i = 2, r − 1 . Thus, ~
y = z . According to
Lemma, there exists µ > 0 such that for any µ ≤ µ0 the
following condition holds
lim z (t ) < δ1 ,
lim ~y (t ) − z (t ) < δ 4 ,
Assume that the coefficients of operators Q(p), R(p) and
coefficient k > 0 belong to a known compact set Ξ. The
polynomial R(λ) is Hurwitz, where λ is a complex variable.
(2)
t →∞
where δ1 > 0 is a required accuracy, hereinafter |⋅| is an
Euclidean norm.
3. HIGH FREQUENCY FILTERING ALGORITHM
Reject signal w from signal y. To this end, introduce the
following algorithm
µξ (t ) = Gξ(t ) + By (t ), yˆ (t ) = Lξ(t ), ξ(0) = 0,
where
− σ 1−1
−1
σ2
G= 0
0
[
0
0
− σ −21
0
−1
3
σ
0
]
−σ
0
−1
3
0
0
0
σ
−1
r
(3)
ξ = [ξ1, ξ2, ..., ξr]T,
0
0
0 ,
σi > 0,
− σ −r 1
B = σ , 0, ..., 0 , µ > 0 is a sufficient small coefficient,
L = [0, ..., 0, 1].
Th eo re m 1 . Let signal z be bounded and the following
condition holds
t
µ −1 lim ∫ eµ
t →∞
where δ4 > 0 is a sufficiently small coefficient.
Consider the signal y consisting of the signal z and the noise
w. Find a condition such that (5) will be hold.
~
Taking into account (3) and (6), rewrite the error ζ = ξ − ξ in
the following form
1
1
ζ (t ) = Gζ (t ) + Bw(t ), yˆ (t ) − ~
y (t ) = Lζ (t ).
µ
µ
−1
G (t − s )
Bw( s )ds < δ 2 ,
(4)
0
where δ2 > 0 is a sufficiently small coefficient. Then there
exists a coefficient µ0 > 0 such that for any µ ≤ µ0 the
following condition holds
lim yˆ (t ) − z (t ) < δ3 .
t →∞
(5)
Here δ3 > 0 is a sufficiently small coefficient.
Pro o f o f Th eo rem 1 . Consider plant (3) with input signal
z:
~
~
~
~
µ ξ (t ) = G ξ (t ) + Bz (t ), ~
y (t ) = L ξ (t ), ξ (0) = 0.
(6)
For analysis of plant (6) let us use Lemma (Furtat, 2014;
Furtat et al., 2014; Furtat et al., 2015).
Le m ma. Consider a plant model
(9)
The solution of the first equation of (9) is
ζ(t ) = e µ
t
−1
Gt
∫
ζ(0) + µ −1 e µ
−1
G (t −s )
Bw( s )ds
0
t
∫
= µ −1 e µ
T
−1
1
(8)
t →∞
−1
G (t −s )
(10)
Bw( s )ds.
0
If condition (4) holds, then we have from (10) that
lim ζ (t ) < δ 2 .
(11)
t →∞
Consider the following relations
yˆ (t ) − z (t ) ≤ yˆ (t ) − ~y (t ) + ~y (t ) − z (t )
≤ ζ(t ) + ~y (t ) − z (t ) .
(12)
Obviously, the upper bounds of (12) will be satisfied
y (t ) − z (t ) .
lim yˆ (t ) − z (t ) ≤ lim ζ (t ) + lim ~
t →∞
t →∞
t →∞
(13)
Let δ2 + δ4 ≤ δ3. Taking into account (8) and (11), we get
estimate (5) from (13). Theorem 1 is proved.
Let noise w be sinusoidal signals
w(t ) = ∑ Ai sin (ω i t + ϕ i ),
v
(14)
i =1
where Ai, ωi and ϕi are the amplitude, the frequency and the
phase accordingly.
Th eo re m 2 . Let z be bounded function and noise w be
signal (14). Then there exists µ0 > 0 such that for any µ ≤ µ0
the following condition holds
4. SYNTHESIS OF CONTROL SYSTEM
Let us use the algorithm (Furtat, 2015) to synthesize the
control system. According to Problem Statement, only the
output signal y(t) is available for measurement. Therefore,
introduce the control law as follows
r
v
1
lim yˆ (t ) − z (t ) ≤ ∑ Ai ∏
(15)
+ δ4 .
t →∞
2
2
2
i =1
j =1 1 + ωi µ σ j
γ −1
(20)
u (t ) = −α d i y ( i ) (t ) ,
Proof of Theorem 2. It follows from the proof of Theorem 1,
i =0
that there exists µ > 0 for system (6) such that for any µ ≤ µ0
the condition (8) holds. Taking into account (3) and (6), write where α > 0 and d0, d1, …, d γ −1 are chosen such that the
~
the error ζ = ξ − ξ in the form (9). Rewrite system (9) as
D(λ) = d γ −1λγ −1 + d γ − 2 λγ − 2 + ... + d1λ + d 0
is
polynomial
∑
ζ 1 (t ) = −µ −1σ1−1ζ1 (t ) + µ −1σ1−1 ∑ Ai sin (ωi t + ϕi ),
v
i =1
(16)
ζ j (t ) = −µ −1σ −j1ζ j (t ) + µ −1σ −j1ζ j −1 (t ), j = 2, r.
Substituting (20) into (1), we get
The solution of the first equation in (16) is
ζ1 (t ) = µ −1σ1−1
t
v
∑A ∫e
i
i =1
+e
−µ −1σ1−1t
F ( p ) z (t ) = αkR( p ) g (t ) + αkR( p ) D( p )ψ (t ) + f (t )
sin (ωi s + ϕ i )ds
∑
γ −1
g (t ) = D( p) yˆ (t ) − ∑ di y (i ) (t ) , ψ(t ) = z (t ) − yˆ (t ) . The value of
i =0
Ai
1 + ωi2 µ 2 σ12
i =1
ϑ1,i = arccos
sin(ωi t + ϕ i + ϑ1,i ),
1
1 + ωi2 µ 2 σ12
(17)
.
ζ 2 (t ) = e −µ
v
∑
i =1
−1 −1
σ2 t
ζ 2 ( 0)
t
Ai
1 + ωi2 µ 2 σ12
∫
eµ
y (t ) = yˆ (t ),
(22)
y ( j −1) (t ) − y ( j −1) (t − h)
, j = 1, γ − 1.
h
Substituting (22) into (20), rewrite the control law (20) in the
form
−1 −1
σ2 ( s −t )
y ( j ) (t ) =
0
× sin(ωi s + ϕ i + ϑ1,i )ds
v
=
∑ A (1 + ω µ σ )(1 + ω µ σ )
1
i
2 2 2
1
i
i =1
2 2
i
× sin (ωi t + ϕ i + ϑ1,i + ϑ 2,i ), ϑ 2,i = arccos
γ −1
d
u (t ) = −α ∑ ii
i =0 h
2
2
1
1+
ωi2 µ 2 σ 22
where Ci j =
.
Therefore, the solution of the rth equation of (16) is
r
1
ζ r (t ) = ∑ Ai ∏
i =1
j =1 1 + ωi2 µ 2 σ 2j
v
where ϑ j ,i = arccos
r
ϑ j ,i ),
sin(ωi t + ϕi + ∑
j =1
1
1 + ωi2µ 2σ 2j
g(t) depends on estimation quality of the signal yˆ (t ) and its
derivatives, the value of ψ(t) depends on quality of (3). Since
the set Ξ is known, then there exist the number α and the
polynomial D(λ) such that the polynomial F(λ) is Hurwitz.
To implement the control law (20) we use the following
observer
Substituting (17) into the second equation of (16), we get
+ µ −1σ 2−1
(21)
F(p) = Q(p) + αkR(p)D(p),
where
0
v
ζ 1 ( 0) =
µ −1σ1−1 ( s −t )
Hurwitz, γ = n – m is a relative degree of (1), y (i ) (t ) is an
estimate of the ith derivative signal yˆ (t ) , i = 0, 1, ..., γ – 1.
(18)
.
Upper bound of (18) is
r
v
1
ζ r (t ) ≤ ∑ Ai ∏
(19)
.
2
2
2
i =1
j =1 1 + ωi µ σ j
Takin into account (8), (13) and (19), we get estimate (15).
Theorem 2 is proved.
i
∑ (−1) j Ci j yˆ (t − jh) ,
j =0
(23)
i!
.
j!⋅(i − j )!
Th eo re m 3 . Let w be a bounded signal. Additionally let (γ –
r – 1)th derivatives of w be bounded, if r ≤ γ . Then there
exist coefficients α > 0 and h > 0 such that the control system
consisting of filtering algorithm (3) and control law (23)
ensures goal (2).
Proof of Theorem 3. Transform system (21) to the form
ε (t ) = Aε(t ) + αkB1 g (t ) + αkB2ψ (t ) + B3 f (t ), z (t ) = Jε(t ) , (24)
where ε = [ε1, ε2, ..., εn]T, z(i) = εi + 1, i = 0, n − 1 , matrixes A,
B1, B2, B3 and J = [1, 0, …, 0] are obtained at the transition
from (21) to (24). Rewrite system (3) as the following
differential equation
∏ (µσi p + 1)yˆ (t ) = y (t ).
r
i =1
Transform (25) to the state space form
(25)
θ (t ) = Mθ(t ) + Ny (t ), yˆ (t ) = Jθ(t ),
(26)
where θ = [θ1, θ2, ..., θr]T, yˆ (i ) = θi +1 , i = 0, r − 1 , matrix M
and vector N are obtained at the transition from (25) to (26).
Consider two cases.
1) Let r < γ. Rewrite the operator D(p) in the following form
[
D( p ) = ρ1T p γ −1 ,
[
+ρ p
T
2
r −1
,
p γ −2 , ,
p
r −2
]
pr
]
(27)
, , 1 ,
where ρ1 and ρ2 are vectors with corresponding coefficients
of the operator D(p). Taking into account (27), rewrite
function g(t) in the form
γ −1
g (t ) =
∑
γ −1
ρ1, j Jθ ( j ) + ρ T2 θ(t ) −
j =r
d
ii
−
i =1
h
∑h
di
i =0
(28)
( −1) C i Jθ(t − jh ),
j =1
γ −1
i
∑ ∑
j
j
θ( j ) = M j θ +
∑M
j −i −1
+
∑M
j −i −1
j −1
Ny (i ) = M j θ +
i =0
∑M
Nw
∑M
N , Ο, ..., Ο],
= M θ + G jε +
j
j −i −1
Nz (i )
[
where G j = N , MN , ..., M
(29)
j −i −1
d
0 (−1) j +1 αkB1 ii Ci j J
, i = 1, γ − 1, j = 1, i,
Fij =
h
0
0
γ −1 j −1 αkρ B JM j − i −1 N
B3
0
1, j 1
(i )
ϑ = w + ∑ ∑
w + f .
j =r i =0
0
N
0
γ −1 i
ε p (t ) = Ap ε p (t ) + ∑ ∑ Fij ε p (t − jh) + ϑ(t ),
γ −1 i
Nw ,
O is the zero
γ −1
∑∑ (ε
∑
γ −1
i
γ −1 d
i
ε (t ) = Aε(t ) − αkB1
i
i =1 h
where χ > 0.
γ −1
+ ρ T2 θ(t ) −
∑h
i =0
di
i
− jh )
− jh ) N ij ε p (t − jh )
)
(34)
Consider upper bounds of terms in (34)
∑ ∑
i
j
j =1
j −1
C i j Jθ(t − jh )
∑M
i =0
ij p ( t
T
T
p ( t ) N ij ε p ( t ) − ε p (t
2ε Tp (t ) P
∑
∑∑ F ε
+ 2ε Tp (t ) Pϑ(t ).
M j −i −1 Nw (i ) + ρ T2 θ(t )
i =0
γ −1
+ αkB1 ρ1, j J M j θ + G j ε +
j =r
i
i
j −1
∑ ∑ (−1)
γ −1
i =1 j =1
γ −1
∑
ApT P + PAp = −Q,
i =1 j =1
(30)
d i
di
ii
Jθ(t ) −
−
( −1) j Ci j Jθ(t − jh ).
i
h
h
i =1
j =1
i =0
It follows from (30), that derivatives of the signal w should
be bounded up to the (γ – r – 1)th order for r ≤ γ . Taking into
account (23) and (30), transform equation (24) to the form
γ −1
(33)
along trajectory of system (32), we get
V = −ε Tp (t )Qε p (t ) + 2ε Tp (t ) P
Substituting (29) into (28), we get
∑
0
i =1 j =1 − jh
+
g=
ρ1, j J M j θ + G j ε +
j =r
(32)
V = ε Tp (t ) Pε p (t ) + ∑ ∑ ∫ ε Tp (t + s ) N ij ε p (t + s )ds,
(i )
matrix.
γ −1
T
NJ
,
M
T
Q = QT > 0, N ij = N ij > 0 . Taking the derivative of (33)
i =0
j −1
γ −1
A + αkB1 ∑ ρ1, j JG j + αkB2 J
j =r
Ap =
γ −1
αkB ρ JM j + ρT − γ −1 di J − αkB J
∑ i
2
1 ∑ 1, j
2
i =0 h
j =r
where P = PT > 0 is the solution of
i =0
j −1
(i )
T
Consider Lyapunov-Krasovskii functional
i =0
j −1
]
i =1 j =1
where ρ1,j is the jth element of the vector ρ1. Taking into
account (26), we find the jth derivative (j ≥ 1) of θ in the
following form
j −1
[
ε p (t ) = ε T (t ), θT (t ) ,
Here ϑ is a bounded signal. Taking into account the
notations, rewrite systems (26) and (31) in the following form
Jθ(t )
i
Denote
j −i −1
Nw (i )
Jθ(t ) + αkB2 Jε(t ) − αkB2 Jθ(t )
+ B3 f (t ).
∑∑ F ε
ij p ( t
− jh ) ≤ 0.5γ (γ − 1)χεTp (t ) P 2 ε p (t )
i =1 j =1
+ χ −1
γ −1
i
∑∑ ε
T
p (t
− jh )FijT Fij ε p (t − jh ),
i =1 j =1
2ε (t ) Pϑ(t ) ≤ χε Tp (t ) P 2 ε p (t ) + χ −1 ϑ(t ) ,
2
T
p
Taking into account upper bounds, rewrite (34) in the
following form
(31)
γ −1 i
V ≤ −ε Tp (t )Wε p (t ) − ∑ ∑ ε Tp (t − jh) Rij ε p (t − jh) + χ −1τ,
i =1 j =1
where
γ −1 i
W = Q − 0.5γ (γ − 1)χP 2 − χP 2 − ∑ ∑ N ij ,
i =1 j =1
−1
Rij = N ij − χ F Fij ,
T
ij
2
τ = sup ϑ(t ) .
t
(35)
Obviously, there exist coefficients α and χ such that W > 0
and Rij > 0. We get upper bound of (35) in the form
V ≤ −λ min (W )εTp (t )ε p (t ) + χ −1τ .
(36)
Here λ min (W ) is the smallest eigenvalue of matrix W. It
follows
from
(36),
that z (t ) ≤ ε p (t ) ≤ χ −1τ λ min (W ) ,
therefore, δ1 ≤ χ τ λ min (W ) .
y (t ) = yˆ (t ),
y (t ) = 20[ y (t ) − y (t − 0.05)],
(1)
[
]
p γ −2 , , 1 .
D( p ) = ρT p γ −1 ,
Taking into account (33), rewrite the signal g(t) as follows
[
(t ) = 20[y
]
(t − 0.05)].
y (t ) = 20 y (1) (t ) − y (1) (t − 0.05) ,
y ( 3)
2) Let r ≥ γ. Rewrite the operator D(p) in the following form
(41)
We use observer (22) for estimation of derivatives in (41).
Let h = 1/20. Then the observer (22) is rewritten in the form
( 2)
−1
)
(
u(t ) = −7 0.9 y (t ) + 1.5 y (1) (t ) + 2 y ( 2 ) (t ) + 0.5 y ( 3) (t ) .
( 2)
(t ) − y ( 2 )
(42)
In addition, compare algorithm (3), (41), (42) with the
classical high-gain observer (Esfandiary and Khalil, 1992)
and modified high-gain observer (Astolfi and Marconi, 2015;
Wang et al., 2015). The control laws (41) are the same in all
algorithms.
γ −1 d
i
di
Jθ(t ) − ∑ ii ∑ (−1) j Ci j Jθ(t − jh) . (37) 1) Introduce high-gain observer (Esfandiary and Khalil,
i
i =1 h j =1
i =0 h
1992):
Substituting (37) into (24), transform equation (24) to the
110
0 1 0 0
form
0 0 1 0
110 2 ⋅ 0.35
ξ(t ) +
γ −1 d
ξ (t ) =
110 3 ⋅ 0.05
0 0 0 1
ε (t ) = Aε(t ) + αkB1 ρ T θ(t ) − ∑ ii Jθ(t )
i =0 h
0 0 0 0
110 4 ⋅ 0.0024
(43)
γ −1 d i i
j
j
× ( y (t ) − [1 0 0 0]ξ(t ) ),
(38)
− αkB1 ∑ i ∑ (−1) C i Jθ(t − jh) + αkB 2 Jε(t )
i =1 h j =1
y (t ) = [1 0 0 0]ξ(t ), y (1) (t ) = [0 1 0 0]ξ(t ),
− B 2 Jθ(t ) + B3 f (t ).
y ( 2 ) (t ) = [0 0 1 0]ξ(t ), y ( 3) (t ) = [0 0 0 1]ξ(t );
Rewrite systems (26) and (38):
2) Consider modified high-gain observer (Astolfi and
γ −1 i
(39) Marconi, 2015; Wang et al., 2015):
ε p (t ) = Ap ε p (t ) + ∑ ∑ Fij ε p (t − jh) + ϑ(t ),
γ −1
g (t ) = ρ T θ(t ) − ∑
i =1 j =1
where
A + B2 J
Ap =
NJ
γ −1
d
αkB1 ρT − ∑ ii J − αkB2 J
,
h
0
=
i
M
B
ϑ = 3 f is a bounded function, matrix Fij structure
0
corresponds to matrix Fij one in (32).
Since system (39) structure is similar to system (32) one, than
further proof of the second case is similar to the first one.
Theorem 3 is proved.
5. EXAMPLE
4
)
+ q3 p 3 + q2 p 2 + q1 p + q0 z (t ) = u (t ) + f (t ),
y (t ) = z (t ) + w(t ).
0 0.5
110
+
([0
2
0 110 0.0525
0
η 3 (t ) =
0
1]η1 (t ) − [1 0]η2 (t ) ),
(44)
1
η3 ( t )
0
0 0.5
110
+
([0 1]η2 (t ) − [1 0]η3 (t ) ),
2
0 110 0.0171
Consider the plant in the following form
(p
0 1
0 0
η 1 (t ) =
η1 (t ) +
η2 (t )
0 0
0 1
0 0.5
110
+
( y (t ) − [1 0]η1 (t ) ),
2
0 110 0.16
0 1
0 0
η 2 (t ) =
η2 (t ) +
η3 ( t )
0 0
0 1
(40)
The set Ξ of parameters possible values in (40) is given by
inequalities:
–1 ≤ q3 ≤ 0.1, –2 ≤ q2 ≤ 2, –3 ≤ q1 ≤ 3, –1 ≤ q0 ≤ 1.
Additionally, |f(t)| ≤ 1.
We choose σi = 1 and µ = 0.01 in (3). The parameter r in (3)
will be determined in Table 1.
Let α = 7, d0 = 0.9, d1 = 1.5, d2 = 2 and d3 = 0.5. Then control
law (23) could be rewritten in the following form
y (t ) = [1 0]η1 (t ), y (1) (t ) = [1 0]η2 (t ),
y ( 2 ) (t ) = [1 0]η3 (t ), y ( 3) (t ) = [0 1]η3 (t ).
Let q3 = 0, q2 = 1, q1 = 1, q0 = 0, f(t) = 0, w(t) = sin(0.5⋅103t)
and z(0) = 1, z (0) = z(0) = 0 , z(0) = −1 in (40). Table 1
shows the maximum errors of estimation of the signal z(t)
derivatives at steady-state mode using control system (3),
(41), (42), control system (41), (43) and control system (41),
(44).
Table 1. The value of e (i ) (t ) = sup z (i ) (t ) − y (i ) (t ) , i = 0,3 at
t
steady-state mode (after 8 s) for the proposed algorithm,
algorithms (43) and algorithm (44)
Control system
e(t )
e (1) (t )
e (2 ) (t )
e (3 ) (t )
Control system
(41),
(43)
(high-gain
observer)
0.22
8.41
132.3
698
Control system
(41),
(44)
(modified highgain observer)
0.26
z
a
z
t, s
4.9
31.1
266.1
z
Proposed
control system
(3), (41), (42)
when r = 2 in
(3)
0.04
Proposed
control system
(3), (41), (42)
when r = 5 in
(3)
3⋅10–3
0.2
2
b
57
z
7.5⋅10–4
2⋅10–3
t, s
5⋅10–3
Table 1 shows that the proposed control algorithm can
significantly reduce the estimate error of derivatives of signal
z. However, the dynamical order of the proposed algorithm
for r = 2 is one less than the dynamical order of the algorithm
(44). Furthermore, it follows from Table 1 that increasing the
parameter r can improve the quality of derivative estimates.
q2 = 2,
q1 = 3,
q0 = 1,
f(t) = sin t,
Let
q3 = 0.1,
z(0) = 1,
w(t) = sin(0.5⋅103t) + sin(103t) + sin(104t),
z (0) = z(0) = 0 , z(0) = −1 in (40). Fig. 1 shows the
simulation result of z(t), z (t ) , z(t ) и z(t ) using the
proposed control algorithm (3), (41), (42) for r = 5 in (3). The
simulation results of z(t) and z(t ) are represented by
continuous curves and the simulation results of z (t ) and
z(t ) are represented by dashed ones.
The simulation results show (Fig. 1) that after 10 (s) the
(t ) do
absolute values of the signals z(t), z(t ) , z(t ) and z
not exceed 0.014. However, the absolute values of estimate
errors of z(t), z(t ) , z(t ) and z(t ) do not exceed 2⋅10–3.
6. CONCLUSION
In this paper the robust control algorithm under parametric
uncertainties, external bounded disturbances and highfrequency noises in measurement signal was proposed. For
synthesis of control algorithm we used the approach that
allows one to control independently the quality of noise
filtering and the quality of the error of stabilization of the
Fig. 1. The simulation results z(t), z (t ) (Fig. 1, a) and z(t ) ,
z(t ) (Fig. 1, b).
output variable. The simulation results show the effectiveness
of the proposed algorithm as compared as standard high-gain
observer (Esfandiary and Khalil, 1992) and modified highgain observer (Astolfi and Marconi, 2015; Wang et al., 2015).
REFERENCES
Ahrens J. and Khalil K. (2009). High-gain observers in the
presence of measurement noise: A switched-gain
approach. Automaica, 45, 936-943.
Astolfi D. and Marconi L. (2015). A high-gain nonlinear
observer with limited gain power. IEEE Trans.
Automatic Control, 60(11), 3059-3064.
Bobtsov, A.A. (2002). Robust output-control for a linear
system with uncertain coefficients. Automation and
remote control, 63(11), 1794–1802.
Boizot N., Busvelle E. and Gauthier J. (2010). An adaptive
high-gain observer for nonlinear systems. Automatica,
46, 1483-1488.
Esfandiary F. and H.K. Khalil (1992). Output feedback
stabilization of fully linearizable systems. Int. J. Control,
56(5), 1007-1037.
Furtat I., Fradkov A. and Tsykunov A. (2014). Robust
synchronization of linear dynamical networks with
compensation of disturbances. International Journal of
Robust and Nonlinear Control, 24(17), 2774-2784.
Furtat I.B. (2014). Adaptive predictor-free control of a plant
with delayed input signal. Automation and remote
control, 75(1), 139-151.
Furtat I.B. (2015). Robust static control algorithm for linear
objects, Automation and Remote Control, 76(3), 446457.
Furtat I.B., Fradkov A.L. and Liberzon D. (2015)
Compensation of disturbances for MIMO systems with
quantized output. Automatica, 60, 239–244.
Gauthier J., Hammouri H. and Othman S. (1992). A simple
observer for nonlinear systems application to bioreactors.
IEEE Trans. Automat. Contr., 37(6). 875-880.
Prasov A. and Khalil H. (2013). A nonlinear high-gain
observer for systems with measurement noise in a
feedback control framework. IEEE Trans. Automat.
Contr., 58(3), 569-580.
Sanfelice R. and Praly L. (2011). On the performance of
high-gain observers with gain adaptation under
measurement noise. Automatica, 47, 2165-2176.
Teel A.R. and Praly L. (1994). Tools for semiglobal
stabilization by partial state and output feedback. SIAM
Journal on Control and Optimization, 33(5), 1443-1488.
Tsykunov A.M. (2008) An algorithm of robust control of a
non-stationary
linear
plant
with
perturbation
compensation, Journal of Computer and Systems
Sciences International, 47(4), 527-534.
Vasiljevic L. and Khalil H. (2008). Error bounds in
differentiation of noisy signals b high-gain observers.
Syst. Contr. Lett., 57, 856-862.
Wang L., Astolfi D., Hongye S., Marconi L. and Isidori A.
(2015). Output stabilization for a class of nonlinear
systems via high-gain observer with limited gain power.
Proc. 1st IFAC Conference on Modelling, Identification
and Control of Nonlinear Systems, MICNON 2015, Saint
Petersburg, Russia. IFAC-PapersOnLine, 48(11), 730735.
| 3cs.SY
|
A COUNTEREXAMPLE TO QUESTIONS ABOUT BOUNDARIES,
STABILITY, AND COMMENSURABILITY
arXiv:1705.03984v3 [math.GR] 26 Jan 2018
JASON BEHRSTOCK
A BSTRACT. We construct a family of right-angled Coxeter groups which provide
counter-examples to questions about the stable boundary of a group, one-endedness
of stable subgroups, and the commensurability types of right-angled Coxeter groups.
I NTRODUCTION
In this short note we construct right-angled Coxeter groups with some interesting
properties. These examples show that a number of questions in geometric group
theory have more nuanced answers than originally expected. In particular, these
examples resolve the following questions in the negative:
‚ (Charney and Sisto): As is the case for right-angled Artin groups, do all (nonrelatively hyperbolic) right-angled Coxeter group have totally disconnected
contracting boundary?
‚ (Taylor): Given that all known quasigeodesically stable subgroups of the
mapping class group are virtually free, does it hold that in any (non-relatively
hyperbolic) group that all quasigeodesically stable subgroups have more
than one end?
‚ (Folk question): If a right-angled Coxeter group has quadratic divergence,
must it be virtually a right-angled Artin group?
We describe a family of graphs, any one of which is the presentation graph of
a right-angled Coxeter group which provides a counterexample to all three of the
above questions. We expect that in special cases, and perhaps in general with
appropriate modifications, there are interesting positive answers to these questions;
we hope this note will encourage the careful reader to formulate and prove such
results.
The construction we give was inspired by thinking about the simplicial boundary
for the Croke–Kleiner group, see [BH, Example 5.12] and [Tat2]. In the process we
give a quick introduction to a few topics of recent interest in geometric group theory,
for further details on these topics see also [ABD, BDM, Cha, CS, Cor2, DT, Tra].
Acknowledgments. Thanks to Ruth Charney, Alessandro Sisto, and Sam Taylor
for sharing with me their interesting questions and to the referee for some helpful
comments which led to Remark 2.3. Also, thanks to Mark Hagen for many fun
discussions relating to topics in this note.
Date: January 29, 2018.
1
RACG COUNTEREXAMPLE
2
1. T HE CONSTRUCTION
Let Γn be a graph with 2n vertices built in the following inductive way. Start with
a pair of vertices a1 , b1 with no edge between them. Given the graph Γn´1 , obtain
the graph Γn by adding a new pair of vertices an , bn to the graph Γn´1 and adding
four new edges, one connecting each of tan´1 , bn´1 u to each of tan , bn u. See Figure 1.
F IGURE 1. The graphs Γ14 (left) and Γ (right).
Note that Γn is a join if and only if n ď 3. More generally, ai , aj are contained in a
common join if and only if |i ´ j| ď 2.
For any m ě 5 choose n sufficiently large so that there exists a set of points
P “ tp1 , . . . , pm u Ă Γn with the property that for each 1 ď i ă j ď m the points
pi and pj are not contained in a common join in Γn . For example, in Γ14 we could
choose the vertices P “ ta1 , a4 , a7 , a10 , a13 u. For each 1 ď i ă m add an edge between
pi and pi`1 ; also, add an edge between pm and p1 . Call this new graph Γ. There are
many choices of Γ depending on our choices of n, m, and P; for the following any
choice will work.
Associated to any graph, one can construct the right-angled Coxeter group with
that presentation graph, this is the group whose defining presentation is given by:
an order-two generator, for each vertex of the graph, and a commutation relation,
between each of the generators associated to a pair of vertices connected by an edge.
Let W denote the right-angled Coxeter group whose presentation graph is the Γ
constructed above. In the next section, we record some key properties about the
group W and then, in the final section, apply this to the questions in the introduction.
2. P ROPERTIES
2.1. Quadratic divergence.
Proposition 2.1. The group W has quadratic divergence. In particular, this group is not
relatively hyperbolic.
Proof. It is easily seen that the graph Γn has the property that each vertex is contained in at least one induced square. It is also easy to verify that given any
pair of induced squares S, S 1 in Γn , there exists a sequence of induced squares
S “ S1 , S2 , S3 , . . . , Sk “ S 1 where for each 1 ď i ă k the squares Si and Si`1 share 3
vertices in common. This property, that there are enough squares to chain together
RACG COUNTEREXAMPLE
3
any pair of points, is called CFS; it was introduced in [DaT] and studied further in
[BFRHS, Lev].
Since Γ has the same vertex set as Γn , and every induced square in Γn is still
induced in Γ, it follows that Γ also has the CFS property.
Given any graph with the CFS property and which is not a join, the associated
right-angled Coxeter group has exactly quadratic divergence, see [DaT, Theorem 1.1]
and [BFRHS, Proposition 3.1].
The second statement in the proposition follows from the fact that any relatively
hyperbolic group has divergence which is at least exponential [Sis, Theorem 1.3].
2.2. Stable surface subgroups. An undistorted subgroup is said to be (quasi-geodesically)
stable if each pair of points in that subgroup are connected by uniformly Morse quasigeodesics, see [DT].
Proposition 2.2. W contains a closed hyperbolic surface subgroup which is stable.
Proof. Recall that in any right-angled Coxeter group an induced subgraph yields a
subgroup isomorpic to the right-angled Coxeter group of the associated subgraph.
Also, note that the right-angled Coxeter group associated to a cycle of length at
least 5 is a 2–dimensional hyperbolic orbifold group. Thus, the subgraph spanned
by P, which is a cycle of length m ě 5, yields a subgroup, H, which is isomorphic to
a 2–dimensional hyperbolic orbifold group.
By construction, the subgraph P doesn’t contain any pair of non-adjacent vertices
in a common join of Γ (by assumption we use the term join to mean non-degenerate
join, in the sense that both parts of the join have diameter at least 2); thus no special
subgroup of H is a direct factor in a non-hyperbolic special subgroup of W .
It is proven in [ABD, Theorem B] that, a subgroup of a hierarchically hyperbolic
space is stable if and only if it is undistorted and has uniformly bounded projections to the curve graph of every proper domain in some hierarchically hyperbolic
structure; in turn, this property is equivalent to the orbit of the subgroup being
quasi-isometrically embedded in the curve graph of the nest-maximal domain.
A hierarchically hyperbolic structure on right-angled Coxeter groups was built in
[BHS]; in that structure the curve graphs are contact graphs which were first defined
in [Hag1] where it was proven they are all quasi-trees. In the present context, a
domain can be thought of as a certain convex subcomplex of the Davis complex
arising from special subgroups and orthogonality holds when the Davis complex
contains the direct product of a given such pair.
Since a closed surface group can not be quasi-isometrically embedded in a quasitree, to verify the proposition a different hierarchically hyperbolic structure must be
used; for this we now rely on a method developed in [ABD] for modifying structures.
The construction given in [ABD, Theorem 3.11] modifies the structure, by removing
certain domains and augmenting the curve graphs associated to domains in which
they are nested. Indeed, passing to the new structure and back, it follows that to
verify a given subgroup has uniformly bounded projections in a particular structure,
it suffices to verify that the subgroup has uniformly bounded projections to the set of
domains W with the property that for each W P W, there exists a domain U which
is orthogonal to W and with infinite diameter curve graph (since all other domains
RACG COUNTEREXAMPLE
4
will be removed in the new structure while only changing the curve graph of the
nest-maximal domain).
In the standard hierarchically hyperbolic structure the only domains to which H
has unbounded projections are contained in the subcomplex of the Davis complex
associated to the special subgroup H. As noted above, since no pair of vertices in P
are contained in a common join in Γ, these domains all have the property that any
domains orthogonal to one of them must have uniformly bounded diameter. Thus,
applying the results of [ABD] just discussed, it follows that H is stable.
Now take a cover of the orbifold to get the desired closed hyperbolic surface.
Remark 2.3. We note that in the subgroup H constructed in Proposition 2.2, it follows easily from [KL, Proposition 3.3] (or, more explicitly, from [Tra, Theorem 1.1]),
that each infinite order element in H acts as a rank-one isometry of the Davis complex of W and is thus Morse. To see that such an argument alone is not enough
to prove stability of the subgroup, we note that in [OOS] they construct lacunary
hyperbolic groups which provide an obstruction. In particular, in [OOS, Theorem 1.12] it is proven that there exist infinite finitely generated non-hyperbolic
groups in which every proper subgroup is infinite cyclic and generate uniformly
Morse quasigeodesics.
The next result follows from Proposition 2.2 and [Cor1, Proposition 4.2].
Corollary 2.4. The right-angled Coxeter group W contains a topologically embedded circle
in its Morse boundary.
3. A PPLICATIONS
3.1. Morse boundaries. Charney and Sultan introduced a boundary for CAT(0)
groups which captures aspects of the negative curvature of the group [CS]. Their
construction was then generalized by Cordes to a framework which exists for all
finitely generated groups [Cor1]; in this general context it is known as the Morse
boundary. Charney and Sultan built examples of relatively hyperbolic right-angled
Coxeter groups whose boundaries are not totally disconnected [CS]. More generally,
it is now known that for hyperbolic groups, the Morse boundary coincides with
the hyperbolic boundary [Cor1, Main Theorem (3)]; using this it is easy to produce
many examples of hyperbolic and relatively hyperbolic right-angled Coxeter groups
whose boundary are not totally disconnected.
On the other hand, the Morse boundary of any right-angled Artin group AΓ
is totally disconnected. The two-dimensional case of this is implicit in [CS]. In
general, this fact follows from the fact that the contact graph has totally disconnected
boundary (since it is a quasi-tree) and [CH, Theorem F]. To see this recall that [CH,
Theorem F] provides a continuous map from the Morse boundary of AΓ to the
boundary of the contact graph of the universal cover of the Salvetti complex of AΓ ; at
which point the result follows since Morse geodesic rays lying at infinite Hausdorff
distance cannot fellow-travel in the contact graph, so this continuous boundary map
is injective.
Accordingly, Ruth Charney and Alessandro Sisto raised the question of whether
outside of the relatively hyperbolic setting, right-angled Coxeter groups all have
totally disconnected Morse boundary. The group W constructed above shows the
RACG COUNTEREXAMPLE
5
answer is no, since it is not relatively hyperbolic by Proposition 2.1 and its boundary
is not totally disconnected, by Corollary 2.4.
3.2. Stable subgroups. Examples of stable subgroups are known both in the mapping class group [Beh, DT] and in right-angled Artin groups [KMT]. In both of these
classes, all known examples of stable subgroups are virtually free; in the relatively
hyperbolic setting on the other hand, it is easy to construct one-ended stable subgroups. Sam Taylor asked whether there exist non-relatively hyperbolic groups with
one-ended stable subgroups. The example W is a non-relatively hyperbolic group
with one-ended stable subgroups by Proposition 2.1 and Proposition 2.2.
3.3. Commensurability. A well-known construction of Davis–Januszkiewicz [DJ]
shows that every right-angled Artin group is commensurable to some right-angled
Coxeter group. The following is a well-known problem:
Question 3.1. Which right-angled Coxeter groups are commensurable to rightangled Artin groups?
It is known that any right-angled Artin groups either has divergence which is
linear (if it is a direct product) or quadratic, see [BC] or [ABD` ]. Since divergence
is invariant under quasi-isometry, and hence under commensurability as well, this
puts a constraint on the answer to Question 3.1. Several people have raised the
question of whether every right-angled Coxeter group with quadratic divergence
is quasi-isometric to some right-angled Artin group. The group W shows that the
answer is no, since [Cor1, Main Theorem (2)] proves that the Morse boundary is
invariant under quasi-isometries, but Propositions 2.1 and 2.2 show that the group W
is a right-angled Coxeter group with quadratic divergence whose Morse boundary
contains an embedded circle, while the Morse boundary of any right-angled Artin
group is totally disconnected.
4. F URTHER QUESTIONS
In [BFRHS] it was established that for a large range of density functions that
asymptotically almost surely the random graph yields a right-angled Coxeter group
with quadratic divergence. More generally, it is known that among n–vertex graphs
with density greater than 1{n and bounded away from 1, that asymptotically almost surely any random graph of this type will contain a large induced polygon.
Accordingly, we expect that the proof of Corollary 2.4 could be used to verify:
Conjecture 4.1. For any density greater than 1{n and bounded away from 1, asymptotically
almost surely the random right-angled Coxeter group contains circles in its Morse boundary.
In particular, it is not virtually a right-angled Artin group.
The prevalence of right-angled Coxeter groups with quadratic divergence makes
the following an appealing (but likely very difficult) question:
Question 4.2. Classify right-angled Coxeter groups with quadratic divergence up
to quasi-isometry. Classify them up to commensurability.
The proof of Proposition 2.2 and Remark 2.3 indicate that there are some nontrivial constraints on what the set of domains could be in any potential hierarchically
hyperbolic structure on a lacunary hyperbolic group. Accordingly we ask:
RACG COUNTEREXAMPLE
6
Question 4.3. Is a lacunary hyperbolic group hierarchically hyperbolic if and only if
it is hyperbolic?
Further, in light of Proposition 2.2 and Remark 2.3 it would be interesting if the
following was true in general or under some moderate hypotheses:
Question 4.4. Let G be a hierarchically hyperbolic group with a subgroup H. If all
infinite cyclic subgroups of H are uniformly Morse, does that imply that H is stable?
R EFERENCES
[ABD]
Carolyn Abbott, Jason Behrstock, and Matthew G Durham. Largest acylindrical actions
and stability in hierarchically hyperbolic groups. AR X IV:1705.06219.
[ABD` ] A. Abrams, N. Brady, P. Dani, M. Duchin, and R. Young. Pushing fillings in right-angled
Artin groups. J. Lond. Math. Soc. (2) 87 (2013), 663–688.
J. Behrstock. Asymptotic geometry of the mapping class group and Teichmüller space.
[Beh]
Geometry & Topology 10 (2006), 1523–1578.
[BC]
J. Behrstock and R. Charney. Divergence and quasimorphisms of right-angled Artin groups.
Math. Ann. 352 (2012), 339–356.
[BDM] J. Behrstock, C. Druţu, and L. Mosher. Thick metric spaces, relative hyperbolicity, and
quasi-isometric rigidity. Math. Ann. 344 (2009), 543–595.
[BFRHS] J. Behrstock, V. Falgas-Ravry, M. Hagen, and T. Susse. Global structural properties of
random graphs. Int. Math. Res. Not. (2016). To appear.
J. Behrstock and M.F. Hagen. Cubulated groups: thickness, relative hyperbolicity, and
[BH]
simplicial boundaries. Geometry, Groups, and Dynamics 10 (2016), 649–707.
[BHS]
J. Behrstock, M.F. Hagen, and A. Sisto. Hierarchically hyperbolic spaces I: curve complexes
for cubical groups. Geometry & Topology, 21:1731–1804, 2017.
[Cha]
Ruth Charney. An introduction to right-angled Artin groups. Geom. Dedicata 125 (2007),
141–158.
Ruth Charney and Harold Sultan. Contracting boundaries of CATp0q spaces. J. Topol. 8
[CS]
(2015), 93–117.
[Cor1]
Matthew Cordes. Morse boundaries of proper geodesic metric spaces. AR X IV:1502.04376.
[Cor2]
Matthew Cordes. A survey on Morse boundaries & stability. AR X IV:1704.07598.
[CH]
Matthew Cordes and David Hume. Stability and the Morse boundary. To appear in Groups,
Geometry, and Dynamics (2016).
[DaT]
Pallavi Dani and Anne Thomas. Divergence in right-angled Coxeter groups. Trans. Amer.
Math. Soc. 367 (2015), 3549–3577.
[DJ]
Michael W. Davis and Tadeusz Januszkiewicz. Right-angled Artin groups are commensurable with right-angled Coxeter groups. J. Pure Appl. Algebra 153 (2000), 229–235.
[DT]
Matthew Gentry Durham and Samuel J. Taylor. Convex cocompactness and stability in
mapping class groups. Algebr. Geom. Topol. 15 (2015), 2839–2859.
[Hag1] Mark F. Hagen. Weak hyperbolicity of cube complexes and quasi-arboreal groups. J. Topol.
7 (2014), 385–418.
[KL]
M. Kapovich and B. Leeb. 3-manifold groups and nonpositive curvature. Geom. Funct. Anal.
8 (1998), 841–852.
[KMT] Thomas Koberda, Johanna Mangahas, and Samuel J Taylor. The geometry of purely loxodromic subgroups of right-angled Artin groups. AR X IV:1412.3663.
[Lev]
Ivan Levcovitz. Divergence of CAT(0) Cube Complexes and Coxeter Groups.
arXiv:1611.04378.
[OOS]
Alexander Yu. Ol1 shanskii, Denis V. Osin, and Mark V. Sapir. Lacunary hyperbolic groups.
Geom. Topol. 13(2009), 2051–2140. With an appendix by Michael Kapovich and Bruce Kleiner.
[Sis]
A. Sisto. On metric relative hyperbolicity. AR X IV:1210.8081.
[Tat2]
Mark F. Hagen. Tattoo. In preparation.
[Tra]
Hung Cong Tran. Purely loxodromic subgroups in right-angled Coxeter groups.
AR X IV:1703.09032.
RACG COUNTEREXAMPLE
L EHMAN C OLLEGE AND T HE G RADUATE C ENTER , CUNY, N EW Y ORK , N EW Y ORK , USA
E-mail address: [email protected]
7
| 4math.GR
|
arXiv:1603.07713v3 [math.GR] 18 May 2016
CHIRAL POLYHEDRA AND FINITE SIMPLE GROUPS
DIMITRI LEEMANS AND MARTIN W. LIEBECK
Abstract. We prove that every finite non-abelian simple group acts as
the automorphism group of a chiral polyhedron, apart from the groups
P SL2 (q), P SL3 (q), P SU3 (q) and A7 .
Keywords: finite simple groups, abstract chiral polyhedra, regular maps.
1. Introduction
Polyhedra and their generalisations to higher ranks, polytopes, are certain ranked partially ordered sets generalising geometric objects that have
been studied since the Greeks (see [24, Chapter 1]). Those polytopes whose
automorphism group acts transitively on maximal flags are called regular.
They have maximum possible rotational and reflectional symmetries. Those
that are chiral have maximum rotational symmetries but no reflections. It
is already known which finite simple groups are automorphism groups of
abstract regular polyhedra (see below for more details). The main purpose
of this article is to determine which finite simple groups are automorphism
groups of chiral polyhedra.
In order to state the main results we require precise definitions of the
above terms. Following [24], an (abstract) polytope (P, ≤) of rank n is
a partially ordered set with a rank function ranging from −1 to n and
satisfying the following properties. The elements of rank i are called the
i-faces of P. There exists a unique least face F−1 and a unique greatest face
Fn . The flags are the maximal totally ordered subsets of P and they must
all contain exactly n + 2 faces. Two flags Φ and Ψ are called adjacent if
they differ in exactly one face. They are called i-adjacent if this face is an
i-face. The poset P must be strongly connected, that is, every pair of flags
must be connected by a path of adjacent flags in P. Finally we require that
for any (i − 1)-face F and any (i + 1)- face G of P such that F ≤ G, there
are exactly two i-faces between F and G. If the rank of P is 3, we call P a
polyhedron.
An automorphism of (P, ≤) is a bijection of the faces of P that preserves
the order ≤. The set of all automorphisms of (P, ≤) with composition forms
a group called the automorphism group of (P, ≤) and denoted Γ(P). If Γ(P)
has a unique orbit on the flags of P, we say that P is regular. If Γ(P) has
two orbits such that any two adjacent flags belong to distinct orbits, we say
that P is chiral.
Date: May 20, 2016.
1
2
DIMITRI LEEMANS AND MARTIN W. LIEBECK
As defined for instance in [24], a C-group is a group G generated by pairwise distinct involutions ρ0 , . . . , ρn−1 which satisfy the following intersection
property:
∀J, K ⊆ {0, . . . , n − 1}, hρj | j ∈ Ji ∩ hρk | k ∈ Ki = hρj | j ∈ J ∩ Ki.
A C-group (G, {ρ0 , . . . , ρn−1 }) is a string C-group if its generators satisfy
the following relations:
(ρj ρk )2 = 1 ∀j, k ∈ {0, . . . n − 1} with | j − k |≥ 2.
In [24] it is shown that string C-groups and abstract regular polytopes are in
one-to-one correspondence. Every string C-group gives an abstract regular
polytope and, given an abstract regular polytope and one of its base flags,
one can construct a set of distinguished generators that, together with the
automorphism group of the polytope, give a string C-group. In particular,
the automorphism group of an abstract regular polyhedron is generated
by three involutions ρ0 , ρ1 , ρ2 , two of which commute (namely, ρ0 , ρ2 ). In
1980, it was asked in the Kourovka Notebook (Problem 7.30) which finite
simple groups have this property. This was solved by Nuzhin and others in
[25, 26, 27, 28, 23]: every non-abelian finite simple group can be generated
by three involutions, two of which commute, with the following exceptions:
P SL3 (q), P SU3 (q), P SL4 (2n ), P SU4 (2n ),
A6 , A7 , M11 , M22 , M23 , M cL.
Thus every finite simple group, apart from the above exceptions, is the
automorphism group of an abstract regular polyhedron.
Similarly, in [29], it is shown that for a finite group G, the chiral polyhedra
having G as automorphism group are in bijective correspondence with pairs
x, t ∈ G satisfying the following conditions:
(i) G = hx, ti;
(ii) t is an involution;
(iii) there is no involution α ∈ Aut(G) such that xα = x−1 , tα = t.
Our main result classifies those finite simple groups G possessing such generators:
Theorem 1.1. Let G be a non-abelian finite simple group, not A7 , P SL2 (q),
P SL3 (q) or P SU3 (q). Then there exist x, t ∈ G such that the following hold:
(i) G = hx, ti;
(ii) t is an involution;
(iii) there is no involution α ∈ Aut(G) such that xα = x−1 , tα = t.
As a consequence, for every nonabelian finite simple group G except A7 ,
P SL2 (q), P SL3 (q) or P SU3 (q), there exists an abstract chiral polyhedron
having automorphism group G. This result was already known to be true
for the Suzuki groups [16, 13], the small Ree groups [14], the alternating
groups (see Section 5 for details) and some small sporadic groups [12].
CHIRAL POLYHEDRA AND FINITE SIMPLE GROUPS
3
We now discuss the exceptions in Theorem 1.1. The groups P SL2 (q)
do not have pairs of elements x, t satisfying (i)–(iii) of the theorem; this
a consequence of a result of Macbeath [22], as observed by Singerman [31,
Theorem 3]. For the group A7 , an exhaustive computer search shows that
no pair of elements x, t satisfying (i)–(iii) of Theorem 1.1 exists – see [12].
Thus it remains to consider the groups P SL3 (q) and P SU3 (q); it will be
shown that these also do not possess generators x, t satisfying conditions
(i)–(iii) in a forthcoming paper [19].
The rest of the paper is devoted to the proof of Theorem 1.1. This is
divided into four cases: namely, the case where G is of exceptional Lie type
(Section 3), the case where G is classical (Section 4), the case where G is
an alternating group (Section 5), and the case where G is a sporadic group
(Section 6).
Observe that chiral polyhedra (and regular polyhedra) may be seen also
as regular maps. In a recent paper [15], Gareth Jones has studied much
further the link between automorphism groups of edge-transitive maps and
finite simple groups.
2. Preliminaries
In this section we prove some lemmas needed for the proof of Theorem 1.1,
The first two lemmas are straightforward hence we omit their proofs.
Lemma 2.1. Let G be a finite group, and suppose x, t ∈ G satisfy xt = x−1 .
Then the set {y ∈ G : xy = x−1 } is contained in the coset CG (x)t.
Lemma 2.2. Let G be a finite group with a subgroup A. Suppose M is a
maximal subgroup of G containing A, such that any two G-conjugates of A
that are contained in M are M -conjugate. Then the number of G-conjugates
of M containing A is |NG (A) : NM (A)|.
For a finite group G and a positive integer r, denote by Ir (G) the set of
elements of order r in G, and let ir (G) = |Ir (G)|.
The next two lemmas provide upper and lower bounds for the numbers
of involutions i2 (G) in groups of Lie type. Recall that a simple group of
Lie type over a finite field Fq can be written as (ḠF )′ , the derived group of
the fixed point group of a Frobenius endomorphism F of the corresponding
simple adjoint algebraic group Ḡ over F̄q .
We use the notation Lǫn (q) (where ǫ = ±) to denote P SLn (q) when ǫ = +,
and P SUn (q) when ǫ = −. Similarly, E6ǫ (q) denotes E6 (q) when ǫ = + and
2E (q) when ǫ = −.
6
Lemma 2.3. Let G be a finite simple group of Lie type over Fq , and write
G = (ḠF )′ as above. Let r be the rank of Ḡ, and N the number of positive
roots in the root system of Ḡ. Define
1
2
2
2
2 (N + r), if G is of type F4 , G2 or B2 ,
MG =
N + r, otherwise.
4
DIMITRI LEEMANS AND MARTIN W. LIEBECK
Then i2 (Aut(G)) < 2 q MG + q MG −1 .
Proof. This is [18, Prop. 1.3].
Lemma 2.4. Let G be a finite simple group of Lie type over Fq .
(i) If G is of classical type, then i2 (G) > 14 q NG , where NG is defined as
follows:
G
NG
ǫ
Ln (q) (ǫ = ±)
[ 12 n2 ]
P Sp2m (q), P Ω2m+1 (q) m2 + m
P Ωǫ2m (q) (ǫ = ±)
m2 − 1
(ii) If G is of exceptional type, then with one exception i2 (G) > 21 q NG ,
where NG is defined below; the exception is G = E7 (q), in which case
i2 (G) > 14 q NG .
G
E8 (q) E7 (q) E6ǫ (q) F4 (q) G2 (q) 2F4 (q) 2G2 (q) 2B2 (q) 3D4 (q)
NG 128
70
40
28
8
14
4
3
16
Proof. An asymptotic version of this result is proved in [21, Props. 4.1,
4.3]. The lemma is easily verified by keeping track of the constants in that
proof.
Note that the exponents MG and NG in Lemmas 2.3 and 2.4 are equal,
except in cases where G possesses an involutory graph automorphism (types
Lǫn , P Ωǫ2m , E6ǫ ).
3. Proof of Theorem 1.1 for G of exceptional type
Let G be a finite simple group of exceptional Lie type over Fq (i.e. of
type E8 , E7 , E6ǫ , F4 , G2 , 2F4 , 2G2 , 2B2 or 3D4 ). Assume that q > 2 when
G is of type G2 or 2F4 , and q > 3 for type 2G2 (for the excluded groups,
G2 (2)′ ∼
= L2 (8) and these will be dealt with as classical
= U3 (3) and 2G2 (3)′ ∼
groups in the next section).
Write G = (ḠF )′ , the derived group of the fixed point group of a Frobenius
endomorphism F of the corresponding simple adjoint algebraic group Ḡ over
F̄q . Let d = |ḠF : G|, so that
(2, q − 1), if G = E7 (q),
(3, q − ǫ), if G = E6ǫ (q),
d=
1, otherwise.
By [17, Section 2], there is a cyclic maximal torus hxi of G of order as
given in Table 1. In the table, Φn (q) denotes the nth cyclotomic polynomial
evaluated at q. Morover, T = CḠF (x) is a maximal torus of ḠF of order
d|hxi|, and also CAut(G) (x) = T (see [30, 2.8(iii)]).
Let w0 be the longest element of the Weyl group W (Ḡ). Suppose that
w0 = −1 (i.e. Ḡ 6= E6 ). Then the involutions in NḠ (T ) that invert x are
conjugates of t0 , a preimage in N (T ) of w0 . Now CḠ (t0 ) has dimension
CHIRAL POLYHEDRA AND FINITE SIMPLE GROUPS
5
equal to the number of positive roots in the root system of Ḡ. Hence we see
from the lists of possibilities for involution centralizers in G, in [10, Section
4.5] for q odd, and in [1] for q even, that CG (t0 ) is as given in Table 1.
Similarly, if Ḡ = E6 , the involutions in Aut(G) inverting x are conjugate
to t0 , a preimage in NGhτ i (T ) of w0 τ = −1, where τ is an involutory graph
automorphism of G, and again CG (t0 ) is as in Table 1.
Table 1. Torus hxi and centralizer CG (t0 ), G exceptional
CG (t0 )
D8 (q), q odd
[q 84 ].B4 (q), q even
1
E7 (q)
Aǫ7 (q).2, q ≡ ǫ mod 4
d Φ18 (q)Φ2 (q), q ≥ 3
129, q = 2
[q 42 ].B3 (q), q even
1
ǫ
E6 (q)
C4 (q), q odd
d Φ9 (q), ǫ = +
1
15
[q ].C3 (q), q even
d Φ18 (q), ǫ = −
F4 (q)
Φ12 (q), q ≥ 3
(A1 (q)C3 (q)).2, q odd
17, q = 2
[q 18 ].A1 (q)2 , q even
G2 (q)
Φ6 (q), q ≥ 4
(A1 (q)A1 (q)).2, q odd
13, q = 3
[q 3 ].A1 (q), q even
3D (q)
Φ12 (q)
(A1 (q)A1 (q 3 )).2, q odd
4
[q 9 ].A1 (q), q even
p
√
2F (q) q 2 +
2q 3 √
+ q + 2q + 1
[q 9 ].A1 (q)
4
2G (q)
q + √3q + 1
A1 (q) × 2
2
2B (q)
q + 2q + 1
[q 2 ]
2
G
E8 (q)
|hxi|
Φ30 (q)
Table 2. Maximal overgroups of hxi, G exceptional
G
|M(x)|
Groups in M(x)
E8 (q)
0
−
2E (q). q+1
E7 (q), q ≥ 3
1
6
d
2A (2).2
E7 (2)
1
7
E6ǫ (q)
1
Aǫ2 (q 3 ).3
3D (q).3
F4 (q), q ≥ 3
(2, q)
4
F4 (2)
2
B4 (2)
G2 (q), q ≥ 5
1
SU3 (q).2
G2 (4)
2
SU3 (4).2, L2 (13).2
G2 (3)
3
SL3 (3).2, SL3 (3).2, L2 (13)
3D (q), 2F (q)
0
−
4
4
2G (q), 2B (q)
0
−
2
2
6
DIMITRI LEEMANS AND MARTIN W. LIEBECK
Let Inv(x) be the set of involutions in Aut(G) that invert x. By Lemma
2.1 we have Inv(x) ⊆ T t0 . Define M(x) to be the set of maximal subgroups
of G containing x such that M 6= NG (hxi). It is shown in the proof of [11,
Prop. 6.2] that the set M(x) is as given in Table 2.
Define
[
[
S=
I2 (CG (α)) ∪
I2 (M ).
α∈Inv(x)
M ∈M(x)
We claim that
(3.1)
i2 (G) > |S|.
Given (3.1), there exists an involution t ∈ G such that
(a) t lies in no maximal subgroup of G containing x, and
(b) t centralizes no involution that inverts x.
It then follows that x, t satisfy conditions (i)-(iii) of Theorem 1.1, completing
the proof of the theorem for exceptional groups.
So it remains to prove (3.1). By Lemma 2.1 we have |Inv(x)| ≤ |T |, and
also from the above discussion we know that CG (α) is conjugate to CG (t0 )
for all α ∈ Inv(x). Hence
X
(3.2)
|S| ≤ |T | i2 (CG (t0 )) +
i2 (M ).
M ∈M(x)
Using Lemma 2.3 (and also the slightly stronger bound i2P
(A1 (q)) < q 2 ), we
obtain the following upper bounds for i2 (CG (t0 )) and for M ∈M(x) i2 (M ):
P
i2 (M ) ≤
i2 (CG (t0 )) <
G
84
20
19
E8 (q)
2q (q + q ) 0
E7 (q)
2q 42 (q 12 + q 11 ) 2(q + 1)(q 42 + q 41 )
ǫ
2q 15 (q 12 + q 11 ) 2(q 15 + q 12 )
E6 (q)
F4 (q)
q 22
2(q 20 + q 19 )
4(q 5 + q 4 )
G2 (q), q ≥ 5 q 5
11
3D (q)
q
0
4
2F (q), q > 2 q 11
0
4
2G (q), q > 3 2q 2
0
2
In all these cases we check that the consequent upper bound for |S| using
(3.2) is less than the lower bound for i2 (G) given by Lemma 2.4, proving
(3.1).
This leaves the following groups to deal with: G = 2B2 (q), G2 (3), G2 (4)
and 2F4 (2)′ . In the first case, we see using [32] that i2 (CG (t0 )) = q − 1, while
i2 (G) = (q − 1)(q 2 + 1), so (3.1) holds. In the other
P cases we can use [9] to
obtain the precise values of i2 (G), i2 (CG (t0 )) and
i2 (M ), and again check
that (3.1) holds.
This completes the proof of Theorem 1.1 when G is an exceptional group
of Lie type.
CHIRAL POLYHEDRA AND FINITE SIMPLE GROUPS
7
4. Proof of Theorem 1.1 for G classical
In this section we prove Theorem 1.1 in the case where G is classical.
Let G be a finite simple classical group over Fq , and exclude L2 (q), L3 (q)
and U3 (q). So G is one of the groups
Lǫn (q) (n ≥ 4), P Sp2m (q) (m ≥ 2), P Ω2m+1 (q) (q odd), P Ωǫ2m (q) (m ≥ 4).
Let V be the natural module for G. As in the previous section, write G =
(ḠF )′ , where F is a Frobenius endomorphism of the corresponding simple
adjoint algebraic group Ḡ over F̄q . Define
(n, q − ǫ), if G = Lǫn (q)
d=
(2, q − 1), if G = P Sp2m (q), P Ω2m+1 (q) or P Ωǫ2m (q).
For convenience we handle first the following groups G:
(4.1)
Lǫ4 (2), Lǫ4 (3), Lǫ4 (4), Lǫ5 (2), Lǫ6 (2),
Sp4 (2), P Sp4 (3), Sp4 (4), Sp6 (2), P Sp6 (3), Sp8 (2),
ǫ
Ω7 (3), Ωǫ8 (2), P Ω+
8 (3), Ω10 (2).
For these groups, generators x, t as in Theorem 1.1 can be found by a search
using Magma [4]. So suppose from now on that G is not one of the groups
in (4.1).
The proof follows along the same lines as the previous section. There is an
element x ∈ G of order given in Table 3; in all cases we take x to act either
irreducibly on the natural module V , or irreducibly on both summands of
an orthogonal decomposition V = Vi +Vi⊥ , where dim Vi = i ≤ 2. Then T =
CḠF (x) is a maximal torus of order d|hxi| and also CAut(G) (T ) = T . Again,
the involutions in Aut(G) that invert x are conjugates of t0 , a preimage in
N (T ) of −1 = w0 or w0 τ (where τ is an involutory graph automorphism).
The possibilities for CG (t0 ) are also given in Table 3. (In the symplectic
and orthogonal cases, CG (t0 ) is either the given group, or the given group
quotiented by the scalars h−Ii.)
As in the previous section, define M(x) to be the set of maximal subgroups of G containing x such that M 6= NG (hxi). We claim that M(x) is
as in Table 4. When hxi is a Singer subgroup (i.e. the intersection with G of
a cyclic subgroup generated by a Singer cycle of P SL(V )), this follows from
the main theorem of [3]. This covers the cases where G = Ln (q), Un (q) with
n odd, P Sp2m (q) and P Ω−
2m (q). For the other cases we use [2, Thm. 3.1],
which classifies subgroups of classical groups of orders divisible by numbers
of the form Φd (q), where d > 12 dim V . Working through the possible subgroups, we find that the only ones containing x are those in Table 4. We
calculate the number of groups in M(x) using Lemma 2.2.
Again let Inv(x) be the set of involutions in Aut(G) that invert x, and
define
[
[
S=
I2 (CG (α)) ∪
I2 (M ).
α∈Inv(x)
M ∈M(x)
8
DIMITRI LEEMANS AND MARTIN W. LIEBECK
Table 3. Torus hxi and centralizer CG (t0 ), G classical
G
Ln (q),
n≥4
Un (q),
n≥4
|hxi|
1 q n +1
d q+1 , n odd
1 n−1
+ 1), n even
d (q
P Sp2m (q),
m≥2
1 m
d (q
+ 1)
P Ω2m+1 (q),
m ≥ 3, q odd
P Ω+
2m (q),
m≥4
1 m
d (q
+ 1)
P Ω−
2m (q),
m≥4
CG (t0 )
P SOnǫ (q), q or n odd
[q n−1 ]Spn−2 (q), q and n even
P SOn (q)
P SOnǫ (q), q odd
n−1
[q
]Spn−2 (q), q even
ǫ
GLm (q).2, q ≡ ǫ mod 4
1
(m2 +3m−2)
2
[q
].Spm−2 (q), q and m even
1
2
[q 2 (m +m) ].Spm−1 (q), q even, m odd
ǫ (q) × O ǫ′
(Om
m+1 (q)) ∩ G
1 q n −1
d q−1
1 m−1
+ 1)(q + 1), m odd
d (q
1 m−1
+ 1), m even
d (q
1 m
d (q
+ 1)
ǫ (q))2 .2 ∩ G, q odd
(Om
[q
].Spm−2 (q), q and m even
].Spm−1 (q), q even, m odd
[q
ǫ (q) × O −ǫ (q)) ∩ G, q odd
(Om
m
1
2
(m +m−2)
2
[q
].Spm−2 (q), q and m even
1
2
[q 2 (m −m) ].Spm−1 (q), q even, m odd
1
(m2 +m−2)
2
1
(m2 −m)
2
We aim to show that i2 (G) > |S|, which will complete the proof of Theorem
1.1 when G is classical.
P
As in (3.2) we have |S| ≤ |T | i2 (CG (t0 )) + M ∈M(x) i2 (M ), and so it
suffices to prove that
(4.2)
i2 (G) > |T | i2 (CG (t0 )) +
X
i2 (M ).
M ∈M(x)
P
Lemma 2.3 gives the upper bounds for i2 (CG (t0 )) and for M ∈M(x) i2 (M )
in Table 5, where d(n) denotes the number of prime divisors of n. We can use
these bounds together with Lemma 2.3 to get an upper bound for the right
hand side of (4.2); and Lemma 2.4 gives a lower bound for i2 (G). In this
way we check easily that (4.2) holds with the following possible exceptions
(recalling that we have already excluded the groups in (4.1)):
(1) G = Lǫ4 (q), P Sp4 (q) or P Ωǫ8 (q);
(2) G = Lǫ5 (3), Lǫ6 (3), Lǫ8 (2), Sp6 (4), P Sp8 (3), Sp8 (4) or Ωǫ12 (2).
For the groups under (1), we show that (4.2) still holds, by improving the
lower bound on i2 (G) from Lemma 2.4 and the upper bound on i2 (CG (t0 ))
CHIRAL POLYHEDRA AND FINITE SIMPLE GROUPS
9
Table 4. Maximal overgroups of hxi, G classical
Groups in M(x)
(GL nr (q r ).r) ∩ G,
r prime, r|n
Un (q)
(GU nr (q r ).r) ∩ G,
n odd, r prime, r|n
GUn−1 (q), n even
P Sp2m (q)
Sp 2m (q r ).r,
r
r prime, r|m
−
O2m
(q), q even
−
P Ω2m+1 (q)
O2m (q) ∩ G
−
−
P Ω+
(q)
(O
2m
2m−2 (q) × O2 (q)) ∩ G
(GUm (q).2) ∩ G, m even
Ω7 (q) (irred.), m = 4
−
r
P Ω2m (q)
(O −
2m (q ).r) ∩ G,
G
Ln (q)
Number
1 for each r
1 for each r
1
1 for each r
1
1
1
2
d
1 for each r
r
r prime, r|m
(GUm (q).2) ∩ G, m odd
Table 5. Upper bounds for i2 (CG (t0 )) and
1
P
i2 (M ), G classical
P
i2 (CG (t0 )) <
i2 (M ) ≤
1
2
n−1
q
i2 (Spn−2 (q))
2(q + 1)2 q 2 (n −n−4)
1
2
i2 (P SOn (q))
2d(n) (q 3 + 1)q 6 (n +3n)
1
2
−
q 2 (m +3m−2) i2 (Spm−2 (q)) i2 (O2m
(q) + d(m) i2 (Spm (q 2 ).2)
1
2 +m)
−
(m
q2
i2 (Spm−1 (q))
i2 (O2m (q)) + d(m) i2 (Sp2m/r (q r )),
r largest prime divisor of m
1
2
(m2 +m−2)
ǫ
2
P Ω2m (q), m ≥ 6 even
q
i2 (Spm−2 (q))
2(q + 1)2 q (m−1) −1 +
1
2
2(q + 1)2 q 2 (m −m+4)
1
2
P Ωǫ2m (q), m ≥ 5 odd
q 2 (m −m) i2 (Spm−1 (q))
as above
′
−
ǫ
ǫ
P Ω2m+1 (q), m ≥ 3, q odd
i2 (Om (q) × Om+1 (q))
i2 (O2m
(q))
G
ǫ
Ln (q), n even
Lǫn (q), n odd
P Sp2m (q), m even
P Sp2m (q), m ≥ 3 odd
in Table 5 by direct calculation in G, as follows:
G
i2 (G) >
1 8
ǫ
L4 (q)
2q
1 5
P Sp4 (q)
2 q (q − 1)
ǫ
P Ω8 (q), q even 12 q 16
P Ωǫ8 (q), q odd 18 q 16
i2 (CG (t0 )) <
2q 4
2q 3
q 11
4q 8
10
DIMITRI LEEMANS AND MARTIN W. LIEBECK
Using these improved bounds, it is straightforward to check that (4.2) again
holds.
Finally, for the groups under (2) above, we again improve the bounds on
i2 (G) and i2 (CG (t0 )) by direct calculation to show that (4.2) holds.
This completes the proof of Theorem 1.1 when G is a classical group.
5. Proof of Theorem 1.1 for G an alternating group
The existence of a pair x, t in An satisfying (i)–(iii) of Theorem 1.1 for
every n ≥ 8 can be extracted from papers of Conder [5, 6], a more recent
paper by Conder et al. [7], and some easy Magma computations. Gareth
Jones recently gave the following pairs x, t in his plenary lecture at the
conference ”Symmetries and Covers of Discrete Objects” (Queenstown, New
Zealand, February 2016).
For even n ≥ 8, take
x = (2, 3, . . . , n) and t = (1, 2)(3, 4)
And for odd n ≥ 9, take
x = (1, 2, . . . , n) and t = (1, 2)(3, 6)
Elementary arguments show that hx, ti = An , and it is an easy exercise to
show that x, t also satisfy (iii) of Theorem 1.1.
6. Proof of Theorem 1.1 for G sporadic
In this section, we show that each of the 26 sporadic simple groups has
at least one pair x, t satisfying (i)–(iii) of Theorem 1.1. We shall use the
fact that such pairs give abstract chiral polyhedra, as explained in the preamble to the theorem. When we give a pair x, y of generators of G, the
corresponding pair x, t := xy is the one satisfying conditions (i)–(iii) of
Theorem 1.1.
6.1. Mathieu groups. The groups Mi with i = 11, 12, 22 were fully investigated in [12]. They respectively have 66, 118 and 242 non-isomorphic
chiral polyhedra, hence they have that many pairs x, t satisfying (i)–(iii) of
Theorem 1.1.
The following generators of M23 give a chiral polyhedron of type {11, 15}
(the type being {p, q} where x, y have orders p, q):
x := (1, 14, 17, 21, 10, 5, 2, 16, 18, 12, 8)(3, 6, 19, 22, 15, 9, 20, 23, 4, 7, 11),
y := (1, 8, 6, 10, 21, 22, 19, 12, 11, 7, 4, 5, 3, 18, 9)(2, 23, 20, 16, 13)(14, 15, 17),
and t := xy. The pair x, t satisfies (i)–(iii) of Theorem 1.1.
The following generators of M24 give a chiral polyhedron of type {23, 15}:
x := (1, 17, 23, 21, 2, 7, 3, 15, 4, 20, 10, 6, 16, 13, 19, 22, 11, 18, 5, 14, 9, 8, 12),
y := (1, 20, 2)(3, 7, 4, 17, 21, 5, 18, 24, 11, 22, 19, 9, 14, 23, 15)(8, 13, 16, 10, 12),
and t := xy. The pair x, t satisfies (i)–(iii) of Theorem 1.1.
CHIRAL POLYHEDRA AND FINITE SIMPLE GROUPS
11
6.2. Janko groups. The groups J1 and J2 were investigated in [12]. They
have respectively 1056 and 888 non-isomorphic chiral polyhedra.
A Magma search gave a chiral polyhedron of type {19, 8} for J3 . We do
not give its generators here as these are permutations on 6516 points.
The group J4 has i2 (J4 ) = 51, 747, 149, 311. It also has a unique class
of maximal subgroups containing elements of order 29. Take σ ∈ J4 of
order 29. The normalizer NJ4 (hσi) = C29 : C28 is maximal in J4 . We have
i2 (C29 : C28 ) = 29, and all the 29 involutions are conjugate in J4 . There are
two conjugacy classes 2A and 2B of involutions in J4 . Using the character
table in Atlas [9], we compute that the structure constant for the classes
2B, 2B, 29A is 1, and hence involutions in C29 : C28 are of type 2B.
Now, the centralizer of an involution of type 2B has structure 211 : (M22 :
2). This subgroup has exactly 280831 involutions. Since 29 · 280831 <<
i2 (J4 ), there must exist at least one chiral polyhedron with automorphism
group J4 .
6.3. Conway groups. A non-exhaustive computer search with Magma
gives a chiral polyhedron of type {23, 23} for Co23 , one of type {14, 23}
for Co2 and one of type {3, 60} for Co1 .
6.4. Fischer groups. The group F i′24 has Out = 2. In F i′24 , take σ of order
29. We have NAut(F i′24 ) (hσi) = C29 : C28 . This contains 29 involutions, all in
F i′24 . These involutions belong to class 2B, and their centralizer C in F i24
satisfies i2 (C) = 5741695. Therefore i2 (C) · 29 << i2 (F i′24 ) and so F i′24 is
the automorphism group of at least one chiral polyhedron.
The group F i23 has Out = 1, and for an element σ of order 23, NF i23 (hσi) =
C23 : C11 . This latter group does not contain any involutions. Moreover,
there is obviously at least one involution that will, with σ, generate the
whole of F i23 .
Finally, a non-exhaustive computer search with Magma gives a chiral
polyhedron of type {11, 13} for F i22 .
6.5. The Monster and the Baby Monster. The Monster M has Out = 1
and a unique class of maximal subgroups of order divisible by 71, namely
subgroups L2 (71). Moreover, NL2 (71) (C71 ) = C71 : C35 , a group of odd
order. Therefore, no element of order 71 in M is conjugate to its inverse.
Take x of order 71 in M . The x is contained in a unique subgroup L2 (71)
of M . Therefore, picking t an involution of M not in the L2 (71) containing
hxi, we have ht, xi = M . The pair x, t satisfies (i)–(iii) of Theorem 1.1.
The Baby Monster BM has Out = 1 and a unique class of maximal
subgroups containing elements of order 47. Take x an element of order 47
in BM . We have NBM (hxi) = C47 : C23 . Any involution t of BM will give
hx, ti = BM . The pair x, t satisfies (i)–(iii) of Theorem 1.1.
6.6. The remaining sporadics. The Thompson group T h has Out = 1,
and for an element x of order 31, NT h (hxi) = C31 : C15 . This latter group
12
DIMITRI LEEMANS AND MARTIN W. LIEBECK
does not contain any involutions. Moreover, there is obviously at least one
involution that will, with x, generate the whole of T h.
The Lyons group Ly has Out = 1 and a unique class of maximal subgroups
of order divisible by 37, namely groups C37 : C18 . Moreover, Ly has a unique
conjugacy class of involutions, and these have centralizers 2 · A11 . We have
i2 (2 · A11 ) = 34650. Therefore i2 (2 · A11 ) · 37 << i2 (Ly) and Ly is the
automorphism group of at least one chiral polyhedron.
For the O’Nan group O ′ N , we refer to [8] where all possible types of chiral
polyhedra for O′ N have been determined.
Finally, a non-exhaustive computer search with Magma gives a chiral
polyhedron of type {19, 20} for HN , one of type {5, 7} for He, one of type
{14, 29} for Ru, one of type {13, 24} for Suz, one of type {11, 15} for M cL,
and one of type {11, 6} for HS.
This completes the proof of Theorem 1.1.
7. Concluding remarks
It would be interesting to prove similar results to Theorem 1.1 and also
that of Nuzhin described in the Introduction, for almost simple (rather than
just simple) groups. Some results in this vein are known. For instance
in [13], it is proved that every almost simple group G with socle Sz(q) is the
automorphism group of at least one abstract chiral polyhedron. And in [20],
it is shown that the only almost simple groups with socle L2 (q) that are not
automorphism groups of abstract chiral polyhedra are L2 (q), P GL(2, q), and
a group of the form L2 (9).2.
8. Acknowledgements
This research was supported by a Marsden grant (UOA1218) of the Royal
Society of New Zealand.
References
[1] M. Aschbacher and G.M. Seitz, Involutions in Chevalley groups over fields of even
order, Nagoya Math. J. 63:1–91, 1976.
[2] J. Bamberg and T. Penttila, Overgroups of cyclic Sylow subgroups of linear groups,
Comm. in Algebra 36:2503–2543, 2008.
[3] A. Bereczky, Maximal overgroups of Singer elements in classical groups, J. Algebra
234:187–206, 2000.
[4] W. Bosma, J. Cannon, and C. Playoust. The Magma Algebra System. I. The User
Language. J. Symbolic Comput., 24(3-4):235–265, 1997.
[5] M. D. E. Conder. Generators for alternating and symmetric groups. J. London Math.
Soc. (2), 22(1):75–86, 1980.
[6] M. D. E. Conder. More on generators for alternating and symmetric groups. Quart.
J. Math. Oxford Ser. (2), 32(126):137–163, 1981.
[7] M. D. E. Conder, V. Hucı́ková, R. Nedela and J. Širáň. Chiral polyhedra of given
hyperbolic type. Bull. Lond. Math. Soc., 48(1):38–52, 2016.
[8] T. Connor and D. Leemans. Algorithmic enumeration of regular maps. Ars Math.
Contemp. 10:211–222, 2016.
CHIRAL POLYHEDRA AND FINITE SIMPLE GROUPS
13
[9] J.H. Conway, R.T. Curtis, S.P. Norton, R.A. Parker, and R.A. Wilson, Atlas of Finite
Groups. Oxford U.P., 1985.
[10] D. Gorenstein, R. Lyons and R. Solomon, The classification of the finite simple groups.
Number 3. Part I. Chapter A. Almost simple K-groups. Mathematical Surveys and
Monographs, 40.3. American Mathematical Society, Providence, RI, 1998.
[11] R.M. Guralnick and W.M. Kantor, Probabilistic generation of finite simple groups,
J. Algebra 234:743–792, 2000.
[12] M.I. Hartley, I. Hubard and D. Leemans, Two atlases of abstract chiral polytopes for
small groups. Ars Math. Contemp., 5(2):371–382, 2012.
[13] I. Hubard, and D. Leemans. Chiral polytopes and Suzuki simple groups. In R. Connelly et al. (eds.), Rigidity and Symmetry, Fields Institute Communications 70:155–
175, 2014.
[14] G. A. Jones. Ree groups and Riemann surfaces. J. Algebra, 165(1):41–62, 1994.
[15] G. A. Jones. Automorphism groups of edge-transitive maps. Preprint, 2016. 62pp.
[16] G. A. Jones and S. A. Silver. Suzuki groups and surfaces. J. London Math. Soc. (2),
48(1):117–125, 1993.
[17] W.M. Kantor and Á. Seress, Prime power graphs for groups of Lie type, J. Algebra
247:370–434, 2002.
[18] R. Lawther, M.W. Liebeck and G.M. Seitz. Fixed point ratios in actions of finite
exceptional groups of Lie type. Pacific J. Math. 205:393–464, 2002.
[19] D. Leemans and M. W. Liebeck, The non-existence of chiral polyhedra with automorphism group L3 (q) or U3 (q), in preparation.
[20] D. Leemans and J. Moerenhout. Chiral polyhedra arising from almost simple groups
with socle P SL(2, q). J. Algebraic Combin., to appear.
[21] M.W. Liebeck and A. Shalev, Classical groups, probabilistic methods, and the (2, 3)generation problem, Annals of Math. 144:77–125, 1996.
[22] A.M. Macbeath, Generators of the linear fractional groups. Number Theory (Proc.
Sympos. Pure Math., Vol. XII, Houston, Tex., 1967) pp. 14–32 Amer. Math. Soc.,
Providence, R.I, 1969.
[23] V. D. Mazurov. On the generation of sporadic simple groups by three involutions,
two of which commute. Sibirsk. Mat. Zh., 44(1):193–198, 2003.
[24] P. McMullen and E. Schulte, Abstract regular polytopes, volume 92 of Encyclopedia
of Mathematics and its Applications. Cambridge University Press, Cambridge, 2002.
[25] Y.N. Nuzhin, Generating triples of involutions of Chevalley groups over a finite field
of characteristic 2, Algebra i Logika 29: 192–206, 1990.
[26] Y.N. Nuzhin, Generating triples of involutions of alternating groups, Mat. Zametki
4:91–95, 1990.
[27] Y.N. Nuzhin, Generating triples of involutions for Lie-type groups over a finite field
of odd characteristic. I, Algebra i Logika 36:77–96, 1997.
[28] Y.N. Nuzhin, Generating triples of involutions for Lie-type groups over a finite field
of odd characteristic. II, Algebra i Logika 36:422–440, 1997.
[29] E. Schulte and A.I. Weiss, Chiral polytopes. In Applied geometry and discrete mathematics, volume 4 of DIMACS Ser. Discrete Math. Theoret. Comput. Sci., pages
493–516. Amer. Math. Soc., Providence, RI, 1991.
[30] G.M. Seitz, The root subgroups for maximal tori in finite groups of Lie type, Pacific
J. Math. 106:153–244, 1983.
[31] D. Singerman, Symmetries of Riemann surfaces with large automorphism group,
Math. Ann. 210:17–32, 1974.
[32] M. Suzuki, On a class of doubly transitive groups, Annals of Math. 75:105–145, 1962.
14
DIMITRI LEEMANS AND MARTIN W. LIEBECK
Dimitri Leemans, University of Auckland, Department of Mathematics, Private Bag 92019, Auckland, New Zealand
E-mail address: [email protected]
Martin W. Liebeck, Department of Mathematics, Imperial College, London
SW7 2AZ, UK
E-mail address: [email protected]
| 4math.GR
|
ActiveMonitor: Non-blocking Monitor Executions for Increased
Parallelism
arXiv:1408.0818v1 [cs.DC] 4 Aug 2014
Wei-Lun Hung
Himanshu Chauhan∗
Vijay K. Garg†
The University of Texas at Austin
{wlhung,himanshu}@utexas.edu [email protected]
We present a set of novel ideas on design and implementation of monitor objects for multi-threaded
programs. Our approach has two main goals: (a) increase parallelism in monitor objects and thus
provide performance gains (shorter runtimes) for multi-threaded programs, and (b) introduce constructs
that allow programmers to easily write monitor-based multi-threaded programs that can achieve these
performance gains. We describe the concepts of our framework, called ActiveMonitor, and its prototype
implementation using futures[11]. We evaluate its performance in terms of runtimes of multi-threaded
programs on linked-list, bounded-buffer, and other fundamental problems implemented in Java. We
compare the runtimes of our implementation against implementations using Java’s reentrant locks [29],
recently proposed automatic signaling framework AutoSynch [23], and some other techniques from the
literature. The results of of the evaluation indicate that monitors based on our framework provide
significant gains in runtime performance in comparison to traditional monitors implemented using
Java’s reentrant locks.
1
Introduction
Monitors are the prevalent programming technique for synchronization between threads in sharedmemory parallel programs. They were designed [21] with two primary goals: (i) to ensure safety of
shared data by enforcing mutual exclusion in critical sections, and (ii) to enable conditional synchronization using the wait-notify mechanism. Monitors ensure that while a thread is executing a critical
section, other threads that cannot access the critical section wait for it to become available. Waiting
threads are notified by some thread exiting the critical section. We argue that this monitor design has
certain drawbacks:
(a) For executions on multi-core CPUs, a thread’s update to the shared data may cause a cache
invalidation for some other thread running on a different core, specially if the thread was waiting
for the critical section. Thus, the penalty incurred due to cache misses would slow down the
program on multi-core CPUs.
(b) If threads do not require the result of their updates on shared data immediately after performing
these updates, then the blocking executions of critical sections for these updates may become a
bottleneck, and could reduce the program throughput. For example, consider a thread that inserts
items in a linked-list. If this thread does not need the result of the insert operation then blocking
insert operations lead to slower runtimes.
∗
Contact Author. Address: ACES 5.432, 201 E 24 Street, Austin, Texas, USA
Supported in part by the NSF Grants CNS-1346245, CNS-1115808, and Cullen Trust for Higher Education Endowed
Professorship
†
1
In this paper, we focus on monitors methods that perform updates on shared data. Thus, the
solution of using Read/Write locks [29] is not applicable. Observe that the design of monitors to enforce
blocking executions was envisioned in 1970’s when processors were single-core, processing power was
scarce and saving processor cycles was a primary concern for programmers. Under these settings,
making the critical section executions blocking was not a performance issue. The current situation is
drastically different: not only multi-core processors are now ubiquitous, but they are also significantly
cheaper and faster. Therefore, it is important to explore alternate design and implementation of
monitors to exploit the availability of multi-core processors. Given that the blocking executions of
monitor methods seem to form a bottleneck for performance, we pose the following question: can
monitor design and implementation be altered to allow non-blocking executions and thus improve
the throughput of conditionally synchronized multi-threaded programs on multi-core machines? The
motivation here is that non-blocking executions can increase parallelism by allowing more instructions
to be executed in fixed time from the perspective of the thread(s). For this purpose, we propose two
novel techniques:
1. A monitor object is instantiated as a thread, and invocation of monitor methods by application
(worker) threads is replaced by submissions of equivalent tasks to the monitor thread. The monitor
thread assumes the responsibility of executing these tasks and returning the result (if required)
to the application threads.
2. The monitor methods, replaced by tasks in representation, can be executed in non-blocking manner. The non-blocking executions are explicitly indicated in the code.
These two changes together tackle the drawbacks discussed above. Under our proposed design,
the worker threads do not directly access the shared data, only monitor thread does, which helps in
improving the data cache locality; and non-blocking executions reduce runtime due to parallelism. This
paper presents our experience with design and implementation of these ideas. We propose a framework,
called ActiveMonitor, that explores non-blocking executions of monitor methods — possibly at the
expense of cheap and abundant processor and memory resources.
Fig. 1 compares the implementation of a bounded-buffer monitor written using: (a) Java’s Reentrant locks [29], and (b) our proposed framework and its constructs. We propose two new keywords:
activemonitor and nonblocking. The activemonitor keyword, in class declaration at line 1 of Fig. 1(b), is
used to declare that this bounded-buffer is implemented as an ‘active’ monitor object, that is: it would
exist as an independent thread. The nonblocking keyword in the method signature of put(), at line 6,
indicates that the execution of this method can be non-blocking. As our experiments show (Section 4),
this implementation provides ∼ 33% faster runtime in comparison to the conventional implementation
of Fig. 1(a). Lastly, our implementation is shorter, cleaner, and more intuitive. This is because we use
the waituntil construct (at lines 9,15 in Fig. 1(b)) that provides implicit/automatic signaling mechanism
for threads [23]. This keyword and the automatic signaling1 is not a contribution of this paper; we
merely use it in our framework’s implementation to exploit the performance benefits it provides.
An immediate criticism of our approach is the need for additional resources for its realization.
Creating a new thread for a monitor object consumes additional memory in the form of its stack
allocation, as well as increases the demand for CPU time slices. However, it is well known that the
processing power of the modern multi-core CPUs remains commonly under utilized2 . Additionally, with
the trends of ever increasing data and computation size, most users tend to favor faster runtimes even at
the cost of somewhat increased consumption of processor and memory resources. Our approach provides
much faster runtimes for many multi-threaded use cases. In addition, we claim that with careful resource
1
Appendix A briefly covers the automatic signaling mechanism
we collected resource usage data, at 10 min. intervals, of Stampede [1] supercomputer over a period of three days.
The mean CPU idle time was 86%.
2
2
(a) Conventional Java Implementation
return x;
30
}
31
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
class BoundedBuffer<V> {
V[] items;
int putPtr, takePtr, item_count, size;
Lock mutex = new ReentrantLock();
Condition notFull = mutex.newCondition();
Condition notEmpty = mutex.newCondition();
public BoundedBuffer(int size) {
this.size = size;
items = new Object[size];
}
void put(V item) {
mutex.lock();
while (item_count == size)
notFull.await();
items [putPtr++] = item;
putPtr = putPtr % size;
++item_count;
notEmpty.signal();
mutex.unlock();
}
V take() {
mutex.lock();
while (item_count == 0)
notEmpty.await();
V x = items [takePtr++];
takePtr = takePtr % size;
--item_count;
notFull.signal();
mutex.unlock();
32
}
(b) ActiveMonitor Implementation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
activemonitor class BoundedBuffer<V> {
V[] items;
int putPtr, takePtr, item_count, size;
public BoundedBuffer(int size) {
this.size = size;
items = new V[size];
}
nonblocking void put(V item) {
waituntil(item_count < size);
items [putPtr++] = item;
putPtr = putPtr % size;
++item_count;
}
V take() {
waituntil(item_count > 0);
V x = items [takePtr++];
takePtr = takePtr % size;
--item_count;
return x;
}
}
Figure 1: Comparsion of Bounded-Buffer programs written using (a) Java’s locks and (b) our approach
management by the means of a dynamic thread creation policy that considers the processor/memory
utilization, coupled with the automatic thread signaling mechanism [23], our approach’s performance
improvements come at an acceptable cost of additional CPU and memory usage. In short, the key
contributions of this paper are following:
• we propose a new monitor design, based on two new keywords: activemonitor and nonblocking, that
facilitates non-blocking executions of data updates within their critical sections. We describe the
concepts required for this design’s implementation in Java.
• we define rules to ensure correctness in presence of non-blocking operations, and show that with
our proposed design, despite additional concurrent executions, the executions are linearizable [20]
as per their non-blocking interpretation.
• we evaluate the runtimes of some fundamental multi-threaded problems using our prototype
implementation and Java’s Reentrant locks [29]. For majority of these problems, our design and
implementation provides ∼ 30 − 40% faster runtimes in comparison to reentrant locks.
The rest of this paper is organized as follows. Section 2 briefly covers the background concepts.
Section 3 discusses the key concepts of the framework and correctness of executions. Section 4 presents
experimental evaluation of our approach. We discuss some limitations of our approach, and future work
in Section 5. Section 6 presents the related work, and Section 7 offers concluding remarks.
3
2
Background
We use the concept of futures [11] to achieve non-blocking executions of critical sections.
Futures: Futures [11] were proposed as a mechanism to facilitate task submissions (to threads) in
shared memory parallel programs. In a task submission based model some threads, called executors,
provide task execution as a service. Under this model, a thread makes an asynchronous call to some
executor thread, and a future object is returned to the caller thread as a pointer to the computation
and its possible result. The executor thread performs a task for the caller thread, and updates the
returned future object with the result of the computation of the task. The caller thread can collect the
result by evaluating the future. Evaluation of a future, i.e. the collection of the result, is performed by
invoking a blocking method on the future object. Hence, if needed the thread would have to wait for
the task computation to complete. In our framework, we use futures to realize non-blocking executions
of monitor methods by treating method invocations as task submissions.
Blocking and Non-Blocking methods: By design, methods provided by monitor objects are usually
blocking (synchronous). If a thread invokes a blocking method m, then the thread can not execute its
next instruction until it has exited m; whether by normally returning from the method or exiting the
method due to an exception. Whereas, if a thread invokes a non-blocking method m′ then it would be
able to execute its next instruction immediately after, while the non-blocking method is being executed
in parallel, possibly by another thread.
3
ActiveMonitor Concepts
In this section, we discuss the key concepts of the ActiveMonitor framework. The relevant implementation details are presented in Appendix D. We first explain our approach using the bounded-buffer
monitor of Fig. 1(b). Observe that apart from using the implicit-signal construct of waituntil, the
code uses two additional keywords that we propose: activemonitor and nonblocking. By including the
activemonitor keyword in the declaration (eg. line 1), the programmer indicates that this monitor is an
‘active’ object, i.e. it should be instantiated as a thread. However, if there are many monitor objects
in a user’s program we limit the number of monitor threads — this is an implementation detail, and
discussed in Appendix D.2. In the conventional monitor design, worker threads, i.e. threads explicitly
created by the user program, execute the instructions of the critical section and update the data protected by the monitor. In contrast, as per our design, once the monitor is instantiated as a thread, it
executes its own critical sections on behalf of worker threads. For this, at compile time method calls
made to the monitor are treated as task submissions to the monitor thread. Thus, for the program
in Fig. 1b, a producer thread’s call to the put method, as well as a consumer thread’s call to the take
method, are converted to submission of equivalent tasks to the monitor thread by our framework. The
worker threads now have the option to perform blocking (synchronous) or non-blocking (asynchronous)
executions of monitor methods. Asynchronous executions are made possible by using our second keyword proposal, nonblocking, that allows for non-blocking (asynchronous) execution of instructions of the
method using this keyword. We first describe how monitor tasks are generated from monitor methods,
and then discuss their asynchronous or synchronous executions.
3.1
Monitor Tasks
In our framework, every monitor method call corresponds to an equivalent task. At compile-time,
invocations of monitor methods by worker threads are replaced by submissions of the equivalent tasks
to the monitor thread.
4
Definition 1 (Monitor Task). A monitor task t consists of a boolean predicate P and a set of statements
S. If the precondition defined by P is true then t is ‘executable’ and statements in S can be executed to
complete t. Otherwise, t is ‘unexecutable’.
Observe that the syntax of a monitor task is similar to that of a guarded command [8]. Consider
the put method (lines 8 − 13) of the bounded-buffer in Fig. 1b. For this monitor method, the equivalent
monitor task T is defined by the code of lines 9−12. Here, the precondition P is (itemCount < buffer_size);
and it checks if the buffer has any space to insert the item. If this condition is false, the waituntil construct
ensures that any thread trying to complete this task has to wait unless the buffer is not full. Lines 10
and 11 together form the set of statements S. As indicated by the nonblocking keyword in the method
signature at line 8, the generated task would be submitted for a non-blocking execution to the monitor
thread. Removal of this keyword from the method signature makes the execution blocking.
The details of compile time generation of tasks for monitor method calls are in Appendix D.1.
Appendix B discusses some special cases of monitor tasks. Our current prototype cannot convert some
particular types of monitor methods to equivalent tasks – for example, methods with waituntil not at
the beginning of the critical section, or waituntil in a conditional branch. They are also discussed in
Appendix D.1.
3.2
Monitor Tasks: Execution and Correctness Guarantees
In the ActiveMonitor framework, monitor threads are responsible for executing monitor tasks. Based
on the availability of resources, the framework tries to create one thread per monitor object. The
implementation details for limiting the monitor thread count are in Appendix D.2.
Monitor threads
execute monitor tasks while observing the following rules.
Rule 1 (Mutex Invariant). All the tasks of a single monitor object are executed by the same monitor
thread.
Thus, Rule 1 maintains the mutual exclusion of critical sections of a monitor. However, for correctness of executions, we require additional rules. Let proc(t) denote the thread that submits the task t to
the a monitor; sub(t) be the time at which the task t is submitted to the monitor, and exe(t) indicate
the time when the monitor thread starts executing t.
Rule 2. For a pair of tasks s and t submitted to a monitor M , if proc(s) = proc(t), then
sub(s) < sub(t) ⇒ exe(s) < exe(t).
Rule 3. Let n1 , m2 be two successive method invocations by a worker thread on two different monitors
M1 and M2 in the user program, and let t1 , t2 be the corresponding task submissions at runtime. Then,
at runtime t1 must be completed before t2 ’s submission.
The notions of method invocation and response used to define linearizability [20] need a different
interpretation under non-blocking executions. In short, invocation now corresponds to submission
of the equivalent task to monitor thread, and response corresponds to this tasks completion. With
this interpretation, the following result establishes correctness of executions by proving linearizable
executions.
Lemma 1. With rules 1, 2, and 3 we get an execution that is equivalent to a lock-based execution.
Proof. Sketched in Appendix C.1.
Observe that Rule 3 forces a worker thread to wait for the previous task to finish even if that task
was on a different monitor object. In many applications, a programmer may not require this constraint
5
on different monitor objects. Hence, to improve performance, we can drop Rule 3 for such applications.
We show that the execution is still linearizable with the new interpretation of invocation and response.
Note that in the standard model of concurrent history [20], the thread history is always sequential
although the object history may not be sequential. In our model, due to non-blocking executions, we
have the dual property: an object history is always sequential whereas a thread history may not be
so. We first define the thread order in presence of non-blocking operations. Let s1 , s2 , ..., sm be m
operations performed on a monitor object by a worker thread. Let si < sj denote that si was executed
before sj . In the standard model, all operations are blocking and we get that si < si+1 for all i. In
our model, we define the order as follows. If si is blocking, then si < sj for all j > i. If si and sj are
operations on the same object and i < j, then si < sj . If si is non-blocking and the result of si is
required before sk , then si < sj for all j ≥ k.
Lemma 2. With Rules 1 and 2, we get an execution that is linearizable.
Proof. Sketched in Appendix C.2.
Observe that the legal sequential history we get may not preserve the order of invocation of operations, but only the thread order.
4
Evaluation
We now present the experimental evaluation of our proposed approach. We evaluate our prototype
implementation on five multi-threaded problems. The first problem is a bounded-buffer problem with
multiple producers and consumers, the second is a sorted linked-list (of integers) on which threads
perform updates. The remaining three are conditional synchronization problems involving varying levels
of complexity in their conditional predicates. We implemented the problems using the ActiveMonitor
framework, Java’s reentrant locks, and AutoSynch [23]. Additional use-case specific implementations
(details given in problem descriptions below) are also compared for bounded-buffer, and the linked-list
problem. For all the problems, we use a benchmark in which each worker thread performs 512000
operations on shared data protected by monitors. All the experiments are conducted on a machine
with 1 socket, 2 AMD Opteron 6180 SE 12 Core, 2.5 GHz CPUs (= 24 hardware threads) and 64 GBs
memory running Linux 2.6.18. We vary the number of worker threads, and measure the time required
for all of them to complete their operations. All threads perform a fixed number of warm-up operations
on shared data before starting the time measurements. For each problem, we measure runtimes for 25
runs, and report the mean values after removing the highest and the lowest values.
For all the experiments, we restrict the ActiveMonitor framework to create only one monitor thread.
We now discuss the setup of experiments and their results in a problem specific manner. Across all
results, we denote the implementation techniques with the following notation: LK: Reentrant locks, AS:
AutoSynch[23], AM: ActiveMonitor (this paper). For each problem, all the runtimes reported in their
results exhibit negligible variance (< 5% of the mean value) across runs.
4.1
Bounded-Buffer (BB) problem [6, 7]
Setup: Items are plain Java objects. Every producer’s put invocation is non-blocking, and every consumer’s take is blocking. We also compare runtimes of Java’s ArrayBlockingQueue [29], denoted by ABQ,
based implementation. Number of producers (= number of consumers) is varied from 2 to 24. We
perform three types of experiments, and measure: (a) runtimes wieh varying number of threads for a
buffer of fixed size (=4). (b) runtimes with varying buffer-size for fixed number of producers/consumers
(=16 each). (c) runtimes with varying limit on non-blocking tasks allowed on the buffer of size 4 with
16 producers and consumers each.
6
Runtime(secs)
8
7
6
5
4
3
2
1
0
AM
ABQ
LK
AS
2
4
8
16
# of Producers (= Consumers)
(a) Buffer size = 4
18
16
14
12
10
8
6
4
2
0
LK
AS
AM
ABQ
1
2
4
8
16
32
64 128 256
8
7
6
5
4
3
2
1
0
Buffer size
(b) Varying buffer size
Queue size = x
Blocking
Unbounded Queue
5 10 15 20 25 30 35 40 45 50
Size of task queue
(c) Varying queue size for tasks
Figure 2: Runtimes for Bounded-Buffer problem
Results
Collective results are shown in Fig. 2. Fig. 2a plots the runtimes for experiment (a). Fig. 2b shows the
runtimes for experiment (b). This result highlights the benefits of non-blocking executions. Recall that
only put calls are non-blocking for AM; all take calls are blocking. For small size buffers, LK, AS, and ABQ
are much slower in comparison to AM due to their blocking insert operations. For larger buffer sizes,
LK and ABQ implementations perform slightly faster than AM. On larger buffer sizes, the frequency of
threads getting blocked out is much lower, and both LK and ABQ benefit from greedy implementation
of Java in lock acquisition; but AM ends up performing slightly more work in enforcing order on tasks
(as per Rules 1,2).
Fig. 2c shows the results for experiment (c). ‘Limit’ on the size of task queue means that whenever
the queue maintained by the monitor thread to store tasks, reaches the ‘limit’ value then even a nonblocking task submission to the monitor is force to block. Making all the invocations blocking (queue
size limit = 0) leads to an overall runtime (for completion) of ∼ 5 seconds, whereas with no limit on the
size of task queue the runtime is ∼ 3 seconds. This observations clearly shows that our non-blocking
task based approach is beneficial. However, an unbounded task queue might grow quite large. But as
shown by the plot, even for a short limit (around 20) on the size of this queue, the runtime performance
converges to optimal performance (of unbounded size). Thus, this result establishes that only a small
number of non-blocking tasks can provide significant performance benefit.
4.2
Sorted Linked-List (SLL) Problem
Setup: A linked-list of integers sorted in non-decreasing order. The number of worker threads is varied
from 2 to 24. The list is pre-populated with 5000 integers. Each worker thread inserts or removes
(with equal probability) a random integer. For a fair comparison, seeds are used for randomization so
that the threads generate the same sequence of random integers across runs as well as across different
monitor implementations. We make every thread execute some local instructions, simple mathematical
computations, outside the critical section between any successive calls on the monitor. This is done in
attempt to simulate some practical program behavior in which generation of data takes some time3 .
Both insert and remove operations are non-blocking. Additional techniques used for performance
comparison: lock-free implementation (LF) [18, 13, 32], fine-grained locking (FG) [18], and transactional
memory (TM) implementation in Java [28]. We measured runtimes for two types of experiments: (a)
runtimes with varying number of threads, when each thread performed 250 mathematical computations
3
Moreover, The results for the BB problem already showed that we perform better than conventional monitors for
saturation tests.
7
LK
AS
AM
LF
FG
TM
2
4
8
16
LK Runtime/AM Runtime
Runtime(secs)
18
16
14
12
10
8
6
4
2
0
2T
1.4
4T
8T
16T
24T
1.3
1.2
1.1
1
0.9
0
# of Threads
200
400
600
800
1000
# of Operations outside CS
(a) Runtimes for experiment (a)
(b) Runtime ratio with varying number of operations outside CS
Figure 3: Results for sorted linked-list problem
(integer additions) outside the critical section between successive calls to the monitor. (b) runtimes
with varying number of threads and varying number of operations performed outside the critical section.
Results
Fig. 3a shows the results for experiment (a). For this experiment, the ActiveMonitor implementation
(AM) clearly outperforms the reentrant lock based monitor (LK), and AutoSynch (AS). However, lock-free
(LF) and transactional memory (TM) based implementations perform better. This is because both of
them use optimized algorithm for the linked-list data structure. Whereas, we do not use any linked-list
specific optimization, and only use non-blocking executions on critical sections.
Fig. 3b plots the ratio of runtimes using LK implementation and AM implementations for experiment
(b). The ratio with no operations outside the critical section is close to 1, i.e. our approach provides
no performance benefit. This is expected because no operations can be performed concurrently with
the non-blocking updates to the list, and the opportunity of parallelism is effectively absent. As the
number of operations outside the critical section is increased, we see significant performance gains with
our approach, upto ∼ 30%, due to increased parallelism introduced by non-blocking executions.
4.3
Conditional Synchronization Problems
We used three thread synchronization problems that involve different levels of complexity in their conditional predicates. For all three, we measured the time taken for all the worker threads to complete
the designated number (512000) of operations on the monitor.
Round-Robin (RR) Access: Every worker thread accesses the monitor in round-robin order. All
operations on the monitor are non-blocking. The number of worker threads is varied from 2 to 24.
Parametrized Bounded-Buffer (PBB) [6, 7]: Producers put a collection of items into a shared
buffer, while consumers remove a number of items from the buffer. For producers, the number of items
to be inserted, and for consumers, the count of items to be removed, both are randomly decided at
runtime. Similar to the SLL problem, seeds are used for randomization for fair evaluation across implementations. The number of producers (= number of consumers) is varied from 2 to 24.
Ticketed-Readers/Writers (TRR) [4, 3, 12]: A ticket is used to maintain the access order for
readers and writers [3]. Every reader/ writer gets a ticket number indicating its arrival order. Readers/writers wait on the monitor for their turn. On their turn, they enter the monitor but do not
8
7
LK
AS
AM
AS
LK
AM
LK
AS
AM
Runtime(secs)
6
5
4
3
2
1
0
2
4
8
16
2
# of Threads
(a) Round-robin problem
4
8
16
# of Consumers
(b) Parametrized bounded-buffer
2/10
4/20
8/40
16/80
24/120
# of Writers/Readers
(c) Ticketed r/w problem
Figure 4: Runtimes for RR, PBB, and TR/W problems: all plots have the same range on y-axis
perform any computation inside the monitor, and immediately exit. Operations for arrival on monitor
and waiting for turn to access the monitor are blocking, whereas operations after getting access to
critical sections are non-blocking. The number of writer threads is varied from 2 to 24. The number of
readers is kept 5 x the number of writers.
Results
Fig. 4 presents the runtimes of LK, AS, and AM based implementations for the three problems. For the
RR problem, the performance of AM is 6 and 5 times faster than that of AS, and LK, respectively.
This is due to better cache locality and fewer context switches. For this problem, after accessing
the critical section, every thread is forced to wait unless all other threads have accessed the critical
section. This constraint leads to poor cache locality and more context switches for conventional monitor
design. However, in our approach, the monitor thread continuously executes all monitor accesses for
each thread; therefore, there is minimal number of context switches between each monitor method
invocation and that results in excellent cache locality.
For the TR/W problems, AM performs poorly in comparison to LK. This is primarily due to the nature of the problem and its underlying implementation requirements. The problem requires conditional
synchronization in the middle of the method calls, and thus our task based approach has to create two
tasks per method call. In fact, any family of problems with such characteristics would lead to much
higher number of tasks than the actual method calls.
4.4
CPU and Memory Consumption
Table 1 presents the results on the CPU and memory usage by LK, AS, and AM for all the evaluated
problems. For each problem, the CPU ratio stands for the ratio of an implementation’s mean (across
runs) CPU consumption and mean (across runs) CPU consumption of LK implementation. As per
memory usage, we report the maximum memory (in MBs) consumed across any of the runs performed
by the implementations for each problem. As expected, our approach consumes more CPU resources.
However, the maximum memory consumed by our approach is only incrementally larger than that
of Java’s reentrant lock based approach. Thus, we have backed our claim that with careful resource
management the cost for the performance benefit is relatively small.
5
Discussion & Future Work
The benefits of our approach are manifold: improved runtime performance, better control over desired
program behavior, and increased opportunities for applying optimizations such as execution re-ordering,
9
Problem
Approach
CPU Ratio
Max Mem.
BB
SLL
RR
PBB
TR/W
LK
AS
AM
LK
AS
AM
LK
AS
AM
LK
AS
AM
LK
AS
AM
1.00
196
1.11
236
1.95
197
1.00
135
0.86
135
3.49
145
1.00
236
1.10
224
2.97
250
1.00
196
2.39
227
2.47
251
1.00
211
1.23
232
1.24
244
Table 1: CPU and memory consumption with Reentrant locks (LK), AutoSynch (AS), and ActiveMonitor
(AM).
operation combining, etc. However, some aspects of our framework are open to criticism. First and
foremost, based on the design assumptions as well as the experimental results, it is evident that at
present, our approach does not suit systems on which processor and memory resources are scarce. Even
though our evaluation results indicate slightly increased memory consumption, the increase in CPU
consumption is more drastic. Hence, an important future work is an improved implementation of our
framework that lowers the CPU usage. Another future work is to address the issue of task generation for
problems that require conditional synchronization in the middle of the method call, eg. Ticket-R/W,
and causes large number of tasks to be created. It is important to note that our approach extends
conventional multi-threaded programs and their constructs, and thus provides a programmer the choice
of an implementation she thinks is most suited for her needs.
We showed that we provide linearlizability of executions — interpreted under the non-blocking
execution model. Although it is desirable in most of the cases, there are situations when this requirements can be relaxed to certain extents for even faster performance. Such relaxations are very much
dependent on the use-cases and the expectations of the programmer. As shown by Kogan et al. in
[25], having ‘weak’ consistency requirements could lead to significant performance gains. In fact, for
specific use-cases programmers may only need eventual consistency [37], which requires that any update to shared data is eventually committed, in their programs. However, allowing weak or eventual
consistency is not possible for all use-cases. For example, consider multi-threaded implementations (in
shared memory model) of Dijkstra’s single-source-shortest-path algorithm [5, 31]. Weak or eventual
consistency based implementations for this algorithm cannot guarantee correct results. However, for a
large number of problems, approximate solutions are good enough, and their parallel implementations
exhibit significant gains in performance. Thus, this opens up two interesting future problems for our
work. First of them is evaluating performance of our non-blocking execution approach for problems
that admit approximate solutions, and the second is to extend our evaluation to the weak consistency
model of [25].
When it comes to evaluating the performance of our proposed design and its implementation, the
exploration space is large. Not only there is a large number of problems related to multi-threaded
updates on data structures, but the set of parameters that affect the runtime for these problems is
also large. We also plan to perform experiments by varying more parameters such as load distributions
of methods, size of shared data, etc. Another key experiment is to analyze the cache locality of our
approach — in terms of cache miss rates per operation and per instruction — for different experiments,
and comparing it to cache locality achieved by other techniques.
Lastly, note that our current implementation does not use ideas such as operation combining [15],
execution re-ordering, and batch updates [25]. Intuitively, incorporating these strategies in our implementation could lead to even better results than those observed in this paper.
6
Related Work
The idea of having monitor objects execute as independent threads is not completely novel. Hoare
proposed a similar mechanism, in which all objects are active, of communicating sequential processes
(CSP) in [22] long ago. However, the proposal was a formal language for interactions between processes.
10
Its constructs are still mostly used for formal analysis. Our work focuses on making parallel programs
faster by using futures to provide selective non-blocking executions on monitor methods.
Research efforts to increase parallelism in programs have been continuous almost since the advent
of computing. With the arrival of multi-core processors, improving multi-threaded programming with
syntactic changes, as well as with novel designs of data structures and algorithms, has been the focus of
many research efforts in the last decade. Transactional memory [17, 34] is a well-known research effort
that proposes modified syntax for ease of writing multi-threaded programs. The prominent contributions in algorithmic direction are: lock-free data structures [13, 32, 16], and wait-free data structures
[26, 27, 36]. Some other research efforts have explored increased parallelism for data structures by
operation re-ordering and combining complementary operations (insert and remove) [14].
Transactional memory [17, 34] introduced an elegant and much needed design for writing multithreaded programs on shared memory. The promising feature in transactional memory proposal was to
provide simplified syntax that guarantees correctness, and delegates the handling of mutual exclusion
to the library providing the implementation. Our approach and ideology is similar on these goals.
However, constructs for conditional waiting under transactional memory are limited [35, 30, 9]. Hence,
writing many conditional synchronization based multi-threaded programs is rather difficult. Also,
unlike transactional memory, our approach merely transfers the responsibility of data manipulation
to a single monitor thread and does not require any complicated rollback mechanism for resolving
conflicting updates on the shared data.
Designing lock-free and wait-free algorithms is a challenging problem for any use-case. In addition, lock-free and wait-free implementations are currently available for a small set of data structures
[10, 19]. Even though there is a lot of ongoing research to grow this set, the growth is slow. By using
our approach, the end user benefits by an additional level of abstraction and does not have to deal
with the complications involved in understanding and correctly implementing wait-free and lock-free
algorithms.
We use the concept of futures [11, 29] to realize the idea of non-blocking/asynchronous executions.
The work by Kogan et al. [25] is similar to our approach, and makes use of futures for non-blocking
executions. However, there are few key differences between our work and [25]. We explore changes
to the paradigm of monitors, especially for conditional synchronization problems, whereas [25] focuses
on throughputs, under saturation scenarios, of data structures such as stacks, queues, and linked-lists.
Hence, our approach provides a more generic tool for performing safe updates and synchronization
between threads.
Frameworks such as Legion [2] and Galois [33] focus on improving runtime performance of parallel
programs on scientific simulations and large graphs, respectively. Thus, they do not directly relate
to thread synchronization and notification based use-cases such as bounded-buffer. Secondly, both of
these libraries require advanced skills and familiarity with the features and syntactic details exposed
by the libraries. Hence, the performance gains achieved by using them are highly correlated with a
programmer’s knowledge of the involved optimization parameters needed to tweak the executions for
optimal performance.
7
Conclusion
Proposing changes to a well established programming paradigm is an ambitious goal. Moreover, convincing programmers to adopt a new programming framework in favor of the conventional one, usually
requires the proposed approach to offer gains in terms of runtime performance as well as programming ease/productivity. The ActiveMonitor framework is a step in this direction. In almost all of the
problems used in our evaluation, monitors with non-blocking executions outperform Java’s reentrant
lock based monitors. Additionally, this gain is achieved by a simple mechanism of using only two new
11
keywords in existing Java programs. At present, our prototype implementation is costly in terms of
processor usage. However, with the ubiquity of multi-core processors, optimized implementations of
our proposed technique could lead to significant advantages at reduced costs. In this paper, our initial
exploration in a new direction, that is: non-blocking executions on monitors, has showed promising
results. We hope that with further research in this direction, our approach would lead to even stronger
results.
References
[1] Texas advanced computing center stampede. https://www.tacc.utexas.edu/stampede/.
[2] M. Bauer, S. Treichler, E. Slaughter, and A. Aiken. Legion: Expressing locality and independence
with logical regions. In Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis, SC ’12, pages 66:1–66:11, Los Alamitos, CA, USA, 2012.
IEEE Computer Society Press.
[3] P. A. Buhr and A. S. Harji. Implicit-Signal Monitors. ACM Transactions on Programming Languages and Systems, 27(6):1270–1343, Nov. 2005.
[4] P. J. Courtois, F. Heymans, and D. L. Parnas. Concurrent control with “readers” and “writers”.
Communications of the ACM, 14(10):667–668, 1971.
[5] E. W. Dijkstra. A note on two problems in connexion with graphs. Numerische Mathematik,
1(1):269–271, 1959.
[6] E. W. Dijkstra. Cooperating Sequential Processes, Technical Report EWD-123. Technical report,
1965.
[7] E. W. Dijkstra. Hierarchical ordering of sequential processes. Acta Informatica, 1(2):115–138,
1971.
[8] E. W. Dijkstra. Guarded commands, nondeterminacy and formal derivation of programs. Communications of the ACM, 18(8):453–457, Aug. 1975.
[9] P. Dudnik and M. Swift. Condition variables and transactional memory: Problem or opportunity?
In The 4th ACM SIGPLAN Workshop on Transactional Computing, 2009.
[10] F. Fich, D. Hendler, and N. Shavit. On the inherent weakness of conditional synchronization primitives. In Proceedings of the Twenty-third Annual ACM Symposium on Principles of Distributed
Computing, PODC ’04, pages 80–87, New York, NY, USA, 2004. ACM.
[11] R. H. Halstead. Multilisp: A language for concurrent symbolic computation. ACM Trans. Program.
Lang. Syst., 7(4):501–538, Oct. 1985.
[12] P. B. Hansen. Structured Multiprogramming. Communications of the ACM, 15(7):574–578, 1972.
[13] T. L. Harris. A Pragmatic Implementation of Non-blocking Linked-Lists. In Proceedings of the
15th International Conference on Distributed Computing, pages 300–314, London, UK, UK, 2001.
Springer-Verlag.
[14] S. Heller, M. Herlihy, V. Luchangco, M. Moir, W. I. Scherer, and N. Shavit. A Lazy Concurrent
List-Based Set Algorithm. In J. Anderson, G. Prencipe, and R. Wattenhofer, editors, Principles
of Distributed Systems, pages 3–16. Springer Berlin Heidelberg, 2006.
12
[15] D. Hendler, I. Incze, N. Shavit, and M. Tzafrir. Flat Combining and the Synchronizationparallelism Tradeoff. In Proceedings of the Twenty-second Annual ACM Symposium on Parallelism
in Algorithms and Architectures, pages 355–364, New York, NY, USA, 2010. ACM.
[16] D. Hendler, N. Shavit, and L. Yerushalmi. A scalable lock-free stack algorithm. In Proceedings of
the Sixteenth Annual ACM Symposium on Parallelism in Algorithms and Architectures, SPAA ’04,
pages 206–215, New York, NY, USA, 2004. ACM.
[17] M. Herlihy and J. E. B. Moss. Transactional memory: Architectural support for lock-free data
structures. In Proceedings of the 20th Annual International Symposium on Computer Architecture,
ISCA ’93, pages 289–300, New York, NY, USA, 1993. ACM.
[18] M. Herlihy and N. Shavit. The Art of Multiprocessor Programming. Morgan Kaufmann Publishers
Inc., San Francisco, CA, USA, 2008.
[19] M. P. Herlihy. Impossibility and universality results for wait-free synchronization. In Proceedings
of the Seventh Annual ACM Symposium on Principles of Distributed Computing, PODC ’88, pages
276–290, New York, NY, USA, 1988. ACM.
[20] M. P. Herlihy and J. M. Wing. Linearizability: A Correctness Condition for Concurrent Objects.
ACM Transactions on Programming Languages and Systems, 12(3):463–492, 1990.
[21] C. A. R. Hoare. Monitors: An Operating System Structuring Concept. Communications of the
ACM, 17(10):549–557, 1974.
[22] C. A. R. Hoare. Communicating sequential processes, volume 178. Prentice-hall Englewood Cliffs,
1985.
[23] W.-L. Hung and V. K. Garg. AutoSynch: An Automatic-signal Monitor Based on Predicate
Tagging. In PLDI, pages 253–262, 2013.
[24] V. Kodaganallur. Incorporating Language Processing into Java Applications: A JavaCC Tutorial.
IEEE Softw., 21(4):70–77, 2004.
[25] A. Kogan and M. Herlihy. The future(s) of shared data structures. In PODC, 2014.
[26] A. Kogan and E. Petrank. Wait-free queues with multiple enqueuers and dequeuers. In Proceedings
of the 16th ACM Symposium on Principles and Practice of Parallel Programming, PPoPP ’11,
pages 223–234, New York, NY, USA, 2011. ACM.
[27] A. Kogan and E. Petrank. A methodology for creating fast wait-free data structures. In Proceedings of the 17th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming,
PPoPP ’12, pages 141–150, New York, NY, USA, 2012. ACM.
[28] G. Korland, N. Shavit, and P. Felber. Noninvasive concurrency with Java STM. In MultiProg
2010: Third Workshop on Programmability Issues for Multi-Core Computers, 2010.
[29] D. Lea. The Java.Util.Concurrent Synchronizer Framework. Sci. Comput. Program., 58(3):293–
309, 2005.
[30] V. Luchangco and V. J. Marathe. Revisiting condition variables and transactions. In The 6th
ACM SIGPLAN Workshop on Transactional Computing, 2011.
[31] U. Meyer and P. Sanders. δ-stepping: A parallelizable shortest path algorithm. J. Algorithms,
49(1):114–152, Oct. 2003.
13
[32] M. M. Michael. High Performance Dynamic Lock-free Hash Tables and List-based Sets. In Proceedings of the Fourteenth Annual ACM Symposium on Parallel Algorithms and Architectures, pages
73–82, New York, NY, USA, 2002. ACM.
[33] K. Pingali, D. Nguyen, M. Kulkarni, M. Burtscher, M. A. Hassaan, R. Kaleem, T.-H. Lee,
A. Lenharth, R. Manevich, M. Méndez-Lojo, D. Prountzos, and X. Sui. The Tao of Parallelism in
Algorithms. In Proceedings of the 32Nd ACM SIGPLAN Conference on Programming Language
Design and Implementation, pages 12–25, New York, NY, USA, 2011. ACM.
[34] N. Shavit and D. Touitou. Software transactional memory. In Proceedings of the Fourteenth Annual
ACM Symposium on Principles of Distributed Computing, PODC ’95, pages 204–213, New York,
NY, USA, 1995. ACM.
[35] A. Skyrme and N. Rodriguez. From locks to transactional memory: Lessons learned from porting a
real-world application. In The 8th ACM SIGPLAN Workshop on Transactional Computing, 2013.
[36] S. Timnat, A. Braginsky, A. Kogan, and E. Petrank. Wait-free linked-lists. In Proceedings of the
17th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPoPP ’12,
pages 309–310, New York, NY, USA, 2012. ACM.
[37] W. Vogels. Eventually Consistent. Communications of the ACM, 52(1):40–44, 2009.
Appendix
A
Implicit/Automatic signal monitors:
Monitors can be divided into two categories according to the different implementations of conditional
synchronization: explicit-signal monitors and implicit-signal monitors [3]. In explicit-signal monitors,
condition variables are used along with wait, and notfiy/notifyAll statements for conditional synchronization between threads. Programmers need to associate assertions with condition variables manually.
When a thread wants to execute a set of instructions under a critical section, it first checks some
condition variable(s), and waits if the predicate is not true. Once a thread enters the waiting state,
another thread on detecting that the conditional predicate has become true, explicitly notifies the
waiting thread. The bounded-buffer implementation in Fig. 1(a) that uses Java’s reentrant locks is an
explicit-signal monitor. In fact, almost all of the prevalent programming languages, including Java, use
the explicit-signal design for monitors.
Implicit-signal monitors, also called automatic-signal monitors, require the underlying system to
handle thread synchronization and wait/notification. Recently, Hung et al. [23] showed that implicitsignal monitors can be beneficial when implemented on modern multi-core processors. In an implicitsignal design, programmers need to use waituntil statements (line 9, 15 in automatic-signal program in
Fig. 1(b)) instead of condition variables for synchronization. In this monitor design, a thread will wait
as long as the condition of a waituntil statement is false, and execute the remaining instructions only
after the condition becomes true. The responsibility of identifying a waiting thread to signal, and then
signaling it, is that of the underlying system/framework rather than of the program logic explicitly put
in the thread by a programmer. The details of design and implementations of such monitors is beyond
the scope of this paper, and we refer the interested reader to [21, 23].
14
B
Monitor Task Examples
Here, we highlight some key aspects of our framework’s handling of monitor tasks. First, note that
the set of statements, S, of a task may be empty. Such a task serves as a barrier; and is legal in our
framework. Second, the precondition can either be absent or appear later, and not as the first statement,
in the monitor method. When a monitor method has no precondition, our framework creates a task
with the precondition as tautology, indicating that the task can be executed at any time. If a monitor
method does not start with a waituntil statement but has some such statement in between, then the
precondition of the first derived task is a tautology. In Fig. 5, the method bar() has two corresponding
tasks T1 and T2 , where T 1 has the precondition as tautology and the set of statements S1 ; T2 has the
precondition P and the set of statements S2 .
Furthermore, monitor tasks are compositional in nature. Consider the method foo() in Fig. 5. If
a method declares n sets of preconditions and statements, then the framework would generate n tasks
such that each task Ti , 1 ≤ i ≤ n, has a precondition Pi and a corresponding set of statements Si .
Appendix B presents some detailed examples of these cases.
void foo() {
waituntil(P1 );
S1 ;
waituntil(P2 );
S2 ;
...
waituntil(Pn );
Sn ;
}
void bar() {
S1 ;
waituntil(P );
S2 ;
}
Figure 5: Examples of methods leading to compositional tasks and tautology as precondition
C
C.1
Lemma Proofs
Proof Sketch for Lemma 1
We show that for any execution in our model there exists an equivalent lock-based execution. Since all
tasks of any monitor object are executed by a single thread due to Rule 1, mutual exclusion is preserved
just as in any lock-based execution. We only need to show that the order of execution of the tasks
corresponds to a schedule in which worker threads execute the tasks.
It is sufficient to show that all tasks submitted by a single worker thread execute in the order of
submissions. Let s and t be two consecutive tasks submitted by a worker thread. If they are submitted
for the same monitor, then the Rule 2 preserves the order. If s is a blocking task, then by definition of
blocking task, t cannot be submitted before s is completed. Hence, execution of s precedes execution
of t. If s is a non-blocking task and is on a different monitor object from t, then due to Rule 3 we wait
for s to finish before submission of t.
C.2
Proof Sketch for Lemma 2
It is sufficient to show that for every thread history there exists an equivalent sequential thread history
that is consistent with the execution by the monitor thread. We get this sequence by considering as
linearization point for an operation the instant at which the monitor thread finishes executing the
corresponding task. We show that this order is consistent with the thread order.
15
Let s < t denote that operation(i.e. its corresponding task) s was executed before t. First consider
the case when s < t because s is a blocking operation. In this case, task t cannot be submitted before
task s is executed. Hence, the order of execution of monitor tasks preserves the thread order. Second,
consider the case when s < t because they are operations on the same monitor object and s is invoked
before t. In this case Rule 2 guarantees that the order of execution is s followed by t. Finally, consider
the case when s and t are tasks submitted by the same thread, s is non-blocking but its results are used
by t. In this case, the execution would have to block for collecting the results of s before executing any
further. Hence t would be submitted later than the execution of s.
D
Implementation Details
The translation of the code written with our two proposed keywords to an equivalent Java code is
performed by using a pre-processor. For our implementation, we use JavaCC [24] pre-processor. We
convert all the monitor methods to equivalent tasks, and use a threadpool based executor service
framework for monitor threads. Here, we first discuss the pre-processing steps.
D.1
Pre-processing
We briefly summarize the concepts applied for compiling the code, using the ActiveMonitor framework,
in which the two proposed keywords are used.
First, our pre-processor identifies the waituntil statements that capture the preconditions of monitor
tasks. Recall that the waituntil keyword is used for automatic signaling between threads, and the
predicate provided as its argument forms the precondition for execution of monitor tasks. In the
corresponding generated code (in Java) every predicate is created as an inner class with a method
called isT rue(), which returns a boolean value as the result of the predicate evaluation. At runtime,
our system invokes the isT rue() method to evaluate the predicate when deciding which task should be
executed.
Next, the pre-processor creates a Callable [29] or Runnable [29] object that contains the set of statements for each task inside its call() or run() method. If the return value of the original monitor
method is void , then we create a Runnable object; otherwise, we create a Callable object. Note that if a
non-blocking monitor method has been divided into multiple tasks, the resulting tasks may have some
shared variables. To handle these variables, our system creates a data object to store them and all the
tasks can access them through this data object.
Finally, we replace the invocations of monitor methods by submitting monitor tasks to threadpool
executor (Appendix D.2). The executor returns f, an instance of Java’s Future object [29], to the
invoker thread when a task is submitted. If the task is blocking, then the thread needs to wait for
the computation in f to finish; this is done by evaluating f within the task and returning the result.
Otherwise, f is registered with the threadpool executor for exception checking and handling.
In Section 3, we mentioned that our current implementation does not admit a particular family
of methods for task based executions. Specifically, it does not admit recursive blocking calls, monitor
methods in which conditional synchronization guards using automatic thread signaling (waituntil) are
not initial statements of the program, or are present in conditional branches. This does not mean
that programs requiring such method implementations are not allowed in our framework; just that at
present we do not support task generation of such methods. A programmer can implement them using
conventional Java syntax, and still use our framework for other methods.
16
D.1.1
Discussion: Thread Dependent Variables and Functions
In our current implementation, thread dependent variables and functions within a monitor method
cannot be used directly in the Runnable or Callable object that is used in task generation by our approach.
This is because the tasks are executed by the monitor thread and not by the worker thread. For example,
suppose there is a monitor method that invokes Thread.currentThread(), if we directly add this statement
to the generated Runnable object (in the task), then this method’s invocation at runtime will return the
reference to the monitor thread when it is executed. However, it is obvious that the intent of this call
inside the monitor method was to refer to the worker thread. To handle this situation, currently, we
require the programmer to perform reference copy and storage and storage in thread-local variables.
For read operations of thread dependent variables and functions, the worker thread would need to
evaluate them outside the monitor, and store the result with final variables. These final variables can
be accessed by the runnable and callable objects. An additional constraint/limitation applies for the
case of write operation on thread dependent variables. For write operations, if the monitor method is
non-blocking then the results can be stored as intermediate data. The worker thread then writes these
results back to its local variable after the task is executed.
D.1.2
Discussion: Blocking recursive method
Our current pre-processing implementation does not support a blocking recursive method on monitors.
This is because the number of the method invocations to be made at the runtime is non-deterministic.
Thus, we cannot know how many tasks we need to create at pre-processing time. In addition, since the
method is blocking, the monitor thread will get blocked when it recurs.
D.2
Limiting Monitor Threads using ThreadPools
A threadpool is a collection of threads to execute tasks that are units of computation. These threads are
usually created together at the start-up , and remain in the pool to provide executor service. For task
execution based programs, the use of a threadpool can significantly improve the runtime performance
by having an already existing thread ready to execute the tasks as they arrive. Generally, tasks are
stored in a collection, and free threads are responsible for finishing the unexecuted tasks. If there is
no task that is eligible for execution, the threads in the pool wait for one to arrive. This approach is
especially beneficial when the number of tasks is greater than the number of threads in a pool. Thus,
the size of the threadpool is a crucial factor as the timing of creating or destroying a thread may have a
significant impact on performance. Given that we spawn a new thread for a monitor object, programs
with a large number of monitors could be adversely affected by having too many additional threads in
the ActiveMonitor framework.
To keep the overhead of monitor threads relatively low, we control the number of monitor threads
created by querying the operating system for available resources and managing the monitor threads
dynamically. A component, called Monitor ThreadPool Executor, manages this process. This component instantiates monitor threads as threads of a threadpool at the start of the execution. The
pre-processing phase collects the information about number of monitor objects in the program, and the
threadpool executor uses this information for creating a threadpool balancing the resource consumption
and performance benefit.
D.2.1
Exception Handling
For a non-blocking method invocation, after submitting its corresponding task to the executor, the
invoker does not need to wait for the completion of the task. The task is executed in parallel by
a thread of the monitor threadpool. Thus, if an exception occurs during its execution, the thread
17
that submitted it must be notified of this exception. The ActiveMonitor framework has an exception
handler that keeps a log of every exception and provides different mechanisms for programmers to
handle exceptions in the non-blocking method. The users may choose to ignore the exceptions or they
can specify a maximum number of times a task may be considered for automatic re-tries. Furthermore,
our system also provides a hook so that the programmer can write their custom exception handler.
18
| 6cs.PL
|
PORTRAIT GROWTH IN CONTRACTING, REGULAR BRANCH
GROUPS
arXiv:1710.02902v1 [math.GR] 9 Oct 2017
ZORAN ŠUNIĆ AND JONE URIA-ALBIZURI
Abstract. We address a question of Grigorchuk by providing both a system
of recursive formulas and an asymptotic result for the portrait growth of the
first Grigorchuk group. The results are obtained through analysis of some features of the branching subgroup structure of the group. More generally, we
provide recursive formulas for the portrait growth of any finitely generated,
contracting, regular branch group, based on the coset decomposition of the
groups that are higher in the branching subgroup structure in terms of the
lower subgroups. Using the same general approach we fully describe the portrait growth for all non-symmetric GGS-groups and for the Apollonian group.
1. Introduction
Regular branch groups acting on rooted trees have been extensively studied since
the 1980s. The interest in these groups is due to their remarkable properties. For
instance, some of the groups in this class provide counterexamples to the General
Burnside Problem, many of them are amenable but not elementary amenable, and
the first example of a group of intermediate word growth was of this type. The initial
examples, constructed in the early 1980s, were the Grigorchuk 2-groups [Gri80,
Gri84] and the Gupta-Sidki p-groups [GS83]. Many other examples and different
generalizations have been introduced since then.
The notion of word growth for a finitely generated group was introduced by
A.S. Švarc [Šva55] and later, independently, by J. Milnor [Mil68a, Mil68b]. Given
a finitely generated group, one can define the word metric with respect to a finite
generating set. It is natural to ask what is the word growth function of a particular
group, that is, what is the number of elements that can be written as words up
to a given length (equivalently, what is the volume of the ball of a given radius
in the word metric). A lot of work has been done in this direction. For instance,
it is known that groups of polynomial growth are exactly the virtually nilpotent
groups [Gro81]. Since the free group of rank at least 2 has exponential growth,
the word growth of every finitely generated group is at most exponential. In 1968,
Milnor [Mil68c] asked if every finitely generated group has either polynomial or
exponential growth, and Grigorchuk [Gri83, Gri84] showed in the early 1980s that
groups of intermediate growth exist. In particular, the growth of the first Grigorchuk group [Gri80], introduced in 1980 as an example of a Burnside 2-group, is
superpolynomial, but subexponential.
The first Grigorchuk group is a self-similar, contracting, regular branch group,
and so are the Gupta-Sidki p-groups. The study of such groups often relies on
the representation of their elements through portraits. For instance, the known
estimate [Bar98] of word growth for the first Grigorchuk group is based on an
estimate of the number of portraits of elements of given length. Thus, in the
2010 Mathematics Subject Classification. Primary 20E08.
J. Uria-Albizuri acknowledge financial support from the Spanish Government, grants
MTM2011-28229-C02 and MTM2014-53810-C2-2-P, and from the Basque Government, grants
IT753-13, IT974-16 and the predoctoral grant PRE-2014-1-347.
1
2
ZORAN ŠUNIĆ AND JONE URIA-ALBIZURI
context of self-similar, contracting groups acting on regular rooted trees, it is of
interest to study the portrait growth of a group with respect to a finite generating
set. The necessary definitions are given in Section 2, but let us provide here a
rough description. Given a finitely generated group acting on a rooted tree, one
can consider the action of the group on the subtrees below each vertex of the tree.
If the action on such subtrees is always given by elements of the group itself, the
group is called self-similar. Moreover, if the elements describing the action at the
subtrees on each level are getting shorter (with respect to the word metric) with the
level, with finitely many exceptions, the group is said to be contracting. Thus, in a
finitely generated, contracting group G, there is a finite set M such that, for every
element g in G, there exists a level such that all elements describing the action of g
on the subtrees below that level come from the finite set M. The first level where
this happens is called the depth of the element g (with respect to M), a concept
introduced by Sidki [Sid87]. In this setting, the portrait growth function (growth
sequence) counts the number of elements up to a given depth and one may ask
what is the growth rate of such a function. This question was specifically raised by
Grigorchuk [Gri05] for the first Grigorchuk group.
We provide a system of recursive formulas for the portrait growth of the first
Grigorchuk group (see Theorem 4.2) and the following asymptotic result.
Theorem 1.1. There exists a positive constant γ such that the portrait growth
sequence {an }∞
n=0 of the first Grigorchuk group G satisfies the inequalities
n
1 γ2n
e
≤ an ≤ 4eγ2 ,
4
for all n ≥ 0. Moreover, γ ≈ 0.71.
We note here that the recursive formulas from Theorem 4.2, together with the
results of Lemma 3.2, may be used to estimate γ with any desired degree of accuracy.
More generally, we provide recursive formulas (see (2)) for the portrait growth
sequence of any finitely generated, contracting, regular branch group. The formulas
are directly based on the branching subgroup structure of the group.
As a further application, we also study the portrait growth for all non-symmetric
GGS-groups and for the Appollonian group and, in each case, we resolve the obtained recursion and provide a straightforward description of the portrait growth.
Theorem 1.2. Let G be a GGS-group defined by a non-symmetric vector e ∈ Fp−1
.
p
The portrait growth sequence of G {an }∞
is
given
by
n=0
a0 = 1 + 2(p − 1)
n−1
an = p(x1 + (p − 1)y1 )p
,
where x1 and y1 are the number of solutions in Fp of
((n0 , . . . , np−1 )C(0, e)) ⊙ (n0 , . . . , np−1 ) = (0, . . . , 0),
with n0 + · · · + np−1 = 0 and n0 + · · · + np−1 = 1, respectively.
For instance, for the so called Gupta-Sidki 3-group, which corresponds to the
GGS-group defined by e = (1, −1) with p = 3, the portrait growth sequence is
n−1
given by a0 = 5 and an = 3 · 93 , for n ≥ 1.
Theorem 1.3. The portrait growth sequence {an }∞
n=0 of the Apollonian group is
given by
n
n
3n −1
1 √ 3
.
an = 3 2 · 7 3 = √ 7 3
3
PORTRAIT GROWTH IN CONTRACTING, REGULAR BRANCH GROUPS
3
Note that, in all three cases, the portrait growth is doubly exponential even
though the word growth function is intermediate for the first Grigorchuk group,
mostly unknown for the GGS-groups, and exponential for the Apollonian group. We
conjecture that the portrait growth is doubly exponential for all finitely generated,
contracting, regular branch groups.
The paper is organized as follows: in Section 2 we provide the basic definitions regarding groups acting on regular rooted trees and we describe a procedure
yielding recursive relations for the portrait growth sequence of a finitely generated,
contracting, regular branch group. In Section 3 we state some useful observations
about sequences of doubly exponential growth. Finally in Sections 4, 5 and 6 we
describe the cases of the first Grigorchuk group, non-symmetric GGS-groups, and
the Apollonian group.
2. Portrait growth sequence on a regular branch contracting
group
A regular rooted tree T is a graph whose vertices are the words over a finite
alphabet X, and two vertices u and v are joined by an edge if v = ux for some
x ∈ X. The empty word, denoted ∅, represents the root of the tree and the tree
is called d-adic if X consists of d elements. The vertices represented by words of
a fixed length constitute a level, that is, the words of length n constitute the nth
level of the tree. The ternary rooted tree based on X = {1, 2, 3} is displayed in
Figure 1.
∅
1
11
12
2
..
.
13
21
22
3
..
.
23
31
32
..
.
33
Figure 1. A ternary rooted tree
An automorphism of the tree T is a permutation of the vertices preserving incidence. Such a permutation, by necessity, must also preserve the root and all
other levels of the tree. The automorphisms of T form the automorphism group
Aut T under composition. A group acting faithfully on a regular rooted tree may
be regarded as a subgroup of Aut T . Every automorphism g can be fully described
by indicating how g permutes the d vertices at level 1 below the root, and how it
acts on the subtrees hanging from each vertex at level 1. Observe that each subtree hanging from a vertex is isomorphic to the whole tree T , so that the following
description makes sense. Namely, we decompose any automorphism g ∈ Aut T as
(1)
g = (g1 , . . . , gd )α,
where α ∈ Sd is the permutation describing the action of g on the d vertices at
level 1, and each gi ∈ Aut T describes the action of g on the ith subtree below
level 1, for i = 1, . . . , d. This process can be repeated at any level, and the element
4
ZORAN ŠUNIĆ AND JONE URIA-ALBIZURI
describing the action of g at a particular vertex u is called the section of g at the
vertex u. Given a group G ≤ Aut T we say that G is self-similar if the sections of
each element in G belong again to the group G.
We say that a self-similar group G ≤ Aut T is contracting if there is a finite
set M of elements in G such that, for every element g, there exists some level n
such that all sections of g at the vertices at and below the nth level belong to
M. The smallest set M among such finite sets is called the nucleus of G and we
denote it by N (G) (or by N , when G is clear from the context). The definition of
a contracting group, in this form, is due to Nekrashevych [Nek05].
For any finitely generating group G, generated by a finite symmetric set S = S −1 ,
we define the word length of g, denoted ∂(g), to be the length of the shortest word
over the alphabet S representing the element g. A self-similar group G generated
by a finite symmetric set S is contracting if and only if there exist constants λ < 1
and C ≥ 0, and level n such that for every g ∈ G and every vertex u on level n
∂(gu ) ≤ λ∂(g) + C.
Note that, if ∂(g) >
C
1−λ ,
then
∂(gu ) ≤ λ∂(g) + C < ∂(g),
that is, the sections gu are strictly shorter than g. Thus, the elements that possibly
have no shortening in their sections are the ones that belong to the finite set
C
g ∈ G | ∂(g) ≤
.
1−λ
In particular, the nucleus is part of this finite set. The metric approach to contracting groups precedes the nucleus definition of Nekrashevych and was used by
Grigorchuk in his early works on families of groups related to the first Grigorchuk
group [Gri80, Gri84]. Nekrasevych [Nek05] showed that he nucleus definition and
the metric definition are equivalent in the case of finitely generated groups.
Given a contracting group G acting on a d-ary tree and an element g in G,
the nucleus portrait is a finite tree, whose interior vertices are decorated by
permutations from Sd and whose leaves are decorated by elements of the nucleus,
describing the action of the element g. The portrait is constructed recursively as
follows. If g is an element of the nucleus the tree consist only of the root decorated
by g. If g is not an element of the nucleus, then we consider its decomposition (1).
The portrait of g is obtained by decorating the root by the permutation α and by
attaching the portraits of the section g1 , . . . , gd at the corresponding vertices of the
first level. Since G is contracting this recursive procedure must end at some point,
and we obtain the portrait of the element g. (A concrete example in the case of
the first Grigorchuk group is provided in Figure 2.)
Let us denote by d(g) the depth of the portrait of an element g ∈ G, that is,
the length of the largest ray from the root to a leaf in the portrait of g. For each
n ∈ N, the set {g ∈ G | d(g) ≤ n} is finite, and the function a : N → N given by
a(n) = |{g ∈ G | d(g) ≤ n}|
is called the portrait growth function, or portrait growth sequence, of G (with
respect to the nucleus).
We now focus on regular branch groups, since their structure gives us a way to
describe the portrait growth function of a contracting group in a recursive way.
Given G ≤ Aut T , the elements of G fixing level n form a normal subgroup of
G, called the nth level stabilizer and denoted stG (n). For every n ∈ N, we have an
PORTRAIT GROWTH IN CONTRACTING, REGULAR BRANCH GROUPS
5
injective homomorphism
ψn : stG (n) −→ Aut T × · · · × Aut T ,
|
{z
}
dn
sending each element g ∈ stG (n) to the dn -tuple consisting of the sections of g at
n
level n. Note that, if the group is self-similar, ψn (stG (n)) ≤ G × .d. . × G. For
simplicity we write ψ = ψ1 . A group G ≤ Aut T is level transitive if it acts
transitively on every level of the tree. A level transitive group G ≤ Aut T is called
regular branch if there exists a normal subgroup K of finite index in G such that
ψ(K ∩ stG (1)) ≥ K × · · · × K.
Observe that, since K is of finite index in G, the group ψ −1 (K × · · · × K) has finite
index in K ∩ stG (1), and hence in G.
We now describe a procedure yielding recursive formulas for the portrait growth
of any finitely generated, contracting, regular branch group G, branching over its
normal subgroup K of index k. Consider a left transversal T = {t1 , . . . , tk } for K
in G and denote by pn (ti ) = |{g ∈ ti K | d(g) ≤ n}| and pn = |{g ∈ G | d(g) ≤ n}|
the sizes of the sets consisting of the elements of depth at most n in the coset ti K
Pk
and in the whole group G, respectively. We have pn = i=1 pn (ti ).
Let S = {s1 , . . . , sℓ } be a left transversal for ψ −1 (K × · · · × K) in K. For
i = 1, . . . , k and j = 1, . . . , ℓ we have
ti sj = (g1 , . . . , gd )α ≡ (tij1 , . . . , tijd )α
(mod K × · · · × K),
for some tijr ∈ T , r = 1, . . . , d. Thus, for n ≥ 0,
(2)
pn+1 (ti ) =
ℓ
X
pn (tij1 ) . . . pn (tijd )
j=1
(3)
pn+1 =
k X
ℓ
X
pn (tij1 ) . . . pn (tijd ).
i=1 j=1
The initial conditions p0 (ti ), i = 1, . . . , k, for the recursive formulas can be
obtained simply by counting the members of the nucleus that come from the corresponding coset, while p0 is the size of the nucleus.
The following observation will be helpful later on. A rooted automorphism is an
automorphisms whose sections, other than at the root, are trivial.
Lemma 2.1. Let G be a finitely generated, contracting, regular branch group
branching over the subgroup K and let T be a transversal of K in G. If a ∈ G is a
rooted automorphism then, for every t ∈ T and n ≥ 1, we have
pn (t) = pn (at) = pn (ta) = pn (ta ).
Proof. Observe that for every g ∈ tK and u ∈ Ln , n ≥ 1, we have
(ag)u = au ga(u) = ga(u) ,
(ga)u = gu ag(u) = gu ,
(g a )u = ga−1 (u) .
Thus, there are bijections between the sets of elements of depth n in tK, atK, taK
and ta K.
6
ZORAN ŠUNIĆ AND JONE URIA-ALBIZURI
3. Doubly exponential growth
We begin by defining sequences of doubly exponential growth.
Definition 3.1. A sequence of positive real numbers {an }n∈N grows doubly exponentially if there exist some positive constants α, β and some γ, d > 1 such that
n
n
αeγd ≤ an ≤ βeγd ,
for every n ∈ N.
In order to show that the portrait growth sequences in our examples are doubly
exponential we need the following auxiliary result.
∞
Lemma 3.2. Let {an }n=0 be a sequence of positive real numbers and d a constant
with d > 1. The following are equivalent.
(i) There exist positive constants A and B such that, for all n ≥ 0,
Aadn ≤ an+1 ≤ Badn .
(ii) There exist positive constants α, β, and γ such that, for all n ≥ 0,
n
n
αeγd ≤ an ≤ βeγd .
∞
Moreover, in case (i) is satisfied, the sequence lndann n=0 is convergent, we may
set
ln an
γ = lim
n→∞ dn
−M
and α and β can be chosen to be e
and eM , respectively, where
1
M=
max{| ln A|, | ln B|}.
d−1
The error of the approximation γ ≈ γn =
ln an
dn
is no greater than M/dn .
Proof. (ii) implies (i). We have, for all n,
β γdn d
β
α d
α γdn d
γdn+1
γdn+1
βe
=
αe
≤
a
≤
βe
=
αe
≤ d adn .
a
≤
n+1
n
d
d
d
β
β
α
α
≤ ln B, and therefore
(i) implies (ii). For all i, we have ln A ≤ ln aai+1
d
i
ln
ai+1
≤ max{| ln A|, | ln B|} = (d − 1)M.
adi
For n ≥ 0, let
rn =
∞
X
ai+1
1
ln d .
i+1
d
ai
i=n
The series rn isPabsolutely convergent and we have the estimate |rn | ≤ M/dn , by
M
1
comparison to ∞
i=n di+1 (d − 1)M = dn .
Let
∞
X
1
ai+1
γ = ln a0 + r0 = ln a0 +
ln d .
i+1
d
ai
i=0
Since r0 is a convergent series, γ is well defined. We have
1 a1
γ = ln a0 + r0 = ln a0 + ln d + r1
d a0
a2
ln a1
1
ln a1
+ r1 =
+ 2 ln d + r2
=
d
d
d
a1
ln a2
= 2 + r2 = . . .
d
PORTRAIT GROWTH IN CONTRACTING, REGULAR BRANCH GROUPS
7
Thus, for all n,
ln an
+ rn .
dn
∞
Since |rn | ≤ M/dn , we see that the sequence lndann n=0 converges to γ. Moreover,
the inequalities γ − M/dn ≤ lndann ≤ γ + M/dn yield
γ=
n
n
e−M eγd ≤ an ≤ eM eγd .
We end this section with a simple combinatorial observation that shows that an
upper bound of the form required in condition (i) of the lemma always exists for
regular branch groups.
Lemma 3.3. Let G be a finitely generated, contracting, regular branch group
acting on the d-adic tree and let {an }n≥0 be the portrait growth sequence of G.
Then
an+1 ≤ |G : stG (1)|adn , for n ≥ 0.
Proof. Since every element of depth at most n + 1 has sections at the first level
that have depth at most n, the number of possible decorations at level 1 and below
for portraits of depth at most n + 1 is at most adn . On the other hand, the number
of possible labels at the root is |G : stG (1)|, and we obtain the inequality.
4. Portrait growth in the first Grigorchuk group
Denote by G the first Grigorchuk group, introduced in [Gri80]. In his treatise [Gri05] on solved and unsolved problems centered around G, Grigorchuk asked
what is the growth of the sequence counting the number of portraits of given size
in G (Problem 3.5).
The first Grigorchuk group is defined as follows.
Definition 4.1. Let T be the binary tree. The first Grigorchuk group G is the
group generated by the rooted automorphism a permuting the two subtrees on
level 1, and by b, c, d ∈ stG (1), where b, c and d are defined recursively by
ψ(b) = (a, c)
ψ(c) = (a, d)
ψ(d) = (1, b)
Already in his early works in 1980s Grigorchuk observed that G is contracting
with nucleus N (G) = {1, a, b, c, d}. Since G is a contracting group, its elements have
well defined portraits, which are finite decorated trees. For instance, the portrait
of the element bacac is provided in Figure 2.
Grigorchuk also showed that G is a regular branch group, branching over the
subgroup K = h[a, b]iG of index |G : K| = 16. An accessible account can be found
in Chapter VIII of [dlH00]. A transversal for K in G is given by
T = { 1, d, ada, dada, a, ad, da, dad, b, c, aca, cada, ba, ac, ca, cad }.
and a transversal for ψ −1 (K × K) in K is given by
S = {1, abab, (abab)2, baba}.
Theorem 4.2. The portrait growth sequence {an }∞
n=0 of the first Grigorchuk group
G is given recursively by
a0 = 5
an = 2xn + 4yn + 2zn + 2Xn + 4Yn + 2Zn , for n ≥ 1,
8
ZORAN ŠUNIĆ AND JONE URIA-ALBIZURI
()
()
(1 2)
1
b
(1 2)
c
d
a
Figure 2. portrait of the element bacac
where xn , yn , zn , Xn , Yn , and Zn , for n ≥ 1, satisfy the system of recursive relations
xn+1 = x2n + 2yn2 + zn2 ,
yn+1 = xn Yn + Yn zn + Xn yn + yn Zn ,
zn+1 = Xn2 + 2Yn2 + Zn2
Xn+1 = 2xn yn + 2yn zn ,
Yn+1 = xn Xn + 2yn Yn + zn Zn ,
Zn+1 = 2Xn Yn + 2Yn Zn ,
with initial conditions
x1 = y1 = z1 = Y1 = 1,
X1 = 2,
Z1 = 0.
Proof. Denote by pn (t) the number of portraits of depth no greater than n in Γ
that represent elements in the coset tK.
By Lemma 2.1, we have
(4)
pn+1 (t) = pn+1 (at) = pn+1 (ta ) = pn+1 (ta), for n ≥ 0, t ∈ T.
Thus we only need to exhibit recursive formulas for the 6 coset representatives in
the set {1, c, dada, b, d, cada}. We have
PORTRAIT GROWTH IN CONTRACTING, REGULAR BRANCH GROUPS
ψ(1) = (1, 1)
ψ(abab) = (ca, ac)
2
ψ((abab) ) = (dada, dada)
ψ(baba) = (ac, ca)
ψ(c) = (a, d)
ψ(cabab) = (aca, cad)
2
ψ(c(abab) ) = (dad, ada)
ψ(dada) = (b, b)
2
ψ(dada(abab) ) = (cada, cada)
9
ψ(cbaba) = (c, ba)
ψ(dadaabab) = (da, ad)
ψ(dadababa) = (ad, da)
ψ(b) = (a, c)
ψ(babab) = (aca, dad)
2
ψ(b(abab) ) = (dad, aca)
ψ(bbaba) = (c, a)
ψ(d) = (1, b)
ψ(dabab) = (ca, ad)
2
ψ(d(abab) ) = (dada, cada)
ψ(cada) = (ba, d)
2
ψ(cada(abab) ) = (cad, ada)
ψ(dbaba) = (ac, da)
ψ(cadaabab) = (ada, cad)
ψ(cadababa) = (d, ba),
where the sections are already written modulo K by using representatives in T (the
spacing into 6 groups of 4 indicates how each of the 6 cosets of K with representatives 1, c, dada, b, d, and cada splits into 4 cosets of K × K).
Thus, for n ≥ 0,
pn+1 (1) = pn (1)2 + 2pn (ac)pn (ca) + pn (dada)2 ,
pn+1 (c) = pn (a)pn (d) + pn (dad)pn (ada) + pn (c)pn (ba) + pn (aca)pn (cad),
pn+1 (dada) = pn (b)2 + 2pn (ad)pn (da) + pn (cada)2
(5)
pn+1 (b) = 2pn (a)pn (c) + 2pn (dad)pn (aca),
pn+1 (d) = pn (1)pn (b) + pn (ac)pn (da) + pn (ca)pn (ad) + pn (dada)pn (cada),
pn+1 (cada) = 2pn (d)pn (ba) + 2pn (ada)pn (cad),
with initial conditions
p0 (1) = p0 (a) = p0 (b) = p0 (c) = p0 (d) = 1,
p0 (t) = 0, for t ∈ T \ {1, a, b, c, d}.
Direct calculations, based on (5), give
p1 (b) = 2
p1 (cada) = 0
p1 (t) = 1, for t ∈ {1, c, d, dada}.
If we denote, for n ≥ 1,
xn = pn (1) = pn (a),
yn = pn (c) = pn (ac) = pn (aca) = pn (ca),
zn = pn (dada) = pn (dad),
Xn = pn (b) = pn (ba),
Yn = pn (d) = pn (ad) = pn (ada) = pn (da),
Zn = pn (cada) = pn (cad)
we obtain, for n ≥ 1,
an = 2xn + 4yn + 2zn + 2Xn + 4Yn + 2Zn ,
10
ZORAN ŠUNIĆ AND JONE URIA-ALBIZURI
where xn , yn , zn , Xn , Yn , and Zn satisfy the recursive relations and initial conditions as claimed, which follows from (5).
Theorem 4.3. There exists a positive constant γ such that the portrait growth
sequence {an }∞
n=0 of the first Grigorchuk group G satisfies the inequalities
n
1 γ2n
e
≤ an ≤ 4eγ2 ,
4
for all n ≥ 0. Moreover, γ ≈ 0.71.
Proof. Following Lemma 3.2, we first determine positive constants A an B such
that for each n ∈ N we have
Aa2n ≤ an+1 ≤ Ba2n .
By Lemma 3.3 we may take B = |G : stG (1)| = 2.
For the other inequality, we need a constant A such that
an+1 − Aa2n ≥ 0.
Using Theorem 4.2 we may express, for n ≥ 1, both an+1 and an in terms of
xn , yn , zn , Xn , Yn , Zn and if we set A = 14 , we obtain
an+1 − Aa2n = (xn − zn + Xn − Zn )2 ≥ 0.
1
1
Since M = d−1
max{| ln A|, | ln B|} = 2−1
max{| ln 1/4|, | ln 2|} = ln 4, we obtain
−M
M
α=e
= 1/4 and β = e = 4. Finally, the approximation of γ ≈ 0.71 can be
calculated by using the recursion given by Theorem 4.2 and Lemma 3.2.
5. Portrait growth in non-symmetric GGS-groups
The GGS-groups (named after Grigorchuk, Gupta, and Sidki) form a family of
groups generalizing the Gupta-Sidki examples [GS83] (which were in turn inspired
by the first Grigorchuk group [Gri80]).
Definition 5.1. For a prime p, p ≥ 3, and a vector e = (e1 , . . . , ep−1 ) ∈ Fp−1
,
p
the GGS-group defined by e is the group of p-ary automorphisms generated by
the rooted automorphism a, which permutes the subtrees on level 1 according to
the permutation (1 . . . p), and the automorphism b ∈ st(1) defined recursively by
b = (b, ae1 , . . . , aep−1 ).
Set S = S −1 = {a, a2 , . . . , ap−1 , b, . . . , bp−1 }. It is easy to see that G is contracting with nucleus N (G) = S ∪ {1}.
Let
0
e1 e2
. . . ep−1
ep−1 0 e1
. . . ep−2
C(0, e) = .
.
.
..
..
..
..
..
.
.
e1
e2 . . . ep−1
0
be the circulant matrix of the vector (0, e1 , . . . , ep−1 ). We say that the vector
e = (e1 , . . . , ep−1 ) is symmetric if ei = ep−i , for i = 1, . . . , p − 1 (that is, the vector
is symmetric precisely when the corresponding circulant matrix is symmetric).
.
Theorem 5.2. Let G be a GGS-group defined by a non-symmetric vector e ∈ Fp−1
p
The portrait growth sequence {an }∞
of
G
is
given
by
n=0
a0 = 1 + 2(p − 1)
n−1
an = p(x1 + (p − 1)y1 )p
,
PORTRAIT GROWTH IN CONTRACTING, REGULAR BRANCH GROUPS
11
where x1 and y1 are the number of solutions in Fpp of
((n0 , . . . , np−1 )C(0, e)) ⊙ (n0 , . . . , np−1 ) = (0, . . . , 0),
with n0 + · · · + np−1 = 0 and n0 + · · · + np−1 = 1, respectively, where ⊙ denotes
the product by coordinates.
Proof. Fernández-Alcober and Zugadi-Reizabal [FAZR14] showed that a GGS-group
defined by a non-symmetric vector is regular branch over G′ , whose index in G is
p2 . A left-transversal for G′ in G is given by
T = {ai bj | i, j = 0, . . . , p − 1}.
For each pair (i, j) ∈ {0, . . . , p − 1}2 denote by pn (i, j) the number of portraits
of depth no greater than n in the coset ai bj G′ .
We have
2
ai bj ≡ ai bn0 (ba )n1 (ba )n2 . . . (ba
where j = n0 + · · · + np−1 in Fp . And then,
ai bj ≡ ai (ai0 bn0 , . . . , aip−1 bnp−1 )
p−1
)np−1
(mod G′ ),
(mod G′ × · · · × G′ ),
where (i0 , . . . , ip−1 ) = (n0 , . . . , np−1 )C(0, e). We obtain that,
X
pn+1 (i, j) =
n0 +···+np−1 =j
p−1
Y
pn (ir , nr ).
r=0
Observe that the decomposition of pn+1 (i, j) does not depend on i, so we can
set pn+1 (i, j) = Pn+1 (j) and we have
(6)
X
pn+1 (j) =
n0 +···+np−1 =j
p−1
Y
pn (ir , nr ),
r=0
Pp−1
and then, for n ≥ 1, we have an = p j=0 pn (j), where we multiply by p because
we have to sum for each i = 0, . . . , p − 1.
Since the nucleus is S ∪ {1}, the initial conditions are the given by
p0 (0, 0) = p0 (i, 0) = p0 (0, j) = 1 for i, j ∈ {1, . . . , p − 1},
p0 (i, j) = 0 otherwise.
In other words, p0 (i, j) = 1 if ij = 0 and p0 (i, j) = 0, otherwise. By (6), p1 (0) is
the number of solutions in Fpp of
(7) (n0 , . . . , np−1 )C(0, e) ⊙ (n0 , n1 , . . . , np−1 ) = (i0 n0 , . . . , ip−1 np−1 ) = (0, . . . , 0),
with n0 + · · · + np−1 = 0, and p1 (j) is the number of solutions of the same equation,
but with n0 + · · · + np−1 = j.
We prove by induction that pn (1) = pn (j), for n ≥ 1 and j 6= 0. Observe that,
for n = 1, if (n0 , . . . , np−1 ) is a solution of (7), with n0 + · · · + np−1 = 1, then
(jn0 , . . . , jnp−1 ) is also a solution, but with n0 + · · · + np−1 = j. Similarly, if we
multiply a solution that sums up to j by the multiplicative inverse j −1 of j in Fp , we
obtain a solution that sums up to 1. Thus, there is a bijection between the solutions
and hence p1 (1) = p1 (j), for j 6= 0. Let us now assume that pn (1) = pn (j), for
n ≥ 1 and j 6= 0, and let us prove the equality for n + 1. By (6) and the assumption
that n ≥ 1, we have that pn (i, j) = pn (j) and
pn+1 (j) =
X
n0 +···+np−1 =j
p−1
Y
r=0
pn (nr ).
12
ZORAN ŠUNIĆ AND JONE URIA-ALBIZURI
By the inductive hypothesis we have pn (nr ) = pn (j −1 nr ) (this is true regardless of
whether nr is 0 or not). Thus,
pn+1 (j) =
p−1
Y
X
n0 +···+np−1 =j
pn (nr ) =
n0 +···+np−1 =j
r=0
X
=
j −1 n0 +···+j −1 np−1 =1
X
p−1
Y
pn (j −1 nr )
r=0
X
pn (j −1 nr ) =
r=0
p−1
Y
n′0 +···+n′p−1 =1
p−1
Y
pn (n′r ) = pn+1 (1).
r=0
We now resolve the recursion (6).
Denote xn = pn (0) and yn = pn (1), for n ≥ 1, so that an = p(xn + (p − 1)yn ).
The fact that pn (ni ) = yn whenever ni 6= 0, together with (6), implies that
X
Y
Y
xn+1 =
xn
yn ,
yn+1 =
n0 +···+np−1 =0
ni =0
X
Y
n0 +···+np−1 =1
ni 6=0
xn
ni =0
Y
yn
ni 6=0
Thus, by making all possible choices of ℓ coordinates, ℓ = 0, . . . , p, in (n0 , . . . , np−1 )
that are equal to 0, we obtain
p
X
p−ℓ ℓ p
(8)
zℓ
xn yn
xn+1 =
ℓ
ℓ=0
p
X
ℓ p
yn+1 =
(9)
z′ ,
xp−ℓ
y
n
n
ℓ ℓ
ℓ=0
where zℓ is the number of solutions of n′1 + · · · + n′ℓ = 0 such that none of n′1 , . . . , n′ℓ
is 0 and zℓ′ the number of solutions of n′1 + · · · + n′ℓ = 1 such that none of n′1 , . . . , n′ℓ
is 0.
For zℓ and zℓ′ , ℓ ≥ 1, we have the relations
zℓ+1 = (p − 1)zℓ′
′
zℓ+1
= zl + (p − 2)zℓ′ ,
with initial conditions z1 = 0 and z1′ = 1. The solution to this system is
1
((p − 1)ℓ − (−1)ℓ−1 (p − 1)),
p
1
zℓ′ = ((p − 1)ℓ − (−1)ℓ ),
p
zℓ =
which, by (8) and (9), gives
1
(xn + (p − 1)yn )p +
p
1
= (xn + (p − 1)yn )p −
p
xn+1 =
yn+1
p−1
(xn − yn )p ,
p
1
(xn − yn )p .
p
Finally, we obtain
xn+1 + (p − 1)(yn+1 ) = (xn + (p − 1)yn )p ,
and we conclude that
n−1
an = p(x1 + (p − 1)y1 )p
.
PORTRAIT GROWTH IN CONTRACTING, REGULAR BRANCH GROUPS
13
6. Portrait growth in the Apollonian group
The Apollonian group is a subgroup of the Hanoi Towers group. The Hanoi
Towers group was introduced by Grigorchuk and the first author [GŠ06] and the
Apollonian group was introduced later by Grigorchuk, Nekrashevych and the first
author [GNŠ06].
Definition 6.1. The Appolonian group A acting on the ternary tree is the group
generated by the automorphisms
x = (1, y, 1)(1 2),
y = (x, 1, 1)(1 3),
z = (1, 1, z)(2 3).
Set S = {x, y, z, x−1 , y −1 , z −1 . It is easy to see that A is contracting with nucleus
N (A) = S ∪ {1}.
Theorem 6.2. The portrait growth sequence {an }∞
n=0 of the Apollonian group is
given, for n ≥ 0, by
n
3n −1
n
1 √ 3
.
an = 3 2 · 7 3 = √ 7 3
3
Proof. Denote by E the subgroup of index 2 in A consisting of the elements in A
that are represented by words of even length over the alphabet {x± , y ± , z ± }. A left
transversal for E in A is given by T = {1, x}. It is known [GŠ07] that the Hanoi
Towers group H is a regular branch group over its commutator H ′ , which is of index
8 in H, and that the index of H ′ × H ′ × H ′ in H ′ is 12. The Apollonian group A
has index 4 in H and contains the commutator H ′ . Moreover, E = H ′ , implying
that A is a regular branch group, branching over E. The index of E × E × E in E
is 12, and a transversal is given by
T ′ = {1, yx, (yx)2 , x2 , y 2 , z 2 , x2 yx, y 3 x, z 2 yx, x2 (yx)2 , y 2 (yx)2 , z 2 (yx)2 }.
(1, 1, 1)
=1
(x, y, 1)(1 3 2)
= yx
≡ 1,
2
(x, yx, y)(1 2 3)
= (yx)
(y, y, 1)
= x2
(x, 1, x)
=y
2
=z
2
(1, z, z)
2
2
(yx, y , 1)(1 3 2) = x yx
2
(x , y, x)(1 3 2)
(x, zy, z)(1 3 2)
3
=y x
2
= z yx
≡ x,
(1, y, 1)(1 2)
=x
≡ 1,
(y, yx, 1)(2 3)
= xyx
(yx, yx, y)(1 3)
= x(yx)
≡ 1,
(y, y 2 , 1)(1 2)
= x3
≡ 1,
≡ 1,
(1, yx, x)(1 2)
≡ 1,
(z, y, z)(1 2)
≡ 1,
2
≡ 1,
≡ 1,
2
2
2
= xy
2
= xz
2
3
(y , y x, 1)(2 3) = x yx
(y, yx , x)(2, 3)
(zy, yx, z)(2 3)
3
= xy x
2
= xz yx
≡ x,
≡ x,
≡ x,
≡ x,
≡ x,
≡ x,
≡ x,
≡ x,
(yx, y 2 x, y)(1 2 3) = x2 (yx)2 ≡ 1,
(y 2 x, y 2 x, y)(1 3) = x3 (yx)2 ≡ x,
(x, zyx, zy)(1 2 3) = z 2 (yx)2 ≡ 1,
(zyx, yx, zy)(1 3) = xz 2 (yx)2 ≡ x,
(x2 , yx, xy)(1 2 3) = y 2 (yx)2 ≡ 1,
(yx, yx2 , xy)(1 3) = xy 2 (yx)2 ≡ x,
Table 1. The cosets of E × E × E decomposing the cosets of E
14
ZORAN ŠUNIĆ AND JONE URIA-ALBIZURI
Denote by Xn and Yn the number of portraits of depth at most n in the cosets
1E and xE respectively. The coset decomposition provided in Table 1 implies that
Xn+1 = 3Xn3 + 9Xn Yn2 ,
Yn+1 = 3Yn3 + 9Xn2 Yn .
which then yields
an+1 = Xn+1 + Yn+1 = 3(Xn + Yn )3 = 3a3n .
Taking into account the initial condition a0 = 7, we obtain an = 3
induction.
3n −1
2
n
73 by
References
[Bar98]
Laurent Bartholdi. The growth of Grigorchuk’s torsion group. Internat. Math. Res.
Notices, (20):1049–1054, 1998.
[dlH00]
Pierre de la Harpe. Topics in geometric group theory. Chicago Lectures in Mathematics.
University of Chicago Press, Chicago, IL, 2000.
[FAZR14] Gustavo A. Fernández-Alcober and Amaia Zugadi-Reizabal. GGS-groups: order of congruence quotients and Hausdorff dimension. Trans. Amer. Math. Soc., 366(4):1993–
2017, 2014.
[GNŠ06] Rostislav Grigorchuk, Volodymyr Nekrashevych, and Zoran Šunić. Hanoi towers groups.
In Topological and Geometric Methods in Group Theory, volume 3 issue 2 of Oberwolfach Reports. 2006.
[Gri80]
R. I. Grigorčuk. On Burnside’s problem on periodic groups. Funktsional. Anal. i
Prilozhen., 14(1):53–54, 1980.
[Gri83]
R. I. Grigorchuk. On the Milnor problem of group growth. Dokl. Akad. Nauk SSSR,
271(1):30–33, 1983.
[Gri84]
R. I. Grigorchuk. Degrees of growth of finitely generated groups and the theory of
invariant means. Izv. Akad. Nauk SSSR Ser. Mat., 48(5):939–985, 1984.
[Gri05]
Rostislav Grigorchuk. Solved and unsolved problems around one group. In Infinite
groups: geometric, combinatorial and dynamical aspects, volume 248 of Progr. Math.,
pages 117–218. Birkhäuser, Basel, 2005.
[Gro81]
Mikhael Gromov. Groups of polynomial growth and expanding maps. Inst. Hautes
Études Sci. Publ. Math., (53):53–73, 1981.
[GS83]
Narain Gupta and Saı̈d Sidki. On the Burnside problem for periodic groups. Math. Z.,
182(3):385–388, 1983.
[GŠ06]
Rostislav Grigorchuk and Zoran Šuniḱ. Asymptotic aspects of Schreier graphs and
Hanoi Towers groups. C. R. Math. Acad. Sci. Paris, 342(8):545–550, 2006.
[GŠ07]
Rostislav Grigorchuk and Zoran Šunić. Self-similarity and branching in group theory.
In Groups St. Andrews 2005. Vol. 1, volume 339 of London Math. Soc. Lecture Note
Ser., pages 36–95. Cambridge Univ. Press, Cambridge, 2007.
[Mil68a] J. Milnor. A note on curvature and fundamental group. J. Differential Geometry, 2:1–7,
1968.
[Mil68b] John Milnor. Growth of finitely generated solvable groups. J. Differential Geometry,
2:447–449, 1968.
[Mil68c] John Milor. Problem 5603. Amer. Math. Monthly, 75:685–686, 1968.
[Nek05]
Volodymyr Nekrashevych. Self-similar groups, volume 117 of Mathematical Surveys
and Monographs. American Mathematical Society, Providence, RI, 2005.
[Sid87]
Said Sidki. On a 2-generated infinite 3-group: subgroups and automorphisms. J. Algebra, 110(1):24–55, 1987.
[Šva55]
A. S. Švarc. A volume invariant of coverings. Dokl. Akad. Nauk SSSR (N.S.), 105:32–34,
1955.
Department of Mathematics, Hofstra University, Hempstead, NY 11549, USA
E-mail address: [email protected]
Department of Mathematics, University of the Basque Country, UPV/EHU, Leioa,
Bizkaia, Spain.
E-mail address: [email protected]
| 4math.GR
|
On Zariski’s Cancellation Problem in positive
characteristic
Neena Gupta
Stat-Math Unit, Indian Statistical Institute
arXiv:1309.1368v1 [math.AC] 5 Sep 2013
203 B.T. Road, Kolkata 700108, India
e-mail: [email protected], [email protected]
Abstract
In this paper we shall show that when k is a field of positive characteristic the affine
space Ank is not cancellative for any n ≥ 3.
1
Introduction
Let k be an algebraically closed field. The Zariski Cancellation Problem for Affine Spaces
asks whether the affine space Ank is cancellative, i.e., if V is an affine k-variety such that
V × A1k ∼
, does it follow that V ∼
= An+1
= Ank ? Equivalently, if A is an affine k-algebra such
k
that A[X] is isomorphic to the polynomial ring k[X1 , . . . , Xn+1 ], does it follow that A is
isomorphic to k[X1 , . . . , Xn ]?
The affine line A1k was shown to be cancellative by S. S. Abhyankar, P. Eakin and W. J.
Heinzer ([1]) and the affine plane A2k was shown to be cancellative by T. Fujita, M. Miyanishi
and T. Sugie ([4], [7]) in characteristic zero and by P. Russell ([8]) in positive characteristic.
However, in [5], the author showed when ch. k > 0, the affine space A3k is not cancellative by
proving that a threefold constructed by Asanuma in [2] is not isomorphic to the polynomial
ring k[X1 , X2 , X3 ].
In this paper, we shall show when ch. k > 0, the affine space Ank is not cancellative
for any n ≥ 3 (Theorem 3.6). This completely settles the Zariski’s Cancellation problem in
positive characteristic.
2
Preliminaries
We shall use the notation R[n] for a polynomial ring in n variables over a ring R.
We shall also use the following term from affine algebraic geometry.
Definition. An element f ∈ k[Z, T ] is called a line if k[Z, T ]/(f ) = k [1] . A line f is called
a non-trivial line if k[Z, T ] 6= k[f ][1] .
1
We recall the definition of an exponential map (a formulation of the concept of Ga -action)
and associated invariants.
Definition. Let A be a k-algebra and let φ : A → A[1] be a k-algebra homomorphism. For
an indeterminate U over A, let the notation φU denote the map φ : A → A[U]. φ is said to
be an exponential map on A if φ satisfies the following two properties:
(i) ε0 φU is identity on A, where ε0 : A[U] → A is the evaluation at U = 0.
(ii) φV φU = φV +U , where φV : A → A[V ] is extended to a homomorphism φV : A[U] →
A[V, U] by setting φV (U) = U.
The ring of φ-invariants of an exponential map φ on A is a subring of A given by
Aφ = {a ∈ A | φ(a) = a}.
An exponential map φ is said to be non-trivial if Aφ 6= A. For an affine domain A over a
field k, let EXP(A) denote the set of all exponential maps on A. The Derksen invariant of
A is a subring of A defined by
DK(A) = k[f | f ∈ Aφ , φ a non-trivial exponential map].
We recall below a crucial observation (cf. [5, Lemma 2.4]).
Lemma 2.1. Let k be a field and A = k [n] , where n > 1. Then DK(A) = A.
We shall also use the following result proved in [5, Lemma 3.3].
Lemma 2.2. Let B be an affine domain over an infinite field k. Let f ∈ B be such that
f − λ is a prime element of B for infinitely many λ ∈ k. Let φ be a non-trivial exponential
map on B such that f ∈ B φ . Then there exist infinitely many β ∈ k such that each f −β is a
prime element of B and φ induces a non-trivial exponential map φ̂ on B/(f − β). Moreover,
B φ /(f − β)B φ is contained in (B/(f − β))φ̂ .
We shall also use the following result proved in [6, Theorem 3.7].
Theorem 2.3. Let k be a field and A be an integral domain defined by
A = k[X, Y, Z, T ]/(X mY − F (X, Z, T )), where m > 1.
Set f (Z, T ) := F (0, Z, T ). Let x, y, z and t denote, respectively, the images of X, Y , Z and
T in A. Suppose that DK(A) 6= k[x, z, t]. Then the following statements hold.
(i) There exist Z1 , T1 ∈ k[Z, T ] and a0 , a1 ∈ k [1] such that k[Z, T ] = k[Z1 , T1 ] and f (Z, T ) =
a0 (Z1 ) + a1 (Z1 )T1 .
(ii) If k[Z, T ]/(f ) = k [1] , then k[Z, T ] = k[f ][1] .
2
An exponential map φ on a graded ring A is said to be homogeneous if φ : A → A[U]
becomes homogeneous when A[U] is given a grading induced from A such that U is a homogeneous element.
We state below a result on homogenization of exponential maps due to H. Derksen, O.
Hadas and L. Makar-Limanov ([3], cf. [5, Theorem 2.3]).
Theorem 2.4. Let A be an affine domain over a field k with an admissible proper Z-filtration
and gr(A) the induced Z-graded domain. Let φ be a non-trivial exponential map on A. Then
φ induces a non-trivial homogeneous exponential map φ̄ on gr(A) such that ρ(Aφ ) ⊆ gr(A)φ̄ .
3
Main Theorems
Throughout the paper, k will denote a field (of any characteristic unless otherwise specified)
and A an integral domain defined by
A = k[X1 , . . . , Xm , Y, Z, T ]/(X1r1 · · · Xm rm Y − F (X1 , . . . , Xm , Z, T )),
where ri > 1 for each i, 1 ≤ i ≤ m. The images of X1 , . . . , Xm , Y , Z and T in A will be
denoted by x1 , . . . , xm , y, z and t respectively.
Set B := k[x1 , . . . , xm , z, t](= k [m+2] ). We note that B ֒→ A ֒→ B[(x1 · · · xm )−1 ]. For
each m-tuple (q1 , . . . , qm ) ∈ Zm , consider the Z-grading on B[(x1 · · · xm )−1 ] given by
M
M
Bi , where Bi =
k[z, t]x1 i1 x2 i2 . . . xm im .
B[(x1 · · · xm )−1 ] =
i∈Z
(i1 ,...,im )∈Zm ,q1 i1 +q2 i2 +···+qm im =i
P
Each element a ∈ B[(x1 · · · xm )−1 ] can be uniquely written as a =
ℓa ≤j≤ua aj , where
aj ∈ Bj . (Note that if a ∈ B then aj ∈ B for each j, ℓa ≤ j ≤ ua .) We call ua the degree of
a and aua the leading homogeneous summand of a.
Using this gradingL
on B[(x1 · · · xm )−1 ], we can define a proper Z-filtration {An }n∈Z on A
by setting An := A ∩ i≤n Bi . Then xj ∈ Aqj \ Aqj −1 , 1 ≤ j ≤ m and z, t ∈ A0 \ A−1 . Since
A is an integral domain, we have F 6= 0 and hence FuF 6= 0. Thus, y L
∈ Ab \ Ab−1 , where
b = uF − (q1 r1 + · · · + qm rm ). Let gr A denote the induced graded ring n∈Z An /An−1 .
Note that each element h ∈ A can be uniquely written as sum of monomials of the form
x1 i1 · · · xm im z j1 tj2 y ℓ ,
(1)
where ij ≥ 0, 1 ≤ j ≤ m and ℓ ≥ 0 satisfying that if ℓ > 0 then is < rs for at least one s,
1 ≤ s ≤ m. Therefore, it can be seen that, if h ∈ An \An−1 then h can be uniquely written as
sum of monomials of the form as in equation (1) and each of these monomials lies in An . Thus,
the filtration defined on A is admissible with the generating set Γ := {x1 , . . . , xm , y, z, t} and
so gr A is generated by image of Γ in gr A (cf. [5, Remark 2.2 (2)]).
We now exhibit a structure of gr A when FuF is not divisible by xj for any j.
Lemma 3.1. Suppose that FuF is not divisible by xj for any j, 1 ≤ j ≤ m, then gr A is
isomorphic to
k[X1 , . . . , Xm , Y, Z, T ]
D=
.
r1
(X1 · · · Xm rm Y − F (X1 , . . . , Xm , Z, T )uF )
3
Proof. For a ∈ A, let gr(a) denote the image of a in gr A. Then, as discussed above, gr A
is generated by gr(x1 ), . . . , gr(xm ), gr(y), gr(z) and gr(t). Note that if a ∈ B(⊆ A),
then gr(a) = gr(aua ). As x1 r1 · · · xm rm y(= F ) ∈ AuF \ AuF −1 and hence x1 r1 · · · xm rm y −
F (x1 , . . . , xm , z, t)uF ∈ AuF −1 . Therefore,
gr(x1 )r1 · · · gr(xm )rm gr(y) − gr(FuF ) = 0 in gr A.
Since FuF is not divisible by xj for any j, 1 ≤ j ≤ m, D is an integral domain. As gr A
can be identified with a subring of gr(B[(x1 · · · xm )−1 ]) ∼
= B[(x1 · · · xm )−1 ], we see that the
elements gr(x1 ), . . . , gr(xm ), gr(z), gr(t) of gr A are algebraically independent over k. Hence
gr A ∼
= D.
Lemma 3.2. We have B(= k[x1 , . . . , xm , z, t]) ⊆ DK(A).
Proof. Define φ1 by φ1 (xj ) = xj for each j, 1 ≤ j ≤ m, φ1 (z) = z, φ1 (t) = t + x1 r1 · · · xm rm U
and
F (x1 , . . . , xm , z, t + x1 r1 · · · xm rm U)
φ1 (y) =
= y + Uα(x1 , . . . , xm , z, t, U)
x1 r1 · · · xm rm
and define φ2 by φ2 (xj ) = xj for each j, 1 ≤ j ≤ m, φ2 (t) = t, φ2 (z) = z + x1 r1 · · · xm rm U,
φ2 (y) =
F (x1 , . . . , xm , z + x1 r1 · · · xm rm U, t)
= y + Uβ(x1 , . . . , xm , z, t, U).
x1 r1 · · · xm rm
Note that α(x1 , . . . , xm , z, t, U), β(x1 , . . . , xm , z, t, U) ∈ k[x1 , . . . , xm , z, t, U] and that k[x1 , . . . , xm , z]
and k[x1 , . . . , xm , t] are algebraically closed in A of transcendence degree m + 1 over k. It
then follows that φ1 and φ2 are nontrivial exponential maps on A with Aφ1 = k[x1 , . . . , xm , z]
and Aφ2 = k[x1 , . . . , xm , t]. Hence k[x1 , . . . , xm , z, t] ⊆ DK(A).
We now prove a generalisation of Theorem 2.3.
Proposition 3.3. Suppose that f (Z, T ) := F (0, . . . , 0, Z, T ) 6= 0 and that DK(A) = A.
Then the following statements hold.
(i) There exist Z1 , T1 ∈ k[Z, T ] and a0 , a1 ∈ k [1] such that k[Z, T ] = k[Z1 , T1 ] and f (Z, T ) =
a0 (Z1 ) + a1 (Z1 )T1 .
(ii) Suppose that k[Z, T ]/(f ) = k [1] . Then k[Z, T ] = k[f ][1] .
Proof. (i) We prove the result by induction on m. The result is true for m = 1 by Theorem
2.3. Suppose that m > 1. Set B := k[x1 , x2 , . . . , xm , z, t](⊆ A). Since DK(A) = A, there
exists an exponential map φ on A such that Aφ * B. Let g ∈ Aφ \ B. Since g ∈
/ B, by
equations (1), there exists a monomial in g which is of the form x1 i1 . . . xm im z j1 tj2 y ℓ where
ℓ > 0 and is < rs for some s, 1 ≤ s ≤ m. Without loss of generality, we may assume that
s = 1.
Consider the proper Z-filtration on A with respect to (−1, 0, . . . , 0) ∈ Zm and let A
denote the induced graded ring. For h ∈ A, let h̄ denote the image of h in A. Since
4
A is an integral domain, we have F (0, x2 , . . . , xm , z, t) 6= 0 and so, F (x1 , . . . , xm , z, t)uF =
F (0, x2 , . . . , xm , z, t) = G (say). Since f (Z, T ) 6= 0, by Lemma 3.1, we have
A∼
= k[X1 , . . . , Xm , Y, Z, T ]/(X1r1 · · · Xm rm Y − G).
By Theorem 2.4, φ induces a non-trivial homogeneous exponential map φ̄ on A such that
φ̄
φ̄
ḡ ∈ A . By the choice of g and the filtration defined on A, we have ȳ | ḡ. Since A is
φ̄
factorially closed in A (cf. [5, Lemma 2.1 (i)]), it follows that ȳ ∈ A .
We now consider the proper Z-filtration on A with respect to (−1, −1, . . . , −1) ∈ Zm and
e denote the induced graded ring. For h̄ ∈ A, let e
e Since
let A
h denote the image of h̄ in A.
f (Z, T ) 6= 0, we have GuG = f (z, t) and hence by Lemma 3.1,
e∼
A
= k[X1 , . . . , Xm , Y, Z, T ]/(X1r1 · · · Xm rm Y − f (Z, T )).
e such
Again by Theorem 2.4, φ̄ induces a non-trivial homogeneous exponential map φe on A
e
e
φ
φ
e
e
that ye ∈ A . Since tr. degk A = m + 1, there exist m algebraically independent elements in
eφe over k[e
A
y ].
e
φ
e
eφe is algebraically closed in A,
e we have A
eφe =
If A ⊆ k[e
y , ze, e
t], then since m ≥ 2 and A
rm
eφe is
eφe as A
e(= f (e
z, e
t)) ∈ k[e
y , ze, e
t], we have xe1 , . . . , xf
e∈ A
k[e
y , ze, e
t]. Since xe1 r1 · · · xf
m, y
m y
e This would contradict the fact that φe is non-trivial. Thus, there
factorially closed in A.
eφe \ k[e
y , ze, e
t]. Hence, h contains a monomial which is
exists an homogeneous element h ∈ A
divisible by xei for some i, 1 ≤ i ≤ m. Without loss of generality we may assume that xe2
divides a monomial of h.
e with respect to the m-tuple (0, 1, 0, . . . , 0) ∈
Now again consider the proper Z-filtration on A
e For
e of A.
e Then gr A
e∼
Zm defined as in Lemma 3.1 and the induced graded ring gr A
= A.
e let gr(a) denote the image of e
e Then gr(x2 ) | gr(h) in gr A. Again by
e
a ∈ A,
a in gr A.
e such that
Theorem 2.4, φe induces a non-trivial homogeneous exponential map gr φ on gr A
gr φ
gr φ
gr(y) and gr(h) ∈ gr A . Since gr(x2 ) | gr(h) in gr A, we have gr(x2 ) ∈ gr A .
Let F be an algebraic closure of the field k. Then gr φ induces a non-trivial exponential
map ψ on
e ⊗k F ∼
E = gr A
=
F [X1 , . . . , Xm , Y, Z, T ]
= F [xˆ1 , . . . , xˆm , ẑ, t̂, ŷ],
(X1 r1 · · · Xm rm Y − f (Z, T ))
such that F [ŷ, xˆ2 ] ⊆ E ψ . Since xˆ2 − λ is a prime element in E for every λ ∈ F ∗ , by Lemma
2.2, ψ induces a non-trivial exponential map on
E/(xˆ2 − λ)E(∼
=
F [X1 , X3 , . . . , Xm , Y, Z, T ]
(λr2 X1 r1 X3 r3 · · · Xm rm Y − f (Z, T ))
for some λ ∈ F ∗ such that the image of ŷ lies in the ring of invariants. Therefore, using
Lemma 3.2, we have DK(E/(xˆ2 − λ)E) = E/(xˆ2 − λ)E. We are thus through by induction
on m.
(ii) Suppose that f (Z, T ) is a line in k[Z, T ]. Then A/(x1 , . . . , xm ) = k[Y, Z, T ]/(f (Z, T )) ∼
=
[2]
∗
∗
k and hence (A/(x1 , . . . , xm )) = k . By (i) above, there exist Z1 , T1 ∈ k[Z, T ] and
5
a0 , a1 ∈ k [1] such that k[Z, T ] = k[Z1 , T1 ] and f (Z, T ) = a0 (Z1 ) + a1 (Z1 )T1 . If a1 (Z1 ) = 0,
then f (Z, T ) = a0 (Z1 ) must be a linear polynomial in Z1 and hence a variable in k[Z, T ].
Now suppose a1 (Z1 ) 6= 0. As f (Z, T ) is irreducible in k[Z, T ], a0 (Z1 ) and a1 (Z1 ) are coprime
1
in k[Z1 ]. Hence A/(x1 , . . . , xm ) ∼
][1] and since (A/(x1 , . . . , xm ))∗ = k ∗ , we have
= k[Z1 , a1 (Z
1)
a1 (Z1 ) ∈ k ∗ . This again implies that f (Z, T ) is a variable in k[Z, T ].
We now deduce a result for a field of positive characteristic.
Corollary 3.4. Suppose that ch. k > 0 and that f (Z, T ) ∈ k[Z, T ] is a non-trivial line in
k[Z, T ]. Then, A ≇ k [m+2] .
Proof. Suppose, if possible that A ∼
= k [m+2] . Then, by Lemma 2.1, DK(A) = A. Therefore,
by Proposition 3.3(ii), k[Z, T ] = k[f ][1] . This contradicts the given hypothesis.
Theorem 3.5. Let k be a field of any characteristic and A an integral domain defined by
A = k[X1 , . . . , Xm , Y, Z, T ]/(X1r1 · · · Xm rm Y − f (Z, T )),
where ri > 1 for each i. Suppose that f (Z, T ) is a line in k[Z, T ]. Then
A[1] = k[X1 , . . . , Xm ][3] ∼
= k [m+3] .
Proof. Let h ∈ k[Z, T ] be such that k[Z, T ] = k[h] + (f )k[Z, T ]. Let P, Q ∈ k [1] be such that
Z = P (h) + f P1 (Z, T ) and T = Q(h) + f Q1 (Z, T ) for some P1 , Q1 ∈ k[Z, T ]. Let W be an
indeterminate over A. Set
W1 : = X1 r1 · · · Xm rm W + h(Z, T )
Z1 : = (Z − P (W1 ))/X1 r1 · · · Xm rm
T1 : = (T − Q(W1 ))/X1 r1 · · · Xm rm .
Let y be the image of Y in A. We show that A[W ] = k[X1 , . . . , Xm , Z1, T1 , W1 ]. Set
B := k[X1 , . . . , Xm , Z1 , T1 , W1 ]. We have
Z = P (W1 ) + X1 r1 · · · Xm rm Z1 ,
T = Q(W1 ) + X1 r1 · · · Xm rm T1 ,
f (Z, T )
f (X1 r1 · · · Xm rm Z1 + P (W1 ), X1 r1 · · · Xm rm T1 + Q(W1 ))
y =
=
X1 r 1 · · · Xm r m
X1 r 1 · · · Xm r m
f (P (W1 ), Q(W1 ))
=
+ α(X1 , . . . , Xm , Z1 , T1 , W1 ),
X1 r 1 · · · Xm r m
W1 − h(X1 r1 · · · Xm rm Z1 + P (W1 ), X1 r1 · · · Xm rm T1 + Q(W1 ))
W1 − h(Z, T )
=
W =
X1 r 1 · · · Xm r m
X1 r 1 · · · Xm r m
W1 − h(P (W1 ), Q(W1 ))
+ β(X1 , . . . , Xm , Z1 , T1 , W1 )
=
X1 r 1 · · · Xm r m
6
for some α, β ∈ B. Since f (P (W1 ), Q(W1 )) = 0 and h(P (W1 ), Q(W1 )) = W1 , we see that
y, W ∈ B. Hence, A[W ] ⊆ B. We now show that B ⊆ A[W ]. We have,
Z − P (W1 )
Z − P (X1 r1 · · · Xm rm W + h(Z, T ))
=
X1 r 1 · · · Xm r m
X1 r 1 · · · Xm r m
f (Z, T )P1(Z, T )
Z − P (h(Z, T ))
+ γ(X1 , . . . , Xm , Z, T, W )
=
r1
rm + γ(X1 , . . . , Xm , Z, T, W ) =
X1 · · · Xm
X1 r 1 · · · Xm r m
= P1 (Z, T )y + γ(X1 , . . . , Xm , Z, T, W ),
Z1 =
and
T − Q(W1 )
T − Q(X1 r1 · · · Xm rm W + h(Z, T ))
=
X1 r 1 · · · Xm r m
X1 r 1 · · · Xm r m
f (Z, T )Q1(Z, T )
T − Q(h(Z, T ))
+ δ(X1 , . . . , Xm , Z, T, W )
=
r1
rm + δ(X1 , . . . , Xm , Z, T, W ) =
X1 · · · Xm
X1 r 1 · · · Xm r m
= Q1 (Z, T )y + δ(X1 , . . . , Xm , Z, T, W )
T1 =
for some γ, δ ∈ A[W ]. Thus, Z1 , T1 ∈ A[W ]. Hence B ⊆ A[W ]. Since B = k[X1 , . . . , Xm ][3] ,
the result follows.
Theorem 3.6. When k is a field of positive characteristic, Zariski’s Cancellation Conjecture
does not hold for the affine n-space Ank for any n ≥ 3.
Proof. There exists non-trivial line in k[Z, T ] when ch k = p > 0. For instance, we may take
2
the Nagata’s line f (Z, T ) = Z p + T + T qp , where q is a prime other than p. The result now
follows from Theorem 3.5 and Corollary 3.4.
References
[1] S. Abhyankar, P. Eakin and W. Heinzer, On the uniqueness of the coefficient ring in a
polynomial ring, J. Algebra 23 (1972) 310–342.
[2] T. Asanuma, Polynomial fibre rings of algebras over Noetherian rings, Inventiones Math.
87 (1987) 101–127.
[3] H. Derksen, O. Hadas and L. Makar-Limanov, Newton polytopes of invariants of additive
group actions, J. Pure and Applied Algebra 156 (2001) 187–197.
[4] T. Fujita, On Zariski problem, Proc. Japan Acad. 55(A) (1979) 106–110.
[5] Neena Gupta, On the Cancellation Problem for the Affine Space A3 in characteristic p, Inventiones Math. DOI 10.1007/s00222-013-0455-2, available online at
http://link.springer.com/article/10.1007%2Fs00222-013-0455-2
[6] Neena Gupta, On the family of affine threefolds xm y = F (x, z, t), preprint.
[7] M. Miyanishi and T. Sugie, Affine surfaces containing cylinderlike open sets, J. Math.
Kyoto Univ. 20 (1980) 11–42.
[8] P. Russell, On Affine-Ruled rational surfaces, Math. Annalen 255(3) (1981) 287–302.
7
| 0math.AC
|
Stability and performance analysis of linear positive systems with
delays using input-output methods
arXiv:1703.00405v1 [math.OC] 1 Mar 2017
Corentin Briat∗
Abstract
It is known that input-output approaches based on scaled small-gain theorems with constant Dscalings and integral linear constraints are non-conservative for the analysis of some classes of linear
positive systems interconnected with uncertain linear operators. This dramatically contrasts with the
case of general linear systems with delays where input-output approaches provide, in general, sufficient
conditions only. Using these results we provide simple alternative proofs for many of the existing results
on the stability of linear positive systems with discrete/distributed/neutral time-invariant/-varying delays
and linear difference equations. In particular, we give a simple proof for the characterization of diagonal
Riccati stability for systems with discrete-delays and generalize this equation to other types of delay
systems. The fact that all those results can be reproved in a very simple way demonstrates the importance
and the efficiency of the input-output framework for the analysis of linear positive systems. The approach
is also used to derive performance results evaluated in terms of the L1 -, L2 - and L∞ -gains. It is also
flexible enough to be used for design purposes.
1
Introduction
Positive systems [1] are a class of systems that are able to represent important processes arising, among
others, in epidemiology, biology, biochemistry, ecology; see e.g. [1–7]. They also naturally arise in the design
of interval observers, a class of observers whose error dynamics is purposely governed by a positive system and
which allows to estimate upper- and lower-bounds on the state of the system; see e.g. [8–13]. Finally, they can
be used as comparison systems for the analysis of more complex systems, notably, for the analysis of systems
with delays; see e.g. [14, 15]. Besides these applicability properties, they have been shown to exhibit very
interesting theoretical properties. For instance, structured state-feedback controllers and certain instances
of the static output-feedback controllers can be designed in a non-conservative way by solving tractable
linear programs [16–18]. The L1 -, L2 - and L∞ -gains of such systems can be also easily characterized in
terms of linear [19, 18, 20, 21] or semidefinite programs [22]. The robust stability analysis of such systems
subject to parametric uncertainties can be exactly performed using scaled small-gain results with constant
D-scalings [18, 23] or integral linear constraints [19, 18, 24], the latter being the linear counterpart of the
integral quadratic constraints [25, 26]. Finally, it also got recently proved that the scaled-small gain theorem
in the L2 -framework states a necessary and sufficient condition for the stability of interconnections in the
special case of positive systems [23], a fact that does not hold true for general linear systems affected by
time-invariant parametric uncertainties; see e.g. [27]. A possible workaround to this problem is to consider
instead the L∞ -framework [28, 29] where the scaled-small gain theorem with constant D-scalings states a
necessary and sufficient condition for the robust stability of linear systems.
The influence of delays on the dynamics of linear positive systems and certain classes of nonlinear monotone systems have been well studied and several necessary and sufficient conditions for the stability have
been obtained using various approaches; see e.g. [30, 31, 18, 32–34]. We propose here to reprove many of the
existing result pertaining on linear systems using a different approach, namely, using input-output approaches
∗
Swiss Federal Institute of Technology–Zürich, Department of Biosystems Science and Engineering, Basel, Switzerland.
Email: [email protected],[email protected]; url: http://www.briat.info
1
and, more specifically, using scaled-small gain results with D-scalings specialized to linear positive systems
[23, 19, 18] and integral linear constraints results [19, 18, 24]. Albeit popular (see e.g. [35–47, 33, 48, 49]),
input/output methods do not seem to have been applied so far for the analysis of linear positive systems with
delays. We notably show that the following statements are rather immediate consequences of scaled-small
gain results and integral linear constraint results:
(i) A linear positive system with discrete constant time-delay is stable if and only if the same system with
the delay set to 0 is also stable [30, 18].
(ii) A linear positive system with bounded discrete time-varying delay is stable if and only if the same
system with constant delay is also stable [31, 18]. This is generalized in [50] to the case of time-varying
distributed delays and to the case of arbitrarily large discrete-delays in [34].
(iii) A linear positive system with constant discrete delay is stable if and only if the associated Riccati
equation has diagonal solutions [32, 51].
(iv) A linear positive coupled differential-difference equation with a single time-varying discrete delay is
stable if and only if the same system with the delay set to 0 is also stable [52].
(v) A linear positive system with discrete time-varying delays is stable if and only if two conditions (which
will stated later), known to be only necessary for the stability of general time-delay systems, are satisfied
[33].
(vi) A linear positive system with distributed time-varying delay is stable if and only if the sum of the matrix
acting on the non-delayed state and the integral of the distributed-delay kernel is Hurwitz stable [50].
(vii) A linear positive neutral system is stable if and only if the system with zero delay is also stable and
it is strongly stable [53, 54]. In particular, it is shown that the strong stability of the difference
equation together with the stability of the retarded part is equivalent to the stability of the neutral
delay equation.
In this regard, the contribution of the paper is not only the development of some new stability results but
also to provide a different, simple and flexible approach for the analysis of linear positive systems with delays. The approach can then be extended to cope with additional uncertainties (e.g. additional parametric
uncertainties, sector-nonlinearities, etc.) and can be used for design purposes (e.g. for the design of interval
observers [8–13]).
Outline. The structure of the paper is as follows: definitions and preliminary results are given in Section
2. General stability results for uncertain linear positive systems are presented in Section 3 and are applied
to linear positive systems with discrete delays in Section 4, to linear positive delay-difference equations in
Section 5, to linear positive coupled differential-difference equations in Section 6, to linear positive systems
with distributed delays in Section 7, to linear positive neutral systems in Section 8.
Notations. The cone of positive and nonnegative vectors of dimension n are denoted by Rn>0 and Rn≥0 ,
respectively. The set of positive integers is given by Z>0 . For two real full matrices A, B having the same
dimension, the inequalities A > (≥)B are componentwise while for two real symmetric matrices A, B having
the same dimension, the relation A ≺ B means that A − B is negative definite. We denote the set of n × n
positive definite diagonal matrices by Dn≻0 . We denote by ρ(A) the spectral radius of the square matrix
A. The n-dimensional vector of ones is denoted by 1n . For a vector v ∈ Rn , ||v||p denotes the standard
vector p-norm while for a matrix M ∈ Rn×m , ||M ||p := max||v||p =1 ||M v||p is the matrix induced p-norm.
For some matrices M1 , . . . , Mn of appropriate dimensions, we define rowN
i=1 {Mi } := M1 . . . MN and
T
T T
.
colN
. . . MN
i=1 {Mi } := M1
2
2
Preliminaries
2.1
System definition
Let us consider the following linear system:
ẋ(t) =
z(t) =
Ax(t) + Ew(t), x(0) = x0
Cx(t) + F w(t)
(1)
where x, x0 ∈ Rn , w ∈ Rq and z ∈ Rq are the state of the system, the initial condition, the input and the
output, respectively. When x0 = 0, the above system defines a linear time-invariant convolution operator
Σ : w 7→ z given by
Z
t
z(t) =
h(s)w(t − s)ds
(2)
0
where h(t) = CeAt E + F δ(t) where δ(t) is the Dirac distribution and whose transfer function is given by
b
Σ(s)
:= C(sI − A)−1 E + F.
(3)
We then have the following proposition [1]:
Proposition 1 The following statements are equivalent:
(i) The system (1) is (internally) positive; i.e. for any x0 ≥ 0 and any w(t) ≥ 0, we have that x(t) ≥ 0
and z(t) ≥ 0 for all t ≥ 0.
(ii) The matrix A is Metzler (i.e. all the off-diagonal elements are nonnegative) and the matrices E, C, F
are nonnegative (i.e. all the entries are nonnegative).
2.2
Norms and gains
Let us start with the definition of the Lp -norms for signals [55]:
Definition 2 Let w : R≥0 → Rn , then its Lp -norm is given by
Z ∞
1/p
p
||w(t)||p dt
when p ∈ Z>0
.
||w||Lp :=
0
ess sup ||w(t)||∞
when p = ∞
(4)
t≥0
We say that w ∈ Lp if ||w||Lp is finite.
The Lp -gain of the convolution operator (2) (or equivalently of the linear system (1) with x0 = 0) defined
as
||Σ||Lp −Lp :=
sup
||Σw||Lp ,
(5)
||w||Lp =1
is finite if and only if A is Hurwitz stable. In particular, when the system (1) is positive, then we have that
b
||Σ||Lp −Lp = ||Σ(0)||
p
(6)
for any p ∈ {1, 2, ∞}. Note that it is often considered that inputs need to be nonnegative. However, it
is immediate to see that for positive systems, the worst-case inputs are necessarily nonnegative since the
impulse response is nonnegative as well. Therefore, imposing this restriction is not necessary when defining
the Lp -gain of a positive system. Also, it is interesting to note that the same definition also holds for
externally positive systems, those systems for which the impulse response h(t) is nonnegative at all times
but which are not internally positive.
We finally have the following result that is due to [56]:
3
q×q
Proposition 3 Let M ∈ R≥0
. Then, for all p ∈ {1, 2, ∞}, we have that
ρ(M ) = infq ||DM D−1 ||p
(7)
D∈D≻0
and the infimum is a minimum whenever M is irreducible.
3
Exact stability results for uncertain linear positive systems and
interconnections of positive systems
The aim of this section is to recall important results regarding the stability of uncertain linear positive
systems and the stability of interconnections of linear positive systems. Both theoretical and computational
results are provided, the latter being stated in terms of linear or semidefinite programs.
3.1
Stability conditions for uncertain systems in LFT form
We are interested here in the stability of the following uncertain systems in linear fractional form
ẋ(t) = Ax(t) + Ew(t)
z(t) = Cx(t) + F w(t)
p
w(t) = ∆z(t), ∆ ∈ B∆
where
p
:=
B∆
with p ∈ {1, 2, ∞} and q =
of constant D-scalings:
PN
∆ ∈ Cq×q
i=1 qi .
D∆ :=
(8)
∆ = diag(∆1 , . . . , ∆N ), ||∆i ||p ≤ 1
∆i ∈ Cqi ×qi , i = 1, . . . , N
(9)
Associated with this uncertainty structure, we define the following set
D ∈ Rq×q
D = diag(d1 Iq1 , . . . , dN IqN )
di > 0, i = 1, . . . , N
.
(10)
The role of the scalings is to capture the structure of the uncertainty set through the property that ∆D = D∆
p
for all ∆ ∈ B∆
and all D ∈ D∆ . Such scalings allow us to reduce the conservatism of the small-gain theorem
and, in some certain cases, make the conservatism vanish completely. This latter effect will happen in the
context of linear positive systems and will allow us to derive nonconservative stability results.
3.1.1
General theoretical result
We have the following result:
Proposition 4 Assume that (A, E, C, F ) is internally positive. Then, the following statements are equivalent:
p
(i) The uncertain system (8) is asymptotically stable for all ∆ ∈ B∆
.
(ii) A is Hurwitz stable and
−1
b
inf ||DΣ(0)D
||p < 1.
D∈D∆
(11)
Moreover, in the repeated scalar uncertainties (i.e. ∆i = δi Iqi , δi ∈ R>0 , i = 1, . . . , N ), then the above
statements are equivalent to
(iii) A is Hurwitz stable and
b
ρ(Σ(0))
< 1.
4
(12)
Proof : The equivalence (i) ⇔ (ii) has been proved in [23] in the case p = 2. The case p = ∞ has been
proved, for instance, in [28, 29]. Finally, the case p = 1 is dual to the case p = ∞. The equivalence (ii) ⇔
(iii) follows from Proposition 3.
♦
Interestingly, in the case p = 2, the internal stability condition on the system can be relaxed into the
condition of positive domination [57, 23]. When p = 1 or p = ∞, the internal positivity condition can be
substituted by an external positivity condition together with an assumption on the initial condition in order
to preserve the positivity of the output (i.e. x0 must be such that CeAt x0 ≥ 0 for all t ≥ 0). Finally,
since eventually positive systems can be used to efficiently represent externally positive systems, some of the
results for internally positive systems are expected to remain true for these systems as well; see e.g. [58, 59].
3.1.2
L1 scaled small-gain theorem
The following result can be seen as an extension of the L1 results in [19, 20, 18]:
Theorem 5 Assume that (A, E, C, F ) is internally positive. Then, the following statements are equivalent:
1
(i) The uncertain system (8) is asymptotically stable for all ∆ ∈ B∆
.
(ii) There exist a positive vector λ ∈ Rn>0 and a matrix D ∈ D∆ such that
λ
D1q
T
A
C
E
< 0.
F − Iq
(13)
−1
b
Proof : Following [18], we have that ||DΣ(0)D
||1 < 1 if and only if there here exists a positive vector
n
λ ∈ R>0 such that the inequalities
T
λ ED
−1
λT A + 1Tq DC
+ 1 DF D−1 − 1Tq
T
<
<
0
0
hold. Right-multiplying the second inequality by D ∈ Dn≻0 yields the result.
3.1.3
(14)
♦
L2 scaled small-gain theorem
The following result, proved in [23], is the positive version of the well-known L2 scaled small-gain theorem
[27, 60] and is based on the Kalman-Yakubovich-Popov (KYP) Lemma for positive systems [61, 22]:
Theorem 6 Assume that (A, E, C, F ) is internally positive. Then, the following statements are equivalent:
2
(i) The uncertain system (8) is asymptotically stable for all ∆ ∈ B∆
.
(ii) There exist matrices P ∈ Dn≻0 and D ∈ D∆ such that
T
A P + PA PE
⋆
−D
⋆
⋆
CT D
F T D ≺ 0.
−D
(15)
Alternative formulations can also be obtained on the basis of the linear KYP lemma for positive systems
proved in [21]:
Theorem 7 Assume that (A, E, C, F ) is internally positive. Then, the following statements are equivalent:
2
(i) The uncertain system (8) is asymptotically stable for all ∆ ∈ B∆
.
5
(ii) There exist λ, µ ∈ Rn>0 , ν ∈ Rq>0 and D ∈ D∆ such that Aλ + Eν < 0 and
T T
λ
C DC
ν
F T DC
hold.
C T DF
+ µT A
F T DF − D
E <0
(16)
−1/2
b
Proof : Applying the linear version of the KYP Lemma from [21] on the scaled system D1/2 Σ(s)D
−1/2
where D ∈ D∆ yields the conditions Aλ + ED
ν̃ < 0 and
λ
C T DC
C T DF D−1/2
(17)
+ µT A
ED−1/2 < 0
T
−1/2
T
−1/2
T
−1/2
ν̃
D
F DC
D
F DF D
−I
for some positive vectors λ, ν and µ. Note that these conditions are equivalent to saying that the LMI
condition in Theorem 6 holds (possibly with a different matrix D). The final result is then obtained by
letting ν = D−1/2 ν̃ and by multiplying the above inequality from the right by the matrix diag(I, D1/2 ). ♦
Unfortunately, the condition (16) is not convex because of the product between λ, ν and D. In this regard,
this condition may not be very convenient to work with for establishing the stability of the uncertain system
2
(8) with ∆ ∈ B∆
.
Finally, it is also interesting to mention the following novel result based on a result in [62]:
Theorem 8 Assume that (A, E, C, F ) is internally positive. Then, the following statements are equivalent:
2
(i) The uncertain system (8) is asymptotically stable for all ∆ ∈ B∆
.
n×q
(ii) There exist ζ ∈ R>0
and D ∈ D∆ such that ζ T A + DC < 0 and
−D
ζ T E + DF
≺0
⋆
−D
(18)
hold.
Proof : To prove this one, we use a result of [62] which exactly characterizes the L2 -gain of a linear positive
−1/2
b
system. By applying it to the scaled system D1/2 Σ(s)D
, we get the conditions ζ̃ T A + D1/2 E < 0 and
−I
ζ̃ T ED−1/2 + D1/2 F −1/2
≺0
(19)
⋆
−I
n×q
. A congruence transformation with respect to the matrix diag(D1/2 , D1/2 ) and the change
for some ζ̃ ∈ R>0
of variables ζ = ζ̃D1/2 yield the result.
♦
Note that unlike the condition in Theorem 7, the condition in Theorem 8 is still convex but not linear. Its
complexity is also higher than the complexity of the condition in Theorem 6.
3.1.4
L∞ scaled small-gain theorem
The following result is the “L∞ counterpart” of Theorem 6 which can also be seen as an extension of the
results in [18] and a version of the scaled small-gain theorem in the L∞ -sense:
Theorem 9 Assume that (A, E, C, F ) is internally positive. Then, the following statements are equivalent:
∞
(i) The uncertain system (8) is asymptotically stable for all ∆ ∈ B∆
.
(ii) There exist a positive vector λ ∈ Rn>0 and a matrix D ∈ D∆ such that
λ
A
E
< 0.
C F − Iq D1q
6
(20)
b
Proof : Following [18], we have that ||D−1 Σ(0)D||
∞ < 1 if and only if there here exists a positive vector
n
λ ∈ R>0 such that the inequalities
Aλ + ED1q
D−1 Cλ + D−1 F D1q − 1q
< 0
< 0
hold. Left-multiplying the second inequality by D yields the result.
3.2
(21)
♦
Stability of interconnections using Integral Linear Constraints
In the current setup, we are interested in the analysis of interconnections of the form
u2
u1
=
=
G1 u1 + d2
G2 u2 + d1
(22)
where G1 : L1 7→ L1 and G2 : L1 7→ L1 are bounded linear positive time-invariant operators with transfer
b1 (s) and G
b 2 (s). Note that since the operators are positive, then we have that G
b1 (0) ≥ 0 and
functions G
b 2 (0) ≥ 0. The signals u1 , u2 are the loop signals and d1 , d2 are the exogenous signals which are all assumed to
G
have dimensions that are compatible with the operators G1 and G2 .The next result is a simplified, specialized
and extended version of the ones in [24] where ILC/separation types results have been formulated. Note
that the statement (iv) has also been reported in [20] whereas the statement (v) seems novel.
Theorem 10 The following statements are equivalent:
(i) The interconnection (22) is well-posed, positive and stable1 .
b1 (0)G
b 2 (0)) < 1.
(ii) We have that ρ(G
p
(iii) There exist some vectors π1 ∈ Rm
>0 and π2 ∈ R such that the conditions
hold.
T
b 2 (0) ≥ 0 and π T G
b
π1T + π2T G
1 1 (0) + π2 < 0
(23)
Moreover, when the internally positive systems G1 and G2 can be represented in terms of the rational transfer
bi (s) = Ci (sI − Ai )−1 Ei + Fi with Ai ∈ Rni ×ni , Ai Metzler, Ci ∈ Rsi ×ni , Ei ∈ Rni ×ri and
functions G
≥0
≥0
si ×ri
Fi ∈ R≥0 , i = 1, 2, s1 = r2 , s2 = r1 , then the above statements are also equivalent to
1
2
(iv) There exist some vectors λi ∈ Rni , µ1 ∈ Rs>0
and µ2 ∈ Rs>0
such that the conditions
λT1 A1 + µT1 C1
+ µT1 F1 − µT2
λT2 A2 + µT2 C2
T
λ2 E2 + µT2 F2 − µT1
λT1 E1
<
<
<
<
0
0
0
0
(24)
hold.
i
i
(v) There exist some matrices Pi ∈ Dn≻0
and Qi ∈ Ss≻0
, i = 1, 2 such that the conditions
1
T
A1 P1 + P1 A1
⋆
⋆
P1 E1
−Q2
⋆
For some additional details about these concepts see [24]
7
C1T Q1
F1T Q1 ≺ 0
−Q1
(25)
and
T
A2 P2 + P2 A2
⋆
⋆
hold.
P2 E2
−Q1
⋆
C2T Q2
F2T Q2 ≺ 0
−Q2
(26)
Proof : The proof that (i) and (ii) are equivalent follows from [24]. The equivalence between (ii) and (iii)
b1 (0)G
b 2 (0)) < 1 if and only if there exists
follows from standard algebraic manipulations and the fact that ρ(G
T b
b
a positive vector λ of compatible dimensions such that λ (G1 (0)G2 (0) − I) < 0. The equivalence between
(iii) and (iv) follows from standard algebraic manipulations together with the change of variables µ1 = π1
and µ2 = −π2 . This statement has also been proven in [20]. To prove the equivalence between (ii) and (v),
first note that from [21], the LMIs are equivalent to saying that
and
b 1 (0) T Q1
G
0
I
0
−Q2
b2 (0) T Q2
G
0
I
0
−Q1
b 1 (0)
G
b 1 (0) − Q2 ≺ 0
b1 (0)T Q1 G
=G
I
b2 (0)
G
b2 (0) − Q1 ≺ 0.
b2 (0)T Q2 G
=G
I
(27)
(28)
b 2 (0)G
b1 (0)−
b1 (0)G
b 2 (0)−Q1 ≺ 0 or G
b1 (0)T G
b 2 (0)T Q2 G
b 2 (0)T G
b 1 (0)T Q1 G
These two inequalities together imply that G
Q2 ≺ 0, which are equivalent to saying that (ii) holds. The converse can be proven by first noting that (ii)
1
b2 (0)T G
b1 (0)T RG
b 1 (0)G
b2 (0) − R ≺ 0.
is equivalent to saying that there exists a matrix R ∈ Rs≻0
such that G
s1
T
T
b 2 (0) G
b 1 (0) (R + S)G
b1 (0)G
b 2 (0) − R ≺
This implies the existence of a sufficiently small S ∈ S≻0 such that G
T
T
b
b
b
b
0. Letting now Q2 = G1 (0) (R + S)G1 (0) and R = Q1 , we get that G2 (0) Q2 G2 (0) − Q1 ≺ 0 and
b 1 (0)−Q2 = −G
b 1 (0)T S G
b 1 (0) 0. In order to prove that opening the inequality sign is not restricb 1 (0)T Q1 G
G
b 1 (0) − Q2 0 then we can always
b 2 (0) − Q1 ≺ 0 and G
b1 (0)T Q1 G
b2 (0)T Q2 G
tive, it is enough to note that if G
T
b 1 (0) − Q2 ≺ 0.
b
b 1 (0)T Q1 G
b
positively perturb the value Q2 so that we have G2 (0) Q2 G2 (0) − Q1 ≺ 0 and G
The proof is completed.
♦
Note that the difference with the results in [19, 19, 20, 23] is that the above result may deal with more
general systems than LTI systems with state-space realization but can deal with any type of linear positive
bounded operators such as bounded integral operators. This will be useful for dealing with systems with
distributed delays. It is also interesting to mention that the above result is not a small-gain result but a
separation result [63, 64, 20, 24]. Note, finally, that even though the matrices Q1 and Q2 are defined to
be general symmetric matrices, one can be chosen to be diagonal without losing the necessity of the result.
However, it is unclear whether they can be both chosen as diagonal without losing necessity.
Remark 11 It is interesting to see that the two last inequalities in (24) can be substituted by µT2 G2 (0)−µT1 <
0 while the LMI (26) can be substituted by G2 (0)T Q2 G2 (0) − Q1 ≺ 0. This will be useful when analyzing
systems with distributed delays in Section 7.
4
4.1
Stability and performance of linear positive systems with discretedelays
Stability analysis – the constant delay case
We start with the following result [38, 46]:
8
Proposition 12 Let p ∈ {1, 2, ∞}. Then, the time-varying delay operator
Dh
: Lp
w(t)
7→ Lp ,
7→ w(t − h), h ≥ 0
(29)
has unitary L1 -, L2 - and L∞ -gains.
Note that by virtue of the Riesz-Thorin interpolation theorem [65], we immediately get that all the Lp -gains,
for p = 1, 2, . . . , ∞, of the above delay operator are equal to 1.
Proposition 13 The linear time-delay system
ẋ(t) = A0 x(t) +
N
X
Ai x(t − hi )
(30)
i=1
coincides with the uncertain system (8) where A = A0 , E = [A1 . . . AN ], C = 1N ⊗ In , F = 0 and
N
−shi
∆ ∈ ∆d := diag(e
In ) : h ≥ 0, ℜ(s) ≥ 0 .
i=1
Proof : The proof follows from direct substitutions.
♦
It is known that the system (30) is positive if and only if A0 is Metzler and Ai is nonnegative for all
i = 1, . . . , N ; see e.g. [30]. We can now state the main result that unifies the results in [30, 18, 51]:
Theorem 14 Assume that the system (30) is positive. Then, the following statements are equivalent:
(i) The system (30) is asymptotically stable.
PN
(ii) The matrix i=0 Ai is Hurwitz stable.
(iii) There exists a vector v ∈ Rn>0 such that v T
P
N
i=0
Ai < 0.
(iv) There exist matrices P, Qi ∈ Dn≻0 , i = 1, . . . , N , such that the Riccati inequality
AT0 P + P A0 +
N
X
T
Qi + P Ai Q−1
i Ai P ≺ 0
i=1
(31)
holds.
P
−1
(v) A0 is Hurwitz stable and −( N
i=1 Ai )A0 is Schur stable.
Proof : The equivalence between the two first statements follows from the application of Theorem 9 on the
system (30). Indeed, by applying the linear programming conditions of Theorem 9 we get that the system
n
is asymptotically stable if and only there exist vectors λ ∈ Rn>0 and ν ∈ RN
>0 such that A0 λ + Eν < 0 and
PN
1N ⊗ λ − ν < 0. This is equivalent to say that we have A0 λ + E(1N ⊗ λ) = ( i=0 Ai )λ < 0, which is equivalent to the first statement of the result. The equivalence between the statements (ii) and (iii) follows from
standard results on the stability of Metzler matrices whereas the equivalence between (i) and (iv) follows
qi
from Theorem 6 where we have set D = diagN
i=1 (Qi ) where Qi ∈ D≻0 . Note that in this case, the uncertainty
structure is diagonal and hence D∆ is the set of diagonal matrices with positive diagonal entries. Finally,
statement (v) is obtained from Proposition 3.
♦
It is interesting to provide few remarks regarding the above result. First of all, we recover the property
that the stability of a linear positive system with constant and discrete-delays does not depend on the delay
values and hence stability is a delay-independent property. Secondly, the second statement of the above
result provides an answer to a particular version of the problem stated in [66, Problem 1.11] by E. Verriest
9
on the Riccati stability of linear time-delay systems with a single discrete and constant delay. This problem is
about finding conditions on the matrices A0 and A1 (i.e. in the case N = 1) for which there exist P, Q1 ∈ Sn≻0
such that (31) holds. The above result provides a solution to this problem for the particular cases of linear
positive and positively dominated systems with delays. Note that the positive systems case has also been
solved in [51, 32] using different approaches. These results have since been extended to some other classes
of systems in [67].
The advantages of the proposed approach over the previously mentioned ones are its simplicity and
its flexibility. Indeed, while the proofs of these results in the above works involve some very technical
developments, the proposed approach allows to retrieve the same results through a very simple application of
the scaled small-gain theorems. Moreover, the approach can be easily extended to other types of uncertainties,
to performance analysis and to design purposes. Finally, the statement (iv) in the above result is also
interestingly as it generalizes the frequency-sweeping results of [38, Section 2.3] where this condition is
stated as necessary for a linear system with delay to be stable. For linear positive systems, this condition is
also sufficient and this result can be interpreted as a consequence of the fact that the maximum value of the
spectral radius is always attained at the zero frequency.
4.2
Performance analysis – the constant delay case
Let us now consider the system
ẋ(t)
= A0 x(t) +
y(t) = C0 x(t) +
N
X
i=1
N
X
Ai x(t − hi ) + Eu u(t)
(32)
Ci x(t − hi ) + Fu u(t)
i=1
where u ∈ Rnu and y ∈ Rny are the input and the output, respectively. It is known that the above system
is positive if and only if A0 is Metzler and C, Eu , Fu , Ai , Ci are nonnegative for all i = 1, . . . , N . We then
have the following result:
Theorem 15 Assume that the system (32) is positive. Then, the following statements hold:
(i) The system (32) is asymptotically stable and has an L1 -gain smaller than γ1 if and only if there exists
a vector λ ∈ Rn>0 such that
#
T
T "PN
0
λ
Ai
Eu
i=0
PN
<
.
(33)
γ1 1nu
1ny
Fu
i=0 Ci
(ii) The system (32) is asymptotically stable and has an L2 -gain smaller than γ2 if and only if there exists
a matrix P, Qi ∈ Dn≻0 , i = 1, . . . , N , such that
N
X
N
row(P Ai )
P Eu
C0T
Qi
AT0 P + P A0 +
i=1
i=1
N
N
T ≺ 0.
(34)
⋆
−
diag
(Q
)
0
col
(C
)
i
i
i=1
i=1
⋆
⋆
−γ2 I
FuT
⋆
⋆
⋆
−γ2 I
(iii) The system (32) is asymptotically stable and has an L∞ -gain smaller than γ∞ if and only if there exists
a vector λ ∈ Rn>0 such that
#
"P
N
λ
0
Ai
Eu
i=0
PN
<
.
(35)
γ∞ 1ny
Fu 1nu
i=0 Ci
10
Proof : The proof of this result is based on the fact that we can also rewrite the performance characterization
problem as a robust stability problem by setting u = M y where M ≥ 0 is a full-block matrix such that
||M ||p = γp−1 (or, equivalently, ||M ||p ≤ γp−1 ). Since for linear positive systems scaled-small gain results are
non-conservative, then the stability of the interconnection is equivalent to the fact that the Lp -gain of the
transfer u 7→ y is at most (or smaller than) γp . Note, however, that the interconnection result applies to
square uncertainties, a condition that is violated when nu 6= ny (recall that the consideration of D scalings
requires the uncertainty to be square). This issue can be easily resolved by suitably augmenting the vector
y or the vector u and appropriately padding the associated matrices with zeros.
We only prove this result in the L1 -gain case, the others are analogous, and we also assume, without loss
of generality, that nu = ny . To this aim, let us consider the interconnection
ẋ
z
y
row(Ai )
=
1N ⊗ In
C0
w
u
Eu x
0
w
u
F
N
A0
i=1
0
N
row(Ci )
u
i=1
= ∆z, ∆ ∈ ∆d
= M y, ||M ||1 ≤ γ1−1 .
(36)
Applying now Theorem 5 on the above system with the extended uncertainty ∆e = diag(∆, M ) with extended
n
scaling De = diag(D, ǫInu ), D ∈ DN
≻0 , ǫ > 0, yields
T
λ̃
µ̃
ǫ1Tny
A0
1N ⊗ In
C0
N
row(Ai )
i=1
0
N
row(Ci )
i=1
T
0
µ̃ .
0
<
ǫγ1 1nu
Fu
Eu
(37)
Solving for µ̃ as in the proof of Theorem 14, dividing everything by ǫ and using the change of variables
λ = λ̃/ǫ yield the result.
♦
4.3
Stability analysis – the time-varying delay case
We extend here the previous results to the case of time-varying discrete-delays. Interestingly, this class of
delays includes, as a special case, scale-delays [68, 46]. A remark will be made in this regard. Note also that
this case is different from the previous one as the gain of the delay operator will be different depending on
the considered norm. This is formalized in the following result [38, 69, 46]:
Proposition 16 Let p ∈ {1, 2, ∞}. Then, the time-varying delay operator
Th
: Lp
w(t)
7→ Lp
7→ w(t − h(t)), h(t) ≥ 0
(38)
has
• an Lp -gain equal to (1 − η)−1/p where ḣ(t) ≤ η < 1, p ∈ Z>0 and
• an L∞ -gain equal to 1.
The above result clearly shows that, unlike for the constant delay-operator, the value of the gain of the timevarying delay operator depends on the considered norm. Note also that while the Lp -gains, p ∈ Z>0 , depend
on the maximum rate of change of the delay, the L∞ -gain does not, a property that makes it appropriate for
the consideration of fast-varying delays. This fact is not surprising since delay operators do not change the
amplitude of the input signal but may change the value of its integral by appropriately distorting time; see
e.g. [42]. Yet, it is possible to determine a finite L2 -gain when the rate of change of the delay may exceed
11
one (see [70]) and an analogous result for the L1 -gain seems to be still missing. Finally, it is interesting to
note that the L∞ -gain is smaller than the others since 1 < (1 − η)−1/p for any p ∈ Z>0 and, in this regard,
this gain may be more interesting to use than the other. This claim will be supported by the main results
of the section but, before proving them, we need to state the following preliminary result:
Proposition 17 The linear delay system
ẋ(t) = A0 x(t) +
N
X
Ai x(t − hi (t))
(39)
i=1
coincides with the uncertain system (8) where A = A0 , E = [A1 . . . AN ], C = 1N ⊗ In , F = 0 and
N
∆ ∈ diag(In ⊗ Thi ) : hi : R≥0 → R≥0 , i = 1, . . . , N .
i=1
We can now state the main result that unifies the results in [30, 31, 18, 34]
Theorem 18 Assume that the system (39) is internally positive and that t − hi (t) → ∞ as t → ∞ for all
i = 1, . . . , N . Then, the following statements are equivalent:
(i) The system (39) is asymptotically stable.
P
(ii) The matrix N
i=0 Ai is Hurwitz stable.
(iii) There exist vectors λ, µi ∈ Rn>0 , i = 1, . . . , N such that
#
"
N
λ
row(Ai ) N
A0
< 0.
i=1
col(µi )
1N ⊗ In
−InN
(40)
i=1
Proof : First note that since the delays are time-varying, then the second statement is necessary for the
stability of the system (39) and, hence, (i) implies (ii). Note also that (ii) and (iii) are equivalent from
Theorem 14. Finally, using Theorem 9 and the fact that the L∞ -gain of the time-varying delay operator is
equal to one, we can conclude that (iii) implies (i), which completes the proof.
♦
Interestingly, we can see that a necessary and sufficient
condition
for
a
linear
positive
time-delay
system
PN
with constant or time-varying discrete-delays is that i=0 Ai be Hurwitz stable. As a consequence, the
existence of diagonal solutions P, Qi ∈ Dn≻0 , i = 1, . . . , N to the Riccati inequality
AT0 P + P A0 +
N
X
i=1
T
Qi + P Ai Q−1
i Ai P ≺ 0
(41)
is also a necessary and sufficient condition for the stability of linear positive systems with time-varying
delays. This result is, however, rather surprising if we take into account the fact that if we were applying
the L2 -scaled small-gain result (i.e. Theorem 6) on the system (39), we would get the following Riccati
inequality
N
X
T
Qi + (1 − ηi )−1 P Ai Q−1
(42)
AT0 P + P A0 +
i Ai P ≺ 0
i=1
where the ηi ’s are such that ḣi (t) ≤ ηi < 1, i = 1, . . . , N , for all t ≥ 0. Therefore, if (42) is feasible, then
so is (41), but the converse is not true in general. In this regard, we would not be able to predict that the
stability of (30) is equivalent to the stability of (39) using Theorem 6, nor even Theorem 5.
More generally, the stability condition for N = 1 obtained in the Lp , p ∈ {1, 2, ∞}, framework is
equivalent to saying that
ρ(−A0−1 A1 ) < (1 − η)1/p
(43)
12
which indicates that the L1 -based result is more conservative than the L2 -based result which is, in turn,
more conservative than the L∞ -based result. Even though this remark seems contradictory with the fact
that scaled small-gain results are nonconservative, it is actually not since we are considering now with
time-varying operators and also because stability in the Lp -sense is analyzed. Indeed, we have the set
of nonnegative continuous functions converging to 0 with finite L1 -norm is strictly included in the set of
nonnegative continuous functions converging to 0 with finite L2 -norm, which is itself strictly included in the
set of nonnegative continuous functions converging to 0 with finite L∞ -norm. The fact that the L∞ -based
result gives the weakest stability condition demonstrates its relevance.
4.4
Performance analysis – the time-varying delay case
Let us now consider the system
ẋ(t)
y(t)
= A0 x(t) +
= C0 x(t) +
N
X
i=1
N
X
Ai x(t − hi (t)) + Eu u(t)
(44)
Ci x(t − hi (t)) + Fu u(t)
i=1
where u ∈ Rnu and y ∈ Rny are the input and the output, respectively. As for system (32), the above system
is internally positive if and only if A0 is Metzler and C, Eu , Fu , Ai , Ci are nonnegative for all i = 1, . . . , N .
We then have the following result:
Theorem 19 Assume that the system (44) is positive. Then, the following statements hold:
(i) Assume that the delays are differentiable and such that ḣi (t) ≤ ηi < 1, i = 1, . . . , N . Then, the system
(44) is asymptotically stable and has an L1 -gain smaller than γ1 if there exists a vector λ ∈ Rn>0 such
that
#
T "
T
PN
A0 + i=1 (1 − ηi )−1 Ai
Eu
λ
0
PN
<
.
(45)
1ny
γ1 1nu
C0 + i=1 (1 − ηi )−1 Ci
Fu
(ii) Assume that the delays are differentiable and such that ḣi (t) ≤ ηi < 1, i = 1, . . . , N . Then, the system
(44) is asymptotically stable and has an L2 -gain smaller than γ2 if there exist matrices P, Qi ∈ Dn≻0 ,
i = 1, . . . , N , such that
N
X
N
T
T
row(P Ai )
P Eu
C0
Qi
A0 P + P A0 +
i=1
i=1
N
N
T ≺ 0.
(46)
⋆
−
diag
((1
−
η
)Q
)
0
col
(C
)
i
i
i
i=1
i=1
⋆
⋆
−γ2 I
FuT
⋆
⋆
⋆
−γ2 I
(iii) Assume that the delays are such that t − hi (t) → ∞ as t → ∞ for all i = 1, . . . , N . Then, the system
(44) is asymptotically stable and has an L∞ -gain smaller than γ∞ if and only if there exists a vector
λ ∈ Rn>0 such that
#
"
P
λ
A0 + N
A
E
0
i
u
i=1
PN
<
.
(47)
γ∞ 1ny
C0 + i=1 Ci
Fu 1nu
Proof : The proof is similar to that of Theorem 15 and is thus omitted.
♦
Remark 20 It is worth mentioning that the condition in statement (i) is necessary and sufficient while it
is unclear, at the moment, whether necessity holds for the conditions in the statements (ii) and (iii).
13
5
Stability and performance of linear positive delay-difference equations
5.1
Stability analysis
We consider in this section, the case of delay-difference equations [71–75, 34] of the form
x(t) =
N
X
Ai x(t − hi )
(48)
i=1
where the delays hi are such that hi > 0 for all i = 1, . . . , N . Clearly, the system is positive if and only
if the matrices Ai are nonnegative for all i = 1, . . . , N . Note that this system can be rewritten as the
interconnection
x(t) = Ew(t)
(49)
z(t) = Cx(t)
where E = A1 . . . An , C = 1N ⊗ In and w = ∆z, ∆ ∈ ∆d . We then have the following result:
Theorem 21 Assume that the system (48) is positive. Then, the following statements are equivalent:
(i) The delay-difference equation is asymptotically stable (or strongly-stable).
!
N
X
Ai e−jωi < 1 for all ωi ∈ R, i = 1, . . . , N .
(ii) ρ
i=1
(iii) ρ
N
X
i=1
Ai
!
< 1.
(iv) There exists a µ ∈ RnN
>0 such that
N
µT −InN + (1N ⊗ In ) row(Ai ) < 0
(50)
i=1
holds.
(v) There exists a µ ∈ RnN
>0 such that
µ
T
N
X
Ai − In
i=1
holds.
!
<0
(51)
(vi) There exist diagonal matrices Qi ∈ Dn≻0 such that the LMI
N
− diag(Qi )
⋆
i=1
≺0
N
N
N
diag(Qi )(1N ⊗ In ) row(Ai ) − diag(Qi )
i=1
i=1
(52)
i=1
holds.
Proof : The equivalence between the two first statements has been proved in [71]. The equivalence between
(iii), (iv) and (v) follows from simple algebraic manipulations and the theory of nonnegative matrices. The
equivalence between (iii) and (iv) follows from the fact that (52) is equivalent to
N
N
N
N
i=1
i=1
i=1
i=1
− diag(Qi ) + ((1N ⊗ In ) row(Ai ))T diag(Qi )(1N ⊗ In ) row(Ai ) ≺ 0
14
(53)
which is, in turn, equivalently, that ρ[(1N ⊗ In ) rowN
i=1 (Ai )] = ρ
P
N
i=1
Ai < 1. The equivalence between
(ii) and (iv) can be proved using a scaled-small gain argument on the system (49).
♦
Remark 22 When the delays are time-varying, the condition in (iv) remains valid as long as the delays
satisfy the condition t − hi (t) → ∞ as t → ∞ for all i = 1, . . . , N . As a result, the stability of the delaydifference equation does not depend on the value of the delays nor on their time-varying nature as long a
L∞ -gain result is considered but will depend on the rate of variation of the delays when L1 -and L2 -gain
results are considered. See the discussion below Theorem 18 for additional details.
5.2
Performance analysis
Let us consider here the following delay-difference equation
x(t)
=
y(t) =
N
X
i=1
N
X
Ai x(t − hi ) + Eu u(t)
(54)
Ci x(t − hi ) + Fu u(t)
i=1
where the delays hi are such that hi > 0 for all i = 1, . . . , N . Clearly, the system is positive if and only if
the matrices Ai , Eu , Ci , Fu are nonnegative for all i = 1, . . . , N . We then have the following result:
Theorem 23 Assume that the system (54) is positive. Then, we have the following results:
(i) The system (54) is asymptotically stable and has an L1 -gain smaller than γ1 if and only if there exists
a µ ∈ RnN
>0 such that the condition
N
N
T
T
(Ai )
col(Eu )
µ
µ
(1N ⊗ In ) row
i=1
i=1
(55)
<
N
1ny
γ1 1nu
row(C )
F
i=1
i
u
holds.
(ii) The system (54) is asymptotically stable and has an L2 -gain smaller than γ2 if and only if there exists
matrices Qi ∈ Dn≻0 such that the condition
N
−
diag
(Q
)
⋆
⋆
⋆
i
i=1
⋆
⋆
0
−γ2 Inu
N
N
(56)
≺0
N
N
col(Qi ) row
(A
)
col
(Q
E
)
−
diag
(Q
)
⋆
i
i
u
i
i=1
i=1
i=1
i=1
N
row(Ci )
Fu
0
−γ2 Iny
i=1
holds.
(iii) The system (54) is asymptotically stable and has an L∞ -gain smaller than γ∞ if and only if there exists
a µ ∈ RnN
>0 such that the condition
N
N
(1
⊗
I
)
row
(A
)
col
(E
)
n
i
u
µ
µ
N
i=1
i=1
<
(57)
N
1nu
γ∞ 1ny
row(Ci )
Fu
i=1
holds.
15
Proof : The proof is based on reformulating the system (54) into an LTI system interconnected with some
delay operators and applying scaled small-gain results.
♦
Remark 24 Interestingly, the L∞ result remains the same when the delays are time-varying and such that
t − hi (t) → ∞ as t → ∞ for all i = 1, . . . , N . As for linear positive systems with discrete-delays, the
conditions for the L1 - and the L2 -gains will be different (i.e. they will depend on the rate of variation of the
delays and become sufficient only).
6
Stability and performance of linear positive coupled differentialdifference equations with delays
We consider here linear positive coupled differential-difference equations with delays which can be seen as
an extension of the systems (30) and (48). Such systems have been, for instance, studied in [76–81] and
in the references therein. In particular, the case of linear positive coupled differential-difference equations
with single time-varying delay has been studied in [52] where necessary and sufficient conditions for their
positivity and stability were obtained. We prove here that these results can be retrieved and extended to
the case of multiple delays and to performance analysis using very simple scaled-small gain arguments.
6.1
Stability analysis
Let us start with the following preliminary result:
Proposition 25 The linear coupled differential-difference system with constant discrete delays
ẋ1 (t) =
A0 x1 (t) +
x2 (t) =
C0 x1 (t) +
N
X
i=1
N
X
Ai x2 (t − hi )
(58)
Ci x2 (t − hi )
i=1
N
N
coincides with the uncertain system (8) where A = A0 , E = row(Ai ), C = 1N ⊗ C0 , F = 1N ⊗ row(Ci ) and
i=1
∆ ∈ ∆d :=
i=1
N
diag(e−shi In ) : h ≥ 0, ℜ(s) ≥ 0 .
i=1
Proof : The proof follows from direct substitutions.
♦
It was proven in [52] that the system (58) is positive if and only if A0 is Metzler and Ai , C0 , Ci are
nonnegative for all i = 1, . . . , N . Note, moreover, that this result is obvious from the description (8) with
the matrices defined in Proposition 25.
We can now state the main result that extends the results in [52] with the difference that constant delays
are considered. Note, however, that the result still holds in the case of time-varying delays in the same way
as in Section 4.3.
Theorem 26 Assume that the system (58) is positive. Then, the following statements are equivalent:
(i) The system (58) is asymptotically stable.
16
(ii) The matrix
N
row(Ai )
A0
M :=
1N ⊗ C0
is Hurwitz stable.
n(N +1)
(iii) There exists a vector v ∈ R>0
i=1
N
1N ⊗ row(Ci ) − InN
i=1
(59)
such that v T M < 0.
n
(iv) There exist some matrices P ∈ Dn≻0 and Q ∈ DN
≻0 such that the generalized Riccati inequality
h N
AT0 P +P A0 + P row(Ai )
i
−1
N
N
T
(1TN ⊗ col(CiT ))Q
col
(A
)P
i=1
≺0
i=1 i
−Q
1N ⊗ C0
−Q
1TN ⊗ C0T
N
Q(1N ⊗ row(Ci ))
i=1
i=1
(60)
holds.
N
(v) The matrices 1N ⊗ row(Ci ) − I and
i=1
are Hurwitz stable.
A0 − A1
A2
N
. . . AN (1N ⊗ row(Ci ) − I)−1 (1N ⊗ C0 )
i=1
(61)
(vi) The matrices A0 and
N
A1
1N ⊗ row(Ci ) − I − (1N ⊗ C0 )A−1
0
i=1
are Hurwitz stable.
(vii) A0 is Hurwitz stable and
PN
−1
i=1 (−C0 A0 Ai
A2
. . . AN
(62)
+ Ci ) is Schur stable.
Proof : The equivalence between the three first statements follows from the application of Theorem 5 on the
input-output formulation of the system (58). The equivalence with the statements (iv) and (vii) comes from
Theorem 6 and the fact that, for two matrices A, B of appropriate dimensions, we have that ρ(AB) = ρ(BA).
The equivalence between (ii), (v) and (vi) follows from Lemma 1 in [20] or Lemma 7.2 in [82].
♦
As for linear systems with discrete-delays, the stability of linear positive coupled differential-difference
equations with delays is equivalent to that one of the same system with all the delays set to 0.
6.2
Performance analysis
Let us now consider the linear coupled differential-difference system with time-varying discrete delays
ẋ1 (t)
x2 (t)
y(t)
= A0 x1 (t) +
= C0 x1 (t) +
= Cy0 x(t) +
N
X
i=1
N
X
i=1
N
X
Ai x2 (t − hi (t)) + E1 u(t)
Ci x2 (t − hi (t)) + E2 u(t)
(63)
Cyi x2 (t − hi (t)) + Fu u(t)
i=1
where u ∈ Rnu and y ∈ Rny are the input and the output, respectively. It is immediate to see that this
system is internally positive if and only if A0 is Metzler and Ai , C0 , Ci , E1 , E2 , Cy0 , Cyi , Fu are nonnegative
for all i = 1, . . . , N . We then have the following result:
17
Theorem 27 Assume that the system (63) is positive. Then, the following statements hold:
(i) Assume that the delays are differentiable and such that ḣi (t) ≤ ηi < 1, i = 1, . . . , N . Then, the system
n
(63) has an L1 -gain smaller than γ1 if there exists some vectors λ ∈ Rn>0 and µ ∈ RN
>0 such that
N
row((1 − ηi )−1 Ai )
E1
A0
T
T
i=1
λ
0
N
N
N
N
µ col (C0 )
µ .
(64)
col(In ) row((1 − ηi )−1 Ci )
col (E2 )
<
i=1
i=1
i=1
i=1
1ny
γ
1
1
n
N
u
row((1 − ηi )−1 Ci )
Fu
Cy0
i=1
(ii) Assume that the delays are differentiable and such that ḣi (t) ≤ ηi < 1, i = 1, . . . , N . Then, the system
(63) has an L2 -gain smaller than γ2 if there exist matrices P, Qi ∈ Dn≻0 , i = 1, . . . , N , such that
T
A0 P
+ P A0 +
N
X
N
row(P Ai )
Qi
⋆
row(C0T )Q
i=1
P Eu
i=1
i=1
N
N
− diag((1 − ηi )Qi )
0
i=1
N
(1N ⊗ row(Ci ))T
i=1
N
⋆
⋆
−γ2 I
row(E2T )
i=1
⋆
⋆
⋆
⋆
⋆
⋆
−Q
⋆
T
Cy0
N
T
col(Cyi
)
i=1
≺ 0.
T
Fu
0
(65)
−γ2 I
(iii) Assume that the delays are such that t − hi (t) → ∞ as t → ∞, i = 1, . . . , N . Then, the system (63)
n
has an L∞ -gain smaller than γ∞ if and only if there exists some vectors λ ∈ Rn>0 and µ ∈ RN
>0 such
that
N
A0
row(Ai )
E1
λ
i=1
0
N
N
N
N
col (C )
µ .
(66)
col(In ) row(Ci )
col(E2 )
µ <
i=1 0
i=1
i=1
i=1
1
γ
1
nu
∞ ny
N
row(Ci )
Fu
Cy0
i=1
As for systems with time-varying discrete delays, the last statement states a necessary and sufficient
condition whereas it is unclear whether the conditions in the two first ones are also necessary. Note also
that in the L1 and L∞ cases, the vector µ can be eliminated from the conditions by explicitly solving it.
However, the benefit of the current conditions is that they are linear in the matrices of the system, thereby
allowing for immediate extensions to uncertain matrices and to design purposes.
7
7.1
Stability and performance of linear positive systems with distributeddelays
Stability analysis – the constant kernel case
We have the following result which be proven using standard manipulations:
Proposition 28 The time-varying distributed-delay operator
Uh
: Lp
w(t)
7→ Lp , p = 2, ∞
Z
1 t
7→
w(θ)dθ, h(t) ∈ [0, h̄]
h̄ t−h(t)
has unitary L1 -, L2 - and L∞ -gains.
18
(67)
Proposition 29 The time-delay system
ẋ(t) = A0 x(t) +
N
X
Ai
i=1
Z
t
x(s)ds
(68)
t−hi (t)
coincides with the uncertain system (8) where A = A0 , E = [A1 . . . AN ], C = 1N ⊗ In , F = 0 and
N
∆ ∈ diag(In ⊗ Uhi ) : hi : R≥0 → [0, h̄i ], i = 1, . . . , N .
i=1
Moreover, it is internally positive if and only if the matrix A0 is Metzler and the matrices Ai , i = 1, . . . , N ,
are nonnegative.
We then have the following result:
Theorem 30 Assume that the system (78) is internally positive. Then, the following statements are equivalent:
(i) The system (78) is asymptotically stable.
P
(ii) A0 + N
i=1 h̄i Ai is Hurwitz stable.
P
(iii) There exists a v ∈ Rn>0 such that v T (A + N
i=1 h̄i Ai ) < 0.
PN
(iv) A is Hurwitz stable and −( i=1 h̄i Ai )A0−1 is Schur stable.
(v) There exist matrices P, Qi ∈ Dn≻0 , i = 1, . . . , N , such that the Riccati inequality
AT0 P + P A0 +
N
X
T
(Qi + h̄2i P Ai Q−1
i Ai P ) ≺ 0
(69)
i=1
holds.
Moreover, when N = 1, then the above statements are also equivalent to
(vi) A0 is Hurwitz stable and h̄1 <
1
.
ρ(A−1
0 A1 )
Proof : This result is proved exactly in the same way as Theorem 14. The last statement can be straightforwardly shown to be equivalent to (iv).
♦
7.2
Performance analysis – the constant kernel case
Let us now consider the system
ẋ(t) =
y(t) =
A0 x(t) +
C0 x(t) +
N
X
i=1
N
X
i=1
Ai
Ci
Z
Z
t
x(s)ds + Eu u(t)
t−hi (t)
(70)
t
x(s)ds + Fu u(t)
t−hi (t)
where u ∈ Rnu and y ∈ Rny are the input and the output, respectively. As for system (32), the above system
is internally positive if and only if A0 is Metzler and C, Eu , Fu , Ai , Ci are nonnegative for all i = 1, . . . , N .
We have the following result:
19
Theorem 31 Assume that the system (70) is internally positive. Then, the following statements hold:
(i) The system (70) is asymptotically stable and has an L1 -gain smaller than γ1 if and only there exists a
vector λ ∈ Rn>0 such that
#
T
T "
PN
Eu
A0 + i=1 h̄i Ai
0
λ
PN
<
.
(71)
γ1 1nu
1ny
C0 + i=1 h̄i Ci
Fu
(ii) The system (70) is asymptotically stable and has an L2 -gain smaller than γ2 if and only if there exist
matrices P, Qi ∈ Dn≻0 , i = 1, . . . , N , such that
AT0 P + P A0 +
⋆
⋆
⋆
N
X
Qi
i=1
N
row(h̄i P Ai )
P Eu
i=1
N
− diag(Qi )
0
⋆
⋆
−γ2 I
⋆
i=1
C0T
N
≺0
col (h̄i CiT )
i=1
FuT
−γ2 I
(72)
(iii) The system (70) is asymptotically stable and has an L∞ -gain smaller than γ∞ if and only if there exists
a vector λ ∈ Rn>0 such that
#
"
PN
Eu
A0 + i=1 h̄i Ai
λ
0
PN
<
.
(73)
γ∞ 1ny
C0 + i=1 h̄i Ci
Fu 1nu
Proof : The proof follows from the same lines as the proof of Theorem 15.
7.3
♦
Stability analysis – the non-constant kernel case
To prove this result, we need to consider the result based on integral linear constraints (Theorem 10) since
results based on gains would be conservative. Let us first consider the following result:
Proposition 32 The distributed-delay operator
VB,h
: Lp
w(t)
7→ Lp
Z 0
7
→
B(θ)w(t + θ)dθ, 0 ≤ h ≤ h̄
(74)
−h
is nonnegative if and only if B(θ) ≥ 0 for all θ ∈ [−h̄, 0], h ∈ [0, h̄]. Moreover, its transfer function is given
by
Z
Vd
B,h (s) =
0
B(θ)esθ dθ
(75)
−h
d
and for any 0 ≤ h ≤ h̄, we have that Vd
B,h (0) ≤ VB,h̄ (0).
Proof : Let z(t) = VB,h (w)(t). To prove the nonnegativity property condition, assume that Bij (s) < 0 for
some s ∈ [−h̄, 0]. Then, it is immediate to see that we can pick a w such that one of the component of the
20
output is negative. Hence, the result follows. The transfer function can be computed as follows
Z ∞
zb(s) =
z(t)e−st dt
Z0 ∞ Z 0
=
B(θ)w(t + θ)e−st dθdt
0
−h
Z ∞
Z 0
=
B(θ)
w(t + θ)e−st dt dθ
−h
0
Z
0
sθ
=
B(θ)e dθ w(s)
b
(76)
−h
=
[
V
b
B,h (s)w(s).
♦
Remark 33 Note that the delay can be made infinite as in [83] as long as B(θ) is integrable on (−∞, 0].
Proposition 34 The linear system with time-varying distributed-delays
ẋ(t) = A0 x(t) +
N Z
X
i=1
0
Ai (θ)x(t + θ)dθ
(77)
−hi (t)
with hi (t) ∈ [0, h̄i ], i = 1, . . . , N , is positive if and only if A0 is Metzler and Ai (θ) ≥ 0 for all θi ∈ [−h̄i , 0]
and all i = 1, . . . , N .
The following result demonstrates that the stability of linear positive systems with distributed delays
does not depend on the nature of the delay (i.e. whether it is time-varying or time-invariant) and only
depends on the delay upper-bound [50]:
Theorem 35 Assume that the system (77) is positive. Then, the statements are equivalent:
(i) The linear positive system (77) with time-varying distributed-delays is asymptotically stable for any
hi : R≥0 7→ [0, h̄i ].
(ii) The linear positive system with constant distributed-delays
Z 0
˙
Ai (θ)x̄(t + θ)dθ
x̄(t)
= A0 x̄(t) +
(78)
−h̄i
is asymptotically stable.
Proof :
The proof can be found in [50] and is thus omitted here.
♦
Proposition 36 The time-delay system (78) coincides with the interconnection (22) with
b 1 (s)
G
where E = 1TN ⊗ In and C = 1N ⊗ In .
b 2 (s)
G
= C(sI − A0 )−1 E
N
= diag(V[
Ai ,h (s))
i=1
We then have the following result:
R0
Theorem 37 Define Āi := −h̄i Ai (θ)dθ. Then, the following statements are equivalent:
21
(79)
(i) The system (78) is asymptotically stable.
PN
(ii) A0 + i=1 Āi is Hurwitz stable.
PN
(iii) There exists a v ∈ Rn>0 such that v T (A0 + i=1 Āi ) < 0.
PN
(iv) A0 is Hurwitz stable and −( i=1 Āi )A−1
0 is Schur stable.
(v) There exist matrices P, Qi ∈ Dn≻0 , i = 1, . . . , N , such that the (diagonal) Riccati inequality
AT0 P + P A0 +
N
X
T
(Qi + P Āi Q−1
i Āi P ) ≺ 0
(80)
i=1
holds.
Proof : Applying Theorem 10 yields that the interconnection is well-posed, positive and stable if and only
b1 (0)G
b 2 (0)) < 1 or, equivalently, if and only if ρ(G
b 2 (0)G
b1 (0)) < 1. Expanding the product yields
if ρ(G
b2 (0)G
b 1 (0) = −(1N ⊗ In )A−1 Ā1
G
0
. . . ĀN
(81)
and hence the spectral radius condition is equivalent to saying that the above nonnegative matrix is Schur
stable. Since ρ(AB) = ρ(BA) for any matrices A, B of compatible dimensions, then we get the result of
statement (iv). By exploiting the similarity of the conditions with those of Theorem 14, the other statements
directly follow.
♦
The last statement in the above result is the distributed-delay analogue of the Riccati inequality for
discrete delays and does not seem to have been reported elsewhere in the literature. It is also interesting
to point out that the results would have been completely different if scaled small-gain theorems would have
been considered. Indeed, such results would have considered the Lp -gain of the operator (74) which are
given by ||V[
Ai ,h (0)||p for p ∈ {1, 2, ∞}. It is clear that, in such a case, the obtained results would have been
conservative.
7.4
Performance analysis – the non-constant kernel case
Let us consider here is the following system with distributed delays
ẋ(t) =
y(t) =
A0 x(t) +
C0 x(t) +
N Z
X
0
Ai (θ)x(t + θ)dθ + Eu u(t)
i=1 −h̄i
N Z 0
X
i=1
(82)
Ci (θ)x(t + θ)dθ + Fu u(t)
−h̄i
where u ∈ Rnu and y ∈ Rny are the input and the output, respectively. As for system (77), the above system
is internally positive if and only if A0 is Metzler, C, Eu , Fu are nonnegative for all i = 1, . . . , N and the
functions Ai (·), Ci (·) are nonnegative on their domain.
Theorem 38 Assume that the system (82) is internally positive and let us define the matrices
Āi :=
Z
0
Ai (s)ds and C̄i :=
Z
0
−h̄i
−h̄i
Then, the following statements hold:
22
Ci (s)ds.
(83)
(i) The system (82) is asymptotically stable and has an L1 -gain smaller than γ1 if and only there exists a
vector λ ∈ Rn>0 such that
#
T
T "
PN
0
A0 + i=1 Āi
Eu
λ
PN
<
.
(84)
γ1 1nu
1ny
Fu
C0 + i=1 C̄i
(ii) The system (82) is asymptotically stable and has an L2 -gain smaller than γ2 if and only if there exist
ny
matrices P, Qi ∈ Dn≻0 , Ri ∈ D≻0
, i = 1, . . . , N , such that
N
N
X
X
T
T
T
T
T
1N ⊗ P
0
P Eu
C0
C̄i Ri C̄i
Āi Qi Āi +
A0 P + P A0 +
i=1
i=1
N
⋆
− diag(Qi )
0
0
0
i=1
≺ 0.
N
⋆
⋆
− diag(Ri )
0
1N ⊗ Iny
i=1
FuT
⋆
⋆
⋆
−γ2 Inu
⋆
⋆
⋆
⋆
−γ2 Iny
(85)
(iii) The system (82) is asymptotically stable and has an L∞ -gain smaller than γ∞ if and only if there exists
a vector λ ∈ Rn>0 such that
#
"
PN
λ
A0 + i=1 Āi
Eu
0
P
<
.
(86)
γ∞ 1ny
C0 + N
Fu 1nu
i=1 C̄i
Proof : Proof of the statement (i). From the Theorem 10 and Remark 11, this is equivalent to saying
N n+N ny
that there exist λ ∈ Rn>0 , µ1 , µ2 ∈ R>0
such that
T
T
1TN ⊗ In
0
Eu
A0
λ
0
1N ⊗ In
0
0
0
< µ2
µ1
(87)
1N ⊗ Iny
0
0
0
1ny
γ
1
1
n
u
C0
0
1TN ⊗ Iny
Fu
N
N
T
T
together with µ2 D − µ1 < 0 where D = diag diag Āi , diag C̄i . This is equivalent to the inequalities
i=1
λT 1TN ⊗ In
i=1
1N ⊗ In
+ 1Tny C0
λT A0 + µT1
1
⊗
I
N
n
y
0n×N ny + 1Tny 0ny ×N n 1TN ⊗ Iny − µT2
λT Eu + 1Tny Fu − γ1 1Tnu
< 0
< 0
< 0.
Using the fact that µT2 D − µT1 < 0, we get that the following equivalent condition
N
col(Āi )
T
λT A0 + µT2 i=1
+ 1ny C0 < 0
N
col (C̄i )
(88)
(89)
i=1
and using the second inequality we can eliminate µ2 to get the equivalent inequalities
PN
PN
λT (A0 + i=1 Āi ) + 1Tny (C0 + i=1 C̄i ) < 0
λT Eu + 1Tny Fu − γ1 1Tnu < 0.
23
(90)
The proof is completed.
Proof of the statement (ii). From the Theorem 10 and Remark 11, this is equivalent
T
A0 P + P A0 1TN ⊗ P
0
P Eu
(1T2N ⊗ In )Z1
C0T
⋆
−Z21
−Z22
0
0
0
3
⋆
⋆
−Z2
0
0
1N ⊗ Iny
≺0
0
FuT
⋆
⋆
⋆
−γ2 Inu
⋆
⋆
⋆
⋆
−Z1
0
⋆
⋆
⋆
⋆
⋆
−γ2 Iny
(91)
N
N
and DT Z2 D − Z1 ≺ 0 where D = diag diag Āi , diag C̄i . Performing a Schur complement and comi=1
i=1
bining these inequalities yields
T
A0 P + P A0 + (1T2N ⊗ In )DT Z2 D(12N ⊗ In ) 1TN ⊗ P
⋆
−Z21
⋆
⋆
⋆
⋆
⋆
⋆
0
−Z22
−Z23
⋆
⋆
P Eu
0
0
−γ2 Inu
⋆
C0T
0
1N ⊗ Iny
≺ 0.
FuT
−γ2 Iny
N
N
i=1
i=1
(92)
Since the system is positive, then we can restrict ourselves to a diagonal Z2 = diag(diag(Qi ), diag(Ri )),
Qi Ri , ∈ Dn≻0 , and hence
N
X
N
X
C̄iT Ri C̄i
ĀTi Qi Āi +
AT0 P + P A0 +
i=1
i=1
⋆
⋆
⋆
⋆
1TN ⊗ P
N
− diag(Qi )
i=1
⋆
0
P Eu
0
0
N
− diag(Ri )
0
⋆
⋆
−γ2 Inu
⋆
i=1
⋆
⋆
C0T
0
≺ 0.
1N ⊗ Iny
T
Fu
−γ2 Iny
Proof of the statement (iii). It is similar to the proof of the statement (i) and it is thus omitted.
8
8.1
(93)
♦
Stability and performance of neutral linear positive systems
Stability analysis
Neutral systems have been extensively studied as they arise, for instance, transmission lines, models of
population dynamics, etc. [84, 85, 76, 36, 86, 87]. The special case of linear positive neutral systems has
been considered in [53, 54] in the single constant delay case. We extend here these stability analysis results
to the case of multiple delays, possibly time-varying, and to performance analysis.
Let us start with the following result:
Proposition 39 The time-delay system
ẋ(t) = A0 x(t) +
N
X
Ar,i x(t − hi ) +
i=1
N
X
i=1
24
An,i ẋ(t − hi )
(94)
N
coincides with the uncertain system (8) where A = A0 , E = [In . . . In ], C = col (An,i A0 + Ar,i ) and
i=1
F =
N
col(1TN
i=1
⊗ An,i ) and
∆ ∈ ∆d :=
N
diag(e
−shi
i=1
In ) : h ≥ 0, ℜ(s) ≥ 0 .
It has been proven in [53, 54] that the system (94) is positive if and only if A0 is Metzler and the matrices
An,i A0 + Ar,i and An,i are nonnegative for all i = 1, . . . , N . We then have the following result:
Theorem 40 Assume that the system (94) is positive. Then, the following statements are equivalent:
(i) The system (94) is asymptotically stable.
PN
(ii) The nonnegative matrix i=1 An,i is Schur stable and the Metzler matrix
In −
N
X
An,i
i=1
!−1
A0 +
N
X
Ar,i
i=1
!
(95)
is Hurwitz stable.
(iii) There exist matrices P, Qi ∈ Dn≻0 , i = 1, . . . , N , such that the LMI
AT0 P + P A0
⋆
⋆
N
N
col(P )
− diag(Qi )
⋆
≺0
i=1
i=1
N
N
N
col (Qi (An,i A0 + Ar,i )) col(Qi (1TN ⊗ An,i )) − diag(Qi )
i=1
i=1
(96)
i=1
holds.
(iv) A0 is Hurwitz stable and the matrices
N
X
An,i
and
−
N
X
i=1
i=1
Ar,i
!
A−1
0
(97)
are Schur stable.
Proof : Proof that (i) is equivalent to (ii). To prove the equivalence between (i) and (ii), first note
that (i) is equivalent to saying that there exist vectors λ ∈ Rn>0 , µi ∈ Rn>0 , i = 1, . . . , N such that
A0
λ
rowN
i=1 (In )
<0
(98)
T
colN
colN
colN
i=1 (An,i A0 + Ar,i )
i=1 (1N ⊗ An,i ) − IN n
i=1 (µi )
N
or, equivalently, that the above Metzler matrix is Hurwitz stable. Let M12 = rowN
i=1 (In ), M21 = col (An,i A0 +
i=1
T
Ar,i ) and M22 = colN
i=1 (1N ⊗ An,i ) − IN n . Then, the stability of the above matrix is equivalent to the
−1
stability of the matrix M22 together with the stability of the matrix A0 − M12 M22
M21 . Clearly, the Metzler
N
T
matrix M22 is Hurwitz stable if and only if the nonnegative matrix coli=1 (1N ⊗ An,i ) is Schur stable. Since
T
colN
i=1 (An,i )(1N × In ) and for two matrices Z1 , Z2 , we have that ρ(Z1 Z2 ) = ρ(Z2 Z1 ), then we get that M22
PN
is Hurwitz stable if and only if ρ( i=1 An,i ) < 1. This proves the first part. To prove the second part, we
−1
need to evaluate A0 − M12 M22
M21 . Using the Sherman-Morrison formula, we get that
!−1
N
X
N
N
−1
row(In ) .
An,i
M22
= − IN n + col(An,i ) In −
(99)
i=1
i=1
25
i=1
−1
Hence, we have that A0 − M12 M22
M21 is equal to
!−1 N
!
N
N
X
X
X
(An,i A0 + Ar,i )
An,i
In −
A0 + In +
An,i
which then simplifies to
A0 +
In −
N
X
An,i
i=1
and to
In −
N
X
An,i
i=1
!−1
!−1
(100)
i=1
i=1
i=1
N
X
(An,i A0 + Ar,i )
(101)
i=1
A0 +
N
X
Ar,i
i=1
!
.
(102)
The proof of the equivalence is completed.
Proof that (i) is equivalent to (iii). This follows from the bounded real lemma.
Proof that (i) is equivalent to (iv). Using the fact that (i) is equivalent to saying that ρ(F ) < 1 and
P
ρ(−CA−1 E + F ) < 1, we get that (i) is equivalent to saying that A0 is Hurwitz stable, that ρ( N
i=1 An,i ) < 1
and that
An,1 A0 + Ar,1
N
−1
..
I
.
.
.
I
+
col
(1TN ⊗ An,i )
(103)
−
A
n
n
0
.
i=1
An,N A0 + Ar,N
is Schur stable. Expanding it yields
Ar,1
In
− ... A−1
0
. . . In
Ar,N
and the proof is completed.
(104)
♦
Interestingly, we can see that, once again, the magnitude of the delays does not affect the stability of
the process and that the stability of the system can be inferred from the stability of the system with all
the delays set to zero. Another interesting point is regarding the concept of strong stability of a difference
equation (see Theorem 21 and [84, 85]). A difference equation of the form
x(t) =
N
X
Mi x(t − hi )
(105)
i=1
is said to be strongly stable if and only if
max
ω∈[0,2π]N
ρ
N
X
Mk e
k=1
−iωk
!
< 1.
(106)
The notion of strong stability has been introduced in [84, 85] for the analysis of neutral delay systems as the
strong stability of the delay-difference equation acting on the derivative of the state is a necessary condition
for the stability of overall neutral delay system and the robustness with respect to arbitrarily small changes
in the values of the delays. In the present case, we have that
!
!
N
N
X
X
−iωk
max ρ
(107)
An,k e
=ρ
An,k
ω∈[0,2π]N
k=1
k=1
26
since the matrices Ank are nonnegative and hence the maximum is attained at θk = 0. Hence, we can
see that the condition of strong stability is encoded in the condition in terms of the well-posedness of the
interconnection of system (8) with the matrices and operators described in Proposition 39.
Finally, it seems interesting to mention that the conditions of the theorem remains true for the stability
of neutral systems with time-varying delays provided that t − hi (t) → ∞ as t → ∞, i = 1, . . . , N .
8.2
Performance analysis
Let us address now the performance analysis of neutral systems. Let us start with the following result:
Proposition 41 The time-delay system
ẋ(t)
=
z(t) =
A0 x(t) +
C0 x(t) +
N
X
i=1
N
X
Ar,i x(t − hi ) +
Cr,i x(t − hi ) +
N
X
i=1
N
X
An,i ẋ(t − hi ) + Eu u(t)
(108)
Cn,i ẋ(t − hi ) + Fu u(t)
i=1
i=1
coincides with the uncertain system (8) where A = A0 , E = 1TN ⊗ In
N
N
col (1T ⊗ An,i )
col
(A
A
+
A
)
r,i
i=1 N
i=1 n,i 0
C =N
F = N T
col(1N ⊗ Cn,i )
col (Cn,i A0 + Cr,i )
i=1
i=1
C0
0
0TN ⊗ In
0
0
Eu ,
0
0
1TN ⊗ In
(109)
Fu
∆ ∈ ∆d := e−shi IN (n+nz ) : h ≥ 0, ℜ(s) ≥ 0 .
Moreover, the system is positive if and only if A0 is Metzler and the matrices E, C and F are nonnegative.
We then have the following result:
Theorem 42 Assume that the system (94) is positive. Then, the following statements are equivalent:
(i) The system (108) is asymptotically stable and has L1 -gain smaller than γ1 if and only if
• the matrix
N
X
An,i is Schur stable and
i=1
• there exists a vector λ ∈ Rn>0 such that
!
N
X
−1
Ar,i
S
A0 +
T
λ
i=1!
!
N
N
N
1ny
X
X
X
−1
Ar,i
Cn,i S
Cr,i +
C0 +
i=1
i=1
where S := I −
N
X
i=1
S
Fu +
−1
N
X
i=1
Eu
!
Cn,i
S −1 Eu
T
0
<
γ1 1nu
(110)
An,i .
i=1
(ii) The system (108) is asymptotically stable and has L2 -gain smaller than γ2 if and only if there exists
N ny
diagonal matrices P ∈ Dn≻0 , Q ∈ DnN
≻0 and R ∈ D≻0 such that the LMI
T
A0 P + P A0 P (1TN ⊗ In )
0
P Eu
T
⋆
−Q
0
0
+ C T diag(Q, R, Iny ) C ≺ 0
(111)
F
⋆
⋆
−R
0
F
⋆
⋆
⋆
−γ22 Inu
where the matrices A, E, C, F are defined in Proposition 41.
27
(iii) The system (108) is asymptotically stable and has L∞ -gain smaller than γ∞ if and only if
• the matrix
N
X
An,i is Schur stable and
i=1
• there exists a vector λ ∈ Rn>0 such that
!
N
X
−1
Ar,i
S
A0 +
i=1
!
!
N
N
N
X
X
X
−1
Ar,i
Cn,i S
Cr,i +
C0 +
where S := I −
N
X
Fu +
i=1
i=1
i=1
S
−1
N
X
Eu
!
Cn,i
i=1
S −1 Eu
λ
0
<
1nu
γ∞ 1ny
(112)
An,i .
i=1
Proof : The statement (ii) can be obtained using the scaled-small gain in the L2 -framework. We now prove
the statement (iii), statement (i) can be proven in exactly the same way. First note that the system (108)
is asymptotically stable and has L∞ -gain smaller than γ∞ if and only if there exist some vectors λ ∈ Rn>0 ,
u
µ1i ∈ Rn>0 and µ2i ∈ Rn>0
, i = 1, . . . , N such that the following inequality
λ
0
1nu γ1ny
rowN
i=1 (Iny )
1 <
colN
0
0
i=1 (µi )
2
0
colN
−Iny N
i=1 (µi )
(113)
is satisfied. Let us denote for simplicity the above matrix by [Mij ]i,j=1,2 . Solving for the µ terms yields
that the above condition is equivalent to saying that
λ
0
<
(114)
M11 − M12 M−1
M
21
22
1nu
γ1ny
A0
Eu
C
Fu
0
N
coli=1 (An,i A0 + Ar,i ) colN
(A
n,i Eu )
i=1
N
colN
(C
A
+
C
)
col
(C
n,i 0
r,i
n,i Eu )
i=1
i=1
rowN
i=1 (In )
0
T
colN
(1
⊗
An,i ) − IN n
i=1
N
T
colN
(1
i=1 N ⊗ Cn,i )
0
together with M22 is Hurwitz stable.P
It is immediate to see that, once again, the latter condition is equivalent
N
T
to the Schur stability of the matrix N
i=1 An,i . For compactness, let us denote now A := coli=1 (1N ⊗ An,i )
N
T
and C := coli=1 (1N ⊗ Cn,i ) and, using this notation, we get that
N
T
col
(1
⊗
A
)
−
I
0
n,i
nN
N
M22 = i=1 N
(115)
−InN
col (1TN ⊗ Cn,i )
i=1
and, hence, we have that
M−1
22 =
(A − InN )−1
C(A − InN )−1
0
−IN n
together with
−1
(A − InN )
Letting now S := I −
PN
i=1
N
= IN n + col(An,i ) In −
i=1
N
X
i=1
An,i
(116)
!−1
An,i , we then get that
#
!
"
N
X
−1
Ar,i λ + Eu 1nu < 0
A0 +
S
i=1
28
N
row(In ).
i=1
(117)
(118)
and similar manipulation gives
!#
!
"
N
N
N
X
X
X
λ+
Ar,i
Cn,i S −1
Cr,i +
C0 +
i=1
i=1
i=1
The proof is now completed.
Fu +
N
X
i=1
Cn,i
!
S −1 Eu
!
1nu < γ1ny .
(119)
♦
As for the other systems, the L∞ condition remains the same in the case of time-varying delays provided
that t − hi (t) → ∞ as t → ∞, i = 1, . . . , N . On the other hand, the other conditions need to be slightly
changed to incorporate the rate of variation of the delays as in the other results.
9
Conclusion
Several recent results regarding the robust stability analysis of uncertain linear positive systems have been
unified in a single formulation using a generalization of the structured singular value. Using this generalization, several necessary and sufficient conditions have been obtained and expressed in terms of scaled
small-gain theorems involving linear or semidefinite programs. These results have been considered for establishing several results for linear positive systems with constant and time-varying delays. It is notably
recalled that the time-varying nature of the delay never deteriorates the asymptotic stability of linear positive
systems but may deteriorate their Lp stability.
Interesting extensions could be concerned with the robust stabilization problem using static/dynamic
output-feedback or state-feedback controllers using ideas from [16, 18, 88, 62] or the extension of the results
to hybrid systems [89]. The design of interval observers is also a potentially interesting follow-up to this
work; see e.g. [9–11, 13, 90].
References
[1] L. Farina and S. Rinaldi. Positive Linear Systems: Theory and Applications. John Wiley & Sons, 2000.
[2] J. D. Murray. Mathematical Biology Part I. An Introduction. 3rd Edition. Springer-Verlag Berlin
Heidelberg, 2002.
[3] C. Briat and E. I. Verriest. A new delay-SIR model for pulse vaccination. Biomedical signal processing
and control, 4(4):272–277, 2009.
[4] C. Briat and M. Khammash. Computer control of gene expression: Robust setpoint tracking of protein
mean and variance using integral feedback. In 51st IEEE Conference on Decision and Control, pages
3582–3588, Maui, Hawaii, USA, 2012.
[5] A. Gupta, C. Briat, and M. Khammash. A scalable computational framework for establishing long-term
behavior of stochastic reaction networks. PLOS Computational Biology, 10(6):e1003669, 2014.
[6] C. Briat, A. Gupta, and M. Khammash. Antithetic integral feedback ensures robust perfect adaptation
in noisy biomolecular networks. Cell Systems, 2:17–28, 2016.
[7] C. Briat, C. Zechner, and M. Khammash. Design of a synthetic integral feedback circuit: dynamic
analysis and DNA implementation. ACS Synthetic Biology, 5(10):1108–1116, 2016.
[8] J. L. Gouzé, A. Rapaport, and M. Z. Hadj-Sadok. Interval observers for uncertain biological systems.
Ecological modelling, 133:45–56, 2000.
[9] F. Mazenc and O. Bernard. Interval observers for linear time-invariant systems with disturbances.
Automatica, 47:140–147, 2011.
29
[10] C. Briat and M. Khammash. Interval peak-to-peak observers for continuous- and discrete-time systems
with persistent inputs and delays. Automatica, 74:206–213, 2016.
[11] D. Efimov, A. Polyakov, and J.-P. Richard. Interval observer design for estimation and control of
time-delay descriptor systems. European Journal of Control, 23:26–35, 2015.
[12] D. Efimov, E. Fridman, A. Polyakov, W. Perruquetti, and J.-P. Richard. On design of interval observers
with sampled measurement. Systems & Control Letters, 96:158–164, 2016.
[13] D. Efimov, E. Fridman, A. Polyakov, W. Perruquetti, and J.-P. Richard. Linear interval observers under
delayed measurements and delay-dependent positivity. Automatica, 72:123–130, 2016.
[14] P. H. A. Ngoc and H. Trinh. Novel criteria for exponential stability of linear neutral time-varying
differential systems. IEEE Transactions on Automatic Control, 61(6):1590–1594, 2016.
[15] F. Mazenc and M. Malisoff. Stability analysis for time-varying systems with delay using linear Lyapunov
functionals and a positive systems approach. IEEE Transactions on Automatic Control, 61(3):771–776,
2016.
[16] M. Ait Rami and F. Tadeo. Controller synthesis for positive linear systems with bounded controls.
IEEE Transactions on Circuits and Systems – II. Express Briefs, 54(2):151–155, 2007.
[17] M. Ait Rami. Solvability of static output-feedback stabilization for LTI positive systems. Systems &
Control Letters, 60:704–708, 2011.
[18] C. Briat. Robust stability and stabilization of uncertain linear positive systems via integral linear constraints - L1 - and L∞ -gains characterizations. International Journal of Robust and Nonlinear Control,
23(17):1932–1954, 2013.
[19] C. Briat. Robust stability analysis of uncertain linear positive systems via integral linear constraints - L1 and L∞ -gains characterizations. In 50th IEEE Conference on Decision and Control, pages 3122–3129,
Orlando, Florida, USA, 2011.
[20] Y. Ebihara, D. Peaucelle, and D. Arzelier. L1 gain analysis of linear positive systems and its applications.
In 50th Conference on Decision and Control, Orlando, Florida, USA, pages 4029–4034, 2011.
[21] A. Rantzer. On the Kalman-Yakubovich-Popov lemma for positive systems. IEEE Transactions on
Automatic Control, 61(5):1346–1349, 2016.
[22] T. Tanaka and C. Langbort. KYP Lemma for internally positive systems and a tractable class of
distributed H-infinity control problems. In American Control Conference, pages 6238–6243, Baltimore,
Maryland, USA, 2010.
[23] M. Colombino and R. S. Smith. A convex characterization of robust stability for positive and positively
dominated linear systems. IEEE Transactions on Automatic Control, 61(7):1965–1971, 2016.
[24] S. Z. Khong, C. Briat, and A. Rantzer. Positive systems analysis via integral linear constraints. In 54th
IEEE Conference on Decision and Control, pages 6373–6378, Osaka, Japan, 2015.
[25] A. Megretski and S. Treil. Power distribution in optimization and robustness of uncertain systems.
Journal of Mathematical Systems, Estimation and Control, 3:301–319, 1993.
[26] A. Megretski and A. Rantzer. System analysis via Integral Quadratic Constraints. IEEE Transactions
on Automatic Control, 42(6):819–830, 1997.
[27] A. Packard and J. C. Doyle. The complex structured singular value. Automatica, 29:71–109, 1993.
30
[28] M. A. Dahleh and I. J. Diaz Bobillo. Control of uncertain systems - A linear programming approach.
Prentice-Hall, 1995.
[29] M. H. Khammash. Necessary and sufficient conditions for the robustness of time-varying systems with
applications to sampled-data systems. IEEE Transactions on Automatic Control, 38(1):49–57, 1993.
[30] W. M. Haddad and V. Chellaboina. Stability theory for nonnegative and compartmental dynamical
systems with time delay. Systems & Control Letters, 51(5):355–361, 2004.
[31] M. Ait Rami. Stability analysis and synthesis for linear positive systems with time-varying delays.
In Positive systems - Proceedings of the 3rd Multidisciplinary International Symposium on Positive
Systems: Theory and Applications (POSTA 2009), pages 205–216. Springer-Verlag Berlin Heidelberg,
2009.
[32] O. Mason. Diagonal Riccati stability and positive time-delay systems. Systems & Control Letters, 61:
6–10, 2012.
[33] J. Zhu and J. Chen. Stability of systems with time-varying delays: An L1 small-gain perspective.
Automatica, 52:260–265, 2015.
[34] J. Shen and J. Lam. ℓ∞ /L∞ -gain analysis for positive linear systems with unbounded time-varying
delays. IEEE Transactions on Automatic Control, 60(3):857–862, 2015.
[35] J. Zhang, C. R. Knospe, and P. Tsiotras. Stability of time-delay systems: Equivalence between Lyapunov
and scaled small-gain conditions. IEEE Transactions on Automatic Control, 46:482–486, 2001.
[36] S. I. Niculescu. Delay effects on stability. A robust control approach, volume 269. Springer-Verlag:
Heidelbeg, 2001.
[37] C. R. Knospe and M. Roozbehani. Stability of linear systems with interval time-delay. In IEEE American
Control Conference, pages 1458–1463, Denver, Colorado, 2003.
[38] K. Gu, V. L. Kharitonov, and J. Chen. Stability of Time-Delay Systems. Birkhäuser, Boston, 2003.
[39] F. Gouaisbaut and D. Peaucelle. Stability of time-delay systems with non-small delay. In Conference
on Decision and Control, San Diego, California, USA, pages 840–845, 2006.
[40] C. R. Knospe and M. Roozbehani. Stability of linear systems with interval time delays excluding zero.
IEEE Transactions on Automatic Control, 51:1271–1288, 2006.
[41] F. Gouaisbaut and D. Peaucelle. Delay dependent robust stability of time delay-systems. In 5th IFAC
Symposium on Robust Control Design, Toulouse, France, 2006.
[42] C. Y. Kao and A. Rantzer. Stability analysis of systems with uncertain time-varying delays. Automatica,
43:959–970, 2007.
[43] Y. Ariba and F. Gouaisbaut. Input-output framework for robust stability of time-varying delay systems.
In 48th Conference on Decision and Control, pages 274–279, Shanghai, China, 2009.
[44] Y. Ariba, F. Gouaisbaut, and K. H. Johansson. Stability interval for time-varying delay systems. In
49th Conference on Decision and Control, pages 1017–1022, Atlanta, Georgia, USA, 2010.
[45] F. Gouaisbaut and Y. Ariba. Delay range stability of a class of distributed time delay systems. Systems
& Control Letters, 60:211–217, 2011.
[46] C. Briat. Linear Parameter-Varying and Time-Delay Systems – Analysis, Observation, Filtering &
Control, volume 3 of Advances on Delays and Dynamics. Springer-Verlag, Heidelberg, Germany, 2015.
31
[47] E. Fridman. Introduction to Time-Delay Systems.
Switzerland, 2014.
Birkhäuser, Springer International Publishing
[48] J. Zhu, T. Qi, and J. Chen. Small-gain stability conditions for linear systems with time-varying delays.
Systems & Control Letters, 81:42–48, 2015.
[49] X. Li, H. Gao, and K. Gu. Delay-independent stability analysis of linear time-delay systems based on
frequency discretization. Automatica, 70:288–294, 2016.
[50] J. Shen and J. Lam. L∞ -gain analysis for positive systems with distributed delays. Automatica, 50:
175–179, 2014.
[51] A. Aleksandrov and O. Mason. Diagonal Riccati stability and applications. Linear Algebra and its
Applications, 492:38–51, 2016.
[52] J. Shen and W. X. Zheng. Positivity and stability of coupled differentialdifference equations with
time-varying delays. Automatica, 57:123–127, 2015.
[53] Y. Ebihara, N. Nishio, and T. Hagiwara.
Stability analysis of neutral type time-delay
positive systems.
In 5th International Symposium on Positive Systems, 2016.
URL
http://www.posta2016.org/slides/4.EbiharaR.pdf.
[54] Y. Ebihara, N. Nishio, and T. Hagiwara. Stability analysis of neutral type time-delay positive systems.
In F. Cacace, L. Farina, R. Setola, and A. Germani, editors, Positive Systems – Theory and Applications
(POSTA 2016). Springer International Publishing, 2017.
[55] C. A. Desoer and M. Vidyasagar. Feedback Systems : Input-Output Properties. Academic Press, New
York, 1975.
[56] J. Stoer and C. Witzgall. Transformations by diagonal matrices in a normed space. Numerische
Mathematik, 4:158–171, 1962.
[57] A. Rantzer. Optimizing positively dominated systems. In 51st IEEE Conference on Decision and
Control, pages 271–277, Maui, Hawaii, USA, 2012.
[58] A. Sootla and A. Mauroy. Properties of eventually positive linear input-output systems. ArXiv e-prints
arXiv:1509.08392, 2015.
[59] C. Altafini. Minimal eventually positive realizations of externally positive. Automatica, 68:140–148,
2016.
[60] G. E. Dullerud and F. Paganini. A course in robust control theory. A convex approach. Springer, New
York, USA, 2000.
[61] R. Shorten, O. Mason, and C. King. An alternative proof of the Barker, Berman, Plemmons result on
diagonal stability and extensions. Linear Algebra and Its Applications, 430:34–40, 2009.
[62] M. Naghnaeian and P. G. Voulgaris. Performance optimization over positive l∞ cones. In American
Control Conference, pages 5645–5650, Portland, USA, 2014.
[63] M. G. Safonov and M. Athans. On stability theory. In 17th Conference on Decision and Control, pages
301–314, San Diego, California, USA, 1978.
[64] T. Iwasaki and S. Hara. Well-posedness of feedback systems: insight into exact robustness analysis and
approximate computations. IEEE Transactions on Automatic Control, 43:619–630, 1998.
[65] L. Hörmander. The analysis of linear partial differential operators I. Springer Verlag, 1985.
32
[66] V. Blondel and A. Megretski. Unsolved Problems in Mathematical Systems and Control Theory. Princeton University Press, 2004.
[67] A. Aleksandrov, O. Mason, and A. Vorob’eva. Diagonal Riccati stability and the Hadamard product.
2016. URL https://arxiv.org/abs/1612.06587v1.
[68] E. I. Verriest. Robust stability, adjoints, and LQ control of scale-delay systems. In 38th IEEE Conference
on decision and control, pages 209–214, Phoenix, Arizona, USA, 1999.
[69] C. Briat. Robust stability analysis in the ∗-norm and Lyapunov-Razumikhin functions for the stability
analysis and control of time-delay systems. In 50th IEEE Conference on Decision and Control, pages
6319–6324, Orlando, Florida, USA, 2011.
[70] E. Shustin and E. Fridman. On delay-derivative-dependent stability of systems with fast-varying delays.
Automatica, 43:1649–1655, 2007.
[71] C. E. Avellar and J. K. Hale. On the zeros of exponential polynomials. Journal of Mathematical Analysis
and Applications, 73:434–452, 1980.
[72] S. Damak, M. Di Loreto, W. Lombardi, and V. Andrieu. Exponential L2 -stability for a class of linear
systems governed by continuous-time difference equations. Automatica, 50:3299–3303, 2014.
[73] S. Damak, M. Di Loreto, W. Lombardi, and V. Andrieu. Stability of linear continuous-time difference
equations with distributed delay: Constructive exponential estimates. International Journal of Robust
and Nonlinear Control, 25(17):3195–3209, 2015.
[74] D. Melchor-Aguilar. A note on stability of functional difference equations. Systems & Control Letters,
67:211–215, 2016.
[75] D. Melchor-Aguilar. Exponential stability of linear continuous time difference systems with multiple
delays. Systems & Control Letters, 62:811–818, 2013.
[76] J. K. Hale and P. M. Amores. Stability in neutral equations. Nonlinear Analysis: Theory, Methods &
Applications, 1(1):161–172, 1977.
[77] S.-I. Niculescu and V. Rasvan. Delay-independent stability in lossless propagation models with applications (i): A complex domain approach. In International Symposium on Mathematical Theory of
Networks and Systems, 2000.
[78] S.-I. Niculescu and V. Rasvan. Delay-independent stability in lossless propagation models with applications (ii): A lyapunov-based approach. In International Symposium on Mathematical Theory of
Networks and Systems, 2000.
[79] P. Pepe and E. I. Verriest. On the stability of coupled delay differential and continuous time difference
equations. IEEE Transactions on Automatic Control, 48(8):1422–1427, 2003.
[80] S.-I. Niculescu, P. Fu, and J. Chen. On the stability of linear delay-differential algebraic systems: Exact
conditions via matrix pencil solutions. In 45th IEEE Conference on Decision & Control, pages 834–839,
2006.
[81] E. I. Verriest and P. Pepe. Time optimal and optimal impulsive control for coupled differential difference
point delay systems with an application in forestry. In Topics in Time Delay Systems, pages 255–265.
Springer Berlin Heidelberg, 2009.
[82] C. Briat. Sign properties of Metzler matrices with applications. Linear Algebra and its Applications,
515:53–86, 2017.
33
[83] O. Solomon and E. Fridman. New stability conditions for systems with distributed delays. Automatica,
49(11):3467–3475, 2013.
[84] J. K. Hale and S. M. Verduyn Lunel. Introduction to Functional Differential Equations. Springer-Verlag,
New York, USA, 1991.
[85] J. K. Hale and S. M. Verdyun Lunel. Strong stabilization of neutral functional differential equations.
IMA Journal of mathematical control and information, 19:5–23, 2002.
[86] A. Bellen and N. Guglielmi. Methods for linear systems of circuit delay differential equations of neutral
type. IEEE Transactions on Circuits and Systems I, 76(1):212–215, 1999.
[87] E. I. Verriest and P. Pepe. Time optimal and optimal impulsive control for coupled differential difference
point delay systems with an application in forestry. In IFAC Workshop on time-delay systems, Nantes,
France, 2007.
[88] Y. Ebihara, D. Peaucelle, and D. Arzelier. Optimal L1 -controller synthesis for positive systems and its
robustness properties. In American Control Conference, pages 5992–5997, Montreal, Canada, 2012.
[89] C. Briat. Dwell-time stability and stabilization conditions for linear positive impulsive and switched
systems. Nonlinear Analysis: Hybrid Systems, 24:198–226, 2017.
[90] C. Briat and M. Khammash. Simple interval observers for linear impulsive systems with applications to
sampled-data and switched systems. In 20th IFAC World Congress (Accepted), Toulouse, France, 2017.
34
| 3cs.SY
|
arXiv:1803.01683v1 [cs.SE] 22 Feb 2018
Investigating the Evolvability of Web Page Load Time
Brendan Cody-Kenny1 , Umberto Manganiello2 , John Farrelly2 , Adrian
Ronayne2 , Eoghan Considine2 , Thomas McGuire2 , and Michael O’Neill1
1
Natural Computing Research and Applications Group (NCRA), Michael
Smurfit Graduate Business School, University College Dublin, Ireland.
2
Fidelity Investments, Ireland.
March 6, 2018
Abstract
Client-side Javascript execution environments (browsers) allow anonymous functions
and event-based programming concepts such as callbacks. We investigate whether a
mutate-and-test approach can be used to optimise web page load time in these environments. First, we characterise a web page load issue in a benchmark web page and
derive performance metrics from page load event traces. We parse Javascript source
code to an AST and make changes to method calls which appear in a web page load
event trace. We present an operator based solely on code deletion and evaluate an existing “community-contributed” performance optimising code transform. By exploring
Javascript code changes and exploiting combinations of non-destructive changes, we can
optimise page load time by 41% in our benchmark web page.
Keywords: Javascript, Performance, Web Applications, Genetic Programming, SearchBased Software Engineering
1
Introduction
Performance characteristics vary across browsers where improvements in one version may
degrade performance in another [13]. Performance characteristics also change frequently as
a Javascript engine is subject to re-design. As a result, performance tuning is a never-ending
task. Javascript developers optimise code for Javascript engines while Javascript engine
developers optimise for how the engine will likely be used.
While a range of work has looked at mutation-based performance [9, 5] and energy improvement [3], no work we are aware of has inspected source code mutation for page load
time in the browser. Related work has looked at web service component selection [4], though
this targets components which are a higher level granularity of software unit.
In this paper we investigate the base mechanisms needed for a Genetic Programming
(GP) code improvement system: fitness measures and operators. We trace page load for
a simple benchmark web app and calculate (i) time, (ii) number of events, and (iii) the
largest depth of event chains found in the trace. We inspect two operators, one which deletes
statements and expressions containing method names found in the trace, and another which
transforms loops to more optimal versions. To validate these mechanisms, we apply each
1
operator iteratively to all source code locations where they are applicable using a greedy
search loop. After applying an operator, if the web page appears as expected, we keep the
source code mutation.
Our code deletion operator was able to reduce (i) page load time by 41%, (ii) total events
by 30% and (iii) event depth by 26%. While these results are encouraging they are relevant
only to our benchmark application, which contains much redundancy by design.
We cover related work in the following section, with Section 3 providing the experimental
setup, the target web application is described in Section 4. We then outline the observations arising from the experiments in Section 5 before drawing conclusions and suggesting
directions for future work (Section 6).
2
Related Work
Previous work on performance improvement has focused on run-time [9, 5] and energy improvement [3]. Program performance improvements frequently result from code deletion [8],
motivating us to initially investigate a deletion operator. Repeatedly applying deletion
produces a sub-program similar to one which could be found using program slicing techniques [2, 15]. Designing operators for performance improvement is an open problem, though
it has been proposed that new operators may be derived by mining existing code and code
generated during GP runs [11]. Program transforms are also written and released by developers in the spirit of making useful transforms reusable [14], adding another potential source
of operators.
Exhaustive mutation has been used to find how robust program functionality is to source
code change. Mutating small programs with fine-grained operators in a relatively statically
typed language such as Java appears to result in relatively low mutational robustness of
30% [5], while larger programs in C++ appear to have high mutational robustness of up
to 89% [10, 12] and 68% in more recent work [7]. As mutational robustness varies depending on the software under evolution, we currently have relatively few data points to draw
comparisons.
When using search algorithms on large programs it is important to focus operators to
reduce search space size [6, 8]. More targeted mutation operators perform program transforms which are highly likely to impact performance while leaving functionality unchanged.
For example, multiple different list implementations which have the same interface can be
evaluated [1].
3
Experimental Setup
We gather performance measures based on (i) page load time, (ii) number of events during
this time as well as (iii) the depth of event chains. We investigate two operators, one written
by the authors which deletes code based on method names and one which has been made
available as a community contribution. The search loop used simply keeps a code change if
the page does not show any error.
2
3.1
Metrics for Web Page Load
We are mainly interested in improving page load time for a web app. We gather traces
of the web app loading via chrome browser’s devtools functionality1 (with caching disabled)
using a client for NodeJS2 . Page load traces list events which can be used to build a call
graph for the entire page load process. The elapsed time for all events to complete and the
depth of call sequences can be calculated from a page trace. We take the time between the
first and last event recorded in the browser to give load time.
We sum browser events as a pseudo-measure of performance, which is subject to less
variation than time-based measures. The number of events in total gives us an idea of how
much work is being performed by the browser. We also take a measure of the most deeply
nested sequence of event calls, which gives us a rough idea of how interdependent events are.
Our intuition is that time can elapse while this call graph is traversed even though more
computation could be performed during the same elapsed time if dependent method calls
could be rearranged.
We assume we are beginning with an “oracle” web app which is considered correct.
During page load, we check what elements have been loaded onto the page, and the page is
only considered fully loaded when a string appears as part of the Document Object Model
(DOM) for the page. We sum the number of pixels by which screenshots differ and use this as
our only measure of functional correctness3 (as this is the only functionality our benchmark
provides). We compare two oracle screenshots to get a measure of acceptable variation while
still considering the web app correct. We use a multiple of this acceptable pixel variability
measure to get a threshold value, above which the page is considered incorrect, that is, too
different to be considered the same as the original. Subsequent screenshots taken of web app
variants are compared with the original oracle screenshot.
A screenshot captures the final state of the page after it has been loaded. The screenshot
does not capture anything about the underlying state of the web page or the structure of
the web page Document Object Model (DOM). As a result, using screenshots to measure
correctness relaxes the constraints on what is considered correct and frees the evolutionary
search process to make changes which affect the underlying structure of the web page HTML.
A screenshot only tells us if the page loaded as expected or not, and does not give us any
gradient or measure of subsequent page functionality. A screenshot gives no way of telling
apart a completely disfunctional page load from a partially functional version of the app.
Some functionality is better than none, and Evolutionary algorithms rely on this gradient. In
future work the HTML state could additionally be used to give partial functionality gradient.
3.2
Operators
We investigate the use of two operators, one which was developed by the authors of this paper,
and the other, which was developed and made freely available by open source contributors.
Both operators build an Abstract Syntax Tree (AST) of Javascript source files which are
searched, manipulated and written back to a file.
1 https://developer.chrome.com/devtools
2 https://github.com/cyrus-and/chrome-remote-interface
3 http://www.imagemagick.org/Usage/compare/#statistics
3
3.2.1
Deletion Operator
The simplest way to reduce page load time is to reduce the amount of computation done.
To inspect this, we delete portions of the web app source code. Deleting code gives us some
indication of the mutational robustness of the software we are targeting. If any portion of
the code can be deleted without affecting the correctness of the resulting web page, then we
can say that there is some level of redundancy in the code. Claims of robustness can only
be made within the context of the operators and test cases used. If the correctness tests
(pixel differences between screenshots) do not capture important features or functionality of
the web app, then deletion may find performance improvements which turn out to have an
undesirable effect on functionality.
A page load trace contains method calls but also which Javascript file the method calls
were made from. We use this information to find where methods are called and defined
within the Javascript source. We should also note that deletion is distinct from dead code
removal. Only method names which appear in the page load trace are considered for deletion.
As these methods appear in the trace, we know they have been executed and would not be
removed by dead code removal.
3.2.2
Loop Optimiser
The “loop optimiser” operator has made freely available4 as a “community” contribution to the
Babel project5 . This source code transform is intended to improve the performance of loops.
A potential issue with this type of operator is that it is not well tested6 . While community
written program transforms can contain useful improvements, they are not written or tested
with the same rigour that might be found in, for example, the transforms performed by a
compiler.
Similar community efforts to produce code transforms gives the Evolutionary Computation community a potential source of operators. Though an operator may perform a beneficial
transform in most contexts, a developer will still need to test and validate the effect this transform has on their code. A potential impediment to the uptake of these transforms is that we
do not have a high level of assurance that these transforms are rigorously semantics preserving. For languages that are dynamically typed, it is more difficult to ensure new transforms
preserve semantics. If program transforms are to be crowd-sourced to some extent, we may
end up with a scenario where the proliferation of transforms becomes an issue in itself for
a developer as time is required to choose, apply and validate transforms. This motivates a
Genetic Programming mutate-and-test approach to the application of community operators.
3.3
Search Loop
The search loop iterates through a list of file names from the web app and applies an operator
to each one. When applying deletion, the loop iterates through a list of method names and
associated files. Method names and files are extracted from a page load trace and used to
focus operators on statements and expressions which contain the method name. For each file,
each mention of a method name in the AST is found. From each element in the AST which
contains the method name, the parent statement or expression is deleted. The app is then
loaded over HTTP via the chrome browser which captures a trace of the page load and a
4 https://github.com/vihanb/babel-plugin-loop-optimizer
5 https://babeljs.io/
6 The
version we used was specifically marked as experimental
4
screenshot. Tracing will timeout after twice the time the oracle page load takes (3 seconds).
Runtime and event counts are extracted from the trace (used in post-analysis only) and the
screenshot is compared with the oracle screenshot to give correctness values. If the page is
considered correct, the mutation is kept, otherwise the mutation is discarded and the next
operator application is made. To be clear, the performance metrics are not used to guide
the search process. The only metric which determines whether a change is kept, is whether
the page loads as expected. The performance metrics are not used directly as part of the
search loop but are used for post-analysis. When all possible operator applications have
been made, we compare the most recent patch, which contains the culmination of all code
changes, to the original. By virtue of the operator we inspect, if the page loads correctly
after many deletions, we can perform more in-depth performance bench-marking to quantify
the performance improvement found.
4
Web Application
In this paper we use an Angular2 web app which is written in typescript and is compiled
to Javascript7 It is this generated Javascript that we target for optimisation. The app is
designed to contain redundancy and as an exemplar app with performance issues. The
functionality provided by the app is to load a paragraph on the screen and display page load
time8 . Page load takes under two seconds even though the resulting page would be far more
effectively delivered as static HTML, instead of being loaded via Angular2 functionality.
Any optimisations found will show a clearly measurable improvement in runtime as well as
being easy to understand and validate. In this scenario, Angular is far too heavyweight a
solution for the resulting web page state. While this appears an easy optimisation target, it
is however a challenging benchmark problem for an evolutionary approach and an excellent
initial problem on which to validate our approach. The challenge lies in disentangling the
extensive sequence of method calls which ultimately produce a static page of text. The oracle
web application triggers over 5000 events to a depth of 242 nested method calls. Although
Angular is designed to deliver more complex functionality, the increased level of abstraction
nonetheless makes it difficult to understand interactions between many layers of sometimes
anonymous and interleaved method calls for this simple app. Our hope is that an evolutionary
algorithm utilising a range of operators which can reduce the dependencies and redundancies
in this benchmark application will also be able to reduce other “real world” web applications
in terms of load time.
5
Results
Figure 1 shows 100 repeated measurements of page load time for the original web app and
an optimised version of the web app which was derived by the exhaustive application of the
deletion operator. Each sequence of traces shows roughly the same pattern, though they
appear offset by roughly 500ms. We notice that there is a warm-up factor when traces are
repeatedly gathered on the same page as initial page load time measures are high in the first
couple of traces when compared to the relatively stable measures gathered subsequently.
Figure 2 shows average values for time, number of events and largest event depth from
1000 samples taken after browser warm-up. On average, there is a 41% saving in page load
7 Full
web app code: https://github.com/mlaval/optimize-angular-app
https://mlaval.github.io/optimize-angular-app/dev
8 Demo:
5
1800
Original
Optimised
Milliseconds
1600
1400
1200
1000
800
600
0
10
20
30
40
50
60
Trace Number
70
80
90
100
6000
270
1400
5000
225
Event Depth
1600
4000
1200
Events
Milliseconds
Figure 1: Trace of Original and Optimised web app versions
1000
800
3000
2000
1000
600
(a) Time
Opt
135
90
45
0
Orig
180
0
Orig
Opt
(b) Events
Orig
Opt
(c) Event Depth
Figure 2: Average Metrics (1000 samples) for Original (Orig) and Optimised (Opt) versions
of the web app
6
time. 30% less events occurred and event depth was 26% lower. Variance in measures was
roughly the same except for event depth which varied significantly in the optimised version
of the app. This may indicate that there are reduced dependencies in this version of the app
which leave the browser more freedom in when and how events are triggered. Although a
saving of 500ms is a lot in terms of page load time, the mean page load time for the optimised
version of the app was still 720ms. Given that the final rendered page is mostly static and
the page is loading from the local machine we should be able to achieve far lower load times
with additional operators and further refinements of our approach leaving us ample room for
future improvement on this benchmark problem.
22% of deletions had no effect on functionality as measured and we can say that the
benchmark app we used is relatively robust to deletions. This is an interesting result as we
expected removing code which appears in the load trace to be more destructive. 8966 lines
were deleted out of a total of 17022 (52%).
Method names found in traces of the app appear 982 times in the app source code.
For this particular benchmark app, as method calls were deleted, other method calls would
appear in the trace. This is due to the redundancy contained in this benchmark application.
We found one iteration of our search loop, which performs parsing, AST traversal, file
writing, page load over HTTPS, trace call tree building, call tree traversal and source code
diffing took between 7 and 13 seconds. We feel this is not prohibitive for evolution in the
browser, especially given that it is likely to be reduced on further refinement of our approach.
We used the Loop Optimiser transform in our search loop but unfortunately no performance improvements were found. During experimentation we also applied the loop optimiser
to all Javascript files in the web app and found that it resulted in a page that did not load.
The existence and availability of such experimental operators justify a search-based approach
which can discover what operators are applicable where.
For GI research, these results reiterate the question as to what is the best way to go about
operator design. We could have a wide range of specific operators which are only applicable
to certain code patterns (loop optimisation), or have a few very general operators (delete,
clone, replace) which are unlikely to leave a program in a fully functional state and only
rarely improve performance.
5.0.1
Limitations
The main limitation of this work is that the benchmark app is very simple and contains a
lot of redundancy by design. As a result, the findings give little indication as to how generalisable the approach is. When considering more real-world applications of this approach
we feel it is unlikely that they will contain the same level of redundancy or be so cleanly
structured that deleting lines of code will improve performance while leaving functionality
similarly unaffected. Additionally, though comparing screenshots is enough to detect a “correct” page load for our benchmark app, a web app which provides user functionality would
need extensive additional testing.
6
Conclusion
We find that improvement of browser-based Javascript using iterative mutation and testing
is possible. We reduce a benchmark app to only what code is necessary to provide the desired
functionality, saving 41% in runtime. We may compare code changes against those derived
through program slicing techniques [15] and investigate if they produce similar performance
7
improvements. Finding performance improvements further to a minimal code “slice” will
likely require more inventive operators. What operators to use is a pertinent open question.
We highlight that community contributed code transforms may make ideal candidate operators for search-based approaches. Our findings also give a good base for expanding the work
toward more complex “real world” applications with more detailed measures of functionality
(e.g. HTML diff or navigation testing) and performance (e.g. user interface jitter).
Acknowledgements This research is based upon works supported by Science Foundation Ireland
under grant 13/IA/1850 and 13/RC/2094 which is co-funded under the European Regional Development Fund through the Southern & Eastern Regional Operational Programme to Lero - the Irish
Software Research Centre (www.lero.ie).
References
[1] Basios, M., Li, L., Wu, F., Kanthan, L., Barr, E.T.: Optimising Darwinian Data Structures
on Google Guava. International Symposium on Search Based Software Engineering (SSBSE)
(2017)
[2] Binkley, D., Harman, M.: A Survey of Empirical Results on Program Slicing. Advances in
Computers 62 (2004)
[3] Bokhari, M.A., Bruce, B.R., Alexander, B., Wagner, M.: Deep Parameter Optimisation on Android Smartphones for Energy Minimisation. In: Proceedings of the Genetic and Evolutionary
Computation Conference Companion (GECCO). Berlin (2017)
[4] Chang, W.C., Wu, C.S., Chang, C.: Optimizing Dynamic Web Service Component Composition
by Using Evolutionary Algorithms. In: International Conference on Web Intelligence, WIC
(2005)
[5] Cody-Kenny, B., Galván-López, E., Barrett, S.: locoGP: Improving Performance by Genetic
Programming Java Source Code. In: Proceedings of the Genetic and Evolutionary Computation
Conference Companion (GECCO) (2015)
[6] Forrest, S., Nguyen, T., Weimer, W., Goues, C.L.: A Genetic Programming Approach to
Automated Software Repair. In: Genetic and Evolutionary Computation Conference (GECCO)
(2009)
[7] Langdon, W.: Evolving better RNAfold C source code. bioRxiv (2017)
[8] Langdon, W.B.: Which is Faster: Bowtie2 GP Bowtie> Bowtie2> BWA. In: Proceedings of
the Genetic and Evolutionary Computation Conference Companion (GECCO) (2013)
[9] Langdon, W.B.: Performance of Genetic Programming Optimised Bowtie2 on Genome Comparison and Analytic Testing (GCAT) Benchmarks. BioData mining 8(1) (2015)
[10] Langdon, W.B., Petke, J.: Software is Not Fragile. In: Complex Systems Digital Campus
E-conference, (CS-DC). Proceedings in Complexity (2015)
[11] Petke, J.: New Operators for Non-functional Genetic Improvement. In: Proceedings of the
Genetic and Evolutionary Computation Conference Companion (GECCO) (2017)
[12] Schulte, E., Fry, Z.P., Fast, E., Weimer, W., Forrest, S.: Software Mutational Robustness.
Genetic Programming and Evolvable Machines (GPEM) (2014)
[13] Selakovic, M., Pradel, M.: Performance issues and optimizations in JavaScript: An empirical
study. In: International Conference on Software Engineering (ICSE) (2016)
[14] Various: jscodeshift. https://github.com/facebook/jscodeshift (2017)
[15] Ye, J., Zhang, C., Ma, L., Yu, H., Zhao, J.: Efficient and Precise Dynamic Slicing for ClientSide JavaScript Programs. In: International Conference on Software Analysis, Evolution, and
Reengineering (SANER). vol. 1 (2016)
8
| 9cs.NE
|
arXiv:1504.00273v3 [math.GR] 24 May 2017
On Alternating and Symmetric Groups
Which Are Quasi OD-Characterizable
Ali Reza Moghaddamfar
Faculty of Mathematics, K. N. Toosi University of Technology,
P. O. Box 16315-1618, Tehran, Iran.
E-mails: [email protected] and [email protected]
April 9, 2018
Abstract
Let Γ(G) be the prime graph associated with a finite group G and D(G)
be the degree pattern of G. A finite group G is said to be k-fold ODcharacterizable if there exist exactly k non-isomorphic groups H such
that |H| = |G| and D(H) = D(G). The purpose of this article is twofold.
First, it shows that the symmetric group S27 is 38-fold OD-charaterizable.
Second, it shows that there exist many infinite families of alternating and
symmetric groups, {An } and {Sn }, which are k-fold OD-characterizable
with k > 3.
Keywords: OD-characterization, alternating group, symmetric group, prime
graph, spectrum, degree pattern, split extension, subdirect product.
1
Introduction
Throughout the article, all the groups under consideration are finite and simple
groups are nonabelian. For a natural number n, we denote by π(n) the set of
all prime divisors of n and put π(G) = π(|G|). The spectrum ω(G) of a group G
is the set of orders of elements in G. The set ω(G) determines the prime graph
Γ(G) whose vertex set is π(G), and two vertices p and q are adjacent if and only
if pq ∈ ω(G). For two vertices p and q we will write (p ∼ q)G to indicate that p
is adjacent to q in Γ(G). Denote by s(G) the number of connected components
of Γ(G) and by πi (G) (i = 1, 2, . . . , s(G)), the set of vertices of its ith connected
component. If 2 ∈ π(G) then we assume that 2 ∈ π1 (G). We recall that the set
of vertices of connected components of all finite simple groups are obtained in
[5] and [14].
As usual, the degree deg(p) of a vertex p ∈ π(G) in Γ(G) is the number of
edges incident on p. We denote the set of all vertices of the prime graph Γ(G)
2000 Mathematics Subject Classification: 20D05, 20D06, 20D08.
1
which are joined to all other vertices by Λ(G). If the prime divisors of |G| are
p1 < p2 < · · · < pk , then we define D(G) := (deg(p1 ), deg(p2 ), . . . , deg(pk )),
which is called the degree pattern of G.
Given a group H, denote by hOD (H) the number of isomorphism classes
of groups G such that |G| = |H| and D(G) = D(H). Clearly, there are only
finitely many isomorphism types of groups of order n, because there are just
finitely many ways that an n × n multiplication table can be filled in. Hence
1 6 hOD (H) < ∞ for any group H. In terms of function hOD , groups H are
classified as follows:
Definition 1.1 Any group H satisfying hOD (H) = k is said to be k-fold ODcharacterizable. Usually, a 1-fold OD-characterizable group is simply called an
OD-characterizable group, and it is called quasi OD-characterizable if it is k-fold
OD-characterizable for some k > 1.
This article is a continuation of my investigations of the OD-characterizability
of alternating and symmetric groups initiated in [4]. We keep the notation created and the conventions made therein.
In a series of articles (see [4], [8], [9], [11] and [17]), it has been shown that
many of the alternating and symmetric groups are n-fold OD-characterizable
for n ∈ {1, 2, 3, 8}. These results are summarized in the following proposition.
Proposition 1.2 The following statements hold:
(a) The alternating groups Ap , Ap+1 , Ap+2 and the symmetric groups Sp and
Sp+1 , where p is a prime number, are OD-characterizable.
(b) The alternating group A10 is 2-fold OD-characterizable, while the symmetric group S10 is 8-fold OD-charaterizable.
(c) The alternating groups Ap+3 , where p 6= 7 is a prime number less than
100, are OD-characterizable.
(d) The symmetric groups Sp+3 , where p 6= 7 is a prime number less than 100,
are OD-characterizable or 3-fold OD-characterizable.
In addition, it was shown in [4, Corollary 1.5] that all alternating groups Am
for which m 6 100, except A10 , are OD-characterizable.
Proposition 1.3 All alternating groups Am , where m is a natural number less
than 100, except A10 , are OD-characterizable.
These observations convinced us to propose the following conjecture in [4]:
Conjecture 1.4 All alternating groups Am , with m 6= 10, are OD-characterizable.
On the other hand, in recent years we have not found any simple group S
with hOD (S) > 3. Therefore, we asked in [4] the following question:
Problem 1.5 Is there a simple group S with hOD (S) > 3?
2
Our recent investigations show that Conjecture 1.4 does not hold in general.
Recently, the authors showed in [6] that the alternating group A125 satisfying
hOD (A125 ) > 3 (see also [7]). Here, we will show that there exist infinite families
of alternating groups Am which are k-fold OD-characterizable with k > 3. We
notice that Problem 1.5 is also answered positively through these examples.
Theorem 1.6 There are infinitely many alternating groups Am which satisfy
hOD (Am ) > 1. In particular, there is no upper bound for hOD (Am ).
It is also worth mentioning that a similar description as Proposition 1.3
is exhibited about OD-characterizability of symmetric groups Sm , where m
is a natural number less than 100 (see [4, Theorem 1.7]). Nevertheless, in
checking the list of such groups, we found out that it contains a mistake (in
fact, Proposition 4.1 in [4] asserts erroneously that the symmetric group S27 is
3-fold OD-charaterizable). Therefore, another result of the present article can
be stated as follows:
Theorem 1.7 The symmetric group S27 is 38-fold OD-characterizable.
Now, we give a revised list of symmetric groups in question.
Corollary 1.8 All symmetric groups Sm , where m is a natural number less than
100, except m = 10, 27, are OD-characterizable or 3-fold OD-characterizable.
Moreover, the symmetric group S10 is 8-fold OD-characterizable, while the symmetric group S27 is 38-fold OD-characterizable.
The following conjecture involving symmetric groups is posed in [4]:
Conjecture 1.9 All symmetric groups Sm , with m 6= 10, are OD-characterizable
or 3-fold OD-characterizable.
It turns out that a negative answer to this conjecture is provided by symmetric group S27 (see also [6, 7]). In addition, we will get many other examples
of symmetric groups which are k-fold OD-characterizable with k > 3.
Theorem 1.10 There are infinitely many symmetric groups Sm which satisfy
hOD (Sm ) > 3. In particular, there is no upper bound for hOD (Sm ).
We conclude the introduction with some further notation and definitions.
Given a natural number m, we denote by lm the largest prime less than or
equal to m and we let △(m) = m − lm . It is clear that lm = m (or equivalently
△(m) = 0) if and only if m is a prime number. Note that if m > 2, then lm is
always an odd prime, and so △(m) is even iff m is odd. In addition, from the
definition, it is easy to see that
lm = lm−1 = lm−2 = . . . = lm−△(m)+1 .
We shall use the notation ν(m) (resp. νa (n)) to denote the number of types
of groups (resp. abelian groups) of order m. Clearly, νa (m) 6 ν(m). We also
denote the set of partitions of m by Par(m). It is known that for any prime p,
νa (pm ) = |Par(m)|. Finally, we use Am and Sm to denote the alternating and
symmetric group on m letters, respectively. In the case when p > 5 is a prime,
we denote by Sp the set of all simple groups with prime divisors at most p.
3
2
Auxiliary results
In this section we give several definitions and auxiliary results to be used later.
The first of them is the following definition of subdirect products.
Definition 2.1 Let n > 2. A subdirect product of the groups G1 , G2 , . . . , Gn is
a subgroup G 6 G1 × G2 × · · · × Gn of the direct product such that the canonical
projections G → Gi are surjective for all i.
One way of obtaining a subdirect product of two groups is via the fibre
product construction. This is illustrated here for two groups. Given some groups
G1 and G2 with normal subgroups N1 and N2 such that G1 /N1 and G2 /N2
are isomorphic, we want to construct a group G having a normal subgroup
N isomorphic to N1 × N2 such that G/N2 is isomorphic to G1 and G/N1 is
isomorphic to G2 . Notice that we will usually identify N1 ×1 with N1 and 1×N2
with N2 . To carry out the construction, let π1 and π2 be homomorphisms from
G1 and G2 onto some group Q. Now let
G := {(g1 , g2 ) ∈ G1 × G2 | π1 (g1 ) = π2 (g2 )} .
It is easy to check that G constitutes a subgroup of G1 × G2 , and the projection
maps onto the coordinates map G onto G1 and G2 , respectively. We call G the
fibre product associated with π1 and π2 (Remak [12] called it the meromorphic
product of G1 and G2 with normal subgroups N1 and N2 .). Also, N1 × N2 is
a normal subgroup of G and the map π on G defined by π((g1 , g2 )) = π1 (g1 )
maps G onto Q with kernel N1 × N2 , so G/(N1 × N2 ) is isomorphic to Q. In
fact, we have
G1 /N1 ∼
= G/(N1 × N2 ) ∼
= G2 /N2 .
It is a basic observation that every subdirect product of G1 and G2 is a fibre
product (or a meromorphic product) of these groups.
Information on the adjacency of vertices in the prime graphs associated with
alternating and symmetric groups can be found in [13, 16]. Consider the function
S : NQ
→ N, defined asP
follows: S(1) = 1 and for m > 1 with prime factorization
αi
i
,
S(m)
=
m = i pα
i
i pi . Then one has [16, Lemma 4 ]:
Lemma 2.2 Let m and n be ntaural numbers. Then there hold:
(1) An has an element of order m if and only if S(m) 6 n for odd m and
S(m) 6 n − 2 for even m.
(2) Sn has an element of order m if and only if S(m) 6 n.
The next two corollaries are the adjacency criteria for two vertices of Γ(An )
and Γ(Sn ), respectively.
Corollary 2.3 Let p, q ∈ π(An ) \ {2}. Then (p ∼ q)An if and only if p + q 6 n,
while (p ∼ 2)An if and only if p + 2 6 n − 2.
Corollary 2.4 Let p, q ∈ π(Sn ). Then (p ∼ q)Sn if and only if p + q 6 n.
4
The Goldbach conjecture says that every even natural number n greater
than 4 can be written as the sum of two odd primes. In what follows, we will
need a stronger conjecture:
Strong Goldbach Conjecture. Every even natural number n greater than six
can be written as the sum of two distinct odd primes.
We can now state the connection between the strong Goldbach conjecture
and the adjacency of vertices in the prime graph of a symmetric group:
Theorem 2.5 [2, Theorem 3] The following statements are equivalent:
(1) the strong Goldbach conjecture is true;
(2) for each even n > 6, Γ(Sn−1 ) $ Γ(Sn ).
Proof. It follows immediately from Lemma 2.2 (2).
The coincidence criterion for pairwise nonisomorphic symmetric groups (statement (1) of the Lemma 2.6) is obtained modulo strong Goldbach conjecture.
Lemma 2.6 Let m and n be natural numbers with 2 6 m < n. The prime
graphs of symmetric groups Sm and Sn are equal if and only if m = n − 1 and
one of the following holds:
(1) both n and n − 2 are non-prime odd numbers.
(2) n = 4 or 6.
Proof.∗ (=⇒) In the case when n 6 6, it is a straightforward verification. In
fact, the equality of the prime graphs Γ(Sn ) and Γ(Sn−1 ), for n ∈ {4, 6}, can be
obviously verified using Corollary 2.4. Assume now that n > 6. We first claim
that m = n − 1. If not, then m < n − 1 and hence one of the numbers n and
n − 1 is even. Since n > 7, it follows from strong Goldbach conjecture that there
exist two distinct odd primes p and q with m 6 n − 2 < p + q 6 n. Hence p is
adjacent to q in Γ(Sn ), while p is nonadjacent to q in Γ(Sm ), so these graphs
are not equal. This contradiction shows m = n − 1 as claimed.
By Theorem 2.5, we may assume that n is an odd number. Assume that
Γ(Sn ) 6= Γ(Sn−1 ). The sets of vertices are distinct if and only if n is a prime.
If n is a composite number, then the sets of vertices are equal, and so the
sets of edges should be distinct. Hence there exist primes p, q ∈ π(Sn ) with
n − 1 < p + q 6 n. Then obviously p + q = n, but n is odd. The only possible
case is {p, q} = {2, n − 2}, and so n − 2 ∈ π(Sn ) and n − 2 is a prime.
(⇐=) The conclusion follows immediately from Corollary 2.4.
We need the following lemma to find some infinite families of alternating
and symmetric groups which are not OD-characterizable.
Lemma 2.7 Let p be an odd prime number. There are infinitely many natural
numbers n such that △(pn ) > 4.
∗ The
idea of proof was borrowed from [18]
5
Proof. Take n to be an even natural number. Clearly, pn −4 is always composite
in this case, and so we need only consider pn − 2. Let a = p2 and b = 2. We will
now show that given any positive integers a, b, we can find an infinite number
of values for N such that aN − b is composite. If aN − b is always composite, we
are done. Otherwise, there exists a positive integer k such that ak − b = q > a
is prime. Then for all positive integers m, we have
ak+(q−1)m ≡ b (mod q),
and certainly ak+(q−1)m −b > q for all m > 0, so ak+(q−1)m −b must be composite
for all positive integers m.
A possible generalization of Lemma 2.7 is the following statement, which
seems intuitive to be true. Let p be an odd prime. Then, there are infinitely
many positive integers n such that
pn − 2, pn − 4, pn − 6, . . . , pn − (p − 1),
are composite. Alternatively, this problem can be formulated as follows.
Problem 2.8 Let p be an odd prime. Do there exist infinitely many positive
integers n such that ∆(pn ) > p.
3
The symmetric group S27
The aim of this section is to find the number of non-isomorphic groups with
the same order and degree pattern as the symmetric group S27 . Indeed, we will
show that there are 38 such groups.
Proof of Theorem 1.7. Let G be a group satisfying the following conditions:
(1) |G| = |S27 | = 223 · 313 · 56 · 73 · 112 · 132 · 17 · 19 · 23 and
(2) D(G) = D(S27 ) = (8, 8, 7, 7, 5, 5, 4, 4, 2).
Under these conditions, we conclude immediately that Γ(G) = Γ(S27 ) (see
also [4, Lemma 2.15]). Letting R be the solvable radical of G, we break the
proof into a number of separate lemmas.
Lemma 3.1 R is a {2, 3, 5, 7, 11}-group. In particular, G is nonsolvable.
Proof. First, we show that R is a 23′ -group. Assume the contrary. Let 23 ∈ π(R)
and let x be an element of R of order 23. Put C = CG (x) and N = NG (hxi).
By the structure of Γ(G), it follows that C is a {2, 3, 23}-group. Since N/C
is embedded in Aut(hxi) ∼
= Z22 , N is a {2, 3, 11, 23}-group. Using Frattini
argument we get G = RN , and so 19 ∈ π(R). Thus R contains a Hall {19, 23}subgroup L of order 19 · 23. Clearly, L is cyclic, and so (19 ∼ 23)G , which is a
contradiction.
Next, we show that R is a q ′ -group for q ∈ {13, 17, 19}. Let q ∈ π(R),
Rq ∈ Sylq (R) and N = NG (Rq ). Again, by Frattini argument G = RN and
6
since R is a 23′ -group we deduce that 23 divides the order of N . Let L be a
subgroup of N of order 23. Since L normalizes Rq , LRq is a subgroup of order
23 · |Rq |, which is abelian. But then (q ∼ 23)G , which is a contradiction.
Finally, R is a {2, 3, 5, 7, 11}-group, and since R 6= G, it follows that G is
nonsolvable. This completes the proof of lemma.
In what follows, we put G = G/R and S = Soc(G). Clearly, S is a direct
product
S = P1 × P2 × · · · × Pm ,
where the Pi are nonabelian simple groups, and we have
S 6 G 6 Aut(S).
We show now that S is a simple group, or equivalently m = 1.
Lemma 3.2 m = 1. In particular, G is an almost simple group.
Proof. By way of contradiction, let m > 2. In this case 23 does not divide |S|,
for otherwise deg(23) > 3, which is a contradiction. Hence, for every i we have
Pi ∈ S19 . Therefore 23 ∈ π(G) ⊆ π(Aut(S)), and so 23 divides the order of
Out(S). But
Out(S) = Out(Q1 ) × Out(Q2 ) × · · · × Out(Qr ),
where Qi is a direct product of ni isomorphic copies of a simple group Pi such
that
S∼
= Q1 × Q2 × · · · × Qr .
Therefore, for some j, 23 divides the order of the outer automorphism group of
Qj of nj isomorphic simple groups Pj . Since Pj ∈ S19 , it follows that |Out(Pj )|
is not divisible by 23 (see [3, 15]). Moreover, since Out(Qj ) = Out(Pj ) ≀ Snj , it
follows that
|Out(Qj )| = |Out(Pj )|nj · (nj )!.
This forces nj > 23, and so 246 must divide the order of G, which is a contradiction. Therefore m = 1 and S = P1 .
Lemma 3.3 There are exactly 38 possibilities for the group G.
Proof. By Lemma 3.2, we have S 6 G 6 Aut(S), where S is a nonabelian
simple group in S23 , and so {13, 17, 19, 23} ∩ π(Out(S)) = ∅, (see [3, 15]). Now,
it follows from Lemma 3.1 and condition (1) that
|S| = 2a · 3b · 5c · 7d · 11e · 132 · 17 · 19 · 23,
where 2 6 a 6 23, 0 6 b 6 13, 0 6 c 6 6, 0 6 d 6 3 and 0 6 e 6 2. Comparing
this with the nonabelian simple groups in S23 , we obtain S ∼
= A26 or A27 . We
refer to [15] for the list of nonabelian simple groups in S23 . In the sequel, we
deal with two cases separately.
7
∼ A27 or
(1) S ∼
= A27 . In this case, we have A27 6 G/R 6 S27 . Thus G/R =
G/R ∼
= S27 . If G/R ∼
= A27 , then |R| = 2. Clearly, R 6 Z(G) and G is
a central extension of R ∼
= Z2 . If G splits over R, then G ∼
= Z2 × A27 ,
∼
otherwise we have G = Z2 · A27 (non-split extension). Next, we assume
G/R ∼
= S27 . Finally, in the case when
= S27 . In this case R = 1 and so G ∼
∼
S = A27 there are three possibilities for G.
(2) S ∼
= A26 . In this case, we have A26 6 G/R 6 S26 , and so G/R ∼
= A26
or G/R ∼
= S26 . If G/R ∼
= A26 , then |R| = 54. We claim that the only
possibilities for G are G = R × A26 , where R is an arbitrary group of order
54, and G = Q × (Z2 · A26 ) where Q is an arbitrary group of order 27.
In the first case, since there are exactly 15 groups of order 54, there are
15 possibilities for G. In the second case, since there are just 5 groups
of order 27, there are 5 possibilities for G. To prove our claim, we first
observe that the automorphism group of R has order smaller than |A26 |.
Now, let C = CG (R). Then G/C is isomorphic to a subgroup of Aut(R),
and so |G/C| < |A26 |. It follows that C is not contained in R, so CR > R
and CR is a normal subgroup of G. Since G/R is simple, it follows that
CR = G. Let D = C ∩ R. Since D 6 R, we see that C centralizes D,
and since D 6 C, we see that R centralizes D, and thus D 6 Z(G). Now
C/D is isomorphic to A26 and D 6 Z(C). Since the Schur multiplier of
A26 has order 2, the only possibilities are that C ′ ∩ D has order 1 or 2.
If |C ′ ∩ D| = 1, then C ′ = A26 and C = D × A26 , and in this case, A26
is a direct factor of G. In the remaining case, C ′ = 2 · A26 , and in this
case, since |D| is 2 times some power of 3, it follows that 2 · A26 is a direct
factor of C and hence of G.
If G/R ∼
= S26 , then |R| = 27. Actually, we want to find (up to isomorphism) all groups G having a normal subgroup R of order 27 such that
G/R is isomorphic to S26 . Let M < G have index 2, where M/R is A26 .
Note that M is unique. Let C be the centralizer of R in M . Then RC is
normal in M , and since M/R is simple, there are only two possibilities:
RC = R or RC = M . We want to show that RC = M . Suppose RC = R.
Now M/C is embedded in Aut(R) and this group faithfully permutes the
26 non-identity elements of R. This action must be transitive or else M/C
is embedded in a direct product of the form Sa × Sb where a + b = 26 and
both a and b are positive. But |M/C| > |A26 | and
|Sa × Sb | = (a!)(b!) < (26!)/2.
This proves that the action is transitive, and it follows that R is elementary
abelian. Now
|Aut(R)| = |GL(3, 3)| < |A26 | 6 |M/C|.
This is a contradiction, so RC = M .
Let Z = Z(R) = R ∩ C. Then C/Z is isomorphic to M/R, which is
A26 . The Schur multiplier of A26 has order 2, and it follows that C ′ ∩ Z
8
is trivial, and thus C ′ is a normal subgroup of M isomorphic to A26 . It
follows that M is the direct product of R and C ′ . Let be write A = C ′ , so
A is isomorphic to A26 . Note that A is characteristic in M , so A is normal
in G. Now G/R ∼
= S26 and G/A has order 54, so G is a subdirect product
of G1 = S26 and G2 = G/A. (Note that, every group of order 54 has a
normal subgroup of index 2.) Actually, since each of these groups has a
unique homomorphism onto the group of order 2, it follows that G can
be constructed as follows: let π1 and π2 be the homomorphisms from G1
and G2 onto the group Z2 . Now, we consider the fibre product associated
with π1 and π2 , that is
G = {(a1 , a2 ) | π1 (a1 ) = π2 (a2 )}.
Clearly, A26 × P , where P is a group of order 27, is a normal subgroup of
G and the map π on G defined by
π((a1 , a2 )) = π1 (a1 ),
maps G onto Z2 with kernel A26 × P , so G/(A26 × P ) is isomorphic to Z2 .
This gives 15 groups, including all direct products of S26 with groups of
order 27.
This completes the proof of lemma and Theorem 1.7.
4
Non OD-characterizable alternating groups
We start this section with a result of M. A. Zvezdina [18, Theorem] which is
concerning simple groups whose prime graphs coincide with the prime graph of
an alternating simple group. More precisely, she proved:
Lemma 4.1 [18, Theorem] Let G be an alternating group An , n > 5, and let
S be a finite simple group. Then the prime graphs of G and S coincide if and
only if one of the following holds:
(a) G = A5 , S = A6 ;
(b) G = A6 , S = A5 ;
(c) G = A7 , S ∈ {L2 (49), U4 (3)};
(d) G = A9 , S ∈ {J2 , S6 (2), O8+ (2)};
(e) G = An , S = An−1 , n is odd, and the numbers n and n − 4 are composite.
Although the groups in the statement (e) of this lemma have the same prime
graph and so the same degree pattern, but they have different orders. In fact,
we have
|An | = |An−1 | × n.
9
Now, if we can choose the number n such that π(n) is contained in the set of
vertices of the prime graph Γ(An ) which are joined to all other vertices, then
the groups An and An−1 × H, where H is an arbitrary group of order n, have
the same order and degree pattern. This will enable us to give a positive answer
to problem 1.5.
Let G be a finite group satisfying |G| = |Am | = m!/2 and D(G) = D(Am ).
By [4, Lemma 2.15], the prime graph Γ(G) coincides with Γ(Am ). Simply, Γ(G)
is a graph with vertex set π(G) = {2, 3, 5, . . . , lm } in which two distinct vertices
r and s are joined by an edge iff r + s 6 m if r and s are odd primes and
r + 2 6 m − 2 if s = 2.
If △(m) 6 2 and p = lm , then we will deal with the alternating groups Ap ,
Ap+1 and Ap+2 , which are OD-charaterizable (see [10, Theorem 1.5]). Therefore,
we may consider the alternating groups Am for which △(m) > 3.
Referring to the already mentioned fact (Proposition 1.3) that all alternating
groups Am , 10 6= m 6 100, are OD-characterizable, we restrict our attention
to the alternating groups Am where m > 100. In what follows, we prove the
following result, which shows that there is an infinite family {Am } of alternating
groups such that hOD (Am ) > 3.
Proposition 4.2 Let m be an odd number satisfying △(m) > 4 and π(m) ⊆
π(△(m)!). Then hOD (Am ) > 2. In particular, if △(3n ) > 4 (resp. △(5n ) > 4),
then hOD (A3n ) > 16 (resp. hOD (A5n ) > 4).
Proof. First of all, it follows by Lemma 2.17 in [4] that π(△(m)!) ⊆ Λ(Am ),
this would mean that every vertex in π(△(m)!) and so in π(m) is adjacent
to all other vertices in π(Am ). Furthermore, since m is an odd number with
△(m) > 4, the prime graphs Γ(Am ) and Γ(Am−1 ) coincide by Lemma 4.1. Now,
if H is an arbitrary group of order m, then the groups Am and Am−1 × H have
the same order and degree pattern, and hence
hOD (Am ) > 1 + ν(m) > 2.
In the case when △(3n ) > 4, it is routine to check that n > 7, and so
hOD (A3n ) > 1 + ν(3n ) > 1 + νa (3n ) = 1 + |Par(n)| > 1 + |Par(7)| = 16.
Similarly, if △(5n ) > 4, then n > 3, and we obtain
hOD (A5n ) > 1 + ν(5n ) > 1 + νa (5n ) = 1 + |Par(n)| > 1 + |Par(3)| = 4.
The proof is complete.
Proof of Theorem 1.6. The result follows immediately from Lemma 2.7 and
Proposition 4.2. Note that, the proof of Proposition 4.2 shows that there is no
upper bound to hOD (Am ).
By Proposition 4.2, we can find many examples of alternating groups Am
with hOD (Am ) > 3. We point out here some of them.
10
(4.a) Some alternating groups Am , m 6 1000, with hOD (Am ) > 3.
In this case, we obtain the following simple groups amongst Am with
hOD (Am ) > 3 (see Table 1):
A125 , A147 , A189 , A539 , A625 , A875 ,
and for each of these groups, we have (see Table 1):
(1) hOD (A125 ) > 1 + ν(125) = 6, (see also [6]),
(2) hOD (A147 ) > 1 + ν(147) = 7,
(3) hOD (A189 ) > 1 + ν(189) = 14,
(4) hOD (A539 ) > 1 + ν(539) = 3,
(5) hOD (A625 ) > 1 + ν(625) = 16,
(6) hOD (A875 ) > 1 + ν(875) = 6.
Table 1.
m
m−4
lm
△(m)
125 = 53
121 = 112
113
147 = 3 · 72
143 = 11 · 13
189 = 3 · 7
539 = 72 · 11
3
625 = 5
4
875 = 53 · 7
π(△(m)!)
ν(m)
12
2, 3, 5, 7, 11
5
139
8
2, 3, 5, 7
6
185 = 5 · 37
181
8
2, 3, 5, 7
13
535 = 5 · 107
523
16
2, 3, 5, 7, 11, 13
2
621 = 3 · 23
619
6
2, 3, 5
15
871 = 13 · 67
863
12
2, 3, 5, 7, 11
5
3
We shall use Lemma 2.7 to show that there are infinite families of alternating groups An , with hOD (An ) > 3.
(4.b) An infinite family of alternating groups A3n , with hOD (A3n ) > 136.
The existence of infinite number of values n for which △(3n ) > 4 is immediate from Lemma 2.7. Indeed, if we take n to be a natural number such
that n ≡ 14 (mod 144), then it follows directly that {7, 17} ⊆ π(3n − 2)
and similarly {5, 19} ⊆ π(3n − 4), so the numbers 3n − 2 and 3n − 4 are
composite. This shows that △(3n ) > 8, for all n ≡ 14 (mod 144) (note
that, this provides an alternate proof of Lemma 2.7 for p = 3). Reasoning
as in the proof of preceding Proposition 4.2, we have
hOD (A3n ) > 1 + ν(3n ) > 1 + νa (3n ) = 1 + |Par(n)| > 1 + |Par(14)| = 136,
where Par(n) denotes the set of all partitions of n. In particular, we have
hOD (A314 ) > 136.
(4.c) An infinite family of alternating groups A5n , with hOD (A5n ) > 4.
By Lemma 2.7 again, there exist infinitely many values of n for which
△(5n ) > 4. Now for every such n, we have hOD (A5n ) > 4 by Proposition
4.2 (see also [7]).
11
5
On the symmetric groups Sm with hOD (Sm ) > 4
In this section we are looking for finite non-isomorphic groups having the same
order and degree pattern as a symmetric group. Suppose that G is a finite
group satisfying |G| = |Sm | = m! and D(G) = D(Sm ), for some natural
number m. First of all, we conclude from [4, Lemma 2.15] that the prime
graph Γ(G) coincides with Γ(Sm ). Actually, Γ(G) is a graph with vertex set
π(G) = {2, 3, 5, . . . , lm } in which two distinct vertices r and s are joined by an
edge iff r + s 6 m. In the case when △(m) 6 1, we deal with the symmetric
groups Sp and Sp+1 , which are OD-charaterizable by [10, Theorem 1.5]. We
now consider the symmetric groups Sm for which △(m) > 2, that is
Sp+2 , Sp+3 , Sp+4 , . . . , Sp+△(m) ,
where p = lm . On the other hand, in view of [4, Theorem 1.7] and Theorem
1.7, it follows that all symmetric groups Sm , where 10, 27 6= m 6 100, are 3-fold
OD-characterizable, and for this reason we now restrict our attention to the the
symmetric groups Sm where m > 100. Here, we first prove the following general
result:
Proposition 5.1 Let m be an odd number satisfying △(m) > 4 and π(m) ⊆
π(△(m)!). Then hOD (Sm ) > 4.
Proof. It follows by Lemma 2.17 in [4] that π(△(m)!) ⊆ Λ(Sm ), this would mean
that every vertex in π(△(m)!) and so in π(m) is adjacent to all other vertices in
π(Sm ). Moreover, since m is an odd number with △(m) > 4, the prime graphs
Γ(Sm ), Γ(Sm−1 ), Γ(Am ) and Γ(Am−1 ) coincide by Corollary 2.4 and Lemma
4.1 (see also [4, Table 7]). Now, if H and K are two arbitrary groups of order
m and 2m, respectively, then the groups Sm , Z2 × Am , Z2 · Am , Sm−1 × H,
(Z2 × Am−1 ) × H, (Z2 · Am−1 ) × H and Am−1 × K, have the same order and
degree pattern, and hence hOD (Sm ) > 4. The proof is complete.
Proof of Theorem 1.10. The result follows immediately from Lemma 2.7 and
Proposition 5.1. Note that, the proof of Proposition 5.1 shows that there is no
upper bound to hOD (Sm ).
Considering Proposition 5.1, we can now find many examples of symmetric
groups Sm satisfying hOD (Sm ) > 4. We point out here some of them.
(5.a) Some symmetric groups Sm , m 6 1000, with hOD (Sm ) > 4.
As before in (4.a), we can obtain the following symmetric groups amongst
Sm , which are k-fold OD-characterizable with k > 4 (see Table 1):
S125 , S147 , S189 , S539 , S625 , S875 .
The case S125 had already been studied in [6].
(5.b) There is an infinite family of symmetric groups Spn , with p ∈ {3, 5}, such
that hOD (Spn ) > 4.
12
Actually, reasoning as before in (4.b) and (4,c), there are an infinite number of values n for which △(pn ) > 4, and the result is now immediate from
Proposition 5.1 (see also [7]).
We conclude this paper with some comments for future works:
(1) Specify the exact value of hOD (An ) (resp. hOD (Sn )) for alternating (resp.
symmetric)) groups of degree n for n ∈ {125, 147, 189, 539, 625, 875}.
(2) It seems that it should be possible to prove that the value of hOD (An )
(resp. hOD (Sn )) is bounded by some function of n.
Acknowledgments
The author would like to express his sincere thanks to Prof. I. M. Isaacs for the
very helpful comments and suggestions to improve the contents and presentation of the manuscript. The author also would like to thank Professor Andrei
V. Zavarnitsine and Professor Robert Styer for their valuable comments and
discussions. Finally, he also wishes to thank Dr. A. Mahmoudifar for pointing
out references [6, 7] and for carefully reading the manuscript.
References
[1] M. Akbari and A. R. Moghaddamfar, Simple groups which are 2-fold ODcharacterizable, Bull. Malays. Math. Sci. Soc., 35 (1)(2012), 65–77.
[2] J. Bamberg, G. Cairns and K. D. Kilminster, The crystallographic restriction, permutations, and Goldbach’s conjecture, Amer. Math. Monthly,
110(3) (2003), 202–209.
[3] J. H. Conway, R. T. Curtis, S. P. Norton, R. A. Parker and R. A. Wilson,
Atlas of Finite Groups. Oxford, Clarendon Press, 1985.
[4] R. Kogani-Moghaddam and A. R. Moghaddamfar, Groups with the same
order and degree pattern, Sci. China Math., 55(4)(2012), 701–720.
[5] A. S. Kondratév, Prime graph components of finite simple groups. (Russian) Math. Sb., 180(6)(1989), 787–797.
[6] A. Mahmoudifar and B. Khosravi, On characterization by order and prime
graph for alternating groups, Siberian Math. J., 56(1)(2015), 125–131.
[7] A. Mahmoudifar and B. Khosravi, The answers to a problem and two conjectures about OD-characterization of finite groups, arXiv: 1409.7903v1,
28 September 2015.
[8] A. R. Moghaddamfar and S. Rahbariyan, More on the ODcharacterizability of a finite group, Algebra Colloq., 18(4)(2011), 663–674.
13
[9] A. R. Moghaddamfar and A. R. Zokayi, OD-Characterization of alternating
and symmetric groups of degrees 16 and 22, Front. Math. China, 4(4)(2009),
669–680.
[10] A. R. Moghaddamfar and A. R. Zokayi, Recognizing finite groups through
order and degree pattern, Algebra Colloq., 15 (3)(2008), 449–456.
[11] A. R. Moghaddamfar, A. R. Zokayi and M. R. Darafsheh, A characterization of finite simple groups by the degrees of vertices of their prime graphs,
Algebra Colloq., 12(3)(2005), 431–442.
[12] R. Remak, Über die Darstellung der endlichen Gruppen als Untergruppen
direkter Produkte, J. Reine Angew. Math., Bd 163, Heft 1 (1930), 1–44.
[13] A. V. Vasilév and E. P. Vdovin, An adjacency criterion in the prime graph
of a finite simple group, Algebra and Logic, 44(6)(2005), 381–406.
[14] J. S. Williams, Prime graph components of finite groups, J. Algebra,
69(2)(1981), 487–513.
[15] A. V. Zavarnitsine, Finite simple groups with narrow prime spectrum, Sib.
Elektron. Mat. Izv., 6 (2009), 1–12.
[16] A. V. Zavarnitsin and V. D. Mazurov, Element orders in coverings of symmetric and alternating groups, Algebra and Logic, 38(3)(1999), 159–170.
[17] L. C. Zhang, W. J. Shi, L. L. Wang and C. G. Shao, OD-Characterization
of A16 , Journal of Suzhou University (Natural Science Edition), 24 (2008),
7–10.
[18] M. A. Zvezdina, On nonabelian simple groups with the same prime graph
as an alternating group, Sib. Math. J., 54(1) (2013), 47–55.
14
| 4math.GR
|
A Constructor-Based Reachability Logic for
Rewrite Theories
arXiv:1709.05045v1 [cs.PL] 15 Sep 2017
Stephen Skeirik, Andrei Stefanescu and José Meseguer
University of Illinois at Urbana-Champaign, USA
Abstract. Reachability logic has been applied to K rewrite-rule-based
language definitions as a language-generic logic of programs. To be able
to verify not just code but also distributed system designs, a new rewritetheory-generic reachability logic is presented and proved sound for a wide
class of rewrite theories. The logic’s automation is increased by means of
constructor-based semantic unification, matching, and satisfiability procedures. New methods for proving invariants of possibly never terminating distributed systems are developed, and experiments with a prototype
implementation illustrating the new proof methods are presented.
Keywords: program verification, rewriting logic, reachability logic
1
Introduction
The main applications of reachability logic to date have been as a languagegeneric logic of programs [14,15]. In these applications, a K specification of a
language’s operational semantics by means of rewrite rules is assumed as the language’s “golden semantic standard,” and a correct-by-construction reachability
logic for a language so defined is automatically obtained [15]. This method has
been effective in proving reachability properties for a wide range of programs.
Although the foundations of reachability logic are very general [14,15], the existing theory does not provide straightforward answers to the following questions:
(1) Could a reachability logic be developed to verify not just conventional programs, but also distributed system designs and algorithms formalized as rewrite
theories in rewriting logic [8]? (2) If so, what would be the most natural way to
conceive such a rewrite-theory-generic logic? A satisfactory answer to questions
(1)–(2) would move the verification game from the level of verifying code to that
of verifying both code and distributed system designs. Since the cost of design
errors can be several orders of magnitude higher than that of coding errors,
answering questions (1) and (2) is of practical software engineering interest.
Although a first step towards a reachability logic for rewrite theories has been
taken in [6], as explained in Section 7 and below, that first step still leaves several
important questions open. The most burning one is how to prove invariants.
Since they are the most basic safety properties, support for proving invariants is
a sine qua non requirement. As explained below, a serious obstacle is what we
call the invariant paradox : we cannot verify in this manner any invariants of a
never-terminating system such as, for example, a mutual exclusion protocol.
2
A second open question is how to best take advantage of the wealth of equational reasoning techniques such as matching, unification, and narrowing modulo
an equational theory (Σ, E), and of recent results on decidable satisfiability of
quantifier-free formulas in initial algebras, e.g., [9] to automate as much as possible reachability logic deduction. In this regard, the very general foundations of
reachability logic —which assume any Σ-algebra A with a first-order-definable
transition relation— provide no help at all for automation. As shown in this
work and its prototype implementation, if we assume instead that the model
in question is the initial model TR of a rewrite theory R satisfying reasonable
assumptions, large parts of the verification effort can be automated.
A third important issue is simplicity. Reachability logic has eight inference
rules [14,15]. Could a reachability logic for rewrite theories be simpler? This work
tackles head on these three open questions to provide a general reachability logic
and a prototype implementation suitable for reasoning about properties of both
distributed systems and programs based on their rewriting logic semantics.
Rewriting Logic in a Nutshell. A distributed system can be designed and
modeled as a rewrite theory R = (Σ, E, R) [8] in the following way: (i) the
distributed system’s states are modeled as elements of the initial algebra TΣ/E
associated to the equational theory (Σ, E) with function symbols Σ and equations E; and (ii) the system’s concurrent transitions are modeled by rewrite rules
R, which are applied modulo E. Let us consider the QLOCK [5] mutual exclusion protocol, explained in detail in Section 2. QLOCK allows an unbounded
number of processes, which can be identified by numbers. Such processes can
be in one of three states: “normal” (doing their own thing), “waiting” for a resource, and “critical,” i.e., using the resource. Waiting processes enqueue their
identifier at the end of a waiting queue and can become critical when their name
appears at the head of the queue. A QLOCK state can be represented as a
tuple < n | w | c | q > where n, resp. w, resp. c, denotes the set of identifiers
for normal, resp. waiting, resp. critical processes, and q is the waiting queue.
QLOCK can be modeled as a rewrite theory R = (Σ, E, R), where E includes
axioms such as associativity-commutativity of multiset union, list associativity,
and identity axioms for ∅ and nil . QLOCK’s behavior is specified by five rewrite
rules R. Rule w2c below specifies a waiting process i becoming critical
w2c : < n | w i | c | i; q > → < n | w | c i | i; q > .
Reachability Logic in a Nutshell. A reachability logic formula has the form
A →⊛ B, with A and B state predicates (see Section 3). Assume for simplicity
that vars(A) ∩ vars(B) = ∅. Such a formula is then interpreted in the initial
model TR of a rewrite theory R = (Σ, E, R), whose states are E-equivalence
classes [u] of ground Σ-terms, and where a state transition [u] →R [v] holds iff
R ⊢ u → v according to the rewriting logic inference system [8] (computation =
deduction). As a first approximation, A →⊛ B is a Hoare logic partial correctness
assertion of the form {A}R{B}, but with the slight twist that B need not hold
of a terminating state, but just somewhere along the way. To be fully precise,
A →⊛ B holds in TR iff for each state [u0 ] satisfying A and each terminating
3
sequence [u0 ] →R [u1 ] . . . →R [un−1 ] →R [un ] there is a j, 0 ≤ j ≤ n such
that [uj ] satisfies B. A key question is how to choose a good language of state
predicates like A and B. Here is where the potential for increasing the logic’s
automation resides. We call our proposed logic constructor-based, because our
choice is to make A and B positive (only ∨ and ∧) combinations of what we call
constructor patterns of the form u | ϕ, where u is a constructor term1 and ϕ a
quantifier-free (QF) Σ-formula. The state predicate u | ϕ holds for a state [u′ ]
iff there is a ground substitution ρ such that [u′ ] = [uρ] and E |= ϕρ.
The Invariant Paradox. How can we prove invariants in such a reachability
logic? For example, mutual exclusion for QLOCK? Paradoxically, we cannot!
This is because QLOCK, like many other protocols, never terminates, that is,
has no terminating sequences whatsoever. And this has the ludicrous trivial
consequence that QLOCK’s initial model TR vacuously satisfies all reachability
formulas A →⊛ B. This of course means that it is in fact impossible to prove any
invariants using reachability logic in the initial model TR . But it does not mean
that it is impossible using some other initial model. In Section 4.1 we give a
systematic solution to this paradox by means of a simple theory transformation
allowing us to prove any invariant in the original initial model TR by proving an
equivalent reachability formula in the initial model of the transformed theory.
Our Contributions. Section 2 gathers preliminaries. The main theoretical contributions of a simple semantics and inference system for a rewrite-theory-generic
reachability logic with just two inference rules and its soundness are developed
in Sections 4 and 5. A systematic methodology to prove invariants by means
of reachability formulas is developed in Section 4.1. The goal of increasing the
logic’s potential for automation by making it constructor-based is advanced in
Sections 3–5. A proof of concept of the entire approach is given by means of
a Maude-based prototype implementation and a suite of experiments verifying
various properties of distributed system designs in Section 6. Related work and
conclusions are discussed in Section 7. Proofs can be found in [13].
2
Many-Sorted Algebra and Rewriting Logic
We present some preliminaries on many-sorted algebra and rewriting logic. For
a more general treatment using order-sorted algebra see [13]. Readers familiar
with many-sorted logic may go directly to Def. 1. We assume familiarity with the
following basic concepts and notation that are explained in full detail in, e.g.,
[10]: (i) many-sorted (MS) signature as a pair Σ = (S, Σ) with S a set of sorts
and Σ an S ∗ × S-indexed family Σ = {Σw,s }(w,s)∈S ∗×S of function symbols,
where f ∈ Σs1 ...sn ,s is displayed as f : s1 . . . sn → s; (ii) Σ-algebra A as a pair
A = (A, A ) with A = {As }s∈S an S-indexed family of sets, and A a mapping
interpreting each f : s1 . . . sn → s as a function in the set [As1 × . . .× Asn → As ].
(iii) Σ-homomorphism h : A → B as an S-indexed family of functions h = {hs :
1
That is, a term in a subsignature Ω ⊆ Σ such that each ground Σ-term is equal
modulo E to a ground Ω-term.
4
As → Bs }s∈S preserving the operations in Σ; (iv) the term Σ-algebra TΣ and
its initiality in the category MSAlgΣ of Σ-algebras when Σ is unambiguous.
An S-sorted set X = {Xs }s∈S of variables, satisfies s 6= s′ ⇒ Xs ∩ Xs′ = ∅,
and the variables in X are always assumed disjoint from all constants in Σ. The
Σ-term algebra on variables X, TΣ (X), is the initial algebra for the signature
Σ(X) obtained by adding to Σ the variables X as extra constants. Since a
Σ(X)-algebra is just a pair (A, α), with A a Σ-algebra, and α an interpretation
of the constants in X, i.e., an S-sorted function α ∈ [X →A], the Σ(X)-initiality
of TΣ (X) means that for each A ∈ MSAlgΣ and α ∈ [X →A], there exists a
unique Σ-homomorphism, α : TΣ (X) → A extending α, i.e., such that for each
s ∈ S and x ∈ Xs we have xαs = αs (x). In particular, when A = TΣ (Y ), an
interpretation of the constants in X, i.e., an S-sorted function σ ∈ [X →TΣ (Y )]
is called a substitution, and its unique homomorphic extension σ : TΣ (X) →
TΣ (Y ) is also
S called a substitution. Define dom(σ) = {x ∈ X | x 6= xσ}, and
ran(σ) = x∈dom(σ) vars(xσ). Given variables Z, the substitution σ|Z agrees
with σ on Z and is the identity elsewhere.
We also assume familiarity with many-sorted first-order logic including: (i)
the first-order language of Σ-formulas for Σ a signature (in our case Σ has only
function symbols and the = predicate); (ii) given a Σ-algebra A, a formula ϕ ∈
Form(Σ), and an assignment α ∈ [Y →A], with Y = fvars(ϕ) the free variables of
ϕ, the satisfaction relation A, α |= ϕ (iii) the notions of a formula ϕ ∈ Form(Σ)
being valid, denoted A |= ϕ, resp. satisfiable in a Σ-algebra A. For a subsignature
Ω ⊆ Σ and A ∈ MSAlgΣ , the reduct A|Ω ∈ MSAlgΩ agrees with A in the
interpretation of all sorts and operations in Ω and discards everything in Σ \ Ω.
If ϕ ∈ Form(Ω) we have the equivalence A |= ϕ ⇔ A|Ω |= ϕ.
An MS equational theory is a pair T = (Σ, E), with E a set of (possibly
conditional) Σ-equations. MSAlg(Σ,E) denotes the full subcategory of MSAlgΣ
with objects those A ∈ MSAlgΣ such that A |= E, called the (Σ, E)-algebras.
MSAlg(Σ,E) has an initial algebra TΣ/E [10]. The inference system in [10] is
sound and complete for MS equational deduction, i.e., for any MS equational
theory (Σ, E), and Σ-equation u = v we have an equivalence E ⊢ u = v ⇔
E |= u = v. For the sake of simpler inference we assume non-empty sorts,
i.e., ∀s ∈ S TΣ , s 6= ∅. Deducibility E ⊢ u = v is abbreviated as u =E v,
called E-equality. An E-unifier of a system of Σ-equations, i.e., a conjunction
φ = u1 = v1 ∧ . . . ∧ un = vn of Σ-equations is a substitution σ such that
ui σ =E vi σ, 1 ≤ i ≤ n. An E-unification algorithm for (Σ, E) is an algorithm
generating a complete set of E-unifiers Unif E (φ) for any system of Σ equations
φ, where “complete” means that for any E-unifier σ of φ there is a τ ∈ Unif E (φ)
and a substitution ρ such that σ =E (τ ρ)|dom (σ)∪dom(τ ) , where =E here means
that for any variable x we have xσ =E x(τ ρ)|dom (σ)∪dom(τ ) . The algorithm is
finitary if it always terminates with a finite set Unif E (φ) for any φ.
We recall some basic concepts about rewriting logic. The survey in [8] gives
a fuller account. A rewrite theory R axiomatizes a distributed system, so that
concurrent computation is modeled as concurrent rewriting with the rules of
R modulo the equations of R. Recall also the following notation from [3]: (i)
5
positions in a term viewed as a tree are marked by strings p ∈ N∗ specifying a
path from the root, (ii) t|p denotes the subterm of term t at position p, and (iii)
t[u]p denotes the result of replacing subterm t|p at position p by u.
Definition 1. A rewrite theory is a 3-tuple R = (Σ, E ∪ B, R) with (Σ, E ∪ B)
an MS equational theory and R a set of conditional Σ-rewrite rules l → r if φ,
with l, r ∈ TΣ (X)s for some s ∈ S, and φ a quantifier-free Σ-formula. We further
assume that: (1) Each equation u = v ∈ B is regular, i.e., vars(u) = vars(v),
and linear, i.e., there are no repeated variables in either u or v. (2) The equations
~ = {u → v if ψ | u = v if ψ ∈
E, when oriented as conditional rewrite rules E
E}, are convergent modulo B, i.e., strictly coherent, confluent, and operationally
terminating as rewrite rules modulo B [7]. (3) The rules R are ground coherent
with the equations E modulo B [4].
Conditions (1)–(2) ensure that the initial algebra TΣ/E∪B is isomorphic to
the canonical term algebra CΣ/E,B , whose elements are B-equivalence classes
~ B-irreducible ground Σ-terms. Define the one-step R, B-rewrite relation
of E,
t →R,B t′ between ground terms as follows. For t, t′ ∈ TΣ,s , s ∈ S, t →R,B t′
holds iff there is a rewrite rule l → r if φ ∈ R, a ground substitution σ ∈ [Y →TΣ ]
with Y the rule’s variables, and a term position p in t such that t|p =B lσ,
t′ = t[rσ]p , and E ∪ B |= φσ. In the context of (1)–(2), condition (3) ensures
~ B-canonical forms before performing R, B-rewriting” is a
that “computing E,
complete strategy. That is, if t →R,B t′ and u = t!E,B , i.e., t →∗E,B
u with u in
~
~
E, B-canonical form (abbreviated in what follows to u = t!), then there exists a
u′ such that u →R,B u′ and t′ ! =B u′ !. Note that vars(r) ⊆ vars(l) is nowhere
assumed for rules l → r if φ ∈ R. This means that R can specify an open
system, in the sense of [11], that interacts with an external, non-deterministic
environment such as, for example, a thermostat.
Conditions (1)–(3) allow a simple description of the initial reachability model
TR [8] of R as the canonical reachability model CR whose states belong to the
canonical term algebra CΣ/E,B , and the one-step transition relation [u] →R [v]
holds iff u →R,B u′ and [u′ !] = [v]. Furthermore, if u →R,B u′ has been performed
with a rewrite rule l → r if φ ∈ R and a ground substitution σ ∈ [Y →TΣ ], then,
assuming B-equality is decidable, checking whether condition E ∪ B |= φσ holds
~ B-canonical form.
is decidable by reducing the terms in φσ to E,
A Running Example. Consider the following rewrite theory R = (Σ, E ∪B, R)
modeling a dynamic version of the QLOCK mutual exclusion protocol [5], where
(Σ, B) defines the protocol’s states, involving natural numbers, lists, and multisets over natural numbers. Σ has sorts S = {Nat, List , MSet, Conf , State, Pred }
with subsorts2 Nat < List and Nat < MSet and operators F = {0 : → Nat , s :
: MSet MSet → MSet , ; :
Nat → Nat, ∅ : → MSet, nil : → List ,
List List → List , dupl : MSet → Pred , tt : → Pred , < > : Conf → State, | | | :
2
As pointed out at the beginning of Section 2, [13] treats the more general order-sorted
case, where sorts form a poset (S, ≤) with s ≤ s′ interpreted as set containment
As ⊆ As′ in a Σ-algebra A.
6
MSet MSet MSet List → Conf }, where underscores denote operator argument
placement. The axioms B are the associativity-commutativity of the multiset
with identity ∅, and the associativity of list concatenation ; with
union
identity nil . The only equation in E is dupl (s i i) = tt. It defines the dupl predicate by detecting a duplicated element i in the multiset s i i (s could be empty).
States of QLOCK are B-equivalence classes of ground terms of sort State.
QLOCK [5] is a mutual exclusion protocol where the number of processes is
unbounded. Furthermore, in the dynamic version of QLOCK presented below,
such a number can grow or shrink. Each process is identified by a number. The
system configuration has three sets of processes (normal, waiting, and critical)
plus a waiting queue. To ensure mutual exclusion, a normal process must first
register its name at the end of the waiting queue. When its name appears at
the front of the queue, it is allowed to enter the critical section. The first three
rewrite rules in R below specify how a normal process i first transitions to a
waiting process, then to a critical process, and back to normal. The last two
rules in R specify how a process can dynamically join or exit the system.
n2w
w2c
c2n
join
exit
:<n i |w |c |q
> → <n |w i |c |q ;i
:<n |w i |c |i ;q > → <n |w |c i |i ;q
:<n |w |c i |i ;q > → <n i |w |c |q
:<n |w |c |q
> → <n i |w |c |q
:<n i |w |c |q
> → <n |w |c |q
>
>
>
> if φ
>
where φ ≡ dupl (n i w c) 6= tt, i is a number, n, w , and c are, respectively, normal,
waiting, and critical process identifier sets, and q is a queue of process identifiers.
It is easy to check that R = (Σ, E ∪ B, R) satisfies requirements (1)–(3). Note
that join makes QLOCK an open system in the sense explained above.
3
Constrained Constructor Pattern Predicates
Given an MS equational theory (Σ, E ∪B), the atomic state predicates appearing
in the constructor-based reachability logic formulas of Section 4 will be pairs
u | ϕ, called constrained constructor patterns, with u a term in a subsignature
Ω ⊆ Σ of constructors, and ϕ a quantifier-free Σ-formula. Intuitively, u | ϕ is
a pattern describing the set of states that are EΩ ∪ BΩ -equal to ground terms
of the form uρ for ρ a ground constructor substitution such that E ∪ B |= ϕρ.
Therefore, u | ϕ can be used as a symbolic description of a, typically infinite, set
of states in the canonical reachability model CR of a rewrite theory R.
Often, the signature Σ on which TΣ/E∪B is defined has a natural decomposition as a disjoint union Σ = Ω ⊎ ∆, where the elements of the canonical term
algebra CΣ/E,B are Ω-terms, whereas the function symbols f ∈ ∆ are viewed
~ B-simplification. Ω (with
as defined functions which are evaluated away by E,
same poset of sorts as Σ) is then called a constructor subsignature of Σ.
~
A decomposition of a MS equational theory (Σ, E ∪ B) is a triple (Σ, B, E)
~ are convergent modulo B. (Σ, B, E)
~ is called sufficiently
such that the rules E
complete with respect to the constructor subsignature Ω iff for each t ∈ TΣ
7
we have: (i) t!E,B
∈ TΩ , and (ii) if u ∈ TΩ and u =B v, then v ∈ TΩ . This
~
ensures that for each [u]B ∈ CΣ/E,B we have [u]B ⊆ TΩ . Sufficient completeness
is closely related to the notion of a protecting inclusion of decompositions.
Definition 2. Let (Σ0 , E0 ∪ B0 ) ⊆ (Σ, E ∪ B) be a theory inclusion such that
~ are respective decompositions of (Σ0 , E0 ∪ B0 ) and
(Σ0 , B0 , E~0 ) and (Σ, B, E)
~ protects (Σ0 , B0 , E~0 )
(Σ, E ∪ B). We then say that the decomposition (Σ, B, E)
′
′
iff (i) for all t, t ∈ TΣ0 (X) we have: (i) t =B0 t ⇔ t =B t′ , (ii) t = t!E~0 ,B0 ⇔
t = t!E,B
~ , and (iii) CΣ0 /E0 ,B0 = CΣ/E,B |Σ0 .
~ iff (i) (Σ, B, E)
~
(Ω, BΩ , E~Ω ) is a constructor decomposition of (Σ, B, E)
~
~
protects (Ω, BΩ , EΩ ), and (ii) (Σ, B, E) is sufficiently complete with respect to
the constructor subsignature Ω. Furthermore, Ω is called a subsignature of free
constructors modulo BΩ iff EΩ = ∅, so that CΩ/EΩ ,BΩ = TΩ/,BΩ .
We are now ready to define constrained constructor pattern predicates.
~ A
Definition 3. Let (Ω, BΩ , E~Ω ) be a constructor decomposition of (Σ, B, E).
constrained constructor pattern is an expression u | ϕ with u ∈ TΩ (X) and
ϕ a QF Σ-formula. The set PatPred (Ω, Σ) of constrained constructor pattern
predicates contains ⊥ and the set of constrained constructor patterns, and is
closed under disjunction (∨) and conjunction (∧). Capital letters A, B, . . . , P, Q, . . .
range over PatPred (Ω, Σ). The semantics of a constrained constructor pattern
predicate A is a subset JAK ⊆ CΣ/E,B defined inductively as follows:
1.
2.
3.
4.
J⊥K = ∅
Ju | ϕK = {[(uρ)!]BΩ ∈ CΣ/E,B | ρ ∈ [X →TΩ ] ∧ E ∪ B |= ϕρ}.
JA ∨ BK = JAK ∪ JBK
JA ∧ BK = JAK ∩ JBK.
Note that for any constructor pattern predicate A, if σ is a (sort-preserving)
bijective renaming of variables we always have JAK = JAσK. Given constructor
patterns u | ϕ and v | ψ with vars(u | ϕ) ∩ vars(v | ψ) = ∅, we say that u | ϕ
subsumes v | ψ iff there is a substitution α such that: (i) v =EΩ ∪BΩ uα, and (ii)
TE∪B |= ψ ⇒ (ϕα). It then follows easily from the above definition
W of Ju | ϕK
that if u | ϕ subsumes v | ψ, then Jv | ψK ⊆ Ju | ϕK. Likewise, i∈I ui | ϕi
subsumes v | ψ iff there is a k ∈ I such that uk | ϕk subsumes v | ψ.
Pattern Predicate Example. Letting n, w , c be multisets of process identifiers
and q be an associative list of process identifiers, recall that QLOCK states have
the form < n | w | c | q >. From the five rewrite rules defining QLOCK, it is
easy to prove that if < n | w | c | q > →∗ < n ′ | w ′ | c ′ | q ′ > and n w c is a set
(has no repeated elements), then n′ w′ c′ is also a set. Of course, it seems very
reasonable to assume that these process identifier multisets are, in fact, sets,
since otherwise we could, for example, have a process i which is both waiting
and critical at the same time. We can rule out such ambiguous states by means
of the pattern predicate < n | w | c | q > | dupl (n w c) 6= tt.
If EΩ ∪ BΩ has a finitary unification algorithm, any constrained constructor
W
pattern predicate A is semantically equivalent to a finite disjunction i ui | ϕi
8
of constrained constructor patterns. This is because: (i) by (3)–(4) in Def. 3
we may assume A in disjunctive
normal form; and (ii) it is easy to check that
S
J(u | ϕ) ∧ (v | φ)K = α∈Unif E ∪B (u,v) Juα | (ϕ ∧ φ)αK, were we assume that
Ω
Ω
vars(u | ϕ) ∩ vars(v | ψ) = ∅, and that all variables in ran(α) are fresh. Pattern
intersection can also be defined when u | ϕ and v | φ share parameters Y =
vars(u | ϕ) ∩ vars(v | φ) = vars(u) ∩ vars(v). [13] defines in detail the notions
of parametric intersection Ju | ϕK ∩Y Jv | φK and of parametric subsumption
v | φ ⊆Y u | ϕ of patterns. These notions are very useful to reason about
parameterized invariants and co-invariants (see Section 4.1 and [13]).
4
Constructor-Based Reachability Logic
The constructor-based reachability logic we define is a logic to reason about
reachability properties of the canonical reachability model CR of a topmost
rewrite theory R where “topmost” intuitively means all rewrites must occur at
the top of the term.3 Many rewrite theories of interest, including those specifying distributed object-oriented systems or the semantics of (possibly concurrent)
programming languages, can be easily made topmost by a theory transformation
(see, e.g., [16]). Formally, we require R = (Σ, E ∪ B, R), besides satisfying the
requirements in Definition 1, also satisfies:
~ and a constructor
1. (Σ, E ∪ B) has a sort State, a decomposition (Σ, B, E),
~
decomposition (Ω, BΩ , EΩ ) where: (i) ∀u ∈ TΩ (X)State , vars(u) = vars(u!);
(ii) BΩ are linear and regular with a finitary EΩ ∪ BΩ -unification algorithm.
2. Rules in R have the form l → r if ϕ with l ∈ TΩ (X). Furthermore, they are
topmost in the sense that: (i) for all such rules, l and r have sort State, and (ii)
for any u ∈ TΩ (X)State and any non-empty position p in u, u|p 6∈ TΩ (X)State .
Requirements (1)–(2) ensure that in the canonical reachability model CR if
[u] →R [v] holds, then the R, B-rewrite u →R,B u′ such that [u′ !] = [v] happens
at the top of u, i.e., uses a rewrite rule l → r if ϕ ∈ R and a ground substitution
σ ∈ [Y →TΩ ], with Y the rule’s variables, such that u =BΩ lσ and u′ = rσ.
We are now ready to define the formulas of our constructor-based reachability
logic for R satisfying above requirements (1)–(2). Let PatPred (Ω, Σ)State denote
the subset of PatPred (Ω, Σ) determined by those pattern predicates A such
that, for all atomic constrained constructor predicates u | ϕ appearing in A, u
has sort State. Reachability logic formulas then have the form: A →⊛ B, with
A, B ∈ PatPred (Ω, Σ)State . The parameters Y of A →⊛ B are the variables in
the set Y = vars(A)∩vars(B), and A →⊛ B is called unparameterized iff Y = ∅.
The reachability logic in [14,15] is based on terminating sequences of state
transitions; when there are no terminating states, all reachability formulas are
vacuously true. Our purpose is to extend the logic in order to verify properties
of general distributed systems specified as rewrite theories R which may never
3
Topmost theories have reachability completeness for narrowing [16]. Our inference
system uses narrowing to symbolically compute successor states in CR .
9
terminate. For this, as explained in Section 4.1, we generalize the all-paths satisfaction relation in [15], which for a theory R we denote by R |=∀ A →⊛ B, to
a relativized satisfaction relation R |=∀T A →⊛ B, where T is a constrained
pattern predicate such that JT K is a set of terminating states. That is, let
Term R = {[u] ∈ CR,State | (6 ∃[v]) [u] →R [v]}. We then require JT K ⊆ Term R .
The standard relation R |=∀ A →⊛ B is then recovered as the special case where
JT K = Term R . Call [u] →∗R [v] a T -terminating sequence iff [v] ∈ JT K.
Definition 4. Given T with
JT K ⊆ Term R , the all-paths satisfaction relaW
tion R |=∀T u | ϕ →⊛ j∈J vj | φj asserts the satisfaction of the formula
W
u | ϕ →⊛ j∈J vj | φj in the canonical reachability model CR of a rewrite
theory R satisfyingWtopmost requirements (1)–(2). It is defined as follows:
W
For u | ϕ →⊛ j∈J vj | φj unparameterized, R |=∀T u | ϕ →⊛ j∈J vj | φj
holds iff for each T -terminating sequence [u0 ] →R [u1 ] . . . [un−1 ] →R [un ] with
[u0 ] ∈ Ju |WϕK there exist k, 0 ≤ k ≤ n and j ∈ J such that W
[uk ] ∈ Jvj | φj K. For
u | ϕ →⊛ j∈J vj | φj with parameters Y , R |=∀T u | ϕ →⊛ j∈J vj | φj holds if
W
R |=∀T (u | ϕ)ρ →⊛ ( j∈J vj | φj )ρ holds for each ρ ∈ [Y →TΩ ].
Since a constrained pattern predicate is equivalent to a disjunction of atomic
ones, we
Wcan define satisfaction onVgeneral reachability logic formulas as follows:
R |=∀T 1≤i≤n ui | ϕi →⊛ A iff 1≤i≤n R |=∀T ui | ϕi →⊛ A, assuming same
parameters Yi = vars(ui | ϕi ) ∩ vars(A), i.e., Yi = Yi′ for 1 ≤ i < i′ ≤ n.
R |=∀T A →⊛ B is a partial correctness assertion: If state [u] satisfies
“precondition” A, then “postcondition” B is satisfied somewhere along each
T -terminating sequences from [u], generalizing a Hoare formula {A}R{B} [13].
Recall that rewrite rules l → r if φ are assumed to have l ∈ TΩ (X). For
symbolic reasoning purposes it will be very useful to also require that r ∈ TΩ (X).
This can be achieved by a theory transformation R 7→ R̂. Stated formally, if
R = (Σ, E ∪ B, R), then R̂ = (Σ, E ∪ B, R̂), where the rules R̂ are obtained
from the rules R by transforming each l → r if φ in R into the rule l → r′ if φ∧ θ̂,
where: (i) r′ is the Ω-abstraction of r obtained by replacing each length-minimal
position p of r such that t|p 6∈ TΩ (X) by a fresh variable xp whose sort is the
V
least sort of t|p , (ii) θ̂ = p∈P xp = tp , where P is the set of all length-minimal
positions in r such that t|p 6∈ TΩ (X).
The key semantic property about this transformation is:
Lemma 1. The canonical reachability models CR and CR̂ are identical.
4.1
Invariants, Co-Invariants, and Never-Terminating Systems
The notion of an invariant applies to any transition system S = (S, →S ) with
states S and transition relation →S ⊆ S×S. The set Reach(S0 ) of states reachable
from S0 ⊆ S is defined as Reach(S0 ) = {s ∈ S | (∃s0 ∈ S0 ) s0 →∗S s}, where →∗S
denotes the reflexive-transitive closure of →S . An invariant about S with initial
states S0 can be specified in two ways: (i) by a “good” property P ⊆ S, the
invariant, that always holds from S0 , i.e., such that Reach(S0 ) ⊆ P , or (ii) as
10
a “bad” property Q ⊆ S, the co-invariant, that never holds from S0 , i.e., such
that Reach(S0 ) ∩ Q = ∅. Obviously, P is an invariant iff S \ P is a co-invariant.
Suppose we have specified a distributed system by a topmost rewrite theory
R, and constrained pattern predicates S0 and P , and we want to prove that
JP K is an invariant of the system (CR,State , →R ) from JS0 K. Can we specify such
invariant or co-invariant by means of reachability formulas and use the inference
system of Section 5 to try to prove such formulas?
The answer to the above question is not obvious. Suppose R specifies a neverterminating system, i.e., Term R = ∅. For example, QLOCK and other mutual
exclusion protocols are never-terminating. Then, no reachability formula can
characterize and invariant holding by means of the satisfaction relation R |=∀T
A →⊛ B. The reason for this impossibility is that, since Term R = ∅, R |=∀T
A →⊛ B holds vacuously for all reachability formulas A →⊛ B.
Is then reachability logic useless to prove invariants? Definitely not. We need
to first perform a simple theory transformation. Call an invariant specifiable by
constrained pattern predicates S0 and P if JP K is an invariant of (CR,State , →R )
from JS0 K. To ease the exposition, we explain the transformation for the case
where Ω has a single state constructor operator, say, h , . . . , i : s1 , . . . , sn →
State. The extension to several such operators is straightforward. The theory
transformation is of the form R 7→ Rstop , where Rstop is obtained from R by
just adding: (1) a new state constructor operator [ , . . . , ] : s1 , . . . , sn → State
to Ω, and (2) a new rewrite rule stop : hx1 :s1 , . . . , xn :sn i → [x1 :s1 , . . . , xn :sn ]
to R. Also, let [ ] denote the pattern predicate [x1 :s1 , . . . , xn :sn ] | ⊤. Likewise,
for any atomic constrained pattern predicate B = hu1 , . . . , un i | ϕ we define
the pattern predicate [B] = [u1 , . . . , un ] | ϕ and extend this notation to any
union Q of atomic predicates. Since h , . . . , i : s1 , . . . , sn → State is the only
state constructor, we can assume without loss of generality that any atomic
constrained pattern predicate in R is semantically equivalent to one of the form
hu1 , . . . , un i | ϕ. Likewise, any pattern predicate will be semantically equivalent
to a union of atomic predicates of such form, called in standard form.
Theorem 1. For S0 , P ∈ PatPred (Ω, Σ) constrained pattern predicates in standard form with vars(S0 ) ∩ vars(P ) = ∅, JP K is an invariant of (CR,State , →R )
from JS0 K iff Rstop |=∀[ ] S0 →⊛ [P ].
The notion of a parametric invariant can be reduced to the unparameterized
one: if Y = vars(S0 ) ∩ vars(P ), then JP K is an invariant of (CR,State , →R )
from JS0 K with parameters Y iff Rstop |=∀[ ] S0 →⊛ [P ]. That is, iff JP ρK is an
(unparameterized) invariant of (CR,State , →R ) from JS0 ρK for each ρ ∈ [Y →TΩ ].
In this way, Theorem 1 extends to parametric invariants.
Specifying Invariants for QLOCK. Consider the QLOCK specification from
Sections 2 and 3. QLOCK nonterminating is never terminating. However, we
can apply the theory transformation in Theorem 1 by adding an operator [ ] :
Conf → State and a rule stop : < t > → [t ] for t : Conf . Define the set of
initial states by the pattern predicate S0 = < n ′ | ∅ | ∅ | nil > | dupl (n′ ) 6= tt.
Since QLOCK states have the form < n | w | c | q >, mutual exclusion means
11
|c| ≤ 1, which is expressible by the pattern predicate < n | w | i | i ; q > ∨
< n | w | ∅ | q >. But we need also to ensure our multisets are actually sets.
Thus, the pattern predicate P = < n | w | i | i ; q > | dupl (n w i ) 6= tt ∨
< n | w | ∅ | q > | dupl (n w ) 6= tt specifies mutual exclusion. By Theorem 1,
QLOCK ensures mutual exclusion from JS0 K iff Rstop |=∀[ ] S0 →⊛ [P ].
The following easy corollary can be very helpful in proving invariants. It can,
for example, be applied to prove the mutual exclusion of QLOCK.
Corollary 1. Let S0 , P ∈ PatPred (Ω, Σ) be constrained pattern predicates in
standard form with vars(S0 )∩vars(P ) = ∅. JP K is an invariant of (CR,State , →R )
from JS0 K if: (i) S0 ⊆ P , and (ii) Rstop |=∀[ ] P →⊛ [P σ], where σ is a sortpreserving bijective renaming of variables such that vars(P ) ∩ vars(P σ) = ∅.
Corollary 1 can be extended to parametric invariants (see [13]). The treatment of co-invariants is similar and can also be found in [13].
5
A Sound Inference System
We present our inference system for all-path reachability for any R satisfying
topmost requirements (1)–(2), with rules R = {lj → rj if φj }j∈J such that
lj , rj ∈ TΩ (X), j ∈ J. Variables of rules in R are always assumed disjoint
from variables in reachability formulas; this can be ensured by renaming. The
inference system has two proof rules. The Step∀ + Subsumption proof rule
allows taking one step of (symbolic) rewriting along all paths according to the
rules in R. The Axiom proof rule allows the use of a trusted reachability formula
to summarize multiple rewrite steps, and thus to handle repetitive behavior.
W
These proof rules derive sequents of the form [A, C] ⊢T u | ϕ −→⊛ i vi | ψi ,
where A and C are finite sets of reachability formulas and T a pattern predicate
defining a set of T -terminating ground states. Formulas in A are called axioms
and those in C are called circularities.
We furthermore assume that in all reachW
ability formulas u | ϕ −→⊛ i vi | ψi we have vars(ψi ) ⊆ vars(vi ) ∪ vars(u | ϕ)
for each i. According to the implicit quantification of the semantic relation |=∀T
this means that any variable in ψi is either universally quantified and comes
from the precondition u | ϕ, or is existentially quantified and comes from vi
only. This property is an invariant preserved by the two inference rules.
Proofs always begin with a set C of formulas that we want to simultaneously
prove, so that the proof effort only succeeds if all formulas in C are eventually
proved. C contains the main properties we want to prove as well as any auxiliary
lemmas that may be needed to carry out the proof. The initial set of goals
we want to prove is [∅,WC] ⊢T C, which is a W
shorthand for the set of goals
{[∅, C] ⊢T u | ϕ −→⊛ i vi | ψi (u | ϕ −→⊛ i vi | ψi ) ∈ C}. Thus, we start
without any axioms A, but we shall be able to use the formulas in C as axioms
in their own derivation after taking at least on step with the rewrite rules
W in R.
A very useful feature is that sequents [∅, C] ⊢T u | ϕ −→⊛ i vi | ψi ,
whose formulas C have been postulated (as the conjectures to be proved), are
transformed by Step∀ + Subsumption into sequents of the form [C, ∅] ⊢T
12
W
u′ | ϕ′ −→⊛ i vi′ | ψi′ , where now the formulas in C can be assumed valid, and
can be used in derivations with the Axiom rule.
Verifying QLOCK’s Mutual Exclusion. By Corollary 1, QLOCK’s mutual
exclusion can be verified by: (i) using pattern subsumption to check the trivial
inclusion JS0 K ⊆ JP K, and (ii) proving Rstop |=∀[ ] P σ →⊛ [P ], where σ is a sortpreserving bijective renaming of variables such that vars(P ) ∩ vars(P σ) = ∅.
But, since for QLOCK, P is a disjunction, in our inference system this means
proving from Rstop that [∅, C] ⊢[] C, where C are the conjectures:
< n ′ | w ′ | i ′ | i ′ ; q ′ > | ϕ′ →⊛ [< n | w | i | i ; q > | ϕ ∨ < n | w | ∅ | q > | ψ]
< n ′ | w ′ | ∅ | q ′ > | ψ ′ →⊛ [< n | w | i | i ; q > | ϕ ∨ < n | w | ∅ | q > | ψ].
where ϕ ≡ dupl (n w i) 6= tt, ψ ≡ dupl (n w ) 6= tt, and ϕ′ , ψ ′ are their obvious
renamings.
Before explaining the Step∀ + Subsumption proof rule we W
introduce some
notational conventions. Assume T is the pattern predicate T = j tj | χj , with
vars(χj ) ⊆ vars(tj ), and let R = {lj → rj if φj }j∈J , we then define:
match(u, {vi }i∈I ) ⊆ {(i, β) | β ∈ [vars (vi ) \ vars(u) → TΩ (X)] s.t. u =EΩ ∪BΩ vi β}
a complete set of (parameter-preserving) EΩ ∪ BΩ -matches of u against the vi ,
unify(u | ϕ′ , R) ≡ {(j, α) | α ∈ UnifEΩ ∪BΩ (u, lj ) and (ϕ′ ∧φj )α satisfiable in TΣ/E∪B }
a complete set of EΩ ∪ BΩ -unifiers of a pattern u | ϕ′ with the lefthand-sides
of the rules in R with satisfiable associated constraints.4 Consider now the rule:
Step∀ + Subsumption
^
[A ∪ C, ∅] ⊢T (rj | ϕ′ ∧ φj )α −→⊛
i
(j,α)∈unify(u|ϕ′ , R)
[A, C] ⊢T u | ϕ −→⊛
_
(vi | ψi )α
_
vi | ψi
i
V
where ϕ′ ≡ ϕ ∧ (i,β)∈match(u, {vi }) ¬(ψi β). This inference rule allows us to
take one step with the rules in R. Intuitively, u | ϕ′ characterizes the states
satisfying u | ϕ that are not subsumed by any vi | ψi ; that is, states in the
lefthand side of the current goal that have not yet reached the righthand side.
4
In the current prototype implementation (see Section 6), variant satisfiability makes
~ convergent modconstraint checking decidable. Future versions will only assume E
ulo B for the equational part E ∪ B of R, so that satisfiability of such constraints will in general be undecidable. Unifiers whose associated constraints cannot
be proved unsatisfiable will then be included in unify(u | ϕ′ , R) as a safe overapproximation. The same approach will apply to the, in general undecidable, checking of satisfiability/validity for other constraints involved in the application of the
Step∀ +Subsumption or Axiom rules below: they will be either over-approximated,
or will become proof obligations to be discharged by an inductive theorem prover.
13
W
Note that, according
to Definition 4, u | ϕ −→⊛ i vi | ψi is semantically valid
W
iff u | ϕ′ −→⊛ i vi | ψi is valid. Thus, this inference rule only unifies u | ϕ′
with the lefthand
W sides of rules in R. We impose on this inference rule a side
condition that j,γ∈UnifE ∪B (u,tj ) (ϕ′ ∧ χj )γ is unsatisfiable in TΣ/E∪B , where
Ω
Ω
W
T = j tj | χj is the pattern predicate characterizing the chosen T -terminating
states. This condition ensures that any state in u | ϕ′ has an R-successor.
Thus,
W
a state in u | ϕ′ reaches on all T -terminating paths a state in i vi | ψi if all
its successors do so. Each R-successor is covered by one of (rj | ϕ′ ∧ φj )α. As
an optimization, we check that (ϕ′ ∧ φj )α is satisfiable and we Wdrop the ones
which are not. Finally, we also
W assume that vars((u | ϕ)α)∩vars(( i vi | ψi )α) =
vars((rj | ϕ′ ∧ φj )α)∩vars (( i vi | ψi )α). This parameter preservation condition
ensures correct implicit quantification. Note that formulas in C are added to A,
so that from now on they can be used by Axiom. By using EΩ ∪ BΩ -unification,
this inference rule performs narrowing of u | ϕ′ with rules R [16].
Axiom
^
j
[{u′ | ϕ′ −→⊛
_
vj′ | ψj′ } ∪ A, ∅] ⊢T vj′ α | ϕ ∧ ψj′ α −→⊛
vi | ψi
i
j
[{u′ | ϕ′ −→⊛
_
_
j
vj′ | ψj′ } ∪ A, ∅] ⊢T u | ϕ −→⊛
_
vi | ψi
i
if ∃α such that u =EΩ ∪BΩ u′ α and TΣ/E∪B |= ϕ ⇒ ϕ′ α. This inference rule
allows us to use a trusted formula in A to summarize multiple transition steps.
This is similar to how several transition steps would apply to a ground term,
except that for ground terms we would check that ϕ′ α is valid, whereas here we
check that the condition ϕ implies ϕ′ α. Since ϕ is stronger
ϕ′ α, we add ϕ
W ′ than
′
′
′
′
′
⊛
to (vj | ψj )α (the result of using axiom u | ϕ −→
j vj | ψj ). We assume that
W
W
u | ϕ −→⊛ i vi | ψi and u′ | ϕ′ −→⊛ j vj′ | ψj′ do not share variables, which
can always be guaranteed by renaming. For correct implicit quantification, as
in Step∀ + Subsumption, W
we assume for each j the parameter preservation
W
condition vars(u | ϕ)∩vars ( i vi | ψi ) = vars(vj′ α | ϕ ∧ ψj′ α)∩vars ( i vi | ψi ).
On a practical note, in order to be able to find the α, our implementation requires
that vars(ϕ′ ) ⊆ vars(u′ ), so that all the variables in vars(ϕ′ ) are matched.
The soundness of Step∀ + Subsumption plus Axiom is now the theorem:
Theorem 2. (Soundness) Let R be a rewrite theory, and C a finite set of reachability formulas. If R proves [∅, C] ⊢T C then R |=∀T C.
Investigating completeness of the logic is left as future work.
6
Prototype Implementation and Experiments
We have implemented the reachability logic proof system in Maude [1]. Our
prototype takes as input (i) a rewrite theory R = (Σ, E ∪ B, R, φ) and (ii) a set
of reachability formulas C = {Ai →⊛ Bi }i∈I to be simultaneously proved.
To mechanize the two proof rules we use a finitary B-unification algorithm
as well as an SMT solver to discharge E ∪ B constraints. For SMT solving we
14
use variant satisfiability [9,12], which allows us to handle any rewrite theory
R = (Σ, E ∪ B, R) satisfying topmost requirements (1)–(2) and such that the
equational theory (Σ, E ∪ B) has a convergent decomposition satisfying the
finite variant property [2] and protects a constructor subtheory which we assume
consists only of commutative and/or AC and/or identity axioms BΩ . Thus,
both validity and satisfiability of QF formulas in the initial algebra TΣ/E∪B are
decidable [9]. Future implementations will support more general rewrite theories,
add other decision procedures, and use an inductive theorem prover backend.
We have verified properties for a suite of examples of rewrite theories specifying distributed systems such as communication or mutual exclusion protocols
and real-time systems. Table 1 summarizes these experiments. For further details
plus runnable code see http://maude.cs.illinois.edu/tools/rltool/.
Table 1. Examples Verified in the Prototype Implementation
Example
Description of the System/Property
Choice
Nondeterministically throws away elements from a
multiset/eventually only one element left
Comm. Protocol 1
Simple communication protocol/received data is always a
prefix of the data to be sent
Comm. Protocol 2
Fault-tolerant communication protocol/all data is eventually
received in-order
Dijkstra
Dijkstra’s mutual exclusion alg./mutual exclusion
Fixed-Size Token Ring 2-Token ring mutual exclusion alg./mutual exclusion
QLOCK
QLOCK mutual exclusion alg./mutual exclusion
Readers/Writers
Readers-writers mutual exclusion alg./mutual exclusion
Lamport’s Bakery
Unbounded Lamport’s bakery/mutual exclusion
Thermostat
Open system that dynamically responds to
temperature/temperature remains in preset bounds
To illustrate how the tool works in practice, Fig. 1 shows a partial derivation of
a sequent. Recall that for QLOCK we had to prove [∅, C] ⊢[] C, where C was two
already-discussed reachability formulas Gi ≡ Pi′ → [P1 ] ∨ [P2 ] for i ∈ {1, 2} with
respective preconditions the renamed disjuncts Pi′ , 1 ≤ i ≤ 2 in invariant P1 ∨P2 ,
and postcondition [P1 ] ∨ [P2 ], where P1 ≡ < n | w | i | i ; q > | dupl(n w i) 6= tt
and P2 ≡ < n | w | ∅ | q > | dupl(n w) 6= tt. Now, consider [∅, C] ⊢[] P2′ →⊛ [P ].
In the proof fragement below, the initial sequent must apply the step rule. The
result of step(n2w, θ) is the goal resulting from unifying the head of the sequent
with the lefthand side of the rule n2w using the unifier θ = {n 7→ n′′ p, w 7→
w′ , c 7→ ∅, q 7→ q ′ }. The next inference axiom(G2 , α) applies axiom G2 using
the substitution α ⊇ {n 7→ n3 , w 7→ w3 , i 7→ i3 , q 7→ q 3 }. Since G2 has two
constrained patterns in its succedent, we derive two new goals, represented by
proof trees T1 and T2 . In either case, we can immediately subsume by noting
15
T1 ≡
(
[C, ∅] ⊢[] [n3 | w3 | ∅ | q 3 ] | dupl (n ′′ w ′ p) 6= tt ∧
dupl (n 3 w 3 ) 6= tt →⊛ [P1 ] ∨ [P2 ]
T2 ≡
(
[C, ∅] ⊢[] [n3 | w3 | i3 | i3 ; q 3 ] | dupl (n ′′ w ′ p) 6= tt ∧
dupl(n 3 w 3 i 3 ) 6= tt →⊛ [P1 ] ∨ [P2 ]
T1
′′
···
T2
′
′
[C, ∅] ⊢[] < n | w p | ∅ | q > | dupl (n ′′ w ′ p) 6= tt
···
→⊛ [P1 ] ∨ [P2 ]
[∅, C] ⊢[] < n ′ | w ′ | ∅ | q ′ > | dupl(n ′ w ′ ) 6= tt
→⊛ [P1 ] ∨ [P2 ]
sub(P1 , α)
sub(P2 , α)
axiom(G2 , α)
step(n2w, θ)
Fig. 1. Partial Proof Tree for QLOCK
that our reachability formula’s antecedent is an instance of either [P1 ] or [P2 ]
using substitution α, thus terminating the proof.
7
Related Work and Conclusions
This work extends reachability logic [14,15] to a rewrite-theory-generic logic to
reason about both distributed system designs and programs. This extension is
non-trivial. It requires: (i) relativizing terminating sequences to a chosen subset
JT K of terminating states; (ii) solving the “invariant paradox,” to reason about
invariants and co-invariants and characterizing them by reachability formulas
through a theory transformation; and (iii) making it possible to achieve higher
levels of automation by systematically basing the state predicates on positive
Boolean combination of patterns of the form u | ϕ with u a constructor term.
In contrast, standard reachability logic [14,15] uses matching logic, which
assumes a first-order model M and its satisfaction relation M |= ϕ in its reachability logic proof system. As discusses in Section 3, we choose TΣ/E∪B as the
model and →R for transitions, rather than some general M and systematically
exploit the isomorphism TΣ/E∪B |Ω ∼
= TΩ/EΩ ∪BΩ , allowing us to use unification, matching, narrowing, and satisfiability procedures based on the typically
much simpler initial algebra of constructors TΩ/EΩ ∪BΩ . This has the advantage
that we can explicitly give the complete details of our inference rules (e.g. how
Step∀ + Subsumption checks the subsumption, or ensures that states have at
least a successor), instead of relying on a general satisfaction relation |= on some
M. The result is a simpler logic with only two rules (versus eight in [14,15]).
We agree with the work in [6] on the common goal of making reachability logic
rewrite-theory-generic, but differ on the methods used. Main differences include:
(1) [6] does not give an inference system but a verification algorithm. (2) the
theories used in [6] assume restrictions like those in [11] for “rewriting modulo
SMT,” which limit the class of equational theories. (3) Matching is used in [6]
16
instead of unification. Thus, unless a formula has been sufficiently instantiated,
no matching rule may exist, whereas unification with some rule is always possible
in our case. (4) No method for proving invariants is given in [6].
In conclusion, the goal of making reachability logic a rewrite-theory-generic
verification logic has been advanced. Feasibility has been validated with a prototype and a suite of examples. Building a robust and highly effective reachability
logic tool for rewrite theories is a more ambitious future goal.
Acknowledgements. Partially supported by NSF Grant CNS 14-09416.
References
1. M. Clavel, F. Durán, S. Eker, J. Meseguer, P. Lincoln, N. Martı́-Oliet, and C. Talcott. All About Maude – A High-Performance Logical Framework. Springer LNCS
Vol. 4350, 2007.
2. H. Comon-Lundh and S. Delaune. The finite variant property: how to get rid of
some algebraic properties. In Proc RTA’05, Springer LNCS 3467, 294–307, 2005.
3. N. Dershowitz and J.-P. Jouannaud. Rewrite systems. In Handbook of Theoretical
Computer Science, Vol. B, pages 243–320. North-Holland, 1990.
4. F. Durán and J. Meseguer. On the Church-Rosser and coherence properties of
conditional order-sorted rewrite theories. J. Log. Algebr. Program., 81(7-8):816–
850, 2012.
5. K. Futatsugi. Fostering proof scores in CafeOBJ. In Proc. ICFEM 2010, volume
6447, pages 1–20. Springer LNCS, 2010.
6. D. Lucanu, V. Rusu, A. Arusoaie, and D. Nowak. Verifying reachability-logic
properties on rewriting-logic specifications. In Logic, Rewriting, and Concurrency,
volume 9200, pages 451–474. Springer LNCS, 2015.
7. S. Lucas and J. Meseguer. Normal forms and normal theories in conditional rewriting. J. Log. Algebr. Meth. Program., 85(1):67–97, 2016.
8. J. Meseguer. Twenty years of rewriting logic. J. Algebraic and Logic Programming,
81:721–781, 2012.
9. J. Meseguer. Variant-based satisfiability in initial algebras. In C. Artho and
P. Ölveczky, editors, Proc. FTSCS 2015, pages 1–32. Springer CCIS 596, 2016.
10. J. Meseguer and J. Goguen. Initiality, induction and computability. In Algebraic
Methods in Semantics, pages 459–541. Cambridge UP, 1985.
11. C. Rocha, J. Meseguer, and C. Muñoz. Rewriting modulo SMT and open system
analysis. J. Logic and Algebraic Methods in Programming, 86:269–297, 2017.
12. S. Skeirik and J. Meseguer. Metalevel algorithms for variant-based satisfiability.
In Proc. WRLA 2016, volume 9942, pages 167–184. Springer LNCS, 2016.
13. S. Skeirik, A. Stefanescu, and J. Meseguer. A constructor-based reachability logic
for rewrite theories. Technical report. http://hdl.handle.net/2142/95770.
14. A. Stefanescu, Ş. Ciobâcă, R. Mereuta, B. M. Moore, T. Serbanuta, and G. Rosu.
All-path reachability logic. In Proc. RTA-TLCA 2014, volume 8560, pages 425–440.
Springer LNCS, 2014.
15. A. Stefanescu, D. Park, S. Yuwen, Y. Li, and G. Rosu. Semantics-based program
verifiers for all languages. In Proc. OOPSLA 2016, pages 74–91. ACM, 2016.
16. P. Thati and J. Meseguer. Symbolic reachability analysis using narrowing and
its application to the verification of cryptographic protocols. J. Higher-Order and
Symbolic Computation, 20(1–2):123–160, 2007.
| 6cs.PL
|
Inferring protein-protein interaction and protein-DNA
interaction directions based on cause–effect pairs in undirected
and mixed networks1
Mehdy Roayaei2 and MohammadReza Razzazi3
Department of Computer Engineering and Information Technology,
AmirKabir University of Technology,
Tehran, Iran
Abstract
We consider the following problem: Given an undirected (mixed) network and a set of ordered
source-target, or cause-effect pairs, direct all edges so as to maximize the number of pairs that
admit a directed source-target path. This is called maximum graph orientation problem, and has
applications in understanding interactions in protein-protein interaction networks and proteinDNA interaction networks. We have studied the problem on both undirected and mixed
networks. In the undirected case, we determine the parameterized complexity of the problem (for
non-fixed and fixed paths) with respect to the number of satisfied pairs, which has been an open
problem. Also, we present an exact algorithm which outperforms the previous algorithms on
trees with bounded number of leaves. In addition, we present a parameterized-approximation
algorithm with respect to a parameter named the number of backbones of a tree. In the mixed
case, we present polynomial-time algorithms for the problem on paths and cycles, and an FPTalgorithm based on the combined parameter the number of arcs and the number of pairs on
general graphs.
Keywords: protein-protein interaction, protein-DNA interaction, cause–effect pairs, fixedparameter tractable, W[1]-hardness
1
This research was in part supported by a grant from I.P.M.
Corresponding author ([email protected]). Tel: +982164542732, P.O. Box 15875-4413, #424 Hafez Avenue,
AmirKabir University of Technology, Tehran, Iran
3
[email protected]
2
1
1
Introduction
Protein-protein interactions (PPIs) form the skeleton of signal transduction in the cell. These
interactions carry directed signaling information. However, current technologies [8], [11] cannot
decide the direction in which the signal flows. Inferring the directions of these interactions is
fundamental to our understanding of how these networks work. Perturbation experiments [16]
provide additional information for possible direction of information in these networks. In this
experiment, a gene is perturbed (cause) and as a result other genes change their expression levels
(effects), to guide the orientation inference. It is assumed that there must be a directed path in the
network from the causal gene (source) to the affected gene (target). The resulting combinatorial
problem, which is called the maximum graph orientation, is to orient the edges of the network
such that a maximum number of cause-effect pairs admit a directed path from the causal to the
affected gene. When studying a PPI network in isolation, the input network is undirected.
However, the more biologically relevant variant considers also protein-DNA interactions as these
are necessary to explain the expression changes. Moreover, the directionality of some PPIs is
known in advance [6]. Therefore, generally, the input network is considered as a mixed graph
containing both directed and undirected edges. In this paper, we consider the maximum graph
orientation on both undirected and mixed graphs.
The maximum graph orientation problem on undirected graphs and mixed graphs is defined as
follows:
Definition 1 (Maximum undirected graph orientation problem- MUGO). An undirected
graph 𝐺 = (𝑉, 𝐸), and a set 𝑃 = {(𝑠𝑖 , 𝑡𝑖 ): 1 ≤ 𝑖 ≤ 𝑝} of ordered source-target pairs are given,
where 𝐸 is the set of edges. Direct all edges so as to maximize the number of pairs that admit a
directed source-target path.
Definition 2 (Maximum mixed graph orientation problem- MMGO). A mixed graph 𝐺 =
(𝑉, 𝐸, 𝐴), and a set 𝑃 = {(𝑠𝑖 , 𝑡𝑖 ): 1 ≤ 𝑖 ≤ 𝑝} of ordered source-target pairs are given, where 𝐸 is
the set of edges and 𝐴 is the set of arcs. Direct all edges so as to maximize the number of pairs
that admit a directed source-target path.
In the remainder of the paper, let P be the set of input pairs, |𝑃| = 𝑝, V the set of vertices of a
tree or a graph, and |𝑉| = 𝑛. By edge, we mean an undirected edge, and by arc, we mean a
directed edge. A pair is satisfied, if it admits a directed source-target path.
It is shown than an MUGO problem can be converted to an equivalent problem on a tree, which
is obtained by contracting cycles of the input graph [12]. Thus, the interesting case is when the
input graph is a tree. This problem is called maximum tree orientation (MTO). MTO is NP-hard,
even on stars, caterpillars, and binary trees [12], but it is polynomial-solvable on paths. The best
approximation ratio obtained for MTO is Ω(
log log 𝑛
log 𝑛
2
) [10]. It is NP-hard to approximate MTO
within a factor of
11
.
12
Also, MTO has been studied from the parameterized complexity point of
view. It is shown that this problem is fixed-parameter tractable with respect to the parameters the
maximum number of pairs passing through a vertex, and the maximum number of cross pairs
passing through a vertex (the cross pair is defined as a source-target pair whose corresponding
path is directed either towards the root or towards the leaves, but do not change its direction.) [4].
However, it is not fixed-parameter tractable w.r.t. the maximum number of pairs passing through
an edge.
MMGO is also NP-hard. Furthermore, although MTO, the feasibility version of MMGO, the
problem of deciding whether a mixed graph G can be oriented so that the resulting directed graph
contains a directed source-target path for all input pairs, is NP-complete [1]. The best
approximation ratio obtained for MMGO is Ω(
within a factor of
7
8
1
(𝑛|𝑃|)1/3
) [9]. It is NP-hard to approximate MMGO
[6]. To our knowledge, there has been no significant results on the
parameterized complexity of MMGO problem.
An orientation of a graph is an assignment of a direction to each edge. We say that the pair
(𝑠𝑗 , 𝑡𝑗 ) conflicts with the pair (𝑠𝑖 , 𝑡𝑖 ) if there exist no orientation of the input graph for which
both (𝑠𝑗 , 𝑡𝑗 ) and (𝑠𝑖 , 𝑡𝑖 ) are satisfied.
We call an tree with exactly one vertex of degree more than two, a star-like tree. Split vertex is a
vertex of degree two, whose incident arcs have different directions. After orienting an undirected
path between the vertices v and w, the nearest split vertex to the vertex v on that path, is denoted
by 𝑠𝑝𝑙𝑖𝑡(𝑣, 𝑤).
A path between two vertices a and b is denoted by a-to-b or [a-to-b]. If this path does not contain
the vertex a (b) and its incident edge, it is denoted by (a-to-b] ([a-to-b)). A path between a vertex
v of degree more than two and a vertex of degree one, is called a branch incident to the vertex v.
A vertex v is far-adjacent to a vertex w, if by ignoring the vertices of degree two, the vertex v is
adjacent to the vertex w. Let T1 and T2 be two subtrees of the tree T. The graph resulting from the
merging of these two subtrees is denoted by T1 + T2. A graph is called K4-free, if it does not
contain any clique of size 4.
Parameterized computation is a new approach dealing with NP-hard problem [2], [17], and [3].
A fixed-parameter tractable algorithm (FPT-algorithm) is an algorithm that solves a problem of
input size n and a parameter k in 𝑓(𝑘). 𝑛𝑂(1) time, in which 𝑓 is a computable function
depending only on the parameter k [13]. If a problem is W[1]-hard with respect to a parameter k,
then it means there is no FPT-algorithm for it (unless FPT = 𝑊[1]).
The remainder of the paper is organized as follows. In section 2, the MUGO problem on
undirected graphs is studied. We determine the parameterized complexity of the problem (for
3
non-fixed and fixed paths) w.r.t. the number of satisfied pairs, which has been an open problem.
Also, we present an exact algorithm which outperforms the previous algorithms on trees with a
limited number of leaves. In addition, we present a parameterized-approximation algorithm w.r.t.
a parameter named the number of backbones of a tree. In section 3, the MMGO on mixed graphs
problem is studied. We present polynomial-time algorithms for paths and cycles, and an FPTalgorithm based on the combined parameter the number of arcs and the number of pairs for
general graphs. We conclude our paper in section 4 by introducing some open problems.
2
Undirected Networks
In this section, we study the complexity of the MUGO problem w.r.t. the number of satisfied
pairs, the number of leaves of the input tree, and a parameter named the number of backbones of
the input tree.
2.1 Number of satisfied pairs
The parameterized complexity of MUGO problem w.r.t. the number of satisfied pairs has been
an open problem [4]. In this subsection, we determine the parameterized complexity of this
problem for fixed and non-fixed paths. The fixed-path variant of MUGO is identical to MUGO
with the exception that each pair (𝑠𝑗 , 𝑡𝑗 ) ∈ 𝑃 is associated with a fixed path 𝑝𝑗 from 𝑠𝑗 to 𝑡𝑗 in the
graph. Hence, a pair (𝑠𝑗 , 𝑡𝑗 ) is satisfied only if the edges of the path 𝑝𝑗 is oriented from the vertex
𝑠𝑗 towards the vertex 𝑡𝑗 .
First, we study the problem for non-fixed paths. Since paths are non-fixed, we can assume that
the input graph is a tree.
Each instance of the MTO problem can be modeled as an instance of the Maximum Independent
Set (MIS) problem. An independent set is a set of vertices in a graph, no two of which are
adjacent. A maximum independent set is an independent set of largest possible size for a given
graph. Each pair (𝑠𝑗 , 𝑡𝑗 ) of MTO instance is considered as a vertex 𝑣𝑗 in the MIS instance. There
is an edge between two vertices of MIS instance, if and only if the corresponding pairs conflict
with each other. We call the resulting graph a conflict graph. It is clear that finding the optimal
solution of the MIS problem on the conflict graph is equivalent to finding the optimal solution of
the corresponding MTO instance.
Lemma 2.1.1 [4]. The resulting conflict graph of an MTO instance is K4-free.
Now, we model the MIS problem as the Party problem [15]: find the minimum number of guests
that must be invited so that at least 𝛼 guests will know each other (a clique of size 𝛼) or at least 𝛽
guests will not know each other (an independent set of size 𝛽). The solution of the problem is
known as the ramsey number 𝑅(𝛼, 𝛽). Thus, if we consider each vertex of the conflict graph as a
4
guest, and each edge between two vertices as the corresponding guests knowing each other, the
rumsey number 𝑛 = 𝑅(𝛼, 𝛽) returns the minimum number of vertices such that a graph with at
least n vertices contains a clique of size 𝛼 or an independent set of size 𝛽. Ramsey's theorem
states that such a number exists for all 𝛼 and 𝛽.
𝛼+𝛽−2
Lemma 2.1.2 [7]. 𝑅(𝛼, 𝛽) ≤ ( 𝛽−1 ).
According to the lemma 2.1.1, the conflict graph whose the number of vertices is equal to or
more than 𝑅(4, 𝛽), has an independent set of size at least 𝛽. According to lemma 2.1.2,
𝑅(4, 𝛽) ≤
(𝛽+2)(𝛽+1)(𝛽)
(𝛽+2)(𝛽+1)(𝛽)
6
6
. Hence, a conflict graph for which 𝑛 ≥
, has an independent
set of size at least 𝛽.
Theorem 2.1.1. The MTO problem is fixed-parameter tractable with respect to the maximum
number of satisfied pairs.
Proof. Let 𝛽 be the maximum number of satisfied pairs. First, create the conflict graph 𝑇𝑐 of the
MTO problem. If 𝑝 >
(𝛽+2)(𝛽+1)(𝛽)
(𝛽+2)(𝛽+1)(𝛽)
6
6
, we can remove arbitrary 𝑝 −
vertices from 𝑇𝑐 .
According to Ramsey's theorem, 𝑇𝑐 has still an independent set of size at least 𝛽. Thus,
regardless of the size of 𝑇𝑐 , it can be reduced to a kernel of size at most
(𝛽+2)(𝛽+1)(𝛽)
6
such that
there is an independent set of size 𝛽 in the kernel, if and only if there is an independent set of
size 𝛽 in 𝑇𝑐 . Since we have reduced the input instance to a polynomial-size kernel w.r.t. to the
parameter 𝛽, MTO is fixed-parameter tractable w.r.t. the maximum number of satisfied pairs. □
We show that, despite the non-fixed version, the MUGO problem with fixed paths is W[1]-hard
w.r.t. the number of satisfied pairs. We reduce from the K-clique problem: Given an undirected
graph 𝐺𝑐 = (𝑉𝑐 , 𝐸𝑐 ), and a parameter K, is there a clique (a set of vertices that are pairwise
adjacent) of size K in 𝐺𝑐 ? The K-clique problem is W[1]-hard w.r.t. the parameter K [5].
Given an instance 𝐺𝑐 = (𝑉𝑐 , 𝐸𝑐 ) of K-clique problem, we construct an instance 𝐺 = (𝑉, 𝐸) of
MUGO as follows. For each vertex 𝑣𝑖 ∈ 𝑉𝑐 , we create two vertices 𝑠𝑖 and 𝑡𝑖 such that the pair
(𝑠𝑖 , 𝑡𝑖 ) is connected by a fixed path 𝑝𝑖 . If there is no edge between the two vertices 𝑣𝑖 , 𝑣𝑗 ∈ 𝑉𝑐 ,
intersect the corresponding two paths 𝑝𝑖 and 𝑝𝑗 as shown in Figure 1 (a). In this case, the path 𝑝𝑖
contains the edge 𝑢𝑖𝑗 -𝑣𝑖𝑗 , and the path 𝑝𝑗 contains the subpath 𝑢𝑗𝑖 -𝑣𝑖𝑗 -𝑢𝑖𝑗 -𝑣𝑗𝑖 . Otherwise, if there
is an edge between two vertices 𝑣𝑖 , 𝑣𝑗 ∈ 𝑉𝑐 , pass one of them above the other, as shown in Figure
1 (b). In this case, the path 𝑝𝑖 contains the edge 𝑢𝑖𝑗 -𝑣𝑖𝑗 , and the path 𝑝𝑗 contains the edge 𝑢𝑗𝑖 -𝑣𝑗𝑖 .
5
( )
( )
Figure 1. Intersection of two paths in an MUGO instance
In Figure 2, a graph of size five, which contains a 3-clique, and its corresponding MUGO
instance is depicted. One can see that the three vertices 𝑣2 , 𝑣3 , and 𝑣5 of the 3-clique in K-clique
instance are one-to-one corresponding to the three paths 𝑝2 , 𝑝3 , and 𝑝5 that can be satisfied at the
same time in the MUGO instance. The following theorem is the conclusion of the discussion
above.
Theorem 2.1.2. The maximum graph orientation problem with fixed paths is W[1]-hard w.r.t.
the number of satisfied pairs.
Proof. According to the discussion above, we reduce from the K-clique problem. Since the Kclique problem is W[1]-hard w.r.t. the parameter K, and the number of satisfied pairs in the
MUGO instance is equal to the size of a clique in the K-clique instance, the MUGO problem
with fixed paths is W[1]-hard w.r.t. the number of satisfied pairs. □
6
1
12
5
2
1
21
21
12
3
2
23
13
3
31
31
23
1
2
1
1
32
32
13
1
3
2
2
3
3
2
15
3
5
35
25
5
5
51
51
15
52
52
25
53
53
5
35
1
2
3
5
5
Figure 2. An instance of clique problem and its corresponding MUGO instance
2.2 Number of leaves
In [4], the parameterized complexity of MTO was studied w.r.t. to the maximum signal flow
over vertices or edges. They defined the notion of cross pair as a source-target pair whose
corresponding path is directed either towards the root or towards the leaves, but do not change its
direction. They showed that MTO is fixed-parameter tractable with respect to the maximum
number of cross pairs passing through a vertex, denoted by 𝑞𝑣 , and presented an 𝑂(2𝑞𝑣 𝑞𝑣 𝑛2 )
time algorithm to solve it.
In this subsection, we present an exact algorithm which outperforms the mentioned algorithm on
the trees with bounded number of leaves. In fact, We show that the MTO problem is tractable for
a constant number of leaves.
We use the following straightforward lemmas in this subsection.
7
Lemma 2.2.1. Maximum tree orientation can be solved on paths in 𝑂(𝑛2 ) time [4].
Lemma 2.2.2. The number of vertices of degree more than two in a tree is at most the number of
leaves minus two.
Consider the star-like tree in Figure 3 (a), which has three leaves.
Figure 3. star-like trees with three leaves
Without loss of generality, assume that 𝑛 = 3𝑖 + 1, for some even integer i; and there are 𝑖
vertices on each branch of the tree. Regardless of the vertex selected as the root, the maximum
number of cross pairs passing through a vertex is 𝑂(𝑛). Hence, using the FPT-algorithm based
on the parameter 𝑞𝑣 , this instance is solved in 𝑂(2𝑛 ) time. We present another algorithm that
solves this instance in 𝑂(𝑛5 ).
We illustrate the core idea of our algorithm in Figure 3 (b). Consider the path between the vertex
r and the leaf 𝑣1 . In an arbitrary orientation of the tree, when we walk away from the vertex r to
the vertex 𝑣1 , all arcs are in the same direction until we reach to an arc with the different
direction (that is, until a split vertex appears). If there is not a split vertex in the path r-to-𝑣1 , it
means that all arcs on this path have the same direction. Now, as in Figure 3 (b), assume that the
first split vertex on the path (from r-to-𝑣1 ) is the vertex a. One can see that there cannot be a
satisfied pair (𝑠𝑗 , 𝑡𝑗 ) such that 𝑠𝑗 ∈ path (a-to-𝑣1 ] and 𝑡𝑗 ∉ path [a-to-𝑣1 ] or vice versa (that is,
𝑡𝑗 ∈ path (a-to-𝑣1 ] and 𝑠𝑗 ∉ path [a-to-𝑣1 ]). Hence, we can process the subpath a-to-𝑣1
independently.
8
This approach can be applied on all branches of the tree. For each branch, there are at most 𝑂(𝑛)
possible choices for choosing the first split vertex on that branch. If there is no split vertex on a
branch, the leaf of the branch is considered as its first split vertex. Hence, there are at most
𝑂(𝑛3 ) possible combinations for choosing the first split vertices of the branches of the tree. For
each combination, the tree is decomposed into three independent paths (a-to-𝑣1 , b-to-𝑣2 , and cto-𝑣3 ), called separated paths, and a star-like tree with the leaves a, b, and c, where all edges of
each branch have the same direction. The branches r-to-𝑎, r-to-𝑏, and r-to-𝑐 are called new
branches of the tree. According to lemma 2.2.1, the separated paths can be processed
simultaneously in 𝑂(𝑛2 ) time. Also, the newly generated star-like tree can be processed in
polynomial time. Therefore, the MTO problem on a tree with three leaves can be solved in
𝑂(𝑛5 ) time. This algorithm can be easily generalized for all star-like trees with k leaves, in
𝑂(𝑛𝑘+2 . 2𝑘 . 𝑘) time.
We show how our algorithm can be generalized for all trees. Consider the situation, where there
is more than one vertex of degree more than two, as in Figure 4(a). First, consider the vertex 𝑟1.
Assume that the vertices a, b, and c are the split vertices of the branches incident to 𝑟1. As in the
previous case, we decompose the tree at these vertices, and orient each of the new branches (𝑟1 −
𝑎, 𝑟1 − 𝑏, and 𝑟1 − 𝑐). Note that for each separated path, the set of input pairs contains the pairs
for which both endpoints are on that path. Also, the remaining subtree can be processed
independently. However, the set of input pairs P must be updated for this subtree. For each pair
(𝑠𝑗 , 𝑡𝑗 ) ∈ 𝑃, one of the following cases occurs:
Pairs such as (𝑠𝑗 , 𝑡𝑗 ), for which 𝑠𝑗 and 𝑡𝑗 are on one of the separated subpaths. Such
pairs must be deleted from the set P.
Pairs such as (𝑠𝑗 , 𝑡𝑗 ), for which 𝑠𝑗 (or 𝑡𝑗 ) belongs to one of the separated subpaths (as
(𝑠1 , 𝑡1 ) in Figure 4 (a)). Such pairs cannot be satisfied, thus, are removed form P.
Pairs such as (𝑠𝑗 , 𝑡𝑗 ), for which 𝑠𝑗 (or 𝑡𝑗 ) belongs to one of the new branches, but
cannot be satisfied according to the current orientation (as (𝑠2 , 𝑡2 ) in Figure 4(a)).
Such pairs are also removed from P.
Pairs such as (𝑠𝑗 , 𝑡𝑗 ), for which 𝑠𝑗 (or 𝑡𝑗 ) belongs to one of the new branches, and may be
satisfied according to the current orientation. (as (𝑠3 , 𝑡3 ) in Figure 4 (a)). In this case, 𝑠𝑗
(or 𝑡𝑗 ) is transformed to the corresponding vertex of degree more than two (as 𝑠3 is
transformed to 𝑟1 in Figure 4 (b)). In this case, the pair (𝑠𝑗 , 𝑡𝑗 ) in P is replaced by the new
pair.
Other pairs, which are remained in P.
After updating the set P of source-target pairs for each combination of split vertices and new
branches orientation, the new branches are removed from the subtree obtained from the
decomposition of the tree, as in Figure 4 (b).
9
Thus, after selecting the split vertices of the branches of 𝑟1, decomposing the tree at the split
vertices, and then orienting its new branches, the remaining subtree, which has one fewer vertex
of degree more than two, can be processed independently. Note that there is always a vertex of
degree more than two, which is far-adjacent to exactly one vertex of degree more than two. We
start the processing of the tree by choosing such vertex (as 𝑟1 in Figure 4 (a)). Then, this process
is iteratively performed until the whole tree is processed.
1
1
2
3
3
5
2
1
2
3
1
6
(a)
2
2
3
5
2
1
3
1
(b)
6
Figure 4. Illustration of the algorithm on general graphs
Theorem 2.2.1. Maximum tree orientation can be solved on trees in 𝑂((2𝑛)2𝑘 ) time, where k is
the number of leaves of the tree.
Proof. As described above, in each iteration of the algorithm, we choose a vertex of degree more
than two that is far-incident to exactly one vertex of degree more than two. We determine the
split vertices on its branches, decompose the tree at these vertices, orient the new branches,
update the source-target pairs, and remove the new branches. This process continues until the
whole tree is processed.
The number of split vertices that are selected for branches of the tree during the algorithm, are
corresponding to the number vertices of degree one or degree more than two. According to
lemma 2.2.2, the total number of these split vertices is at most 2𝑘 − 2. Thus, there are 𝑂(𝑛2𝑘−2 )
10
combinations for selecting the split vertices. Also, there are two possible orientations for each
new branch which is obtained after the decomposition of the tree at a split vertex. Therefore,
there are 𝑂(22𝑘−2 ) combinations for selecting the orientations of new branches. Altogether, the
algorithm runs in 𝑂(𝑛2𝑘 . 22𝑘−2 ) time. Note that we have omitted some polynomial factors from
the time complexity of the algorithm to simplify it. This algorithm is depicted in Figure 5. □
11
Figure 5. The algorithm based on the number of leaves
12
2.3 Number of backbones
As stated before, the best approximation ratio obtained for MTO is Ω(
log log 𝑛
log 𝑛
) [10], and there has
been no constant-ratio approximation algorithm for it. In this section, we introduce the notion of
the backbone in a tree, and present a parameterized-approximation algorithm with respect to the
number of backbone of a tree, which provides a constant ratio for trees with bounded number of
backbones.
A caterpillar is a tree in which all vertices are within distance one of a central path. We define a
caterpillar-like tree as a tree in which all vertices of degree more than two are on the same path.
We call this central path, a backbone.
Consider the star-like tree and the caterpillar-like tree in Figure 6.(a). In the star-like tree, orient
each branch randomly. Since each pair passes through at most two branches, thus, each pair is
satisfied with a probability of at least 1/4. In the caterpillar-like tree, orient the central path (the
backbone) 𝑣1 -to-𝑣2 and each of the branches incident to it randomly. Analogously, each pair is
satisfied with a probability of at least 1/8.
1
2
(a)
3
1
2
(b)
Figure 6. (a) a star-like and a caterpillar-like tree (b) a tree with two backbones
13
In general trees, vertices of degree more than two may be on more than one path (backbones).
Note that these backbones must be edge-disjoint so that we can orient each of them randomly.
For example, consider the tree in Figure 6 (b), in which, the vertices of degree more than two are
at least on the two backbones 𝑣1 -to-𝑣2 and 𝑣3 -to-𝑣 . Assume that the vertices of degree more
than two of a tree reside on b edge-disjoint backbones. Then, by a random orientation of
backbones and branches incident to the backbones, each pair is satisfied with a probability of at
1
least 2𝑏+2 .
Now, we must find the minimum number of edge-disjoint backbones that cover all vertices of
degree more than two. The first step is to contract all edges that are incident to a vertex of degree
one or two. The resulting tree, which we call the backbone tree 𝑇𝑏 , shows the far-adjacency of
the vertices of degree more than two. It is obvious that a decomposition of 𝑇𝑏 into b edge-disjoint
paths provides b edge-disjoint backbones that cover all vertices of degree more than two in T. So,
it is sufficient to decompose 𝑇𝑏 into the minimum number of edge-disjoint paths.
Lemma 2.3.1. Let T be a tree with 2𝑏, 𝑏 > 0, vertices of odd degree. Then, T can be
decomposed into b edge-disjoint paths. Also, any decomposition of T into edge-disjoint paths
contains at least b paths.
Proof. The proof is by induction on b. If 𝑏 = 0, T has no edge. Assume that a tree with 2𝑏 − 2
vertices of odd degree can be decomposed into 𝑏 − 1 paths. Let a tree T be a tree with 2𝑏
vertices of odd degree. T has at least two vertices of degree one, namely 𝑣1 and 𝑣2 . Deleting the
edges of the path 𝑣1 -to-𝑣2 , does not change the parity of the vertices except 𝑣1 and 𝑣2 . Let the
resulting tree be 𝑇𝑏−1 . 𝑇𝑏−1 is a tree with 2𝑏 − 2 vertices of odd degree, which can be
decomposed into 𝑏 − 1 paths. Thus, using the path 𝑣1 -to-𝑣2 , T can be decomposed into b edgesdisjoint paths.
Since adding an edge-disjoint path to a tree will increase the number of vertices of odd degree at
most by 2, any decomposition contains at least b edge-disjoint paths; otherwise, the resulting tree
has less than 2𝑏 vertices of odd degree. □
The algorithm is illustrated in Figure 7.
14
Figure 7. The algorithm based on the number of backbones
It is easy to use the method of conditional expectations to obtain a deterministic algorithm from
the algorithm in Figure 7 that produces an orientation for a graph with b backbones that satisfies
1
at least 2𝑏+2 of the pairs.
Theorem 2.3.1. There is an approximation algorithm with the approximation ratio
1
2𝑏+2
for the
MTO problem, where b is the minimum number of edge-disjoint backbones (paths) that cover all
vertices of degree more than two.
3
Mixed Networks
Unlike the MTO problem, there may be more than one path between the source and the target of
a pair of P in an MMGO instance. Thus, there has been no efficient algorithm for determining
conflicts between the pairs in P. Therefore, we state the following conjecture:
Conjecture 3.1. MMGO is not fixed-parameter tractable w.r.t. the number of pairs.
When the input is restricted to trees, because each pair corresponds to a path, MMGO is fixedparameter tractable w.r.t. the number of pairs.
Proposition 3.1. MMGO on trees is fixed-parameter tractable w.r.t. the number of pairs.
3.1 Paths
Assume that a path is considered from left to right, and the vertices are numbered from 1 to n.
For all 𝑣, 𝑤 ∈ 𝑉, where 𝑣 ≤ 𝑤, 𝑆(𝑣, 𝑤) is the maximum number of pairs with both endpoints on
the path 𝑣-to-𝑛 that can be satisfied on the path 𝑣-to-𝑛 such that the subpath 𝑣-to-𝑤 is oriented
from 𝑣 to 𝑤. Analogously, 𝑆(𝑤, 𝑣) is the maximum number of pairs with both endpoints on the
path 𝑣-to-𝑛 that can be satisfied on the path 𝑣-to-𝑛 such that the subpath 𝑣-to-𝑤 is oriented from
𝑤 to 𝑣. Also, 𝐴(𝑣, 𝑤) is the number of pairs with both endpoints on the path 𝑣-to-𝑤 that are
satisfied when orienting the path from 𝑣 to 𝑤.
15
When there is an arc on the path 𝑣-to-𝑤 whose direction is toward 𝑣, then 𝐴(𝑣, 𝑤) = 𝑆(𝑣, 𝑤) =
0. Analogously, when there is an arc on the path 𝑣-to-𝑤 whose direction is toward 𝑤, then
𝐴(𝑤, 𝑣) = 𝑆(𝑤, 𝑣) = 0. Also, 𝐴(𝑣, 𝑣) = 0 for all 𝑣.
Then, 𝑆(𝑣, 𝑤) and 𝑆(𝑤, 𝑣) can be calculated as follows:
𝑆(𝑣, 𝑤) = 𝐴(𝑣, 𝑤) + max{𝑆(𝑢, 𝑤), 𝑆(𝑣, 𝑢) − 𝐴(𝑣, 𝑤)} (3.1.1),
𝑆(𝑤, 𝑣) = 𝐴(𝑤, 𝑣) + max{𝑆(𝑤, 𝑢), 𝑆(𝑢, 𝑣) − 𝐴(𝑤, 𝑣)} (3.1.2),
where u is the right-hand side vertex of w, that is 𝑢 = 𝑤 + 1. Note than when 𝑤 = 𝑛, then
𝑆(𝑣, 𝑤) = 𝐴(𝑣, 𝑤), and 𝑆(𝑤, 𝑣) = 𝐴(𝑤, 𝑣).
The main idea behind the recurrence relations (3.1.1) and (3.1.2) is that when a split vertex
appears on a path, the path can be decomposed at that vertex and the two resulting subpaths can
be processed independently. The following theorem is straightforward.
Theorem 3.1.1. MMGO on paths can be solved in 𝑂(𝑛2 ) time.
Proof. In the first step, initialization, as stated above, is done in 𝑂(𝑛2 ) time. The matrix A can be
easily computed in 𝑂(𝑛2 ) time. Also, the matrix S is computed in 𝑂(𝑛2 ) time. 𝑆(1,1) returns the
value of the optimal solution of the MMGO problem on the input path. □
3.2 Cycles
If a cycle has no arc, or has one arc, or all of its arcs have the same direction, it can be oriented
such that all pairs are satisfied. In this case, it is sufficient to orient all edges such that all edges
(and all arcs) have the same direction.
Consider the cycle in Figure 8, in which there are two arcs with opposite directions. Because of
these two arcs, there must be at least two split vertices in any optimal orientation. There are
𝑂(𝑛2 ) combinations for choosing these two split vertices. Assume that a and b are those two
split vertices. We can decompose the cycle at the split vertices into two subpaths between a and
b. The two subpaths are independent of each other except for the pair (𝑎, 𝑏). The pair (𝑎, 𝑏) is
satisfied if and only if the path is oriented from 𝑎 to 𝑏.
{ |𝑎𝑏1 + | + |𝑎𝑏2 + | − 𝛼(𝑎,𝑏) , |𝑎𝑏1 + | + |𝑎𝑏2 |𝑃−(𝑎,𝑏) ,
|𝑎𝑏| = max
(3.2.1),
|𝑎𝑏2 + | + |𝑎𝑏1 |𝑃−(𝑎,𝑏) , |𝑎𝑏1 |𝑃−(𝑎,𝑏) + |𝑎𝑏2 |𝑃−(𝑎,𝑏) }
where |𝑎𝑏1 |𝑃−(𝑎,𝑏) (|𝑎𝑏2 |𝑃−(𝑎,𝑏) ) is the value of the optimal solution of the MMGO on the path
𝑎𝑏1 (𝑎𝑏2 ) with input pairs 𝑃 − (𝑎, 𝑏). Also, 𝛼(𝑎,𝑏) = 1, if the pair (𝑎, 𝑏) ∈ 𝑃, otherwise 𝛼(𝑎,𝑏) =
0.
16
2
1
3
Figure 8. Two arcs with opposite directions in a cycle
This approach can be easily generalized for all cycles.
Theorem 3.2.1. Given a cycle 𝐶 = (𝑉, 𝐸, 𝐴), and a set 𝑃 = {(𝑠𝑖 , 𝑡𝑖 ): 1 ≤ 𝑖 ≤ 𝑝} of source-target
pairs. The MMGO problem can be solved on C in 𝑂(𝑛 ) time.
Proof. The algorithm is depicted in Figure 9. The algorithm is similar to the example showed
above. The only difference is that a path 𝑎𝑏𝑗 (𝑗 = 1,2) may have an arc with the direction
opposite to the direction from a to b. In this case, we assign |𝑎𝑏𝑗 + | = 0.
There are 𝑂(𝑛2 ) combinations for choosing the two split vertices. Since the two subpaths are
independent of each other, for each combination, the solution can be calculated in 𝑂(𝑛2 ) time.
Thus, the MMGO problem on cycles can be solved in 𝑂(𝑛 ) time. □
17
Figure 9. The algorithm for mixed cycles
3.3 Number of arcs and pairs
According to conjecture 3.1, the MMGO is probably not fixed-parameter tractable w.r.t. the
number of pairs. On the other hand, it is clear that MMGO is not fixed-parameter tractable w.r.t.
the number of arcs, |𝐴|, otherwise, MTO could be solved in polynomial time. We show that
MMGO is fixed-parameter tractable w.r.t. the combined parameter the number of pairs and the
number of arcs. In the remainder of this section, let |𝐴| = 𝑘.
Without loss of generality, we assume that the input graph G is a mixed-acyclic graph [14]. One
can see that the graph G can be considered as a set of undirected trees which are connected using
the arcs in A. We call each undirected tree an undirected component. For an illustration, consider
the mixed graph in Figure 10, where the arcs are shown as dashed arrows.
18
1
2
1
3
2
1
2
5
6
Figure 10. Illustration of a mixed graph as connections of some undirected components
The pairs in P can be divided into two types 𝑃𝑈 and 𝑃𝐷 . A pair (𝑠, 𝑡) ∈ 𝑃𝑈 if both 𝑠, 𝑡 are in the
same undirected component. Other pairs belong to 𝑃𝐷 . Satisfying each type of pair can be
handled as follows:
Pair (𝑠, 𝑡) ∈ 𝑃𝑈 : in this case, there is only one path corresponding to the pair. Thus, this
path must be oriented from 𝑠 to 𝑡 to satisfy the pair (such as (𝑠2 , 𝑡2 ) in Figure 10).
Pair (𝑠, 𝑡) ∈ 𝑃𝐷 : in this case, there may be more than one path corresponding to the pair.
Thus, it is sufficient to orient one of them from 𝑠 to 𝑡 to satisfy the pair (such as (𝑠1 , 𝑡1 )
in Figure 10). Each path may pass through one or more undirected components. If for
each of this component, we determine the vertex through which a path enters the
undirected component (the input vertex), and the vertex from which a path exits the
undirected component (the output vertex), in fact, we have determined the path from 𝑠 to
𝑡, because in each undirected component there is only one path between the input vertex
and the output vertex. Also, the remainder of the path is constituted by arcs.
Note that the input and output vertices are endpoints of the arcs that belong to an undirected
component, thus, there are at most 2𝑘 of them. A vertex of an undirected component that is the
head (tail) of an arc, is an input (output) vertex. For each vertex of an undirected component
which is the endpoint of an arc, we must determine that the paths of which pairs pass through it.
For example, the path corresponding to the pair (𝑠1 , 𝑡1 ) may pass through the vertices
𝑣1 , 𝑣2 , 𝑣3 , 𝑣 , 𝑣5 , 𝑣6 . Note that the vertex 𝑠𝑖 is considered as the input vertex through which the
pair (𝑠𝑖 , 𝑡𝑖 ) enters to the component containing 𝑠𝑖 . Analogously, the vertex 𝑡𝑖 is considered as the
output vertex from which the pair (𝑠𝑖 , 𝑡𝑖 ) exits the component containing 𝑡𝑖 .
For each consistent assignment of the pairs to the input and output vertices, some of the pairs in
𝑃𝐷 are satisfied. An assignment is consistent if for all components, the path from the input vertex
to the output vertex, for all pairs that pass through that component, can be oriented. After
19
orienting a consistent assignment, each undirected component has been converted to a mixed
component. Since the pairs of 𝑃𝑈 in a component are independent of the pairs of 𝑃𝑈 in other
components, the mixed components can be processed independently. According to proposition
3.1, the maximum number of pairs of 𝑃𝑈 that can be satisfied in each mixed component can be
computed in 𝑂(2|𝑃𝑈 | ) time. The maximum number of total satisfied pairs among all consistent
assignments of the pairs to the input and output vertices is the optimal solution to the problem.
The algorithm is depicted in Figure 11.
Figure 11. The algorithm for mixed graphs based on the number of pairs and arcs
Theorem 3.3.1. Given a mixed graph 𝐺 = (𝑉, 𝐸, 𝐴) and a set 𝑃 = {(𝑠𝑖 , 𝑡𝑖 ): 1 ≤ 𝑖 ≤ 𝑝} of sourcetarget pairs, where |𝐴| = 𝑘. The MMGO problem is fixed-parameter tractable w.r.t. the
combined parameter (𝑝, 𝑘).
Proof. There are 𝑂(22𝑝𝑘 ) assignments of the pairs to the input and output vertices. Consistency
checking of each assignment can be done in polynomial time. According to proposition 3.1, the
MMGO problem can be solved on each mixed component in 𝑂(2𝑝 ) time. Thus, the problem can
be solved in 𝑂(2(2𝑘+1).𝑝 ) time. Note that we have omitted some polynomial factors from the
time complexity of the algorithm to simplify it. □
20
4
Conclusion
In this paper, we studied the maximum graph orientation problem on undirected and mixed
graphs. In the undirected case, we determined the parameterized complexity of the problem (for
non-fixed and fixed paths) w.r.t. the number of satisfied pairs, which was an open problem. Also,
we presented an exact algorithm based on the number of leaves of a tree. In addition, we
presented a parameterized-approximation algorithm w.r.t. a parameter named number of
backbones of a tree. In the mixed case, we presented polynomial-time algorithms for paths and
cycles, and an FPT-algorithm based on the combined parameter the number of arcs and the
number of pairs for general graphs.
There are still some open problems for future investigations, some of which are:
5
What is the parameterized complexity of the MMGO problem w.r.t. the parameter
“number of pairs”?
What is the parameterized complexity of the MTO problem w.r.t. the parameter “number
of all pairs minus the number of input pairs”?
Are there constant-ratio approximation algorithms for MTO and MMGO problems?
References
[1] E. Arkin and R. Hassin, “A note on orientations of mixed graphs,” Discret. Appl. Math., pp.
1–10, 2002.
[2] J.E. Chen, “Parameterized computation and complexity: a new approach dealing with NPhardness,” J. Comput. Sci. Technol., vol. 20, no. 1, pp. 18–37, 2005.
[3] J. Chen and Q.L. Feng, “On Unknown Small Subsets and Implicit Measures: New
Techniques for Parameterized Algorithms,” J. Comput. Sci. Technol., vol. 29, no. 5, pp. 870–
878, 2014.
[4] B. Dorn, F. Hüffner, D. Krüger, R. Niedermeier, and J. Uhlmann, “Exploiting bounded signal
flow for graph orientation based on cause – effect pairs,” Algorithms Mol. Biol., vol. 6, no. 1, p.
21, Jan. 2011.
[5] R. G. Downey and M. R. Fellows, Fundamentals of parameterized complexity, vol. 4.
Springer, 2013.
[6] M. Elberfeld and D. Segev, “Approximation algorithms for orienting mixed graphs,” Theor.
Comput. Sci., vol. 483, pp. 96–103, Apr. 2011.
[7] P. Erdös and G. Szekeres, “A combinatorial problem in geometry,” Compos. Math., vol. 2,
pp. 463–470, 1935.
21
[8] S. Fields, “High-throughput two-hybrid analysis,” FEBS J., vol. 272, no. 21, pp. 5391–5399,
2005.
[9] Gamzu and M. Medina, “Improved approximation for orienting mixed graphs,” Struct. Inf.
Commun. Complex., vol. 7355, pp. 243–253, 2012.
[10] Gamzu, D. Segev, and R. Sharan, “Improved orientations of physical networks,” Lect. Notes
Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics), vol. 6293
LNBI, pp. 215–225, 2010.
[11] A.C. Gavin, M. Bösche, R. Krause, P. Grandi, M. Marzioch, A. Bauer, J. Schultz, J. M.
Rick, A.-M. Michon, C.-M. Cruciat, and others, “Functional organization of the yeast proteome
by systematic analysis of protein complexes,” Nature, vol. 415, no. 6868, pp. 141–147, 2002.
[12] Medvedovsky, V. Bafna, U. Zwick, and R. Sharan, An algorithm for orienting graphs based
on cause-effect pairs and its applications to orienting protein networks. Springer Berlin
Heidelberg, 2008.
[13] R. Niedermeier, Invitation to Fixed Parameter Algorithms. Oxford University Press, 2006.
[14] D. Silverbush, “Optimally orienting physical networks,” J. Comput. Biol., vol. 18, no. 11,
pp. 1437–48, Nov. 2011.
[15] D. B. West and others, Introduction to graph theory, vol. 2. Prentice hall Upper Saddle
River, 2001.
[16] C.H. Yeang, T. Ideker, and T. Jaakkola, “Physical network models,” J. Comput. Biol., vol.
11, no. 2–3, pp. 243–262, 2004.
[17] X. Zhao and D. Ding, “Fixed-parameter tractability of disjunction-free default reasoning,” J.
Comput. Sci. Technol., vol. 18, no. 1, pp. 118–124, 2003.
22
| 8cs.DS
|
Programming the Interactions of Collective Adaptive Systems by Relying on
Attribute-based Communication
arXiv:1711.06092v3 [cs.PL] 29 Nov 2017
YEHIA ABD ALRAHMAN, IMT School for Advanced Studies, Lucca, Italy
ROCCO DE NICOLA, IMT School for Advanced Studies, Lucca, Italy
MICHELE LORETI, Università di Camerino, Camerino, Italy
Collective adaptive systems are new emerging computational systems consisting of a large number of interacting components
and featuring complex behaviour. These systems are usually distributed, heterogeneous, decentralised and interdependent, and are
operating in dynamic and possibly unpredictable environments. Finding ways to understand and design these systems and, most of all,
to model the interactions of their components, is a difficult but important endeavour. In this article we propose a language-based
approach for programming the interactions of collective-adaptive systems by relying on attribute-based communication; a paradigm
that permits a group of partners to communicate by considering their run-time properties and capabilities. We introduce AbC, a
foundational calculus for attribute-based communication and show how its linguistic primitives can be used to program a complex
and sophisticated variant of the well-known problem of Stable Allocation in Content Delivery Networks. Also other interesting case
studies, from the realm of collective-adaptive systems, are considered. We also illustrate the expressive power of attribute-based
communication by showing the natural encoding of other existing communication paradigms into AbC.
CCS Concepts: • Computing methodologies → Distributed programming languages; • Theory of computation → Distributed computing models; Operational semantics; Process calculi;
Additional Key Words and Phrases: Collective-adaptive systems, Attribute-based communication, distributed programming
ACM Reference format:
Yehia Abd Alrahman, Rocco De Nicola, and Michele Loreti. 0. Programming the Interactions of Collective Adaptive Systems by Relying
on Attribute-based Communication. ACM Trans. Softw. Eng. Methodol. 0, 0, Article 32 ( 0), 36 pages.
https://doi.org/0000000.0000000
1
INTRODUCTION
The ever increasing complexity of modern software systems has changed the perspective of software designers who
now have to consider a broad range of new classes of systems, consisting of a large number of interacting components
and featuring complex interaction mechanisms, e.g. Software-Intensive Systems [Broy 2006], IoT Systems [Kopetz 2011],
and Collective Adaptive Systems (CAS) [Anderson et al. 2013; Ferscha 2015]. These systems are usually distributed,
heterogeneous, decentralised and interdependent, and are operating more and more in dynamic and often unpredictable
This research has been partially supported by the European projects IP 257414 ASCENS and STReP 600708 QUANTICOL, and by the Italian project PRIN
2010LHT4KM CINA.
Author’s addresses: Y.A. Alrahman and R. De Nicola, IMT School for Advanced Studies, Lucca, Italy; M. Loreti, Università di Camerino, Camerino, Italy;
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not
made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components
of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on
servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected].
© 0 Copyright held by the owner/author(s). Publication rights licensed to Association for Computing Machinery.
Manuscript submitted to ACM
Manuscript submitted to ACM
1
2
Y. Abd Alrahman et al.
environments. There exist different kinds of complexity in the development of software systems. Historically, as software
systems grew larger, the focus shifted from the complexity of developing algorithms to the complexity of structuring
large systems, and to the inevitable complexities of building distributed and concurrent systems [Wirsing and Holzl
2006]. We are now facing another level of complexity arising from the fact that systems have to operate in large, open
and non-deterministic environments where the complexity of interaction is increased.
Most of the current communication models and programming frameworks still handle the interaction between
distributed components by relying on their identities, like in the Actor model [Agha 1986], or by relying on the
communication channels on which they are listening, like in point-to-point channel-based communication [Sangiorgi and
Walker 2003], multicast with explicit addressing [Holbrook and Cheriton 1999], and broadcast communication [Prasad
1991]. These models and frameworks use names or addresses, that are totally independent from the run-time properties
and capabilities of the interacting components, as first-class citizens to model the interaction. This makes it hard to
program complex behaviours and coordination actions (reconfiguration, adaptation, opportunistic interactions, . . . )
that depend on the components’ status rather than on their identities or addresses. The complexity of these models
arises because they are communication centric; the behaviour mainly consists of a process implementing a specific
communication protocol and the programmer has to manage the channels or the addresses on which components
interact.
In our view, to deal with CAS, it is important to consider more data centric interaction primitives that abstract
from the underlying communication infrastructure (i.e., infrastructure-agnostic) and relies on anonymous one-to-many
interactions to send messages to potential receivers. In this article we study the impact of a new paradigm that permits a
group of partners to interact by considering their properties and capabilities at run-time. The findings we report in this
article have been triggered by our interest in CAS, and by the recent attempts to define appropriate linguistic primitives
to deal with such systems, see e.g. Intelligent Ensembles [Krijt et al. 2017], FORMS [Weyns et al. 2012], Rainbow [Garlan
et al. 2004], SOA-based [Calinescu 2008], Swarm of Cognitive Agents [Park and Tran 2012], UNDERSEA [Gerasimou
et al. 2017], TOTA [Mamei and Zambonelli 2004], SCEL [De Nicola et al. 2014] and the calculi presented in [Alrahman
et al. 2015] and [Viroli et al. 2013].
Collective-Adaptive Systems (CAS)
CAS [Anderson et al. 2013] consist of a large number of interacting components which combine their behaviours, by
forming collectives, to achieve specific goals depending on their attributes, objectives, and functionalities. Decisionmaking in these systems is complex and components interaction may lead to unexpected behaviours. CAS are inherently
scalable [Sanderson et al. 2015] and the boundaries between different CAS are fluid in the sense that components
may enter or leave the collective at any time and may have different (potentially conflicting) objectives; so they need
to dynamically adapt to their environmental conditions and contextual data. The need of engineering techniques to
address the challenges of developing, integrating, and deploying CAS is advocated in [Sommerville et al. 2012]. Also
the development of theoretical foundation for CAS is very important to understand their distinctive features [Cheng
et al. 2009].
Clearly, the existing communication models do not scale with the high level of dynamicity of CAS, and, in our opinion,
a change of perspective, that takes into account run-time properties, status, and capabilities of communicating systems,
is on demand. The key concepts of CAS should be the basis for guiding the development of the new communication
modalities.
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
3
In the following, we summarise the main concepts of CAS as reported in [Anderson et al. 2013; Cheng et al. 2009;
Ferscha 2015; Sanderson et al. 2015]:
Distribution: components are distributed on a shared environment and evolve independently, without any
centralised control.
Awareness: components are aware of their run-time status, characteristics, capabilities, and have some (partial)
view of their surroundings.
Adaptivity: components adapt their behaviours and their interaction policies in response to the changes in their
contextual conditions and to the collected awareness data.
Interdependence: any change in the shared environment might influence the behaviour of components.
Collaboration: components collaborate and combine their behaviours to achieve system-level goals in response to
changes in the environment or because of their predefined roles.
Anonymity: components are able to communicate and exchange data without previously knowing the existence
of each other; for example, the identity of a service provider is not relevant, only its features and its ability to
provide the service are important.
Scalability and Open-endedness: components may join or leave a system without disturbing the overall system
behaviour; senders emit messages without being aware of the presence of receivers and receivers do not rely
on specific senders.
Attribute-based Communication
In order to capture the above mentioned concepts, we have developed attribute-based communication, a communication
paradigm that permits a group of partners to communicate by considering the predicates over the attributes they
expose. Communication takes place anonymously in an implicit multicast fashion without a prior agreement between
the communicating partners. Anonymity of interaction allows programmers to secure scalability, dynamicity, and
open-endedness more easily. Sending operations are non-blocking while receiving operations are blocking. This breaks
synchronisation dependencies between interacting partners, and permits modelling systems where communicating
partners can enter or leave a group at any time without endangering the overall behaviour. Moreover, attributes make it
easy to model awareness by locally reading the values of the attributes that may represent either the component status
(e.g., the battery level of a robot) or the external environment (e.g., the external humidity). Groups are dynamically
formed at the time of interaction by considering the interested receiving components that satisfy sender’s predicates,
and any run-time changes of attribute values allow opportunistic interactions between components. By parameterising
the interaction predicates with local attributes, groups can be implicitly changed and adaptation is naturally captured.
Security mechanisms can be placed on top of the attribute-based framework following standard approaches.
Modeling opportunistic interaction in classical communication paradigms like channel-based communication, e.g.,
π -calculus [Milner et al. 1992], is definitely more challenging because components have to agree on specific names
or channels to interact. Channels have no connection with the component attributes or characteristics; they are
specified as addresses where the exchange should happen. Names and channels are static and changing them locally at
run-time requires explicit communication and intensive use of scoping mechanisms which do affect readability and
compositionality of programs.
It is also worth mentioning that Multi-Agent System (MAS) [Wooldridge 2000] is usually used for modeling distributed
and adaptive systems, e.g., [Leitão et al. 2012; Lepuschitz et al. 2013]. Since MAS originates as a sub-field of distributed
Manuscript submitted to ACM
4
Y. Abd Alrahman et al.
artificial intelligence where logic is the reference formal framework, MAS models provide little support to formally
validate an implementation against its specifications [Viroli and Omicini 2005]. Logics, however, provide powerful tools
to declaratively model the features of complex systems as MAS [Wooldridge 2000]. In our approach we provide an
operational model that precisely describes the step-by-step admissible evolutions of the system. Operational models
ensure more straightforwardly that implementations are correct with respect to their specifications and based on the
evolution rules, system properties can be verified [Viroli and Omicini 2005]. Different attempts have been carried out to
bridge the gap between MAS and process algebra approaches, see [Bergenti et al. 2003; de Boer et al. 2005; Kinny 2005],
however this is out the scope of this article.
We would also like to stress that an attribute-based system is more than just the result of the parallel composition of
its interacting components; it also takes into account the environment where components are executed. Indeed, the
environment has a great impact on components behaviours and allows modelling interdependence, i.e., a situation
where components influence each other unintentionally. For example, in an ant foraging system [Jackson and Ratnieks
2006], when an ant disposes pheromone in the shared space to keep track of its way back home, it influences other
ants behaviour as they are programmed to follow traces of pheromone with higher concentration. In this way, the ant
unintentionally influences the behaviour of the other ants by only modifying the environment.
In this article we introduce AbC, a calculus comprising a minimal set of primitives that permits attribute-based
communication and is the outcome of different attempts towards modeling the interaction of CAS scenarios. From our
experience, we learnt that any attribute-based paradigm, tailored for modeling the interaction of CAS systems, should
at least provide support for the following notions, that are the key ingredients of AbC:
• Attribute Environment: to provide a collection of attributes whose values represent the status of a component.
These values can be used to influence the behaviour of a component at run-time.
• Attribute-based send and receive operations: to establish the communication links between different components
based on the satisfaction of predicates over components attributes.
• Attribute update operation: to change attribute values based on contextual conditions and adapt the behaviour
of a component accordingly.
• Awareness construct: to collect awareness data and take decisions based on the changes in the attribute
environment.
In this article, we focus on a language-based approach to program the interactions of CAS and concentrate on the
linguistic constructs of the AbC calculus and on their use. Other more theoretical aspects, concerned with modeling,
with primitives for abstraction, with behavioural theories, and with equational laws, can be found in [Alrahman et al.
2016a, 2017a].
An AbC system is rendered as a set of parallel components, each equipped with a set of attributes whose values
can be modified by internal actions. Communication actions (both send and receive) are decorated with predicates
over attributes that partners have to satisfy to make the interaction possible. Thus, communication takes place in an
implicit multicast fashion, and communication partners are selected by relying on predicates over the attributes in their
interfaces. Unlike IP multicast [Holbrook and Cheriton 1999] where the reference address of the group is explicitly
included in the message, AbC components are unaware of the existence of each other and they receive messages only if
they mutually satisfy the requirements of each other, otherwise messages are discarded.
Our main goal is to permit the construction of high-level and formally verifiable communication APIs for CAS based
on the AbC calculus. These APIs can then be used to reason and verify properties about the interactions of the actual
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
5
deployed systems by relying on the operational semantics of AbC. For example, we developed APIs for Java [Alrahman
et al. 2016b], for Go [Alrahman et al. 2017b], and for Erlang [De Nicola et al. 2017]. The actual implementations of
these APIs fully rely on the formal semantics of AbC. We also provided a one-to-one correspondence between the AbC
primitives and the programming constructs of the above mentioned APIs. The direct correspondence permits exploiting
the AbC linguistic primitives to program the interaction of CAS applications in different host languages as required by
the application domain. At the same time, it enhances the confidence on the behaviour of programs, written by using
these APIs, after they have been analysed via formal methods, which is made possible by relying on the operational
semantics of the AbC calculus.
Contributions. The main contribution of this article is the introduction of a language-based approach for programming
the interactions of collective-adaptive systems. In Section 2, we introduce the AbC calculus. To capture the reader’s
intuition, we present the main features of AbC in a step-by-step fashion by using a distributed variant of the Graph
Colouring Problem [Jensen and Toft 1995] as a running example. In Section 2.3, we illustrate the expressive power of
attribute-based communication by showing the natural encoding of other communication paradigms. In the rest of the
article, we exploit the interaction primitives of the AbC calculus to show how complex and interesting case studies,
from the realm of collective-adaptive systems, can be programmed in an intuitive way. More specifically, in Section 3,
we show how to program a more complex and sophisticated variant of the well-known problem of Stable Allocation in
Content Delivery Networks [Maggs and Sitaraman 2015]. This variant is designed to cover most of CAS features in a
single case study with a well-understood system-level goal. Additional case studies are presented in the appendices A.1
and A.2. In A.1, we show how to program a crowd steering scenario while in Section A.2, we show how to program
swarm of robotics, performing a rescuing mission in a disaster arena.
2
THE ABC CALCULUS
In this section we first present the syntax of AbC, we discuss its formal operational semantics, and then we provide an
evidence of its expressive power.
2.1
Syntax of the AbC Calculus
The syntax of AbC is reported in Table 1. The top-level entities of the calculus are components (C). A component, Γ :I P,
is a process P associated with an attribute environment Γ, and an interface I . An attribute environment Γ : A →
7 V is a
partial map from attribute identifiers1 a ∈ A to values v ∈ V where A ∩ V = ∅. A value could be a number, a name
(string), a tuple, etc. An interface I ⊆ A consists of a set of attributes that are exposed by a component to control the
interactions with other components. We will refer to the attributes in I as public attributes, and to those in dom(Γ) − I
as private attributes. Complex components are built by using parallel operator C 1 ∥C 2 .
Example 2.1 (Distributed Graph Colouring Scenario). We model a distributed variant of the well known Graph
Colouring Problem [Jensen and Toft 1995] using AbC constructs. We render the problem as a typical CAS scenario where
a collective of agents, executing the same code, collaborate to achieve a system-level goal without any centralised
control.
The problem consists of colouring the vertices of a graph in such a way that no two vertices sharing an edge have
the same colour. Formally, we have a set of n vertices, each of which is identified by an id (an integer in our model). A
1 In
the rest of this article, we shall occasionally use the term “attribute” instead of “attribute identifier”.
Manuscript submitted to ACM
6
Y. Abd Alrahman et al.
(Components)
(Processes)
(Updates)
(Predicates)
(Expressions)
C ::= Γ :I P | C 1 ∥C 2
P ::= 0 | Π(x̃).U | (Ẽ)@Π.U | ⟨Π⟩P | P1 + P2 | P 1 |P2 | K
U ::= [a := E]U | P
Π ::= tt | ff | p(Ẽ) | Π 1 ∧ Π 2 | Π 1 ∨ Π2 | ¬Π
E ::= v | x | a | this.a | op(Ẽ)
Table 1. The syntax of the AbC calculus
set of neighbours Ni is also associated with each vertex i, where j ∈ Ni if and only if i ∈ N j . Our goal is to assign to
each vertex i a colour c i such that for each j ∈ Ni , c j , c i .
Each vertex is modelled in AbC as a component of the form Ci = Γi : {id,N} PC . Public attributes id and N are used to
represent the vertex id and the set of neighbours N , respectively. The overall system is defined as the parallel composition
of existing components (i.e., C 1 ∥C 2 ∥, . . . , ∥Cn ).
□
A process can be the inactive process 0, an action-prefixed process, act .U , where act is a communication action and U
is a process possibly preceded by an attribute update, a context aware ⟨Π⟩P process, a nodeterministic choice between
two processes P1 + P2 , a parallel composition of two processes P1 |P 2 , or a process call with a unique identifier K used in
process definition K ≜ P. All of these operators will now be described below. We start by explaining what we mean by
expressions and predicates, then we continue by describing the actual operations on processes.
An expression E is built from constant values v ∈ V, variables x, attribute identifiers a, a reference to the value
of a (this.a) in the component that is executing the code, or through a standard operators op(Ẽ)2 . The evaluation
of expression E under Γ is denoted by JEKΓ . The definition of J·KΓ is standard, the only interesting cases are JaKΓ =
Jthis.aKΓ = Γ(a).
A predicate Π is built from boolean constants, tt and ff, and from an atomic predicate p(Ẽ) by using standard boolean
operators (¬, ∧ and ∨). The precise set of atomic predicates is not detailed here; we only assume that it contains basic
binary relations like >, <, ≤, ≥, =, and the predicates ∈ and <. In what follows, we shall use the notation {Π}Γ to
indicate the closure of a predicate Π under the attribute environment Γ. The closure is also a predicate Π ′ obtained from
Π by replacing the occurrences of the expression this.a with its value Γ(a).
The attribute-based output (Ẽ)@Π is used to send the evaluation of the sequence of expressions Ẽ to the components
whose attributes satisfy the predicate Π.
The attribute-based input Π(x̃) is used to receive messages from a component satisfying predicate Π; the sequence x̃
acts as a placeholder for received values. The action Π(x̃) acts as a binder for names, e.g., x̃ in Π(x̃).U . We will say that
a name x is bound when it occurs under the scope of an input action while it is free when it is not bound. We use bn(P)
and f n(P) to denote the set of bound and free names of P, respectively. We use f v(P) to denote the set of free process
variables of P. We use x, y, . . . to range over names while X , Y , . . . to range over process variables. Notice that names
are used as placeholders for values while process variables are used as placeholders for processes. Our processes are
closed, i.e. without free process variables (f v(P) = ∅) because AbC components can only exchange values, but not code.
2 For
the sake of simplicity, we omit the specific syntax of operators used to build expressions and use Ẽ to denote sequences of expressions.
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
7
Predicates, used in communication actions, can also refer to variable names in x̃ and the received values can be used
to check whether specific conditions are satisfied. For instance, the action
((x = “try”) ∧ (this.id > id) ∧ (this.round = z))(x, y, z)
can be used to receive a message of the form (“try”, c, r ) where the value received on z is equal to this.round and the
value of the interface attribute id of the sending component is less than this.id. Thus, the predicate can be used to
check either the received values or the values of the sending component interface. A predicate can also refer to local
attributes of components. Thus, an action like
(“try”, c, r )@(this.id ∈ N)
can be used to send the message (“try”, c, r ) to all components whose attribute N contains this.id.
An attribute update, [a := E], is used to assign the result of the evaluation of E to the attribute identifier a. The syntax
is devised in such a way that sequences of updates are only possible after communication actions. Actually, updates
can be viewed as side effects of interactions. It should be noted that the execution of a communication action and the
following update(s) is atomic. This possibility allows components to modify their attribute values and thus triggering
new behaviours in response to collected contextual data.
The awareness construct, ⟨Π⟩P, is used to trigger new behaviours (i.e., P) when the status of a component is changed
(i.e., Π |= Γ). It blocks the execution of P until predicate Π satisfies the attribute environment.
The parallel operator, P |Q, models the interleaving between co-located processes, i.e., processes residing within the
same component.
The choice operator, P + Q, indicates a nondeterministic choice among P and Q.
Other process operators can be defined as macros in AbC. Indeed, we will use the following derived operators:
if Π then P1 else P 2 ≜ ⟨Π⟩P 1 + ⟨¬Π⟩P 2
(1)
let x = E in P ≜ P[E/x]
(2)
set(a, E)P ≜ ()@ff.[a := E]P
(3)
[a 1 := E 1 , a 2 := E 2 , . . . , an := En ]P ≜ [a 1 := E 1 ][a 2 := E 2 ] . . . [an := En ]P
(4)
Example 2.2. In Example 2.1 we introduced the structure of the components modelling the vertices in the considered
scenario. Here, we show how the behaviour of our components can be programmed to assign a colour (an integer) to
each of them while avoiding that two neighbours get the same colour.
The proposed algorithm consists of a sequence of rounds for colour selection that goes on until the specified goal is
reached. At the end of each round at least one component is assigned a colour.
Components use messages of the form (“try”, c, r ) to inform their neighbours that at round r they want to select
colour c and messages of the form (“done”, c, r ) to communicate that colour c has been definitely chosen at the end of
round r . At the beginning of a round, each component i selects a colour and sends a try-message to all components
in Ni . Component i also collects try-messages from its neighbours. The selected colour is assigned to a component
only if it has the greatest id among those that have selected the same colour in that round. After the assignment, a
done-message (associated with the current round) is sent to neighbours. A new round starts when a message, associated
with a round r such that this.round < r , is received.
Manuscript submitted to ACM
8
Y. Abd Alrahman et al.
This algorithm can be implemented in AbC by using four processes, F for forwarding try-messages to neighbours, T
for handling try-messages, D for handling done-messages, and A for assigning a final colour. Process PC of Example 2.1
is now defined as the parallel composition of these four processes: PC = F | T | D | A.
The following private attributes, local to each component, are used to control the progress of our algorithm: round,
done, assigned, used, counter, send and constraints. The attribute “round" stores the current round while “constraints"
and “used" are sets, registering the colours used by neighbours. The attribute “counter" counts the number of messages
collected by a component while “send" is used to enable/disable forwarding of messages to neighbours. The private
attributes of each component are all initialised with the following values: round = done = 0, constraints = used = ∅,
send = tt, and assigned = ff.
In process F reported below, when the value of attribute send becomes tt, a new colour is selected, and a message
containing this colour and the current round is sent to all the components having this.id as neighbour. The new colour
is the smallest colour that has not yet been selected by neighbours, that is min{i < this.used }. The guard ¬assigned is
used to make sure that components with assigned colours do not take part in the selection anymore.
F
≜
⟨send ∧ ¬assigned⟩
set(colour, min{i < this.used })
(“try”, this.colour, this.round)@(this.id ∈ N).[send := ff]F
Process T , reported below, receives messages of the form (“try”, c, r ). If r = this.round, as in the first two branches,
then the received message has been originated by another component performing the same round of the algorithm.
The first branch is executed when this.id > id, i.e., the sender has an id smaller than the id of the receiver. In this case,
the message is ignored (there is no conflict), simply the counter of received messages (this.counter) is incremented. In
the second branch, this.id < id, the received colour is recorded to check the presence of conflicts. The value of y is
added to this.constraints and this.counter is incremented by 1.
If r > this.round, as in the last two branches, then the received message has been originated by a component
executing a successive round and two possible alternatives are considered, this.id < id or this.id > id. In both cases,
round is set to r , send and counter are updated accordingly, and this.constraints is set to the value of y if this.id < id.
T
≜
((x = “try”) ∧ (this.id > id) ∧ (this.round = z))(x, y, z).
[counter := counter + 1]T
+ ((x = “try”) ∧ (this.id < id) ∧ (this.round = z))(x, y, z).
[counter := counter + 1, constraints := constraints ∪ {y}]T
+ ((x = “try”) ∧ (this.id > id) ∧ (this.round < z))(x, y, z).
[round := z, send := tt, counter := 1, constraints := ∅]T
+ ((x = “try”) ∧ (this.id < id) ∧ (this.round < z))(x, y, z).
[round := z, send := tt, counter := 1, constraints := {y}]T
Process D, reported below, is used to receive messages of the form (“done”, c, r ). These are sent by components that
have reached a final decision about their colour. When (“done”, c, r ) is received, we have that either this.round ≥ r or
this.round < r . In the first case, the used colour is registered and the counter this.done is incremented. In the second
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
9
case, private attributes are updated to indicate the startup of a new round.
D
((x = “done”) ∧ (this.round ≥ z))(x, y, z).
≜
[done := done + 1, used := used ∪ {y}]D
+ ((x = “done”) ∧ (this.round < z))(x, y, z).
[round := z, done := done + 1, constraints := ∅,
send := tt, counter := 0, used := used ∪ {y}]D
Process A, reported below, is used to manage the definitive selection of a colour and can only be executed when
messages from all “pending” neighbours have been received (this.counter = |this.N| − this.done) and no conflict
has been found (this.colour < this.used ∪ this.constraints). When the above conditions are satisfied, message
(“done”, this.colour, this.round + 1) is sent to neighbours, the assigned attribute is set to true, and the process
terminates.
A ≜ ⟨(this.counter = |this.N| − this.done) ∧ (this.colour < this.constraints ∪ this.used)⟩
(“done”, this.colour, this.round + 1)@(this.id ∈ N).[assigned := tt]0
□
Remark. Example 2.2 shows the main advantages of the attribute-based interaction. In essence, components are
infrastructure-agnostic, i.e., they abstract from the underlying communication infrastructure and rely on anonymous
one-to-many interaction pattern to communicate; this simplifies the design of component behaviours, because there is no
need to manage the communication channels or the addresses on which components interact. The use of communication
predicates, to derive the interaction between different components, permits programming data centric applications by
taking into account the run-time features of the interacting components.
2.2
AbC Operational Semantics
In this section, we provide an overview of the operational semantics of AbC and use fragments of the Distributed Graph
Colouring example to show how the semantics rules work. The operational semantics of AbC is based on two relations.
The transition relation 7−
→ that describes the behavior of single components and the transition relation −
→ that relies on
7−
→ and describes system behaviours.
2.2.1 Operational semantics of components. We use the transition relation 7−
→ ⊆ Comp × CLAB × Comp to define
the local behavior of a component where Comp denotes the set of components and CLAB is the set of transition labels
α generated by the following grammar:
α ::= λ
|
Γ
▷ Π(ṽ)
λ ::= Γ ▷ Π(ṽ)
|
Γ ▷ Π(ṽ)
The λ-labels are used to denote AbC output (Γ ▷ Π(ṽ)) and input (Γ ▷ Π(ṽ)) actions. The former contains the sender’s
predicate Π, that specifies the expected communication partners, the transmitted values ṽ and the portion of the sender
attribute environment Γ that can be perceived by receivers. The latter is just the complementary label selected among
all the possible ones that the receiver may accept. The α-labels include an additional label Γ
▷ Π(ṽ) to model the case
where a component is not able to receive a message. As it will be seen later, this kind of negative labels is crucial to
appropriately handle dynamic operators like choice and awareness.
Manuscript submitted to ACM
10
Y. Abd Alrahman et al.
Γ
Γ
Γ
Γ
Γ
Γ
|= tt for all Γ
̸ |= ff for all Γ
|= p(Ẽ) iff JẼKΓ ∈ JpKΓ
|= Π1 ∧ Π 2 iff Γ |= Π1 and Γ |= Π 2
|= Π1 ∨ Π 2 iff Γ |= Π1 or Γ |= Π2
|= ¬Π iff not Γ |= Π
Table 2. The predicate satisfaction
JẼKΓ = ṽ
{Π 1 }Γ = Π
Brd
Γ↓I ▷Π(ṽ)
Γ ′ |= {Π 1 [ṽ/x̃]}Γ1
Γ :I (Ẽ)@Π.P 7−−−−−−−→Γ :I (Ẽ)@Π.P
Γ1 ↓ I |= Π
Γ′ ▷Π(ṽ)
Γ1 :I Π1 (x̃).U 7−−−−−−−→⦃Γ1 :I U [ṽ/x̃]⦄
Γ |= Π
Γ ′ ̸ |= {Π[ṽ/x̃]}Γ ∨ Γ1 ↓ I ̸ |= Π ′
Rcv
′ ▷Π′ (ṽ)
Γ
λ
Γ :I ⟨Π⟩P 7−
→ Γ ′ :I P ′
Γ |= Π
Γ ̸ |= Π
Aware
FRcv
Γ1 :I Π(ṽ).U 7−−−−−−−→Γ1 :I Π(ṽ).U
λ
Γ :I P 7−
→ Γ ′ :I P ′
FBrd
′ ▷Π′ (ṽ)
Γ
Γ :I (Ẽ)@Π1 .U 7−−−−−−−−→⦃Γ :I U ⦄
′ ▷Π′ (ṽ)
Γ
FAware1
Γ :I ⟨Π⟩P 7−−−−−−−→Γ :I ⟨Π⟩P
′ ▷Π′ (ṽ)
Γ
Γ :I P 7−−−−−−−→Γ :I P
′ ▷Π′ (ṽ)
Γ
FAware2
Γ :I ⟨Π⟩P 7−−−−−−−→Γ :I ⟨Π⟩P
Table 3. Operational Semantics of Components (Part 1)
The transition relation 7−
→ is defined in Table 3 and Table 4 inductively on the syntax of Table 1. For each process
operator we have two types of rules: one describing the actions a term can perform, the other showing how a component
discards undesired input messages.
The behaviour of an attribute-based output is defined by rule Brd in Table 3. This rule states that when an output is
executed, the sequence of expressions Ẽ is evaluated, say to ṽ, and the closure Π of predicate Π1 under Γ is computed.
Hence, these values are sent to other components together with Γ ↓ I . This represents the portion of the attribute
environment that can be perceived by the context and it is obtained from the local Γ by limiting its domain to the
attributes in the interface I :
(Γ ↓ I )(a) =
(
Γ(a)
⊥
a∈I
otherwise
Afterwards, possible updates U , following the action, are applied. This is expressed in terms of a recursive function
⦃C⦄ defined below:
⦃C⦄ =
(
⦃ Γ[a 7→ JEKΓ ] :I U ⦄
Γ :I P
C ≡ Γ :I [a := E]U
C ≡ Γ :I P
where Γ[a 7→ v] denotes an attribute update such that Γ[a 7→ v](a ′ ) = Γ(a ′ ) if a , a ′ and v otherwise.
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
λ
11
λ
Γ :I P1 7−
→ Γ ′ :I P1′
Γ :I P2 7−
→ Γ ′ :I P2′
SumL
λ
Γ :I P1 + P 2 7−
→ Γ ′ :I P1′
′ ▷Π′ (ṽ)
Γ
Γ :I P1 7−−−−−−−→Γ :I P 1
λ
Γ :I P1 + P2 7−
→ Γ ′ :I P2′
′ ▷Π′ (ṽ)
Γ
Γ :I P 2 7−−−−−−−→Γ :I P2
′ ▷Π′ (ṽ)
Γ
SumR
FSum
Γ :I P1 + P2 7−−−−−−−→Γ :I P1 + P 2
λ
Γ :I P 1 7−
→ Γ ′ :I P ′
λ
λ
Γ :I P1 | P2 7−
→ Γ ′ :I P ′ | P 2
′ ▷Π′ (ṽ)
Γ
Γ :I P2 7−
→ Γ ′ :I P ′
IntL
Γ :I P1 7−−−−−−−→Γ :I P 1
λ
Γ :I P1 | P2 7−
→ Γ ′ :I P 1 | P ′
′ ▷Π′ (ṽ)
Γ
Γ :I P 2 7−−−−−−−→Γ :I P2
′ ▷Π′ (ṽ)
Γ
IntR
FInt
Γ :I P 1 | P2 7−−−−−−−→Γ :I P1 | P2
λ
Γ :I P 7−
→ Γ ′ :I P ′
λ
K ≜P
Γ :I K 7−
→ Γ ′ :I P ′
Rec
′ ▷Π′ (ṽ)
Γ
Γ :I P 7−−−−−−−→Γ :I P
′ ▷Π′ (ṽ)
Γ
′ ▷Π′ (ṽ)
Γ
K ≜P
FRec
Γ :I K 7−−−−−−−→Γ :I K
FZero
Γ :I 0 7−−−−−−−→Γ :I 0
Table 4. Operational Semantics of Components (Part 2)
Example 2.3. In Example 2.2, an attribute-based output is used by components to communicate their colour. If
we take component C 1 = Γ1 : {id,N} F 1 , with Γ1 = {id = 0, N = {1, 2}, colour = 0, round = 0, . . .} and F 1 =
(“try”, this.colour, this.round)@(this.id ∈ N).[this.send := f alse]F , it is easy to see that, by using rule Brd, the
following transition can be derived:
{id=0,N={1,2} }▷0∈N(“t ry”,0,0)
C 1 7−−−−−−−−−−−−−−−−−−−−−−−−−−−→ Γ1 [send 7→ ff] : {id,N} F
Rule Brd is not sufficient to fully describe the behaviour of an output action; we need another rule (FBrd) to model
the fact that all inputs are discarded in case only output actions are possible.
Rule Rcv governs the execution of input actions. It states that a message can be received when two communication
constraints are satisfied: the local attribute environment restricted to interface I (Γ1 ↓ I ) satisfies Π, the predicate used by
the sender to identify potential receivers; the sender environment Γ ′ satisfies the receiving predicate {Π1 [ṽ/x̃]}Γ1 . When
these two constraints are satisfied the input action is performed and the update U is applied under the substitution
[ṽ/x̃]. The predicate satisfaction relation, |=, is reported in Table 2. Rule FRcv states that an input is discarded when
the local attribute environment does not satisfy the sender’s predicate, or the receiving predicate is not satisfied by the
sender’s environment.
Manuscript submitted to ACM
12
Y. Abd Alrahman et al.
Example 2.4. Let us consider the component C 2 = Γ1 : {id,N} T ′ where Γ1 (id) = 2, Γ1 (round) = 2, Γ1 (counter) = 0 and
T ′ = ((x = “try”) ∧ (this.id > id) ∧ (this.round = z))(x, y, z).[counter := counter + 1]T
The following transition can be derived by using rules Rcv:
{id=1, ... }▷id∈ {0,2}(“t r y”,1,2)
C 2 7−−−−−−−−−−−−−−−−−−−−−−−−−−→ Γ1 [counter 7→ 1] : {id,N} T
Instead, the following two negative transitions can be derived by using FRcv:
{id=3, ... }▷id∈
{0,2}(“t r y”,1,2)
C 2 7−−−−−−−−−−−−−−−−−−−−−−−−−−→ C 2
{id=1, ... }▷id∈
{0,3}(“t r y”,1,3)
C 2 7−−−−−−−−−−−−−−−−−−−−−−−−−−→ C 2
In the first transition the message is discarded because sender’s id = 3 and predicate this.id > id is not satisfied. While
in the second one message is discarded because sender predicate id ∈ {0, 3} is not satisfied by Γ1 ↓ {id, N}.
The behaviour of a component Γ :I ⟨Π⟩P is the same as of Γ :I P only when Γ |= Π, while the component is inactive
when Γ ̸ |= Π. This is rendered by rules Aware, FAware1 and FAware2.
Rules SumL, SumR, and FSum describe behaviour of Γ :I P1 + P2 . Rules SumL and SumR are standard and just say that
Γ :I P 1 + P2 behaves nondeterministically either like Γ :I P1 or like Γ :I P2 . A message is discarded by Γ :I P1 + P2 if and
only if both P1 and P 2 are not able to receive it. We can observe here that the presence of discarding rules is fundamental
to prevent processes that cannot receive messages from evolving without performing actions. Thus dynamic operators,
that are the ones disappearing after a transition like awareness and choice, persist after a message refusal.
The behaviour of the interleaving operator is described by rules IntL, IntR and FInt. The first two are standard
process algebraic rules for parallel composition while the discarding rule FInt has a similar interpretation as of rule
FSum: a message can be discarded only if both the parallel processes can discard it.
Finally, rules Rec, FRec and FZero are the standard rules for handling process definition and the inactive process.
The latter states that process 0 always discards messages.
2.2.2 Operational semantics of systems. The behaviour of an AbC system is described by means of the transition
relation −
→ ⊆ Comp × SLAB × Comp, where Comp denotes the set of components and SLAB is the set of transition
labels λ which are generated by the following grammar:
λ ::= Γ ▷ Π(ṽ)
|
Γ ▷ Π(ṽ)
The definition of the transition relation −
→ is provided in Table 5 and relies on the rules Comp, FComp, Sync, ComL
and ComR.
Rules Comp and FComp depends on relation 7−
→ and are used to lift the effect of local behaviour to the system level.
The former states that the relations 7−
→ and −
→ coincide when performing either an input or an output actions, while
rule FComp states that a component Γ :I P can discard a message and remain unchanged. However, we would like to
stress that the system level label of FComp coincides with that of Comp in case of input actions, which means that
externally it cannot be percieved whether a message has been accepted or discarded.
Rule Sync states that two parallel components C 1 and C 2 can receive the same message.
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
′ ▷Π′ (ṽ)
Γ
λ
Γ :I P 7−
→ Γ ′ :I P ′
Comp
λ
Γ :I P −
→ Γ ′ :I P ′
13
Γ :I P 7−−−−−−−→Γ :I P
Γ′ ▷Π′ (ṽ)
Γ :I P −−−−−−−→ Γ :I P
Γ▷Π(ṽ)
FComp
Γ▷Π(ṽ)
C 1 −−−−−−→ C 1′ C 2 −−−−−−→ C 2′
Γ▷Π(ṽ)
Sync
C 1 ∥ C 2 −−−−−−→ C 1′ ∥ C 2′
Γ▷Π(ṽ)
Γ▷Π(ṽ)
C 1 −−−−−−→ C 1′ C 2 −−−−−−→ C 2′
Γ▷Π(ṽ)
C 1 ∥ C 2 −−−−−−→ C 1′ ∥ C 2′
Γ▷Π(ṽ)
ComL
Γ▷Π(ṽ)
C 1 −−−−−−→ C 1′ C 2 −−−−−−→ C 2′
Γ▷Π(ṽ)
ComR
C 1 ∥ C 2 −−−−−−→ C 1′ ∥ C 2′
Table 5. Operational Semantics of Systems
Finally, rule ComL (and its symmetric variant ComR) governs communication among two parallel components C 1
and C 2 : If C 1 sends a message then (C 2 can receive it by applying rule Comp). However, C 2 has also the possibility of
discarding the message by applying rule FComp.
2.3
Expressiveness of AbC Calculus
In this section, we would like to provide some evidences of the expressive power of the AbC calculus by showing
how different communication models and interaction patterns can be easily modelled in AbC. Indeed, we think that
attribute-based communication can be used as a unifying framework to encompass a number of communication models.
Further details regarding the actual implementation of the material, presented in this section, can be found in the
a
Webpage of Ab CuS3 : a Java API for the AbC calculus.
Encoding channel-based interaction. We show how one-to-many channel-based interaction can be encoded in the
AbC calculus. It may seem tempting to model a channel name as an attribute in AbC, however it turns out not to be
the case. The reason is that in channel-based communication, a channel, where the exchange happens, is instantly
enabled at the time of interaction and is disabled afterwards. This feature is not present in AbC since attributes are
persistent in the attribute environment and cannot be disabled at any time (i.e., attribute values are always available to
be checked against sender predicates). However, this is not a problem because we can exploit the fact that the receiving
predicates in AbC can check the values in the received message. The key idea is to use structured messages to select
communication partners where the name of the channel is rendered as the first element in the message; receivers only
accept messages with attached channels that match their receiving channels. Actually, attributes do not play any role in
such interaction so we assume components with empty environments and interfaces i.e., ∅ : ∅ P. In what follows, we use
[P] to denote ∅ : ∅ P. Thus a pair of processes, one willing to receive on channel a and the other willing to send on the
same channel, can be modeled as follows:
[(x = a)(x, y).P] ∥ [(a, msд)@(tt).Q]
3
http://lazkany.github.io/AbC/
Manuscript submitted to ACM
14
Y. Abd Alrahman et al.
To show the feasibility of encoding broadcast channel-based calculi into AbC, we have encoded the bπ -calculus [Ene and
Muntean 2001] into AbC. The bπ -calculus has been chosen because it uses broadcast instead of binary communication
as a basic primitive for interaction which makes it a sort of variant of value-passing CBS [Prasad 1991]. Furthermore,
channels in bπ -calculus can be communicated like in the point-to-point π -calculus [Milner et al. 1992] which is
considered as one of the richest paradigms introduced for concurrency so far.
Based on a separation result presented in [Ene and Muntean 1999], it has been proven that bπ -calculus and π -calculus
are incomparable in the sense that there does not exist any uniform, parallel-preserving translation from bπ -calculus
into π -calculus up to any “reasonable” equivalence. On the other hand, in π -calculus a process can non-deterministically
choose the communication partner while in bπ -calculus it cannot. Proving the existence of a uniform and parallelpreserving encoding of bπ -calculus into AbC up to some reasonable equivalence ensures at least the same separation
results between AbC and π -calculus.
The full encoding, the formal definition which specifies what properties are preserved by this encoding, and a proof
for its correctness up to a specific behavioural equivalence can be found in [Alrahman et al. 2017a].
Encoding interaction patterns. Also group-based [Agha and Callsen 1993; Chockler et al. 2001; Holbrook and Cheriton
1999] and publish/subscribe-based [Bass and Nguyen 2002; Eugster et al. 2003] interaction patterns can be naturally
rendered in AbC. In the group-based model, when an agent wants to send a message to all elements of a group, it
attaches the name or a reference to the group in the message and the message is propagated using this reference. In the
publish/subscribe model, there are two types of agents: publishers and subscribers and there is an exchange server
that mediates their interactions. For instance, in topic-based publish/subscribe models [Eugster et al. 2003], publishers
produce messages tagged with topics and send them to the exchange server which is responsible for filtering and
forwarding these messages to interested subscribers. Subscribers simply register their interests to the exchange server
and receive messages according to their interests. Since these interaction patterns do not have formal descriptions, we
proceed by relying on examples.
We start with group-based interaction patterns and show that when using a group name as an attribute in AbC, the
constructs for joining or leaving a group can be modelled as attribute updates, like in the following example, where we
assume that initially we have Γ1 (дroup) = b, Γ2 (дroup) = a, and Γ7 (дroup) = c:
Γ1 : {дr oup } (msд)@(дroup = a).0 ∥
Γ2 : {дr oup } (дroup = b)(x).0 | set(this.дroup, c)0 ∥ . . .
∥ Γ7 : {дr oup } (дroup = b)(x) | set(this.дroup, a)0
Component 1 wants to send the message “msд” to group “a”. Only Component 2 is allowed to receive it as it is the only
member of group “a”. Component 2 can leave group “a” and join “c” by performing an attribute update as reported on
the right hand side of the interleaving operator | . On the other hand, if Component 7 joined group “a” before “msд” is
emitted then both of Component 2 and Component 7 will receive the message.
It is worth mentioning that a possible encoding of group communication into bπ -calculus has been introduced
in [Ene and Muntean 2001]. The encoding is relatively complicated and does not guarantee the causal order of message
reception. “Locality” is neither a first class construct in bπ -calculus nor in AbC. However, “locality” (the group name, in
this case) can be naturally modeled as an attribute in AbC; in bπ -calculus, more effort is needed.
Publish/subscribe interaction patterns can be considered as special cases of the attribute-based ones. For instance, a
natural modeling of the topic-based publish/subscribe model [Eugster et al. 2003] into AbC can be obtained by allowing
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
15
publishers to broadcast messages with true predicates (i.e., satisfied by all subscribers) and requiring subscribers to
check compatibility of the exposed publishers attributes with their subscriptions, like in the example below:
Γ1 : {topic } (msд)@(tt).0 ∥ Γ2 : {subscr ipt ion } (topic = this.subscription)(x).P ∥
. . . ∥ Γn : {subscr ipt ion } (topic = this.subscription)(x).Q
The publisher broadcasts the message “msд” tagged with a specific topic for all subscribers (predicate “tt" is satisfied by
all); subscribers receive the message if the topic matches their subscription.
In the next section, we present a case study inspired by an industrial application of CAS. We focus on the role of the
communication primitives and of the external environment in determining the communication between interacting
components. Moreover, we use the case study to motivate the new communication primitives of AbC and to show that
attribute-based communication is appropriate for handling interactions in CAS.
3
STABLE ALLOCATION IN CONTENT DELIVERY NETWORKS
This case study is based on the distributed stable allocation algorithm adopted by Akamai’s Content Delivery Network
(CDN) [Maggs and Sitaraman 2015]. Akamai’s CDN is one of the largest distributed systems in the world. It has currently
over 170,000 servers located in over 1300 networks in 102 countries and serves 15-30 % of all Web traffic. To avoid
dealing with billions of clients individually, Akamai divides the clients of the global internet into groups, called map
units each having a specific demand, based on their locations and traffic types. Also content servers are grouped into
clusters, and each cluster is rated according to its capacity, latency, etc. Map units prefer highly rated clusters while
clusters prefer low demand map units. The goal of global load balancing is to assign map units to clusters such that
preferences are accounted for and constraints are met.
The allocation algorithm in [Maggs and Sitaraman 2015] is a slight variant of the original Stable Marriage Problem
(SMP), reported in [Gale and Shapley 1962]. The goal of the original algorithm is to find a stable matching between two
equally sized sets of elements given an ordering of preferences of each element of the two sets. Each element in one set
has to be paired to an element in the opposite set in such a way that there are no two elements of different pairs which
both would rather have each other than their current partners. When there are no such pairs of elements, the set of
pairs is deemed stable. A natural and straightforward AbC implementation of the original stable marriage problem can
be found in [Alrahman et al. 2016b].
The variant considered in [Maggs and Sitaraman 2015] allows (1) more map units to be assigned to a single cluster
and (2) map units to rank only the top dozen, or so, clusters that are likely to provide good performance. The first feature
is a typical generalisation [Iwama and Miyazaki 2008] of the original SMP, while the second is a mere simplification of
the problem. Implementing these features in AbC does not pose any challenge, but it would make the example more
verbose. Actually, our implementation of the original problem in [Alrahman et al. 2016b] needs only to be extended
to consider an extra attribute, named capacity, necessary for determining when a cluster should stop engaging with
more map units. Moreover, the map units assigned to a cluster should be ordered according to their demands, so that
a dissolve message goes first to the most demanding map units when necessary. Furthermore, these features do not
add much to the original SMP; they still require map units and clusters to have predefined lists of preferences such
that only ranked elements can participate in the algorithm. Obviously, this implies that one cannot take advantage of
dynamic creation of new clusters.
Manuscript submitted to ACM
16
Y. Abd Alrahman et al.
In this article, we consider a more interesting variant of SMP [Gale and Shapley 1962] that is better suited for the
dynamicity of CDN. In this variant, the arrival of new clusters is considered, it is not required that elements know
each other, and no predefined preference list is assumed. Notice that, in these settings, point-to-point interaction is
not possible because elements are not aware of each other and the choice of implicit multicast is crucial. Indeed, in
our variant, elements express their interests in potential partners by relying on their attributes rather than on their
identities. In essence, an element of one set communicates with elements of the opposite set using predicates. Two
parties that agree on some predicates form a pair, otherwise they keep looking for better partners. A pair splits only
if one of its elements can find a better partner willing to accept its offer. In this way, preferences are represented as
predicates over the values of some attributes of the interacting partners.
In this scenario, we consider the values of attributes demand, for a map unit, and ratinд, for a cluster, as a means to
derive the interaction. For simplicity, these attributes can take two different values: high (H ) and low (L). An element in
the system can be either a Unit or a Cluster. Units start the protocol by communicating with clusters in the quest of
finding an element that satisfies their highest expectations. If no cluster accepts the offer, a unit lowers its expectations
and proposes again until a partner is found. Clusters are always willing to accept proposals from any unit that enhances
their levels of satisfaction. In case of a new arrival, some pairs of elements might dissolve if the new arrival enhances
their levels of satisfaction. This means that not all pairs in the system are required to split on new arrivals; only
those interested will do so. The system level goal (the emergent behaviour) is to construct a set of stable pairs from
elements of different types by combining the behaviour of individual elements in the system through message passing.
Mathematically speaking, the problem consists of computing a function at the system level by combining individual
element behaviours, without relying on a centralised control. Notice that since map units initiate the interaction, the
solution is a “map-unit-optimal", as proved in [Gale and Shapley 1962], which is a property that fits with the CDN’s
goal of maximising performance for clients.
Allowing new arrivals is crucial to guarantee scalability and open-endedness while communicating based on
predicates rather than on identities or ranks is crucial to deal with anonymity. The actual implementation of this
algorithm alongside with experiment results can be found in the Webpage of GoAt4 : a Go API for the AbC calculus.
There we also provide a short tutorial to provide some intuition about how to use this API for programming.
The system in our attribute-based scenario can be modelled in AbC as the parallel composition of existing units
and clusters (i.e., U niti ∥ . . . ∥ U nitn ∥ Cluster i ∥ . . . ∥ Cluster n ). Notice that units and clusters interact in a distributed
setting without any centralised control. Each element is represented as an AbC component. A unit, U niti , has the form
Γi :I P where Γi represents its attribute environment, I represents its interface where I = {demand, idi }, and the process
P represents its behaviour. A cluster, Cluster i , has the form Γr :I ′ Q where Γr represents its attribute environment, I ′
represents its interface where I ′ = {ratinд, idr }, and the process Q represents its behaviour.
In addition to the attributes demand and ratinд, mentioned above, the attribute environments of units and clusters
contain the following attributes:
partner: current partner’s identity; in case they are not engaged, the value is −1;
exPartner: previous partner;
idi and idr : the identity of units and clusters, respectively;
ref: current preference, 0 for high rating and 1 for low rating, initially re f = 0;
4 https://github.com/giulio-garbi/goat
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
17
success: a boolean attribute which is set to true when an accept message from a cluster is received, initially
success = ff;
arrival: a boolean attribute which is set to true when an arrival message is received, initially arrival = ff;
ack: a boolean attribute which is set to false when a negative acknowledgment message is received, initially
ack = tt;
dissolve: a boolean attribute which is set to true when a dissolve message is received, initially dissolve = ff;
rank: an integer flag used to rank the current partner: 0 for high and 1 for low, initially rank = 2;
bof: an integer flag used to rank the new arrival: 0 for high and 1 for low. Initially bo f = 2;
lock: an integer flag used to implement a lock within a single component, initially lock = 0;
timer: a counter implementing a timer of a single component. Initially timer > time_out where time_out is a
constant number representing the number of steps a process can take before specific event can occur.
The behaviour of a unit component is specified by the process P which is the parallel composition of the processes
I, T , M, and N . Process I defines a proposal process, process T defines a timer process, process M defines a message
handler process, and process N defines a negative acknowledgment process. The behaviour of the proposal process, I ,
is defined below:
I ≜ ⟨re f = 0⟩(“propose”, this.demand, this.idi , 0)@Πh .[timer := 0, dissolve := ff]I ′
+
⟨re f = 1⟩(“propose”, this.demand, this.idi , 1)@Πl .[timer := 0, dissolve := ff]I ′
Process I sends a proposal message to all components that either satisfy predicate Πh or predicate Πl , depending on
the current value of the re f attribute. The predicate Πh represents high expectations where Πh = (ratinд = “H ”) while
the predicate Πl represents low expectations where Πl = (ratinд = “L”). Notice that the branches of process I encode
the preferences of a unit and the selection of any of them depends on the run-time value of re f . These branches can be
thought of as context-dependent behavioural variations in Context-Oriented Programming [Hirschfeld et al. 2008].
Since the initial value of re f is 0, the process proceeds with the first branch. The proposal message contains a proposal
label, “propose”, the values of attributes demand, idi , and re f of the unit respectively. The sent value of re f will be
used later to decide if an accept message from a cluster is stale (i.e., the received value of re f is different from the
current value of re f ). By sending a proposal message, the timer and the dissolve attributes are reset. After this step,
process I evolves to I ′ which is a proposal handler process. Reseting the timer attribute will decide how long process I
should wait before proposing again. The timer process is reported below:
T ≜ ⟨timer < time_out⟩()@ff.[timer := timer + 1]T
+
⟨timer = time_out⟩()@ff.[timer := timer + 1, re f := (re f + 1)%2, success := ff]T
T keeps increasing the value of the timer autonomously with silent moves (i.e., ()@ff) as long as its value is less
than the constant time_out. By assuming fairness and a reasonably large value of time_out, we can guarantee that an
accept message from a cluster, that satisfies the sent proposal message (if there is any), is received before a timeout. If a
timeout is reached, the timer process lowers the preference of the unit and resets the value of the success attribute
to false. As we have seen so far, the code of processes is infrastructure-agnostic, i.e., it does not contain addresses or
channel names. It is completely data-centric and relies on the run-time characteristics of the interacting partners.
Manuscript submitted to ACM
18
Y. Abd Alrahman et al.
In the proposal handler process, reported below, we can understand the role of the awareness construct ⟨Π⟩ as an
environmental parameter used to influence the behaviour of a unit at run-time.
I ′ ≜ ⟨lock = 0 ∧ ¬success ∧ timer > time_out⟩I
+
⟨lock = 0 ∧ dissolve⟩I
+
⟨lock = 0 ∧ arrival ∧ (bo f ≤ rank − 1)⟩
(“dissolve”)@(idr = this.partner ).
[arrival := ff, success := ff, re f := bo f , bo f := 2, rank := 2,
exPartner := partner, partner := −1]I
The process blocks executing until one of three events occurs. If no lock is acquired (lock = 0), no accept message
from a cluster is received (¬success), and a timeout occurs (timer > time_out), process I ′ calls the proposal process I
again by considering the new value of re f modified by the timer process. If no lock is acquired and a dissolve message
from the current partner is received (dissolve = true), process I ′ calls the proposal process I again. Finally, if no lock is
acquired and an arrival message is received (arrival = true), and the rank of the new arrival is better than the rank of
the current partner (bo f ≤ rank − 1), the process sends a dissolve message that contains a dissolve label to its partner,
sets the value of attribute exPartner to the value of attribute partner and the value of attribute re f to the value of
attribute bo f , and resets the values of attributes arrival, success, bo f , rank, and partner to their initial values. Process
I ′ calls the proposal process I again.
The negative acknowledgment process is reported below:
N ≜ (x = “accept”∧((z , re f ) ∨ success))(x, y, z).
(“ack”, −1)@(idr = y).0 | N
N ensures that after a successful reception of a first accept message from a cluster (i.e., success = tt) all other accept
messages to this unit are discarded (i.e., it sends an acknowledgement message with id=-1 which is interpreted by
a cluster as a negative acknowledgement). Notice that the process replicates itself every time an accept message is
received to be always able to catch all accept messages; the condition (z , re f ) is crucial to discard stale messages.
The message handler process is reported below:
M ≜ ⟨¬success⟩(x = “accept”∧ z = re f )(x, y, z).
[lock := 1, success := tt, rank := re f , exPartner := partner, timer := time_out + 1]
(“ack”, y)@tt.[lock := 0, parnter := y]M
+
(x = “dissolve” ∧ idi = partner )(x).
[dissolve := tt, success := ff, re f := 0, rank := 2, partner := −1]M
+
⟨partner , −1⟩(x = “arrived” ∧ ratinд = “H ”)(x).[arrival := tt, bo f := 0]M
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
19
+
⟨bo f , 0∧ partner , −1⟩(x = “arrived” ∧ ratinд = “L”)(x).[arrival := tt, bo f := 1]M
M can respond to one of three events: if no accept message is received yet (i.e., ¬success) and a new one arrives, the
process receives the accept message only if the message is not stale (z = re f ), acquires a lock (lock := 1), sets the value
of attribute success to true, the value of attribute rank to the value of attribute re f , the value of attribute exPartner to
the value of attribute partner and turns off the timer. The process proceeds by sending an acknowledgement message
to the cluster, releasing the lock, setting the value of attribute partner to the received cluster identity, and then the
process continues as M. Notice that the lock is important to ensure that the handler process I ′ does not proceed before
all required attributes are assigned the right values. In this case, the first branch of process M is executed atomically
with respect to the co-located processes in the unit component.
If a dissolve message from the current partner is received, the process sets the value of attribute dissolve to true,
resets the values of attributes success, re f , rank, and partner to their initial values and continues as M. If an arrival
message is received, the arrival attribute is set to true and the value of attribute bo f is set to 0 if the attached attribute
value is high ratinд = “H ” otherwise the value of attribute bo f is set to 1 and the process continues as M. Notice
that arrivals with low value of attribute ratinд cannot override the ones with high values. This is guaranteed by the
condition bo f , 0 on the last branch. Also an arrival message can have an effect on the behavior of a unit only if the
unit is already engaged otherwise the arrival message is just discarded which is ensured by the condition partner , −1
in the last two branches of process M.
The coordination among processes running in a single unit is made possible by relying on shared attributes, awareness
constructs, and attribute updates to implement proper lock mechanisms.
The behaviour of a cluster component is specified by the process Q which is the parallel composition of the processes
A, R, and D. Process A defines an arrival process, process R defines a proposal reception process, and process D defines
a dissolve handler process. The arrival process, A, is defined below:
A ≜ (“arrived”)@tt.0
This is the first process that takes a step when a cluster component wants to be involved in the allocation procedure. It
simply sends an arrival message to all components in the system and then terminates. The arrival message contains an
arrival label, “arrived”.
The proposal reception handler is reported below:
R ≜ (x = “propose”)(x, y, z, n).R ′ | R
The process simply receives a proposal message and evolves to the reception handler process R ′ . Notice that process
R replicates itself every time a message is received to ensure that no proposal is lost. The reception handler process,
R ′ reported below, checks if no lock is acquired and the rank of the communicated value of attribute demand of the unit
is less than the rank of the current partner (rank(y) < rank), the process sends an accept message to the unit where the
proposal message came from, addressing it by its identity and acquires a lock, otherwise the process R ′ terminates.
The function rank(y) takes the value of demand as a parameter and returns 1 if y = “H ” and 0 otherwise. The accept
message contains an acceptance label, “accept”, the identity of the current cluster, and the reference of the proposal
message. The process then waits to receive either acknowledgement message or a negative one from the unit. If an
acknowledgement is received, the value of attribute partner is set to the received identity of the unit, the rank is set to
Manuscript submitted to ACM
20
Y. Abd Alrahman et al.
the value returned by rank(y), the lock is released and process R ′ terminates. Notice that a dissolve message is sent to
the current partner before releasing the lock in case a cluster is already engaged. If a negative acknowledgement is
received then the lock is released. The lock is needed to ensure that concurrent proposals are handled sequentially
which is important to guarantee a consistent state of the attribute environment.
R ′ ≜ ⟨lock = 0⟩(
if (rank(y) < rank) then {
(“accept”, idr , n)@(idi = z).[lock := 1](
if (partner , −1) then {
(e = “ack” ∧ f = idr ∧ idi = z)(e, f ).
[exPartner := partner, partner := z, rank := rank(y)]
(“dissolve”)@(idi = this.exPartner ).[lock := 0]0
} else {
(e = “ack” ∧ f = idr ∧ idi = z)(e, f ).
}
[partner := z, rank := rank(y), lock := 0]0
+
(e = “ack” ∧ f , idr ∧ idi = z)(e, f ).[lock := 0]0)
} else 0)
The dissolve handler process is reported below:
D ≜ (x = “dissolve” ∧ idi = partner )(x).
[rank := 2, exPartner := partner, partner := −1]D
When a dissolve message from a cluster’s partner is received, the value of attribute exPartner is set to the value of
attribute partner, the values of attributes rank and partner are reset to their initial values, and the process calls itself.
Remark 3.1. In [Gale and Shapley 1962], the authors showed that their algorithm terminates with a matching that is
stable after no more than n2 proposals, where n is the number of proposing elements, i.e., the algorithm has O(n 2 ) worst-case
complexity. In our variant, it should be clear that the worst case complexity is also O(n 2 ) even after relaxing the assumptions
of the original algorithm, i.e., no predefined preference lists and components are not aware of the existence of each other, so
point-to-point communication is not possible. Interestingly, the complexity is still quadratic even if we consider a blind
broadcast mechanism where proposals are sent to all components in the system except for the sender unit. In this way,
for n-units, n-clusters, and a constant L related to the preferences of a unit where L is computed based on the number of
branches in the proposal process I , we have that each unit can send at most L ∗ (2n − 1) proposals.
Below we first provide a simple interaction fragment to show how the semantics rules are applied in our stable
allocation (Example 3.1). We then show a possible and complete execution of the protocol by considering two pairs of
components (Example 3.2).
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
21
Example 3.1 (Interaction fragment). A unit, say I 1 with re f = 1, demand = “H ”, and idi = 1, can send a proposal and
generate the following transition by applying rule Comp:
Γi : {demand,idi } I | T | M | N
{demand=“H ”, id i =1}▷Πl (“pr opose”, “H ”, 1, 1)
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→
Γi [timer 7→ 0, dissolve 7→ ff] : {demand,idi } I ′ | T | M | N
Any cluster, say r 1 , can receive the message by applying rule Comp and generate this transition:
Γr : {r at inд,idr } R | A | D
{demand =“H ”, id i =1}▷Πl (“pr opose”, “H ”, 1, 1)
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→
Γr : {r at inд,idr } R ′ [“propose”/x, “H ”/y, 1/z, 1/n] | R | A | D
Now the overall system evolves by applying rule ComL as follows:
{demand=“H ”, id i =1}▷Πl (“pr opose”, “H ”, 1, 1)
S −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→ Γi [timer 7→ 0, dissolve 7→ ff] : {demand,idi } I ′ | T | M | N
∥ . . . ∥ Γr : {r at inд,idr } R ′ [“propose”/x, “H ”/y, 1/z, 1/n] | R | A | D
□
Example 3.2. Let us consider two map units M 0 and M 1 and two clusters C 0 and C 1 with attribute environments
defined as follows: Γm0 = {demand = “H ”, idi = m 0 }, Γm1 = {demand = “L”, idi = m 1 }, Γc 0 = {ratinд = “H ”, idr = c 0 },
and Γc 1 = {ratinд = “L”, idr = c 1 }. Below we show the execution of the stable allocation protocol, presented above,
until a stable matching is reached.
• C 0 broadcasts an arrival message through process A by applying Comp.
• C 1 broadcasts an arrival message through process A by applying Comp. Because no unit is engaged yet, all
arrival messages are discarded.
• M 0 proposes for a cluster with a high rating by applying Comp and waits for a response.
• C 0 receives the proposal because its rating is high by applying Comp.
• M 1 proposes for a cluster with a high rating by applying Comp and waits for a response.
• C 0 receives the proposal because its rating is high by applying Comp.
• C 0 handles the first proposal by process R ′ . It acquires a lock and sends an accept message to M 0 by applying
rule Comp. Notice that the second proposal cannot be handled before the lock is released.
• M 0 receives the message by process M, acquires a lock and sets the attributes success, rank and exPartner
appropriately by rule Comp.
• M 0 sends an acknowledgement message to all possible components and includes the identity of C 0 in the
message. It sets its partner to C 0 and releases the lock.
• Clearly, all components are potential receivers for this message, however, only C 0 receives it because it is the
only component waiting an acknowledgement that contains its identity. By doing so, C 0 sets its partner to M 0 ,
its rank to 1 (because rank(“H ”) = 1) and releases the lock.
• C 0 handles the second proposal by process R ′ again. It acquires a lock and sends an accept message to M 1
because rank(“L”) = 0 < 1, by applying rule Comp and waits for an acknowledgement.
• M 1 is the only component that can receive the accept message because it satisfies the sending predicate by
applying rule Comp. It receives the message by process M, acquires a lock and sets the attributes success, rank
and exPartner appropriately by rule Comp.
Manuscript submitted to ACM
22
Y. Abd Alrahman et al.
• M 1 sends an acknowledgement message to all possible components and includes the identity of C 0 in the
message. It sets its partner to C 0 and releases the lock.
• C 0 receives the acknowledgement, sets its exPartner to current partner, its partner to M 1 and its rank to 0 by
rule Comp.
• C 0 sends a dissolve message to its exPartner (i.e., M 0 ) and releases the lock by rule Comp.
• M 0 receives the dissolve message by the second branch of process M which sets the dissolve attribute to true
and resets the attributes success, re f , rank, and partner to their initial values. By doing so, the second branch
of process I ′ will be enabled and M 0 is ready to propose again.
• M 0 proposes again for a cluster with a high rating by applying Comp and waits for a response.
• C 0 receives the proposal because its rating is high by applying Comp.
• C 0 drops the proposal by the last branch of process R ′ , because it is satisfied with its current partner, i.e., its
rank<1.
• No component can actually respond to the proposal. M 0 keeps waiting until a timeout occurs, i.e., the second
branch of process T is enabled. The timer is turned off, the ref attribute value is set to 1 and the success attribute
is turned off. M 0 can now propose for a cluster with a low rating. In this case, only C 1 accepts the proposal in a
similar way as described before. Notice that (M 0 , C 1 ) and (M 1 , C 0 ) are actually stable pairs.
4
DISCUSSION
In this section, we first explain how AbC has helped us in dealing with the collective adaptive features that were evident
in the previous case studies, finally we comment on other features of AbC code, namely Compositionality and Code
Extensibility.
Distribution is naturally obtained because an AbC system has no centralised control; and it is represented as the
parallel composition of independent components that can mutually influence each other only through message-exchange.
For instance in Section 3, the overall system is defined as the parallel composition of existing units and clusters.
Awareness is supported by the attribute environment that plays a crucial role in orchestrating the behaviour of AbC
components. It makes components aware of their own status and provides partial views of the surrounding environment.
Components behave differently under different environmental contexts. This is possible because the behaviour of AbC
processes is parametric with respect to the run-time attribute values of the component in which they are executed. The
awareness construct, ⟨Π⟩, is used as an environmental parameter to influence the behaviour of AbC components at
run-time. For instance, in Section 3, the behaviour of the proposal handler process, I ′ , is totally dependent on the values
of the attributes lock, success, timer , dissolve, arrival, bo f , and rank. In case the value of just one of these attributes
changes, the process will change its behaviour accordingly.
Adaptivity is obtained by means of the interaction predicates (both for sending and receiving) of AbC components
that can be parametrised with their local attribute values; any run-time changes of these values might influence/change
the possible set of interaction partners. Notice that the target of the dissolve message, in the third branch of the proposal
handler process I ′ in Section 3, depends on the identity of the current partner (this.partner ).
Interdependence among co-located processes can be obtained by modifying, with the attribute update construct, the
attribute environment shared among processes within a single component. A branch of process I ′ , in Section 3, is
mainly selected depending on the attribute updates performed by the timer T and the message handler M.
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
23
Collaboration is obtained by combining individual component behaviours, through message exchange, to achieve a
global goal for which a single component would not be sufficient. For instance, in Section 3, the goal was to construct
a set of stable pairs without any centralised control. The goal was achieved by allowing different components to
collaborate through message passing; each component contributed, and the combined behaviour of all components was
necessary, to reach the overall system goal.
Anonymity is obtained by allowing the interaction primitives (both send and receive) to rely on predicates over the
run-time attribute values of the interacting partners rather than on specific names or addresses. Thus the qualification
needed to receive a message no longer depends on available channels or addresses, but on the values of dynamic
attributes. For instance, in Section 3, the process I sends a proposal to a group of clusters whose attributes satisfy a
specific predicate (i.e., Πh or Πl ). There is no prior agreement between the units and the other clusters and the set of
candidate receivers is specified at the time of interaction in the sense that any change in the value of attribute re f
changes the set of targeted components.
Scalability and Open-endedness are guaranteed by the adoption of multiparty communication instead of the binary
one. Actually AbC supports an implicit multicast communication in the sense that the multicast group is not specified
in advance, but rather it is specified at the time of interaction by means of the available set of receivers with attributes
satisfying the sender predicate. The non-blocking nature of the AbC multicast, alongside with anonymity of interaction,
breaks the synchronisation dependencies between senders and receivers and components can join or leave a system
without disrupting its overall behaviour. For instance, new clusters, as defined by process A, Section 3, can join the
allocation procedure at run-time without causing any disruption to the overall system behaviour. Actually, the third
branch of process I ′ of a unit is used as an adaptation mechanism to handle new arrival of clusters. Notice that the
condition bo f ≤ rank − 1 of that branch ensures that a unit responds only to new clusters that enhance its satisfaction.
This means that new arrivals affect only specific components in the system.
It is worth mentioning that the initial behaviour of all components with the same type (e.g., Unit or Cluster in
Section 3) is exactly the same. However, since this behaviour is parametrised with respect to the run-time attribute values
of each component, components might exhibit different behaviours. This means that the context where a component is
executing has a great impact on its behaviour, in the sense that it can disable or enable specific behaviours depending
on the environmental changes. In some sense, the behaviour of a component evolves based on contextual conditions.
Components do not need to have complex behaviour to achieve adaptation at system level. Complex behaviour can be
achieved by combining the local behaviour of individual components.
We would also like to stress that AbC code naturally supports compositionality and extensibility.
Compositionality: a component is the basic building block of AbC programs and any program can be broken down
into a set of individual components which can only interact by exchanging messages. This simplifies verification
because components can be analysed individually and their external behaviour can be assessed by their observable
communication capabilities. Actually, we can abstract from the internal behaviour of an AbC component and only
consider its observable behaviour when it interacts with other components. Thus, a component can be treated as a
blackbox that acts and reacts to its environment. We refer interested readers to our theoretical results in [Alrahman
et al. 2016a, 2017a].
Code Extensibility: we would like to stress that AbC code can be easily extended in the sense that new alternative
behaviours can be added and removed by modifying the attribute environments and the interfaces of components.
Also component behaviours can be adapted without changing the internal structure of their running processes. For
Manuscript submitted to ACM
24
Y. Abd Alrahman et al.
instance in Section 3, a unit can reverse the order of its preferences by directly modifying the required value of attribute
re f in the awareness constructs and the following output actions of process I . Also if we add a public attribute to a
cluster, say location, the code of a unit can be easily adapted to consider new possible preferences. One way is to add
another alternative/branch to process I that consider the location of a cluster. Of course some attribute values have to
be adjusted accordingly. As mentioned before, these branches can be thought of as context-dependent behavioural
variations in Context-Oriented Programming [Hirschfeld et al. 2008].
5
RELATED WORK
In this section, we touch on related works concerning languages and calculi with primitives that model either collective
interaction, or multiparty interaction. We also report on other approaches to interaction in distributed systems and
show how they relate to AbC.
Some of the well-known approaches include: channel-based models [Hoare 1978; Milner 1980; Milner et al. 1992],
group-based models [Agha and Callsen 1993; Chockler et al. 2001; Holbrook and Cheriton 1999], and publish/subscribe
models [Bass and Nguyen 2002; Eugster et al. 2003]. Below we briefly report the main features and limitations of such
approaches.
Channel-based models rely on explicit naming for establishing communication links in the sense that communicating
partners have to agree on channels or names to exchange messages. This implies that communicating partners are
not anonymous to each other. Actually, communicating partners have static interfaces that determine the target of
communication e.g., binary communication CCS [Milner 1980], multiway synchronisation CSP [Hoare 1978], and
broadcast communication CBS [Prasad 1995]. In AbC the interacting partners are anonymous to each other; they do not
need to agree on channels or names to interact, but rather to rely on the satisfaction of predicates over the attributes
they expose. This makes our calculus more suitable for modeling scalable distributed systems as anonymity is a key
factor for scalability. The π -calculus [Milner et al. 1992] was developed as a way to mitigate this problem by allowing the
exchange of channel names and thus providing dynamic interfaces and additional flexibility. However, the dynamicity
of interfaces is still limited because even if generic input or output actions are allowed, they are disabled until they are
instantiated with specific channel names. This means that a process can engage in a communication only when its
actions are enabled. The broadcast bπ -calculus [Ene and Muntean 1999, 2001] is based on CBS and the π -calculus in
that it extends the former with a channel-passing mechanism. Furthermore, π -calculus and most process calculi rely on
synchronous communication that creates dependencies between senders and receivers and affects the overall scalability
of the system. AbC actions are always enabled with respect to the current attribute values of the component where
they are executing. When these values change, the interaction predicates change seamlessly and become available for
other communication partners.
In group-based models, like the one used for IP multicast [Holbrook and Cheriton 1999], the group is specified in
advance in the sense that the reference address of the group is explicitly included in the message. Groups or spaces in
the ActorSpace model [Agha and Callsen 1993] are regarded as passive containers of processes (actors) and may be
created and deleted with explicit constructs. Spaces may be nested or even overlap and can be created dynamically
at run-time. Actually, the notion of space is a first class entity in the ActorSpace model. AbC extends the ActorSpace
pattern-matching mechanism to select partners by predicates on both sides of the communication where not only the
sender can select its partner but also the receiver can decide to either receive or discard the message. The notion of
spaces/collectives in AbC is more abstract and is only specified at run-time.
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
25
In the publish/subscribe model, like the one used in NASA Goddard Mission Services Evolution Center (GMSEC)5 ,
each component can take the role of a publisher or a subscriber. Publishers produce messages and subscribers consume
them. The subscribers are indirectly addressed by the contents of sent messages. That is, a subscriber expresses its
interest independently from the publisher that produces the messages, and then it is asynchronously notified when
a message that matches its interest arrives. The propagation of messages from publishers to subscribers is obtained
by introducing an exchange server that mediates the interaction. The exchange server stores the subscriptions of
subscribers, receives messages from publishers, and forwards the messages to the correct subscribers. The result is
that publishers and subscribers are unaware of the existence of each other. Though the anonymity of interaction is an
effective solution to achieve scalability by allowing participants to enter or leave the system at anytime, scalability
issues are moved to the implementation of the exchange server. In fact, since the exchange server is responsible for
subscriptions and message filtering, it should be able to face a large number of participants with evolving subscriptions
while maintaining an acceptable level of performance.
As discussed before, the publish/subscribe model can be considered as a special case of AbC where publishers can
attach attributes to messages and send them with empty predicates (i.e., satisfied by all). Only subscribers can check the
compatibility of the attached publishers attributes with their subscriptions.
The Actor communication model, originally introduced in [Hewitt 1977; Hewitt et al. 1973], has been used to
support the development of object-based concurrent computation. Actors embody the spirit of objects and extends their
advantages to concurrent computations. As with objects where data and behavior are encapsulated to separate what an
object can do from how it does it, the actors separate the control (where and when) from the logic of a computation. The
early proposals of the Actor model in [Hewitt 1977; Hewitt et al. 1973] were rather informal. However, the definition of
actors that is commonly used today follows Agha’s definition [Agha 1986] that provides an abstract representation of
actor systems in terms of what is called Asynchronous Computation Trees, building on notions borrowed from Milner’s
work [Milner 1989]. An actor system, also called a configuration, consists of autonomous objects, called actors, and of a
collection of messages in transit. Computations in an actor system are carried out in response to received messages.
Each message contains the destination identity and the actual content of the message. As opposed to AbC components,
actors adopt binary communication which does not scale well in large systems. To interact, actors rely on identities;
messages can only be sent to those actors whose identity is known. Thus actors can only interact if they know each
other.
Programming collective and/or adaptive behavior has been studied in different research communities like those interested in Context-Oriented Programming (COP) [Hirschfeld et al. 2008] and in the Component-Based approach [Bruneton
et al. 2004]. In Context-Oriented Programming (COP), a set of linguistic constructs is used to define context-dependent
behavioural variations. These variations are expressed as partial definitions of modules that can be overridden at
run-time to adapt to contextual information. They can be grouped via layers to be activated or deactivated together
dynamically. These layers can be also composed according to some scoping constructs. Our approach is different in
that components adapt their behaviour by considering the run-time changes of the values of their attributes which
might be triggered by either contextual conditions or by local interaction. Another approach that considers behavioural
variations is adopted in the Helena framework [Klarl 2015].
The component-based approach, represented by FRACTAL [Bruneton et al. 2004] and its Java implementation,
JULIA [Bruneton et al. 2006], is an architecture-based approach that achieves adaptation by defining systems that
5 http://opensource.gsfc.nasa.gov/projects/GMSEC_API_30/index.php
Manuscript submitted to ACM
26
Y. Abd Alrahman et al.
are able to adapt their configurations to the contextual conditions. System components are allowed to manipulate
their internal structure by adding, removing, or modifying connectors. However, in this approach interaction is still
based on explicit connectors. In our approach predefined connections simply do not exist: we do not assume a specific
architecture or containment relations between components. The connectivity is always subject to change at any time
by means of attribute updates. In our view, AbC is definitely more adequate when highly dynamic environments have
to be considered.
Finally, the attribute-based interaction has been exploited in the Carma specification language [Bortolussi et al. 2015;
Loreti and Hillston 2016] which was designed based on the linguistic primitives of the AbC calculus to support quantitive
analysis of collective-adaptive systems. Send and receive operations in Carma are associated with execution rates that
can be dynamically adjusted based on environmental conditions. A Carma model can be automatically translated into
a Continuous-Time Markov process that can be simulated and thus its dynamics can be studied quantitatively. AbC
models are non-deterministic and they abstract from timing constraints. They are concerned with qualitative properties
of the system being modelled. Clearly, AbC and Carma have been designed to serve different objectives in the sense
that AbC is designed as a programming framework for CAS interactions while Carma is designed to quantitatively
assess CAS models in terms of their performance.
AbC combines the lessons learnt from the above mentioned languages and calculi, in that it strives for expressiveness
while aiming to preserve minimality and simplicity. The dynamic settings of attributes and the possibility of inspecting/modifying the environment gives AbC greater flexibility and expressiveness while keeping models as natural as
possible.
6
CONCLUDING REMARKS
We have proposed a language-based approach for programming the interaction of collective-adaptive systems by
relying on attribute-based communication. We have introduced the programming constructs of the AbC calculus and
we exploited them to show how complex and interesting case studies, from the realm of collective-adaptive systems, can
be programmed in an intuitive way. We illustrated the expressive power of attribute-based communication by showing
the natural encoding of other existing communication paradigms. We argued that the general concept of attribute-based
communication can be exploited to provide a unifying framework to encompass different communication models and
interaction patterns. Since the focus of this article was to show the expressive power of attribute-based communication
and their applicability in the context of CAS systems, we refrained from presenting theoretical results. However, full
details about behavioural theory, equational laws, and a formal proof of encoding can be found in [Alrahman et al. 2016a,
2017a]. We also provided AbC APIs to enable attribute-based interaction in well-known programming languages, e.g.,
in Java [Alrahman et al. 2016b], in Go [Alrahman et al. 2017b], and in Erlang [De Nicola et al. 2017]. In [Alrahman 2017],
we have also laid the basis for an efficient distributed implementation of the attribute-based interaction; we proposed
three different communication infrastructures, proved their correctness, modelled them as stochastic processes, and
finally evaluated their performance.
For future work, we want to investigate anonymity of interaction at the level of attribute identifiers. Indeed in
our model, components are anonymous, but the “name-dependency” challenge arises at another level, that is, the
attribute environments. In other words, the sender’s predicate should be aware of the identifiers of receiver’s attributes
in order to explicitly use them. For instance, the sending predicate loc = (1, 4) targets the components at location (1, 4),
however, different components might use different identifiers to denote their locations; this means that there should be
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
27
an agreement about the attribute identifiers among components. For this reason, appropriate mechanisms for handling
attribute directories together with identifiers matching/correspondence will be considered. These mechanisms will be
particularly useful when integrating heterogeneous applications.
We plan to develop formal tools based on AbC semantics to analyse the code generated by our AbC APIs for ensuring
safety and liveness properties. We want to study the possibility of using static analysis to discipline the interaction in
AbC and thus producing a correct by construction programs and we will also consider the more challenging problem of
specifying and verifying collective properties of AbC programs.
Acknowledgments. The authors would like to thank Francesco Tiezzi, Mirco Tribastone and Catia Trubiani for their
valuable comments and suggestions to improve the quality of this article. We would also like to thank Giulio Garbi for
debugging and implementing the case studies in GoAT.
REFERENCES
Gul Agha. 1986. Actors: A Model of Concurrent Computation in Distributed Systems. MIT Press, Cambridge, MA, USA.
Gul Agha and Christian J Callsen. 1993. ActorSpace: an open distributed programming paradigm. Vol. 28. ACM.
Yehia Abd Alrahman. 2017. A Foundational Theory for Attribute-based Communication. (2017). http://e-theses.imtlucca.it/id/eprint/214 PhD Thesis, IMT
school for Advanced Studies Lucca.
Yehia Abd Alrahman, Rocco De Nicola, Giulio Garbi, and Michele Loreti. 2017b. Attribute-based Interaction in Google Go. https://github.com/giulio-garbi/
goat. (2017).
Yehia Abd Alrahman, Rocco De Nicola, and Michele Loreti. 2016a. On the Power of Attribute-Based Communication. In Formal Techniques for Distributed
Objects, Components, and Systems - 36th IFIP WG 6.1 International Conference, FORTE 2016, Held as Part of the 11th International Federated Conference on
Distributed Computing Techniques, DisCoTec 2016, Heraklion, Crete, Greece, June 6-9, 2016, Proceedings. 1–18. https://doi.org/10.1007/978-3-319-39570-8_1
Yehia Abd Alrahman, Rocco De Nicola, and Michele Loreti. 2016b. Programming of CAS Systems by Relying on Attribute-Based Communication. In
Leveraging Applications of Formal Methods, Verification and Validation: Foundational Techniques - 7th International Symposium, ISoLA 2016, Imperial,
Corfu, Greece, October 10-14, 2016, Proceedings, Part I. Springer, 539–553. https://doi.org/10.1007/978-3-319-47166-2_38
Yehia Abd Alrahman, Rocco De Nicola, and Michele Loreti. 2017a. A Behavioural Theory for Interactions in Collective-Adaptive Systems. ArXiv e-prints
(Nov. 2017). https://arxiv.org/abs/1711.09762
Yehia Abd Alrahman, Rocco De Nicola, Michele Loreti, Francesco Tiezzi, and Roberto Vigo. 2015. A Calculus for Attribute-based Communication. In
Proceedings of the 30th Annual ACM Symposium on Applied Computing (SAC ’15). ACM, 1840–1845. https://doi.org/10.1145/2695664.2695668
S Anderson, N Bredeche, AE Eiben, G Kampis, and M van Steen. 2013. Adaptive collective systems: herding black sheep. BookSprints for ICT Research
(2013).
Michael A Bass and Frank T Nguyen. 2002. Unified publish and subscribe paradigm for local and remote publishing destinations. (June 11 2002). US
Patent 6,405,266.
Federico Bergenti, Giovanni Rimassa, and Mirko Viroli. 2003. Operational semantics for agents by iterated refinement. In International Workshop on
Declarative Agent Languages and Technologies. Springer, 37–53.
Luca Bortolussi, Rocco De Nicola, Vashti Galpin, Stephen Gilmore, Jane Hillston, Diego Latella, Michele Loreti, and Mieke Massink. 2015. CARMA:
Collective Adaptive Resource-sharing Markovian Agents. Workshop on Quantitative Aspects of Programming Languages and Systems, QAPL 2015 (2015),
16–31.
Manfred Broy. 2006. The’grand challenge’in informatics: engineering software-intensive systems. Computer 39, 10 (2006), 72–80.
Eric Bruneton, Thierry Coupaye, Matthieu Leclercq, Vivien Quéma, and Jean-Bernard Stefani. 2006. The fractal component model and its support in Java.
Software: Practice and Experience 36, 11-12 (2006), 1257–1284.
Eric Bruneton, Thierry Coupaye, and Jean-Bernard Stefani. 2004. The FRACTAL component model. Draft of specification, version 2, 3 (2004).
R. Calinescu. 2008. Implementation of a Generic Autonomic Framework. In Fourth International Conference on Autonomic and Autonomous Systems
(ICAS’08). 124–129. https://doi.org/10.1109/ICAS.2008.21
Betty HC Cheng, Rogerio De Lemos, Holger Giese, Paola Inverardi, Jeff Magee, Jesper Andersson, Basil Becker, Nelly Bencomo, Yuriy Brun, Bojan Cukic,
et al. 2009. Software engineering for self-adaptive systems: A research roadmap. In Software engineering for self-adaptive systems. Springer, 1–26.
Gregory V Chockler, Idit Keidar, and Roman Vitenberg. 2001. Group communication specifications: a comprehensive study. ACM Computing Survey
(CSUR) 33, 4 (2001), 427–469.
Frank S de Boer, Wieke de Vries, John-Jules Ch Meyer, Rogier M van Eijk, and Wiebe van der Hoek. 2005. Process algebra and constraint programming
for modeling interactions in MAS. Applicable Algebra in Engineering, Communication and Computing 16, 2-3 (2005), 113–150.
Manuscript submitted to ACM
28
Y. Abd Alrahman et al.
Rocco De Nicola, Tan Duong, Omar Inverso, and Catia Trubiani. 2017. AErlang: Empowering Erlang with Attribute-Based Communication. In
Coordination Models and Languages - 19th IFIP WG 6.1 International Conference, COORDINATION 2017, Held as Part of the 12th International Federated
Conference on Distributed Computing Techniques, DisCoTec 2017, Neuchâtel, Switzerland, June 19-22, 2017, Proceedings. 21–39. https://doi.org/10.1007/
978-3-319-59746-1_2
Rocco De Nicola, Michele Loreti, Rosario Pugliese, and Francesco Tiezzi. 2014. A formal approach to autonomic systems programming: the SCEL Language.
ACM Transactions on Autonomous and Adaptive Systems (2014), 1–29.
Cristian Ene and Traian Muntean. 1999. Expressiveness of point-to-point versus broadcast communications. In Fundamentals of Computation Theory.
Springer, 258–268.
Christian Ene and Traian Muntean. 2001. A broadcast-based calculus for communicating systems. In Parallel and Distributed Processing Symposium,
International, Vol. 3. IEEE Computer Society, 30149b–30149b.
M. F. Ercan, L. Partawijaya, and Y. F. Fung. 2006. Collective Search and Exploration with a Robot Swarm. In TENCON 2006 - 2006 IEEE Region 10 Conference.
1–4. https://doi.org/10.1109/TENCON.2006.344078
Patrick Th. Eugster, Pascal A. Felber, Rachid Guerraoui, and Anne-Marie Kermarrec. 2003. The Many Faces of Publish/Subscribe. ACM Comput. Surv. 35, 2
(June 2003), 114–131. https://doi.org/10.1145/857076.857078
Alois Ferscha. 2015. Collective adaptive systems. In Adjunct Proceedings of the 2015 ACM International Joint Conference on Pervasive and Ubiquitous
Computing and Proceedings of the 2015 ACM International Symposium on Wearable Computers. ACM, 893–895.
David Gale and Lloyd S Shapley. 1962. College admissions and the stability of marriage. The American Mathematical Monthly 69, 1 (1962), 9–15.
D. Garlan, S. W. Cheng, A. C. Huang, B. Schmerl, and P. Steenkiste. 2004. Rainbow: architecture-based self-adaptation with reusable infrastructure.
Computer 37, 10 (Oct 2004), 46–54. https://doi.org/10.1109/MC.2004.175
Simos Gerasimou, Radu Calinescu, Stepan Shevtsov, and Danny Weyns. 2017. UNDERSEA: an exemplar for engineering self-adaptive unmanned
underwater vehicles. In Proceedings of the 12th International Symposium on Software Engineering for Adaptive and Self-Managing Systems. IEEE Press,
83–89.
Carl Hewitt. 1977. Viewing control structures as patterns of passing messages. Artificial intelligence 8, 3 (1977), 323–364.
Carl Hewitt, Peter Bishop, and Richard Steiger. 1973. A universal modular actor formalism for artificial intelligence. In Proceedings of the 3rd international
joint conference on Artificial intelligence. Morgan Kaufmann Publishers Inc., 235–245.
Robert Hirschfeld, Pascal Costanza, and Oscar Nierstrasz. 2008. Context-oriented programming. Journal of Object Technology 7, 3 (2008).
Charles Antony Richard Hoare. 1978. Communicating sequential processes. Commun. ACM 21, 8 (1978), 666–677.
Hugh W Holbrook and David R Cheriton. 1999. IP multicast channels: EXPRESS support for large-scale single-source applications. In ACM SIGCOMM
Computer Communication Review, Vol. 29. ACM, 65–78.
Kazuo Iwama and Shuichi Miyazaki. 2008. A Survey of the Stable Marriage Problem and Its Variants. In Proceedings of the International Conference on
Informatics Education and Research for Knowledge-Circulating Society (Icks 2008) (ICKS ’08). IEEE Computer Society, 131–136. https://doi.org/10.1109/
ICKS.2008.7
Duncan E Jackson and Francis LW Ratnieks. 2006. Communication in ants. Current biology 16, 15 (2006), R570–R574.
Tommy R Jensen and Bjarne Toft. 1995. Graph coloring problems. Vol. 39. John Wiley & Sons.
David Kinny. 2005. Algebraic specification of agent computation. Applicable Algebra in Engineering, Communication and Computing 16, 2 (2005), 77–111.
Annabelle Klarl. 2015. Engineering Self-Adaptive Systems with the Role-Based Architecture of Helena. In Infrastructure for Collaborative Enterprises,
WETICE 2015, Larnaca, Cyprus, June 15-17, 2015. 3–8. https://doi.org/10.1109/WETICE.2015.32
Hermann Kopetz. 2011. Internet of things. In Real-time systems. Springer, 307–323.
Filip Krijt, Zbynek Jiracek, Tomas Bures, Petr Hnetynka, and Ilias Gerostathopoulos. 2017. Intelligent Ensembles: A Declarative Group Description
Language and Java Framework. In Proceedings of the 12th International Symposium on Software Engineering for Adaptive and Self-Managing Systems
(SEAMS ’17). IEEE Press, Piscataway, NJ, USA, 116–122. https://doi.org/10.1109/SEAMS.2017.17
Paulo Leitão, José Barbosa, and Damien Trentesaux. 2012. Bio-inspired multi-agent systems for reconfigurable manufacturing systems. Engineering
Applications of Artificial Intelligence 25, 5 (2012), 934–944.
Wilfried Lepuschitz, Benjamin Groessing, Munir Merdan, and Georg Schitter. 2013. Evaluation of a multi-agent approach for a real transportation system.
In Industrial Technology (ICIT), 2013 IEEE International Conference on. IEEE, 1273–1278.
Michele Loreti and Jane Hillston. 2016. Modelling and Analysis of Collective Adaptive Systems with CARMA and its Tools. In Formal Methods for the
Quantitative Evaluation of Collective Adaptive Systems - 16th International School on Formal Methods for the Design of Computer, Communication, and
Software Systems, SFM 2016, Bertinoro, Italy, June 20-24, 2016, Advanced Lectures (Lecture Notes in Computer Science), Marco Bernardo, Rocco De Nicola,
and Jane Hillston (Eds.), Vol. 9700. Springer, 83–119. https://doi.org/10.1007/978-3-319-34096-8_4
Bruce M. Maggs and Ramesh K. Sitaraman. 2015. Algorithmic Nuggets in Content Delivery. SIGCOMM Comput. Commun. Rev. 45, 3 (July 2015), 52–66.
https://doi.org/10.1145/2805789.2805800
Marco Mamei and Franco Zambonelli. 2004. Programming pervasive and mobile computing applications with the TOTA middleware. In Pervasive
Computing and Communications, 2004. PerCom 2004. Proceedings of the Second IEEE Annual Conference on. IEEE, 263–273.
Robin Milner. 1980. A calculus of communicating systems. (1980).
Robin Milner. 1989. Communication and concurrency. Prentice-Hall, Inc.
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
29
Robin Milner, Joachim Parrow, and David Walker. 1992. A calculus of mobile processes, II. Information and computation 100, 1 (1992), 41–77.
Hong-Seok Park and Ngoc-Hien Tran. 2012. An autonomous manufacturing system based on swarm of cognitive agents. Journal of Manufacturing
Systems 31, 3 (2012), 337 – 348. https://doi.org/10.1016/j.jmsy.2012.05.002
KVS Prasad. 1991. A calculus of broadcasting systems. In TAPSOFT’91. Springer, 338–358.
Kuchi VS Prasad. 1995. A calculus of broadcasting systems. Science of Computer Programming 25, 2 (1995), 285–327.
D. Sanderson, N. Antzoulatos, J. C. Chaplin, D. Busquets, J. Pitt, C. German, A. Norbury, E. Kelly, and S. Ratchev. 2015. Advanced Manufacturing: An
Industrial Application for Collective Adaptive Systems. In 2015 IEEE International Conference on Self-Adaptive and Self-Organizing Systems Workshops.
61–67. https://doi.org/10.1109/SASOW.2015.15
Davide Sangiorgi and David Walker. 2003. The pi-calculus: a Theory of Mobile Processes. Cambridge university press.
A. Sassi, C. Borean, R. Giannantonio, M. Mamei, D. Mana, and F. Zambonelli. 2015. Crowd Steering in Public Spaces: Approaches and Strategies. In 2015
IEEE International Conference on Computer and Information Technology; Ubiquitous Computing and Communications; Dependable, Autonomic and Secure
Computing; Pervasive Intelligence and Computing. 2098–2105. https://doi.org/10.1109/CIT/IUCC/DASC/PICOM.2015.312
Ian Sommerville, Dave Cliff, Radu Calinescu, Justin Keen, Tim Kelly, Marta Kwiatkowska, John Mcdermid, and Richard Paige. 2012. Large-scale complex
IT systems. Commun. ACM 55, 7 (2012), 71–77.
Mirko Viroli, Ferruccio Damiani, and Jacob Beal. 2013. A calculus of computational fields. In Advances in Service-Oriented and Cloud Computing. Springer,
114–128.
Mirko Viroli and Andrea Omicini. 2005. Process-algebraic approaches for multi-agent systems: an overview. Applicable Algebra in Engineering,
Communication and Computing 16, 2-3 (2005), 69–75.
Danny Weyns, Sam Malek, and Jesper Andersson. 2012. FORMS: Unifying Reference Model for Formal Specification of Distributed Self-adaptive Systems.
ACM Trans. Auton. Adapt. Syst. 7, 1, Article 8 (May 2012), 61 pages. https://doi.org/10.1145/2168260.2168268
Martin Wirsing and Matthias Holzl. 2006. Software intensive systems. Draft Report on ERCIM Beyond the Horizon Thematic Group 6 (2006).
Michael J Wooldridge. 2000. Reasoning about rational agents. MIT press.
Manuscript submitted to ACM
30
A
Y. Abd Alrahman et al.
APPENDIX
To provide further evidences of the expressive power of AbC and of its ability to naturally model the interaction in
CAS, below we present two additional case studies from different application domains, namely crowd steering and
swarm robotics.
A.1 A Smart Conference System
This case study is inspired by work on crowd steering in public spaces during big events [Sassi et al. 2015]. The idea
is to exploit the mobile devices of conference participants to guide them to the presentations they are interested in.
Each participant expresses her topic of interest and the conference venue is responsible for guiding her to the location
where the presentations that match her interest take place. The conference venue is composed of a set of rooms where
the conference sessions are planned. We assume that the name of each room identifies its location, e.g., “1st Floor,
Room.101" and that each participant has a unique id. The conference program and session relocation can be dynamically
adjusted at anytime to handle specific situations, i.e., a crowded session can be moved to a larger room and this should
be done seamlessly without any disruption to the whole conference program. When relocation happens, the updates
should be communicated to the interested participants.
The smart conference system can be modelled in AbC as the parallel composition of the conference venue and the
set of available participants. The conference venue is represented as a set of parallel AbC components, each of them
representing a room (Room 1 ∥ . . . ∥Roomn ) and each Roomi has the following structure Γi :I R where Γi represents the
attribute environment of a room, I = {role, session} represents its interface and R represents its behaviour. A Participanti
instead has the following structure Γp :I ′ P where Γp represents its attribute environment, I ′ = {id, interest } represents
its interface, and P represents its behaviour. The overall system is defined as follows:
Room 1 ∥ . . . ∥Roomn ∥ Participant 1 ∥ . . . ∥ Participantm
The attribute environments for participants and rooms contain the following attributes:
id: the identity of a participant;
interest: the current topic of interest of a participant;
name: the name of a room, e.g., “1st Floor, Room.101";
dest: to store the name of the room where the topic of interest for a participant is presented;
role: the role of a component, each room has a Provider role;
session: the current scheduled session for a room;
prevSession: the session that was supposed to be held previously in a room if any;
newSession: the new session assigned to a room;
relocate: a boolean attribute, that indicates if a room has to start the relocation process.
The behaviour of process P, specifying the behaviour of a participant is defined as follows:
let v = (initialTopic) in P ≜ set(interest, v)
(“request”, this.id)@(role = “Provider ”).
(session = this.interest ∧ x = “interestRply”)(x, y).[dest := y]U pd
U pd ≜ (x = “update” ∧ z = this.interest)(x, y, z, l).[dest := l]U pd
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
31
A participant starts by updating the value of her attribute interest with an initial topic. This will allow the participant
to communicate her topic of interest to the conference venue by sending a session-request to nearby providers; in
our case, this is a room. Notice that a participant does not know room names in advance, and only sends messages to
components with a provider role. The communicated values include a label “request” and the identity of the participant.
Once the message is emitted, the process blocks until a reply notification that matches her interest arrives. The
notification contains an interestRply label, and the name of the room where the session is to be held. The participant is
eligible to receive the message only if its attributes satisfy the sending predicate. Moreover, the participant receives
only from a room with a session that matches her topic of interest (i.e., session = this.interest). By receiving this
notification, the process updates its destination and waits for new possible updates through process U pd. This process
is blocked until it receives an update notification about a session that matches the participant interest. The notification
message contains an update label, the session that was supposed to be held previously in this room, the current session
of the room, and the name of the room where the session of interest has been moved. Once a notification message is
received, the process updates the destination to the new location and waits for future updates.
The behaviour of a room component is obtained as the parallel composition of three different processes:
R ≜ Service
|
Relocation
| U pdatinд
The Service process is responsible for providing a normal service for a room. When a room receives a session-request,
the process Service replies back to the requester with the room location whenever the current session of the room
matches the topic of interest of the requester. The Relocation and the U pdatinд processes are responsible for handling
sessions and participants relocation from one room to another.
The process Service, reported below, shows how the room normally handles participant requests. The Service process
is busy-waiting for session-requests from the available participants. Once a message with a possible session-request is
received, and its attached interest value matches the current room session, the process sends an interest-reply message
to the requester and gets ready to handle new session-requests by replicating itself. The interest-reply message contains
an interestRply label and the name of the room. Notice that, since the predicate of the interestRply message addresses
the requester with its identity, there is only one possible receiver for this message which is the requester herself.
Service ≜ (x = “request” ∧ interest = this.session)(x, y).(“interestRply”, name)@(id = y).0 |
Service
The process Relocation handles unexpected changes of the schedule for a room. The idea is to handle these run-time
changes in a way such that interested participants in the new session and also other rooms where a swap of schedule
should happen are notified. The behaviour of Relocation is triggered by changes in the environmental attributes
N ewSession and relocate. In general, environmental attributes are controlled either by the environment itself or by
other components; these may be humans or sensors that intervene to adapt the system behaviour to guarantee the
expected functioning.
A possible scenario is that the session is becoming too crowded and needs to be relocated to another larger room
with possibly few attendees. Another component that plays a portal role and keeps information about the capacity and
the run-time utilisation of all rooms can propose a new session that best fits with the capacity of this room. Note that
the portal component only proposes suitable sessions for rooms based on their capacities and their run-time utilisation,
but has no control on their behaviours. The decision of relocation is made by the room itself depending on the readings
from sensors which set the value of the attribute relocate to true if the level of overcrowding exceeded a certain
Manuscript submitted to ACM
32
Y. Abd Alrahman et al.
threshold. To achieve relocation we have to steer the crowd from one room to another and vice versa. Actually, three
different parties have to be notified about the changes of the schedule. The participants who are interested in the new
proposed session, the room that is currently assigned the new proposed session and should swap its session with the
crowded room, and the participants who are interested in the crowded session.
The Relocation process is responsible for notifying the room where a swap of schedule should happen and the
participants who are interested in the new proposed session. The U pdatinд process communicates the new location of
the crowded session to the interested participants. An awareness operator is needed to allow the Relocation process
to keep track of environmental changes; it is used to collect run-time data from the attribute environment of the
component where the process Relocation is executed. These data are used to decide future behaviours.
Process Relocation, reported below, is blocked until the value of attribute relocate becomes true; it sends a session/interest update to the participants interested in the new assigned session and also to the room where a swap of schedule
should happen as shown in the “∨" predicate of the send action. The interest update message contains an interest
update label, the current session, the new session, and the name of the room. By sending this message, the process
updates the room’s previous session to the current one and the current session to the new session provided by the
portal component through the value of the attribute newSession. The flag relocate is set to false.
Relocation ≜ ⟨relocate = tt⟩(“update”, this.session, this.newSession, name)
@(interest = this.newSession ∨ session = this.newSession).
[prevSession := session,
session := newSession, relocate := ff]Relocation
To relocate the sessions and steer participants correctly both the involved rooms do collaborate and propagate the
changes in their schedules to the interested participants. The global goal is to make both groups of participants, interested
in one of the two room sessions, aware of the new location of their interest. As seen before, process Relocation is partly
responsible for propagating the changes to the group of participants interested in the newly assigned session. The same
message is used to ask for collaboration to the other room, where a swap of schedule should happen. The other room
has to update its session and propagate the changes to all participants interested in the crowded session. This task is
performed by process U pdatinд:
U pdatinд ≜ (x = “update” ∧ z = this.session)(x, y, z, l).
[prevSession := session, session := y]
(“update”, this.prevSession, this.session, this.name)
@(interest = this.session ∨ session = this.session).0 | U pdatinд
When a room receives a session-update message regarding its own session, the attribute prevSession gets the value of
session and this is assigned the value of the session name communicated by the other room. Once the changes have
been applied, the process sends a session/interest-update message and again the process U pdatinд is made available to
handle session-update messages. The session-update message is sent to the other group of participants so that they
relocate to the new destination. It should be noted that the structure of this message is exactly the same as the one sent
by Relocation. The only difference is that the sent values depend on the current attribute values of the room where the
process U pdatinд is executed.
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
33
Example A.1 (Interaction fragment). Let us assume that a participant, say p1 with id = 1 and with “Theory” as her
initial topic of interest, sent a session-request for nearby providers. The participant p1 applies rule Comp twice. The
first application of Comp, reported below, updates the interest attribute of the participant6 .
Γp :I ′ P
{id=1, int er est =⊥}▷ff()
−−−−−−−−−−−−−−−−−−−−−→
Γp [interest 7→ “Theory”] :I ′ P ′
The second application of Comp is used to send the request to nearby providers:
Γp′ :I ′ P ′
{id=1, int er est =“T heor y”}▷(r ol e=“P r ovider ”)(”r equest ”,1)
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→
Γp′ :I ′ P ′′
If a room, say r 1 scheduled for the session “Theory”, receives this message, it applies rule (Comp) and performs the
following transition:
Γ1 :I Service | Relocation | U pdatinд
{id=1, int er est =“T heor y”}▷(r ol e=“P r ovider ”)(”r equest ”,1)
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→
Γ1 :I Service ′ [“request”/x, 1/y] | Service | Relocation | U pdatinд
All other rooms will just discard the request and apply rule FComp and the overall system, by applying rule ComR, will
evolve as follows:
S
{id=1, int er est =“T heor y”}▷(r ol e=“P r ovider ”)(”r equest ”,1)
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→
(Γ1 :I Service ′ [“request”/x, 1/y] | Service | Relocation | U pdatinд) ∥
Γp′ :I ′ P ′′ ∥ Γ2 :I R 2 ∥ . . . ∥ Γn :I Rn
The components Γ2 :I R 2 , . . . , Γn :I Rn represent the other rooms in the conference venue.
A.2
□
A swarm robotics scenario in AbC
We consider a scenario where a swarm of robots spreads throughout a given disaster area with the goal of locating
and rescuing possible victims. Similar case studies have been considered in [Ercan et al. 2006]. All robots execute
the same code that defines their functional behaviour, and rely on a set of adaptation mechanisms that regulates the
interactions with other robots and with the environment. Initially, all robots play the explorer role to search for victims
in the environment. Once a robot finds a victim, its role changes to “rescuer” and it sends information about the victim
to nearby explorers. The collective (the swarm) starts forming in preparation of the rescuing procedure. As soon as
another robot receives information about the victim, it changes its role to “helper” and joins the rescuers-collective.
The rescuing procedure starts only when the collective formation is complete. During exploration, in case of a critical
battery level, a robot enters a power saving mode until it is recharged.
This model exploits the fact that a process running on a robot can either read the values of attributes provided by its
sensors or read and update other attributes in its environment. Reading the values of the attributes controlled by sensors
provides information about the environment or about the current status of the robot. Thus, the model captures both
self-awareness and context-awareness. For instance, when reading the value of the collision attribute, if Γ(collision) = tt,
the robot becomes aware that a collision with a wall in the arena is imminent and this triggers an adaptation mechanism
6 Please
remember that we have set(int er est, I nit ialT opic) ≡ ()@ff.[int er est := init ialT opic]
Manuscript submitted to ACM
34
Y. Abd Alrahman et al.
to change its direction. On the other hand, if Γ(batteryLevel) = 15%, the robot becomes aware that its power is limited
(e.g., < 20%) and starts an adaptation program to take the robot into the power saving mode.
We assume that each robot has a unique identity (id) and since the robot acquires information about its environment
or its own status by reading the values provided by sensors, no additional assumption about its initial state is needed. It
is worth mentioning that sensors and actuators are not modelled by AbC because they represent the robot internal
infrastructure while the AbC model represents the programmable behaviour of the robot (i.e., its running code).
The robotics scenario is modelled as a set of parallel AbC components, each of which represents a robot (Robot 1 ∥ . . . ∥Robotn )
and each robot has the following form (Γi :I P R ) where I = {id, role}.
The attribute environment for a robot contains the following attributes:
id: the unique identity of a robot;
role: the current role of a robot;
victimPerceived: to indicate if a victim is found;
state: its value triggers an actuation signal to either move or stop the movement;
vPostion: to indicate a victim position;
count: to indicate the required number of robots needed to rescue a victim;
target: its value triggers the actuators to move to a specific location;
direction: its value triggers the actuators to follow a specific direction;
collision: to indicate if a collision is detected;
batteryLevel: to indicate the battery level of the robot 7 .
The behaviour of a single robot is modelled by:
PR
≜ (Rescuer
+ Explorer )| RandW alk | IsMovinд
The robot follows a random walk in exploring the disaster arena. The robot can either become a “rescuer” when he
becomes aware of the presence of a victim by locally reading the value of an attribute controlled by its sensors or
remain as “explorer” and keep sending queries for information about the victim from nearby robots whose role is either
“rescuer ” or “helper ”.
The definition of process Rescuer is reported below. Please notice that the actual values of x 1 , x 2 and x 3 are provided
by external sensors.
Rescuer ≜
⟨this.victimPerceived = tt⟩()@ff.[this.state := stop, this.vPosition := < x 1 , x 2 >,
this.count := x 3 , this.role := rescuer ]
(x = “qry” ∧ role = “explorer ”)(x, y).
(“ack”, this.vPosition, this.count)@(id = y).0
If sensors recognise the presence of a victim (“victim-Perceived” becomes “tt”), the robot updates its “state” to “stop”
(which triggers an actuation signal to halt the actuators and stop movement), computes the position of the victim and
the number of robots necessary to rescue the victim and stores them in “vPosition” and “count”, then it changes its role
7 Notice
that the values of attributes victimPerceived, vPosition, count, collision, and batteryLevel are provided by sensors while the values of
attributes state, target, and direction send actuation signals to the actuators of the robot.
Manuscript submitted to ACM
Programming the Interactions of Collective Adaptive Systems
35
to “rescuer ” and waits for queries from nearby explorers. Once a message from an explorer is received, the robot sends
back the information about the victim to the requesting robot addressing it with its identity “id”.
Alternatively, the robot plays the explorer role and continuously sends queries for information about the victim
to nearby robots whose role is either “rescuer” or “helper ”. The query message contains a label “qry” to indicate the
request type and the identity of the robot “this.id”. When it receives an acknowledgement with victim’s information,
the robot changes its role to “helper ” and starts the helping procedure. Notice that if an acknowledgment message is
not received, the robot can again behave as rescuer or explorer as described above.
Explorer ≜
(“qry”, this.id)@(role = “rescuer ” ∨ role = “helper ”).
(((role = “rescuer ” ∨ role = “helper ”) ∧ x = “ack”)(x, vpos, c).[this.role := helper ]Helper
+ Rescuer
+ Explorer )
The “Helper ” process is reported below:
Helper ≜ ()@ff.[this.vPosition := vpos, this.tarдet := vpos]
( ⟨this.position = this.tarдet⟩set(this.role, rescuer )0
|
⟨c > 1⟩(x = “qry” ∧ role = “explorer ”)(x, y).
(“ack”, this.vPosition, c − 1)@(id = y).0 )
The helping robot stores the victim position in the attribute “vPosition” and updates its target to be the victim
position to tell the actuators to move to the specified location. While moving towards a victim, the robot is ready to
respond to other robots queries, in case more than one robot (c > 1) is needed for the rescuing procedure. Once it
reaches the victim (i.e., its position coincides with the victim position), the robot changes its role to “rescuer ” and joins
the rescuer-collective.
The “RandW alk” process is defined below, it determines the random direction to be followed by the robot. When a
collision is detected by a proximity sensor, a new random direction is computed.
RandW alk ≜ set(this.direction, 2πrand())⟨this.collision = tt⟩RandW alk
Finally, process “IsMovinд” captures the status of the battery level in a robot at any time. Once the battery level
drops into a critical level (≤ 20%), the robot changes its status to “stop” which results in halting the actuators and the
robot enters the power saving mode. The robot stays in this mode until it is recharged (≥ 90%), then it starts moving
again.
IsMovinд ≜ ⟨this.state = move ∧ (this.batteryLevel ≤ 20%)⟩
set(this.state, stop)⟨this.batteryLevel ≥ 90%⟩
set(this.state, move)IsMovinд
For simplifying the presentation, in this scenario we are not modelling the charging task and assume that this task is
accomplished according to some predefined procedure.
Manuscript submitted to ACM
36
Y. Abd Alrahman et al.
Example A.2 (Interaction fragment). Let us assume that the role of Robot 1 is “rescuer ” and Robot 2 is “explorer ”. Robot 2
can send a query to nearby rescuing or helping robots (i.e., Robot 1 ) by using rule Comp and generate this transition:
Robot 2
{id=2,r ol e=“explor er ”}▷(r ol e=“r escuer ”∨r ol e=“helpinд”)(“qr y”, 2)
−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→ Γ2 :I (Explorer ′ | P 3 )
where P3 ≡ RandW alk | IsMovinд
On the other hand, Robot 1 can receive this query by using rule Comp and generate this transition:
{id=2,r ol e=“explor er ”}▷(r ol e=“r escuer ”∨r ol e=“helpinд”)(“qr y”, 2)
Robot 1 −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→ Γ1′ :I (Rescuer ′ [“qry”/x, 2/y]|P 3 )
where P 3 ≡ RandW alk | IsMovinд
Other robots which are not addressed by communication discard the message by applying rule FComp. Now the overall
system evolves by applying rule ComR as follows:
{id=2,r ol e=“explor er ”}▷(r ol e=“r escuer ”∨r ol e=“helpinд”)(“qr y”, 2)
S −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−→
Γ1′ :I (Rescuer ′ [“qry”/x, 2/y]|P3 )∥ Γ2 :I (Explorer ′ | P3 )∥ Γ3 :I P R 3 ∥ . . . ∥ Γn :I P Rn
Manuscript submitted to ACM
□
| 6cs.PL
|
Hard Capacitated Set Cover and Uncapacitated Geometric Set Cover
Rahil Sharma
Department of Computer Science, The University of Iowa, Iowa City, IA 52242
arXiv:1612.03097v1 [cs.DS] 30 Nov 2016
[email protected]
1
Introduction
In the optimization problem, minimum set cover, we are given a set system (X, S), where X is a ground
set of n elements and S is a set of subsets of X. Each set S ∈ S has a cost w(S) > 0 associated with it.
Let R be a subset of S. We say that R is a set cover of X, if each element in X belongs to at least one
set in R. The objective is to find a set cover R that minimize the sum of the cost of all sets in R. The
minimum set cover problem is a classic NP-hard problem [10]. In the example below, X = {1, 2, 3, 4, 5, 6}
and S = {S1 , S2 , S3 , S4 }, where S1 = {1, 2}, S2 = {2, 3, 4}, S3 = {4, 5, 6}, S4 = {5, 6}. The costs of the
four sets are 1, 2, 5, 3 respectively.
w(S1) = 1
S1
w(S2) = 2
w(S3) = 5 w(S4) = 3
S4
S2
S3
1
2
3
4
X
5
6
Figure 1: R = {S2 , S4 } is not a set cover of X, since element 1 ∈ X is not contained in any of the sets
in R. There are 2 possible set covers for this example, R = {S1 , S2 , S3 } and R = {S1 , S2 , S4 } of cost 8
and 6 respectively. Hence R = {S1 , S2 , S4 } is the minimum set cover here.
We are interested in the capacitated version of minimum set cover problem, in which each set S ∈ S has
a capacity k(S) associated with it, such that for each set S ∈ S, at most k(S) elements can be contained
in S. The capacitated covering problems are of 2 types. First, sets with soft capacities where each set has
unbounded number of copies that can be used to cover and the second, sets with hard capacities where
each set has a bound on the number of available copies.
Formally, in a capacitated set cover problem with hard capacities we are given a ground set
of elements X and a collection of its subset S. Each S ∈ S has a positive integral capacity k(S) and
a non-negative cost w(S) associated with it. Let R be a set of subsets of S. Let f : X → R be an
assignment of elements in X to a set in R such that for any x ∈ X, if (a) f (x) = S then x ∈ S (b)
|{x | f (x) = S}| ≤ k(S) for all S ∈ R. We call R a valid set cover if such f exists. Cost of the
Psolution R
is the sum of the cost of all sets in R. The goal is to find a set cover R of minimum cost i.e. S∈R w(S).
The best known result for this problem was given by Wolsey’s algorithm [9]. It gave a set cover of cost
O(log n) of the optimal solution. In the first of the two parts of this report, we show the same result using
Wolsey’s algorithm, but with a different and simpler analysis given in [5]. We also make a key observation
in the analysis given in [5] which allows us to apply the weighted set cover greedy algorithm’s analysis [6]
to hard capacitated set cover problem.
One of the motivating applications is [4]; i.e. GMID (Glyco-molecule ID) which is a chip-based
technology that is used to generate fingerprints which uniquely identify glycomolecules. Each run of the
experiment answers the question: For a given building block A, and for each member B in a set S of
building blocks, does the solution contain a molecule which contains both building blocks A and B? The
size of the set S is restricted, because of the specific technology. Here the information is presented as a
graph where the building blocks are its vertices, and an edge exists between two vertices if the question
regarding their connectivity is required. The device is able to answer |S| = k questions at once if they
share a common vertex. The problem of minimizing the number of GMID experiments needed to cover
the required information graph, is precisely a capacitated vertex cover with hard capacities which is a
special case of the set cover, with each element belongs to two sets [7]. It is easy to see the generalization
of this special case to our problem where the information is provided in a multi-graph.
Another interesting application this problem has is in the area of cellular network coverage. We are
given n users and a set S = {S1 , S2 , ..., Sm } of m antennas. Each antenna has a capacity k(Si ) and a
positive cost of installation and maintenance c(Si ), for all i where, 1 ≤ i ≤ m; associated with it. The
capacity determines the upper-bound on the number of users it can serve at a time, no matter how many
users it covers. This is a real-life bandwidth issue, that needs to be resolved while setting up the network.
The problem here is to find a minimum cost subset of antennas from S, that serves all the n-users,
without violating the capacity constraint. Provided that a feasible solution for the problem exists. Here
the coverage area of the antenna can be viewed as geometric objects like a disk centered at the antenna
or a sector originating from an antenna. This drives our motivation behind exploring geometric set cover
problems where the covering objects/sets are geometric bodies.
base station
(a)
(b)
Figure 2: (a) Antenna located at the center of the circle, which represents the coverage area (b) A single
base station with 3 directional antennas on it and the coverage area resembles a sector
In the geometric set cover problem, we are given a ground set X of n points in <2 and a set of
objects S whose union contains all the points in X. The objects are geometric bodies like fat triangles,
discs, rectangles, etc. The goal is to find the minimum cardinality subset R ⊆ S such that every point in
X is contained in at least one object in R. It is important to note that, there is another class of problems
called as hitting set problem. In the related hitting set problem; the goal will be to select a minimum
cardinality subfamily X 0 ⊆ X, such that every object in S contains at least one point in X 0 .
We can reduce the above two problems to the general set cover problem. The geometric set cover
problem is reduced to the general set cover problem (see Figure 3(a)), by viewing the points in the plane
as elements to be covered and objects as sets to cover them. The hitting set problem is reduced to general
set cover problem (see Figure 3(b)), by viewing each point in the plane as a set and each object as an
element to be covered. The geometric set cover problem is the dual problem for the geometric hitting set
problem. This is an unweighted version of the general set cover problem, where each object in S has no
cost/weight associated with it or they have uniform weight.
(a)
(b)
Figure 3: (a) Geometric set cover instance to general set cover (b) Hitting set instance to general set cover
Let us define -net since it is the central theme used for improving bounds on geometric set cover
problems. Let X be a set of n points in the plane and S be the set of axis parallel rectangles whose
union covers X. For ∈ (0, 1), the -net of X is a set Y of points in a plane, such that any rectangle
S ∈ S that contains at least |X| points from X, contains a point from Y . One can even think of -nets
as a cover for heavily covered points. The connection between the size of -nets and size of the hitting
set is shown by Bronnimann and Goodrich [3], i.e. if for any point set -net of size O(1/ h(1/)) is
computed in polynomial time, then the corresponding geometric hitting set problem has a polynomial
time O(h(OP T )) approximation, where OP T is the size of the optimal cover. Haussler and Welzl [8]
showed the existence of such nets of size O(1/ log (1/)) for general geometric objects. Thus using
the Bronnimann and Goodrich result, for general geometric objects we can achieve a hitting set of size
O(log OP T ) of the optimal solution.
In this report, the second problem that we will be discussing is a geometric hitting set problem where,
X is a ground set of points in the plane and S is a set of axis parallel rectangles [2]. Here we shall show
the existence of -nets of size O(1/ log log 1/). Then applying Bronnimann and Goodrich result [3] we
get the hitting set of size O(log log OP T ) of the optimal solution. This can be extended to axis-parallel
boxes as ranges in 3-dimension, leading to the same result as in 2-dimension. In this report we shall just
focus on the 2 dimensional result.
2
Hard Capacitated Set Cover
Recalling the problem statement; we are given a set system (X, S), where X = {x1 , x2 , ..., xn } is the
ground set of n elements and S is the set of subsets of X. Each set S ∈ S has a cost w(S) > 0 and
a positive integral capacity k(S) > 0 associated with it. Let R be a subset of S. Let f : X → R be
an assignment of elements in X to a set in R such that for any x ∈ X, if (a) f (x) = S then x ∈ S
(b) |{x | f (x) = S}| ≤ k(S) for all S ∈ R. We call R a valid set cover if such f exists. Cost of the
solution R is the sum of the cost of all sets in R. The goal is to find a set cover R of minimum cost. Give
an instance (X, S) of this problem, we first need to check whether there exists a valid set cover. So in
this section, firstly we will set up a directed flow network for our instance of hard capacitated set cover
problem. Using maximum-flow, we will check in polynomial time, whether there exists a valid set cover
for the given problem instance. Secondly, we will describe Wolsey’s algorithm and finally, we will proceed
with our analysis to give a set cover of cost O(log n) of the optimal solution.
2.1
Setting up directed flow network
Flow networks have played an important role in design of algorithms and in other areas of computer
science like network design, computer vision, etc. Many of the covering problems are visualized as flow
network problems [4], since there is a rich literature of network flow problems available. In capacitated
set cover problem with hard capacities there is a ground set X = {1, .., n} of elements and a collection of
its subset S. Each S ∈ S has a capacity k(S) and a non-negative cost w(S) associated with it. Let P be
a subset of sets of S. Denote C ⊆ P ×X as a partial cover iff for each (S, e) ∈ C, e ∈ S. We say e ∈ X
is covered by S in C if (S, e) ∈ C. Let us give an example, P = {a, b, c}, where a = {1, 3, 4}, b = {2, 5, 6},
c = {3, 8, 9, 5} and X = {1, 2, 4, 8}. If C = {(a, 1), (a, 8), (c, 8)} then it is not a valid partial cover, since
(a, 8) ∈ C but 8 ∈
/ a. Where as C = {(a, 1), (a, 4), (c, 8)} is a valid partial cover.
Assume without loss of generality that each element e ∈ X is covered at most once in C, as we can
view covering as assigning element e ∈ X to set (S, e) ∈ C. Partial cover C is feasible if all sets maintain
their capacity constraints. Let value of C be the number of elements it covers, denoted by |C|. Let f (P)
denote the maximal value of the partial cover C over all the feasible partial covers C. We show that a
feasible partial cover C ⊆ P ×X of value f (P), can be computed in polynomial time.
Lemma 2.1. Given an instance of hard capacitated set cover problem and a subset P of sets of S, a cover
C of value f (P ) can be computed in polynomial time i.e. we can establish in polynomial time, whether P
is a valid set cover.
Proof. Let us construct a directed network/graph G = (L, R, E) of P and X. Each set Si ∈ P, for all
1 ≤ i ≤ |S|, represents a vertex in L and each element ej ∈ X, for all 1 ≤ j ≤ n represents a vertex in
R. There is an edge (Si , ej ) ∈ E of capacity 1 iff ej ∈ Si . Add a source vertex S and an edge (S, Si ) of
capacity k(Si ), for all i. Add a target vertex T and an edge (ej , T ) of capacity 1, for all j (see Figure 4).
Here an edge (Si , ej ) implies ej ∈ Si . So a flow in the network is a feasible partial cover. Since we are
using all the edges between Si and ej , the value of maximum flow is at least f (P). In fact, value of the
maximum flow in the network is exactly equal to f (P), since the value of the flow is integral. So P is a
feasible solution to the set cover problem iff f (P) = |X|. Since maximum flow value can be computed in
polynomial time, so can f (P). This completes our proof.
Figure 4: Directed network G = (L, R, E) where; L: Si ∈ P, R: ej ∈ X and E: set of all edges
2.2
Wolsey’s Algorithm and Analysis
Let R ⊆ S be a family of sets. The value of f (R) can be computed in polynomial time from Lemma 1.
Let fR (S) = f (R ∪{S}) − f (R), i.e. the increase in the number of elements covered when set S is added
to subset R.
It is important to note that Wolsey’s greedy algorithm is quite similar to the weighted set cover greedy
algorithm. In greedy algorithm for weighted uncapacitated set cover (see Algorithm 2), we pick the set
Algorithm 1 : Wolsey’s greedy algorithm
Require: Feasible capacitated set cover denoted by P
Initially, P = φ.
while P is not a feasible capacitated set cover do
Let S = arg minS: fP (S)>0 (w(S)/fP (S))
Add S to P
end while
return P
in the cover C to be the one that makes most progress i.e. cover the most uncovered elements per unit
weight. If X is the set of elements that are not covered yet, we add set Si to the cover C, if it minimizes
the quantity wi /(Si ∩ X). In this paper [5], we make a key observation that allows us to use the same
analysis as that of the weighted set cover greedy algorithm [6], for hard capacitated set cover. In the
Section 2.3 we have showed this alternative analysis.
Algorithm 2 : Greedy algorithm for weighted set cover
Require: minimum set cover C
Initially, let ground set of elements be X = n elements, weight wi associated with each Si ∈ S and
cover C = φ; X → Y
while Y 6= φ do
wi
Let Si be the set that minimizes
|Si ∩ Y |
C = C ∪ {Si }
Y = Y \ Si
end while
return C
Let C ⊆ R ×X be a feasible partial cover. For R0 ⊆ R, let fC (R0 ) denote the number of elements
covered by sets of R0 in cover C.
Lemma 2.2. Let R be a valid set cover for an instance of capacitated set cover problem with hard capacities
and R1 R2 be its partition into 2 disjoint subsets. There exists a feasible cover C ⊆ R ×X such that:
1. all elements of X are covered in C.
2. fC (R1 ) = f (R1 )
Proof. For a feasible cover C ⊆ R × X, where each element e ∈ X is covered by some set S ∈ R; let
us assume that fC (R1 ) < f (R1 ). Now, consider any feasible partial cover C 0 ⊆ R1 × X satisfying the
condition fC 0 (R1 ) = f (R1 ). Since all the elements are not covered by cover C 0 , we shall use a subroutine
which will gradually change the cover C by replacing some of its set assignments, with the assignment
in C 0 to achieve the desired cover C, while maintaining feasibility. When the subroutine is executed, the
Algorithm 3 : Subroutine
Require: Cover C such that fC (R1 ) = f (R1 ).
while fC (R1 ) < f (R1 ) do
There exists at least one set S ∈ R1 that covers more elements in C 0 than in C.
If e ∈ X is one of that additional element covered by S in C 0 but by some other set S 0 in C;
remove (S 0 , e) and add (S, e) to C.
end while
return C
pair (S, e) is added to C, only if S covers more elements in C 0 than in C, so the capacity constraint is
maintained. Hence this procedure is carried out without violating the feasibility of cover C. It is easy to
see that, once an assignment is made in the subroutine, it remains there till the subroutine has completed
execution. The maximum number of iterations taken by the subroutine before terminating, is bounded
above by |C 0 | and after termination we get the desired cover C.
7
8
1
2
5
4
6
3
5 elements assigned to S in C
e
(S ′, e) ∈ C
7
6
5
1
2
4
8
3
e
6 elements assigned to S in C’
(S, e) ∈ C ′
Figure 5: S covers more elements (e ∈ E) in C 0 than in C; remove assignment (S 0 , e) and add (S, e) to C
Analysis of Wolsey’s Greedy Algorithm: Let P = {S1 , S2 , ..., Sm } be the solution returned by the
algorithm in the above order. The solution returned by the algorithm in the ith iteration, where 0 ≤ i ≤ m,
is Pi = {S1 , S2 , ..., Si }. Now we re-run the algorithm. Initially, P0 = φ i.e. no set is added to the solution
by the greedy algorithm. Now we apply Lemma 2 in this iteration, which computes a feasible cover
C ⊆ (P0 ∪ OP T ) × X such that each element in X is assigned to exactly one set in OPT. For each set
S ∈ OP T , we define a0 (S) as the number of elements covered by S in OP T . In the ith iteration Si is
added to the solution by the greedy algorithm i.e. Pi = {S1 , S2 , .., Si }. Now we apply Lemma 2 in this
iteration, which computes a feasible cover C ⊆ (Pi ∪OP T )×X. For each S ∈ OP T \Pi , we define ai (S) to
be the number of elements assigned to a set S in C where, each set S has capacity of at most the number
of elements assigned to that set in C ⊆ (Pi−1 ∪ OP T ) × X i.e. ai−1 (S). For the rest of the analysis the
following statement remains constant and we show this inductively: all the elements can be covered by the
sets in OP T ∪ Pi , even when the sets S ∈ OP T \ Pi has a capacity constraint, of at most ai (S).
Here base case holds for P0 and a0 . Assume that, all the elements can be covered by the sets in
OP T ∪ Pi−1 , even when the sets S ∈ OP T \ Pi−1 has a capacity constraint of at most ai−1 (S). Since
OP T \ Pi ⊆ OP T \ Pi−1 , it is safe to say that OP T ∪ Pi is a feasible cover, even when the sets S ∈
OP T \ Pi has a capacity constraint of at most ai−1 (S). From Lemma 2, there exists a feasible cover
C ⊆ (OP T ∪ Pi ) × X, where fC (Pi ) = f (Pi ) and each S ∈ OP T \ Pi covers at most ai−1 (S) elements.
Since a set S covers same or more number of elements in OP T \Pi−1 than in OP T \Pi , so it can cover same
or more number of elements in C when S ∈ OP T \ Pi−1 as to when S ∈ OP T \ Pi . Hence ai−1 (S) ≥ ai (S).
Charging: Let Si be the set added to the solution by the greedy algorithm in the ith iteration. If
the added set Si ∈ OP T , then we don’t charge any set in OP T , since OP T pays for the same as well.
If Si ∈
/ OP T , let the maximum possible increase in the number of elements that can be covered in C,
when Si is added to P after the (i − 1)th iteration be denoted by fPi−1 (Si ) = ni . Summing over all
S ∈ OP T \ Pi , the difference between maximum capacity of a set S i.e. ai−1 (S) and the number of
elements actually covered by S in C i.e. ai (S), gives the maximum possible increase in C on adding Si i.e.
ni . Mathematically;
X
[ai−1 (S) − ai (S)] = ni
(1)
S∈OP T \Pi
We charge a cost of (w(Si )/ni ) × [ai−1 (S) − ai (S)] to each S ∈ OP T \ Pi . Using equation 1, we can see
that in ith iteration the total cost charged to Si ∈ OP T is precisely w(Si ).
Bounding the cost charged to each set in OP T : If the set S ∈ OP T is added to the solution P by
the greedy algorithm, let (j + 1) be the iteration at which this addition occurs. If the set S ∈ OP T
is not included in P, let (j + 1) be the first iteration at which aj (S) = 0. For each of the iterations
i < j, the maximum possible increase in the number of elements covered in C at the beginning of the ith
iteration is more than the maximum capacity of S i.e. fPi−1 (S) ≥ ai−1 (S), otherwise i would have been
the last iteration and not j. Since S ∈
/ P, i.e. not selected by the greedy algorithm in this iteration, hence
w(Si )
w(S)
w(S)
≤
≤
. Here the first inequality follows from the condition in the greedy algorithm
ni
fPi−1 (S)
ai−1 (S)
and the second follows from above. Total cost charged to S is
j
X
i=1
∀ i, 1 ≤ i ≤ j; add
j
X ai−1 (S) − ai (S)
w(Si )
[ai−1 (S) − ai (S)]
≤ w(S)
ni
ai−1 (S)
(2)
i=1
1
1
1
for [ai−1 (S) − ai (S)] times and l ≤ ai−1 (S) implies ≥
ai−1 (S)
l
ai−1 (S)
ai−1 (S)
X
l=ai (S)+1
1
≤
ai−1 (S)
ai−1 (S)
X
l=ai (S)+1
1
.
l
(3)
Substituting (3) in (2) we see that, the value charged to S is bounded above by w(S)H(S). Hence, total
cost of the solution is [OP T + OP T.ln(maxs |S|)].
Theorem 2.3. There exists a greedy algorithm for set cover problem with hard capacities, that gives a
solution of cost O(log n) of that of the optimal solution.
2.3
Alternative Analysis for Wolsey’s Greedy Algorithm
In this paper [5], we make a key observation in their analysis of Wolsey’s algorithm for hard capacitated
set cover problem, that in each iteration of the algorithm there exist an ordering in which the elements
of the ground set X are covered. This observation allows us to apply the same analysis as the weighted
set cover greedy algorithm [6] to hard capacitated set cover. The observation is stated in the following
lemma.
Lemma 2.4. Consider an instance of the set cover problem with hard capacities and let P be a valid set
cover and C ⊆ R × X be a feasible cover for this instance. There exists an ordering {x1 , x2 , ..., xn } in
which the greedy algorithm covers elements in the ground set X, such that for all i > 0,
(a) the union of the sets in Pi covers fC (Pi ) = f (Pi ) elements and
(b) they are indexed based on the order they are covered i.e. {x1 , x2 , .., xfPi }.
Proof. Let P = {S1 , S2 , ..., Sm } be the solution returned by the algorithm in the above order. The solution
returned by the algorithm in the ith iteration, where 0 ≤ i ≤ m, is Pi = {S1 , S2 , ..., Si }. Let C ⊆ P × X
be a feasible cover where each element is covered by some set S ∈ P. Without loss of generality assume
that no element is covered by more than one set in P. Now we shall prove the lemma using an inductive
argument. In the first iteration of the greedy algorithm when the set S1 is added to the solution, denoted
by P1 , it covers fC (P1 ) = f (P1 ) elements and they are indexed based on the order they are covered i.e.
{x1 , x2 , .., xfP1 }. Assume that in the ith iteration of the greedy algorithm, when the set Si is added to the
solution denoted by Pi , it covers fC (Pi ) = f (Pi ) elements and they are indexed based on the order they are
, xfP
, ..., xfPi }. Now let us prove the induction hypothesis
covered i.e. {x1 , x2 , ..., xfPi−1 , xfP
+1
+2
(i−1)
(i−1)
that, in the (i + 1)th iteration of the greedy algorithm, when the set Si+1 is added to the solution, denoted
by Pi+1 , it covers fC (Pi+1 ) = f (Pi+1 ) elements and they are indexed based on the order they are covered
i.e. {x1 , x2 , .., xfPi , .., xfPi+1 } elements. For proving this, assume that fC (Pi+1 ) < f (Pi+1 ). Let C 0 ⊆ P ×X
be a feasible partial cover, such that fC 0 (Pi+1 ) = f (Pi+1 ). Some elements may not be covered by C 0 , since
it is a partial cover. While maintaining the feasibility, we will gradually change the cover C using the
subroutine given below, until the lemma is satisfied.
In the above sub-routine we can see that, for all i > 0, any element that is already covered by assigning
it to a set in the ith iteration, remains covered in the (i + 1)th iteration as well. Only its assignment to a
particular set may change. The union of the sets in Pi+1 cover at least the elements covered by union of
the sets in Pi i.e. f (Pi ) elements. Any element e ∈ X that is covered by the union of the sets in Pi+1 is (a)
Algorithm 4 : Subroutine
Require: Cover C such that fC (Pi+1 ) = f (Pi+1 ).
while fC (Pi+1 ) < f (Pi+1 ) do
There exists at least one set S ∈ Pi+1 that covers more elements in C 0 than in C.
If e ∈ X is one of that additional element covered by S in C 0 but by some other set S 0 in C;
remove (S 0 , e) and add (S, e) to C.
If e ∈ X is not assigned to any set in C but assigned to some set S 00 ∈ C 0 ; add (S 00 , e)
end while
return C
either covered in the previous iteration i.e. by union of the sets in Pi , in which case, from the inductive step
we can say that, it has already been indexed within {x1 , x2 , .., xfPi } (b) or the element e is newly covered
in (i + 1)th iteration and it is indexed based on the order it is covered i.e. {xf(Pi ) +1 , xf(Pi ) +2 , ..., xf(Pi+1 ) }.
This completes the inductive argument. Once an assignment is made in the subroutine, it remains there
till the subroutine has completed execution. The maximum number of iterations taken by the subroutine
before terminating, is bounded above by |C 0 | and after termination we get Pi , for all i > 0 that satisfies
the lemma.
Alternative Analysis of Wolsey’s Greedy Algorithm: Let every element in X be covered by exactly
one set in optimal set cover denoted by OPT. When the greedy algorithm chooses a set Si in the ith
iteration, let it charge the price per element for that iteration, to each of the newly covered elements
i.e.f (Pi ) − f (Pi−1 ). Total weight of the sets chosen by the algorithm equals the total amount charged.
Each element is charged once. Consider a set S 0 ∈ OP T having k < n elements from X, assigned to it
in the optimal cover. Let the greedy algorithm cover the elements of S 0 in the order: yk , yk−1 , .., y1 . At
the beginning of an iteration in which the algorithm covers element yi of S 0 , at least i − 1 elements of S 0
remains uncovered. If the greedy algorithm chooses the set S 0 to be added to the solution in this iteration,
it would pay a cost per element of at most wS 0 /i − 1. Thus in this iteration greedy algorithm pays at most
wS 0 /i − 1 per element covered. Hence element yi pays at most wS 0 /i − 1 to be covered.
P Now, summing
over all the elements in S 0 , the total amount charged to the elements in S 0 is at most ki=1 wS 0 /i − 1 i.e.
w(S 0 )Hk−1 . Now we sum over all the sets S 0 ∈ OP T and noting the fact P
that every element in X is in
some set in OPT, the total amount charged to elements overall is at most S 0 ∈OP T wS0 Hk−1 = Hn .OP T
Theorem 2.5. For hard capacitated set cover problem, the Wolsey’s greedy algorithm returns a set cover
of cost at most Hn times of that of the optimal set cover.
3
Geometric Set Cover: Axis-Parallel Rectangles
In this section, we explore a geometric hitting set problem where, X is a ground set of points in a plane
and S is a set of axis-parallel rectangles and the goal is to select a minimum cardinality subfamily X 0 ⊆ X,
such that every axis-parallel rectangle in S contains one point from X 0 . We will firstly show the existence
of small size -nets of size O(1/ log log 1/) for axis-parallel rectangles. We shall show this in two sub
parts (1) Constructing -nets (2) Estimating the size of the -nets. Then we will use the Bronnimann
and Goodrich result to show that, if the existence of the above sized net can be shown in polynomial time,
then solution of size O(log log OP T ) of the optimal solution, for the corresponding hitting set problem
can be computed in polynomial time. We conclude this section by giving the lower bound example and a
key observation, that motivated our technique.
3.1
Constructing -nets
Let us start by constructing a balanced binary tree (BBT) T , over set P of n points in the plane. We
build the tree based on the x-order of these n points and we terminate the construction when every leaf
node of the tree reaches to a size between the range [n/r, n/2r], where r = 2/. By size here we mean
the number of points under each leaf node. Since its a BBT, the number of nodes at some level i will
be n/2i and due to the termination condition we can say that, the maximum level imax of the tree is
n/2r = n/2imax i.e. (1 + log r) levels.
n/2
n/4
(1 + log r)
levels
n/8
n/2i =[n/r, n/2r]
Figure 6: Balanced binary tree, decomposing the point set P
We randomly select a set R ⊆ P , which will be included in the -net we construct, such that each
point in P that is included in R is selected independently with probability π = s/n, thus E[|R|] = s. Set
s = cr log log r. In figure (7a), for each node v, let Pv be the subset of the points in the sub-tree rooted
at v. Let lv be the vertical line that divides Pv into two subsets Pv1 and Pv2 , which contains points in
the sub-tree rooted at v1 and v2 respectively. Let the two half planes corresponding to Pv1 and Pv2 be
denoted as σv1 and σv2 . Let the plane which contains all the points in the sub-tree rooted at the root
node v, be denoted as σroot . Our sampling of set R is unbiased, since Pv is set before selecting R.
In figure (7b), σv is the right (resp. left) portion of σu bounded by lu on the left hand side (resp.
right). Hence we define lu as the left entry side (resp. right entry side) of σv . Let Q be an axis-parallel
rectangle that contains at least n points of P . Let u be the highest node in the T such that, the vertical
line assigned to it cuts Q into 2 parts. Let Q be one of the 2 parts, which has at least n/2 = n/r points
from P . There has to exist one such partition. Let v be the child of u that contains Q ⊂ σv , hence Q is
said to be anchored at left entry side lu of σv . For the rest of the section, without loss of generality we
will assume that the entry side lu of σv is the left side.
lu
v
v1
Q
v2
Q
Pv 2
lv
σ v1
σ v2
left entry side of σv
σroot = σu
σu
(a)
σv
(b)
Figure 7: (a) lv splits Pv into 2 subsets Pv1 and Pv2 in half plane σv1 and σv2 respectively (b) Half rectangle
Q anchored at left entry side lu of strip σv
We want to stab all the heavy rectangles i.e. we want to construct a subset from P , that intersects
every rectangle Q such that |Q| ≥ n. Since Q ⊂ Q, if any point from our first random sample R ⊆ P is
contained in Q we have stabbed Q successfully and we are done. Let us assume that Q does not contain a
point from R, and hence we will say Q is R-empty or Rv -empty. We define a set Mv for each node v ∈ T ,
that contains all the maximal anchored Rv -empty axis-parallel rectangles in σv . We shall then bound the
size of Mv . Note that for any v ∈ T , Rv is the set of points from our random subset R that is contained
in Mv .
lu
lu
M
(a)
lu
lu
lu
M
M
(b)
(c)
M
M
(d)
(e)
Figure 8: (a) Rectangle M ∈ Mv is defined by 3 points from Rv , one point on each of the 3 unanchored
sides, Mv also contain degenerate rectangles with some or all of these 3 points missing, (b) and (c) M
with 2 points on the two unanchored sides each (its easy to see the missing case similar to case (b)), (d)
M with just one point on the unanchored side (either top-most or bottom-most), (e) M with no points
from Rv , in this case you just have the entire strip as one maximal Rv -empty rectangle.
Lemma 3.1. For each node v ∈ T , the number of rectangles in the set Mv is equal to (2rv + 1), where
rv = |Rv |.
Proof. We shall cover all of the following, possible cases to show the required bound (see Figure 9). In
case (a), when q ∈ Rv lies on the right unanchored side of any rectangle M ∈ Mv then, no other rectangle
in Mv can have q on its right unanchored side. Let us argue this by contradiction. Assume that some
other rectangle M 0 ∈ Mv also has q on its right unanchored side, this will lead to M 0 ⊂ M , hence it is not
a maximal rectangle which contradicts the definition of Mv else, M 0 contains a point that belongs to Rv
which contradicts our assumption that Q does not contain any point from R. Hence maximum number
of such rectangles is rv .
In case (b), the points q1 , q2 ∈ Rv are on the top and bottom unanchored sides of rectangle M , then
these points are consecutive in Rv in y-coordinate, if they are not then M will to contain the points which
lies between these two points in Rv based on y-order, which is a contradiction. There are rv − 1 such pairs.
Finally, we are left with case (c), where there are 2 semi-unbounded rectangles, one bounded from above
and one from below by their highest and lowest points respectively. Thus from the above 3 cases we get
our required bound of |Mv | = (2rv + 1). It is also easy to see that, the bound holds when rv = 0 where,
M is the entire strip and rv = 1 where, we get M as the entire plane and 2 semi-unbounded rectangles.
We are left to show that, Q belongs to at least one rectangle in Mv . Since we have assumed that the
entry side of σv is the left side, we expand Q by pushing its right unanchored side till it touches a point
of Rv or reaches the boundary of σv . Then we extend it vertically in both the directions till it reaches a
point in Rv or we let it extend till ±∞. Thus, the rectangle we get after expansion contains Q and also
belongs to Mv . This completes the proof.
lu
lu
lu
lu
q1
M′
M
M
q
q
M
M
q
lu
q2
(a)
(b)
M
q
(c)
Figure 9: Contradicting cases for (a) q ∈ Rv lies on the right unanchored side of exactly one M ∈ Mv
(b) q1 , q2 ∈ Rv are on the top and bottom unanchored sides of M ∈ Mv , then these points are consecutive
in Rv in y-coordinate, and (c) 2 semi-unbounded rectangles M
For all nodes v ∈ T , overall number of rectangles M ∈ Mv at a fixed level is O(|R| + r0 ), where
r0 is the total number of nodes at a fixed level and over all the levels of T is O(|R|log r + r0 log r) i.e.
O(|R|log r + r). For each v ∈ T and each M ∈ Mv , let tM be the weight factor of M and it is defined as
s|M ∩ P |/n. The rectangles M with tM < s/r = c log log r i.e. s|M ∩ P |/n < s/r i.e. |M ∩ P | < n/r
and these rectangles can be ignored since they have no anchored rectangles Q contained in them. This is
because |Q| ≥ n/2 = n/r points of P. We will just consider rectangles with tM ≥ c log log r.
Using Haussler and Welzl result [8], for each M ∈ Mv having tM ≥ c log log r, there exists NM ⊆
M ∩ Pv of size k tM log tM , that forms (1/tM ) -net for M ∩ Pv , where k is a constant. We are done with
the construction phase of -nets. So the final -net is union of set R that we picked in the first level
of sampling, with the sets NM , ∀ v ∈ T and ∀ M ∈ Mv with tM ≥ c log log r, in the second level of
sampling.
P
P
Lemma 3.2. The constructed net N = R ∪ v∈T M ∈Mv NM is an -net.
Proof. Since our final -net N contains all the points in R, it is enough for us to show that for any
R − empty rectangle Q containing at least n/r points of P , for any M ∈ Mv containing Q; Q ∩ NM 6= φ.
|Q ∩ P |
c log log r
1
n/r
s
=
≥
≥
=
|M ∩ P |
ntM /s
rtM
tM
tM
(4)
The first inequality above comes from the definition of weight factor. The above equation says that,
rectangle Q contains at least (1/tM ) fraction of points in P contained in M . The result we got from the
Haussler and Welzl paper, that NM is (1/tM ) -net of size k tM log tM for M ∩ P , which means if we pick
any (1/tM ) points from M ∩ P , it will surely contain a point from set NM and |Q ∩ P | is at least (1/tM )
of |M ∩ P |. Thus it follows that Q ∩ NM 6= φ. This completes the proof.
3.2
Estimating the size of N
The expected size of N is
E[|R| + k
X
X
v
M ∈Mv
tM ≥c log log r
= cr log log r + k
tM log tM ]
X
X
v
M ∈Mv
tM ≥c log log r
(5)
tM log tM ]
We shall now fix a level i in the tree T for the rest of the analysis. Since T is a balanced binary
tree, for each node v at level i, |Pv | = n/2i . Let the union of the collections of all rectangles M ∈ Mv ,
for all the nodes v at level i be denoted by CT (R). Let CTt (R) ⊆ CT (R) be rectangles with tM ≥ t
where, t > 0. Let R0 be another random sample from P , such that each point from P is chosen in R0
independently with probability π 0 = π/t. Let C be a set of those rectangles M , anchored at the entry side
of σv for all the nodes at level i and has one point of P on each of its 3 unanchored sides. Note that,
all the degenerate rectangles can be handled in analogous manner. For a rectangle M ∈ C, we denote its
defining set D(M ) by the 3 points on its unanchored sides and its killing set K(M ) by the points that
lie in the interior of M and belong to P . We shall now define an axiom, which trivially holds true by our
construction and assumption that, no point of R is contained in Q and thus not contained in Mv .
Axiom 3.3. Every rectangle M ∈ C belongs to the set CT (R) iff D(M ) ⊆ R and (K(M ) ∩ R) = φ.
We need the following exponential decay lemma [1],[2] for our analysis.
Lemma 3.4. E[|CTt (R)|] = O(2−t E[|CT (R0 )|])
Proof. Let Z be a collection of all axis-parallel rectangles anchored at the entry side of σv , at a fixed level
i. Each unanchored sides of these rectangles has one point of Pv (degenerate cases can be handled the
same way). Let Zt ⊆ Z, contain all the rectangles with tM ≥ t. Since CTt (R) ⊆ CT (R) and since we are
summing over all the M ∈ (Zt ⊆ Z) we can write,
X
E[|CTt (R)|] =
P r{M ∈ CT (R)}
(6)
M ∈Zt
E[|CT (R0 )|] =
X
M ∈Z
P r{M ∈ CT (R0 )} ≥
X
M ∈Zt
P r{M ∈ CT (R0 )}
(7)
The above inequality comes from the fact that, Zt ⊆ Z. We are done, if we show that for each M ∈ Zt ;
P r{M ∈ CT (R)}
= O(2−t )
P r{M ∈ CT (R0 )}
Let A be the event that, D(M ) ⊆ R and K(M ) ∩ R = φ and A0 be the event that, D(M ) ⊆ R0
and K(M ) ∩ R0 = φ. From axiom 1 and the fact that all the rectangles of M ∈ C, we can say A is the
event that M ∈ CT (R) and A0 is the event that M ∈ CT (R0 ). Let us now denote ρ = |D(M )| ≤ 3 and
w = |K(M )|. A point from point set P is included in R with probability π and in R0 with probability π 0 .
Hence P r{A} = π ρ (1 − π)w and P r{A0 } = π 0ρ (1 − π 0 )w . π = s/n, π 0 = π/t and w ≥ n/r = tn/s. Hence,
w
P r{M ∈ CT (R)}
P r{A}
π ρ (1 − π)w
ρ 1−π
= O(2−t )
=
t
=
=
P r{M ∈ CT (R0 )}
P r{A0 }
π 0ρ (1 − π 0 )w
1 − π0
This concludes the proof.
We shall now apply Lemma 6, by substituting t = c log log r. For all v, the collection of all the
maximal Rv0 − empty rectangles anchored on their entry side, in their corresponding strips σv , at a fixed
level i, is denoted by CT (R0 ). Applying Lemma 4, on all the nodes of level i we get,
X
(2rv0 + 1)
(8)
E[|CT (R0 )|] =
v
Rv0
R0
rv0
|Rv0 |.
where,
=
∩ σv and =
Since all the nodes at a particular level of a balanced binary tree are
disjoint follows, sets Rv0 at level i are also disjoint. So we get,
X
rv0 = |R0 |
(9)
v
Since we terminate the construction of the balanced binary tree when the size of every leaf node ranges
between [n/r, n/2r], we can say that the maximum number of nodes at level imax (imax is the level which
has the maximum nodes) is at most
n
n
(10)
=
=⇒ 2imax = 2r
2imax
2r
Substituting equation (9) and (10) in (8) we get
X
X
E[|CT (R0 )|] =
(2rv0 + 1) ≤ 2|R0 | +
1 = 2|R0 | + 2r = O(r)
(11)
Substituting the value of
v
0
E[|CT (R )|]
v
in lemma 4, we get
E[|CTt (R)|] = O(2−t E[|CT (R0 )|]) = O(2−c log log r r) = O
r
log c r
(12)
Making it general for any j ≥ t, we can say E[|CTj (R)|] = O(r/2j ). Now we will calculate the number of
nodes from a fixed level i, that contributes to the size of the net N . Let us calculate the second part of
equation (5).
X
X
tM log tM ]
(13)
E[
v at level i M ∈Mv
tM ≥t
= E[
X
X
j log j]
j≥t M ∈CT (R)
tM =j
= E[
X
j≥t
j log j(|CTj (R)| − |CTj+1 (R)|)]
= E[(t log t)|CTt (R)| +
X
(j log j − (j − 1)log(j − 1))|CTj (R)|]
j>t
X r
r
=O
(t
log
t)
+
logj
log c r
2j
j>t
rt log t
=O
log c r
r log log r log log log r
=O
log c r
Since this is just for a fixed level i, so we repeat this analysis for all the (1 + log r) levels,
=O
r log log r log log log r
log c−1 r
Using the above result in equation (5) we get
r log log r log log log r
E[|N |] = cr log log r +
= O(r log log r)
log c−1 r
(14)
1
1
Theorem 3.5. For axis-parallel rectangles, there exists an -net ( > 0) of size O( log log ), for any
point set P of n-points in a plane.
Now using the result by Bronnimann and Goodrich [3], we come up with the following theorem,
Theorem 3.6. For the hitting set problem of axis parallel rectangles, there exists a polynomial time
algorithm which achieves a solution of size O(log log OP T ) of the optimal solution.
3.3
Key idea
It is important to see that, if we make a bad choice of sample R in the first level of sampling, the maximum
number of R − empty rectangles can be θ(s2 ). It can be seen in the quadratic lower bound example in
Figure (10a). Here each point of the lower staircase is matched with its corresponding point in upper
staircase. Hence we use the tree decomposition technique over the balanced binary tree we constructed,
to prune most of those rectangles, such that we are left with O(s log r). In Figure (10b), consider the
points to the left of lu , we show that the number of maximal R − empty rectangles anchored at lu is linear
in |R ∩ σv |, for all v ∈ T .
lu
(a)
(b)
Figure 10: (a) Quadratic lower bound example (b) Decomposition of the point set into canonical subsets
4
Conclusion
The first part of this report describes the following result that, logarithmic approximation factor for hard
capacitated set cover can be achieved from Wolsey’s work [9], using a simpler and more intuitive analysis.
We further show in our work, that O(log n) approximation factor can be achieved for the same problem
by applying analysis of general set cover to analyze Wolsey’s algorithm. This work is based on the key
observation that we make in Lemma 3. The second part of the report describes the geometric hitting
set problem, where X is a ground set of points in a plane and S is a set of axis parallel rectangles. It is
shown that -nets of size O(1/ loglog 1/) can be computed in polynomial time. Applying Bronnimann
and Goodrich result [3] gives the hitting set of size O(loglog OP T ) for this problem. One open problem is
to consider the dual version of the geometric hitting set problem, described in this report. Namely, given
a collection S of n axis parallel rectangles and each rectangle in the subset S contains some point in the
plane. We want to show the existence of a small size -net i.e S 0 ⊆ S whose union contains all the points
contained in n rectangles of S. It is not known whether the same method [2] can also be extended to the
dual version. One of the future directions that I am pursuing, is to apply the notion of hard capacities of
combinatorial set cover, to the geometric set cover problem.
References
[1] Pankaj K. Agarwal, Jiří Matoušek, and Otfried Schwarzkopf. Computing Many Faces in Arrangements of Lines and Segments. In Proceedings of the Tenth Annual Symposium on Computational
Geometry, SCG ’94, pages 76–84, New York, NY, USA, 1994. ACM.
[2] Boris Aronov, Esther Ezra, and Micha Sharir. Small-Size \eps-Nets for Axis-Parallel Rectangles and
Boxes. SIAM Journal on Computing, 39(7):3248–3282, 2010.
[3] Hervé Brönnimann and Michael T Goodrich. Almost optimal set covers in finite VC-dimension.
Discrete & Computational Geometry, 14(1):463–479, 1995.
[4] Jessica Chang and Samir Khuller. A Min-Edge Cost Flow Framework for Capacitated Covering
Problems. In ALENEX, pages 14–25. SIAM, 2013.
[5] Julia Chuzhoy and Joseph Naor. Covering problems with hard capacities. In Foundations of Computer
Science, 2002. Proceedings. The 43rd Annual IEEE Symposium on, pages 481–489. IEEE, 2002.
[6] Vasek Chvatal. A greedy heuristic for the set-covering problem. Mathematics of operations research,
4(3):233–235, 1979.
[7] Sudipto Guha, Refael Hassin, Samir Khuller, and Einat Or. Capacitated vertex covering with applications. In Proceedings of the thirteenth annual ACM-SIAM symposium on Discrete algorithms,
pages 858–865. Society for Industrial and Applied Mathematics, 2002.
[8] David Haussler and Emo Welzl. Epsilon-nets and simplex range queries. Discrete and Computational
Geometry, 2:127?151, 1987.
[9] Laurence A Wolsey. An analysis of the greedy algorithm for the submodular set covering problem.
Combinatorica, 2(4):385–393, 1982.
[10] David Zuckerman. NP-complete problems have a version that’s hard to approximate. In Structure in
Complexity Theory Conference, 1993., Proceedings of the Eighth Annual, pages 305–312. IEEE, 1993.
| 8cs.DS
|
arXiv:1801.04889v1 [math.GR] 15 Jan 2018
On non-amenable embeddable spaces in relation
with free products
Kévin BOUCHER
January 16, 2018
Abstract
In this paper we give sufficient conditions for a free product of residually
finite groups to admit an embeddable box space. This generalizes the constructions of Arzhantseva, Guentner and Spakula in [2] and gives a new class
of non-amenable metric spaces with bounded geometry which coarsely embeds
into Hilbert space.
1
Introduction
The concept of coarse embedding of a metric space into another was inspired by Gromov’s work [12] in relation with the Novikov conjecture on the homotopy invariance
of higher signatures for closed manifolds :
Definition 1. Let (X, dX ) and (Y, dY ) be two metric spaces. We say that X coarsely
embeds into Y if, there exist a map F : X → Y and two proper functions ρ± : R+ →
R+ such that:
ρ− (dX (x, y)) ≤ dY (F (x), F (x′ )) ≤ ρ+ (dX (x, x′ ))
for all x, x′ ∈ X.
This tells us that two points are far from each other in X if and only if they are
far in F (X).
This geometric notion was related by Yu [21] to the coarse Baum-Connes conjecture and in the case of Cayley graphs of groups to the Novikov conjecture.
In that remarkable paper the author introduced a notion of amenability for metric
spaces, called property (A), which implies coarse embedding into Hilbert spaces:
Definition 2. A discrete metric space (X, d) with bounded geometry has property
(A) if and only if for every R > 0 and ε > 0, there exist S > 0 and a function
φ : X → ℓ2 (X) with kφ(x)k = 1 for any x ∈ X such that:
(1) for any x1 , x2 ∈ X with d(x1 , x2 ) ≤ R, |1 − hφ(x1 ), φ(x2 )i| ≤ ε, and
1
(2) for any x ∈ X, {φ(x) 6= 0} ⊂ BS (x).
Note that this definition of property (A) is not the original but an equivalent one
for metric spaces with bounded geometry.
Definition 2 can be compared to the following notion of amenability for discrete
countable groups:
Definition 3. A discrete group G with a proper length function ℓ is amenable if and
only if for every R > 0 and ε > 0, there exists a function ξ ∈ ℓ2 (G) with kξk = 1
such that for all g ∈ B(e, R), kλ(g)ξ − ξk ≤ ε where λ : G → ℓ2 (G) is the left regular
representation.
Indeed, for given ε > 0 and R > 0, the map φ(g) = λ(g)ξ verifies the conditions
of Definition 2 and so property (A) can be thought as a non-equivariant notion of
amenability.
Actually property (A) can be characterized by the amenability of a certain groupoı̈d
associated to the given metric space, called translation groupoı̈d (cf. [19]).
Numbers of coarse properties imply property (A) and make it, in some situations,
easier to check than coarse embeddability. For example: any metric space of finite
asymptotic dimension [10], as hyperbolic spaces [17], has property (A).
In the same way, coarse embeddability of a metric space into a Hilbert space can
also be identified to a weaker notion of amenability, called a-T-menability. Indeed, a
metric space coarsely embeds into Hilbert space if and only if its translation groupoı̈d
is a-T-menable [19]. Let us recall the definition of a-T-menability in the group setting
for the sake of simplicity:
Definition 4. A discrete countable group G is a-T-menable if it acts properly on a
Hilbert space by isometries.
In group theory it is well known that the free group of rank 2 is an example of
finitely generated non-amenable but a-T-menable group. In the case of metric spaces
with bounded geometry, find a metric space without (A) but still embeddable into
a Hilbert space was a crucial question answered in [2]. In that paper the authors
exhibited a box space (cf. below) associated to the rank 2 free group which embeds
and use a girth criterion to justify that it does not have property (A) [20].
Before we introduce box spaces we must talk about coarse disjoint union of metric
spaces.
Let N be the set of integers that we endow with a far away distance df.a given by the
formula df.a(n, m) = |2n − 2m |. The important aspect of df.a is that in the neighborhoods of the infinity the points of N are arbitrarily far away from each other. Indeed
df.a (n, m) ≥ 2max(n,m)−1 for all n, m ∈ N with n 6= m and so df.a (n, m) → ∞ when
n + m → ∞ with n 6= m.
Definition 5. Let ((Xn , dn ))n be a sequence of metric spaces.
The coarse disjoint union (⊔n Xn , d) is the unique metric space (X, d) up to coarse
equivalence such that:
2
(1) there exist a map p : X → N and a proper function ρ : R+ → R+ such that p−1 (n)
is isometric to Xn for all n and df.a (p(x), p(y)) ≤ ρ(d(x, y)) for all x, y ∈ X.
(2) For all R > 0, Un,m = {x ∈ Xn | d(x, Xm ) ≤ R} is a bounded set of Xn for all
n, m.
As it is explained in Section 3 there is a correspondence between sequence of metric spaces with their arrows (cf. Definition 10) and coarse disjoint union (X, d, p, ρ)
with classical coarse arrows that commute with the map p.
Let G be a discrete residually finite group endowed with a proper length ℓ and
N = (Nn )n a sequence of finite index normal subgroups with trivial intersection.
Definition 6. The box space of G with respect to N , denoted by N G, is the coarse
disjoint union ⊔n G/Nn , where each finite quotient Q/Nn is endowed with the quotient
length induced by ℓ (cf. Section 3).
Box spaces are defined up to coarse equivalence and by Proposition 4 and Corollary 1 they do not depend on the choice of ℓ.
One of the key points in box space theory is that geometric properties on a group
box space translate to analytic/harmonic properties on this group:
G is amenable ⇐⇒ N G has property A,
G is a-T-menable ⇐= N G is coarsely embeddable into a Hilbert space,
G has property (T) =⇒ N G is an expander.
(1)
(2)
(3)
The Proofs of the first two facts can be found in [16] and the third in [15].
Recall that property (T) is a strong negation of a-T-menability in the following sense:
Definition 7. A discret countable group G has property (T) if any isometric action
of G on a Hilbert space has a fixed point.
Analogously to the relation between property (T) and a-T-menability, the notion
of expansion, mentioned in the correspondance (3) and explained in Section 5, is an
obstruction to coarse embeddability of a given metric space into Hilbert space (cf. [13]
and [15] for more about expansion).
Let us mention that in some sense the converse of the correspondence (2) exists.
A generalized notion of coarse embedding called fibered coarse embedding was introduced in [7]. Since certain expander sequences can admit such kind of embedding, it
appears to be strictly weaker than the original notion of coarse embedding.
The strength of this generalization is that it can be associated to any geometric action
of a residually finite group on a metric space. Moreover, in the context of Lp spaces,
there is an equivalence between the existence of a proper action on a Lp space and
existence of a fibered coarse embedding of its box space into it [1].
3
In this paper we investigate finite quotients of free products.
Let’s consider A and B two discrete residually finite groups that we suppose to be
finitely generated.
Let G = A ⋆ B be their free product (which is residually finite cf. [4]) and ℓ a proper
length on G.
The main result of this paper is:
Theorem 1. If A and B are amenable, then G admits a box space which coarsely
embeds into Hilbert space.
In other words, as we use this terminology in the rest, G has a faithful sequence
of finite quotients which coarsely embeds into Hilbert space (cf. Lemma 2).
Definition 8. Let G be a residually finite group. A infinite sequence of (right) quotient (Qn )n of G is called faithful if any element of G acts non-trivially by left translation on all but a finite number of these quotients.
Since G is non-amenable when |A|, |B| ≥ 3, Theorem 1 and the box space correspondence (1) described before gives us a new class of non-amenable metric spaces
with bounded geometry which coarsely embeds into Hilbert space.
Moreover, correspondence (2) combined with our result give another proof with a
different flavor of the well known result:
Theorem. Let A and B be two discrete countable amenable groups. Then their free
product G = A ⋆ B is a-T-menable [8].
We are also interested by what can be done beyond the free product of amenable
groups case.
Given a free product G = A ⋆ B, a necessary condition for it to admit an embeddable
box space is that its factors A and B admit such box spaces (cf Proposition 6). So
a natural question is: does the free product of discrete residually finite groups which
have an embeddable box space have itself such an embeddable box space?
It is not clear if our method can be adapted to treat this situation. As it is explained
in the Outline, a part of our approach uses an amenable extension type result adapted
to spaces obtained as coarse disjoint union. But even in the context of classical metric
spaces we cannot expect, in full generality, that an extension of embeddable spaces to
be itself embeddable (cf. [3]).
However we have the following partial answer:
Let A, B and G as above. We assume that A and B are both finitely generated and
we fix a finite generating set S ⊂ G of G.
Proposition 1. If A and B have an embeddable box space then G has a faithful
sequence of irreducible symmetric representations (σn : G → Sym(Xn ))n such that
the coarse disjoint union of Schreier graphs ⊔n SchS (Xn ) coarsely embeds into Hilbert
space.
As explained in Section 5 such symmetric representations are nothing but quotients induced by non-necessary normal subgroups.
4
Even if sequences of symmetric representations still preserve the direct arrow of correspondence (1) and the implication (3). We have to keep in Definition 6 of a box
space the assumption that the sequence of subgroups we consider are normal to preserve some part of the correspondences described before. Indeed the following result,
proven in Section 5, shows that considering symmetric representations instead of finite
quotients breaks the correspondence (1):
Proposition 2. Let F2 be the free group of rank 2. There exists a faithful sequence
of symmetric representations (σn : F2 → Sym(Xn ))n of F2 such that the sequence of
Schreier graphs ⊔n Sch(Xn ) has property (A).
Our investigation on free products was motived by the following observation:
Consider a free product G = A ⋆ B. G acts without bounded orbits on a tree (cf. [18]
and Section 4) and so does not have property (T) [5]. This finds an analogue at finite
scale in box space theory:
Proposition 3. Let A, B be two residually finite groups and G = A ⋆ B their free
product. Then G has a non-expander box space.
In other words, we cannot expect a free product with property (τ ) [13] even if both
of its factors has property (T). The proof, given in Section 5, uses only elementary
considerations but as far as we know has never been written.
We finish this introduction by a question. It deserves to be noted that all known
examples of non-amenable embeddable metric spaces come from groups which admit
(or contain a large subgroup who does) an action without bounded orbits on a tree.
Can we find an example of embeddable non-amenable box space coming from a group
with property (FA) [5]?
Recall that a group has property (FA) if all of its actions on a tree have a fixed point.
2
Outline
Here we explain the idea behind the proof of Theorem 1 and detail the structure of
this paper.
Let us start with the classic case of discrete metric spaces.
We consider the free product of two amenable groups G = A ⋆ B endowed with a
proper length function ℓ.
The group G can be described as an extension of the direct product A × B by a free
subgroup D (cf. Lemma 3):
1→D →G→A×B →1
At first look this exact sequence is only algebraic and apriori does not give any information on the geometry of (G, ℓ). Now if we endow D with the restricted length
ℓ|D and A × B with the quotient length ℓ induced by ℓ. Some aspects of the coarse
geometry of (G, ℓ) as uniform embeddability can be understood through the geometry
5
of (D, ℓ|D ) and (A × B, ℓ).
Indeed, if A, B and so A × B are amenable, they have property (A) as metric groups
with respect to any proper length and by a theorem of [9] the coarse embedding of
(G, ℓ), can be reduced to the coarse embedding of (D, ℓ|D ). Moreover since proper
length functions on a given group are unique up to coarse equivalence. The length
function ℓ|D on D can be replaced by any other proper length function. So a trick
to prove the uniform embedding of D ≃ F∞ endowed with ℓ|D is to realize it as a
subgroup of the rank 2 free group F2 endowed with a word metric on a basis for which
it is known that an uniform embedding exists (cf. [6]) and then use the uniqueness
up to coarse equivalence of proper lengths on F∞ to conclude.
Note that as soon as G is non-trivial and A or B has infinite cardinality, D has infinite
rank (cf. Proposition 7) and so the class of proper length on D are not the natural
one. In particular ℓ|D cannot be replaced by a word length on a generating of D,
since the latter distance is not proper on the infinite rank free group.
Our approach, under the assumption that A and B are residually finite, is to do
something similar at finite scale.
Let A and B be two finite quotients of respectively A and B. We denote G = A ⋆ B
the free product of these two groups.
Then the following commutative diagram holds:
1
1
//
D
//
D′
1
A×B
//
A×B
//
//
G
//
// 1
G
// 1
1
1
where the free subgroup D ′ has, this time, finite rank (cf. Lemma 3).
Our prototype of finite quotients has the form G/H where H E D ′ is a finite index
normal subgroup of D ′ . In particular G/H appears as an extension of A × B by a
large girth subgroup D ′ /H.
This process, which is detailed in Section 4, gives faithful sequence under reasonable
assumptions on the choice of the A, B and H.
Now, let (Gn )n be a faithful sequence of finite quotients of G built as described
before. We endow each Gn with the quotient length induced by ℓ that we denote the
same way. Here we assume that Gn is an extension of a direct product An × Bn of
finite quotients of A and B by a subgroup of large girth Dn .
As in the classic case, our approach is to reduce the coarse geometry of the metric
space ⊔n (Gn , ℓ) to the geometry of ⊔n (Dn , ℓ|Dn ) and ⊔n (An × Bn , ℓ).
The first step is to reduce the proof of Theorem 1 to the uniform embedding of
the sequence ⊔n (Dn , ℓ|Dn ) (cf. Theorem 2) under the assumption that A and B are
amenable. This is done in Subsection 4.1 where we develop gluing techniques inspired
by [10] for metric spaces obtained as coarse disjoint union.
The second step of the proof which concerns the uniform embedding of the space
6
⊔n (Dn , ℓ|Dn ) is a bit more delicate. Indeed if we refer to the preceding notations,
we have to do a suitable choice of the H’s inside of the D ′ ’s to ensure the uniform
embedding of ⊔n (Dn , ℓ|Dn ). The difficulty here is that the proper lengths on the Dn ’s
do not correspond to the natural class of length we would like to consider on them as
soon as G is non-trivial and A or B has infinite cardinality, if we refer to our remark
on the classic case. To solve this part we compare the sequence ((Dn , ℓ|Dn ))n to the
one obtained by taking of each Dn the quotient length induced by D endowed with
a word length on a suitable basis (cf. Proposition 7). We call this new sequence the
combinatorial sequence. Note that, this procedure is not free. Indeed, the compression
function ρn,− of (Dn , ℓ|Dn ) → (Dn , ℓComb ) (cf. Section 3) can be estimated at every
stage n and verifies |An | + |Bn | ≤ ρn,− (1). This inequality tells us, as suspected,
that these two sequences of metric spaces are non-equivalent to each other if G is
non-trivial and A or B has infinite cardinality. In spite of this, some local comparison
is possible and this is enough to conclude.
Let us explain it in the classic case:
Let F(S) be the free on an infinite countable set S endowed with a proper length ℓ
and Fk = F(Sk ) where Sk ⊂ S is a finite subset of S with Sk ⊂ Sk+1 and S = ∪k Sk .
Since ℓ is proper, for all R ≥ 0, there exists K0 such that Bℓ (e, R) ⊂ Fk for all k ≥ K0 .
Since Fk embeds uniformly a direct limit argument can be applied to prove that F(S)
uniformly embeds too (cf. [9]).
An investigation of the geometry of D allows us to build a sequence (Xk )k inside
of ⊔n Dn which plays a role similar to Fk in the classic case. The fundamental point is
that each Xk can be compared, with controlled compression and dilatation function,
to its combinatorial version on all fibers of ⊔n Dn except a finite number of them, in
other words the two metric structures on Xk match at infinity (cf. Corollary 4). As a
consequence it appears that the uniform embedding of ⊔n (Dn , ℓ|Dn ) is implied by the
uniform embedding of its combinatorial version. This is explained in Subsection 4.2
where we introduce a direct limit formalism and establish a stability result for coarse
disjoint unions of groups (cf. Proposition 11). So to conclude Theorem 1 proof it is
enough to treat the combinatorial case. But the uniform embedding problem in this
case can be solved by a slight modification of [2] arguments (cf. Proposition 12).
This paper is organized as follows: we start with Section 3 where we introduce
general results and terminologies that we need in the rest. Section 4 is devoted to
the proof of Theorem 1. We finish with Section 5 where we prove the Proposition 1,
2 and 3.
3
Basics
In this section we expose generalities on metric groups and spaces.
First, let us briefly recall terminologies and constructions about length functions.
Here G is a discrete countable group.
Definition 9. A length function ℓ : G → R+ on G is a function which satisfies:
7
(1) for any g, h ∈ G, ℓ(g.h) ≤ ℓ(g) + ℓ(h).
(2) for any g ∈ G, ℓ(g) = ℓ(g −1 ).
(3) ℓ(g) = 0 if and only if g = e.
We say that ℓ is proper if for any R > 0, ℓ−1 ([0, R]) ⊂ G is a finite set of G, in other
words, each ball of G is finite.
Given a length ℓ on G, we associate a left-invariant metric d by the formula
d(g, h) = ℓ(g −1h). Moreover, starting with the left-invariant metric d, ℓ(g) = d(g, e)
is a length function on G. In the rest we pass to one to the other without distinction.
Let H E G be a normal subgroup of G and ℓ a proper length function on G.
The quotient length on Q = G/H induced by ℓ is defined by the formula: ℓQ (g) =
inf h∈H ℓ(gh) where g ∈ Q is the class of an element g ∈ G.
Since H is normal it is routine to prove that conditions (1) and (2) of Definition 9
are verified.
Note that there is no natural way to build a distance on a G-homogeneous space in
full generality.
Since ℓ is proper, for all ε > 0, the set of h ∈ H such that ℓ(gh) ≤ ℓQ (g) + ε is finite.
This ensures that we can find h0 such that ℓQ (g) = ℓ(gh0) and so for any g ∈ Q,
ℓQ (g) = 0 implies that g ∈ H. Condition (3) is verified. Another consequence is
that for any R ≥ 0, π(BG (g, R)) = BQ (g, R) where π is the quotient map. This set
equality ensures that ℓQ is proper.
Given a symmetric generating set S ⊂ G, the word length on G, ℓ : G → R+ ,
associated to S is defined by:
ℓ(a) = inf{n | a = a1 . . . an , with ai ∈ S}
for a 6= 0 and ℓ(e) = 0. It appears that when S is finite, the length ℓ is proper.
In the case of free groups we assume this generating set to be a basis [14].
Note that, concerning the coarse geometry of a locally finite metric group the
choice of the proper length is not matter as the following proposition shows.
Proposition 4. Let G be a discrete countable group.
Suppose ℓ1 and ℓ2 are two proper length functions on G. Then (G, ℓ1 ) is coarsely
equivalent to (G, ℓ2 ).
Proof. Let’s define ρ− (t) = inf {g|ℓ1 (g)≥t} ℓ2 (g) and ρ+ (t) = sup{g|ℓ1 (g)≤t} ℓ2 (g) which are
well defined because of the properness assumption on ℓ1 . These two increasing functions are called respectively compression and dilatation. They verify the following
inequalities:
ρ− (d1 (g, h)) ≤ d2 (g, h) ≤ ρ+ (d1 (g, h))
for any g, h ∈ G. Moreover, they are respectively the largest and smalest control
function such that the above inequalities hold. Actually, coarse embeddability of a
8
space into another is equivalent to the existence and properness of ρ± .
Using their definitions, in order to verify that ρ− (respectively, ρ+ ) is proper, it
is enough to prove that, for all R > 0, there exists R′ > 0 such that ℓ2 (g) ≤ R
(respectively, ℓ1 (g) ≤ R) implies ℓ1 (g) ≤ R′ (respectively, ℓ2 (g) ≤ R). But since ℓ2
(respectively, ℓ1 ) is proper and G is discrete, the set {g | ℓ2 (g) ≤ R} (respectively,
{g | ℓ1 (g) ≤ R}) is finite and thus any bound of ℓ1 (respectively, ℓ2 ) on this set do
the job.
Let’s introduce some terminologies and results about sequences of metric spaces.
Definition 10. Let X = ((Xn , dXn ))n and Y = ((Yn , dYn ))n be two metric space
sequences. We say X coarsely embeds in Y if there exists a sequence of maps F =
(Fn : Xn → Yn )n and two proper functions ρ± : R+ → R+ such that
ρ− (dXn (x, y)) ≤ dYn (Fn (x), Fn (x′ )) ≤ ρ+ (dXn (x, x′ ))
for any n and x, x′ ∈ Xn .
If moreover (Fn (Xn ))n is uniform coarsely dense in Y, i.e if there exists R > 0 such
that Yn is contain in the R-neighborhood of Fn (Xn ) for any n, we say that F is a
coarse equivalence or X and Y are coarsely equivalent.
In this paper we consider principally embeddings of sequences of metric spaces
into Hilbert space, when there no confusion we just say that the sequence coarsely
embeds.
Lemma 1. Let Hn be a sequence of Hilbert spaces. The space ⊔n Hn coarsely embeds
into Hilbert space.
Proof. Let’s denote X = ⊔n Hn the coarse disjoint union of (Hn )n and ρ, p are as in
Definition 5.
L
We consider H = n Hn ⊕ C and define the map F : X → H given by the formula:
F (vn , n) = (vn ) ⊕ 2n .
We have that:
kvn − v ′ k
if n = m
′
(4)
kF (vn ) − F (vm )k = p n m m 2
2
′
2
|2 − 2 | + kvn k + kvm k otherwise
Now, let’s consider
ρ− (t) =
inf
′
kF (vn ) − F (vm
)k
sup
′
kF (vn ) − F (vm
)k
′ )∈X ×2 |t≤d (v ,v ′ )}
{(vn ,vm
X n m
and
ρ+ (t) =
′ )∈X ×2 |d (v ,v ′ )≤t}
{(vn ,vm
X n m
, respectively, the compression and dilatation functions.
By considering separately the case n = m and n 6= m, the formula (4) implies that,
′
for all R ≥ 0, there exists R′ ≥ 0 such that kF (vn ) − F (vm
)k ≤ R implies that
9
′
dX (vn , vm
) ≤ R′ . In other words ρ− is proper.
′
′
′
In the other hand, since dX (vn , vm
) ≥ ρ(df.a (p(vn ), p(vm
))) for all vn , vm
∈ X, there
exists Nt such that dX (Hn , Hm ) ≥ t + 1 for all n, m ≥ Nt with m =
6 n.
It follows that:
′
)k
ρ+ (t) = sup kF (vn ) − F (vm
At ∪Bt
′
′
) ≤ t} and Bt = {(vn , vn′ ) ∈
with At = {(vn , vm
) ∈ (∪n≤Nt Hn )×2 | dX (vn , vm
(∪n Hn )×2 | dX (vn , vn′ ) ≤ t}. Using part (2) of the Definition 5, there exists Rt ≥ 0
′
such that dX (vn , vm
) ≤ t implies that vn ∈ Bn (0, Rt ) ⊂ Hn and vm ∈ Bm (0, Rt ) ⊂ Hm
for all n 6= m with n, m ≤ Nt .
p
As a consequence ρ+ (t) ≤ max (t, 22Nt +2 + 2Rt2 ), in particular ρ+ is well defined.
The following inequalities imply the properness of ρ+ :
sup kF (vn ) − F (vn′ )k = sup kvn − vn′ k
Bt
Bt
= t ≤ ρ+ (t).
The following lemma allows us to pass to a language to another concerning embeddings of sequences of metric spaces.
Lemma 2. Let X = (Xn )n be a sequence of metric spaces. The following are equivalent:
(1) ⊔n Xn coarsely embeds into Hilbert space.
(2) There exist a sequence of maps F = (Fn : Xn → Hn )n where Hn denote a Hilbert
space and two proper functions ρ± : R+ → R+ , such that:
ρ− (dXn (x, y)) ≤ kFn (x) − Fn (x′ )k ≤ ρ+ (dXn (x, x′ ))
for all n and x, x′ ∈ Xn . In other words, X coarsely embeds into Hilbert spaces.
(3) For all R > 0 and ε > 0, there exists a sequence (ϕn )n of maps ϕn : Xn → Hn
with Hn a Hilbert space, such that kϕn (x)k = 1 and:
(i) for all n ≥ 1 and x, y ∈ Xn with dXn (x, y) ≤ R, kϕn (x) − ϕn (y)k ≤ ε.
(ii) for all ε′ > 0, there exists M > 0, |hϕn (x)|ϕn (y)i| ≤ ε′ whenever dXn (x, y) ≥
M.
Proof. (1) implies (2): This is because Xn embeds isometrically into ⊔n Xn .
(2) implies (1): Let F , ρ± be as in (2) and dX the metric on X = ⊔n Xn .
Let F : (X, dX , ρX , pX ) → (H = ⊔n Hn , dH , ρH , pH ) be the map induced by F given
by the formula F (xn ) = Fn (xn ).
Let’s fix x0,n ∈ Xn for all n. We may assume that ρX = ρH = ρ (otherwise we take
10
the infimum).
We have the following inequalities hold:
ρ(df.a (pX (xn ), pX (xm ))) = ρ(df.a (pH ◦ Fn (xn ), pH ◦ Fm (xm )))
≤ dH (Fn (xn ), Fm (xm ))
(5)
ρ− (dX (xn , x′n )) ≤ dH (Fn (xn ), Fn (x′n )) ≤ ρ+ (dX (xn , x′n ))
(6)
and
for all xn , x′n ∈ Hn and xm ∈ Hm .
Let’s prove the properness of the dilatation function (the proof for the compression
follows the same scheme).
If dX (xn , xm ) ≤ R, in the case n = m by inequalities (6) we have dH (Fn (xn ), Fm (xm )) ≤
R′ = ρ(R). Otherwise n 6= m so by inequalities (5) n, m ≤ N where N depends on
R and the part (2) of Definition 5 ensures that d(x0,n , xn ), d(x0,m , xm ) ≤ M. Using
inequalities (6) once again we have d(Fn (x0,n ), Fn (xn )), d(Fm (x0,m ), Fm (xm )) ≤ M ′ .
In other words the dilatation function is proper. We use the Lemma 1 to conclude.
P
⊗k
(2) implies (3): Take a vector ξ in a Hilbert space H, then exp(ξ) = k≥0 ξn! ∈
L
′
⊗k
has a sense and hexp(ξ)| exp(ξ ′ )i = ehξ|ξ i for any ξ, ξ ′ ∈ H.
k≥0 H
√
2
Now let’s consider F and ρ± as in (2) and define ϕtn (x) = e−tkFn (x)k exp( 2t.Fn (x))
2
with t > 0 which verifies: hϕtn (x)|ϕtn (y)i = e−tkFn (x)−Fn (y)k . This relation implies, for
any t > 0, kϕtn (x)k = 1 and :
e−tρ+ (dXn (x,y)) ≤ |hϕtn (x)|ϕtn (y)i| ≤ e−tρ− (dXn (x,y))
for any x, y ∈ Xn . Since ρ+ is proper, given ε > 0 we can find t0 > 0 such that
e−t0 ρ+ (r) ≥ 1 − 2ε , for any r ≤ R. For such choice of parameter t, part (i) of (3) holds
with ϕtn0 . Furthermore because of the properness of ρ− , condition (ii) of (3) is verified
for any parameter t fixed. So the implication is proven.
√
(3) implies (2): Let’s take εl = 1/l and Rl = l. Let (ϕln )n with ϕln : Xn → Hnl
as in (3) for the constants εl and Rl . Let (Ml )l such that kϕln (x) − ϕln (x′ )k ≥ 1
whenever dXn (x, x′ ) ≥ Ml that we can assume to be an increasing sequence
L which
goes to infinity. We fix a point x0n ∈ Xn for any n and define Fn : Xn → l Hnl by
Fn (xn ) = 21 ((ϕ1n (xn ) − ϕ1n (x0n )) ⊕ (ϕ2n (xn ) − ϕ2n (x0n )) ⊕ . . . ), then:
ρ− (dXn (x, x′ )) ≤ kFn (x) − Fn (x′ )k ≤ dXn (x, x′ ) + 1
√
P
for any n and x, x′ ∈ Xn with ρ− = 21 k≥1 k − 1χ[Mk−1 ,Mk ) where χ[Mk−1 ,Mk ) is the
characteristic function of the interval √
[Mk−1 , Mk ).
√
Indeed let x, x′ ∈ Xn and l such that l − 1 ≤ dXn (x, x′ ) ≤ l. We have:
1X k
1 X
kϕkn (x) − ϕkn (x′ )k2 +
kϕ (x) − ϕkn (x′ )k2
4 k≤l−1
4 k≥l n
1X 1
≤ (l − 1) +
≤ d(x, x′ )2 + 1
2
4 k≥l k
kFn (x) − Fn (x′ )k2 =
11
Similarly if we take l such that Ml−1 ≤ d(x, x′ ) ≤ Ml , then
kFn (x) − Fn (x′ )k2 ≥
1 X
l−1
kϕkn (x) − ϕkn (x′ )k2 ≥
= ρ− (dXn (x, x′ ))2 .
4 k≤l−1
4
The next proposition can be view as coarse analogue of the fundamental theorem
of geometric group theory [11] for box spaces. It finds, in others, applications in the
proof of Theorem 3 (cf Section 5) which is the cornerstone of Proposition 1 and 2 but
also in Corollary 5 which is an important part of Theorem 1 proof.
Proposition 5. Let G be a discrete countable group with a proper left-invariant distance d. Let X be a proper metric on which G acts freely and geometrically (i.e.
properly discontinuously by isometries) and f : G → X a G-equivariant coarse embedding.
If N = (Nn )n is a sequence of normal subgroups of G with trivial intersection, then the
sequence of metric groups (G/Nn )n coarsely embeds into the sequence X = (X/Nn )n .
If moreover f is a coarse equivalence, then (G/Nn )n and X are coarsely equivalent.
Proof. Let G, X, d, f and N be as in the statement. Let ρ± be the two proper
functions, that we may assume increasing, such that:
ρ− (d(g, h)) ≤ dX (f (g), f (h)) ≤ ρ+ (d(g, h))
for any g, h ∈ G.
Since f is G-equivariant, for any g ∈ G, f (g) = g.x0 with x0 = f (e). Moreover f
induces a sequence of maps fn : G/Nn → Xn (since Nn E G is normal, gNn = Nn g
for any g ∈ Nn ) which is nothing but the well defined map fn (x) = f (x).
The distance dX,n on X/Nn is given by the formula dX,n (x, y) = inf h,h′∈Nn dX (h.x, h′ .y).
Let’s define ℓx0 ,n : G/Nn → R+ by the formula ℓx0 ,n (g) = inf h∈Nn dX (gh.x0 , x0 ) =
dX,n (g.x0 , x0 ) and denote ℓn the quotient length induced by d on G/Nn . Since G
acts isometrically it is enough for our purpose to prove the existence of two proper
functions ρ′± such that:
ρ′− (ℓn (g)) ≤ ℓx0 ,n (g) ≤ ρ′+ (ℓn (g)).
(7)
for any n and g ∈ G/Nn .
Let’s prove the left side of the inequality (7). Because ρ− is proper and does not
depend on n, for any R > 0, g ∈ G with ℓx0 ,n (g) ≤ R and ε > 0, there exists R′
independent of n such that: any h ∈ Nn with
ρ− (ℓ(gh)) ≤ dX (gh.x0 , x0 ) ≤ ℓx0 ,n (g) + ε ≤ R + ε
is contained in a ball with center g −1 and radius R′ which is a finite set since ℓ is
proper. In particular for such h ∈ Nn , ℓn (g) ≤ ℓ(gh) ≤ R′ . In other words, for any
R > 0, there exists R′ > 0 such that for any n, ℓx0 ,n (g) ≤ R implies ℓn (g) ≤ R′ .
12
For the right side of inequality (7) we do something similar. Since ℓ is proper,
there exists h0 ∈ Nn such that, ℓn (g) = ℓ(gh0) and so
ℓx0 ,n (g) ≤ dX (gh0 .x0 , x0 ) ≤ ρ+ (ℓ(gh0)) = ρ+ (ℓn (g))
The second part of the proposition is because the R-density of f (G) ⊂ X is preserved
by taking quotient.
Here are some consequences of Proposition 5.
Corollary 1. Let G be a discrete residually finite group and ℓi , i = 1, 2, be two proper
lengths on G.
Let N = (Nn )n be a sequence of finite index normal subgroups of G with trivial intersection, then N ,1 G and N ,2 G are coarsely equivalent with dilation and compression
functions which depend only on ℓ1 and ℓ2 . By N ,i G we mean the box space obtained
by taking the quotient metric associated to ℓi , i = 1, 2, on each G/Nn .
Proof. We apply proposition 5 to the identity map Id : (G, ℓ1 ) → (G, ℓ2 ) which is a
G-equivariant coarse equivalence by Proposition 4.
Corollary 2. Let G be a discrete residually finite group and H ≤ G a finite index
subgroup of G. Suppose there exists N ′ = (Nn′ )n a sequence of finite index normal
subgroups of H with trivial intersection such that N ′ H embeds into a Hilbert space
(or any other metric space). Then G also admits such a sequence.
Proof. Let’s fix a proper length ℓ on G. The inclusion map Id|H : (H, ℓ|H ) → (G, ℓ) is
a H-equivariant isometry so Proposition 5 can apply (remember that, ℓ|H is equivalent
to any other proper length on H by Proposition 4).
As in the classical case, we have heredity on the uniform embedding of box spaces.
Proposition 6. Let G be a discrete residually finite group and H ≤ G a subgroup of
G. Suppose there exists N = (Nn )n a sequence of finite index normal subgroups of
G with trivial intersection such that N G embeds into a Hilbert space (or any other
metric space), then H also admits such a sequence.
Proof. N ∩ H = (H ∩ Nn )n defines a sequence of finite index normal subgroups of H
with trivial intersection. If we endow H with the restriction of a given proper length
ℓ on G then N ∩H H embeds isometrically into N G.
4
Finite quotients of a free product
This section, devoted to the proof of Theorem 1, is organized as follows:
We start by a succinct introduction to Bass-Serre theory [18] that we need in the rest
and which allows us to formulate and prove the Proposition 7.
After that we present a process to obtain faithful sequences of finite quotients of
a given free product of residually finite groups with an interesting decomposition
13
property. As mentioned in the outline, this construction and the sequences that it
produces are in the heart of the proof of Theorem 1.
In Subsection 4.1 we introduce gluing techniques and give a reduction of Theorem
1 to the uniform embedding of a certain sequence of large girth subgroups endowed
with particular metrics. This is Theorem 2.
We finish Theorem 1 proof in Subsection 4.2 which is dedicated to the study of this
large girth sequence. In this subsection we introduce direct limit formalism for sequence of groups and approximate our sequence (cf. Proposition 11) with another
natural metric structure on which we have a better grasp.
Let A, B be two discrete countable groups and G = A ⋆ B their free product. We
can associate to G a tree, called Bass-Serre tree [18], T given by:
V (T ) = G/A ∪ G/B
E(T ) = G
with the endpoint maps, the projections G → G/A and G → G/B.
The action of G on its cosets is combinatorial with respect to T (i.e it respects the
adjacency) and defines an isometric action on T which has two orbits and stabilizers
equal to a conjugate of A or B depending on the family of cosets we consider.
We can deduce from this a simple criterion to characterize normal subgroups of G
which act freely on the Bass-Serre tree T : a normal subgroup H E G acts freely on
T if and only if H has trivial intersection with A and B.
The following lemma is a normal form type result on quotients of a free product:
Lemma 3. Let A and B be two discrete countable groups and G = A ⋆ B their free
product.
Given a quotient Q of G, A (respectively, B) factorizes through a quotient, denoted
A (respectively, B), such that A and B are subgroups of Q and :
Q ≃ A ⋆ B/F
where F E A ⋆ B is a free normal subgroup.
Furthermore if Q is finite then A, B and rk(F ) are finite.
Proof. Let π : G → Q be the canonical projection. If we choose A = A/ker(π|B ) and
B = B/ker(π|B ), π factor through the following diagram:
G
π
// Q
②<<
②
π ②②②
②
②
②
A⋆B
Moreover, since π|A (respectively, π|B ) has trivial kernel by construction, A and B
are subgroups of Q. In particular, if Q is finite then A and B are finite.
The normal subgroup F = ker(π) of G = A ⋆ B does not intersect A and B. So F
14
acts freely on the Bass-Serre tree T of G and this implies that F is a free group [18].
Moreover if Q is finite the action of F on T is cocompact. Since T has bounded
geometry (the geometry of T is controlled by the cardinality of A and B which are
finite), these two facts imply that F has finite rank [18].
Let Y be a graph. It is well known that any maximal tree subgraph of Y contains
every vertices of Y .
Given a group G which acts simplicialy on X, a tree representation of X/G is a
connected subgraph X ∗ of X which is mapped isomorphically by the projection to a
maximal tree subgraph of X/G. Note that such a representation always exists.
Proposition 7. Let A and B be two discrete countable groups and G = A ⋆ B their
free product. Let D be the kernel of the morphism π : G → A × B. Then D has basis
B = {[a, b] | a ∈ A, b ∈ B; a, b 6= 1} and there exist constants C, λ > 0 independent of
A and B such that (D, B) is (C, λ)-quasi-isometric to the Bass-Serre tree associated
to G.
Proof. An orientation on the Bass-Serre tree T of G which is preserved by the action
corresponds to a choice of an orientation of the quotient G\T . In our case there is
only two choices and we assume here that oriented edges on T goes from B-cosets to
A-cosets.
The subgroup D is normal and does not intersect A and B, it follows that D acts
freely on the tree T and so is a free group [18].
Let T ′ = D\T be the quotient T by the action of D, then T ′ is a complete bipartite
graph given by:
V (T ′ ) = (A × B)/A ⊔ (A × B)/B
E(T ′ ) = A × B
with endpoint maps given by the quotient maps on, respectively, A and B.
In particular a tree representation of T ′ is given by the subtree Tr of T with:
V (Tr ) = {aB; a ∈ A} ⊔ {bA; b ∈ B} ⊂ V (T )
E(Tr ) = A ∪ B ⊂ G = E(T )
Let ∂V Tr+ be the set of vertices of T \ Tr which are target of a edge with source a
vertices of Tr . Then the set B = {g ′ 6= 1 | g ′ ∈ D, g ′.Tr ∩ ∂V Tr+ 6= ∅} which is, by
direct computation, equal to {[a, b]; a ∈ A, b ∈ B} forms a basis of the free group D
(this geometric criterion can be found Section 3 [18]). It appears that the rank of D
is finite when A and B are finite.
Since G acts isometrically, to prove the last part of the proposition it is enough
to prove the following inequalities:
1
.dT (g.A, A) ≤ ℓ(g) ≤ dT (g.A, A) + 1
4
15
(8)
with g ∈ D. These will imply that f : D → T with f (g) = g.A is a (4, 1)-quasiembedding but T ′ has diameter equal to 2, in other words, f (G) is 2-dense and by
consequence defines a (4, 1)-quasi-isometry.
Since dT (s.A, A) = 4 for any s ∈ S = B ∪ B−1 , given g = s0 . . . sn ∈ D a reduced
word with si ∈ S and n ≥ 1, we have :
d(g.A, A) ≤
n−1
X
d(s0 . . . si .A, s0 . . . si+1 .A) =
i=0
n−1
X
d(A, si+1 .A) = 4.ℓ(g)
i=0
This prove the left side of the inequalities (8).
For the right side, let g ∈ D and x0 = A, . . . , xn = g.A a sequence of vertices which
describes the geodesic path in T between A and gA. Since D.Tr = T , we can find
g0 = e, g1 , . . . , gn = g such that xi ∈ gi .Tr for any i = 0, . . . , n. If we denote s0 = g1
and si = gi−1 gi+1 for 1 ≤ i ≤ n − 1, then g = s0 . . . . .sn−1 and:
dT (si Tr , Tr ) = dT (gi+1 .Tr , gi .Tr ) ≤ d(xi+1 , xi ) = 1
so si ∈ S for any i.
Indeed, if we denote ∂V Tr− the analogue of ∂V Tr+ for the opposite orientation on T .
An element g ′ ∈ D belongs to S if and only if g ′.Tr ∩∂V Tr+ 6= ∅ or g ′ .Tr ∩∂V Tr− 6= ∅. A
consequence is that d(g ′Tr , Tr ) ≥ 2 whenever g ′ ∈ D \ S. It follows that ℓ(g) ≤ n.
Let us consider A, B two discrete countable groups as above and A, B two quotients of respectively A and B. Let G = A ⋆ B (respectively, G = A ⋆ B) their free
product. We have the following commutative diagram:
1
1
//
D
//
D
1
′
//
A×B
G
//
//
// 1
G
//
A×B
// 1
1
1
Note that, the basis B of D described in Proposition 7 is sent to the basis B′ of D ′
defined analogously by the quotient map.
In the rest of the text we will talk about commutator subgroups to designate
kernel’s of type D (·) .
Let Γ be a residually finite group. Recall that a sequence of finite quotients of Γ is
called faithful if any non-trivial element g ∈ Γ acts non-trivially by left translation
on all but a finite number of these quotients.
Now, if we assume A and B residually finite, it is well know that their free product
G = A ⋆ B is residually finite [4] (the construction described in that paper is recall
and used in the next section).
Here we propose another construction. Let (An )n and (B n )n be two faithful sequence
of finite quotients of respectively A and B. We denote Gn = An ⋆ B n and Dn the
16
commutator subgroup of Gn which has finite rank since An and B n are finite as
explained above. Let (M(n,k) )k be a sequence of finite index characteristic subgroups
of Dn , i.e M(n,k) is normalized by each automorphism of Dn . Moreover we assume
that the subgroups (M(n,k) )k have trivial intersection and that M(n,k+1) is a proper
characteristic subgroup of M(n,k) .
For example we can take M(n,k+1) = h{w s | w ∈ M(n,k) } ∪ {[w, w ′] | w, w ′ ∈ M(n,k) }i
with s ≥ 2 fixed integer. Since Dn has finite rank it is not hard to see that M(n,k) has
finite index in Dn and Proposition 3.3 of [14] ensures that the intersection is trivial.
Proposition 8. The sequence of finite quotients (Gn /M(n,k) )n,k≥n of G is faithful.
Proof. Let E be a finite symmetric set of G with e ∈ E. Let E ′ = E (2) = {x.y | x, y ∈
E} ⊂ G and EA ⊂ A (respectively, EB ⊂ B) be the set of letters in A (respectively,
B) which appears in the normal form of the elements of E ′ . Since E is supposed to
be finite, E ′ , EA and EB are finite. In the other hand A and B are supposed to be
residually finite so there exists N0 such that EA (respectively, EB ) is sent bijectively
into An (respectively, B n ) for any n ≥ N0 , in particular E ′ is sent bijectively into Gn
for n ≥ N0 .
If we denote D the commutator subgroup of G then D ∩ E ′ 6= 1 implies Dn ∩ E ′ 6=
1 (where E ′ = πn (E) and πn : G → Gn is the quotient map) for any n ≥ N0 .
Moreover, dB (e, D ∩ E ′ ) ≥ dBn (e, Dn ∩ E ′ ) where B (respectively, Bn ) denote the basis
of Proposition 7 and d(∗) the word distance associated. Since the sequence (M(n,k) )k
is such that M(n,k+1) is proper and characteristic inside of M(n,k) Proposition 3.2
of [14] ensures that dBn (M(n,k) , e) ≥ k in Dn . So we have that for any k ≥ n ≥
max{N0 , dB (e, D ∩ E ′ ) + 1} each elements of E ′ are sent non-trivially into the finite
quotients Gn /M(n,k) .
Assume there exist x, y ∈ E, x 6= y with same image into Gn /M(n,k) , then x.y −1 ∈ E ′
with x.y −1 6= e is sent trivially into Gn /M(n,k) which is a contradiction. So E is sent
bijectively into Gn /M(n,k) for any k ≥ n ≥ max{N0 , dB (e, D ∩ E ′ ) + 1}.
In order to prove Theorem 1 with a sequence of the above type it seems that
we must find suitable (An )n , (B n )n and (M(n,k) )k ’s. As it is explained in the next
subsection, if we assume A and B amenable the choice of (An )n and (B n )n is not
matter.
4.1
Geometry of metric spaces and gluing techniques
In this subsection we develop gluing techniques that allow us, under the assumption that A and B are amenable, to reduce the uniform embedding of the sequence
(Gn /M(n,k) )n,k≥n introduced above to that of the large girth sequence of subgroups
(Dn /M(n,k) )n,k≥n endowed with the restricted metric structure.
Let us first introduce a variant of property (A) that fits better with our purpose:
Definition 11. A sequence X = (Xn )n of metric spaces is equi-exact or uniformly
exact if for any R > 0, ε > 0 and n ≥ 0, there exists a partition of unity (ϕ(n,i) )i∈In
of Xn subordinated to a cover Un = (U(n,i) )i∈In , i.e. {ϕ(n,i) 6= 0} ⊂ U(n,i) , such that:
17
(1) for any n ≥ 1 and x, y ∈ Xn , dXn (x, y) ≤ R implies
ε.
P
i∈In
|ϕ(n,i) (x) −ϕ(n,i) (y)| ≤
(2) the cover U = {Un }n of X is uniformly bounded, i.e there exists C > 0 such that
diam(U(n,i) ) ≤ C for any n ≥ 0 and i ∈ In .
Lemma 4. Let G be a discrete residually finite group, ℓ a proper length on G and
N = (Nn )n a sequence of finite index normal subgroups with trivial intersection. If
G is amenable, then the sequence of metric spaces X = (Xn )n , where Xn = G/Nn is
endowed with the quotient metric induced by ℓ, forms a equi-exact sequence.
Proof. Since G is amenable by proposition 11.39 [16] N G has property (A). Let
R > 0, ε > 0 be given. Let ξ = (ξi )i∈I : N G → ℓ2 (I) with M > 0 as in Definition 2.
Define ϕ(n,i) : Xn → R+ by ϕ(n,i) = (ξi |Xn )2 and U(n,i) = {ϕ(n,i) 6= 0}. Since
supp(ξ(x)) ⊂ B(x, M), U(n,i) ⊂ B(x,
P M) and then {{U(n,i) }i∈In }n forms an uniformly
bounded cover of X . Moreover i∈In ϕ(n,i) (x) = kξ(x)k2 = 1 and for any x, y ∈ Xn
such that dXn (x, y) = dN G (x, y) ≤ R:
X
|ϕ(n,i) (x) − ϕ(n,i) (y)| ≤ kξ(x) − ξ(y)k ≤ ε.
i∈In
Proposition 9. Let X = (Xn )n be a sequence of metric spaces. Assume that for any
R > 0, ε > 0 and n ≥ 1, there exists a partition of unity (ϕ(n,i) )i∈In of Xn such that:
P
(1) for any x, y ∈ Xn , dXn (x, y) ≤ R implies i∈In |ϕ(n,i) (x) − ϕ(n,i) (y)| ≤ ε.
(2) the cover U = {U(n,i) }n,i∈In with U(n,i) = {ϕ(n,i) 6= 0} of X is uniformly embeddable into a Hilbert space.
Then X is uniformly embeddable.
Proof. We prove that X verifies the criterion (3) of Lemma 2.
Let’s fix R > 0, ε > 0 and consider (ϕ(n,i) )i∈In as above. Since U is uniformly
embeddable, U(R) = {U(n,i) (R)}n,i∈In is also embeddable (these two families are
uniformly quasi-isometric). So we can find ξ(n,i) : U(n,i) (R) → H(n,i) as in Lemma 2
characterization (3) for L
the constants R > 0 and ε > 0.
Let’s define ϕn : Xn → i∈In H(n,i) by the formula ϕn (x) = (ϕ(n,i) (x)1/2 ξ(n,i) (x))i∈In .
It is immediate from the construction that kϕn (x)k = 1 for any x ∈ Xn .
Let x, y ∈ Xn with dXn (x, y) ≤ R, there exists i ∈ In such that x ∈ Uin and so
x, y ∈ U(n,i) (R). In particular kξin (x) − ξin (y)k ≤ ε. It follows:
X
kϕ(n,i) (x)1/2 (ξ(n,i) (x) − ξ(n,i) (y))+
kϕn (x) − ϕn (y)k = [
i∈In
− ϕ(n,i) (y)1/2 )ξ(n,i) (y)k2]1/2 ≤
X
X
|ϕ(n,i) (x)1/2 − ϕ(n,i) (y)1/2 |2 kξ(n,i) (y)k2]1/2 ≤
ϕ(n,i) (x).kξ(n,i) (x) − ξ(n,i) (y)k2]1/2 + [
[
(ϕ(n,i) (x)
1/2
i∈In
i∈In
X
X
|ϕ(n,i) (x) − ϕ(n,i) (y)|]1/2 ≤ ε + ε1/2 .
|ϕ(n,i) (x)1/2 − ϕ(n,i) (y)1/2 |2 ]1/2 ≤ ε + [
ε+[
i∈In
i∈In
18
In other words, for any x, y ∈ Xn , dXn (x, y) ≤ R implies kϕn (x) − ϕn (y)k ≤ ε′ =
ε+ε1/2 . The second part to verify is a simple consequence of the following inequalities:
X
ϕni (x)1/2 ϕni (y)1/2 |hξin (x)|ξin (y)i| ≤
|hϕn (x)|ϕn (y)i| ≤
sup
n,i∈In
i∈In
n
sup{|hξi (x)|ξin (y)i|}
| d(x, y) ≥ M, x, y ∈ Xn }
Proposition 10. Let X = (Xn )n be a sequence of metric spaces. Suppose for any
R > 0, ε > 0 and n ∈ N, there exists a partition of the unity (ϕ(n,i) )i∈In of Xn such
that:
P
(1) for any x, y ∈ Xn , dXn (x, y) ≤ R implies i∈In |ϕ(n,i) (x) − ϕ(n,i) (y)| ≤ ε.
(2) the family U = {U(n,i) }n,i∈In with U(n,i) = {ϕ(n,i) 6= 0} is equi-exact.
Then X is equi-exact.
Proof. Let R > 0, ε > 0 be given. Let (ϕ(n,i) )n,i∈In be the maps subordinated to
U = {U(n,i) }n,i∈In as in the statement.
Since U is equi-exact so is U(R) = {U(n,i) (R)}n,i∈In . In particular we can find ψ(n,i,j) :
Xn → R+ subordinated to V(n,i,j) ⊂PU(n,i) (R) with j ∈ J(n,i) such that for any
x, y ∈ U(n,i) (R) with dXn (x, y) ≤ R, j∈J(n,i) |ψ(n,i,j)(x) − ψ(n,i,j) (y)| ≤ ε and V =
{V(n,i,j) }n,i∈In,j∈J(n,i) is uniformly bounded.
Let’s define the partition of the unity of Xn , η(n,i,j) = ϕ(n,i) ψ(n,i,j) subordinated to V.
Moreover,
X
X
X
|ψ(n,i,j)(x) − ψ(n,i,j) (y)|
ϕ(n,i) (x)
|η(n,i,j)(x) − η(n,i,j) (y)| ≤
i∈In ,j∈J(n,i)
+
≤
X
i∈In
X
i∈In
ϕ(n,i) (x)
X
i∈In
|ϕ(n,i) (x) − ϕ(n,i) (y)|
j∈J(n,i)
j∈J(n,i)
X
ψ(n,i,j) (x)
j∈J(n,i)
|ψ(n,i,j)(x) − ψ(n,i,j)(y)| +
X
i∈In
|ϕ(n,i) (x) − ϕ(n,i) (y)|
Since ϕ(n,i) (x)
P 6= 0 implies that x ∈ U(n,i) and then y ∈ U(n,i) (R), by definition of
(ψ(n,i,j) )j∈J(n,i) , j∈J(n,i) |ψ(n,i,j)(x) − ψ(n,i,j) (y)| ≤ ε.
P
that:
Moreover for any x, y ∈ Xn with dXn (x, y) ≤ R we know P
i∈In |ϕ(n,i) (x) −
ϕ(n,i) (y)| ≤ ε. Combining these two estimations we have:
i∈In ,j∈J(n,i) |η(n,i,j) (x) −
η(n,i,j) (y)| ≤ 2ε as desired.
Corollary 3. Let X = (Xn )n and Y = (Yn )n be two sequences of metric spaces and
F = (Fn : Xn → Yn )n a sequence of contractive maps. Assume that Y is equi-exact
and for any uniformly bounded cover U = {{U(n,i) }i∈In }n of Y, i.e {U(n,i) }i∈In is a partition of Yn and supn,i∈In diam(U(n,i) ) < +∞, the cover F −1 (U) = {{Fn−1 (U(n,i) )}i }n
of X is uniformly embeddable (respectively, equi-exact). Then the family X is uniformly embeddable (respectively, equi-exact).
19
Proof. Let R > 0, ε > 0 be two positive constants. Since Y is an equi-exact sequence,
we can find an uniformly bounded cover U of Y and a subordinated partition of
unity {(ϕ(n,i) )i∈In }n which satisfies Definition 11 with respect to the constants R and
ε. Then {(ϕni ◦ Fn )i∈In }n is a partition of unity on the sequence X subordinated
to F −1(U) which is uniformly embeddable (respectively, equi-exact) by assumption.
Proposition 9 (respectively, Proposition 10) can be applied.
Corollary 4. Let X = (Xn )n and Y = (Yn )n be two sequences of metric spaces, G a
group which acts by isometries on each Xn and Yn for any n and F = (Fn : Xn → Yn )n
a sequence of G-equivariant contractions. Assume that Y is equi-exact and the actions
on Y are uniformly cobounded, i.e there exists C > 0 such that diam(Yn /G) ≤ C for
any n.
If for any n, there exists yn ∈ Yn such that for any R ≥ 0 the sequence of inverse
images (Fn−1 (B(yn , R)))n ⊂ X is uniformly embeddable (respectively, equi-exact) then
X is uniformly embeddable (respectively, equi-exact).
Proof. Let R > 0, ε > 0 be two positive constants. Since Y is an equi-exact sequence,
we can find an uniformly bounded cover U of Y and a subordinated partition of unity
{(ϕ(n,i) )i∈In }n which satisfies Definition 11 with respect to the constants R and ε. Let
yn ∈ Xn as in the statement. Since the actions of G are uniformly cobounded, we can
find a sequence of subspaces Bn ⊂ Yn such that yn ∈ Bn , every orbit of the G-action
on Yn intersects Bn and (Bn )n is a uniformly bounded sequence of metric spaces.
Let R0 = supi,n diam(Uin ) + supn diam(Bn ) < +∞, we claim that for any n ∈ N and
i ∈ In , there exists gin such that gin Uin ⊂ B(yn , R0 ). Indeed, let gin ∈ G be such that
gin Uin intersects Bn and z ∈ sni Uin ∩ Bn , then for any y ∈ Uin we have :
d(gin y, yn) ≤ d(gin y, z) + d(z, yn ) ≤ d(y, (gin)−1 z) + d(z, yn ) ≤ R0 .
It follows gin Fn−1 (Uin ) = Fn−1 (gin Uin ) ⊂ Fn−1 (B(yn , R0 )) which is uniformly embeddable
in n (respectively, equi-exact). So Corollary 3 can be applied
Now we can formulate our reduction result discussed at the beginning of this subsection:
Let A, B be two discrete residually finite groups and (An )n , (B n )n two faithful sequence of finite quotients of respectively A and B.
Let’s denote G = A ⋆ B, Gn = An ⋆ B n their free products and D E G, Dn E Gn the
free commutator subgroups associated.
Let (M(n,k) )k be sequence of finite index characteristic subgroups of Dn with trivial
intersection (we need a bit less assumptions here because we do not require faithfulness of the sequence of quotients). Assume A and B are amenable and fix a proper
length function ℓ on G.
Theorem 2. If the sequence of metric spaces (Dn /M(n,k) )n,k≥n , where each Dn /M(n,k)
is endowed with the restriction of the quotient length on Gn /M(n,k) induced by ℓ,
embeds uniformly into a Hilbert space, then (Gn /M(n,k) )n,k≥n embeds uniformly.
20
Proof. Any free product Γ = Γ1 ⋆ Γ2 is a subgroup of (Γ1 × Γ2 ) ⋆ Z/2. Indeed, let u be
the generator of Z/2, then the homomorphism f : Γ → (Γ1 × Γ2 ) ⋆ Z/2 which sends
g1 ∈ Γ1 7→ g1 and g2 ∈ Γ2 7→ u.g2 .u is a realization.
Moreover, if we assume Γ1 and Γ2 amenable, stability of amenability by direct product
ensures that Γ1 × Γ2 is amenable [6] and heredity on box space uniform embedding
(cf Proposition 6) allows us to assume B = Z/2.
Let Y = (Yn )n with Yn = Gn /Dn = An × Z/2. We have the following exact
sequence:
1 → Dn /M(n,k) → Gn /M(n,k) → An × Z/2 → 1.
G acts on both Xn and Yn transitively which ensures the action to be uniformly
cobounded and the projections πn : Xn → Yn are G-equivariant contractions. Since
A×Z/2 is amenable, Lemma 4 ensures that the sequence Y is equi-exact with respect
to any proper length function on A × Z/2.
Moreover Dn /M(n,k) is R-dense in πn−1 (BYn (e, R)), so the uniform embeddability of
U(R) = (πn−1 (B(e, R)))n follows by our assumption on (Dn /M(n,k) )n,k≥n . Corollary 4
can apply.
4.2
Coarse metric approximation and free product
The purpose of this subsection is the conditions under which the sequence of metric
groups (Dn /M(n,k) )n,k≥n, discussed previously, endowed with a free product metric (cf.
Theorem 2 statement) embeds uniformly into a Hilbert space. It appears that even
if the sequences X = ((Dn /M(n,k) , ℓ))n,k≥n , where ℓ is the quotient metric induces
by a fixed length on G, and XComb = ((Dn /M(n,k) , ℓComb,B )n,k≥n , where ℓComb,B is the
quotient metric induced by the word metric on B of D, are not coarsely equivalent
(cf. remark 1). We can locally approximate X by XComb , in a particular sense, and
be able to conclude on the uniform embedding of X from the uniform embedding of
XComb .
We first need an analogue of the stability by direct limit of the uniform embedding
of box spaces.
Let (G, ℓ) be a discrete group endowed with a proper length ℓ.
Recall that a sequence of subgroups (Hτ )τ of G converges to (G, ℓ) if for all R > 0,
there exists T > 0 such that B(e, R) ⊂ Hτ for all τ ≥ T .
Given a sequence of normal subgroups N = (Nn )n on (G, ℓ), we can associate to
(Hτ )τ a sequence of metric subspaces of X = ⊔N ∈N G/N by considering Xτ =
⊔N ∈N πN (Hτ ) ⊂ X where πN : G → G/N is the quotient map.
This sequence play a role analogue as in the group setting. In particular uniform
embedding is stable by direct limit through these type of sequences as the following
proposition shows.
Proposition 11. Assume (G, ℓ) has a sequence of normal subgroups N = (Nn )n such
that for all τ , the metric space Xτ = ⊔N ∈N πN (Hτ ) uniformly embeds into Hilbert
space.
Then the metric space X = ⊔N ∈N G/N coarsely embeds into Hilbert space.
21
Proof. Since Xτ embeds uniformly into Hilbert space, for all ε > 0 we can find
ϕτn : Hnτ → Hnτ , with Hnτ a Hilbert space, such that kϕn (x)k = 1 and:
(i) for all n ≥ 1 and x, y ∈ Hnτ with dn (x, y) ≤ R, kϕτn (x) − ϕτn (y)k ≤ ε.
(ii) for all ε′ > 0, there exists M > 0 such that |hϕτn (x)|ϕτn (y)i| ≤ ε′ whenever
dn (x, y) ≥ M.
with dn the left-invariant distance associated to ℓn .
Let’s fix σnτ : Qτn → Gn , with Qτn = Gn /Hnτ , a section for all n and τ . Let R > 0 be a
positive constant and T such that Bn (e, R) ⊂ Hnτ for all n and τ ≥ T .
First, observe that, for all x, y ∈ Γn , x ∼ y if and only if σnτ (x) = σnτ (y) and in such
case:
dn (x, y) = dn (σnτ (x)x′ , σnτ (y)y ′) = dn (x′ , y ′ )
with x′ = σnτ (x)−1 x ∈ Hnτ (respectively, y ′ = σnτ (y)−1 y ∈ Hnτ ).
cτn : Γn → Hnτ ⊗ ℓ2 (Qτn ) by the formula ϕ
cτn (x) =
Now, we define the induced maps ϕ
ϕτn (x′ ) ⊗ δx with x′ as above. A direct computation shows that these maps verify :
τ
′
τ
′
τ
τ
cn (x) − ϕ
cn (y)k = kϕn (x ) − ϕn (y )k if x ∼ y
kϕ
2
otherwise
for x, y ∈ Gn .
But if dn (x, y) ≤ R, by our assumption this imply x−1 y ∈ Bn (e, R) ⊂ Hnτ and
consequently x ∼ y. So according to the above observation dn (x, y) = dn (x′ , y ′) ≤ R
cτn (x) − ϕ
cτn (y)k = kϕτn (x′ ) − ϕτn (y ′ )k ≤ ε.
and then we have: kϕ
In a similar way:
τ
′
τ
′
cτn (x)|ϕ
cτn (y)i| = |hϕn (x )|ϕn (y )i| if x ∼ y
|hϕ
(9)
0
otherwise
Let ε′ > 0 be a positive constant and M > 0 such that, |hϕτn (xτ )|ϕτn (yτ )i| ≤
ε′ whenever dn (xτ , yτ ) ≥ M with xτ , yτ ∈ Hnτ . According to the formula (9),
cτn (x)|ϕ
cτn (y)i| ≤ ε′ whenever dn (x, y) ≥ M. Indeed, if x ≁ y there is nothing
|hϕ
to do, otherwise, using the observation we have that dn (x′ , y ′) = dn (x, y) ≥ M and
cτn (x)|ϕ
cτn (y)i| = |hϕτn (x′ )|ϕτn (y ′)i| ≤ ε′ .
|hϕ
This result combined with Lemma 2 prove the Proposition 11.
Let’s get back to free product. As usual A and B are two discrete finitely generated
groups, G = A ⋆ B denote their free product and D E G is the commutator subgroup
of G.
Given two sequences of quotients of A and B denoted respectively (An )n and (B n )n ,
as before, we define a quotient sequence of G by taking Gn = An ⋆ B n and denote
πn : G → Gn the quotient maps associated.
We define Dτ , with τ > 0, the free subgroup of D with basis Bτ = {[a, b] | ℓ(a), ℓ(b) ≤
τ ; a, b 6= 1} ⊂ D (cf. Proposition 7). Hτ acts isometrically by left translation on
(πn (Hτ ), ℓ) where ℓ is the quotient length associated to a proper length ℓ on G.
22
Lemma 5. Let ℓ be a proper length function on G.
We endow Dτ with the word metric, ℓComb,Bτ , associated to the basis Bτ . For all
τ > 0, there exist a rank N ≥ 0 and a pair of constants (C, λ) which depends only on
τ such that the quotient maps πn : (Hτ , ℓComb,Bτ ) → (πn (Hτ ), ℓ) define G-equivariant
(C, λ)-quasi-isometries for all n ≥ N.
Proof. Any proper metric on a discrete countable group are equivalent, we may assume ℓ is a word metric associated to the union of two finite generating sets of
respectively A and B.
Since the sequence of quotients (Gn )n is faithful, there exists a rank N ≥ 0 such that
(2)
Bτ = {x.y −1 | x, y ∈ Bτ } is sent bijectively inside of Dn for n ≥ N. Because of
the freeness of the set πn (B) ⊂ Gn we have that ℓComb,πn (B) (g) = ℓComb,Bτ (g) for all
g ∈ Dτ and n ≥ N where ℓComb,πn (B) is the word distance associated to πn (B) ⊂ Gn
and g = πn (g).
Let Tn be the Bass-Serre tree of Gn and dn its graph metric. Then dn (g.An , An ) ≤ ℓ(g)
for all g ∈ Gn and by the inequality (8) of Proposition 7 we have:
ℓComb,πn(B) (g) = ℓComb,Bτ (g) ≤ dn (g.An , An ) + 1 ≤ ℓ(g) + 1
for all g ∈ Dτ .
Now, let g ∈ Dτ , then g can be written g = si1 . . . sin with sik = [aik , bik ] ∈ Bτ
and n = ℓComb,Bτ (g) in an unique way, so :
ℓ(g) ≤
n
X
k=1
ℓ(sik ) ≤
n
X
k=1
2ℓ(aik ) + 2ℓ(bik ) ≤ 4τ n = 4τ ℓComb,Bτ (g).
We just proved that:
ℓComb,Bτ (g) − 1 ≤ ℓ(g) ≤ 4τ ℓComb,Bτ (g)
for any g ∈ Dτ , with quasi-isometry constants which depend only on τ .
Remark 1. It is interesting to note that the dilatation on all D, ρ+ (t) = supℓComb(g)≤t ℓ(g),
depends on A and B. Indeed, a simple computation shows that diam(A) + diam(B) ≤
ρ+ (1). This justifies what we discussed at the beginning of this subsection and our
approach.
Corollary 5. Let Xτ be the coarse disjoint union of the sequence of metric spaces
((πn (Dτ )/M(n,k) , ℓ))n,k≥n where ℓ is the quotient metric induces by a fixed proper length
ℓ on G. Then Xτ is quasi-isometric to coarse disjoint union Xτ,Comb of the sequence
((πn (Dτ )/M(n,k) , ℓComb,Bτ ))n,k≥n where ℓComb,Bτ is the quotient length associated to the
word distance on Bτ .
In particular the uniform embedding of Xτ reduces to the uniform embedding of
Xτ,Comb .
Proof. This is a direct consequence of Lemma 5 and Proposition 5.
23
Since the commutator subgroups Dn E Gn = An ⋆ B n are finite rank free groups
and , for n large enough, the restriction of the quotient length associated to the word
distance on B of D, ℓComb,B , to Dτ is nothing but the word distance on the basis
Bτ . The following proposition ensures the existence of a suitable sequence of finite
index normal subgroups (M(n,k) )(n,k) such that the Xτ,Comb uniformly embed. This
concludes the proof of Theorem 1.
Proposition 12. For all free group Fω of finite rank ω ∈ N endowed with a word
metric on a basis, there exists a sequence of finite index characteristic subgroups of Fω
with trivial intersection Nω = (N(ω,k) )k with N(ω,k+1) characteristic into N(ω,k) such
that the sequence (Fω /N(ω,k) )ω,k≥ω embeds uniformly into Hilbert space.
Indeed, for all n, there exists Nn = (Nn,k )k such that XComb = ⊔n,k≥n Dn /Nn,k
embeds into Hilbert space. Since Xτ,Comb ⊂ XComb for all τ > 0, the Proposition 11
concludes the proof of Theorem 1.
Let us prove the Proposition 12:
Proof. Let us recall that a characteristic subgroup N of a group G is a normal subgroup stable by all automorphisms of G.
Let Fω be a free group of finite rank ω ∈ N, following [2] idea on the existence of a
free group box space which coarsely embeds, we consider the sequence of characteristic subgroups (Nk )k of Fω , define recursively by the formula N1 = h{w 2; w ∈ Fω }i,
Nk+1 = h{w 2; w ∈ Nk+1 }i E Nk . Since Nk+1 is a finite index subgroup of Nk as far as
the rank of Nk is finite, by induction we deduce that these groups has finite index in
Fω . Moreover Levi theorem (cf [14]) justifies that this sequence has trivial intersection.
The AGS [2] embedding result rests on a construction of a pseudo-metric dW,k on
Xkω = Fω /Nk associated to a wall structure which agrees with the combinatorial metric dωk at the scale of girth(Xkω ). Note that such a pseudo-metric always corresponds
to a Hilbert metric [16].
In details, on any finite quotient Xkω , dωW,k ≤ dω and :
dω (x, y) ≤ girth(Xkω ) if and only if dωW,k (x, y) ≤ girth(Xkω )
for any x, y ∈ Xkω . Furthermore, if one of the above inequalities hold, then dω (x, y) =
dωW,k (x, y) [2].
So to prove our proposition a sufficient condition is to prove that girth(Xkω ) grows
uniformly on the rank ω, i.e there exists a proper function f : N → N such that
girth(Xkω ) ≥ f (k) for all ω ≥ 1 and k ≥ ω. Indeed, take in account the relations
between dωW and dω described above, we just need to prove the existence of a proper
function ρ− : R+ → R+ such that:
ρ− (dω (x, y)) ≤ dωW (x, y)
(10)
for any ω, k ≥ ω and x, y ∈ Xkω .
We must prove that ρ(t) = inf {(x,y);dω (x,y)≥t} dωW,k (x, y) is proper, i.e for any R > 0,
24
there exists R′ > 0 such that for any ω, dωW,k (x, y) ≤ R implies dω (x, y) ≤ R′ .
If such a function f exists. Let K0 such that f (k) > R for any k ≥ K0 and R′ ≥
maxω≤K0 diam(⊔K0 ≥k≥ω Xkω ) + R. We have that for any x, y ∈ Xkω with k ≥ ω, if
k ≤ K0 by construction d(x, y) ≤ maxω≤K0 diam(⊔K0 ≥k≥ω Xkω ) ≤ R′ otherwise k ≥ K0
and girth(Xkω ) ≥ f (k) ≥ R so dW (x, y) = d(x, y) ≤ R ≤ R′ as required.
Now the existence of f follows from the fact that the sequence (N(ω,k) )k is characteristic and N(ω,k+1) contains no basis element of N(ω,k) . Indeed, automorphisms
act transitively on basis, so by contradiction this will implies that N(ω,k+1) = N(ω,k)
which is absurd because the sequence is defined recursively and has trivial intersection. Proposition 3.2 [14] ensures that relatively to any basis of Fω , ℓComb,ω (g) ≥ k
for any g ∈ Nkω and so Girth(Xkω ) ≥ k.
5
Further on finite quotients of free products
In this section the groups A and B we consider are supposed to be finitely generated.
We start this section by presenting another construction of finite quotients of a free
product due to Baumslag [4] that is extensively use here.
First, assume A and B are two finite groups and G = A ⋆ B their free product. Let
U(k) ⊂ G be the set of words with normal form: gl . . . g1 with l ≤ k. For k ≥ 2, U(k)
decomposes as follows:
U(k) = U(k − 1) ∪ (V (k) ∩ A.U(k − 1)) ∪ (V (k) ∩ B.U(k − 1))
with V (k) = U(k) \ U(k − 1) the set of words with normal form: gk . . . g1 .
The action by left multiplication of A on G stabilizes U(k − 1) ∪ (V (k) ∩ A.U(k − 1))
and can be extend to all U(k) by assuming that A acts trivially on V (k) ∩ B.U(k −
1). Something similar can be achieved with B by replacing V (k) ∩ A.U(k − 1) by
V (k) ∩ B.U(k − 1).
This gives two symmetric representations σA (k) : A → Sym(U(k)) and σB (k) :
B → Sym(U(k)) and by the universal property of free product we obtain: σ(k) =
σA (k) ⋆ σB (k) : G → Sym(U(k)). Because A and B are finite U(k) is also finite, in
particular Nk = ker(σ(k)) has finite index in G.
Remark 2. The sequence of finite index normal subgroups (Nk )k has trivial intersection. Indeed, let g = gn0 . . . g1 be the reduced form of g ∈ G with g 6= e. Then
σ(k)(g)(e) = g 6= e for any n ≥ n0 .
Now assume A and B are both residually finite groups. Let (An )n and (B n )n be
two faithful sequence of finite quotients of respectively A and B. Let’s denote Gn =
An ⋆ B n and σn (k) : Gn → Sym(Un (k)) the representation built previously applied
to the free product of finite groups Gn . Then the subgroups N(n,k) = ker(σn (k) ◦ πn )
of G where πn : G → Gn is the quotient map, form a sequence of finite index normal
subgroups with trivial intersection.
25
5.1
Free product box spaces and expansion property
Here we study expansion property of the sequence (G/N(n,k) )n,k≥n.
Let X = (V, E) be a finite graph and A ⊂ V a subset of vertices. The edge
boundary ∂A ⊂ E of A is the set of edges which relate an element A to one of
V \ A.
Definition 12. Given a finite graph X = (V, E), the Cheeger constant h(X) is
defined as minA⊂V ||A|≤ |V | |∂A|
.
|A|
2
A sequence of finite graphs (Xn )n is called expander if there exists c > 0 such
that h(Xn ) ≥ c for any n.
Let G be a group which acts on a set Z and S a generating set of G. We call
Schreir graph and denote it XS (Z) the graph defined by: V (XS (Z)) = Z and
(z, z ′ ) ∈ E(XS (Z)) if z ′ = s.z with s ∈ S.
Given a sequence of finite quotients (Qn )n of G, we construct a sequence of Schreir
graphs XS (Qn ). Expansion of (XS (Qn ))n does not depends on S and we say that
(Qn )n is expander if such a S exists.
Proposition 3 is a consequence of the following:
Proposition 13. Let A and B be two residually finite groups and G = A ⋆ B their
free product. The faithful sequence of finite quotients (Gn /Nn,k )(n,k≥n) of G is nonexpander.
In particular we cannot expect a free product with property (τ ) (cf [13]).
In order to prove Proposition we need a intermediate result.
Lemma 6. Let G be a finite group, S a symmetric generating set of G and H ≤ G
a subgroup of G. Then h(G, S) ≤ h(G/H, S).
Proof. Let X ⊂ G be a subset. Then |X.H| = |X|.|H| if and only if the multiplication
map X × H → G is bijective and this is equivalent to say that (X −1 .X) ∩ H = {e}.
Now let Y ⊂ G/H, there exists X ⊂ G which is sent bijectively to Y by the projection
π : G → G/H such that X −1 .X ∩ H = {e}. Indeed, let’s define Y (2) = {a−1 a′ | a, a′ ∈
Ỹ } with Ỹ ⊂ G a arbitrary set of representative of Y , there exists X0 ⊂ G which
is sent bijectively to Y (2) . In particular X0 .H ∩ X0 = {e} and the representative of
Y ⊂ A′ (2) in X0 do the job.
It’s enough to prove that for any Y ⊂ G/H with |Y | ≤ |G/H|
and a set of represen2
tative X ⊂ G as above:
|∂Y |
|∂X.H|
≤
.
(11)
|X.H|
|Y |
.|H| = |G|
.
Indeed, |X.H| = |X|.|H| ≤ |G/H|
2
2
Observe that for a edge e = {xh, sxh} ∈ E(X.H, G \ (X.H)) between a element of
X.H and G \ (X.H), the edges e.h′ = {xhh′ , sxhh′ } with h′ ∈ H are distincts and
26
still belong to E(X.H, G \ (X.H)). Furthermore they represent the fiber of the edge
e = {x, s.x} ∈ E(Y, (G/H) \ Y ). It follows that |∂X.H| = |∂Y |.|H|. Finally the
equality |X.H| = |X|.|H| implies inequality 11.
Lemma 6 reduces Proposition to the study of the expansion property of Schreir
graphs (XS (Un (k)))n,k≥n. Indeed, Un (k) ≃ G/N(n,k) /St(e) where St(e) is the stabilizer of e ∈ Un (k) .
Proposition 14. The sequence of finite homogenous graphs (XS (Un (k)))n,k≥n is nonexpander.
Proof. We take the Schreir graph structure associated to the finite generating S =
SA ∪ SB of G with SA and SB two finite generating sets of respectively A and B.
Let’s consider PnA (k) ⊂ Un (k) (respectively, PnB (k) ⊂ Un (k)) the set of words with
reduced form gk . . . g1 with k ≤ n and g1 ∈ An (respectively, g1 ∈ B n ). Given
A
w ∈ Un (k) with w 6= e, w belongs to PnA (k) or PnB (k). Assume w ∈ Un,k
, if we denote
A
its set of neighbors by V (w) = {s.w | s ∈ S}, then V (w) ⊂ Un,k . In particular
∂PnA (k) ⊂ V (e) and it follows that |∂PnA (k)| ≤ |S|.
Furthermore we have |PnA (k)| ≥ (|An | − 1)⌊n/2⌋ (|B n | − 1)⌊n/2⌋ which implies:
|∂PnA (k)|
|S|
≤
→0
A
⌊n/2⌋
|Pn (k)|
(|An | − 1)
(|B n | − 1)⌊n/2⌋
when n or k goes to infinity and similar holds for PnB (k). Since the sets PnA (k) and
PnB (k) verify |PnA (k)| + |PnB (k)| = |Un (k)| + 1, one of them has cardinality lower or
equal then |Un2(k)| , so h(Un (k)) → 0 whenever k or n go to infinity.
5.2
Free product symmetric representations and uniform embedding.
Let G be a group, we call irreducible symmetric representation of G an action of
type: G y G/N with N ≤ G a finite index subgroup of G. This denomination comes
from the fact that these actions give symmetric representations: G → Sym(G/N)
and any transitive action on a finite set are obtained this way.
A sequence of irreducible symmetric representations will be for us a sequence of finite
index subgroups, non-necessarily normal, in G. Moreover, we say that it is faithful if
these subgroups has trivial intersection.
Given such a representation N ≤ G and a generating set S of G. We denote the
Schreir graph XS (G/N) associated to N and S simply by XS (N).
Proposition 1. Let A and B be two finitely generated residually finite groups,
G = A ⋆ B their free product and S ⊂ G a finite generating set of G.
Assume A and B admit respectively an embeddable box space. Then there exists
a faithful sequence of irreducible symmetric representation (Nn )n of G such that
⊔n (XS (Nn ))n embeds into a Hilbert space.
27
By using irreducible symmetric representation we relaxed the assumption that the
subgroups we consider are normal but these notion still keep track of certain properties as properties (τ ) [13] and (A). More precisely we can still prove that property
(T) (respectively, amenability) implies property (τ ) (respectively, property (A)).
In a other hand we lose on the approximation properties correspondance:
Proposition 2. There exists a faithful sequence of irreducible symmetric representation (Nn )n of F2 , such that ⊔n F2 /Nn has property (A).
These two Proposition’s are actually consequences of the following theorem: Let
A and B be two discrete finitely generated residually finite groups. Let’s denote
(An )n and (B n )n two faithful sequences of finite quotients of respectively A and B
and Gn = An ⋆ B n their free products.
Theorem 3. If ⊔n An and ⊔n B n embed into Hilbert space (respectively, has property
(A)), then ⊔n Gn embeds into a Hilbert space (respectively, has property (A)).
Let S = SA ∪ SB ⊂ G be a finite generating set of G built by the union of a
generating set of A and B.
By considering Baumslag construction, then Un (k) ≃ G/St(e), with the Schreir graph
structure induced by S, embeds isometrically in (Gn , S) and is our needed sequence
of symmetric representation.
Proposition 15. Let T = {(T, dT )}T be the family of bounded geometry trees endowed
with their graph metric. Then T forms an equi-exact family of metric spaces.
Proof. Let’s fix xT ∈ T on any T ∈ T and denote (x|y)T = 21 [|x|T + |y|T − d(x, y)]
with |x|T = dT (x, xT ) with x, y ∈ T .
Recall that trees are 0-hyperbolic space and so satisfy the following hyperbolic condition:
(x|z)T ≥ min{(x|y)T , (y|z)T }
for any x, y, z ∈ T . In particular given any positive constant L ≥ 0, the relation
x ∼ y if (x|y)T ≥ L is an equivalence relation on T≥L = {x ∈ T | |x|T ≥ L}.
Let L > 0 be a positive constant and AT,L,k = {x ∈ T | kL ≤ |x|T ≤ k(L + 1)} an
annulus of T≥L . The relation on AT,L,k ⊂ T≥L : x ∼ y if (x|y)T ≥ L(k − 1/2) is an
equivalence relation on AT,L,k . Moreover, x ∼ y implies dT (x, y) ≤ 3L and x ≁ y
implies dT (x, y) ≥ L. This induces a partition AT,L,k = ∪i∈IT,L,k CT,L,k,i such that
dT (CT,L,k,i, CT,K,k′,i ) ≥ L whenever k 6= k ′ and diam(CT,K,k,i) ≤ 3L.
Using that decomposition we construct the covering U = U1 ∪U2 with U1 = {CT,L,2k,i |
k ≥ 0, i ∈ IT,L,2k } and U2 = {CT,L,2k+1,i | k ≥ 0, i ∈ IT,L,2k+1} of T with the property
that Ui forms a L-separated, i.e for any C, C ′ ∈ Ui , d(C, C ′) ≥ L, 3L-uniformly
bounded family with i = 1, 2.
Let W = W1 ∪ W2 with Wi = Ui (L/2) = {C(L/2) | C ∈ Ui } which is the covering
obtained by taking the L/2-neighborhood of each element of U. Since at most two
elements of U intersect a ball of radius L/2, at most two elements of W contain a
given point. Moreover any ball of radius L/2 are contained in an element of W. This
28
fact will be use later for estimation.
Let’s define φT,L : T → ℓ1 (∪k∈N IT,L,k ) by the formula
dT (x, T \ WT,L,k,i )
k ′ ∈N,j∈IT,L,k dT (x, T \ WT,L,k ′ ,j )
φT,L (x)(i) = P
for any x ∈ T and i ∈ IT,L,k .
Then for k ∈ N and i ∈ IT,L,k , φT,L (.)(i) is supported in WT,L,k,i which has diameter
lower or equal to 4L and so forms a uniformly bounded supported partition of the
unity. Moreover:
d(x, y)
k ′ ∈N,j∈IT,L,k dT (x, T \ WT,L,k ′ ,j )
|φT,L (x)(i) − φT,L (y)(i)| ≤ P
1
1
−P
′
′
k ′ ∈N,j∈IT,L,k dT (x, T \ WT,L,k ,j )
k ′ ∈N,j∈IT,L,k dT (y, T \ WT,L,k ,j )
X
2d(x, y) 2
≤
+
|dT (x, T \ WT,L,k′,j ) − dT (y, T \ WT,L,k′,j )|
L
L ′
+dT (y, T \ WT,L,k,i) P
k ∈N,j∈IT,L,k
≤
10
d(x, y)
L
It follows that:
kφT,L (x) − φT,L (y)k1 ≤
X
k∈N,i∈IT,L,k
≤
|φT,L (x)(i) − φT,L (y)(i)|
40
d(x, y).
L
This does not depend on the tree T we consider. By taking L > 0 large enough we
can check the Definition 11.
Proposition 16. Let (Xn )n be a sequence of metric spaces with an uniformly embeddable (respectively, equi-exact) cover U = {Un }n . Assume for any L > 0 and n ≥ 1,
there exists Zn ⊂ Xn a subspace of Xn such that (Zn )n is an embeddable sequence
and {U \ Zn | U ∈ Un } is L-separated in Xn for any n. Then (Xn )n is uniformly
embeddable (respectively, equi-exact).
Proof. Let L > 0 be a positive constant and Zn ⊂ Xn as in the statement. Then
W = W1 ∪ W2 with W1 = {Zn (L/2)} and W2 = {(U \ Zn )(L/2) | U ∈ Un } is a cover
such that at most two elements of W contain a same element of Xn and any ball of
radius L/2 is contained in a element of W. As in the proof of Proposition 15 we can
build a O(L−1 )-lipschitz partition of unity subordinated to W and apply Proposition
9 (respectively, Proposition 10) to conclude.
We present a brief introduction to tree of metric spaces formalism (cf. [9]) that
we use to prove Theorem 3.
Let T = (V, E) be an oriented (this is for relieve on notations) connected tree.
29
A tree of metric spaces on T is given by a family of metric spaces X = {(Xv , dv )}v∈V
and a map f : E → ∪v,v′ |(v,v′ )∈E Xv × Xv′ such that f (e) ∈ Xs(e) × Xt(e) .
The total space X associated to (X , f ) is a metric space defined by X = ⊔v∈V Xv
and a metric dX obtained as follows:
Let’s consider the subset D = (∪v∈V Xv ×Xv )∪f (E) of X ×X. Then D defines a ample
domain, i.e for any pair of elements x, y ∈ X, there exists a path x1 = x, . . . , xn = y
such that (xi , xi+1 ) ∈ D for any 1 ≤ i ≤ n − 1. The metric dX is defined as the
unique enveloping metric of the partial metric db : D → R+ given by the formulas
b Xv ×Xv = dv and db ◦ f (e) = 1 for any e ∈ E (cf [9]).
d|
Now, given a free product G = A ⋆ B endowed with a proper left invariant metric d
and its Bass-Serre tree T = (G/A ∪ G/B, G). We associate a tree of metric spaces
by taking XgA = (gA, d|gA) and XgB = (gB, d|gB ) and f : G → G/A × G/B with
f (g) = (gA, gB).
G acts on the total spaces X(G,d) by the formula g ′.(ga) = gg ′a ∈ Xg′ gA with g ′ ∈ G
and ga ∈ XgA and similarly on G/B. Moreover this action is isometric, free and has
two orbits. The formula for the distance on X(G,d) given in Proposition (5.5) of [9]
implies the existence of a pair of positive constants (C, λ) such that any free product
(G, d) of finitely generated groups endowed with a proper left invariant metric d is
(C, λ)-quasi-isometric to X(G,d) . In other words it is equivalent to prove Theorem 2
for (Gn )n or (X(Gn ,dn ) )n (cf. Proposition 5).
We can prove Theorem 3.
Proof. Let Gn as in the statement, Tn its Bass-Serre tree and Xn its tree of metric
spaces. The sequence (Gn )n are uniformly quasi-isometric to X = (Xn )n , so the proof
reduce to show that X embeds uniformly (respectively, has (A)) if (An )n and (B n )n
embed (respectively, has (A)).
The actions of Gn on Tn are uniformly cobounded and Proposition 15 ensures that
T = (Tn )n forms an equi-exact sequence. Moreover pn : Xn → Tn is a sequence
of Gn -equivariant contractions (this is justified by Proposition 5.5 of [9]). To apply
Corollary 4, we must prove that p−1
n (B(An , r)) ⊂ Xn (here An ∈ V (Tn )) embeds
uniformly into a Hilbert space (respectively, has (A)) for any r ≥ 0.
Let’s do this by induction on r. For r = 0, p−1
n (An ) = XAn which is isometric to
An for any n and is an embeddable sequence (respectively, equi-exact sequence) by
assumption.
Assume now r ≥ 1, if we denote Cn (r) = p−1
n (B(An , r)), we have the following
decomposition:
[
Cn (r) = Cn (r − 1) ∪
Xv .
v∈Tn |dTn (An ,v)=r
Let L > 0 be positive constant and Yv ⊂ Xv be the subspace of Xv forms by elements
at distance at most L/2 from Cn (r − 1) for v ∈ Tn . Then Cn′ (r − 1) = Cn (r − 1) ∪
S
v∈Tn |dTn (An ,v)=r Yv is quasi-isometric to Cn (r − 1) and by induction hypothesis forms
an embeddable (respectively, equi-exact) sequence.
S
In the other hand Cn (r) \ Cn′ (r − 1) = v∈Tn |dTn (An ,v)=r Xv \ Yv is L-separated in
Cn (r) and Proposition 16 implies that (Cn (r))n embeds uniformly (respectively, is
30
equi-exact). We can apply Corollary 4 to conclude.
References
[1] S. Arnt. Fibred coarse embeddability of box spaces and proper isometric affine
actions on Lp spaces. Bull. Belg. Math. Soc. Simon Stevin, 23(1):21–32, 2016.
[2] Goulnara Arzhantseva, Erik Guentner, and Ján Spakula. Coarse non-amenability
and coarse embeddings. Geom. Funct. Anal., 22(1):22–36, 2012.
[3] Goulnara Arzhantseva and Romain Tessera. Admitting a coarse embedding is
not preserved under group extensions. 05 2016.
[4] Benjamin Baumslag and Marvin Tretkoff. Residually finite HNN extensions.
Comm. Algebra, 6(2):179–194, 1978.
[5] Bachir Bekka, Pierre de la Harpe, and Alain Valette. Kazhdan’s property (T),
volume 11 of New Mathematical Monographs. Cambridge University Press, Cambridge, 2008.
[6] Nathanial P. Brown and Narutaka Ozawa. C ∗ -algebras and finite-dimensional
approximations, volume 88 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2008.
[7] Xiaoman Chen, Qin Wang, and Xianjin Wang. Characterization of the Haagerup
property by fibred coarse embedding into Hilbert space. Bull. Lond. Math. Soc.,
45(5):1091–1099, 2013.
[8] Pierre-Alain Cherix, Michael Cowling, Paul Jolissaint, Pierre Julg, and Alain
Valette. Groups with the Haagerup property, volume 197 of Progress in Mathematics. Birkhäuser Verlag, Basel, 2001. Gromov’s a-T-menability.
[9] Marius Dadarlat and Erik Guentner. Constructions preserving Hilbert space
uniform embeddability of discrete groups. Trans. Amer. Math. Soc., 355(8):3253–
3275, 2003.
[10] Marius Dadarlat and Erik Guentner. Uniform embeddability of relatively hyperbolic groups. J. Reine Angew. Math., 612:1–15, 2007.
[11] Benson Farb and Dan Margalit. A primer on mapping class groups, volume 49
of Princeton Mathematical Series. Princeton University Press, Princeton, NJ,
2012.
[12] M. Gromov. Asymptotic invariants of infinite groups. In Geometric group theory,
Vol. 2 (Sussex, 1991), volume 182 of London Math. Soc. Lecture Note Ser., pages
1–295. Cambridge Univ. Press, Cambridge, 1993.
31
[13] Alex Lubotzky. What is. . . property (τ )? Notices Amer. Math. Soc., 52(6):626–
627, 2005.
[14] Roger C. Lyndon and Paul E. Schupp. Combinatorial group theory. Classics in
Mathematics. Springer-Verlag, Berlin, 2001. Reprint of the 1977 edition.
[15] G. A. Margulis. Explicit constructions of expanders. Problemy Peredači Informacii, 9(4):71–80, 1973.
[16] John Roe. Lectures on coarse geometry, volume 31 of University Lecture Series.
American Mathematical Society, Providence, RI, 2003.
[17] John Roe. Hyperbolic groups have finite asymptotic dimension. Proc. Amer.
Math. Soc., 133(9):2489–2490, 2005.
[18] Jean-Pierre Serre. Trees. Springer-Verlag, Berlin-New York, 1980. Translated
from the French by John Stillwell.
[19] G. Skandalis, J. L. Tu, and G. Yu. The coarse Baum-Connes conjecture and
groupoids. Topology, 41(4):807–834, 2002.
[20] Rufus Willett and Guoliang Yu. Higher index theory for certain expanders and
Gromov monster groups, I. Adv. Math., 229(3):1380–1416, 2012.
[21] Guoliang Yu. The coarse Baum-Connes conjecture for spaces which admit a
uniform embedding into Hilbert space. Invent. Math., 139(1):201–240, 2000.
32
| 4math.GR
|
Robust Estimators in High Dimensions
without the Computational Intractability
arXiv:1604.06443v1 [cs.DS] 21 Apr 2016
Ilias Diakonikolas∗
CS, USC
[email protected]
Jerry Li §
EECS, MIT
[email protected]
Gautam Kamath†
EECS, MIT
[email protected]
Ankur Moitra¶
Math & CSAIL, MIT
[email protected]
Daniel Kane‡
CSE & Math, UCSD
[email protected]
Alistair Stewartk
CS, USC
[email protected]
April 22, 2016
Abstract
We study high-dimensional distribution learning in an agnostic setting where an adversary is allowed to arbitrarily corrupt an ε-fraction of the samples. Such questions have a rich history spanning
statistics, machine learning and theoretical computer science. Even in the most basic settings, the only
known approaches are either computationally inefficient or lose dimension-dependent factors in their
error guarantees. This raises the following question: Is high-dimensional agnostic distribution learning
even possible, algorithmically?
In this work, we obtain the first computationally efficient algorithms with dimension-independent
error guarantees for agnostically learning several fundamental classes of high-dimensional distributions:
(1) a single Gaussian, (2) a product distribution on the hypercube, (3) mixtures of two product distributions (under a natural balancedness condition), and (4) mixtures of spherical Gaussians. Our algorithms
achieve error that is independent of the dimension, and in many cases scales nearly-linearly with the
fraction of adversarially corrupted samples. Moreover, we develop a general recipe for detecting and
correcting corruptions in high-dimensions, that may be applicable to many other problems.
∗
Part of this work was performed while the author was at the University of Edinburgh. Research supported in part by EPSRC
grant EP/L021749/1, a Marie Curie Career Integration Grant, and a SICSA grant.
†
Supported by NSF Award CCF-0953960 (CAREER) and ONR grant N00014-12-1-0999. This work was done in part while
the author was an intern at Microsoft Research Cambridge and visiting the Simons Institute for the Theory of Computing.
‡
Part of this work was performed while visiting the University of Edinburgh.
§
Supposed by NSF CAREER Award CCF-1453261, a Google Faculty Research Award, and an NSF Fellowship. This work was
done in part while the author was an intern at Microsoft Research Cambridge.
¶
Supported by NSF CAREER Award CCF-1453261, a grant from the MIT NEC Corporation, and a Google Faculty Research
Award.
k
Part of this work was performed while the author was at the University of Edinburgh. Research supported in part by EPSRC
grant EP/L021749/1.
Contents
1 Introduction
1.1 Background . . . . . . . . .
1.2 Our Techniques . . . . . . .
1.3 Our Results . . . . . . . . .
1.4 Discussion and Related Work
1.5 Organization . . . . . . . . .
1
1
2
3
5
6
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2 Preliminaries
2.1 Types of Adversaries . . . . .
2.2 Basic Notation . . . . . . . .
2.3 Distributions of Interest . . . .
2.4 Agnostic Hypothesis Selection
2.5 Bounds on TV Distance . . . .
2.6 Miscellaneous Lemmata . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6
. 6
. 7
. 8
. 8
. 9
. 11
3 Some Natural Approaches, and Why They Fail
11
4 Agnostically Learning a Gaussian, via Convex Programming
4.1 The Set SN,ε . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Concentration Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2.1 Empirical Estimates of First and Second Moments . . . . . . . . . . . . . . . . .
4.2.2 Estimation Error in the Frobenius Norm . . . . . . . . . . . . . . . . . . . . . . .
4.2.3 Understanding the Fourth Moment Tensor . . . . . . . . . . . . . . . . . . . . . .
4.2.4 Concentration of the Fourth Moment Tensor . . . . . . . . . . . . . . . . . . . . .
4.3 Finding the Mean, Using a Separation Oracle . . . . . . . . . . . . . . . . . . . . . . . .
4.3.1 Naive Pruning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.2 The Separation Oracle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.3.3 The Full Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.4 Finding the Covariance, Using a Separation Oracle . . . . . . . . . . . . . . . . . . . . .
4.4.1 The Full Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.5 Learning an Arbitrary Gaussian Agnostically . . . . . . . . . . . . . . . . . . . . . . . .
4.5.1 From Unknown Mean, Unknown Covariance, to Zero Mean, Unknown Covariance
4.5.2 From Unknown Mean, Approximate Covariance, to Approximate Recovery . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
13
14
14
14
15
16
17
18
18
19
22
23
29
30
31
31
.
.
.
.
.
.
.
.
.
.
.
32
33
33
34
35
36
36
37
38
39
39
40
5 Agnostically Learning a Mixture of Spherical Gaussians, via Convex Programming
5.1 Notation and Norms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Concentration Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Naive Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5 Estimating the Covariance Using Convex Programming . . . . . . . . . . . . . . .
5.5.1 Properties of Our Convex Set . . . . . . . . . . . . . . . . . . . . . . . .
5.5.2 Separation Oracle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6 Spectral Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.7 Exhaustive Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.8 Applying the Tournament Lemma . . . . . . . . . . . . . . . . . . . . . . . . . .
5.8.1 Guessing the Mixing Weights and Variance . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
5.9
5.8.2 Pruning Our Hypotheses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Handling Unequal Variances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
6 Agnostically Learning Binary Product Distributions, via Filters
6.1 The Balanced Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.1.1 Algorithm F ILTER -BALANCED -P RODUCT: Proof of Proposition 6.8 .
6.1.2 Setup and Basic Structural Lemmas . . . . . . . . . . . . . . . . . .
6.1.3 The Case of Small Spectral Norm . . . . . . . . . . . . . . . . . . .
6.1.4 The Case of Large Spectral Norm . . . . . . . . . . . . . . . . . . .
6.2 Agnostically Learning Arbitrary Binary Product Distributions . . . . . . . .
6.2.1 Algorithm F ILTER -P RODUCT: Proof of Proposition 6.24 . . . . . . .
6.2.2 Chi-Squared Distance and Basic Reductions . . . . . . . . . . . . . .
6.2.3 Setup and Basic Structural Lemmas . . . . . . . . . . . . . . . . . .
6.2.4 The Case of Small Spectral Norm . . . . . . . . . . . . . . . . . . .
6.2.5 The Case of Large Spectral Norm . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
42
43
45
46
48
49
51
52
53
55
58
60
7 Agnostically Learning Mixtures of Two Balanced Binary Products, via Filters
62
7.1 The Full Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.2 Mixtures of Products Whose Means Differ Significantly in One Coordinate . . . . . . . . . 63
7.3 Mixtures of Products Whose Means Are Close in Every Coordinate . . . . . . . . . . . . . 71
8 Agnostically Learning a Gaussian, via Filters
8.1 Learning a Gaussian With Unknown Mean . . . . . . . . . . . . . . . . . . . . . .
8.1.1 Algorithm F ILTER -G AUSSIAN -U NKOWN -M EAN: Proof of Proposition 8.5
8.1.2 Proof of Correctness of F ILTER -G AUSSIAN -U NKOWN -M EAN . . . . . . .
8.2 Learning a Gaussian With Unknown Covariance . . . . . . . . . . . . . . . . . . .
8.2.1 Analysis of Filter-based Algorithm: Proof of Proposition 8.17 . . . . . . .
A Deferred Proofs from Section 2
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
81
81
83
83
86
88
99
B Deferred Proofs from Section 4
100
B.1 Proof of Theorem 4.17 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
C Deferred Proofs from Section 5
106
C.1 Proof of Theorem 5.12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
D Deferred Proofs from Section 6
111
E Deferred Proofs from Section 7
112
F Deferred Proofs from Section 8
113
1 Introduction
1.1 Background
A central goal of machine learning is to design efficient algorithms for fitting a model to a collection of
observations. In recent years, there has been considerable progress on a variety of problems in this domain,
including algorithms with provable guarantees for learning mixture models [FOS05, KMV10, MV10, BS10,
HK13], phylogenetic trees [CGG02, MR05], HMMs [AHK12], topic models [AGM12, AGHK13], and
independent component analysis [AGMS15]. These algorithms crucially rely on the assumption that the
observations were actually generated by a model in the family. However, this simplifying assumption is not
meant to be exactly true, and it is an important direction to explore what happens when it holds only in an
approximate sense. In this work, we study the following family of questions:
Question 1.1. Let D be a family of distributions on Rd . Suppose we are given samples generated from the
following process: First, m samples are drawn from some unknown distribution P in D. Then, an adversary
is allowed to arbitrarily corrupt an ε-fraction of the samples. Can we efficiently find a distribution P ′ in D
that is f (ε, d)-close, in total variation distance, to P ?
This is a natural formalization of the problem of designing robust and efficient algorithms for distribution
estimation. We refer to it as (proper) agnostic distribution learning and we refer to the samples as being
ε-corrupted. This family of problems has its roots in many fields, including statistics, machine learning, and
theoretical computer science. Within computational learning theory, it is related to the agnostic learning
model of Haussler [Hau92] and Kearns et al. [KSS94], where the goal is to learn a labeling function whose
agreement with some underlying target function is close to the best possible, among all functions in some
given class. In the even more challenging malicious noise model [Val85, KL93], an adversary is allowed to
corrupt both the labels and the samples. A major difference with our setting is that these models apply to
supervised learning problems, while here we will work in an unsupervised setting.
Within statistics and machine learning, inference problems like Question 1.1 are often termed “estimation under model misspecification”. The usual prescription is to use the maximum likelihood estimator
[Hub67, Whi82], which is unfortunately hard to compute in general. Even ignoring computational considerations, the maximum likelihood estimator is only guaranteed to converge to the distribution P ′ in D that is
closest (in Kullback-Leibler divergence) to the distribution from which the observations are generated. This
is problematic because such a distribution is not necessarily close to P at all.
A branch of statistics – called robust statistics [HR09, HRRS86] – aims to tackle questions like the
one above. The usual formalization is in terms of breakdown point, which (informally) is the fraction of
observations that an adversary would need to control to be able to completely corrupt an estimator. In lowdimensions, this leads to the prescription that one should use the empirical median instead of the empirical
mean to robustly estimate the mean of a distribution, and interquartile range for robust estimates of the
variance. In high-dimensions, the Tukey depth [Tuk75] is a high-dimensional analogue of the median that,
although provably robust, is hard to compute [JP78]. Similar hardness results have been shown [Ber06,
HM13] for essentially all known estimators in robust statistics.
Is high-dimensional agnostic distribution learning even possible, algorithmically? The difficulty is that
corruptions are often hard to detect in high dimensions, and could bias the natural estimator by dimensiondependent factors. In this work, we study agnostic distribution learning for a number of fundamental classes
of distributions: (1) a single Gaussian, (2) a product distribution on the hypercube, (3) mixtures of two
product distributions (under a natural balancedness condition), and (4) mixtures of k Gaussians with spherical covariances. Prior to our work, all known efficient algorithms (e.g. [LT15, BD15]) for these classes
required the error guarantee, f (ε, d), to depend polynomially in the dimension d. Hence, previous efficient
estimators could only tolerate at most a 1/ poly(d) fraction of errors. In this work, we obtain the first efficient algorithms for the aforementioned problems, where f (ε, d) is completely independent of d and depends
1
polynomially (often, nearly linearly) in the fraction ε of corrupted samples. Our work is just a first step in
this direction, and there are many exciting questions left to explore.
1.2 Our Techniques
All of our algorithms are based on a common recipe. The first question to address is the following: Even
if we were given a candidate hypothesis P ′ , how could we test if it is ε-close in total variation distance to
P ? The usual way to certify closeness is to exhibit a coupling between P and P ′ that marginally samples
from both distributions, where the samples produced from each agree with probability 1 − ε. However,
we have no control over the process by which samples are generated from P , in order to produce such
a coupling. And even then, the way that an adversary decides to corrupt samples can introduce complex
statistical dependencies.
We circumvent this issue by working with an appropriate notion of parameter distance, which we use
as a proxy for the total variation distance between two distributions in the class D. Various notions of
parameter distance underly several efficient algorithms for distribution learning in the following sense. If θ
and θ ′ are two sets of parameters that define distributions Pθ and Pθ′ in a given class D, a learning algorithm
often relies on establishing the following type of relation1 between dTV (Pθ , Pθ′ ) and the parameter distance
dp (θ, θ ′ ):
(1)
poly(dp (θ, θ ′ ), 1/d) ≤ dTV (Pθ , Pθ′ ) ≤ poly(dp (θ, θ ′ ), d) .
Unfortunately, in our agnostic setting, we cannot afford for (1) to depend on the dimension d at all. Any
such dependence would appear in the error guarantee of our algorithm. Instead, the starting point of our
algorithms is a notion of parameter distance that satisfies
poly(dp (θ, θ ′ )) ≤ dTV (Pθ , Pθ′ ) ≤ poly(dp (θ, θ ′ ))
(2)
which allows us to reformulate our goal of designing robust estimators, with distribution-independent error
guarantees, as the goal of robustly estimating θ according to dp . In several settings, the choice of the
parameter distance is rather straightforward. It is often the case that some variant of the ℓ2 -distance between
the parameters works2 .
Given our notion of parameter distance satisfying (2), our main ingredient is an efficient method for
robustly estimating the parameters. We provide two algorithmic approaches which are based on similar
principles. Our first approach is faster, requiring only approximate eigenvalue computations. Our second
approach relies on convex programming and achieves much better sample complexity, in some cases matching the information-theoretic limit. Notably, either approach can be used to give all of our concrete learning
applications with nearly identical error guarantees. In what follows, we specialize to the problem of robustly
learning the mean µ of a Gaussian whose covariance is promised to be the identity, which we will use to
illustrate how both approaches operate. We emphasize that what is needed to learn the parameters in more
general settings requires many additional ideas.
Our first algorithmic approach is an iterative greedy method that, in each iteration, filters out some of
the corrupted samples. Given a set of samples S ′ that contains a set S of uncorrupted samples, an iteration
of our algorithm either returns the sample mean of S ′ or finds a filterPthat allows us to efficiently compute
a set S ′′ ⊂ S ′ that is much closer to S. Note the sample mean µ
b= N
i=1 (1/N )Xi (even after we remove
1
For example, the work of Kalai et al. [KMV10] can be reformulated as showing that for any pair of mixtures of two Gaussians
(with suitably bounded parameters), the following quantities are polynomially related: (1) discrepancy in their low-order moments,
(2) their parameter distance, and (3) their total variation distance. This ensures that any candidate set of parameters that produce
almost identical moments must itself result in a distribution that is close in total variation distance.
2
This discussion already points to why it may be challenging to design agnostic algorithms for mixtures of arbitrary Gaussians
or arbitrary product distributions: It is not clear what notion of parameter distance is polynomially related to the total variation
distance between two such mixtures, without any dependence on d.
2
√
points that are obviously outliers) can be Ω(ε d)-far from the true mean in ℓ2 -distance. The filter approach
shows that either the sample mean is already a good estimate for µ or else there is an elementary spectral test
that rejects some of the corrupted points and almost none of the uncorrupted ones. The crucial observation
is that if a small number of corrupted points are responsible for a large change in the sample mean, it must
be the case that many of the error points are very far from the mean in some particular direction. Thus, we
obtain our filter by computing the top absolute eigenvalue of a modified sample covariance matrix.
Our second algorithmic approach relies on convex programming. Here, instead of rejecting corrupted
samples,
PNwe compute appropriate weights wi for the samples Xi , so that the weighted empirical average
µ
bw = i=1 wi Xi is close to µ. We work with the convex set:
Cδ =
wi | 0 ≤ wi ≤ 1/((1 − ε)N ),
N
P
wi = 1,
i=1
N
P
i=1
T
wi (Xi − µ)(Xi − µ) − I
2
≤δ
.
P
We prove that any set of weights in Cδ yields a good estimate µ
bw = N
i=1 wi Xi in the obvious way. The
catch is that the set Cδ is defined based on µ, which is unknown. Nevertheless, it turns out that we can use
the same types of spectral arguments that underly the filtering approach to design an approximate separation
oracle for Cδ . Combined with standard results in convex optimization, this yields an algorithm for robustly
estimating µ.
The third and final ingredient is some new concentration bounds. In both of the approaches above, at
best we are hoping that we can remove all of the corrupted points and be left with only the uncorrupted
ones, and then use standard estimators (e.g., the empirical average) on them. However, an adversary could
have removed an ε-fraction of the samples in a way that biases the empirical average of the remaining
uncorrupted samples. What we need are concentration bounds that show for sufficiently large N , for samples
X1 , X2 , . . . , XN from a Gaussian with mean µ and identity covariance, that every (1 − ε)N set of samples
produces a good estimate for µ. In some cases, we can derive such concentration bounds by appealing to
known concentration inequalities and taking a union bound. However, in other cases (e.g., concentration
bounds for degree two polynomials of Gaussian random variables) the existing concentration bounds are not
strong enough, and we need other arguments to prove that every set of (1 − ε)N samples produces a good
estimate.
1.3 Our Results
We give the first efficient algorithms for agnostically learning several important distribution classes with
dimension-independent error guarantees. Our first main result is for a single arbitrary Gaussian with mean
µ and covariance Σ, which we denote by N (µ, Σ). In the previous subsection, we described our convex
programming approach for learning the mean vector when the covariance is promised to be the identity. A
technically more involved version of the technique can handle the case of zero mean and unknown covariance. More specifically, consider the following convex set, where Σ is the unknown covariance matrix:
N
N
P
−1/2 P
T
−1/2
Cδ = wi | 0 ≤ wi ≤ 1/((1 − ε)N ),
wi = 1, Σ
wi Xi Xi Σ
−I
≤δ .
i=1
i=1
F
We design an approximate separation oracle for this unknown convex set, by analyzing the spectral properties of the fourth moment tensor of a Gaussian. Combining these two intermediate results, we obtain our first
e (d, ε, τ )) when remain result (below). Throughout this paper, we will abuse notation and write N ≥ Ω(f
ferring to our sample complexity, to signify that our algorithm works if N ≥ Cf (d, ε, τ )polylog(f (d, ε, τ ))
for a large enough universal constant C.
3
Theorem 1.2. Let µ, Σ be arbitrary and unknown, and let ε, τ > 0. There is a polynomial
3 2 time
algorithm
d
log
1/τ
e
which given ε, τ, and an ε-corrupted set of N samples from N (µ, Σ) with N ≥ Ω
, produces µ
b
ε2
b so that with probability 1 − τ we have dTV (N (µ, Σ), N (b
b ≤ O(ε log 3/2 (1/ε)).
and Σ
µ, Σ))
We can alternatively establish Theorem 1.2 with a slightly worse sample complexity via our filtering technique. See Section 8.
Our second agnostic learning result is for a product distribution on the hypercube – arguably the most
fundamental discrete high-dimensional distribution. We solve this problem using our filter technique, though
our convex programming approach would also yield similar results. We start by analyzing the balanced case,
when no coordinate is very close to being deterministic. This special case is interesting in its own right and
captures the essential ideas of our more involved analysis for the general case. The reason is that, for two
balanced product distributions, the ℓ2 -distance between their means is equivalent to their total variation
distance (up to a constant factor). This leads to a clean and elegant presentation of our spectral arguments.
For an arbitrary product distribution, we handle the coordinates that are essentially deterministic separately.
Moreover, we use the χ2 -distance between the means as the parameter distance and, as a consequence, we
need to apply the appropriate corrections to the covariance matrix. Formally, we prove:
Theorem 1.3. Let Π be an unknown binary product distribution, and let ε, τ > 0. There
time
6is a polynomial
d log(1/τ )
, produces
algorithm which given ε, τ, and an ε-corrupted set of N samples from Π with N ≥ Ω
ε3
p
e so that with probability 1 − τ , we have dTV (Π, Π)
e ≤ O( ε log 1/ε).
a binary product distribution Π
For the sake of simplicity in the presentation, we did not make an effort to optimize the sample complexity of our robust estimators. We also
p remark that for the case of balanced binary product distributions,
our algorithm achieves an error of O(ε log(1/ε)).
Interestingly enough, the above two distribution classes are trivial to learn in the noiseless case, but in
the agnostic setting the learning problem turns out to be surprisingly challenging. Using additional ideas, we
are able to generalize our agnostic learning algorithms to mixtures of the above classes under some natural
conditions. We note that learning mixtures of the above families is rather non-trivial even in the noiseless
case. First, we study 2-mixtures of c-balanced products, which stipulates that the coordinates of the mean
vector of each component are in the range (c, 1 − c). We prove:
Theorem 1.4 (informal). Let Π be an unknown mixture of two c-balanced binary product distribution, and
let ε, τ > 0. There is a polynomial
time algorithm which given ε, τ, and an ε-corrupted set of N samples
d4 log 1/τ
e so that with
from Π with N ≥ Ω
, produces a mixture of two binary product distributions Π
ε13/6
e ≤ Oc (ε1/6 ).
probability 1 − τ , we have dTV (Π, Π)
This generalizes the algorithm of Freund and Mansour [FM99] to the agnostic setting. An interesting open
question is to improve the ε-dependence in the above bound to (nearly) linear, or to remove the assumption
of balancedness and obtain an agnostic algorithm for mixtures of two arbitrary product distributions.
Finally, we give an agnostic learning algorithm for mixtures of spherical Gaussians.
Theorem 1.5. Let M be a mixture of k Gaussians with spherical covariances, and let ε, τ > 0 and k be a
constant. There is a polynomial time (for constant k) algorithm which given ε, τ , and an ε-corrupted set of
N samples from M with N ≥ poly(k, d, 1/ε, log(1/τ )), outputs an M′ so that with probability 1 − τ , we
√
have dTV (M, M′ ) ≤ Õ(poly(k) · ε).
Our agnostic algorithms for (mixtures of) balanced product distributions and for (mixtures of) spherical
Gaussians are conceptually related, since in both cases the goal is to robustly learn the means of each
component with respect to ℓ2 -distance.
4
In total, these results give new robust and computationally efficient estimators for several well-studied
distribution learning problems that can tolerate a constant fraction of errors independent of the dimension.
This points to an interesting new direction of making robust statistics algorithmic. The general recipe we
have developed here gives us reason to be optimistic about many other problems in this domain.
1.4 Discussion and Related Work
Our results fit in the framework of density estimation and parameter learning which are both classical
problems in statistics with a rich history (see e.g., [BBBB72, DG85, Sil86, Sco92, DL01]). While these
problems have been studied for several decades by different communities, the computational complexity of learning is still not well understood, even for some surprisingly simple distribution families. Most
textbook estimators are hard to compute in general, especially in high-dimensional settings. In the past
few decades, a rich body of work within theoretical computer science has focused on designing computationally efficient distribution learning algorithms. In a seminal work, Kearns et al. [KMR+ 94] initiated
a systematic investigation of the computational complexity of distribution learning. Since then, efficient
learning algorithms have been developed for a wide range of distributions in both low and high-dimensions
[Das99, FM99, AK01, VW02, CGG02, MR05, BV08, KMV10, MV10, BS10, DDS12a, DDS12b, CDSS13,
DDO+ 13, CDSS14a, CDSS14b, HP15, ADLS15, DDS15b, DDKT16, DKS15, DKS16].
We will be particularly interested in efficient learning algorithms for mixtures of high-dimensional Gaussians and mixtures of product distributions, as this is the focus of our algorithmic results in the agnostic setting. In a pioneering work, Dasgupta [Das99] introduced the problem of parameter estimation of a Gaussian
mixture to theoretical computer science, and gave the first provably efficient algorithms under the assumption that the components are suitably well-separated. Subsequently, a number of works improved these separation conditions [AK01, VW02, BV08] and ultimately removing them entirely [KMV10, MV10, BS10].
In another line of work, Freund and Mansour [FM99] gave the first polynomial time algorithm for properly
learning mixtures of two binary product distributions. This algorithm was substantially generalized to phylogenetic trees [CGG02] and to mixtures of any constant number of discrete product distributions [FOS05].
Given the vast body of work on high-dimensional distribution learning, there is a plethora of problems where
one could hope to reconcile robustness and computational efficiency. Thus far, the only setting where robust
and efficient algorithms are known is on one-dimensional distribution families, where brute-force search or
some form of polynomial regression often works. In contrast, essentially nothing is known about efficient
agnostic distribution learning in the high-dimensional setting that we study here.
Question 1.1 also resembles learning in the presence of malicious errors [Val85, KL93]. There, an algorithm is given samples from a distribution along with their labels according to an unknown target function.
The adversary is allowed to corrupt an ε-fraction of both the samples and their labels. A sequence of works
studied this problem for the class of halfspaces [Ser01, Ser03, KLS09] in the setting where the underlying
distribution is a Gaussian, culminating in the work of Awasthi et al. [ABL14], who gave an efficient algorithm that finds a halfspace with agreement O(ε). Our work and theirs are not directly comparable, since we
work in an unsupervised setting. Moreover, their algorithms need to assume that the underlying Gaussian
distribution is in isotropic position. In fact, our results are complementary to theirs: One could use our
algorithms (on the unlabeled examples) to learn an affine transformation that puts the underlying Gaussian
distribution in approximately isotropic position, even in the presence of malicious errors, so that one can
then directly apply the [ABL14] algorithm.
Another connection is to the work on robust principal component analysis (PCA). PCA is a transformation that (among other things) is often justified as being able to find the affine transformation Y =
Σ−1/2 (X − µ) that would place a collection of Gaussian random variables in isotropic position. One can
think of our results on agnostically learning a Gaussian as a type of robust PCA that tolerates gross corruptions, where entire samples are corrupted. This is different than other variants of the problem where random
5
sets of coordinates of the points are corrupted [CLMW11], or where the uncorrupted points were assumed
to lie in a low-dimensional subspace to begin with [ZL14, LMTZ12]. Finally, Brubaker [Bru09] studied the
problem of clustering samples from a well-separated mixture of Gaussians in the presence of adversarial
noise. The goal of [Bru09] was to separate the Gaussian components from each other, while the adversarial
points are allowed to end up in any of clusters. Our work is orthogonal to [Bru09], since even if such a
clustering is given, the problem still remains to estimate the parameters of each component.
1.5 Organization
The structure of this paper is as follows: In Section 2, we introduce basic notation and a number of useful
facts that will be required throughout the paper, as well as the formal definition of our adversary model. In
Section 3, we discuss several natural approaches to high-dimensional agnostic learning, all of which lose
polynomial factors that depend on the dimension, in terms of their error guarantee.
The main body of the paper is in Sections 4–8. Sections 4 and 5 illustrate our convex programming
framework, while Sections 6–8 illustrate our filter framework.
More specifically, in Section 4 we apply our convex programming framework to a single Gaussian
with unknown mean and unknown covariance. In Section 5, we generalize these techniques to obtain an
agnostic algorithm for mixtures of spherical Gaussians with unknown means. In Section 6, we apply our
filter techniques to a binary product distribution. In Section 7 we generalize the techniques of Section 6 to
obtain an agnostic learning algorithm for a mixture of two balanced binary product distributions. Finally,
in Section 8 we apply our filter technique to obtain an alternate agnostic learning algorithm for a single
Gaussian with unknown mean and unknown covariance.
We note that for some of the more advanced applications of our frameworks, the technical details can
get in the way of the fundamental ideas. For the reader who is interested in seeing the details of our most
basic application of the convex programming framework, we recommend reading the case a Gaussian with
unknown mean, in Section 4.3. Similarly, for the filter framework, we suggest either the balanced product
distribution in Section 6.1 or the Gaussian with unknown mean in Section 8.1.
2 Preliminaries
2.1 Types of Adversaries
In this paper, we will consider a powerful model for agnostic distribution learning that generalizes many
other existing models. The standard setup involves an oblivious adversary who chooses a distribution that
is close in total variation distance to an unknown distribution in some class D.
Definition 2.1. Given ε > 0 and a class of distributions D, the oblivious adversary chooses a distribution
P so that there is an unknown distribution D ∈ D with dTV (P, D) ≤ ε. An algorithm is then given m
independent samples X1 , X2 , . . . , Xm from P .
b in D, where dTV (D, D)
b is small.
The goal of the algorithm is to return the parameters of a distribution D
We refer to the above adversary as oblivious because it fixes the model for noise before seeing any of the
samples. In contrast, a more powerful adversary is allowed to inspect the samples before corrupting them,
both by adding corrupted points and deleting uncorrupted points. We refer to this as the full adversary:
Definition 2.2. Given ε > 0, and a class of distributions D, the full adversary operates as follows: The
algorithm specifies some number of samples m. The adversary generates m samples X1 , X2 , . . . , Xm from
some (unknown) distribution D ∈ D. It then draws m′ from an appropriate distribution. This distribution is
allowed to depend on X1 , X2 , . . . , Xm , but when marginalized over the m samples satisfies m′ ∼ Bin(ε, m).
6
The adversary is allowed to inspect the samples, and removes m′ of them, and replaces them with arbitrary
points. The set of m points is given (in any order) to the algorithm.
We remark that there are no computational restrictions on the adversary. As before, the goal is to return
b in D, where dTV (D, D)
b is small. The reason we allow the draw m′ to
the parameters of a distribution D
depend on the samples X1 , X2 , . . . , Xm is because our algorithms will tolerate this extra generality, and it
will allow us to show that the full adversary is at least as strong as the oblivious adversary (this would not
necessarily be true if m′ were sampled independently from Bin(ε, m)).
We rely on the following well-known fact:
Fact 2.3. Let P, D be two distributions so that dTV (P, D) = ε. Then there are distributions N1 and N2 so
that (1 − ε1 )P + ε1 N1 = (1 − ε2 )D + ε2 N2 , where ε1 + ε2 = ε.
Now we can describe how the full adversary can corrupt samples from D to get samples distributed
according to P .
Claim 2.4. The full adversary can simulate any oblivious adversary.
Proof. We draw m samples X1 , X2 , . . . , Xm from D. We delete each sample Xi independently with probability ε2 and replace it with an independent sample from N2 . This gives a set of samples Y1 , Y2 , . . . , Ym
that are independently sampled from (1 − ε2 )D + ε2 N2 . Since the distributions (1 − ε1 )P + ε1 N1
and 1 − ε2 )D + ε2 N2 are identical, we can couple them to independent samples Z1 , Z2 , . . . , Zm from
(1 − ε1 )P + ε1 N1 . Now each sample Zi that came from N1 , we can delete and replace with an independent
sample from P . The result is a set of samples that are independently sampled from P where we have made
m′ edits and marginally m′ ∼ Bin(ε1 + ε2 , m), although m′ has and needs to have some dependence on the
original samples from D.
The challenge in working with the full adversary is that even the samples that came from D can have biases. The adversary can now choose how to remove uncorrupted points in a careful way so as to compensate
for certain other biases that he introduces using the corrupted points.
Throughout this paper, we will make use of the following notation and terminology:
Definition 2.5. We say a a set of samples X1 , X2 , . . . , Xm is an ε-corrupted set of samples generated by
the oblivious (resp. full) adversary if it is generated by the process described above in the definition of the
oblivious (resp. full) adversary. If it was generated by the full adversary, we let G ⊆ [m] denote the indices
of the uncorrupted samples, and we let E ⊆ [m] denote the indices of the corrupted samples.
In this paper, we will give a number of algorithms for agnostic distribution learning that work in the full
adversary model. In our analysis, we will identify a set of events that ensure the algorithm succeeds and
will bound the probability that any of these events does not occur when m is suitably large. We will often
explicitly invoke the assumption that |E| ≤ 2εm. We can do this even though the number
that are
of points
log 1/τ
, we know
corrupted is itself a random variable, because by the Chernoff bound, as long as m ≥ O
ε
that |E| ≤ 2εm holds with probability at least 1 − O(τ ). Thus, making the assumption that |E| ≤ 2εm
costs us an additional additive O(τ ) term in our union bound, when bounding the failure probability of our
algorithms.
2.2 Basic Notation
Throughout this paper, if v is a vector, we will let kvk2 denote its Euclidean norm. If M is a matrix, we
will let kM k2 denote its spectral norm, and kM kF denote its Frobenius norm. For a discrete distribution
P , we will denote by P (x) the probability mass at point x. For a continuous distribution, let it denote the
probability density function at x. Let S be a multiset over {0, 1}d . We will write X ∈u S to denote that X
is drawn from the empirical distribution defined by S.
7
2.3 Distributions of Interest
One object of study in this paper is the Gaussian (or Normal) distribution.
Definition 2.6. A Gaussian distribution N (µ, Σ) with mean µ and covariance Σ is the distribution with
probability density function
1
T −1
−d/2
1/2
f (x) = (2π)
|Σ| exp − (x − µ) Σ (x − µ) .
2
We will also be interested in binary product distributions.
Definition 2.7. A (binary) product distribution is a probability distribution over {0, 1}d whose coordinate
random variables are independent. Note that a binary product distribution is completely determined by its
mean vector.
We will also be interested in mixtures of such distributions.
Definition 2.8. A mixture P of distributions P1 , . . . , Pk with mixing weights α1 , . . . , αk is the distribution
defined by
X
P (x) =
αj Pk (x),
where αj ≥ 0 for all j and
P
j∈[k]
j∈[k] αj
= 1.
2.4 Agnostic Hypothesis Selection
Several of our algorithms will return a polynomial-sized list of hypotheses at least one of which is guaranteed
to be close to the target distribution. Usually (e.g. in a non-agnostic setting), one could use a polynomial
number of additional samples to run a tournament to identify the candidate hypothesis that is (roughly) the
closest to the target distribution. In the discussion that follows, we will refer to these additional samples
as test samples. Such hypothesis selection algorithms have been extensively studied [Yat85, DL96, DL97,
DL01, DK14, AJOS14, SOAJ14, DDS15a, DDS15b]. Unfortunately, against a strong adversary we run into
a serious technical complication: the training samples and test samples are not necessarily independent.
Moreover even if we randomly partition our samples in training and test, a priori there are an unbounded set
of possible hypotheses that the training phase could output, and when we analyze the tournament we cannot
condition on the list of hypotheses and assume that the test samples are sampled anew. Our approach is to
require our original algorithm to return only hypotheses from some finite set of possibilities, and as we will
see this mitigates the problem.
Lemma 2.9. Let C be a class of probability distributions. Suppose that for some N, ε, τ > 0 there exists a
polynomial time algorithm that given N independent samples from some Π ∈ C, of which up to a 2ε-fraction
have been arbitrarily corrupted, returns a list of M distributions whose probability density functions are
explicitly computable and which can be effectively sampled from so that with 1 − τ /2 probability there exists
a Π′ ∈ M with dTV (Π′ , Π) < δ. Suppose furthermore that the distributions returned by this algorithm
are all in some fixed set M. Then there exists another polynomial time algorithm, which given O(N +
(log(|M|) + log(1/τ ))/ε2 ) samples from Π, an ε-fraction of which have been arbitrarily corrupted, returns
a single distribution Π′ so that with 1 − τ probability dTV (Π′ , Π) < O(δ + ε).
Proof. Firstly, we randomly choose a subset of N of our samples and a disjoint subset of C(log(|M|) +
log(1/τ ))/ε2 of our samples for some sufficiently large C. Note that with high probability over our randomization, at most a 2ε-fraction of samples from each subset are corrupted. Thus, we may instead consider the stronger adversary who sees a set S1 of N independent samples from Π and another set, S2 , of
8
C(log(|M|) + log(1/τ ))/ε2 samples from Π and can arbitrary corrupt a 2ε-fraction of each, giving sets
S1′ ,S2′ .
With probability at least 1 − τ /2 over S1 , the original algorithm run on S1′ returns a set M satisfying
the desired properties.
For two distributions P and Q in M we let AP Q be the set of inputs x where PrP (x) > PrQ (x). We note
that we can test membership in AP Q as, by assumption, the probability density functions are computable.
We also note that dTV (P, Q) = PrP (AP Q ) − PrQ (AP Q ). Our tournament will depend on the fact that if P
is close to the target and Q is far away, that many samples will necessarily lie in AP Q .
We claim that with probability at least 1 − τ /2 over the choice of S2 , we have that for any P, Q ∈ M
that
Pr (x ∈ AP Q ) = Pr (x ∈ AP Q ) + O(ε).
x∈u S2
x∼Π
This follows by Chernoff bounds and a union bound over the |M|2 possibilities for P and Q. Since the total
variational distance between the uniform distributions over S2 and S2′ is at most 2ε, we also have for S2′ that
Pr (x ∈ AP Q ) = Pr (x ∈ AP Q ) + O(ε).
x∈u S2′
x∼Π
Suppose that dTV (P, Π) < δ and dTV (Q, Π) > 5δ + Cε. We then have that
Pr (x ∈ AP Q ) = Pr (x ∈ AP Q ) + O(ε) ≥ Pr (x ∈ AP Q ) + O(ε) − δ ≥ Pr (x ∈ AP Q ) + δ + Cε/5.
x∈u S2′
x∼Π
x∼P
x∼Q
On the other hand, if dTV (Π, Q) < δ then
Pr (x ∈ AP Q ) = Pr (x ∈ AP Q ) + O(ε) < Pr (x ∈ AP Q ) + δ + Cε/5.
x∈u S2′
x∼Π
x∼Q
Therefore, if we throw away any Q in our list for which there is a P in our list so that
Pr (x ∈ AP Q ) ≥ Pr (x ∈ AP Q ) + δ + Cε/5,
x∈u S2′
x∼Q
we have thrown away all the Q with dTV (Q, Π) > 5δ + Cε, but none of the Q with dTV (Q, Π) < δ.
Therefore, there will be a Q remaining, and returning it will yield an appropriate Π′ .
2.5 Bounds on TV Distance
The Kullbeck-Liebler divergence (also known as relative entropy, information gain, or information divergence) is a well-known measure of distance between two distributions.
Definition 2.10. Let P, Q be two probability distributions on Rd . Then the KL divergence between P and
Q, denoted dKL (P kQ), is defined as
Z
dP
log
dKL (P kQ) =
dP .
dQ
d
R
The primary interest we have in this quantity is the fact that (1) the KL divergence between two Gaussians has a closed form expression, and (2) it can be related (often with little loss) to the total variation
distance between the Gaussians. The first statement is expressed in the fact below:
Fact 2.11. Let N (µ1 , Σ1 ) and N (µ2 , Σ2 ) be two Gaussians so that det(Σ1 ), det(Σ2 ) 6= 0. Then
det(Σ1 )
1
−1
T −1
tr(Σ2 Σ1 ) + (µ2 − µ1 ) Σ2 (µ2 − µ1 ) − d − ln
.
dKL (N (µ1 , Σ1 )||N (µ2 , Σ2 )) =
2
det(Σ2 )
(3)
9
The second statement is encapsulated in the well-known Pinsker’s inequality:
Theorem 2.12 (Pinsker’s inequality). Let P, Q be two probability distributions over Rd . Then,
r
1
dTV (P, Q) ≤
dKL (P, Q) .
2
With this, we can show the following two useful lemmas, which allow us to relate parameter distance
between two Gaussians to their total variation distance. The first bounds the total variation distance between
two Gaussians with identity covariance in terms of the Euclidean distance between the means:
Corollary 2.13. Let µ1 , µ2 ∈ Rd be arbitrary. Then dTV (N (µ1 , I), N (µ2 , I)) ≤ 21 kµ2 − µ1 k2 .
Proof. In the case where Σ1 = Σ2 = I, (3) simplifies to
1
dKL (N (µ1 , I), N (µ2 , I)) = kµ2 − µ1 k22 .
2
Pinsker’s inequality (Theorem 2.12) then implies that
r
1
1
dKL (N (µ1 , I), N (µ2 , I)) = kµ2 − µ1 k2 ,
dTV (N (µ1 , I), N (µ2 , I)) ≤
2
2
as desired.
The second bounds the total variation distance between two mean zero Gaussians in terms of the Frobenius norm of the difference between their covariance matrices:
−1/2
Corollary 2.14. Let δ > 0 be sufficiently small. Let Σ1 , Σ2 so that kI − Σ2
−1/2
Σ1 Σ2
kF = δ. Then,
dTV (N (0, Σ1 ), N (0, Σ2 )) ≤ O(δ) .
−1/2
Proof. Let M = Σ2
−1/2
Σ1 Σ2
. Then (3) simplifies to
dKL (N (µ1 , Σ1 )||N (µ2 , Σ2 )) =
1
(tr(M ) − d − ln det(M )) .
2
Since both terms in the last line are rotationally invariant, we may assume without loss of generality that M
is diagonal. Let M
between the two distributions is
P= diag(1 + λ1 , . . . , 1 + λd ). Thus, the KL divergenceP
given exactly by di=1 λi − log(1 + λi ), where we are guaranteed that ( di=1 λ2i )1/2 = δ. By the linear
approximation to ln(1 + x), for x small, we have that for δ sufficiently small,
!
d
d
X
X
λ2i = Θ(δ2 ) .
λi − log(1 + λi ) = Θ
i=1
i=1
Thus, we have shown that for δ sufficiently small, dKL (N (µ1 , Σ1 )||N (µ2 , Σ2 )) ≤ O(δ2 ). The result now
follows by an application of Pinsker’s inequality (Theorem 2.12).
Our algorithm for agnostically learning an arbitrary Gaussian will be based on solving two intermediate
problems: (1) We are given samples from N (µ, I) and our goal is to learn µ, and (2) We are given samples
from N (0, Σ) and our goal is to learn Σ. The above bounds on total variation distance will allow us to
conclude that our estimate is close in total variation distance to the unknown Gaussian distribution in each
of the two settings.
We will also need the following lemma, bounding the total variation distance between two product
distributions by the chi-squared distance between the corresponding mean vectors.
10
Lemma 2.15. Let P, Q be binary product distributions with mean vectors p, q ∈ (0, 1)d . We have that
d2TV (P, Q) ≤ 2
d
X
i=1
(pi − qi )2
.
(pi + qi )(2 − pi − qi )
The proof is deferred to Section A.
2.6 Miscellaneous Lemmata
We will frequently use the following version of the Chernoff bound:
Theorem 2.16. Let Z1 , . . . , Zn be independent random variables with Zi supported on [ai , bi ]. Let Z =
P
d
i=1 Zi . Then for any T > 0,
!
−2T 2
Pr(|Z − E[Z]| > T ) ≤ 2 exp Pd
.
2
i=1 (bi − ai )
3 Some Natural Approaches, and Why They Fail
Many of the agnostic distribution learning problems that we study are so natural that one would immediately
wonder why simpler approaches do not work. Here we detail some other plausible approaches, and what
causes them to lose dimension-dependent factors (if they have any guarantees at all!). For the discussion
that follows, we note that by Fact 2.13 in order to achieve an estimate that is O(ε)-close in total variation
distance (for a Gaussian when µ is unknown and Σ = I) it is necessary and sufficient that kµ̂ − µk2 = O(ε).
Learn Each Coordinate Separately
One plausible approach for agnostic learning in high dimensions is to agnostically learn along each coordainate separately. For instance, if our goal is to agnostically learn the mean of a Gaussian with known
covariance I, we could try to learn each coordinate of the mean separately. But since a ε-fraction
√ of the
samples are corrupted, our estimate can be off by ε in each coordinate and would be off by ε d in high
dimensions.
Maximum Likelihood
Given a set of samples X1 , . . . , XN , and a classQof distributions D, the maximum likelihood estimator
(MLE) is the distribution F ∈ D that maximizes m
i=1 F (Xi ). Equivalently, F minimizes the negative log
likelihood (NLL), which is given by
NLL(F, X1 , . . . , XN ) = −
N
X
log F (Xi ) .
i=1
d
In particular, if D = {N (µ, I) : µ ∈ R } is the set of Gaussians with unknown mean and identity covariance, we see that for any µ ∈ Rd , the NLL of the set of samples is given by
N
X
1 −kXi −µk22 /2
log √ e
NLL(N (µ, I), X1 , . . . , XN ) = −
2π
i=1
= N log
N
√
1X
kXi − µk22 ,
2π +
2
i=1
11
and so the µ which minimizes NLL(N (µ, I), X1 , . . . , XN ) is the mean of the samples X
Pi ,Nsince for any
set of vectors v1 , . . . , vN , the average of the vi ’s is the minimizer of the function h(x) = i=1 kvi − xk22 .
Hence if an adversary places an ε-fraction of the points at some very large distance, then the estimate for
the mean would need to move considerably in that direction. By placing the corruptions further and further
away, the MLE can be an arbitrarily bad estimate. Hence even though it is well known [Hub67, Whi82] that
the MLE converges to the distribution F ∈ D that is closest in KL-divergence to the distribution from which
our samples were generated (i.e. after the adversary has added corruptions), F is not necessarily close to the
uncorrupted distribution.
Geometric Median
In one dimension, it is well-known that the median provides a provably robust estimate for the mean in
aPnumber of settings. The mean of a set of points a1 , . . . , aN is the minimizer of thePfunction f (x) =
N
N
2
i=1 (ai − x) , and in contrast the median is the minimizer of the function f (x) =
i=1 |ai − x|. In
higher dimensions, there are many natural definitions for the median that generalize the one-dimensional
case. The Tukey median is one such notion, but as we discussed it is hard to compute [JP78] and the best
known algorithms run in time exponential in d. Motivated by this, the geometric median is another highdimensional notion of a median. It often achieves better robustness than the mean, and can be computed
quickly [CLM+ 16]. The formal definition is:
X
med(S) , min
kx − vk2 .
v
x∈S
Next, we show that the geometric median also does not give dimension independent error guarantees
for learning the mean of a Gaussian with known covariance. Suppose we have a set S1 of (1 − ε)N samples
from N (0, I) and a set S2 of εN samples ∈ {0, 1}d . Each element x ∈ S2 is generated independently by
choosing a random set Qx ⊂ [d] of cardinality d/2 and setting xi = 1 if i ∈ Qx , and 0 otherwise. Call this
distribution D. P
Let f (v) = x∈S1 ∪S2 kxj − vk2 , and let v ∗ be the value which minimizes this function, which is also
the geometric median of this set of points. We know that the gradient
is 0 at this point, and thus, for all i,
√
(∇f (v ∗ ))i = 0. Assume that kv ∗ k2 ≤ ε, where we set ε ≥ Ω(1/ d). We will show to be a contradiction.
Writing the expression for the gradient, we have
(∇f (v))i =
1 X vi − x i
1 X vi − xi
+
.
2
kx − vk2 2
kx − vk2
x∈S1
x∈S2
We will bound each term of this in turn at v = v ∗ , assuming kv ∗ k ≤ ε. First, observe that
X
1
vi∗
1
∗
− vi E
→0
(1 − ε)N
kx − v ∗ k2
kx − v ∗ k2
x∈S1
as N → ∞ by a Chernoff bound, and by basic concentration, we have that E
12
h
1
kx−v∗ k2
i
√
= Θ(1/ d), if
kv ∗ k ≤ O(1). We also have
X
X
1
1
xi
xi
qP
qP
−
d
∗ 2
(1 − ε)N
(1 − ε)N
∗ 2
x∈S1
x∈S1
j6=i (xj − vj )
j=1 (xj − vj )
qP
qP
d
∗ 2
∗ )2 −
(x
−
v
X
j
j=1 (xj − vj )
j6=i
j
1
≤
|xi | · r
P
Pd
(1 − ε)N
∗ 2
∗ 2
x∈S1
j=1 (xj − vj )
j6=i (xj − vj )
X
e (1)
1
O
1
e
|xi |
≤O
,
≤
e
(1 − ε)N
d
Ω(d)
x∈S1
from basic Gaussian concentration. Moreover, we have
X
1
qP
(1 − ε)N
x∈S1
xi
∗ 2
j6=i (xj − vj )
→0
as N → ∞, by another Chernoff bound. Putting these terms together, we have that
X v ∗ − xi
1
1
∗
i
e 1 + o(1) .
≤O
− vi
E
∗
(1 − ε)N
kx − vk2
d
x∼N (0,I) kx − v k2
x∈S1
Similarly, we also have that since Ex∼D [xi ] = 1/2,
1 X vi∗ − xi
1
1
∗
e
− (vi + 1/2) E
≤O
+ o(1) ,
∗
x∼D kx − v k2
εN
kx − vk2
d
x∈S2
√
= Θ 1/ d . Thus, altogether we have that
1
1
1
1
∗
∗
∗
0=
∇f (v )i = vi · Θ √
+ o(1) ,
+ ε(vi + 1/2) · Θ √
+O
(1 − ε)N
d
d
d
√
which we see is impossible for our choice of ε ≥ Ω(1/ d) unless |vi∗ | = Ω(ε). In particular,
this means that
√
∗
each coordinate of the geometric median must be at least Ω(ε), and so kv k2 ≥ Ω(ε d), which contradicts
our original assumption. This finishes the analysis.
and moreover, Ex∼D
h
1
kx−v∗ k2
i
4 Agnostically Learning a Gaussian, via Convex Programming
e
In this section we give a polynomial time algorithm to agnostically learn a single Gaussian up to error O(ε).
Our approach is based on the following ingredients: First, in Section 4.1, we define the set SN,ε , which
will be a key algorithmic object in our framework. In Section 4.2 we give key, new concentration bounds
on certain statistics of Gaussians. We will make crucial use of these concentration bounds throughout this
section. In Section 4.3 we give an algorithm to agnostically learn a Gaussian with unknown mean and whose
covariance is promised to be the identity via convex programming. This will be an important subroutine in
our overall algorithm, and it also helps to illustrate our algorithmic approach without many of the additional
complications that arise in our later applications. In Section 4.4 we show how to robustly learn a Gaussian
with mean zero and unknown covariance again via convex programming. Finally, in Section 4.5 we show
how to combine these two intermediate results to get our overall algorithm.
13
4.1 The Set SN,ε
An important algorithmic object for us will be the following set:
Definition 4.1. For any 21 > ε > 0 and any integer N , let
X
Sm,ε = (w1 , . . . , wN ) :
wi = 1, and 0 ≤ wi ≤
1
, ∀i .
(1 − 2ε)N
Next, we motivate this definition. For any J ⊆ [N ], let wJ ∈ RN be the vector which is given by
1
for i ∈ J and wiJ = 0 otherwise. Then, observe that
wiJ = |J|
SN,ε = conv wJ : |J| = (1 − 2ε)N )
,
and so we see that this set is designed to capture the notion of selecting a set of (1 − 2ε)N samples from N
samples.
Given w ∈ SN,ε we will use the following notation
X
X
wg =
wi and wb =
wi
i∈G
i∈E
to denote the total weight on good and bad points respectively. The following facts are immediate from
|E| ≤ 2εN and the properties of SN,ε .
2ε
. Moreover, the renormalized weights w′ on good
Fact 4.2. If w ∈ SN,ε and |E| ≤ 2εN , then wb ≤ 1−2ε
wi
′
′
points given by wi = wg for all i ∈ G, and wi = 0 otherwise, satisfy w′ ∈ SN,4ε .
4.2 Concentration Inequalities
Throughout this section and in Section 5 we will make use of various concentration bounds on low moments
of Gaussian random variables. Some are well-known, and others are new but follow from known bounds
and appropriate union bound arguments.
4.2.1
Empirical Estimates of First and Second Moments
Here we will give rates of convergence for various statistics for Gaussians with covariance matrix I that we
will make use of later. First, we will require the following well-known tail bounds:
Lemma4.3 ([Ver10],
Theorem 5.50). Let Yi ∼ N (0, I) be independent, for i = 1, . . . , N , where we set
d+log(1/τ )
N =Ω
. We have that with probability 1 − τ ,
δ2
N
X
1
Yi YiT − I
m
i=1
2
≤δ.
Lemma 4.4 (folklore). With the same setup as above, with probability 1 − τ ,
N
1 X
Yi
N
i=1
2
14
≤δ.
The above bounds have the property that as we take more samples, we can take δ to be arbitrarily small.
We will also be interested in how well various statistics concentrate around their expectation, when we take
the worst-case set of weights in SN,ε . This is more subtle because as we take more samples, any fixed
statistic (e.g. taking the uniform distribution over the samples) concentrates better but the size of SN,ε (e.g.
the number of sets of (1 − 2ε)N samples) grows too. We defer the proof to Section B.
p
Lemma 4.5. Fix ε ≤ 1/2 and τ ≤1. There is a δ1 = O(ε log 1/ε) so that if Y1 , . . . , YN are independent
)
, then
samples from N (0, I) and N = Ω d+log(1/τ
δ12
#
"
m
X
T
(4)
Pr ∃w ∈ Sm,ε :
wi Yi Yi − I ≥ δ1 ≤ τ .
i=1
2
A nearly identical argument (Use Hoeffding instead of Bernstein in the proof of Theorem 5.50 in
[Ver10]) yields:
p
Lemma 4.6. Fix ε and τ as above. There is aδ2 = O(ε log 1/ε) so that if Y1 , . . . , YN are independent
)
, then
samples from N (0, I) and N = Ω d+log(1/τ
δ22
#
"
N
X
wi Yi ≥ δ2 ≤ τ .
(5)
Pr ∃w ∈ SN,ε :
i=1
2
Note that by Cauchy-Schwarz, this implies:
p
Corollary 4.7. Fix ε and τ as above.
There
is
a
δ
=
O(ε
log 1/ε) so that if Y1 , . . . , YN are independent
2
d+log(1/τ )
, then
samples from N (0, I) and N ≥ Ω
δ22
!
#
"
m
X
T
d
wi Yi v
(6)
≥ δ2 kvk2 ≤ τ .
Pr ∃v ∈ R , ∃w ∈ SN,ε :
i=1
2
We will also require the following, well-known concentration, which says that no sample from a Gaussian deviates too far from its mean in distance.
Fact
p4.8. Fix τ > 0. Let X1 , . . . , XN ∼ N (0, I). Then, with probability 1 − τ , we have that kXi k2 ≤
O
d log(N/τ ) for all i = 1, . . . , N .
4.2.2
Estimation Error in the Frobenius Norm
Let X1 , ...XN be N i.i.d. samples from N (0, I). In this section we demonstrate a tight bound on how
b denote the
many samples are necessary so that the sample covariance is close to I in Frobenius norm. Let Σ
empirical covariance, defined to be
N
X
b= 1
Σ
Xi XiT .
N
i=1
By self-duality of the Frobenius norm, we know that
D
E
b − IkF = max
b − I, U
kΣ
Σ
kU kF =1
=
max
kU kF =1
m
1 X
tr(Xi XiT U ) − tr(U ) .
m
i=1
We first state the following lemma, whose proof is deferred to Section B:
15
2 +log 1/τ ′
Lemma 4.9. Let U 0 with kU kF = 1. Then, for any τ ′ ≥ 0 and any δ ≥ 0, if N = Ω( d
with probability 1 − τ ′ .
δ2
) then
N
1 X
tr(Xi XiT U ) − tr(U ) ≤ δ
N
i=1
2
Since there is a 1/4-net over all PSD matrices with Frobenius norm 1 of size 9d , the Vershynin-type
union bound argument immediately gives us the following:
Corollary 4.10. For any δ, τ > 0, if
N =Ω
d2 + log 1/τ
δ2
,
b − IkF ≤ δ with probability 1 − τ .
then kΣ
By the same union bound technique as used in the proof of Lemma 4.5, we obtain:
p
Corollary 4.11. Fix ε, τ > 0. There is a δ1 = O(ε log 1/ε) so that if X1 , . . . , XN are independent
samples from N (0, I), with
2
d + log 1/τ
N =Ω
,
δ12
then
"
Pr ∃w ∈ SN,ε :
N
X
i=1
wi Xi XiT − I
F
#
≥ δ1 ≤ τ .
In fact, the proof technique there can be used to show something slightly stronger, which we will require
later. The technique actually shows that if we take any set of size at most εm, and take the uniform weights
over that set, then the empirical covariance is not too far away from the truth. More formally:
p
Corollary 4.12. Fix ε, τ > 0. There is a δ2 = O(ε log 1/ε) so that if X1 , . . . , XN are independent
samples from N (0, I), with
2
d + log 1/τ
N =Ω
,
δ22
then
"
X 1
Xi XiT − I
Pr ∃T ⊆ [N ] : |T | ≤ εN and
|T |
4.2.3
i∈T
F
#
N
≥ O δ2
≤τ .
|T |
Understanding the Fourth Moment Tensor
Our algorithms will be based on understanding the behavior of the fourth moment tensor of a Gaussian when
restricted to various subspaces. Let ⊗ denote the Kronecker product on matrices. We will make crucial use
of the following definition:
2
Definition 4.13. For any matrix M ∈ Rd×d , let M ♭ ∈ Rd denote its canonical flattening into a vector in
2
2
Rd , and for any vector v ∈ Rd , let v ♯ denote the unique matrix M ∈ Rd×d so that M ♭ = v.
16
We will also require the following definitions:
2
Definition 4.14. Let Ssym = {M ♭ ∈ Rd : M is symmetric}, and let S ⊆ Ssym be the subspace given by
S = {v ∈ Ssym : tr(v ♯ ) = 0} ,
and let ΠS and ΠS ⊥ denote the projection operators onto S and S ⊥ respectively. Finally let
kvkS = kΠS vk2 and kvkS ⊥ = kΠS ⊥ vk2 .
2 ×d2
Moreover, for any M ∈ Rd
, let
kM kS =
vT M v
.
v∈S −{0} kvk2
sup
In fact, the projection of v = A♭ onto S where A is symmetric can be written out explicitly. Namely, it
is given by
tr(M )
tr(M )
M= M−
I +
I.
d
d
By construction the flattening of the first term is in S and the flattening of the second term is in S ⊥ . The
)|
√
expression above immediately implies that kvkS ⊥ = |tr(M
.
d
The key result in this section is the following:
Theorem 4.15. Let X ∼ N (0, Σ). Let M be the d2 × d2 matrix given by M = E[(X ⊗ X)(X ⊗ X)T ].
Then, as an operator on Ssym , we have
T
.
M = 2Σ⊗2 + Σ♭ Σ♭
It is important to note that the two terms above are not the same; the first term is high rank, but the
second term is rank one. The proof of this theorem will require Isserlis’s theorem, and is deferred to Section
B.
Theorem 4.16 (Isserlis’s theorem). Let a1 , . . . , ak ∈ Rd be fixed vectors. Then if X ∼ N (0, I), we have
#
" k
Y
XY
hai , Xi =
hai , aj i ,
E
where the
4.2.4
PQ
i=1
is over all matchings of {1, . . . , k}.
Concentration of the Fourth Moment Tensor
We also need to show that the fourth moment tensor concentrates:
Theorem 4.17. Fix ε, τ > 0. Let Yi ∼ N (0, I) be independent, for i = 1, . . . , m, where we set
2
d + log 1/τ
e
,
N =Ω
δ32
Let Zi = Yi⊗2 . Let M4 = E[Zi ZiT ] be the canonical flattening of the true fourth moment tensor. There is a
δ3 = O(ε log2 1/ε) so that if Y1 , . . . , YN , and Z1 , . . . , Zm be as above, then we have
#
"
N
X
wi Zi ZiT − M4 ≥ δ3 ≤ τ .
Pr ∃w ∈ Sm,ε :
i=1
S
17
To do so will require somewhat more sophisticated techniques than the ones used so far to bound spectral
deviations. At a high level, this is because fourth moments of Gaussians have a sufficiently larger variance
that the union bound techniques used so far are insufficient. However, we will show that the tails of degree
four polynomials of Gaussians still sufficiently concentrate so that removing points cannot change the mean
by too much. The proof appears in Section B.1.
4.3 Finding the Mean, Using a Separation Oracle
In this section, we consider the problem of approximating µ given N samples from N (µ, I) in the full
adversary model. Our algorithm will be based on working with the following convex set:
)
(
N
X
wi (Xi − µ)(Xi − µ)T − I ≤ δ .
Cδ = w ∈ SN,ε :
i=1
2
It is not hard to show that Cδ is non-empty for reasonable values of δ (and we will show this later). Moreover
we will show that for any set of weights w in Cδ , the empirical average
µ
b=
N
X
wi Xi
i=1
will be a good estimate for µ. The challenge is that since µ itself is unknown, there is not an obvious way
to design a separation oracle for Cδ even though it is convex. Our algorithm will run in two
√ basic steps.
First, it will run a very naive outlier detection to remove any points which are more than O( d) away from
the good points. These points are sufficiently far away that a very basic test can detect them. Then, with
the remaining points, it will use the approximate separation oracle given below to approximately optimize
with respect to Cδ . It will then take the outputted set of weights and output the empirical mean with these
weights. We will explain these steps in detail below.
Our results will hold under the following deterministic conditions:
p
d log(N/τ ) , ∀i ∈ G ,
(7)
kXi − µk2 ≤ O
X
i∈G
wi (Xi − µ)(Xi − µ)T − wg I
X
i∈G
wi (Xi − µ)
2
2
≤ δ1 ∀w ∈ SN,4ε , and
(8)
≤ δ2 ∀w ∈ SN,4ε .
(9)
The concentration bounds we gave earlier were exactly bounds on the failure probability of either of these
conditions, albeit for Sm,ε instead of Sm,4ε .
4.3.1
Naive Pruning
The√first step of our algorithm will be to remove points which have distance which is much larger than
O( d) from the mean. Our algorithm is very naive: it computes
√ all pairwise distances between points,
and throws away all points which have distance more than O( d) from more than an 2ε-fraction of the
remaining points.
Then we have the following fact:
Fact 4.18. Suppose that (7) holds. Then NAIVE P RUNE removes
points, and moreover, if Xi
p no uncorrupted
d log(N/τ ) .
is not removed by NAIVE P RUNE, we have kXi − µk2 ≤ O
18
Algorithm 1 Naive Pruning
1: function NAIVE P RUNE(X1 , . . . , XN )
2:
For i, j = 1, . . . , N , define δi,j = kXi − Xj k2 .
3:
for i = 1, . . . , j do
p
4:
Let Ai = {j ∈ [N ] : δi,j > Ω( d log(N/τ ))}
5:
if |Ai | > 2εN then
6:
Remove Xi from the set.
7:
return the pruned set of samples.
Proof. That no uncorrupted point follows directly from (7) and the fact that there can be at most 2εm
corrupted points. Similarly, if Xp
i is not removed by NAIVE P RUNE, that means there must be an uncorrupted
Xj so that kXi − Xj k2 ≤ O( d log(N/τ )). Then the desired property follows from (7) and a triangle
inequality.
Henceforth, for simplicity we shall assume
p that no point was removed by NAIVE P RUNE, and that for all
i = 1, . . . , m, we have kXi − µk2 < O( d log(m/τ )). Otherwise, we can simply work with the pruned
set, and it is evident that nothing changes.
4.3.2
The Separation Oracle
Our main result in this section is an approximate separation oracle for Cδ . Our first step is to show that any
set of weights that does not yield a good estimate for µ cannot be in the set Cδ :
P
Lemma 4.19. Suppose that (8)-(9) holds. Let δ = max(δ1 , δ2 ). Let w ∈ SN,ε and set µ
b= N
i=1 wi Xi and
∆=µ−µ
b. Further, suppose that k∆k ≥ Ω(δ). Then
N
X
i=1
wi (Xi − µ)(Xi − µ)T − I
Proof. By Fact 4.2 and (9) we have k
X
i∈E
wi (Xi − µ)
2
P
wi
i∈G wg Xi
2
≥Ω
k∆k22
ε
.
− µk2 ≤ δ2 . Now by the triangle inequality we have
X
≥ k∆k2 −
i∈G
wi (Xi − µ) − wg µ
2
≥ Ω(k∆k2 )
Using the fact that variance is nonnegative we have
X wi
(Xi − µ)(Xi − µ)T
wb
i∈E
!
!T
X wi
X wi
,
(Xi − µ)
(Xi − µ)
wb
wb
i∈E
i∈E
and therefore
X
i∈E
T
wi (Xi − µ)(Xi − µ)
N
≥Ω
k∆k22
wb
≥Ω
k∆k22
ε
.
On the other hand,
X
i∈G
wi (Xi − µ)(Xi − µ)T − I
2
≤
X
i∈G
wi (Xi − µ)(Xi − µ)T − wg I
19
2
+ wb ≤ δ1 + wb .
where in the last inequality we have used Fact 4.2 and (8). Hence altogether this implies that
m
X
i=1
T
wi (Xi − µ)(Xi − µ) − I
2
≥Ω
k∆k22
ε
− wb − δ1 ≥ Ω
k∆k22
ε
,
as claimed.
As a corollary, we find that any set of weights in Cδ immediately yields a good estimate for µ:
Corollary 4.20. Suppose that (8) and (9) hold. Let w ∈ Cδ for δ = O(ε log 1/ε). Then
p
k∆k2 ≤ O(ε log 1/ε)
Our main result in this section is an approximate separation oracle for Cδ with δ = O(ε log 1/ε).
Theorem 4.21. Fix ε > 0, and let δ = O(ε log 1/ε). Suppose that (8) and (9) hold. Let w∗ denote the
weights which are uniform on the uncorrupted points. Then there is a constant c and an algorithm so that:
1. (Completeness) If w = w∗ , then it outputs “YES”.
2. (Soundness) If w 6∈ Ccδ , the algorithm outputs a hyperplane ℓ : Rm → R so that ℓ(w) ≥ 0 but
ℓ(w∗ ) < 0. Moreover, if the algorithm ever outputs a hyperplane ℓ, then ℓ(w∗ ) < 0.
These two facts imply that the ellipsoid method with this separation oracle will terminate
in poly(d, 1/ε)
p
′
′
steps, and moreover with high probability output a w so that kw − w k < ε/(m d log(m/τ )), for some
w ∈ Ccδ . Moreover, it will do so in polynomially many iterations.
Remark 4.22. The conditions that the separation oracle given here satisfy are slightly weaker than the
traditional guarantees, given, for instance, in [GLS88]. However, the correctness of the ellipsoid algorithm
with this separation oracle follows because outside Ccδ , the separation oracle acts exactly as a separation
oracle for w∗ . Thus, as long as the algorithm continues to query points outside of Ccδ , the action of the
algorithm is equivalent to one with a separation oracle for w∗ . If
The separation oracle is given in Algorithm 2.
Algorithm 2 Separation oracle sub-procedure for agnostically learning the mean.
1: function S EPARATION O RACLE U NKNOWN M EAN (w, X1 , . . . , XN )
P
2:
Let µ
b= N
i=1 wi Xi .
3:
For i = 1, . . . , m, define Yi = Xi − µ
b.
P
T
4:
Let λ be the eigenvalue of largest magnitude of M = N
i=1 wi Yi Yi − I
5:
Let v be its associated eigenvector.
6:
if |λ| < 2c δ then
7:
return “YES”.
8:
else if λ > 2c δ then
P
N
2 − 1 − λ.
9:
return the hyperplane ℓ(w) =
w
hY
,
vi
i=1 i i
10:
else
P
N
2−1 .
11:
return the hyperplane ℓ(w) = λ −
w
hY
,
vi
i
i
i=1
Next, we prove correctness for our approximate separation oracle:
20
Proof of Theorem 4.21. Again, let ∆ = µ − µ
b. By expanding out the formula for M , we get:
N
X
wi Yi YiT
i=1
−I =
=
N
X
i=1
N
X
i=1
=
N
X
i=1
wi (Xi − µ + ∆)(Xi − µ + ∆)T − I
wi (Xi − µ)(Xi − µ)T − I +
N
X
i=1
wi (Xi − µ)∆T + ∆
N
X
i=1
wi (Xi − µ)T + ∆∆T
wi (Xi − µ)(Xi − µ)T − I − ∆∆T .
Let us now prove completeness.
Claim 4.23. Suppose w = w∗ . Then kM k2 < 2c δ.
Proof. Recall that w∗ are the weights that are uniform on the uncorrupted points. Because |E| ≤ 2εN we
have that w∗ p
∈ SN,ε . We can now use (8) to conclude that w∗ ∈ Cδ1 . Now by Corollary 4.20 we have that
k∆k2 ≤ O(ε log 1/ε). Thus
N
X
i=1
wi∗ (Xi − µ)(Xi − µ)T − I − ∆∆T
2
N
X
≤
i=1
wi∗ (Xi − µ)(Xi − µ)T − I
2
+ k∆∆T k2
cδ
.
≤ δ1 + O(ε log 1/ε) <
2
2
We now turn our attention to soundness.
Claim 4.24. Suppose that w 6∈ Ccδ . Then |λ| > 2c δ.
Proof. By the triangle inequality, we have
N
X
i=1
wi (Xi − µ)(Xi − µ)T − I − ∆∆T
p
2
N
X
≥
i=1
wi (Xi − µ)(Xi − µ)T − I
2
− ∆∆T
2
.
Let us now split into two cases. If k∆k2 ≤ cδ/10, then the firstpterm above is at least cδ by definition and
we can conclude that |λ| > cδ/2. On the other hand, if k∆k2 ≥ cδ/10, by Lemma 4.19, we have that
N
X
k∆k22
k∆k22
2
T
T
− k∆k2 = Ω
.
(10)
wi (Xi − µ)(Xi − µ) − I − ∆∆
≥Ω
ε
ε
i=1
2
which for sufficiently small ε also yields |λ| > cδ/2.
Now by construction ℓ(w) ≥ 0. All that remains is to show that ℓ(w∗ ) < 0 always holds. We will
only consider the case where the top eigenvalue λ of M is positive. The other case (when λ < − 2c δ) is
symmetric. We will split the analysis into two parts.
1 X
(Xi − µ
b)(Xi − µ
b)T − I
|G|
i∈G
=
2
1 X
(Xi − µ + ∆)(Xi − µ + ∆)T − I
|G|
i∈G
2
1 X
1 X
(Xi − µ)(Xi − µ)T − I + 2k∆k2
(Xi − µ) +k∆k22
≤
|G|
|G|
i∈G
i∈G
2
{z
} |
{z
}2
|
≤δ1
≤2δ2 k∆k2
21
(11)
Suppose k∆k2 ≤
p
cδ/10. By (11) we immediately have:
ℓ(w∗ ) ≤ δ1 + 2δ2 k∆k2 + k∆k22 − λ ≤
since λ > cδ/2. On the other hand, if k∆k2 ≥
together we have:
ℓ(w∗ ) ≤
p
cδ
−λ<0,
5
cδ/10 then by (10) we have λ = Ω
k∆k22
ε
. Putting it all
1 X
(Xi − µ
b)(Xi − µ
b)T − I −λ ,
|G|
i∈G
2
{z
}
|
≤δ1 +2δ2 k∆k2 +k∆k22
where in the last line we used the fact that λ > Ω
the proof.
4.3.3
k∆k22
ε
, and k∆k22 ≥ Ω(ε2 log 1/ε). This now completes
The Full Algorithm
This separation oracle, along with the classical theory of convex optimization, implies that we have shown
the following:
p
Corollary 4.25. Fix ε, τ > 0, and let δ = O(ε log 1/ε). Let X1 , . . . , XN be a ε-corrupted set of points
satisfying (8)-(9), for δ1 , δ2 ≤ δ. Let c be a sufficiently large constant. Then, there is an algorithm
L EARNA PPROX M EAN (ε, τ, X1 , . . . , XN ) which runs in time poly(N, d, 1/ε,
p log 1/τ ), and outputs a set
of weights w′ ∈ SN,ε so that there is a w ∈ Ccδ so that kw − w′ k∞ ≤ ε/(N d log(N/τ )).
This algorithm, while an extremely powerful primitive, is technically not sufficient. However, given
this, the full algorithm is not too difficult to state: simply run NAIVE P RUNE, then optimize
P over Ccδ
using this separation oracle, and get some w which is approximately in Ccδ . Then, output N
i=1 wi Xi .
For completeness, the pseudocode for the algorithm is given below. In the pseudocode, we assume that
E LLIPSOID (S EPARATION O RACLE U NKNOWN M EAN , ε′ ) is a convex optimization routine, which given the
S EPARATION O RACLE U NKNOWN M EAN separation oracle and a target error ε′ , outputs a w′ so that kw −
w′ k∞ ≤ ε′ . From the classical theory of optimization, we know such a routine exists and runs in polynomial
time.
Algorithm 3 Convex programming algorithm for agnostically learning the mean.
1: function L EARN M EAN (ε, τ, X1 , . . . , XN )
2:
Run NAIVE P RUNE (X1 , . . . , XN ). Let {Xi }i∈I be the pruned set of samples.
3:
For simplicity assume I = [N ].
4:
Let w′ ←
L EARNA PPROX M EAN (ε, τ, X1 , . . . , XN ).
PN
5:
return i=1 wi′ Xi .
We have:
p
Theorem 4.26. Fix ε, τ > 0, and let δ = O(ε log 1/ε). Let X1 , . . . , XN be an ε-corrupted set of samples,
where
d + log 1/τ
N =Ω
.
δ2
Let µ
b be the output of L EARN M EAN (X1 , . . . , XN ). Then, we have kb
µ − µk2 ≤ δ.
22
Proof. By Fact 4.8, Lemma 4.5, and Lemma 4.6, we know that (7)-(9) hold with probability 1 − τ , with
δ1 , δ2 ≤ δ. Condition on the event that this event holds. After NAIVE P RUNE, byp
Fact 4.18 we may assume
that no uncorrupted points are removed, and all points satisfy kXi − µk2 ≤ p
O( d log(N/τ )). Let w′ be
the output of the algorithm, and let w ∈ Ccδ be so that kw − w′ k∞ < ε/(N d log(N/τ )). By Corollary
P
4.20, we know that k N
i=1 wi Xi − µk2 ≤ O(δ). Hence, we have
N
X
i=1
wi′ Xi − µ
2
≤
N
X
i=1
wi Xi − µ
+
N
X
i=1
2
|wi − wi′ | · kXi − µk2 ≤ O(δ) + ε ,
so the entire error is at most O(δ), as claimed.
4.4 Finding the Covariance, Using a Separation Oracle
In this section, we consider the problem of approximating µ given m samples from N (0, Σ) in the full
adversary model. Let Ui = Σ−1/2 Xi so that if Xi ∼ N (0, Σ) then Ui ∼ N (0, I). Moreover let Zi = Ui⊗2 .
Our approach will parallel the one given earlier in Section 4.3. Again, we will work with a convex set
!
)
(
m
X
≤δ .
wi Xi XiT Σ−1/2 − I
Cδ = w ∈ SN,ε : Σ−1/2
i=1
F
and our goal is to design an approximate separation oracle. Our results in this section will rely on the
following deterministic conditions:
X
i∈G
kUi k22 ≤ O (d log(N/τ )) , ∀i ∈ G
wi Ui UiT − wg I
X 1
Ui UiT − I
|T |
i∈T
X
i∈G
wi Zi ZiT − wg M4
F
F
S
≤ δ1 ,
N
≤ O δ2
|T |
≤ δ3 ,
(12)
(13)
, and
(14)
(15)
for all w ∈ SN,ε , and all sets T ⊆ G of size |T | ≤ 2εN . As before, by Fact 4.2, the renormalized weights
over the uncorrupted points are in SN,4ε . Hence, we can appeal to Fact 4.8, Corollary 4.11, Corollary 4.12,
and Theorem 4.17 with SN,4ε instead of SN,ε to bound the probability that this event does p
not hold. Let
∗
w be the set of weights which are uniform over the uncorrupted points; by (13) for δ ≥ Ω(ε log 1/ε) we
have that w∗ ∈ Cδ .
Theorem 4.27. Let δ = O(ε log 1/ε). Suppose that (13), (14), and 15 hold for δ1 , δ2 ≤ O(δ) and δ3 ≤
O(δ log 1/ε). Then, there is a constant c and an algorithm so that, given any input w ∈ SN,ε we have:
1. (Completeness) If w = w∗ , the algorithm outputs “YES”.
2. (Soundness) If w 6∈ Ccδ , the algorithm outputs a hyperplane ℓ : Rm → R so that ℓ(w) ≥ 0 but we
have ℓ(w∗ ) < 0. Moreover, if the algorithm ever outputs a hyperplane ℓ, then ℓ(w∗ ) < 0.
By the classical theory of convex optimization this implies that we will find a point w so that kw − w′ k∞ ≤
ε
′
poly(N ) for some w ∈ Ccδ , using polynomially many calls to this oracle.
23
The pseudocode for the separation oracle is given in Algorithm 4. Observe briefly that this algorithm
does indeed run in polynomial time. Lines 2-7 require only taking top eigenvalues and eigenvectors, and
so
can be done in
polynomial time. For any ξ ∈ {−1, +1}, line 8 can be run by sorting the samples by
√
kYi k2
wi √d − d and seeing if there is a subset of the top 2εN samples satisfying the desired condition,
and line 9 can be executed similarly.
Algorithm 4 Convex programming algorithm for agnostically learning the covariance.
1: function S EPARATION O RACLE U NKNOWN C OVARIANCE (w)
b = PN wi Xi X T .
2:
Let Σ
i
i=1
b −1/2 Xi and let Zi = (Yi )⊗2 .
3:
For i = 1, . . . , N , let Yi = Σ
P
T
4:
Let v be the top eigenvector of M = N
i=1 wi Zi Zi − 2I restricted to S, and let λ be its associated
eigenvalue.
5:
if |λ| > Ω(ε log2 1/ε) then
6:
Let ξ = sgn(λ).
7:
return the hyperplane
!
N
X
2
wi hv, Zi i − 2 − λ .
ℓ(w) = ξ
i=1
8:
9:
else if there exists a sign ξ ∈ {−1, 1} and a set T of samples of size at most 2εN so that
X kYi k2 √ (1 − ε)αδ
√ 2− d >
,
α=ξ
wi
2
d
i∈T
then
return the hyperplane
ℓ(w) = ξ
X
i∈T
10:
wi
kYi k22 √
√ − d −α,
d
return “YES”.
We now turn our attention to proving the correctness of this separation oracle. We require the following
technical lemmata.
Lemma 4.28. Fix δ < 1 and suppose that M is symmetric. If kM − IkF ≥ δ then kM −1 − IkF ≥ 2δ .
Proof. We will prove this lemma in the contrapositive, by showing that if kM −1 − IkF < 2δ then kM −
IkF < δ. Since the Frobenius normP
is rotationally invariant, we may assume that M −1 = diag(1 +
ν1 , . . . , 1 + νd ), where by assumption
νi2 < δ2 /4. By our assumption that δ < 1, we have |νi | ≤ 1/2.
Thus
d
X
i=1
where we have used the inequality |1 −
proof.
1
1−
1 + νi
1
1+x |
2
≤
d
X
4νi2 < δ ,
i=1
≤ |2x| which holds for all |x| ≤ 1/2. This completes the
24
Lemma 4.29. Let M, N ∈ Rd×d be arbitrary matrices. Then kM N kF ≤ kM k2 kN kF .
Proof. Let N1 , . . . , Nd be the columns of N . Then
kM N k2F =
d
X
i=1
kM N k22 ≤ kM k22
d
X
i=1
kNi k22 = kM k22 kN k2F ,
so the desired result follows by taking square roots of both sides.
Lemma 4.30. Let wi for i = 1, . . . , N be a set of non-negative weights so that
be arbitrary. Then
!2
N
N
X
X
.
ai wi2 ≥
ai wi
i=1
PN
i=1 wi
= 1, and let ai ∈ R
i=1
PN
Proof. Let P be the distribution where ai is chosen with probability wi . Then EX∼P [X] =
i=1 ai wi
PN
2
2
2
2
and EX∼P [X ] = i=1 ai wi . Since VarX∼P [X] = EX∼P [X ] − EX∼P [X] is always a non-negative
quantity, by rearranging the desired conclusion follows.
T
≤ kM − Ik2F .
Lemma 4.31. Let M ∈ Rd×d . Then, M ♭ M ♭
S
Proof. By the definition of k · kS , we have
T T
T
M ♭ M ♭ A♭ = sup hA, M i2 .
M♭ M♭
= sup A♭
S
A∈S
kAkF =1
A♭ ∈S
kAkF =1
By self duality of the Frobenius norm, we know that
hA, M i = hA, M − Ii ≤ kM − IkF ,
since I ♭ ∈ S ⊥ . The result now follows.
Proof of Theorem 4.27. Let us first prove completeness. Observe that by Theorem 4.15, we know that
restricted to S, we have that M4 = 2I. Therefore, by (15) we will not output a hyperplane in line 7.
Moreover, by (14), we will not output a hyperplane in line 8. This proves completeness.
Thus it suffices to show soundness. Suppose that w 6∈ Ccδ . We will make use of the following elementary
fact:
b −1/2 and B = Σ
b −1/2 ΣΣ
b −1/2 . Then
Fact 4.32. Let A = Σ−1/2 ΣΣ
kA−1 − IkF = kB − IkF
b −1 Σ1/2 . Using this expression and the fact that all the matrices involved
Proof. In particular A−1 = Σ1/2 Σ
are symmetric, we can write
kA−1 − Ik2F = tr (A−1 − I)T (A−1 − I)
b −1 ΣΣ
b −1 Σ1/2 − 2Σ1/2 Σ
b −1 Σ1/2 − I
= tr Σ1/2 Σ
b −1/2 ΣΣ
b −1 ΣΣ
b −1/2 − 2Σ
b −1/2 ΣΣ
b −1/2 − I
= tr Σ
= tr (B − I)T (B − I) = kB − Ik2F
where in the third line we have used the fact that the trace of a product of matrices is preserved under cyclic
shifts.
25
This allows us to show:
Claim 4.33. Assume (13)–(15) hold with δ1 , δ2 ≤ O(δ) and δ3 ≤ O(δ log 1/ε), and assume furthermore
that kA − IkF ≥ cδ. Then, if we let δ′ = (1−ε)c
2 δ = Θ(δ), we have
X
i∈E
wi Zi − wb I ♭
+
S
X
i∈E
wi Zi − wb I ♭
S
⊥
≥ δ′ .
(16)
Proof. Combining Lemma 4.28 and Fact 4.32 we have
kA − IkF ≥ cδ ⇒ kB − IkF ≥
cδ
.
2
(17)
P
We can rewrite (13) as the expression i∈G wi Xi XiT = wg Σ1/2 (I + R)Σ1/2 where R is symmetric and
P
b we have that N wi Yi Y T = I, and so
satisfies kRkF ≤ δ1 . By the definition of Σ
i
i=1
X
i∈E
wi Yi YiT − wb I
=
F
X
i∈G
wi Yi YiT − wg I
b −1/2 Σ1/2 (I + R)Σ1/2 Σ
b −1/2 − I
= wg Σ
F
Furthermore we have
b −1/2 Σ1/2 RΣ1/2 Σ
b −1/2
Σ
F
b −1/2 ΣΣ
b −1/2
≤ δ1 Σ
2
F
,
by applying Lemma 4.29. And putting it all together we have
X
i∈E
wi Yi YiT − wb I
F
≥ wg
b −1/2 ΣΣ
b −1/2 − I
Σ
F
b −1/2 ΣΣ
b −1/2
− δ1 Σ
2
b −1/2 ΣΣ
b −1/2 − IkF ≥ cδ, then
It is easily verified that for c > 10, we have that for all δ, if kΣ
Hence all this implies that
b −1/2 ΣΣ
b −1/2 − IkF ≥ 2δkΣ
b −1/2 ΣΣ
b −1/2 k2 .
kΣ
X
i∈E
where δ′ =
c(1−ε)
2 δ
wi Yi YiT − wb I
F
≥ δ′ ,
= Θ(δ). The desired result then follows from Pythagorean theorem.
Claim 4.33 tells us that if w 6∈ Ccδ , we know that one of the terms in (17) must be at least 12 δ′ . We first
show that if the first term is large, then the algorithm outputs a separating hyperplane:
Claim 4.34. Assume that (13-(15) hold with δ1 , δ2 ≤ O(δ) and δ3 ≤ O(δ log 1/ε). Moreover, suppose that
X
i∈E
wi Zi − wb I ♭
S
1
≥ δ′ .
2
Then the algorithm outputs a hyperplane in line 7, and moreover, it is a separating hyperplane.
26
Proof. Let us first show that given these conditions,
P then the algorithm indeed outputs a hyperplane in♭ line
♭
⊥
7. Since I ∈ S , the first term is just equal to
i∈E wi Zi S . But this implies that there is some M ∈ S
♭
so that kM k2 = kM kF = 1 and so that
X
i∈E
wi hM ♭ , Zi i ≥
1 ′
δ ,
2
which implies that
X wi
1 δ′
hM ♭ , Zi i ≥
.
wb
2 wb
i∈E
The wi /wb are a set of weights satisfying the conditions of Lemma 4.30 and so this implies that
!
X
δ′ 2
♭
2
wi hM , Zi i ≥ O
wb
i∈E
!
δ′ 2
≥O
ε
(18)
e =Σ
b −1 Σ. By Theorem 4.15 and (15), we have that
Let Σ
⊗2
⊗2
X
T
T
♭
♭
⊗2
1/2
1/2
e
e
e + Σ
e
e
wi Zi Zi = wg
Σ
Σ
+ 2Σ
R Σ
,
i∈G
where kRk2 ≤ δ3 . Hence,
X
i∈G
wi Zi ZiT − 2I
= wg
S
e♭
Σ
e♭
Σ
T
⊗2
⊗2
e ⊗2 − I + (1 − wg )I + Σ
e 1/2
e 1/2
+2 Σ
R Σ
e − Ik2 + (1 − wg ) + kRkkΣk
e 2
e − Ik2 + 2kΣ
≤ kΣ
F
e − Ik2F + δkΣk
e 2 + O(ε) .
≤ 3kΣ
2
≤ O δ′ + δ′ ,
S
(19)
e 2 ≤ O(kΣ
e − IkF ) as long as kΣ
e − IkF ≥ Ω(δ), which it is by (17).
since it is easily verified that δkΣk
Equations 18 and 19 then together imply that
N
X
i=1
♭ T
wi (M )
Zi ZiT (M ♭ )
♭ T
♭
− (M ) IM ≥ O
δ2
ε
,
and so the top eigenvalue of M is greater in magnitude than λ, and so the algorithm will output a hyperplane
in line 7. Letting ℓ denote the hyperplane output by the algorithm, by the same calculation as for (19), we
must have ℓ(w∗ ) < 0, so this is indeed a separating hyperplane. Hence in this case, the algorithm correctly
operates.
Now let us assume that the first term on the LHS is less than 21 δ′ , so that the algorithm does not necessarily output a hyperplane in line 7. Thus, the second term on the LHS of Equation 16 is at least 21 δ′ . We
now show that this implies that this implies that the algorithm will output a separating hyperplane in line 9.
27
Claim 4.35. Assume that (13-(15) hold. Moreover, suppose that
X
i∈E
wi Zi − wb I ♭
S⊥
1
≥ δ′ .
2
Then the algorithm outputs a hyperplane in line 9, and moreover, it is a separating hyperplane.
Proof. By the definition of S ⊥ , the assumption implies that
√
tr(Z ♯ )
1
wi √ i − Mb d ≥ δ′ ,
2
d
i∈E
X
which is equivalent to the condition that
ξ
X
i∈E
wi
kYi k22 √
(1 − ε)δ′
√ − d ≥
,
2
d
for some ξ ∈ {−1, 1}. In particular, the algorithm will output a hyperplane
X kYi k2 √
√ 2 − d −λ
ℓ(w) = ξ
wi
d
i∈S
in Step 9, where S is some set of size at most εN , and λ = O(δ′ ). Since it will not affect anything, for
without loss of generality let us assume that ξ = 1. The other case is symmetrical.
It now suffices to show that ℓ(w∗ ) < 0. Let T = S ∩ G. By (14), we know that
X 1
e 1/2 (I + A) Σ
e 1/2 − I ,
Yi YiT − I = Σ
|T |
i∈T
N
where kAkF = O δ |T
| . Hence,
X
i∈T
|T |
1
Yi YiT −
I
(1 − ε)N
(1 − ε)N
|T |
e 1/2 (I + A) Σ
e 1/2 − I
Σ
(1 − ε)N
F
|T |
e − IkF + kAkF kΣk
e 2
≤
kΣ
(1 − ε)N
|T |
e − IkF + O(δ)kΣk
e 2
kΣ
≤
(1 − ε)N
≤ O(δδ′ + δ) ,
=
F
as long as δ′ ≥ O(δ). By self-duality of the Frobenius norm, using the test matrix
X
i∈T
√
1
kYi k2 − d ≤ O(δδ′ + δ) < α
(1 − ε)N
and hence ℓ(w∗ ) < 0, as claimed.
These two claims in conjunction directly imply the correctness of the theorem.
28
√1 I,
d
this implies that
4.4.1
The Full Algorithm
As before, this separation oracle and the classical theory of convex optimization shows that we have demonstrated an algorithm F INDA PPROX C OVARIANCE with the following properties:
Theorem 4.36. Fix ε, τ > 0, and let δ = O(ε log 1/ε). Let c > 0 be a universal constant which is sufficiently large. Let X1 , . . . , XN be an ε-corrupted set of points satisfying (13-(15), for δ1 , δ2 ≤ O(δ) and δ3 ≤
O(δ log 1/ε). Then F INDA PPROX C OVARIANCE (ε, τ, X1 , . . . , XN ) runs in time poly(N, d, 1/ε, log 1/τ ),
and outputs a u so that there is some w ∈ Ccδ so that kw − uk∞ ≤ ε/(N d log(N/τ )).
As before,P
this is not quite sufficient to actually recover the covariance robustly. Naively, we would just
T
−1/2 X k
like to output N
i 2
i=1 ui Xi Xi . However, this can run into issues if there are points Xi so that kΣ
is extremely large. We show here that we can postprocess the u so that we can weed out these points. First,
observe that we have the following lemma:
Lemma 4.37. Assume X1 , . . . , XN satisfy (13). Let w ∈ SN,ε . Then
N
X
i=1
wi Xi XiT (1 − O(δ1 ))Σ .
P
Proof. This follows since P
by (13), we have that i∈G wi Xi XiT wg (1 − δ1 )Σ (1 − O(δ1 ))Σ. The
lemma then follows since i∈E wi Xi XiT 0 always.
Now, for any set of weights w ∈ SN,ε , let w
e− ∈ RN be the vector given by w
ei− = max(0, wi −
ε/(N d log(N/τ ))), and let w− be the set of weights given by renormalizing w
e− . It is a straightforward
−
calculation that for any w ∈ SN,ε , we have w ∈ SN,2ε . In particular, this implies:
P
−
T
Lemma 4.38. Let u be so that there is w ∈ Ccδ so that ku−wk∞ ≤ ε/(N d log(N/τ )). Then, N
i=1 ui Xi Xi
(1 + O(δ))Σ.
P
Proof. By the definition of Ccδ , we must have that N
Xi XiT (1 + cδ)Σ. Moreover, we must
i=1 wiP
−
T
have u
e ≤ wi for every index i ∈ [N ]. Thus we have that N
e−
i Xi Xi (1 + cδ)Σ, and hence
i=1 u
i wP
PN i −
P
N
N
−
T
T
T
e−
i=1 ui wi Xi Xi (1 + cδ)Σ, since
i=1 ui wi Xi Xi (1 + O(ε))
i=1 u
i wi Xi Xi .
We now give the full algorithm. The algorithm proceeds as follows: first run F INDA PPROX C OVARI to get some set of weights u which is close to some element of Ccδ . We then compute the empirical
P
−1/2
T
covariance Σ1 = N
Xi k22 which
i=1 ui Xi Xi with the weights u, and remove any points which have kΣ1
are too large. We shall show that this removes no good points, and removes all corrupted points which
have kΣ−1/2 Xi k22 which are absurdly large. We then rerun F INDA PPROX C OVARIANCE with this pruned
set of points, and output the empirical covariance with the output of this second run. Formally, we give the
pseudocode for the algorithm in Algorithm 5.
We now show that this algorithm is correct.
ANCE
Theorem 4.39. Let 1/2 ≥ ε > 0, and τ > 0. Let δ = O(ε log 1/ε). Let X1 , . . . , XN be a ε-corrupted set
of samples from N (0, Σ) where
3
d log2 1/τ
e
.
N =Ω
ε2
b be the output of L EARN C OVARIANCE (ε, τ, X1 , . . . , XN ). Then with probability 1−τ , kΣ−1/2 ΣΣ
b −1/2 −
Let Σ
IkF ≤ O(δ).
29
Algorithm 5 Full algorithm for learning the covariance agnostically
1: function L EARN C OVARIANCE(ε, τ, X1 , . . . , XN )
2:
Let u ← FP
INDA PPROX C OVARIANCE (ε, τ, X1 , . . . , XN ).
−
T
3:
Let Σ1 = N
i=1 ui Xi Xi .
4:
for i = 1, . . . , N do
−1/2
5:
if kΣ1 Xi k22 ≥ Ω(d log N/τ ) then
6:
Remove Xi from the set of samples
7:
Let I be the set of pruned samples.
8:
Let u′ ←
INDA PPROX C OVARIANCE (ε, τ, {Xi }i∈I ).
PF
N
9:
return i=1 u′i Xi XiT .
Proof. We first condition on the event that we satisfy (12)-(15) with δ1 , δ2 ≤ O(δ) and δ3 ≤ O(δ log 1/ε).
By our choice of N , Fact 4.8, Corollary 4.10, Corollary 4.12, and Theorem 4.17, and a union bound, we
know that this event happens with probability 1 − τ .
By Theorem 4.36, Lemma 4.37, and Lemma 4.38, we have that since ε is sufficiently small,
1
Σ Σ1 2Σ .
2
In particular, this implies that for every vector Xi , we have
1 −1/2
−1/2
kΣ
Xi k22 ≤ kΣ1 Xi k22 ≤ 2kΣ−1/2 Xi k22 .
2
Therefore, by (12), we know that in line 6, we never throw out any uncorrupted points, and moreover, if Xi
is corrupted with kΣ−1/2 Xi k22 ≥ Ω(d log N/τ ), then it is thrown out. Thus, let I be the set of pruned points.
Because no uncorrupted point is thrown out, we have that |I| ≥ (1 − 2ε)N , and moreover, this set of points
still satisfies (13)-(15)3 and moreover, for ever i ∈ I, we have kΣ−1/2 Xi k22 ≤ O(d log N/τ ). Therefore, by
Theorem 4.36, we have that there is some u′′ ∈ Cc|I| so that ku′ − u′′ k∞ < ε/(N d log(N/τ )). But now if
′
T
b=P
Σ
i∈|I| ui Xi Xi , we have
b −1/2 − IkF ≤
kΣ−1/2 ΣΣ
X
i∈I
u′′i Σ−1/2 Xi XiT Σ−1/2 − I
≤ cδ + O(ε) ≤ O(δ) ,
+
F
X
i∈I
|u′i − u′i |kΣ−1/2 Xi k22
which completes the proof.
4.5 Learning an Arbitrary Gaussian Agnostically
We have shown how to agnostically learn the mean of a Gaussian with known covariance, and we have
shown how to agnostically learn the covariance of a mean zero Gaussian. In this section, we show how to
use these two in conjunction to agnostically learn an arbitrary Gaussian. Throughout, let X1 , . . . , XN be an
ε-corrupted set of samples from N (µ, Σ), where both µ and Σ are unknown. We will set
3
d log2 1/τ
e
N =Ω
.
ε2
3
Technically, the samples satisfy a slightly different set of conditions since we may have thrown out some corrupted points, and
so in particular the number of samples may have changed, but the meaning should be clear.
30
4.5.1
From Unknown Mean, Unknown Covariance, to Zero Mean, Unknown Covariance
We first show a simple trick which, at the price of doubling the amount of error, allows us to assume that
the mean is zero, without
√ changing the covariance. We do so as follows: for each i = 1, . . . , N/2, let
Xi′ = (Xi − XN/2+i )/ 2. Observe that if both Xi and XN/2+i are uncorrupted, then Xi′ ∼ N (0, Σ).
Moreover, observe that Xi′ is corrupted only if either Xi or XN/2+i is corrupted. Then we see that if
′
X1 , . . . , XN is ε-corrupted, then the X1′ , . . . , XN/2
is a N/2-sized set of samples which is 2ε-corrupted.
b so that
Thus, by using the results from Section 4.4, with probability 1 − τ , we can recover a Σ
b −1/2 − IkF ≤ O(ε log 1/ε) ,
kΣ−1/2 ΣΣ
(20)
b N (0, Σ)) ≤ O(ε log 1/ε) .
dTV (N (0, Σ),
(21)
which in particular by Corollary 2.14, implies that
4.5.2
From Unknown Mean, Approximate Covariance, to Approximate Recovery
b −1/2 Xi . Then, for Xi which is not corrupted, we have that X ′′ ∼ N (0, Σ1 ), where
For each Xi , let Xi′′ = Σ
i
−1/2
−1/2
b
e
Σ1 = Σ
ΣΣ
. By Corollary 2.14 and Lemma 4.28, if (20) holds, then we have
b −1/2 µ, Σ1 ), N (Σ
b −1/2 µ, I)) ≤ O(ε log 1/ε) .
dTV (N (Σ
By Claim 2.4, this means that if (20) holds, the uncorrupted set of Xi′′ can be treated as an O(ε log 1/ε)b −1/2 µ, I). Thus, if (20) holds, the entire set of samples X ′′ , . . . , X ′′ is
corrupted set of samples from N (Σ
m
1
b −1/2 µ, I). Then, by using results from Section 4.3, with
a O(ε log 1/ε)-corrupted set of samples from N (Σ
b −1/2 µk2 ≤ O(ε log3/2 (1/ε)).
probability 1 − τ , assuming that 20 holds, we can recover a µ
b so that kb
µ−Σ
Thus, by Corollary 2.13, this implies that
or equivalently,
b −1/2 µ, I)) ≤ O(ε log 3/2 (1/ε)) ,
dTV (N (b
µ, I), N (Σ
b 1/2 µ
b N (µ, Σ))
b ≤ O(ε log3/2 (1/ε)) ,
dTV (N (Σ
b, Σ),
which in conjunction with (21), implies that
b 1/2 µ
b N (µ, Σ)) ≤ O(ε log3/2 (1/ε)) ,
dTV (N (Σ
b, Σ),
and thus by following this procedure, whose formal pseudocode is given in Algorithm 6, we have shown the
following:
Algorithm 6 Algorithm for learning an arbitrary Gaussian robustly
1: function R ECOVER ROBUST G UASSIAN (ε, τ, X1 , . . . , XN )
√
2:
For i = 1, . . . , N/2, let Xi′ = (Xi − XN/2+i )/ 2.
b ← L EARN C OVARIANCE (ε, τ, X ′ , . . . , X ′ ).
3:
Let Σ
1
N/2
b −1/2 Xi .
4:
For i = 1, . . . , N , let Xi′′ = Σ
′′ ).
5:
Let µ
b ← L EARN M EAN (ε, τ, X1′′ , . . . , XN
−1/2
b
b
6:
return the Gaussian with mean Σ
µ
b, and covariance Σ.
31
Theorem 4.40. Fix ε, τ > 0. Let X1 , . . . , XN be an ε-corrupted set of samples from N (µ, Σ), where µ, Σ
are both unknown, and
3
d log2 1/τ
e
N =Ω
.
ε2
There is a polynomial-time algorithm R ECOVER ROBUST G AUSSIAN (ε, τ, X1 , . . . , XN ) which with probab µ
bility 1 − τ , outputs a Σ,
b so that
b 1/2 µ
b N (µ, Σ)) ≤ O(ε log3/2 (1/ε)) .
dTV (N (Σ
b, Σ),
5 Agnostically Learning a Mixture of Spherical Gaussians, via Convex Programming
In this section, we give an algorithm to agnostically learn a P
mixture of k Gaussians with identical spherical
√
e
covariance matrices up to error O(poly(k)
· ε). Let M = j∈[k] αj N (µj , σ 2 I) be the unknown k-GMM
each of whose components are spherical. For X ∼ M, we write X ∼j M if X was drawn from the jth
component of M.
Our main result of this section is the following theorem:
Theorem
P 5.1. Fix ε, τ > 0, and k ∈ N. Let X1 , . . . , XN be an ε-corrupted set of samples from a k-GMM
M = j∈[k] αj N (µj , σj2 I), where all αj , µj , and σj2 are unknown, and
e (poly (d, k, 1/ε, log(1/τ ))) .
N =Ω
There is an algorithm which with probability 1 − τ , outputs a distribution M′ such that
√
e
dTV (M, M′ ) ≤ O(poly(k)
· ε) .
2
The running time of the algorithm is poly(d, 1/ε, log(1/τ ))k .
Our overall approach will be a combination of our method for agnostically learning a single Gaussian
and recent work on properly learning mixtures of multivariate spherical Gaussians [SOAJ14, LS15]. At a
high level, this recent work relies upon the empirical covariance matrix giving an accurate estimate of the
overall covariance matrix in order to locate the subspace in which the component mean lie. However, as
we have observed already, the empirical moments do not necessarily give good approximations of the true
moments in the agnostic setting. Therefore, we will use our separation oracle framework to approximate the
covariance matrix, and the rest of the arguments follow similarly as previous methods.
The organization of this section will be as follows. We define some of the notation we will be using
and the Schatten top-k norm in Section 5.1. Section 5.2 states the various concentration inequalites we
require. In Section 5.3, we go over our overall algorithm in more detail. Section 5.4 describes a first naive
clustering step, which deals with components which are very well separated. Section 5.5 contains details
on our separation oracle approach, allowing us to approximate the true covariance. Section 5.6 describes
our spectral clustering approach to cluster components with means separated more than Ωk (log 1/ε). In
Section 5.7, we describe how to exhaustively search over a particular subspace to obtain a good estimate
for the component means. In Section 5.8, we go over how to limit the set of hypotheses in order to satisfy
the conditions of Lemma 2.9. For clarity of exposition, all of the above describe the algorithm assuming
all σj2 are equal. In Section 5.9, we discuss the changes to algorithm which are required to handle unequal
variances.
For clarity of presentation, many of the proofs are deferred to Section C.
32
5.1 Notation and Norms
Recall the definitionPof SN,ε from Section 4.1, which we will use extensively in this section. We will use
the notation µ =
j∈[k] αj µj to denote the mean of the unknown GMM. Also, we define parameters
2
γj = αj kµj − µk2 and let γ = maxj γj . And for ease of notation, let
f (k, γ, ε) = k1/2 ε + kγ 1/2 ε + kε2 and h(k, γ, ε) = k1/2 ε + kγ 1/2 ε + kγε + kε2 = f (k, γ, ε) + kγε.
Finally, we use the notation
Q=
X
j∈[k]
αj (µj − µ)(µj − µ)T .
(22)
to denote the covariance of the unknown GMM. Our algorithm for learning spherical k-GMMs will rely
heavily on the following, non-standard norm:
Definition 5.2. For any symmetric matrix M ∈ Rd×d with singular values σ1 ≥ σ2 ≥ . . . σd , let the
Schatten top-k norm be defined as
k
X
σi ,
kM kTk =
i=1
i.e., it is the sum of the top-k singular values of M .
It is easy to see that the Schatten top-k norm is indeed a norm, as its name suggests:
Fact 5.3. kM kTk is a norm on symmetric matrices.
Proof. The only property that is not immediate, that we need to verify, is that it satisfies the triangle inequality. Indeed, we can lower bound kM kTk in terms of its dual characterization
√
kM kTk = max Tr(X T M T M X) ,
X∈Rd×k
where the maxima is taken over all X with orthonormal columns. And now the fact that the triangle inequality holds is immediate.
5.2 Concentration Inequalities
In this section, we will establish some concentration inequalities that we will need for our algorithm for
agnostically learning mixtures of spherical Gaussians. Recall the notation as described in Section 5.1. The
following two concentration lemmas follow from the same proofs as for Lemmas 42 and 44 in [LS15].
P
Lemma 5.4. Fix ε, δ > 0. If Y1 , . . . , YN are independent samples from the GMM with PDF j∈[k] αj N (µj , Σj )
where αj ≥ Ω(ε) for all j, and N = Ω d+logε2(k/δ) then with probability at least 1 − O(δ),
N
1 X
(Yi − µ)(Yi − µ)T − I − Q
N
i=1
2
≤ O (f (k, γ, ε)) ,
where Q is defined as in equation (22).
P
Lemma 5.5. Fix ε, δ > 0. If Y1 , . . . , YN are independent samples from the GMM with PDF j∈[k] αj N (µj , Σj )
where αj ≥ Ω(ε) for all j, and N = Ω d+logε2(k/δ) then with probability at least 1 − O(δ),
N
1 X
Yi − µ
N
i=1
2
33
≤ O k1/2 ε .
From the same techniques as before, we get the same sort of union bounds as usual over the weight
vectors:
p
Lemma 5.6. Fix ε ≤ 1/2 and τ ≤ 1. There
P is a δ = O(ε log 1/ε) so that if Y1 , . . . , YN are independent samples from the GMM with PDF j∈[k] αj N (µj , Σj ) where αj ≥ Ω(ε) for all j, and N =
Ω d+logδ2(k/τ ) , then
1
"
Pr ∃w ∈ SN,ε :
N
X
i=1
T
wi (Yi − µ)(Yi − µ) − I − Q
2
#
≥ f (k, γ, δ1 ) ≤ τ ,
(23)
where Q is defined as in equation (22).
p
Lemma 5.7. Fix ε ≤ 1/2 and τ ≤ 1. There
P is a δ = O(ε log 1/ε) so that if Y1 , . . . , YN are independent samples from the GMM with PDF j∈[k] αj N (µj , Σj ) where αj ≥ Ω(ε) for all j, and N =
Ω d+logδ2(k/τ ) , then
2
#
"
N
X
1/2
(24)
Pr ∃w ∈ SN,ε :
wi Yi − µ ≥ k δ2 ≤ τ.
i=1
2
5.3 Algorithm
Our approach is based on a tournament, as used in several recent works [DK14, SOAJ14, DDS15a, DDS15b,
DKT15, DDKT16]. We will generate a list S of candidate hypotheses (i.e. of k-GMMs) of size |S| =
2
poly(d, 1/ε, log(1/τ ))k with the guarantee that there is some M∗ ∈ S such that dTV (M, M∗ ) ≤ Õ(poly(k)·
√
ε). We then find (roughly) the best candidate hypothesis on the list. It is most natural to describe the algorithm as performing several layers of guessing. We will focus our discussion on the main steps in our
analysis, and defer a discussion of guessing the mixing weights, the variance σ 2 and performing naive clustering until later. For reasons we justify in Section 5.8, we may assume that the mixing weights and the
variance are known exactly, and that the variance σ 2 = 1.
Our algorithm is based on the following deterministic conditions:
|{Xi ∈ G, Xi ∼j M : kXi − µj k22 ≥ Ω(d log k/ε)}|
≤ ε/k , ∀j = 1, . . . , N
|{Xi ∈ G, Xi ∼j M}|
X
wi (Xi − µ)(Xi − µ)T − wg I − wg Q ≤ f (k, γ, δ1 ) ∀w ∈ SN,4ε , and
i∈G
(25)
(26)
2
X
i∈G
wi (Xi − µ)
2
≤ k1/2 δ2 ∀w ∈ SN,4ε .
(27)
(25) follows from basic Gaussian concentration, and (26) and (27) follow from the results in Section 5.2 for
N sufficiently large. Note that these trivially imply similar conditions for the Schatten top-k norm, at the
cost of an additional factor of k on the right-hand side of the inequalities. For the rest of this section, let
δ = max(δ1 , δ2 ).
At this point, we are ready to apply our separation oracle framework. In particular, we will find a weight
vector w over the points such that
N
X
i=1
wi (Xi − µ)(Xi − µ)T − I −
X
j∈[k]
34
αj (µj − µ)(µj − µ)T
≤ η,
2
for some choice of η. The set of such weights is convex, and concentration implies that the true weight
vector will have this property. Furthermore, we can describe a separation oracle given any weight vector
not contained in this set (as long as η is not too small). At this point, we use classical convex programming
methods to find a vector which satisfies these conditions. Further details are provided in Section 5.5.
After this procedure, Lemma 5.13 shows that the weighted empirical covariance is spectrally close to
the true covariance matrix. We are now in the same regime as [SOAJ14], which obtains their results as
a consequence of the empirical covariance concentrating about the true covariance matrix. Thus, we will
appeal to their analysis, highlighting the differences between our approach and theirs. We note that [LS15]
also follows a similar approach and the interested reader may also adapt their arguments instead.
First, if γ is sufficiently large (i.e., Ωk (log(1/ε))), this implies a separation condition between some
component mean and the mixture’s mean. This allows us to cluster the points further, using a spectral
method. We take the top eigenvector of the weighted empirical covariance matrix and project in this direction, using the sign of the result as a classifier. In contrast to previous work, which requires that no points are
misclassified, we can tolerate poly(ε/k) misclassifications, since our algorithms are agnostic. This crucially
allows us to avoid a dependence on d in our overall agnostic learning guarantee. Further details are provided
in Section 5.6.
Finally, if γ is sufficiently small, we may perform an exhaustive search. The span of the means is the top
k − 1 eigenvectors of the true covariance matrix, which we can approximate with our weighted empirical
covariance matrix. Since γ is small, by trying all points within a sufficiently tight mesh, we can guess a
set of candidate means which are sufficiently close to the true means. Combining the approximations to the
means with Corollary 2.13 and the triangle inequality, we can guarantee that at least one of our guesses is
sufficiently close to the true distribution. Additional details are provided in Section 5.7.
To conclude our algorithm, we can apply Lemma 2.9. We note that this hypothesis selection problem
has been studied before (see, i.e. [DL01, DK14]), but we must adapt it for our agnostic setting. This allows
us to select a hypothesis which is sufficiently close to the true distribution, thus concluding the proof. We
note that the statement of Lemma 2.9 requires the hypotheses to come from some fixed finite set, while there
are an infinite number of Gaussian mixture models. In Section 5.8, we discuss how to limit the number of
hypotheses based on the set of uncorrupted samples in order to satisfy the conditions of Lemma 2.9.
5.4 Naive Clustering
We give a very naive clustering algorithm, the generalization of NAIVE P RUNE , which recursively allows us
to cluster components if they are extremely far away. The algorithm is very simple: for each Xi , add all
points within distance O(d log(k/ε)) to a cluster Si . Let C be the set of clusters which contain at least 4εN
points, and let the final clustering be C1 , . . . , Ck′ be formed by merging clusters in C if they overlap, and
stopping if no clusters overlap. We give the formal pseudocode in Algorithm 7.
Algorithm 7 Naive clustering algorithm for spherical GMMs.
1: function NAIVE C LUSTER GMM(X1 , . . . , Xn )
2:
for i = 1, . . . , N do
3:
Let Si = {i′ : kXi − Xi′ k22 ≤ Θ(dk log 1/ε)}.
Let C = {Si : |Si | ≥ 4εN }.
4:
while ∃C, C ′ ∈ C so that C 6= C ′ and C ∪ C ′ 6= ∅ do
5:
Remove C, C ′ from C
6:
Add C ∪ C ′ to C
7:
return the set of clusters C
We prove here that this process (which may throw away points) throws away only at most a ε fraction of
35
good points, and moreover, the resulting clustering only misclassifies at most an O(ε)-fraction of the good
points, assuming (25).
Theorem 5.8. Let X1 , . . . , Xm be a set of samples satisfying (25). Let C1 , . . . , Ck′ be the set of clusters
returned. For each component j, let ℓ(j) be the ℓ so that Cℓ contains the most points from j. Then:
1. Then, for each ℓ, there is some j so that ℓ(j) = ℓ.
2. For all j, we have
|{Xi ∈ G, Xi ∼j M}| − |{Xi ∈ G, Xi ∼j M, Xi ∈ Cℓ(j) }| ≤ O
ε
k
|{Xi ∈ G, Xi ∼j M}| .
3. For all j, j ′ , we have that if ℓ(j) = ℓ(j ′ ), then kµj − µj ′ k22 ≤ O(dk log k/ε)
4. If Xi , Xj ∈ Cℓ , then kXi − Xj k22 ≤ dk log 1/ε.
Thus, we have that by applying this algorithm, given an ε-corrupted set of samples from M, we may
cluster them in a way which misclassifies at most an ε/k fraction of the samples from any component, and
so that within each cluster, the means of the associated components differ by at most dk log k/ε. Thus, each
separate cluster is simply a ε-corrupted set of samples from the mixture restricted to the components within
that cluster; moreover, the number of components in each cluster must be strictly smaller than k. Therefore,
we may simply recursively apply our algorithm on these clusters to agnostically learn the mixture for each
cluster, since if k = 1, this is a single Gaussian, which we know how to learn agnostically.
Thus, for the remainder of this section, let us assume that for all j, j ′ , we have kµj −µ′j k22 ≤ O(dk log 1/ε).
Moreover, we may assume that there are no points j, j ′ (corrupted or uncorrupted), so that kXj − Xj ′ k22 ≥
Ω(dk log 1/ε).
5.5 Estimating the Covariance Using Convex Programming
In this section, we will apply our separation oracle framework to estimate the covariance matrix. While in the
non-agnostic case, the empirical covariance will approximate the actual covariance, this is not necessarily
true in our case. As such, we will focus on determining a weight vector over the samples such that the
weighted empirical covariance is a good estimate for the true covariance.
We first define the convex set for which we want an interior point:
N
X
X
Cη = w ∈ SN,ε :
wi (Xi − µ)(Xi − µ)T − I −
αj (µj − µ)(µj − µ)T ≤ η .
i=1
j∈[k]
2
In Section 5.5.1, we prove lemmas indicating important properties of this set. In Section 5.5.2, we give
a separation oracle for this convex set. We conclude with Lemma 5.13, which shows that we have obtained
an accurate estimate of the true covariance.
5.5.1
Properties of Our Convex Set
We start by proving the following lemma, which states that for any weight vector which is not in our set, the
weighted empirical covariance matrix is noticeably larger than it should be (in Schatten top-k norm).
Lemma 5.9. Suppose that (26) holds, and w 6∈ Cckh(k,γ,δ). Then
N
X
i=1
wi (Xi − µ)(Xi − µ)T − I
36
Tk
≥
X
j∈[k]
γj +
3ckh(k, γ, δ)
.
4
We also require the following lemma, which shows that if a set of weights poorly approximates µ, then
it is not in our convex set.
P
Lemma 5.10. Suppose that (26) and (27) hold. Let w ∈ Sm,ε and set µ
b= m
b.
i=1 wi Xi and ∆ = µ − µ
Furthermore, suppose that k∆k2 ≥ Ω(h(k, γ, δ)). Then
N
X
i=1
T
wi (Xi − µ)(Xi − µ) − I −
X
j∈[k]
T
≥Ω
αj (µj − µ)(µj − µ)
2
k∆k22
ε
.
By contraposition, if a set of weights is in our set, then it provides a good approximation for µ:
Corollary 5.11. Suppose that (26) and (27) hold. Let w ∈ Ch(k,γ,δ) for δ = Ω(ε log 1/ε). Then
5.5.2
Separation Oracle
p
k∆k2 ≤ O(ε log 1/ε).
In this section, we provide a separation oracle for Cη . In particular, we have the following theorem:
Theorem 5.12. Fix ε > 0, and let δ = Ω(ε log 1/ε). Suppose that (26) and (27) hold. Let w∗ denote the
weights which are uniform on the uncorrupted points. Then there is a constant c and an algorithm so that:
1. (Completeness) If w = w∗ , then it outputs “YES”.
2. (Soundness) If w 6∈ Cckh(k,γ,δ), the algorithm outputs a hyperplane ℓ : Rm → R so that ℓ(w) ≥ 0 but
ℓ(w∗ ) < 0.
These two facts imply that the ellipsoid method with this separation oracle will terminate in poly(d, 1/ε)
steps, and moreover, will with high probability output a w′ so that kw − w′ k∞ ≤ ε/(N dk log 1/ε) for some
w ∈ Cckh(k,γ,δ). Moreover, it will do so in polynomially many iterations.
The proof is deferred to Section C.1.
Algorithm 8 Separation oracle sub-procedure for agnostically learning the span of the means of a GMM.
1: function S EPARATION O RACLE GMM(w)
P
2:
Let µ
b= N
i=1 wi Xi .
3:
For i = 1,P
. . . , N , define Yi = Xi − µ
b.
N
T
4:
Let M = i=1 wi Yi Yi − I.
P
5:
if kM kTk < j∈[k] γj + ckh(k,γ,δ)
then
2
6:
return “YES”.
7:
else
8:
Let Λ = kM kTk .
9:
Let U be a d × k matrix with orthonormal
the top k eigenvectors of M .
Pcolumns which span
N
T
T
10:
return the hyperplane ℓ(w) = Tr U
i=1 wi Yi Yi − I U − Λ > 0
After running this procedure, we technically do not have a set of weights in Cckh(k,γ,δ) . But by the
same argument as in Section 4.3, because the maximum distance between two points within any cluster is
37
bounded, and we have the guarantee that kXi − Xj k2 ≤ O(dk log 1/ε) for all i, j, we may assume we have
a set of weights satisfying
N
X
i=1
wi (Xi − µ)(Xi − µ)T − I −
X
j∈[k]
αj (µj − µ)(µj − µ)T
≤ 2ckh(k, γ, δ).
2
We require the following lemma, describing the accuracy of the empirical covariance matrix with the
obtained weights.
P
Lemma 5.13. Let µ
b= N
i=1 wi Xi . After running the algorithm above, we have a vector w such that
N
X
i=1
wi (Xi − µ
b)(Xi − µ
b)T − I −
X
j∈[k]
αj (µj − µ)(µj − µ)T
≤ 3ckh(k, γ, δ).
2
Proof. By triangle inequality and Corollary 5.11,
N
X
i=1
≤
wi (Xi − µ
b)(Xi − µ
b)T − I −
N
X
i=1
X
j∈[k]
wi (Xi − µ)(Xi − µ)T − I −
αj (µj − µ)(µj − µ)T
X
j∈[k]
2
+ k∆k22
αj (µj − µ)(µj − µ)T
2
≤ 2ckh(k, γ, δ) + O(δ) ≤ 3ckh(k, γ, δ)
5.6 Spectral Clustering
Now that we have a good estimate of the true covariance matrix, we will perform spectral clustering while
γ is sufficiently large. We will adapt Lemma 6 from [SOAJ14], giving the following lemma:
Lemma 5.14. Given a weight vector w as output by Algorithm 8, if γ ≥ Ω(poly(k) · log 1/ε), there exists
an algorithm which produces a unit vector v with the following guarantees:
• There exists a non-trival partition of [k] into S0 and S1 such that v T µj > 0 for all j ∈ S0 and
v T µj < 0 for all j ∈ S1 ;
• The probability of a sample being misclassified is at most O(poly(ε/k)), where a misclassification is
defined as a sample X generated from a component in S0 having v T X < 0, or a sample generated
from a component in S1 having v T X > 0.
The algorithm will be as follows. Let v be the top eigenvector of
N
X
i=1
wi (Xi − µ
b)(Xi − µ
b)T − I.
For a sample X, cluster it based on the sign of v T X. After performing this clustering, recursively perform
our algorithm from the start on the two clusters.
The proof is very similar to that of Lemma 6 in [SOAJ14]. Their main concentration lemma is Lemma
30, which states that they obtain a good estimate of the true covariance matrix, akin to our Lemma 5.13.
38
Lemma 31 argues that the largest eigenvector of this estimate is highly correlated with the top eigenvector
of the true covariance matrix. Since γ is large, this implies there is a large margin between the mean and the
hyperplane. However, by standard Gaussian tail bounds, the probability of a sample landing on the opposite
side of this hyperplane is small.
We highlight the main difference between our approach and theirs. For their clustering step, they require
that no sample is misclustered with high probability. As such, they may perform spectral clustering while
γ = Ω (poly(k) · log(d/ε)). We note that, in the next step of our algorithm, we will perform an exhaustive
search. This will result in an approximation which depends on the value of γ at the start of the step,
and as such, using the same approach as them would result in an overall approximation which depends
logarithmically on the dimension.
We may avoid paying this cost by noting that our algorithm is agnostic. They require that no sample
is misclustered with high probability, while our algorithm tolerates that a poly(ε/k)-fraction of points are
misclustered. As such, we can continue spectral clustering until γ = O (poly(k) · log(1/ε)).
5.7 Exhaustive Search
The final stage of the algorithm is when we know that all γi ’s are sufficiently small. We can directly apply
Lemma 7 of [SOAJ14], which states the following:
Lemma 5.15. Given a weight vector w as output by Algorithm 8, then the projection of
space orthogonal to the span of the top k − 1 eigenvectors of
N
X
i=1
has magnitude at most
wi (Xi − µ
b)(Xi − µ
b)T − I
µj −µ
kµj −µk2
onto the
2
!
√
p
ε
log(1/ε)
1/2
O poly(k) · h(k, γ, δ)/γi
.
= O poly(k) ·
1/2
γi
At this point, our algorithm is identical to the exhaustive search of [SOAJ14]. We find the span of the
top k − 1 eigenvectors by considering the (k − 1)-cube with side length 2γ centered at µ
b. By taking an
η-mesh over the points in this cube (for η = poly(ε/dk) sufficiently small), we obtain a set of points M̃ .
Via identical arguments as in the proof of Theorem 8 of [SOAJ14], for each j ∈ [k], there exists some point
µ̃j ∈ M̃ such that
√
ε log(1/ε)
.
kµ̃j − µj k2 ≤ O poly(k) ·
√
αi
By taking a k-wise Cartesian product of this set, we are guaranteed to obtain a vector which has this guarantee simultaneously for all µj .
5.8 Applying the Tournament Lemma
In this section, we discuss details about how to apply our hypothesis selection algorithm. First, in Section
5.8.1, we describe how to guess the mixing weights and the variance of the components. Then in Section
5.8.2, we discuss how to ensure our hypotheses come from some fixed finite set, in order to deal with
technicalities which arise when performing hypothesis selection with our adversary model.
39
5.8.1
Guessing the Mixing Weights and Variance
The majority of our algorithm is focused on generating guesses for the means of the Gaussians. In this
section, we guess the remaining parameters: the mixing weights and the variance. While most of these
guessing arguments are standard, we emphasize that we reap an additional benefit because our algorithm
is agnostic. In particular, most algorithms must deal with error incurred due to misspecification of the
parameters. Since our algorithm is agnostic, we can pretend the misspecified parameter is the true one, at
the cost of increasing the value of the agnostic parameter ε. If our misspecified parameters are accurate
enough, the agnostic learning guarantee remains unchanged.
Guessing the mixing weights is fairly straightforward. For some ν = poly(ε/k) sufficiently small, our
algorithm generates a set of at most (1/ν)k = poly(k/ε)k possible mixing weights by guessing the values
{0, ε, ε + ν, ε + 2ν, . . . , 1 − ν, 1} for each αj . Note that we may assume each weight is at least ε, since
components with weights less than this can be specified arbitrarily at a total cost of O(kε) in total variation
distance.
Next, we need to guess the variance σ 2 of the components. To accomplish this, we will take k + 1
samples (hoping to find only uncorrupted ones) and compute the minimum distance between any pair of
them. Since we assume k << 1/ε, we can repeatedly draw k + 1 samples until we have the guarantee that
at least one set is uncorrupted. If none of the k+1 samples are corrupted, then at least two of them came from
the same component, and in our high-dimensional
setting the distance between any pair of samples from
√
the same component concentrates around 2dσ. After rescaling this distance, we can then multiplicatively
enumerate around this value with granularity poly(ε/dk) to get an estimate for σ 2 that is sufficiently good
for our purposes. Applying Corollary 2.14 bounds the cost of this misspecification by O(ε). By rescaling
the points, we may assume that σ 2 = 1.
5.8.2
Pruning Our Hypotheses
In this section, we describe how to prune our set of hypotheses in order to apply Lemma 2.9. Recall that
this lemma requires our hypotheses to come from some fixed finite set, rather than the potentially infinite
set of GMM hypotheses. We describe how to prune and discretize the set of hypotheses obtained during the
rest of the algorithm to satisfy this condition. For the purposes of this section, a hypothesis will be a k-tuple
of d-dimensional points, corresponding only to the means of the components. While the candidate mixing
weights already come from a fixed finite set (so no further work is needed), the unknown variance must be
handled similarly to the means. The details for handling the variance are similar to (and simpler than) those
for handling the means, and are omitted.
More precisely, this section will describe a procedure to generate a set of hypotheses M, which is
exponentially large in k and d, efficiently searchable, and comes from a finite set of hypotheses which
are fixed with respect to the true distribution. Then, given our set of hypotheses generated by the main
algorithm (which is exponentially large in k but polynomial in d), we iterate over this set, either replacing
each hypothesis with a “close” hypothesis from M (i.e., one which is within O(ε) total variation distance),
or discarding the hypothesis if none exists. Finally, we run the tournament procedure of Lemma 2.9 on the
resulting set of hypotheses.
At a high level, the approach will be as follows. We will take a small set of samples, and remove any
samples from this set which are clear outliers (due to having too few nearby neighbors). This will give us a
set of points, each of which are within a reasonable distance from some component mean. Taking a union
of balls around these samples will give us a space that is a subset of a union of (larger) balls centered at the
component centers. We take a discrete mesh over this space to obtain a fixed finite set of possible means,
and round each hypothesis so that its means are within this set.
We start by taking N = O(k log(1/τ )/ε2 ) samples, which is sufficient to ensure that the number of
40
(uncorrupted) samples from component j will be (wj ± Θ(ε))N for all j ∈ [k] with probability 1 − O(τ ).
Recall that we are assuming that wj = Ω(ε) for all j, as all other components may be defined arbitrarily at
the cost of O(kε) in total variation distance. This implies that even after corruption, each component has
generated at least εN uncorrupted samples.
By standard Gaussian concentration bounds, we know that if N samples are taken from
p a Gaussian, the
maximum distance between a sample and the Gaussian’s mean will be at most ζ = O( d log(N/τ )) with
probability 1−τ . Assume this condition holds, and thus each component’s mean will have at least εN points
within distance ζ. We prune our set of samples by removing any point with fewer than εN other points at
distance less than 2ζ. This will not remove any uncorrupted points, by the above assumption, and triangle
inequality. However, this will remove any corrupted points at distance at least 3ζ from all component means,
due to the fact that the adversary may only move an ε-fraction of the points, and reverse triangle inequality.
Now, we consider the union of the balls of radius 3ζ centered at each of the remaining points. This set
contains all of the component means, and is also a subset of the union of the balls of radius 6ζ centered at the
ε
component means. We discretize this set by taking its intersection with a lattice of side-length √
. We note
k d
that any two points in this discretization are at distance at most ε/k. By a volume argument, the number of
√ d
points in the intersection is at most k 12ζkε d . Each hypothesis will be described by the k-wise Cartesian
√ kd
product of these points, giving us a set M of at most kk 12ζkε d
hypotheses.
Given a set of hypotheses H from the main algorithm, we prune it using M as a reference. For each
h ∈ H, we see if there exists some h′ ∈ M such that the means in h are at distance at most ε/k from the
corresponding means in h′4 . If such an h′ exists, we replace h with h′ – otherwise, h is simply removed. By
Corollary 2.13 and the triangle inequality, this replacement incurs a cost of O(ε) in total variation distance.
At this point, the conditions of Lemma 2.9 are satisfied and we may run this procedure to select a sufficiently
accurate hypothesis.
5.9 Handling Unequal Variances
In this section, we describe the changes required to allow the algorithm to handle different variances for
the Gaussians. The main idea is to find the minimum variance of any component and perform clustering
so we only have uncorrupted samples from Gaussians with variances within some known, polynomiallywide interval. This allows us to grid within this interval in order to guess the variances, and the rest of the
algorithm proceeds with minor changes.
The first step is to locate the minimum variance of any component. Again using standard Gaussian
concentration, in sufficiently high dimensions, if N samples are taken√
from a Gaussian with variance σ 2 I,
the distance between any two samples will be concentrated around σ( 2d − Θ(d1/4 )). With this in hand,
we use the following procedure to estimate the minimum variance. For each sample i, record the√distance
to the (εN + 1)st closest sample. We take the (εN + 1)st smallest of these values, rescale it by 1/ 2d, and
similar to before, guess around it using a multiplicative (1+poly(ε/kd)) grid, which will give us an estimate
2
σ̂min
for the smallest variance. We note that discarding the smallest εN fraction of the points prevents this
2
statistic from being grossly corrupted by the adversary. For the remainder of this section, assume that σmin
is known exactly.
At this point, we partition the points into those that come from components with small variance, and
those with large variance. We will rely upon the following concentration inequality from [SOAJ14], which
gives us the distance between samples from different components:
4
We observe that the complexity of this step is polynomial in d and k, not exponential, if one searches for the nearest lattice
point in the sphere surrounding each unpruned sample, rather than performing a naive linear scan over the entire list.
41
Lemma 5.16 (Lemma 34 from [SOAJ14]). Given N samples from a collection of Gaussian distributions,
with probability 1 − O(τ ), the following holds for every pair of samples X, Y :
s
2
N
log τ
,
kX − Y k22 ∈ d(σ12 + σ22 ) + kµ1 − µ2 k22 1 ± 4
d
where X ∼ N (µ1 , σ12 I) and Y ∼ N (µ2 , σ22 I).
Assume the event that this condition holds. Now, let
the set of all points with at least εN points
r Hℓ be !
2
ℓ−1 2
log Nτ
σmin 1 + 4
at squared-distance at most 2 1 + k1
, for ℓ ∈ [k]. Note that Hℓ ⊆ Hℓ+1 . Let
d
ℓ∗ be the minimum ℓ such that Hℓ = Hℓ+1 , or k if no such ℓ exists, and partition the set of samples into
Hℓ∗ and H ℓ∗ . This partition will contain all samples from components with variance at most some threshold
2
t, where t ≤ eσmin
in Hℓ∗ . All samples from components with variance at least t will fall into H ℓ∗ . We
continue running the algorithm with Hℓ∗ , and begin the algorithm recursively on H ℓ∗ 5
This procedure works due to the following argument. When we compute H1 , we are guaranteed that it
2 , by the upper bound in Lemma 5.16. However,
will contain all samples from components with variance σmin
2 , for
it may also contain samples from other components – in particular, those with variance at most γσmin
s
s
2
2
log Nτ
log Nτ
1 − 4
≤ 1 + 1,
γ ≤ 1 + 16
d
d
k
where the second inequality follows for d sufficiently large. Therefore, we compute H2 , which contains all
samples from such components. This is repeated up to at most Hk , since if a set Hℓ+1 is distinct from Hℓ , it
k
must have added at least one component, and we have only k components. Note that 1 + k1 ≤ e, giving
the upper bound on variances in Hℓ∗ .
After this clustering step, the algorithm follows similarly to before. The main difference is in the convex
programming steps and concentration bounds. For instance, before, we considered the set
N
X
X
Cη = w ∈ SN,ε :
wi (Xi − µ)(Xi − µ)T − σ 2 I −
αj (µj − µ)(µj − µ)T ≤ η .
i=1
j∈[k]
2
P
Now, to reflect the different expression for the covariance of the GMM, we replace σ 2 I with j∈[k] αj σj2 I;
for example:
N
X
X
X
Cη = w ∈ SN,ε :
wi (Xi − µ)(Xi − µ)T −
αj σj2 I −
αj (µj − µ)(µj − µ)T ≤ η .
i=1
j∈[k]
j∈[k]
2
We note that since all variances in each cluster are off by a factor of at most e, this will only affect our
concentration and agnostic guarantees by a constant factor.
6 Agnostically Learning Binary Product Distributions, via Filters
In this section, we study the problem of agnostically learning a binary product distribution. Such a distribution is entirely determined by its coordinate-wise mean, which we denote by the vector p, and our first
5
We require an additional guess of “k1 and k2 ”: the split into how many components are within Hℓ∗ and H ℓ∗ respectively.
42
e
goal is to estimate p within ℓ2 -distance O(ε).
Recall that the approach for robustly learning the mean of
an identity covariance Gaussian, sketched in the introduction, was to compute the top absolute eigenvalue
of a modified empirical covariance matrix. Our modification was crucially based on the promise that the
covariance of the Gaussian is the identity. Here, it turns out that what we should do to modify the empirical
covariance matrix is subtract off a diagonal matrix whose entries are p2i . These values seem challenging
to directly estimate. Instead, we directly zero out the diagonal entries of the empirical covariance matrix.
e
Then the filtering approach proceeds as before, and allows us to estimate p within ℓ2 -distance O(ε),
as we
wanted. In the case when p has no coordinates that are too biased towards either zero or one, our estimate
e
is already O(ε)
close in total variation distance. We give an agnostic learning algorithm for this so-called
balanced case (see Definition 6.2) in Section 6.1.
However, when p has some very biased coordinates, this need not be the case. Each coordinate that is
biased needs to be learned multiplicatively correctly. Nevertheless, we can use our estimate for p that is close
in ℓ2 -distance as a starting point for handling binary product distributions that have imbalanced coordinates.
Instead, we control the total variation distance via the χ2 -distance between the mean vectors. Let P and Q
be two product distributions whose means are p and q respectively. From Lemma 2.15, it follows that
dTV (P, Q)2 ≤ 4
P (pi − qi )2
.
i qi (1 − qi )
So, if our estimate q is already close in ℓ2 -distance to p, we can interpret the right hand side above as giving a
renormalization of how we should measure the distance between p and q so that being close (in χ2 -distance)
implies that our estimate is close in total variation distance. PWe can then set up a corrected eigenvalue
problem using our initial estimate q as follows. Let χ2 (v)q = i vi2 qi (1 − qi ). Then, we compute
max v T Σv ,
χ2 (v)q =1
e √ε)
where Σ is the modified empirical covariance. Ultimately, we show that this yields an estimate that is O(
close in total variation distance. See Section 6.2 for further details.
6.1 The Balanced Case
The main result of this section is the following theorem:
Theorem 6.1. Let P be a binary product distribution in d dimensions and ε, τ > 0. Let S be a multiset of
Θ(d4 log(1/τ )/ε2 ) independent samples from P and S ′ be a multiset obtained by arbitrarily changing an
ε-fraction of the points in S. There exists a polynomial time algorithm
P′
p that returns a product distribution
′
′
so that, with probability at least 1 − τ , we have kp − p k2 = O(ε log(1/ε)), where p and p are the mean
vectors of P and P ′ respectively.
Note that Theorem 6.1 applies to all binary product distributions, and its performance guarantee relates
the ℓ2 -distance between the mean vectors of the hypothesis P ′ and the target product distribution P . If P
is balanced, i.e., it does not have coordinates that are too biased towards 0 or 1, this ℓ2 -guarantee implies a
similar total variation guarantee. Formally, we have:
Definition 6.2. For 0 < c < 1/2, we say that a binary product distribution is c-balanced if the expectation
of each coordinate is in [c, 1 − c].
For c-balanced binary product distributions, we have the following corollary of Lemma 2.15:
Fact 6.3. Let P and Q be c-balanced binary
product distributions with mean vectors p and q. Then, we have
−1/2
that dTV (P, Q) = O c
· kp − qk2 .
43
That is, for two c-balanced binary product distributions, where c is a fixed constant, the ℓ2 -distance
between their mean vectors is a good proxy for their total variation distance. Using Fact 6.3, we obtain the
following corollary of Theorem 6.1:
Corollary 6.4. Let P be a c-balanced binary product distribution in d dimensions, where c > 0 is a fixed
constant, and ε, τ > 0. Let S be a multiset of Θ(d4 log(1/τ )/ε2 ) independent samples from P and S ′
be a multiset obtained by arbitrarily changing an ε-fraction of the points in S. There exists a polynomial
timep
algorithm that returns a product distribution P ′ so that with probability at least 1 − τ , dTV (P ′ , P ) =
√
O(ε log(1/ε)/ c).
We start by defining a condition on the uncorrupted set of samples S, under which our algorithm will
succeed.
Definition 6.5 (good set of samples). Let P be an arbitrary distribution on {0, 1}d and ε > 0. We say that
a multiset S of elements in {0, 1}d is ε-good with respect to P if for every affine function L : {0, 1}d → R
we have | PrX∈u S (L(X) ≥ 0) − PrX∼P (L(X) ≥ 0)| ≤ ε/d.
The following simple lemma shows that a sufficiently large set of independent samples from P is ε-good
(with respect to P ) with high probability.
Lemma 6.6. Let P be an arbitrary distribution on {0, 1}d and ε, τ > 0. If the multiset S is obtained by
taking Ω((d4 +d2 log(1/τ ))/ε2 ) independent samples from P, it is ε-good with respect to P with probability
at least 1 − τ.
Proof. For a fixed affine function L : {0, 1}d → R, an application of the Chernoff bound yields that
after drawing N samples from P, we have that | PrX∈u S (L(X) ≥ 0) − PrX∼P (L(X) ≥ 0)| > ε/d
2
with probability at most 2 exp(−N ε2 /d2 ). Since there are at most 2d distinct linear threshold functions on
{0, 1}d , by the union bound, the probability that there exists an L satisfying the condition | PrX∈u S (L(X) ≥
2
0) − PrX∼P (L(X) ≥ 0)| > ε/d is at most 2d +1 exp(−N ε2 /d2 ), which is at most τ for N = Ω((d4 +
d2 log(1/τ ))/ε2 ).
The following definition quantifies the extent to which a multiset has been corrupted.
Definition 6.7. Given finite multisets S and S ′ we let ∆(S, S ′ ) be the size of the symmetric difference of S
and S ′ divided by the cardinality of S.
Note that the definition of ∆(S, S ′ ) is not symmetric in its two arguments.
Our agnostic learning algorithm establishing Theorem 6.1 is obtained by repeated application of the
efficient procedure whose performance guarantee is given in the following proposition:
Proposition 6.8. Let P be a binary product distribution with mean vector p and ε > 0 be sufficiently small.
Let S be ε-good with respect to P , and S ′ be any multiset with ∆(S, S ′ ) ≤ 2ε. There exists a polynomial
time algorithm F ILTER -BALANCED -P RODUCT that, given S ′ and ε > 0, returns one of the following:
p
(i) A mean vector p′ such that kp − p′ k2 = O(ε log(1/ε)).
(ii) A multiset S ′′ ⊆ S ′ such that ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − 2ε/d.
We start by showing how Theorem 6.1 follows easily from Proposition 6.8.
44
Proof of Theorem 6.1. By the definition of ∆(S, S ′ ), since S ′ has been obtained from S by corrupting an
ε-fraction of the points in S, we have that ∆(S, S ′ ) ≤ 2ε. By Lemma 6.6, the set S of uncorrupted samples
is ε-good with respect to P with probability at least 1 − τ. We henceforth condition on this event.
Our algorithm iteratively applies the F ILTER -BALANCED -P RODUCT
procedure of Proposition 6.8 until
p
it terminates returning a mean vector p′ with kp − p′ k2 = O(ε log(1/ε)). We claim that we need at
most d + 1 iterations for this to happen. Indeed, the sequence of iterations results in a sequence of sets
S0 = S ′ , S1′ , . . . , so that ∆(S, Si′ ) ≤ ∆(S, S ′ ) − i · (2ε/d). Thus, if the algorithm does not terminate in the
first d iterations, we have Sd′ = S, and in the next iteration we output the sample mean of S.
6.1.1
Algorithm F ILTER -BALANCED -P RODUCT: Proof of Proposition 6.8
In this section, we describe the efficient procedure establishing Proposition 6.8 followed by its proof of
correctness. Our algorithm F ILTER -BALANCED -P RODUCT is very simple: We consider the empirical dis′
tribution defined by the (corrupted) sample multiset S ′ . We calculate its mean vector µS and covariance
′
matrix M . If the matrix M has no large eigenvalues, we return µS . Otherwise, we use the eigenvector v ∗
′
corresponding to the maximum magnitude eigenvalue λ∗ of M and the mean vector µS to define a filter.
We zero out the diagonal elements of the covariance matrix for the following reason: The diagonal elements
could contribute up to Ω(1) to the spectral norm, even without noise. This would prevent us from obtaining
e
the desired error of O(ε).
Our efficient filtering procedure is presented in detailed pseudocode below.
Algorithm 9 Filter algorithm for a balanced binary product distribution
1: procedure F ILTER -BALANCED -P RODUCT (ε, S ′ )
input: A multiset S ′ such that there exists an ε-good S with ∆(S, S ′ ) ≤ 2ε
output: Multiset S ′′ or mean vector p′ satisfying Proposition 6.8
′
2:
Compute the sample mean µS = EX∈u S ′ [X] and the sample covariance M with zeroed diagonal,
′
′
3:
i.e., M = (Mi,j )1≤i,j≤d with Mi,j = EX∈u S ′ [(Xi − µSi )(Xj − µSj )], i 6= j, and Mi,i = 0.
4:
Compute approximations for the largest absolute eigenvalue of M , λ∗ := kM k2 , and the associated
unit eigenvector v ∗ .
′
5:
if kM k2 ≤ O(ε log(1/ε)), then return µS .
p
6:
Let δ := 3 εkM k2 . Find T > 0 such that
′
Pr (|v ∗ · (X − µS )| > T + δ) > 8 exp(−T 2 /2) + 8ε/d.
X∈u S ′
7:
′
return the multiset S ′′ = {x ∈ S ′ : |v ∗ · (x − µS )| ≤ T + δ}.
Tightness of our Analysis.pWe remark that the analysis of our filter-based algorithm is tight, and more
generally our bound of O(ε log(1/ε)) is a bottleneck for filter-based approaches.
More specifically, we note that our algorithm will never successfully add points back to S after they have
been removed by the adversary. Therefore, if an ε-fraction of the points in S are changed, our algorithm
may be able to remove these outliers from S ′ , but will not be able top
replace them with their original values.
These changed values can alter the sample mean by as much as Ω(ε log(1/ε)).
To see this, consider the following example. Let P be the product distribution with mean p, where
pi = 1/2 for all i. Set ε = 2−(d−1) . We draw a Θ(d4 log(1/τ )/ε2 ) size multiset S which we assume
is ε-good. The fraction of times the all-zero vector appears in S is less than 2−(d−1) . So, the adversary
is allowed to corrupt all such zero-vectors. More specifically, the adversary replaces each occurrence of
the all-zero vector with fresh samples from P, repeating if any all-zero vector is drawn. In effect, this
45
procedure generates samples from the distribution Pe, defined as P conditioned on not being the all-zero
vector. Indeed, with high probability, the set S ′ is ε-good for Pe. So, with high probability, the mean of S ′
in each √
coordinate is at least
1/2 + 2−(d+2) . Thus, the ℓ2 -distance between the mean vectors of P and Pe is
p
at least d2−(d+2) = Θ(ε log(1/ε)). Note that for any affine function L, we have that PrX∈u S ′ (L(X) ≥
0) ≤ PrX∈u S (L(X) ≥ 0)/(1 − ε) + 2ε/d, which means that no such function can effectively distinguish
between S ′ \ S and S, as would be required by a useful filter.
The rest of this section is devoted to the proof of correctness of algorithm F ILTER -BALANCED -P RODUCT.
6.1.2
Setup and Basic Structural Lemmas
By definition, there exist disjoint multisets L, E, of points in {0, 1}d , where L ⊂ S, such that S ′ = (S \
′
L) ∪ E. With this notation, we can write ∆(S, S ′ ) = |L|+|E|
|S| . Our assumption ∆(S, S ) ≤ 2ε is equivalent
′
to |L| + |E| ≤ 2ε · |S|, and the definition of S directly implies that (1 − 2ε)|S| ≤ |S ′ | ≤ (1 + 2ε)|S|.
Throughout the proof, we assume that ε is a sufficiently small constant. Our analysis will make essential
use of the following matrices:
′
′
• MP denotes the matrix with (i, j)-entry EX∼P [(Xi − µSi )(Xj − µSj )], but 0 on the diagonal.
′
′
• MS denotes the matrix with (i, j)-entry EX∈u S [(Xi − µSi )(Xj − µSj )], but 0 on the diagonal.
′
′
• ME denotes the matrix with (i, j)-entry EX∈u E [(Xi − µSi )(Xj − µSj )].
′
′
• ML denotes the matrix with (i, j)-entry EX∈u L [(Xi − µSi )(Xj − µSj )].
Our first claim follows from the Chernoff bound and the definition of a good set:
Claim 6.9. Let w ∈ Rd be any unit vector, then for any T > 0,
′
′
Pr (|w · (X − µS )| > T + kµS − pk2 ) ≤ 2 exp(−T 2 /2) + ε/d.
X∈u S
and
′
′
Pr (|w · (X − µS )| > T + kµS − pk2 ) ≤ 2 exp(−T 2 /2).
X∼P
Proof. Since S is ε-good, the first inequality follows from the second one. To prove the second inequality, it
′
′
suffices to bound the probability that |w · (X − µS ) − E[w · (X − µS )]| > T , X ∼ P , since the expectation
′
′
in question is w · (p − µS ), whose absolute value is at most kµS − pk2 , by Cauchy-Schwarz. Note that
′
′
w · (X − µS ) is a sum of independent random variables wi (Xi − µSi ), each supported on an interval of
length 2|wi |. An application of the Chernoff bound completes the proof.
The following sequence of lemmas bound from above the spectral norms of the associated matrices.
Our first simple lemma says that the (diagonally reduced) empirical covariance matrix MS , where S is the
set of uncorrupted samples drawn from the binary product distribution P, is a good approximation to the
matrix MP , in spectral norm.
Lemma 6.10. If S is ε-good, kMP − MS k2 ≤ O(ε).
Proof. It suffices to show that |(MP )i,j − (MS )i,j | ≤ O(ε/d) for all i 6= j. Then, we have that
kMP − MS k2 ≤ kMP − MS kF ≤ O(ε).
46
Let ei denote the standard basis vector in the i-th direction in Rd . For i 6= j we have:
′
′
(MP )i,j = E [(Xi − µSi )(Xj − µSj )]
X∼P
= E [Xi Xj ] − µSi
X∼P
′
E [Xj ] − µSj
′
X∼P
= Pr ((ei + ej ) · X ≥ 2)
X∼P
′
− µSi
′
E [Xi ] + µSj µSi
′
X∼P
′
′
′
Pr (ej · X ≥ 1) − µSj Pr (ei · X ≥ 1) + µSj µSi .
X∼P
X∼P
A similar expression holds for MS except with probabilities for X ∈u S. Since S is ε-good with respect to
′
′
P , we have |(MP )i,j − (MS )i,j | ≤ ε/d + µSi ε/d + µSj ε/d ≤ 3ε/d. This completes the proof.
As a simple consequence of the above lemma, we obtain the following:
Claim 6.11. If S is ε-good, kM − (1/|S ′ |)(|S|MP + |E|ME − |L|ML )k2 = O(ε).
Proof. First note that we can write |S ′ |M = |S|MS + |E|ME0 − |L|ML0 , where ME0 and ML0 are obtained
from ME and ML by zeroing out the diagonal. Observe that |E| + |L| = O(ε)|S ′ |. This follows from
the assumption that ∆(S, S ′ ) ≤ 2ε and the definition of S ′ . Now note that the matrices ME − ME0 and
ML − ML0 are diagonal with entries at most 1, and thus have spectral norm at most 1. The claim now follows
from Lemma 6.10.
′
Recall that if µS = p, MP would equal the (diagonally reduced) covariance matrix of the product
distribution P, i.e., the identically zero matrix. The following simple lemma bounds from above the spectral
norm of MP by the ℓ22 -norm between the corresponding mean vectors:
′
Lemma 6.12. We have that kMP k2 ≤ kµS − pk22 .
′
′
Proof. Note that (MP )i,j = (µSi − pi )(µSj − pj ) for i 6= j and 0 otherwise. Therefore, MP is the difference
′
′
′
of (µS − p)(µS − p)T and the diagonal matrix with entries (µSi − pi )2 . This in turn implies that
′
′
′
(µS − p)(µS − p)T MP Diag(−(µSi − pi )2 ) .
′
Note that both bounding matrices have spectral norm at most kµS − pk22 , hence so does MP .
The following lemma, bounding from above the spectral norm of ML , is the main structural result of
this section. This is the core result needed to establish that the subtractive error cannot change the sample
mean by much:
′
Lemma 6.13. We have that kML k2 = O(log(|S|/|L|) + kµS − pk22 + ε · |S|/|L|), hence
′
(|L|/|S ′ |) · kML k2 = O(ε log(1/ε) + εkµS − pk22 ).
Proof. Since L ⊆ S, for any x ∈ {0, 1}d , we have that
|S| · Pr (X = x) ≥ |L| · Pr (X = x) .
X∈u S
X∈u L
(28)
Since ML is a symmetric matrix, we have kML k2 = maxkvk2 =1 |v T ML v|. So, to bound kML k2 it suffices
to bound |v T ML v| for unit vectors v. By definition of ML , for any v ∈ Rd we have that
|v T ML v| =
′
E [|v · (X − µS )|2 ].
X∈u L
47
The RHS is in turn bounded from above as follows:
S′
2
E [|v · (X − µ )| ] = 2
X∈u L
≤2
≪
Z
Z
√
d
Pr
0
0
√
X∈u L
d
′
|v · (X − µS )| > T · T dT
S′
min 1, |S|/|L| · Pr |v · (X − µ )| > T
T dT
X∈u S
Z 4√log(|S|/|L|)+kµS ′ −pk2
0
+ (|S|/|L|)
Z
√
d
√
4
T dT
log(|S|/|L|)+kµS ′ −pk2
′
exp(−(T − kµ
S′
≪ log(|S|/|L|) + kµS − pk22 + ε · |S|/|L| ,
− pk2 ) /2)T + εT /d dT
2
where the second line follows from (28) and the third line follows from Claim 6.9. This establishes the first
part of the lemma.
′
The bound (|L|/|S|)kML k2 = O(ε log(1/ε) + εkµS − pk22 ) follows from the previously established
bound using the monotonicity of the function x ln(1/x), and the fact that |L|/|S| ≤ 2ε. The observation
|S|/|S ′ | ≤ 1 + 2ε ≤ 2 completes the proof of the second part of the lemma.
Claim 6.11 combined with Lemmas 6.12 and 6.13 and the triangle inequality yield the following:
′
Corollary 6.14. We have that kM − (|E|/|S ′ |)ME k2 = O(ε log(1/ε) + kµS − pk22 ).
We are now ready to analyze the two cases of the algorithm F ILTER -BALANCED -P RODUCT.
6.1.3
The Case of Small Spectral Norm
′
We start by analyzing the case where the mean vector µS is returned. This corresponds to the case that
the spectral norm of M is appropriately small, namely kM k2 ≤ O(ε log(1/ε)). We start with the following
simple claim:
′
Claim 6.15. Let µE , µL be the mean vectors of E and L respectively. Then, kµE − µS k22 ≤ kME k2 and
′
kµL − µS k22 ≤ kML k2 .
Proof. We prove the first inequality, the proof of the second being identical. Note that ME is a symmetric
matrix, so kME k2 = maxkvk2 =1 |v T ME v|. Moreover, for any vector v we have that
v T ME v =
′
′
E [|v · (X − µS )|2 ] ≥ |v · (µE − µS )|2 .
X∈u E
′
Let w = µE − µS and take v = w/kwk2 . We conclude that kME k2 ≥ kwk22 , as desired.
′
The following crucial lemma, bounding from above the distance kµS − pk2 as a function of ε and
kM k2 , will be important for both this and the following subsections.
p
p
′
Lemma 6.16. We have that kµS − pk2 ≤ 2 εkM k2 + O(ε log(1/ε)).
Proof. First we observe that the mean vector µS of the uncorrupted sample set S is close to p. Since S is
ε-good, this follows from the fact that for any i ∈ [d], we have
|µSi − pi | = | Pr [ei · X ≥ 1] − Pr [ei · X ≥ 1]| ≤ ε/d.
X∈u S
X∼P
48
√
Therefore, we get that kµS − pk2 ≤ ε/ d.
Consider µE and µL , the mean vectors of E and L, respectively. By definition, we have that
′
|S ′ |µS = |S|µS + |E|µE − |L|µL ,
and thus by the triangle inequality we obtain
√
′
kµS − pk2 ≤ k(|E|/|S ′ |)(µE − p) − (|L|/|S ′ |)(µL − p)k2 + ε/ d .
Therefore, we have the following sequence of inequalities:
√
′
′
′
′
kµS − pk2 ≤ (|E|/|S ′ |) · kµE − µS k2 + (|L|/|S ′ |) · kµL − µS k2 + O(ε) · kµS − pk2 + ε/ d
p
p
√
′
≤ (|E|/|S ′ |) · kME k2 + (|L|/|S ′ |) · kML k2 + O(ε) · kµS − pk2 )) + ε/ d
p
p
√
′
≤ O(ε log(1/ε)) + (3/2) εkM k2 + O( ε) · kµS − pk2
p
p
′
≤ O(ε log(1/ε))) + (3/2) εkM k2 + kµS − pk2 /4 ,
where the first line follows from the triangle inequality, the second uses Claim 6.15, while the third uses
Lemma 6.13 and Corollary 6.14. Finally, the last couple of lines assume that ε is sufficiently small. The
proof of Lemma 6.16 is now complete.
We can now deduce the correctness of Step 5 of the algorithm F ILTER -BALANCED
-P RODUCT, since
p
′
for kM k2 ≤ O(ε log(1/ε)), Lemma 6.16 directly implies that kµS − pk2 = O(ε log(1/ε)).
6.1.4
The Case of Large Spectral Norm
We next show the correctness of the algorithm F ILTER -BALANCED -P RODUCT if it returns a filter (rejecting
an appropriate subset of S ′ ) in Step 6. This corresponds to the case that kM k2 ≥ Cε log(1/ε), for a
sufficiently large universal constant C > 0. We will show that the multiset S ′′ ⊂ S ′ computed in Step 6
satisfies ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − 2ε/d.
We start by noting that, as a consequence of Lemma 6.16, we have the following:
p
′
Claim 6.17. We have that kµS − pk2 ≤ δ := 3 εkM k2 .
p
′
Proof. By Lemma 6.16, we have that kµS − pk2 ≤ 2δ/3
p + O(ε log(1/ε)). Recalling that kM k2 ≥
Cε log(1/ε), if C > 0 is sufficiently large, the term O(ε log(1/ε)) is at most δ/3.
By construction, v ∗ is the unit eigenvector corresponding to the maximum magnitude eigenvalue of M.
Thus, we have (v ∗ )T M v ∗ = kM k2 = δ2 /(9ε). We thus obtain that
′
E [|v ∗ · (X − µS )|2 ] = (v ∗ )T ME v ∗ ≥
X∈u E
δ2 |S ′ |
,
20ε|E|
(29)
where the equality holds by definition, and the inequality follows from Corollary 6.14 and Claim 6.17 using
the fact that ε is sufficiently small and the constant C is sufficiently large (noting that the constant in the
RHS of Corollary 6.14 does not depend on C).
We show that (29) implies the existence of a T > 0 with the properties specified in Step 6 of the
algorithm F ILTER -BALANCED -P RODUCT. More specifically, we have the following crucial lemma:
Lemma 6.18. If kM k2 ≥ Cε log(1/ε), for a sufficiently large constant C > 0, there exists a T > 0
satisfying the property in Step 6 of the algorithm F ILTER -BALANCED -P RODUCT, i.e., such that
′
Pr (|v ∗ · (X − µS )| > T + δ) > 8 exp(−T 2 /2) + 8ε/d .
X∈u S ′
49
Proof. Assume for the sake of contradiction that this is not the case, i.e., that for all T > 0 we have that
′
Pr (|v ∗ · (X − µS )| ≥ T + δ) ≤ 8 exp(−T 2 /2) + 8ε/d .
(30)
X∈u S ′
Since E ⊆ S ′ , for all x ∈ {0, 1}d , we have that |S ′ | PrX∈u S ′ [X = x] ≥ |E| PrY ∈u E [Y = x]. This fact
combined with (30) implies that for all T > 0
′
Pr (|v ∗ · (Y − µS )| ≥ T + δ) ≪ (|S ′ |/|E|)(exp(−T 2 /2) + ε/d) .
(31)
Y ∈u E
Using (29) and (31), we have the following sequence of inequalities:
′
δ2 |S ′ |/(ε|E|) ≪
E [|v ∗ · (Y − µS )|2 ]
Y ∈u E
Z ∞
′
Pr |v ∗ · (Y − µS )| ≥ T · T dT
=2
0
′
Y ∈u E
Z
≪ (|S |/|E|)
Z √
′
4
≪
√
O( d)
0
log(|S |/|E|)+δ
min |E|/|S ′ |, exp(−(T − δ)2 /2) + ε/d T dT
T dt +
Z
∞
√
4
0
≪ log(|S ′ |/|E|) + δ2 +
ε|S ′ |
|E|
log(|S ′ |/|E|)+δ
(|S ′ |/|E|) exp(−(T − δ)2 /2)T dT +
Z
√
O( d)
0
.
This yields the desired contradiction
recalling that the assumption kM k2 ≥ Cε log(1/ε) and the definition
p
of δ imply that δ ≥ C ′ ε log(1/ε) for an appropriately large C ′ > 0.
The following simple claim completes the proof of Proposition 6.8:
Claim 6.19. We have that ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − 2ε/d .
Proof. Recall that S ′ = (S \ L) ∪ E, with E and L disjoint multisets such that L ⊂ S. We can similarly
write S ′′ = (S \ L′ ) ∪ E ′ , with L′ ⊇ L and E ′ ⊂ E. Since
∆(S, S ′ ) − ∆(S, S ′′ ) =
|E \ E ′ | − |L′ \ L|
,
|S|
it suffices to show that |E \ E ′ | ≥ |L′ \ L| + 2ε|S|/d. Note that |L′ \ L| is the number of points rejected by
the filter that lie in S ∩ S ′ . By Claim 6.9 and Claim 6.17, it follows that the fraction of elements x ∈ S that
are removed to produce S ′′ (i.e., satisfy |v ∗ · (x − µS ′ )| > T + δ) is at most 2 exp(−T 2 /2) + ε/d. Hence, it
holds that |L′ \ L| ≤ (2 exp(−T 2 /2) + ε/d)|S|. On the other hand, Step 6 of the algorithm ensures that the
fraction of elements of S ′ that are rejected by the filter is at least 8 exp(−T 2 /2) + 8ε/d. Note that |E \ E ′ |
is the number of points rejected by the filter that lie in S ′ \ S. Therefore, we can write:
|E \ E ′ | ≥ (8 exp(−T 2 /2) + 8ε/d)|S ′ | − (2 exp(−T 2 /2) + ε/d)|S|
≥ (8 exp(−T 2 /2) + 8ε/d)|S|/2 − (2 exp(−T 2 /2) + ε/d)|S|
≥ (2 exp(−T 2 /2) + 3ε/d)|S|
≥ |L′ \ L| + 2ε|S|/d ,
where the second line uses the fact that |S ′ | ≥ |S|/2 and the last line uses the fact that |L′ \ L|/|S| ≤
(2 exp(−T 2 /2) + ε/d). This completes the proof of the claim.
50
ε|S ′ |
T dT
d|E|
6.2 Agnostically Learning Arbitrary Binary Product Distributions
In this subsection, we build on the approach of the previous subsection to show the following:
Theorem 6.20. Let P be a binary product distribution in d dimensions and ε, τ > 0. There is a polynomial
time algorithm that, given ε and a set of Θ(d6 log(1/τ )/ε3 ) independent samples from P, an ε fraction of
e
which have been arbitrarily corrupted, outputs (the
p mean vector of) a binary product distribution P so that,
with probability at least 1 − τ , dTV (P, Pe) ≤ O( ε log(1/ε)).
By Lemma 2.15, the total variation distance between two binary product distributions can be bounded
from above by the square root by the χ2 -distance between the corresponding means. For the case of balanced product distributions, the χ2 -distance and the ℓ2 -distance are within a constant factor of each other.
Unfortunately, this does not hold in general, hence the guarantee of our previous algorithm is not sufficient
to get a bound on the total variation distance. Note, however that the χ2 -distance and the ℓ2 -distance can
be related by rescaling each coordinate by the standard deviation of the corresponding marginal. When we
rescale the covariance matrix in this way, we can use the top eigenvalue and eigenvector as before, except
that we obtain bounds that involve the χ2 in place of ℓ2 -distance. The concentration bounds we obtain with
this rescaling are somewhat weaker, and as a result, our quantitative guarantees for the general case are
somewhat weaker than in the balanced case.
Similarly to the case of balanced product distributions, we will require a notion of a “good” set for our
distribution. For technical reasons, the definition in this setting turns out to be more complicated. Roughly
speaking, this is to allow us to ignore coordinates for which the small fraction of errors is sufficient to
drastically change the sample mean.
Definition 6.21 (good set of samples). Let P be a binary product distribution and ε, η > 0. We say that a
multiset S of elements in {0, 1}d is (ε, η)-good with respect
to P if for every affine function L : {0, 1}d → R
P
and every subset of coordinates T ⊆ [d] satisfying i∈T pi (1 − pi ) < η the following holds: Letting ST
be the subset of points in S that have their ith coordinate equal to the most common value under P for all
i ∈ T, and letting PT be the conditional distribution of P under this condition, then
| Pr (L(X) ≥ 0) − Pr (L(X) ≥ 0)| ≤ ε3/2 /d2 .
X∈u ST
X∼PT
We note that a sufficiently large set of samples from P will satisfy the above properties with high
probability:
Lemma 6.22. If S is obtained by taking Ω(d6 log(1/τ )/ε3 ) independent samples from P, it is (ε, 1/5)-good
with respect to P with probability at least 9/10.
The proof of this lemma is deferred to Section D.
We will also require a notion of the number of coordinates on which S non-trivially depends:
Definition 6.23. For S a multiset of elements in {0, 1}d , let supp(S) be the subset of [d] consisting of indices
i so that the ith coordinate of elements of S is not constant.
Similarly to the balanced case, our algorithm is obtained by repeated application of an efficient filter
procedure, whose precise guarantee is described below.
Proposition 6.24. Let P be a binary product distribution in d dimensions and ε > 0. Suppose that S is an
(ε, η)-good multiset with respect to P with η > 10ε and S ′ be any multiset with ∆(S, S ′ ) ≤ 2ε. There exists
a polynomial time algorithm which, given ε and S ′ , returns one of the following:
p
(i) The mean vector of a product distribution P ′ with dTV (P, P ′ ) = O( ε ln(1/ε)).
51
(ii) A multiset S ′′ of elements of {0, 1}d so that there exists a product distribution Pe with mean pe and a
multiset Se that is (ε, η − kp − pek1 )-good with respect to Pe such that
e S ′′ ) + ε|supp(S ′′ )|/d + kp − pek1 /6 < ∆(S, S ′ ) + ε|supp(S ′ )|/d .
∆(S,
Our agnostic learning algorithm is then obtained by iterating this procedure. Indeed, the proof of Theorem 6.20 given Proposition 6.24 is quite simple.
Proof of Theorem 6.20. We draw Θ(d6 log(1/τ )/ε3 )) samples forming a set S, which is (ε, 1/5) good with
probability 1 − τ by Lemma 6.22. We condition on this event. The adversary corrupts an ε-fraction of S
producing a set S ′ with ∆(S, S ′ ) ≤ 2ε. The iterations of the algorithm produce a sequence of sets S0 =
S, S1 , . . . , Sk , where Si is (ε, ηi )-good for some binary product distribution Pi and some sets Si′ . We note
that ∆(Si , Si′ ) + ε|supp(Si′ )|/d is monotonically decreasing and never more than 3ε. Since |µPi − µPi+1 | ≤
th
dTV (P
Pi , Pi+1 ), in the i iteration, the aforementioned quantity decreases by at least dTV (Pi , Pi+1 ), and′
thus i dTV (Pi , Pi+1 ) ≤
p3ε. Therefore, dTV (P, Pk ) ≤ 3ε. Our algorithm outputs a product distribution P
with dTV (Pk , P ′ ) = O( ε ln(1/ε)). The triangle inequality completes the proof.
6.2.1
Algorithm F ILTER -P RODUCT: Proof of Proposition 6.24
In this section, we describe and analyze the efficient routine establishing Proposition 6.24. Our efficient
filtering procedure is presented in detailed pseudocode below.
Algorithm 10 Filter algorithm for an arbitrary binary product distribution
1: procedure F ILTER -P RODUCT(ε, S ′ )
input: ε > 0 and multiset S ′ such that there exists an ε-good S with ∆(S, S ′ ) ≤ 2ε
output: Multiset S ′′ or mean vector p′ satisfying Proposition 6.24
′
2:
Compute the sample mean µS = EX∈u S ′ [X] and the sample covariance matrix M
′
′
3:
i.e., M = (Mi,j )1≤i,j≤d with Mi,j = EX∈S ′ [(Xi − µSi )(Xj − µSj )].
′
′
4:
if there exists i ∈ [d] with 0 < µSi < ε/d or 0 < 1 − µSi < ε/d, then
5:
let S ′′ be the subset of elements of S ′ in which those coordinates take their most common value.
6:
return S ′′ .
/* For the later steps, we ignore any coordinates not in supp(S ′ ). */
7:
Compute approximations
for the largest magnitude eigenvalue λ′ of DM D, λ′ := kDM Dk2 , where
q
D = Diag(1/
8:
9:
′
′
µSi (1 − µSi )), and the associated unit eigenvector v ′ .
′
if kDM Dk2 < O(log(1/ε)), then return µS (re-inserting all coordinates affected by Step 6).
p
Let δ := 3 εkDM Dk2 . Find T > 0 such that
′
Pr (|v ∗ · (X − µS )| > T + δ) > 20/T 2 + 4ε3/2 /d2 ,
X∈u S ′
where v ∗ := Dv ′ .
′
10:
return the multiset S ′′ = {x ∈ S ′ : |v ∗ · (x − µS )| ≤ T + δ} .
This completes the description of the algorithm. We now proceed to prove correctness.
52
6.2.2
Chi-Squared Distance and Basic Reductions
As previously mentioned, our algorithm will use the chi-squared distance between the mean vectors as a
proxy for the total variation distance between two binary product distributions. Since the mean vector of the
target distribution is not known to us, we will not be able to use the symmetric definition of the chi-squared
distance used in Lemma 2.15 We will instead require the following asymmetric version of the chi-squared
distance:
2
def P
i −yi )
.
Definition 6.25. The χ2 -distance of x, y ∈ Rd is defined by χ2 (x, y) = di=1 x(xi (1−x
i)
The following fact follows directly from Lemma 2.15.
Fact
p6.26. Let P, Q be binary product distributions with mean vectors p, q respectively. Then, dTV (P, Q) =
O( χ2 (p, q)).
There are two problems with using the chi-squared distance between the mean vectors as a proxy for
the total variation distance. The first is that the chi-squared distance between the means is a very loose
approximation of the total variational distance when the means are close to 0 or 1 in some coordinate. To
circumvent this obstacle, we remove such coordinates via an appropriate pre-processing in Step 6. The second is that the above asymmetric notion of the χ2 -distance may be quite far from the symmetric definition.
To overcome this issue, it suffices to have that qi = O(pi ) and 1 − qi = O(1 − pi ). To ensure this condition
is satisfied, we appropriately modify the target product distribution (that we aim to be close to). Next, we
will show how we deal with these problems in detail.
Before we embark on a proof of the correctness of algorithm F ILTER -P RODUCT, we will make a few
reductions that we will apply throughout. First, we note that if some coordinate in Step 6 exists, then removing the uncommon values of that coordinate increases ∆(S, S ′ ) by at most ε/d but decreases |supp(S ′ )| by
at least 1. We also note that, if N is the set of coordinates outside of the support of S ′ , the probability that
an element in S ′ has a coordinate in N that does not take its constant value is 0. Note that this is at most
O(ε)P
away from the probability that an element taken from P has this property, and thus we can assume
that i∈N min{pi , 1 − pi } = O(ε). Therefore, after Step 6, we can assume that all coordinates i have
ε/d ≤ pi ≤ 1 − ε/d.
The next reduction will be slightly more complicated and depends on the following idea: Suppose that
there is a new product distribution Pe with mean pe and an (ε, η − kp − pek1 )-good multiset Se for Pe so that
e S ′ ) + kp − pek1 /5 ≤ ∆(S, S ′ ).
∆(S,
Then, it suffices to show that our algorithm works for Pe and Se instead of P and S (note that the input to the
algorithm, S ′ and ε in the same in either case). This is because the conditions imposed by the output in this
′
case would be strictly stronger. In particular, we may assume that µSi ≥ pi /3 for all i:
′
′
Lemma 6.27. There is a product distribution Pe whose mean vector pe satisfies µSi ≥ pei /3 and 1 − µSi ≥
(1 − pei /3) for all i, and a set Se ⊆ S that is (ε, η − kp − pek1 )-good for Pe and satisfies
e S ′ ) + kp − pek1 /5 ≤ ∆(S, S ′ ).
∆(S,
′
′
Proof. If all coordinates i have µSi ≥ pi /3 and 1 − µSi ≥ (1 − pi /3), then we can take Pe = P and Se = S.
′
Suppose that the ith coordinate has µSi < pi /3. Let Pe be the product whose mean vector pe has pei = 0
and pej = pj for j 6= i. Let Se be obtained by removing from S all of the entries with 1 in the ith -coordinate.
e S ′ ) + pi /5 ≤ ∆(S, S ′ ). Note that here we
Then, we claim that Se is (ε, η − pi )-good for Pe and has ∆(S,
have kp − pek1 = pi .
53
e
e
P First, we show that S is (ε, η − pi )-good for P . For any affine function L(x) and set T ⊆ [d] with
ej (1 − pej ) ≤ η − pi , we need to show that
j∈T p
| Pr (L(X) > 0) − Pr (L(X) > 0)| ≤ ε3/2 /d2 .
eT
X∈u S
X∼PeT
Let Te = T ∪ {i}. We may or may not have i ∈ T but, from the definition of pe,
X
X
X
pej (1 − pej ) =
pej (1 − pej ) =
pj (1 − pj ).
j∈T
Thus,
X
j∈Te
j∈T \{i}
pj (1 − pj ) = pi (1 − pi ) +
Since S is good for P , we have that
j∈T \{i}
X
j∈T
pej (1 − pej ) ≤ η − pi + pi (1 − pi ) ≤ η.
| Pr (L(X) ≥ 0) − Pr (L(X) ≥ 0)| ≤ ε3/2 /d2 .
X∈u STe
X∼PTe
Moreover, note that STe = SeT and PTe = PeT . Thus, Se is (ε, η − pi )-good for Pe.
e S ′ ) + pi /5 ≤ ∆(S, S ′ ). We write S = Se \ L
e ∪ E.
e We write S1 , L1 , S ′ for
Next, we show that ∆(S,
1
the subset of S, L, S ′ respectively, where the ith coordinate is 1. Since S is (ε, η)-good for P, we have that
|µSi − pi | ≤ ε3/2 /d2 . Recall that we are already assuming that pei ≥ ε/d. Thus, µSi ≥ 29pi /30. Therefore,
we have that |S1 | ≥ 29pi |S|/30. On the other hand, we have that |S1′ | ≤ pi |S ′ |/3 ≤ 11pi |S|/30. Thus,
e S ′ )) = O(ε). However, E
e = E ∪ S ′ and
|L1 | = |S1 \ S1′ | ≥ 18pi |S|/30. This means that pi = O(∆(S,
1
e = L \ L1 . This gives
L
e S′) =
∆(S,
≤
e + |L|
e
|E|
|E| + |S1′ | + |L| − |L1 |
≤
e
e
|S|
|S|
|E| + |L| − 7pi |S|/30
|E| + |L| − 7pi /30
=
e
|S|(1 − µSi )
|S|
∆(S, S ′ ) − 7pi /30
= ∆(S, S ′ ) − 7pi /30 + O(εpi )
1 − 31pi /30
≤ ∆(S, S ′ ) − pi /5 .
≤
′
Similarly, suppose that instead the ith -coordinate has 1 − µSi < (1 − pi )/3. Let Pe be the product whose
mean vector pe has pei = 1 and pej = pj for j 6= i. Let Se be obtained by removing from S all of the entries
with 0 in the ith -coordinate. Then, by a similar proof we have that Se is (ε, η − (1 − pi ))-good for Pe and has
e S ′ ) + (1 − pi )/5 ≤ ∆(S, S ′ ). Note that here we have kp − pek1 = 1 − pi .
∆(S,
′
′
By an easy induction, we can set all coordinates i with µSi ≥ pei /3 and 1 − µSi ≥ (1 − pei /3) to 0 or 1
respectively, giving an Se and Pe such that Se is (ε, η − kp − pek1 )-good for Pe and
e S ′ ) + kp − pek1 /5 ≤ ∆(S, S ′ ) ,
∆(S,
as desired.
In conclusion, throughout the rest of the proof we may and will assume that for all i,
′
• ε/d ≤ µSi ≤ 1 − ε/d.
′
′
• µSi ≥ pi /3 and 1 − µiS ≥ (1 − pi )/3.
54
6.2.3
Setup and Basic Structural Lemmas
As in the balanced case, we can write S ′ = (S \ L) ∪ E for disjoint multisets L and E. Similarly, we define
the following matrices:
′
′
• MP to be the matrix with (i, j)-entry EX∼P [(Xi − µSi )(Xj − µSj )],
′
′
• MS to be the matrix with (i, j)-entry EX∈u S [(Xi − µSi )(Xj − µSj )],
′
′
• ME to be the matrix with (i, j)-entry EX∈u E [(Xi − µSi )(Xj − µSj )], and
′
′
• ML to be the matrix with (i, j)-entry EX∈u L [(Xi − µSi )(Xj − µSj )].
Note that we no longer zero-out the diagonals of MP and MS . We start with the following simple claim:
P
′
′
Claim 6.28. For any v ∈ Rd satisfying di=1 vi2 µSi (1 − µSi ) ≤ 1, the following statements hold:
p
′
(i) VarX∼P [v · X] ≤ 9 and |v · (p − µS )| ≤ χ2 (µS ′ , p), and
p
′
(ii) PrX∼P |v · X − µS | ≥ T + χ2 (µS ′ , p) ≤ 9/T 2 .
Proof. Recall that p denotes the mean vector of the binary product P. To show (i), we use the fact that
Xi ∼ Ber(pi ) and the Xi ’s are independent. This implies that
#
" d
d
d
d
X
X
X
X
′
′
2
2
vi2 µSi (1 − µSi ) ≤ 9 ,
vi pi (1 − pi ) ≤ 9
vi Var[Xi ] =
Var
vi Xi =
X∼P
i=1
i=1
′
i=1
i=1
′
where we used that pi ≤ 3µSi , (1 − pi ) ≤ 3(1 − µSi ) and the assumption in the claim statement. For the
second part of (i), note that
v
u d
d
q
q
q
S′
uX
X
p
−
µ
′
′
′
′
i
S′
i
S
S
≤t
vi2 µSi (1 − µSi )· χ2 (µS ′ , p) ≤ χ2 (µS ′ , p) ,
vi µi (1 − µi ) · q
|v·(p−µ )| =
′
′
µSi (1 − µSi )
i=1
i=1
where the firstP
inequality is Cauchy-Schwarz, and the second follows from the assumption in the claim
′
′
statement that di=1 vi2 µSi (1 − µSi ) ≤ 1. This proves (i).
To prove (ii), we note that Chebyshev’s inequality gives
Pr (|v · (X − p)| ≥ T ) ≤ Var [v · X]/T 2 ≤ 9/T 2 ,
X∼P
X∼P
where the second inequality follows from (i). To complete the proof note the inequality
q
′
|v · (X − µS )| ≥ T + χ2 (µS ′ , p)
implies that
′
′
|v · (X − p)| ≥ |v · (X − µS )| − |v · (p − µS )| ≥ T ,
where we used the triangle inequality and the second part of (i).
Let Cov[S] denote the sample covariance matrix with respect to S, and Cov[P ] denote the covariance
matrix of P. We will need the following lemma:
Lemma 6.29. We have the following:
55
(i)
p
p
χ2 (µS ′ , µS ) − χ2 (µS ′ , p) ≤ O(ε/d), and
√
(ii) kD (Cov[S] − Cov[P ]) Dk2 ≤ O( ε) .
Proof. For (i): Since S is good, for any i ∈ [d], we have
|µSi − pi | =
Pr (ei · X ≥ 1) − Pr (ei · X ≥ 1) ≤ ε3/2 /d2 .
X∈u S
X∼P
Therefore, by the triangle inequality we get
v
s
u d
q
q
S − p )2
u
X
(µ
d · (ε3 /d4 )
i
i
≤
≤ O(ε/d) ,
χ2 (µS ′ , µS ) − χ2 (µS ′ , p) ≤ t
′
′
S
S
ε/(2d)
µ
(1
−
µ
)
i
i
i=1
′
′
where the second inequality uses the fact that µSi (1 − µSi ) ≥ ε/(2d).
For (ii): Since S is good, for any i, j ∈ [d], we have
E [Xi Xj − pi pj ] =
X∈u S
Pr [(ei + ej ) · X ≥ 1] − Pr [(ei + ej ) · X ≥ 1] ≤ ε3/2 /d2 .
X∈u S
X∼P
Combined with the bound |µSi − pi | ≤ ε3/2 /d2 above, this gives
|Cov[S]i,j − Cov[P ]i,j | ≤ O(ε3/2 /d2 ) .
We thus obtain
kCov[S] − Cov[P ]k2 ≤ k(Cov[S] − Cov[P ])kF ≤ O(ε3/2 /d) .
q
p
′
′
S
S
Note that kDk2 = maxi 1/ µi (1 − µi ) ≤ 2d/ε. Therefore,
√
kD (Cov[S] − Cov[P ]) Dk2 ≤ O( ε) .
Combining Claim 6.28 and Lemma 6.29 we obtain:
P
′
′
Corollary 6.30. For any v ∈ Rd with di=1 vi2 µSi (1 − µSi ) ≤ 1, we have:
p
′
(i) VarX∈u S [v · X] ≤ 10 and |v · (µS − µS )| ≤ χ2 (µS ′ , p) + O(ε/d), and
p
′
(ii) PrX∈u S |v · X − µS | ≥ T + χ2 (µS ′ , p) ≤ 9/T 2 + ε3/2 /d2 .
Proof. We have that
| Var [v · X] − Var [v · Y ]| = v T (Cov[S] − Cov[P ]) v
X∈u S
Y ∼P
≤ kD −1 vk22 · kD (Cov[S] − Cov[P ]) Dk2
√
≤ O( ε)
≤1,
P
′
′
where the second line uses Lemma 6.29 (ii), and the assumption kD −1 vk22 = di=1 vi2 µSi (1 − µSi ) ≤ 1,
and the third line holds for small enough ε. Thus, using Claim 6.28 (i), we get that
Var [v · X] ≤ Var [v · Y ] + 1 ≤ 10 .
X∈u S
Y ∼P
56
By the Cauchy-Schwarz inequality and Lemma 6.29, we get
q
q
′
S
S′
2
S
S
|v · (µ − µ )| ≤ χ (µ , µ ) ≤ χ2 (µS ′ , p) + O(ε/d) .
This proves (i).
Part (ii) follows directly from Claim 6.28 (ii) and the assumption that S is good for P.
Lemma 6.31. For all i, j ∈ [d], we have that |(MS )i,j − (MP )i,j | ≤ O(ε3/2 /d2 ) and therefore
|S ′ |M − |S|MP + |E|ME − |L|ML
2
= O(|S ′ |ε3/2 /d) .
Proof. The bound |(MS )i,j − (MP )i,j | ≤ O(ε3/2 /d2 ) comes from expanding the LHS in terms of the
differences of linear threshold functions on S and P in the same way as the proof of Lemma 6.29. The
lemma then follows from the fact that that |S ′ |M = |S|MS + |E|ME − |L|ML and the observation |S| ≤
|S ′ |/(1 − 2ε) ≤ 2|S ′ |.
Combining the above we obtain:
Corollary 6.32. We have that kD(|S ′ |M − |S|MP − |E|ME + |L|ML )Dk2 = O(|S ′ | ·
q
p
′
′
Proof. Note that kDk2 = maxi 1/ µSi (1 − µSi ) ≤ 2d/ε, and therefore
D(|S ′ |M − |S|MP − |E|ME + |L|ML )D
2
√
ε) .
≤ kDk22 · |S ′ |M − |S|MP − |E|ME + |L|ML
′
3/2
≤ (2d/ε) · O(|S |ε
√
= O(|S ′ | · ε) .
2
/d)
We have the following lemma:
′
Lemma 6.33. We have that kDMP Dk2 ≤ 9 + χ2 (µS , p).
′
′
T
′
′
S
S
Proof. Note that M
P = (µ − p)(µ − p) + Diag(pi (1 − pi )). For any v with kv2 k ≤ 1, the vector
P
′
′
d
v = Dv ′ satisfies i=1 vi2 µSi (1 − µSi ) ≤ 1. Therefore, we can write
′
v ′T DMP Dv ′ = v T MP v = (v · (µS − p))2 + v T Diag(pi (1 − pi ))v .
Using Claim 6.28 (i), we get
′
′
(v.(µS − p))2 ≤ χ2 (µS , p)
and
|v T Diag(pi (1 − pi ))v| = | Var (v · (X − p))| ≤ 9 .
X∼P
This completes the proof.
The following crucial lemma bounds from above the contribution of the subtractive error:
′
Lemma 6.34. The spectral norm kDML Dk2 = O(|S ′ |/|L| + χ2 (µS , p)).
57
Proof. Similarly, we need to bound from above the quantity |v ′T DML Dv ′ | for all v ′ ∈ Rd with kv ′ k2 ≤ 1.
′
Note
|v ′T DML Dv ′ | = |v T ML v| = EX∈u L [|v · (X − µS )|2 ], where the vector v = Dv ′ satisfies
Pd that
S′
2 S′
i=1 vi µi (1 − µi ) ≤ 1. The latter expectation is bounded from above as follows:
′
′
E [(v · (X − µS ))2 ] ≤ 2 E [(v · (X − p))2 ] + 2(v · (µS − p))2
X∈u L
X∈u L
′
≤ 2 E [(v · (X − p))2 ] + 2χ2 (µS , p)
X∈u L
′
≤ (2|S|/|L|) · E [(v · (X − p))2 ] + 2χ2 (µS , p)
X∈u S
′
≤ 20|S|/|L| + 2χ2 (µS , p)
′
≤ 21|S ′ |/|L| + 2χ2 (µS , p) ,
where the first line uses the triangle inequality, the second line uses Claim 6.28 (i), the third line follows
from the fact that L ⊆ S, the fourth line uses Corollary 6.30 (i), and the last line uses the fact that ε is small
enough.
The above lemmas and the triangle inequality yield the following:
′
Corollary 6.35. We have that kD (M − (|E|/|S ′ |)ME ) Dk2 = O(1 + χ2 (µS , p)) .
We are now ready to analyze the two cases of the algorithm F ILTER -P RODUCT.
6.2.4
The Case of Small Spectral Norm
′
We start by considering
the case where the vector µS is returned. It suffices to show that in this case
p
′
dTV (P, P ) = O( ε log(1/ε)).
Let N be the set of coordinates not in supp(S ′ ). We note that only an ε fraction of the points in S could
have that any coordinate in N does not have its most common value. Therefore, at most a 2ε fraction of
samples from P have this property. Hence, the contribution to the variation distance coming from these
coordinates is O(ε). So, it suffices to consider only the coordinates not in N and show that dTV (PN , PN′ ) =
p
O( ε log(1/ε)). Thus, we may assume for the sake the analysis below that N = ∅.
We begin by bounding various χ2 -distances by the spectral norm of appropriate matrices.
′
Lemma 6.36. Let µE , µL be the mean vector of E and L, respectively. Then, χ2 (µS , µE ) ≤ kDME Dk2
′
and χ2 (µS , µL ) ≤ kDML Dk2 .
Proof. We prove the first inequality, the proof of the second being very similar.
′
′
Note that for any vector v, v T ME v = EX∈u E [|v · (X − µS )|2 ] ≥ |v · (µE − µS )|2 . Let v ∈ Rd be the
vector defined by
S′
µE
i − µi
p
vi =
.
′
′
µSi (1 − µSi ) χ2 (µS ′ , µE )
We have that
kD −1 vk22 =
d
X
i=1
′
′
vi2 µSi (1 − µSi ) =
d
S′ 2
X
(µE
i − µi )
= 1.
′
S ′ (1 − µS ′ )
χ2 (µS , µE )
µ
i
i
i=1
Therefore,
1
′
′
kDME Dk2 ≥ v T ME v ≥ |v · (µE − µS )|2 = χ2 (µS , µE ) .
58
We can now prove that the output in Step 8 has the desired guarantee:
p
p
√
Lemma 6.37. We have that χ2 (µS ′ , p) ≤ 2 εkDM Dk2 + O( ε).
′
Proof. Since S ′ = (S \ L) ∪ E, we have that |S ′ |µS = |S|µS + |E|µE − |L|µL . Recalling that L, E are
disjoint, the latter implies that
q
q
q
(|S|/|S ′ |) χ2 (µS ′ , µS ) ≤ (|E|/|S ′ |) χ2 (µS ′ , µE ) + (|L|/|S ′ |) χ2 (µS ′ , µL ) .
(32)
p
p
′
′
First note that, by Lemma 6.29, | χ2 (µS , µS ) − χ2 (µS , p)| ≤ O(ε/d). Lemma 6.36 and Corollary
6.35 give that
′
′
(|E|/|S ′ |)2 χ2 (µS , µE ) ≤ (|E|/|S ′ |)2 kDME Dk2 + O(ε) ≤ (|E|/|S ′ |)kDM Dk2 + O(ε(1 + χ2 (µS , p))) .
Thus,
q
q
p
√
′
2
S
E
′
(|E|/|S |) χ (µ , µ ) ≤ (|E|/|S |)kDM Dk2 + ε · O(1 + χ2 (µS ′ , p)) .
′
Lemmas 6.34 and 6.36 give that
′
′
(|L|/|S ′ |)2 χ2 (µS , µL ) ≤ (|L|/|S ′ |)2 kDML Dk2 ≤ O((|L|/|S ′ |)2 χ2 (µS , p) + ε) .
Thus,
q
q
√
′
′
2
S
L
(|L|/|S |) χ (µ , µ ) ≤ O((|L|/|S |) χ2 (µS ′ , p)) + O( ε) .
′
Substituting these into (32), yields
q
q
p
√
′
′
′
2
S
′
2
S
(|S|/|S |) χ (µ , p) ≤ (|E|/|S |)kDM Dk2 + O
ε 1 + χ (µ , p)
.
For ε sufficiently small, we have that the
p
χ2 (µS ′ , p) terms satisfy
√
√
1
(|S|/|S ′ |) − O( ε) ≥ 1 − 2ε − O( ε) ≥ √ .
2
Recalling that |E|/|S ′ | ≤ ∆(S, S ′ )|S|/|S ′ | ≤ (5/2)ε, we now have:
q
p
√
χ2 (µS ′ , p) ≤ (5/2) εkDM Dk2 + O( ε) ,
as required.
p
p
p
Corollary 6.38. Forp
some universal constant C, if δ ≤ C ε ln(1/ε), then χ2 (µS ′ , p) ≤ O( ε log(1/ε)).
Otherwise, we have χ2 (µS ′ , p) ≤ δ.
p
p
Proof. From the algorithm we have δ := 3 ε|λ| = 3 εkDM Dk2 . By Lemma 6.37, we have that
q
√
5
χ2 (µS ′ , p) ≤ δ + O( ε) .
6
p
p
√
If C is sufficiently large, when δ > C ε ln(1/ε), this O( ε) is at most C ε ln(1/ε)/6.
p
Claim 6.39. If the algorithm terminates at Step 8, then we have dTV (P, P ′ ) ≤ O( ε ln(1/ε), where P ′ is
′
the product distribution with mean vector µS .
p
p
′
Proof. By Corollary 6.38, we have that χ2 (µS , p) ≤ O( ε log(1/ε)). Thus, p
by Corollary 6.26, the total
′
variation distance between the product distributions with means p and µS is O( ε log(1/ε)).
59
6.2.5
The Case of Large Spectral Norm
We next need to show the correctness of the algorithm if it returns a filter, If we reach thisp
step, then we have
kDM Dk2 = Ω(1), indeed |v ′ DM Dv ′T | = Ω(1), and by Corollary 6.38, it follows that χ2 (µS ′ , p) ≤ δ.
P
P
′
′
′2
Since kv ′ k2 = 1, Dv ′ satisfies di=1 (Dv ′ )2i µSi (1 − µSi ) = m
i=1 vi = 1. Thus, we can apply Corollary 6.30 to it.
′
We will show that PrX∈u E (|v ∗ · (X − µS )| > T + δ) is larger than this quantity for some T > 0.
Corollary 6.40. If we do not terminate at Step 8, then v ∗T ME v ∗ ≥ δ2 |S ′ |/(2ε|E|).
Proof. By Corollary 6.35, we have that
′
kD(M − (|E|/|S ′ |)ME )Dk2 = O(1 + χ2 (µS , p)) .
Since kD −1 vk2 ≤ 1, we have that
′
v ∗T (M − (|E|/|S ′ |)ME )v ∗ = O(1 + χ2 (µS , p)) = O(1 + δ) .
Recall that we can assume that δ2 /ε ln(1/ε) is a sufficiently large constant. So,
(|E|/|S ′ |)v ∗T ME v ∗ ≥ v ′T DM Dv ′ − O(1 + δ) = δ2 /ε − O(1 + δ) ≥ δ2 /2ε .
Lemma 6.41. If kDM Dk2 = Ω(1), there exists a T > 0 satisfying the property in Step 9 of the algorithm,
′
i.e., PrX∈u S ′ [v ∗ · (X − µS )| > T + δ] ≥ 20/T 2 + 4ε3/2 /d2 .
Proof. Assume for the sake of contradiction that this is not the case, i.e., that for all T > 0 we have that
′
Pr ′ (|v ∗ · (X − µS )| ≥ T + δ) < 20/T 2 + 4ε3/2 /d2 .
X∈u S
This implies that for all T > 0
′
|E| Pr (|v ∗ · (Y − µS )| > T + δ) ≪ |S ′ |(1/T 2 + ε3/2 /d2 ).
Y ∈u E
Since
for all i, we have
p
p
′
|vi (Yi − µSi )| ≤ 1/ pi (1 − pi ) ≤ 2 d/ε
√
′
Pr [|v ∗ · (Y − µS )|) > 2d3/2 / ε] = 0 .
Y ∈u E
60
Therefore, we have that
δ2 /ε ≪ (|E|/|S ′ |)(v ∗ )T ME v ∗
′
= (|E|/|S ′ |) E [|v ∗ · (Y − µS )|2 ]
Y ∈u E
Z ∞
′
Pr (|v ∗ · (Y − µS )| > T )T dT
= 2(|E|/|S ′ |)
≪
≤
≪
Z
Z
√
2d3/2 /
0
ε
0
√
2d3/2 / ε
0
Z √1/ε+δ
0
Y ∈u E
min{(|E|/|S ′ |), 1/(T − δ)2 + (ε/d)2 }T dT
min{(5/2)ε, 1/(T − δ)2 + (ε/d)2 }T dT
Z
∞
εT dt + √
2
1/ε+δ
T /(T − δ) dT +
√
≪ (1 + δ ε) + (log(1/ε) + δ ε) + ε/ d
√
≪ log(1/ε) + δ ε + δ2 ε .
2
√
This yields the desired contradiction recalling that δ ≥ C
Z
√
2d3/2 / ε
ε3/2 /d2 T dT
0
p
ε log(1/ε) for an appropriately large C > 0.
The following simple claim completes the proof of Proposition 6.24, and thus proves Theorem 6.20.
Claim 6.42. If we return an S ′′ it will satisfy ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − ε3/2 /d2 .
Proof. We can write S ′′ = (S \ L′′ ) ∪ E ′′ for disjoint multisets L′′ and E ′′ . Note that S ′′ is produced from
′
S ′ by removing points with |v · X − µS | ≥ T + δ. Thus, E ′′ ⊂ E and L′′ ⊇ L.
′
By our choice of T , the fraction of x ∈ S ′ with |v · X − µS | ≥ T + δ is at least 20/T 2 + 4ε3/2 /d2 .
Thus,
|E \ E ′′ | + |L′′ \ L| ≥ (20/T 2 + 4ε3/2 /d2 ) · |S ′ | ≥ (18/T 2 + 3ε3/2 /d2 )|S| .
On the other hand, the fraction of x ∈ S with the desired property is
′
Pr [|v ∗ · X − µS | ≥ T + δ] ≤ 9/T 2 + ε3/2 /d2 ,
X∈u S
by Corollary 6.30 (ii). Thus,
|L′′ \ L| ≤ (9/T 2 + ε3/2 /d2 )|S| ,
and therefore
|E \ E ′′ | ≥ (9/T 2 + 2ε3/2 /d2 )|S|.
In summary, we have that
∆(S, S ′ ) − ∆(S, S ′′ ) =
|E| + |L| − |E ′′ | − |L′′ |
≥ ε3/2 /d2 .
|S|
61
7 Agnostically Learning Mixtures of Two Balanced Binary Products, via
Filters
In this section, we study the problem of agnostically learning a mixture of two balanced binary product
distributions. Let p and q be the coordinate-wise means of the two product distributions. Let u = p2 − 2q .
Then, when there is no noise, the empirical covariance matrix is Σ = uuT + D, where D is a diagonal
(pi −qi )2
i
matrix whose entries are pi +q
. Thus, it can already have a large eigenvalue. Now in the presence
2 −
4
of corruptions it turns out that we can construct a filter when the second absolute eigenvalue is also large.
But even if only the top absolute eigenvalue is large, we know that both p and q are close to the line µ + cv,
where µ is the empirical mean and v is the top eigenvector. And by performing a grid search over c, we will
find a good candidate hypothesis.
Unfortunately, bounds on the top absolute eigenvalue do not translate as well into bounds on the total
variation distance of our estimate to the true distribution, as they did in all previous cases (e.g., if the top
absolute eigenvalue is small in the case of learning the mean of a Gaussian with identity covariance,
√ we can
just use the empirical mean, etc). In fact, an eigenvalue λ could just mean that p and q differ by λ along
the direction v. However, we can proceed by zeroing out the diagonals. If uuT has any large value along
the diagonal, this operation can itself produce large eigenvalues. So, this strategy only works when kuk∞ is
appropriately bounded. Moreover, there is a strategy to deal with large entries in u by guessing a coordinate
whose value is large and conditioning on it, and once again setting up a modified eigenvalue problem. Our
overall algorithm then follows from balancing all of these different cases, and we describe the technical
components in more detail in the next subsection.
7.1 The Full Algorithm
This section is devoted to the proof of the following theorem:
Theorem 7.1. Let Π be a mixture of two c-balanced binary product distributions in d dimensions. Given
ε > 0 and poly(d, 1/ε) log(1/τ ) independent samples from Π, an ε-fraction of which have been arbitrarily
corrupted, there is a polynomial time algorithm that, with probability at least 1 − τ , outputs a mixture of
√
two binary product distributions Π′ such that dTV (Π, Π′ ) = O(ε1/6 / c).
Recall that our overall approach is based on two strategies that succeed under different assumptions. Our
first algorithm (Section 7.2) assumes that there exists a coordinate in which the means of the two component
product distributions differ by a substantial amount. Under this assumption, we can use the empirical mean
vectors conditioned on this coordinate being 0 and 1. We show that the difference between these conditional
mean vectors is almost parallel to the difference between the mean vectors of the product distributions.
Considering eigenvectors perpendicular to this difference will prove a critical part of the analysis of this
case. Our second algorithm (Section 7.3) succeeds under the assumption that the mean vectors of the two
product distributions are close in all coordinates. This assumption allows us to zero out the diagonal of the
covariance matrix without introducing too much error.
Both these algorithms give an iterative procedure that produces filters which improve the sample set
until they produce an output. We note that these algorithms essentially only produce a line in Rd so that
both mean vectors of the target product distributions are guaranteed to be close to this line in ℓ2 -distance.
The assumption that our product distributions are balanced implies that Π is close in variation distance to
some mixture of two products whose mean vectors lie exactly on the given line. Given this line, we can
exhaustively compare Π to a polynomial number of such mixtures and run a tournament to find one that is
sufficiently close.
We note that together these algorithms will cover all possible cases. Our final algorithm runs all of
these procedures in parallel, obtaining a polynomial number of candidate hypothesis distributions, so that at
62
least one is sufficiently close to Π. We then run the tournament described by Lemma 2.9 in order to find a
particular candidate that is sufficiently close to the target. To ensure that all the distributions returned are in
some finite set M, we round each of the probabilities of each of the products to the nearest multiple of ε/d,
and similarly round the mixing weight to the nearest multiple of ε. This introduces at most O(ε) additional
error.
Algorithm 11 Filter algorithm for agnostically learning a mixture of two balanced products
1: procedure L EARN P RODUCT M IXTURE(ε, τ, S ′ )
input: a set of poly(d, 1/ε) log(1/τ ) samples of which an ε-fraction have been corrupted.
output: a mixture of two balanced binary products that is O(ε1/6 )-close to the target
2:
Run the procedure F ILTER -BALANCED -P RODUCT(2ε1/6 , S1′ ) for up to d + 1 iterations on a set S1′
of corrupted samples of size Θ(d4 log(1/τ )/ε1/3 ).
3:
for each 1 ≤ i∗ ≤ d do
′ ) for up to d+1 iterations
4:
Run the procedure F ILTER -P RODUCT-M IXTURE -A NCHOR(i∗ , ε, S2,i
∗
4
13/6
′
).
on a set S2,i∗ of corrupted samples of size Θ(d log(1/τ )/ε
Run the procedure F ILTER -P RODUCT-M IXTURE -C LOSE(ε, S3′ , δ := ε1/6 ) for up to d + 1 iterations
on a set S3′ of corrupted samples of size Θ(d4 log(1/τ )/ε13/6 ).
6:
Run a tournament among all mixtures output by any of the previous steps. Output the winner.
5:
7.2 Mixtures of Products Whose Means Differ Significantly in One Coordinate
We will use the following notation. Let Π be a mixture of two c-balanced binary product distributions. We
will write Π as αP + (1 − α)Q, where P, Q are binary product distributions with mean vectors p, q, and
α ∈ [0, 1]. In this subsection, we prove the following theorem:
Theorem 7.2. Let Π = αP + (1 − α)Q be a mixture of two c-balanced binary product distributions in d
dimensions, with ε1/6 ≤ α ≤ 1 − ε1/6 , such that there exists 1 ≤ i∗ ≤ d with pi∗ ≥ qi∗ + ε1/6 . There is an
algorithm that, given i∗ , ε > 0 and Θ(d4 log(1/τ )/ε3 ) independent samples from Π, an ε-fraction of which
have been arbitrarily corrupted, runs in polynomial time and, with probability at least 1 − τ , outputs a set
√
R of candidate hypotheses such that there exists Π′ ∈ R satisfying dTV (Π, Π′ ) = O(ε1/6 / c).
For simplicity of the analysis, we will assume without loss of generality that i∗ = d, unless otherwise
specified. First, we determine some conditions under which our sample set will be sufficient. We start by
recalling our condition of a good set for a balanced binary product distribution:
Definition 7.3. Let P be a binary product distribution in d dimensions and let ε > 0. We say that a multiset
S of elements of {0, 1}d is ε-good with respect to P if for every affine function L : Rd → R it holds
| Pr (L(X) > 0) − Pr (L(X) > 0)| ≤ ε/d.
X∈u S
X∼P
We will also need this to hold after conditioning on the last coordinate.
Definition 7.4. Let P be a binary product distribution in d dimensions and let ε > 0. We say that a
multiset S of elements of {0, 1}d is (ε, i)-good with respect to P if S is ε-good with respect to P , and
def
S j = {x ∈ S : xi = j} is ε-good for the restriction of P to xi = j, for j ∈ {0, 1}.
Finally, we define the notion of a good set for a mixture of two balanced products.
63
Definition 7.5. Let Π = αP + (1 − α)Q be a mixture of two binary product distributions. We say that a
multiset S of elements of {0, 1}d is (ε, i)-good with respect to Π if we can write S = SP ∪ SQ , where SP is
P|
(ε, i)-good with respect to P , SQ is (ε, i)-good with respect to Q, and | |S|S|
− α| ≤ ε/d2 .
We now show that taking random samples from Π produces such a set with high probability.
Lemma 7.6. Let Π = αP + (1 − α)Q be a mixture of binary product distributions, where P, Q are
binary product distributions with mean vectors p, q. Let S be a set obtained by taking Ω(d4 log(1/τ )/ε13/6 )
independent samples from Π. Then, with probability at least 1 − τ , S is (ε, i)-good with respect to Π for all
i ∈ [d].
The proof of this lemma is deferred to Section E.
We claim that given a good set with an ε-fraction of its entries corrupted, we can still determine Π from
it. In particular, this is achieved by iterating the following proposition.
Proposition 7.7. Let Π = αP + (1 − α)Q be a mixture of two c-balanced binary products, with pd ≥
qd + ε1/6 and ε1/6 < α < 1 − ε1/6 . Let S be an (ε, d)-good multiset for Π, and let S ′ be any multiset with
∆(S, S ′ ) ≤ 2ε. There exists an algorithm which, given S ′ and ε > 0, runs in polynomial time and returns
either a multiset S ′′ with ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − 2ε/d, or returns a list of mixtures of two binary products
√
S so that there exists a Π′ ∈ S with dTV (Π, Π′ ) = O(ε1/6 / c).
We note that iteratively applying this algorithm until it outputs a set R of mixtures gives Theorem 7.2.
Notation. All vectors in this section should be assumed to be over the first d − 1 coordinates only. We will
write p−d and q−d for the first d − 1 coordinates of p and q, but for other vectors we will use the similar
notation to that used elsewhere to denote (d − 1)-dimensional vectors.
The algorithm, written in terms of i∗ instead of d for generality, is as follows:
Algorithm 12 Filter algorithm for a mixture of two binary products whose means differ significantly in
some coordinate
1: procedure F ILTER -P RODUCT-M IXTURE -A NCHOR (i∗ , ε, S ′ )
2:
Let µ be the sample mean of S ′ without the i∗ coordinate. Let Σ be the sample covariance matrix of
S ′ without the i∗ row and column.
3:
Let S0′ and S1′ be the subsets of S ′ with a 0 or 1 in their i∗ coordinates, respectively.
4:
Let µ(j) be the sample mean of Sj′ without the i∗ coordinate.
5:
Let u = µ(1) − µ(0) . Compute the unit vector v ∗ ∈ Rd−1 with v ∗ · u = 0 that maximizes v T Σv and
let λ = v ∗T Σv ∗ .
/* Note that v ∗ is the unit vector maximizing the quadratic form v T Σv over the subspace u · v = 0,
and thus can be approximated using standard eigenvalue computations.*/
6:
if λ ≤ O(1) then
7:
Let√L be the set of points µ + i(ε1/6 /kuk2 )u truncated to be in [c, 1 − c]d for i ∈ Z with
|i| ≤ 1 + d/ε1/6 .
8:
return the set of distributions Π′ = α′ P ′ + (1 − α′ )Q′ with the means of P ′ and Q′ , p′ , q ′ with
′
1/6 .
′
′
′
p′−i∗ , q−i
∗ ∈ L and pi∗ , qi∗ ∈ [c, 1 − c], α ∈ [0, 1], multiples of ε
√
9:
Let δ = C(ε1/6 λ + ε2/3 log(1/ε)) for a sufficiently large constant C.
10:
Find a real number T > 0 such that
Pr (|v ∗ · (X−i∗ − µ)| > T + δ) > 8 exp(−T 2 /2) + 8ε/d .
X∈u S ′
11:
return the set S ′′ = {x ∈ S ′ : |v · (x−i∗ − µ)| ≤ T + δ}.
64
We now proceed to prove correctness. We note that given S = SP ∪ SQ , we can write
′
S ′ = SP′ ∪ SQ
∪ E.
′ ⊆ S and E is disjoint from S \ S ′ and S \ S ′ . Thus, we have
where SP′ ⊆ SP , SQ
Q
P
Q
P
Q
∆(S, S ′ ) =
′ | + |E|
|SP \ SP′ | + |SQ \ SQ
|S|
.
′
We use the notation µSP , µSP , µE ∈ Rd−1 etc., for the means of SP , SP′ , E, etc., excluding the last coordinate.
We next need some basic lemmata relating the means of some of these distributions.
Lemma 7.8. Let P be a binary product distribution with mean vector p. Let S be an ε-good multiset for P
in the sense of Definition 7.3. Let S̃ be a subset of S with |S| − |S̃| = O(ε|S|). Let µS̃ be the mean of S̃.
p
Then, kp − µS̃ k2 ≤ O(ε log(1/ε)).
√
Proof. Since S is ε-good, kµS − pk2 ≤ ε/ d. Let L = S \ S̃. We can apply appropriate lemmata from
′
Section 6.1. Note that Lemma 6.13 and Claim 6.15, only depend on µS as far as it appears in the definition
′
of ML , and we may treat it as a parameter that we will set to p. By Lemma 6.13 with µS := p, we have
′
k EX∈u L [(X − p)(X − p)T ]k2 ≤ O (log(|S|/|L|) + ε|S|/|L|). By Claim 6.15 again with µS := p, it
p
follows that (|L|/|S|)kµL − pk2 ≤ O(ε log(1/ε)). Since |S|µS = |S̃|µS̃ + |L|µL , we have µS − µS̃ =
−(|L|/|S̃|)(µL − µS ) and so
√
p
kµS − µS̃ k2 ≤ (|L|/|S̃|)kµL − µS k2 ≤ O(ε2 / d) + O(1 + ε)(|L|/|S|)kµL − pk2 ≤ O(ε log(1/ε)).
p
p
√
By the triangle inequality, kp − µS̃ k2 ≤ ε/ d + O(ε log(1/ε)) = O(ε log(1/ε)).
We next show that µ(1) − µ(0) is approximately parallel to p−d − q−d . Note that if we had S = S ′ and
µSP = p−d , µSQ = q−d , then µ(1) − µ(0) would be a multiple of p−d − q−d . Since S is ε-good, we can
bound the error introduced by µSP − p, µSQ − q−d and Lemma 7.8 allow us to bound the error in taking
′
′
µSP , µSQ instead of p−d , q−d . However, we still have terms in the conditional means of E:
Lemma 7.9. For some scalars a = O(ε), b0 = O(|E 0 |/|S ′ |), b1 = O(|E 1 |/|S ′ |), we have
0
1
k(1 − µd )µd u − (α(1 − α)(pd − qd ) + a)(p−d − q−d ) − b0 (µE − µ) − b1 (µE − µ)k2 ≤ O(ε log(1/ε)) ,
j
where E j is the subset of E with last entry j, µE is the mean of E j with dth coordinate removed.
′j
Proof. Let SP′j , SQ
, E j , S ′j denote the subset of the appropriate set in which the last coordinate is j. Let
′ j
′ j
′j
, and E j with the last entry truncated, respectively.
µSP , µSQ , µE denote the means of SP′j , SQ
We note that
′
S ′j = SP′ ∪ SQ
∪ Ej .
j
Taking the means of the subsets of S ′j , we find that
′ j
′ j
j
′j SQ
|µ
+ |E j |µE .
|S ′j |µ(j) = |SP̃ j |µSP + |SQ
Therefore, using this and Lemma 7.8, we have that
j
′j
|S ′j |µ(j) = |SP′j |p−d + |SQ
|q−d + |E j |µE + O(ε log(1/ε)|S j |),
65
where O(ε) denotes a vector of ℓ2 -norm O(ε).
Thus, we have
|S ′0 ||S ′1 |(µ(1) − µ(0) ) = (|S ′0 ||SP′1 | − |S ′1 ||SP′0 |)p−d
′1
′0
+ (|S ′0 ||SQ
| − |S ′1 ||SQ
|)q−d
1
0
+ |E 1 ||S ′0 |µE − |E 0 ||S ′1 |µE + O(ε log(1/ε)(|S 1 ||S ′0 | + |S 0 ||S ′1 |)) .
(33)
′j
| + |E j |, we have:
Since |S ′j | = |SP′j | + |SQ
0 = |S ′0 ||S ′1 | − |S ′1 ||S ′0 | = (|S ′0 ||SP′1 | − |S ′1 ||SP′0 |)
′1
′0
+ (|S ′0 ||SQ
| − |S ′1 ||SQ
|) + |E 1 ||S ′0 | − |E 0 ||S ′1 | .
Thus, the sum of the coefficients of the p−d and q−d terms in Equation (33) is |E 0 ||S ′1 | − |E 1 ||S ′0 |, which
is bounded in absolute value by |E||S ′ | ≤ O(ε|S|2 ). Meanwhile, the p−d coefficient of Equation (33) has:
|S ′0 ||SP′1 | − |S ′1 ||SP′0 |
= |S ′0 ||SP1 | − |S ′1 ||SP0 | + O(ε|S ′ |2 ) = |S ′0 ||S|αpd − |S ′1 ||S|α(1 − pd ) + O(ε|S ′ |2 )
= |S 0 ||S|αpd − |S 1 ||S|α(1 − pd ) + O(ε|S ′ |2 )
= ((α(1 − pd ) + (1 − α)(1 − qd ))αpd − (αpd + (1 − α)qd )α(1 − pd ) + O(ε))|S ′ |2
= (α(1 − α)(1 − qd )pd − α(1 − α)qd (1 − pd ) + O(ε))|S ′ |2 = (α(1 − α)(pd − qd ) + O(ε))|S ′ |2 .
Noting that (|E 1 ||S ′0 | − |E 0 ||S ′1 |)α = O(ε|S ′ |2 ) and (|E 1 ||S ′0 | − |E 0 ||S ′1 |)(1 − α) = O(ε|S ′ |2 ), we can
write Equation (33) as:
|S ′0 ||S ′1 |(µ(1) − µ(0) ) = (α(1 − α)(pd − qd ) + O(ε))|S ′ |2 (p−d − q−d )
+ (|E 1 ||S ′0 | − |E 0 ||S ′1 |)(αp−d + (1 − α)q−d )
1
0
+ |E 1 ||S ′0 |µE − |E 0 ||S ′1 |µE + O(ε log(1/ε)|S ′ |2 ) .
We write µΠ = αp−d + (1 − α)q−d and so, dividing by |S ′ |2 and recalling that |E|/|S ′ | ≤ O(ε), we get
1
µd (1 − µd )(µ(1) − µ(0) ) = (α(1 − α)(pd − qd ) + O(ε))(p−d − q−d ) + O(|E 1 |/|S ′ |)(µE − µΠ )
0
+ O(|E 0 |/|S ′ |)(µE − µΠ ) + O(ε log(1/ε)) .
(34)
If µΠ = µ, then we would be done. So, we must bound the error introduced by making this substitution. We
can express µ as
′
′
′
|µSQ + |E|µE
|S ′ |µ = |SP′ |µSP + |SQ
1
0
= |S|µΠ + O(ε|S|)(p−d − q−d ) + O(ε log(1/ε)|S ′ |) + |E 1 |µE + |E 0 |µE ,
and so
1
0
|S|(µΠ − µ) = O(ε|S|)(p−d − q−d ) + O(ε log(1/ε)|S|) + |E 1 |(µE − µ) + |E 0 |(µE − µ) .
Thus, we have
1
0
µΠ = µ + O(ε)(p−d − q−d ) + O(ε log(1/ε)) + O(|E 1 |/|S ′ |)(µE − µ) + O(|E 0 |/|S ′ |)(µE − µ) .
Substituting this into Equation (34), gives the lemma.
66
We now show that, for any vector v perpendicular to u, if the variance of S ′ in the v-direction is small,
then v · p−d and v · q−d are both approximately v · µ.
Lemma 7.10. For any v with kvk2 = 1, v · u = 0, we have that |v · (p−d − µ)| ≤ δ and |v · (q−d − µ)| ≤ δ
for δ := C(ε1/6 kΣk2 + ε2/3 log(1/ε)) for a sufficiently large constant C as defined in the algorithm.
Proof. We begin by noting that
v T Σv = Var ′ (v · X) =
X∈u S
j
≥ (|E |/|S ′ |)
E [|v · (X − µ)|2 ]
X∈u S ′
E [|v · (X − µ)|2 ]
X∈u E j
j
≥ (|E j |/|S ′ |)|v · (µE − µ)|2 .
Next, since v · u = 0, we have by Lemma 7.9 that
|v · (p−d − q−d )| ≤
1
1
0
≤
· O(|E 0 |/|S ′ |)v · (µE − µ) + O(|E 1 |/|S ′ |)v · (µE − µ) + O(ε log(1/ε))kvk2
α(1 − α)(pd − qd )
q
1
=O
ε(v T Σv) + ε log(1/ε) .
α(1 − α)(pd − qd )
′
′
′ |µSq + |E|µE + |S ′ |O(ε log(1/ε)), and so
However, we have that |S ′ |µ = |SP′ |µSp + |SQ
′
′
′
′
(|S ′ | − |E|)(µ − µSp ) = |SQ
|(µSQ − µSP ) + |E|(µE − µ) + |S ′ |O(ε log(1/ε)) .
Now, we have:
µ − p−d = (1 − α + O(ε))(q−d − p−d ) + O(|E|/|S ′ |)(µE − µ) + O(ε log(1/ε)) .
Thus,
|v · (p−d − µ)| = O(v · (p−d − q−d )) + O(|E|/|S ′ |)(v · (µE − µ) − v · (µ − p−d )) + O(ε log(1/ε)).
Therefore,
|v · (p−d − µ)| = O
1
α(1 − α)(pd − qd )
q
ε(v T Σv)
+ ε log(1/ε) .
Inserting our assumptions that α(1 − α) ≥ ε1/6 /2, and pd − qd ≥ ε1/6 gives
p
|v · (p−d − µ)| = O(ε1/6 kΣk2 + ε2/3 log(1/ε)) ≤ δ ,
when C is sufficiently large.
The other claim follows symmetrically.
We can now show that if we return R, some distribution returned is close to Π. First, we show that there
are points on L close to p−d and q−d .
Lemma 7.11. There are c, d ∈ R such that p̃ = µ + cu and q̃ = µ + du have kp̃ − p−d k2 , kq̃ − q−d k2 ≤ δ.
Proof. If we take the c that minimizes kp̃ − p−d k2 , then u · (p̃ − p−d ) = 0. Thus, we can apply Lemma
7.10, giving that |(p̃ − p−d ) · (p−d − µ)| ≤ kp̃ − p−d k2 δ.
However, p̃ − µ = cu so we have (p̃ − p−d ) · (p̃ − µ) = 0 and thus
kp̃ − pk22 = |(p̃ − p−d ) · (p−d − µ)| ≤ kp̃ − p−d k2 δ.
Therefore, kp̃ − p−d k2 ≤ δ.
67
It is clear that even discretizing c and d, we can still find such a pair that satisfies this condition.
Lemma 7.12. There are p′ , q ′ ∈ L such that kp−d − p′ k2 , kq−d − q ′ k2 ≤ δ + O(ε1/6 ) .
Proof. By Lemma 7.11, there exist points p̃ = µ + (a/kuk2 )u and q̃ = µ + (b/kuk2 )u with a, b ∈ R that
have kp̃ − p−d k2 , kq̃ − q−d k2 ≤ δ.
Letting iε1/6 be the nearest integer multiple of ε1/6 to a, we have that p′ := µ + i(ε1/6 /kuk2 )u has
kp−d − p′ k2 ≤ kp̃−d − pk2 + kp′ − p̃k2 ≤ δ + ε1/6 . √
√
√
Note that we√have kp−d − p̃k2 ≤ kp−d − µk2 ≤ dkp−d − µk∞ ≤ d, which implies that a ≤ d.
Thus, |i| ≤ 1 + d/ε1/6 . If p′ ∈
/ [c, 1 − c], then replacing any coordinates less than c with c and more than
1 − c with 1 − c can only decrease the distance to p since p ∈ [c, 1 − c]d . Thus, there is a point p′ ∈ L with
kp−d − p′ k2 ≤ δ + O(ε1/6 ).
Similarly, we show that there is a q ′ ∈ L such that kq − q ′ k2 ≤ δ + O(ε1/6 ).
Corollary 7.13. If the algorithm terminates at Step 8, then there is a Π′ ∈ R with with dTV (Π′ , Π) =
√
O(ε1/6 / c).
Proof. By Lemma 7.12, there exists p̃, q̃ ∈ L such that kp−d − p̃k2 , kq−d − q̃k2 ≤ δ+O(ε1/6 ). But now there
is a distribution Π′ ∈ R, where Π′ = α′ P ′ + (1 − α′ )Q′ for binary products P ′ and Q′ , whose mean vectors
′ −q k ≤ O(ε1/6 ) and |p′ −p |, |q ′ −q | = O(ε1/6 ).
are p′ , q ′ and with |α′ −α| ≤ ε1/6 , kp′−d −p−d k2 , kq−d
d
d
−d 2
d
d
Note that this implies that kp′ − pk2 , kq ′ − qk2 = O(ε1/6 ).
√
√
Since P and Q are c-balanced, we have dTV (P, P ′ ) ≤ O(kp − p′ k2 / c) ≤ O(ε1/6 / c) and
√
√
dTV (Q, Q′ ) ≤ O(kq − q ′ k2 / c) ≤ O(ε1/6 / c).
√
√
Thus, dTV (Π′ , Π) ≤ δ + O(ε1/6 / c). Since we terminated in Step 8, λ ≤ O(1), and so δ = C(ε1/6 λ +
ε2/3 log(1/ε)) = O(ε1/6 ).
Now, we are ready to analyze the second part of our algorithm. The basic idea will be to show that if λ
is large, then a large fraction of the variance in the v-direction is due to points in E.
Lemma 7.14. If λ ≥ Ω(1), then
Var ′ [v ∗ · X] ≪
X∈u S
|E| EY ∈u E [|v ∗ · (Y − µ)|2 ]
.
|S ′ |(α(1 − α)(pd − qd ))2
Proof. We have that
|S| Var ′ [v ∗ · X] = |SP′ |
X∈u S
′
+ |SQ
|
′
Var ′ [v ∗ · X] + |v ∗ · (µSP − µ)|2
X∈u SP
Var [v ∗ · X] + |v ∗ · (µ
′
X∈u SQ
+ |E| E [|v ∗ · (X − µ)|2 ]] .
X∈u E
68
′
SQ
− µ)|2
!
Since SP and SQ are ε-good, we have that
Var ′ [v ∗ · X] =
X∈u SP
′
E
′
X∈u SP
[(v · X − v ∗ · µSP )2 ]
≤ (|SP |/|SP′ |)
=
(|SP |/|SP′ |)
′
[(v ∗ · X − v ∗ · µSP )2 ]
∗
∗
Var [v · X] + (v · (µ
X∈u SP
SP
−µ
′
SP
2
))
p
′
Var [v ∗ · X] + (v ∗ · (p−d − µSP ) + O(ε log(1/ε)))2
X∼P
p
≤ (1 + O(ε/α)) · (O(1) + O(ε log(1/ε))2 ) ≤ O(1) ,
≤ (|SP |/|SP′ |)
and similarly,
E
X∈u SP
Var [v ∗ · X] = O(1) .
′
X∈u SQ
Thus, we have:
|S ′ | Var ′ [v ∗ · X] ≤ |E| E [|v ∗ · (X − µ)|2 ] + O(1 + |v ∗ · (p−d − µ)|2 + |v ∗ · (q−d − µ)|2 )|S ′ | .
X∈u E
X∈u S
By Lemma 7.9, we have
|v · (p−d − µ)|, |v ∗ · (q−d − µ)| ≤ O(1/(α(1 − α)(pd − qd )))
0
1
· (O(|E0 |/|S ′ |)|v ∗ · (µE − µ)| + O(|E 1 |/|S ′ |)|v ∗ · (µE − µ)| + O(ε log(1/ε)))
r
≤ (|E|/|S ′ |) E [|v ∗ · (Y − µ)|2 ] + O(ε log(1/ε)) .
Y ∈u E
However,
λ = Var ′ [v ∗ · X] ≪
X∈u S
|E| EY ∈u E [|v ∗ · (Y − µ)|2 ]
+ O(1) .
|S|(α(1 − α)(pd − qd ))2
Since λ is larger than a sufficiently large constant, this completes the proof.
We next show that the threshold T > 0 required by our algorithm exists.
Lemma 7.15. If λ ≥ Ω(1), there exists a T > 0 such that
Pr (|v ∗ · (X − µ)| > T + δ) > 8 exp(−T 2 /2) + 8ε/d .
X∈u S ′
Proof. Assume for the sake of contradiction that this is not the case, i.e., that for all T > 0 we have that
Pr (|v ∗ · (X − µ)| ≥ T + δ) ≤ 8 exp(−T 2 /2) + 8ε/d .
X∈u S ′
Using the fact that E ⊂ S ′ , this implies that for all T > 0
|E| Pr (|v ∗ · (Y − µ)| > T + δ) ≪ |S ′ |(exp(−T 2 /2) + ε/d) .
Y ∈u E
69
Therefore, we have that
E [|v ∗ · (Y − µ)|2 ] ≪ δ2 + E [min(0, |v ∗ · (Y − µ)| − δ)2 ]
Y ∈u E
≪ δ2 +
Y ∈u E
√
d
Z
0
Z
√
d
Pr (|v ∗ · (Y − µ)| > T + δ)T dT
Z √
′
Y ∈u E
2
log(|S |/|E|)
≪ δ2 +
(ε/d)T dT +
T dT
0
0
Z ∞
+ √
(|S ′ |/|E|) exp(−T 2 /2)T dT
2
log(|S ′ |/|E|)
≪ δ2 + ε + log(|S ′ |/|E|) .
On the other hand, we know that
E [|v ∗ · (Y − µ)|2 ] ≫ (α(1 − α)(pd − qd ))2 λ|S ′ |/|E| ≫ log(|S ′ |/|E|) .
Y ∈u E
Combining with the above we find that
δ2 = O(ε1/3 λ) ≫ (α(1 − α)(pd − qd ))2 λ|S ′ |/|E| .
Or in other words,
ε4/3 ≥ ε1/3 |E|/|S ′ | ≫ (α(1 − α)(pd − qd ))2 ≥ ε2/3 ,
which provides a contradiction.
Finally, we show that S ′′ is closer to S than S ′ was.
Claim 7.16. If the algorithm returns S ′′ then ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − 2ε/d.
′′ ∪ E ′′ for S ′′ ⊆ S ′ , S ′′ ⊆ S and E ′′ ⊂ E, where E ′′
Proof. Since S ′′ ⊂ S, we can write S ′′ = SP′′ ∪ SQ
Q
P
P
Q
′′ \ S . Thus, we need to show that
has disjoint support from SP′′ \ SP and SQ
Q
′
′′
|E ′′ \ E| ≥ 2ε|S|/d + |SP′ \ SP′′ | + |SQ
\ SQ
|.
We have that
|S ′ \ S ′′ | = Pr ′ (|v · (X − µ)| ≥ T + δ)|S ′ |
X∈u S
≥ (8 exp(−T 2 /2) + 8ε/d)|S ′ | ≥ (4 exp(−T 2 /2) + 4ε/d)|S| .
By Hoeffding’s inequality, we have that
Pr (|v ∗ · (X − p−d )| ≥ T ) ≤ 2 exp(−T 2 /2) .
X∼P
By Lemma 7.10, we have that |v ∗ · (µ − p−d )| ≤ δ and so
Pr (|v ∗ · (X − µ)| ≥ T + δ) ≤ 2 exp(−T 2 /2) .
X∼P
Since S is (ε, d)-good, we have
Pr (|v ∗ · (X − µ)| ≥ T + δ) ≤ 2 exp(−T 2 /2) + ε/d .
X∈u SP
70
We get the same bound for X ∈u SQ , and so
Pr (|v ∗ · (X − µ)| ≥ T + δ)
X∈u S
= (|SP |/|S|) Pr (|v ∗ · (X − µ)| ≥ T + δ) + (|SQ |/|S|) Pr (|v ∗ · (X − µ)| ≥ T + δ)
X∈u SP
X∈u SQ
2
≤ 2 exp(−T /2) + ε/d .
′ \ S ′′ ) has v ∗ · (x − µ) ≥ T + δ, we have that
Since L′′P ∪ L′′Q ⊆ S but any x ∈ (SP′ \ SP′′ ) ∪ (SQ
Q
′
′′
|SP′ \ SP′′ | + |SQ
\ SQ
| ≤ Pr (|v ∗ · (X − µ)| ≥ T + δ)|S|
X∈u S
≤ (2 exp(−T 2 /2) + ε/d)|S| .
Finally, we have that:
′
′′
|E \ E ′ | = |S ′ \ S ′′ | − |SP′ \ SP′′ | − |SQ
\ SQ
|
≥ (4 exp(−T 2 /2) + 4ε/d)|S| − (2 exp(−T 2 /2) + ε/d)|S|
≥ (2 exp(−T 2 /2) + 3ε/d)|S|
′
′′
≥ |SP′ \ SP′′ | + |SQ
\ SQ
| + 2ε|S|/d ,
which completes the proof.
7.3 Mixtures of Products Whose Means Are Close in Every Coordinate
In this section, we prove the following theorem:
Theorem 7.17. Let ε, τ > 0 and let Π = αP + (1 − α)Q be a d-dimensional mixture p
of two c-balanced
product distributions P and Q whose means p and q satisfy kp − qk∞ ≤ δ, for δ ≥ ε log(1/ε), and
c ≤ pi , qi ≤ 1 − c for i ∈ [d]. Let S be a multiset of Ω(d4 log(1/τ )/(ε2 δ)) independent samples from Π.
Let S ′ be obtained by adversarially changing an ε-fraction of the points in S. There exists an algorithm that
runs in polynomial time and, with probability at least 1 − τ , returns set of distributions R such that some
√
Π′ ∈ R has dTV (Π, Π′ ) ≤ O(δ/ c).
We will assume without loss of generality that α ≤ 1/2. We may also assume that α > 10δ ≥ 10ε since
otherwise, we can make use of our algorithm for learning a single product distribution.
In this context, we require the following slightly different definition of a good set:
Definition 7.18. Let S be a multiset in {0, 1}d . We say that S is ε-good for the mixture Π if there exists a
P|
− α ≤ ε and that SP and SQ are ε/6-good for the component product
partition S = SP ∪ SQ so that |S|S|
distributions P and Q, respectively.
Lemma 7.19. If Π has mixing weights δ ≤ α ≤ 1 − δ, with probability at least 1 − τ , a set S of
Ω(d4 log 1/τ /(ε2 δ)) samples drawn from Π is good for Π.
The proof of this lemma is in Section E.
Our theorem will follow from the following proposition:
Proposition 7.20. Let Π be as above and S be a good multiset for Π. Let S ′ be any multiset with ∆(S, S ′ ) ≤
2ε. There exists a polynomial time algorithm that, given S ′ , ε > 0 and δ, returns either a multiset S ′′ with
∆(S, S ′′ ) ≤ ∆(S, S ′ ) − 2ε/d or a set of parameters of binary product distributions of size O(d/(εδ2 ))
√
which contains the parameters of a Π′ with dTV (Π, Π′ ) ≤ O(δ/ c).
71
Algorithm 13 Filter algorithm for mixture of two binary products whose means are close in every coordinate
1: procedure F ILTER -P RODUCT-M IXTURE -C LOSE (ε, S ′ , δ)
2:
Compute µ, the sample mean of S ′ , and Cov0 (S ′ ). Let C be a sufficiently large constant.
3:
if Cov0 (S ′ ) has at most one eigenvector with absolute eigenvalue more than Cδ2 then
4:
Let v ∗ be the unit eigenvector of Cov0 (S ′ ) with largest absolute eigenvalue.
√
5:
Let L be the set of points µ + iδv ∗ truncated to be in [c, 1 − c]d , for i ∈ Z with |i| ≤ 1 + d/δ.
6:
return the set of distributions of the form Π′ = α′ P ′ + (1 − α′ )Q′ with the means of P ′ and Q′
in L and α′ is a multiple of ε in [10ε, 1/2].
7:
Let v ∗ and u∗ be orthogonalpeigenvectors with eigenvalues more than Cδ2 .
8:
Find a number t ≥ 1 + 2 ln(1/ε) and θ a multiple of δ2 /d so that r = (cos θ)u∗ + (sin θ)v ∗
satisfies
Pr ′
Pr ′ (|r · (X − Y )| < t) < 2ε > 12 exp(−t2 /4) + 3ε/d .
X∈u S
Y ∈u S
return the set S ′′ = {x ∈ S ′ | PrY ∈u S ′ (|r · (x − Y )| < t) ≥ 2ε} .
9:
Before we present the algorithm, we give one final piece of notation. For S a set of points, we let
Cov(S) be the sample covariance matrix of S and Cov0 (S) be the sample covariance matrix with zeroed out
diagonal. Our algorithm is presented in detailed pseudocode below.
To analyze this algorithm, we begin with a few preliminaries. Firstly, we recall that S = SP ∪ SQ . We
′ ∪ E, where S ′ ⊂ S , S ′ ⊂ S , and
can write S ′ = SP′ ∪ SQ
P
Q
P
Q
′
|S|∆(S, S ′ ) = |SP \ SP′ | + |SQ \ SQ
| + |E| .
′
′
′ , respectively.
Let µSP and µSQ be the sample means of SP′ and SQ
p
′
′
Lemma 7.21. We have that αkp − µSP k2 , (1 − α)kq − µSQ k2 = O(ε log(1/ε)) .
Proof. This follows from Lemma 7.8.
A key part in our analysis involves showing that the matrix Cov0 (S ′ ) is close to being PSD:
′
Lemma 7.22. Let T be the multiset obtained from S ′ by replacing all points of SP′ with copies of µSP and
′
′ with copies of µSQ
all points of SQ
. Then, kCov0 (S ′ ) − Cov(T )k2 = O(δ2 ) .
Proof. Noting that the mean of T is µ and |T | = |S ′ |, we have:
|S ′ |Cov(S ′ ) = |SP′ |
E
′
X∈u SP
′
[(X − µ)(X − µ)T ] + |SQ
|
E
′
X∈u SQ
[(X − µ)(X − µ)T ]
+ |E| E [(X − µ)(X − µ)T ]
X∈u E
′
′
′
′
′
| Cov(SP′ ) + (µSQ − µ)(µSQ − µ)T
= |SP′ | Cov(SP′ ) + (µSP − µ)(µSP − µ)T + |SQ
+ |E| E [(X − µ)(X − µ)T ]
=
X∈u E
′
|SP |Cov(SP′ )
′
′
) + |S ′ |Cov(T ) .
+ |SQ
|Cov(SQ
(35)
′ ) can have large diagonal elements but small
Since P and Q are product distributions, Cov(SP′ ) and Cov(SQ
off-diagonal ones. On the other hand, we bound the elements on the diagonal of Cov(T ), but kCov(T )k2
may still be large due to off-diagonal elements.
72
By the triangle inequality, and Equation (35) with zeroed diagonal, we have:
kCov0 (S ′ ) − Cov(T )k2 ≤ kCov0 (S ′ ) − Cov0 (T )k2 + kCov0 (T ) − Cov(T )k2
!
′
′ |
|S
|SP |
Q
′
≤
kCov0 (SQ
)k2 + kCov0 (T ) − Cov(T )k2 .
kCov0 (SP′ )k2 +
|S ′ |
|S ′ |
(36)
We will bound each of these terms separately.
Note that Cov0 (T ) − Cov(T ) is a diagonal matrix and its non-zero entries are
(Cov0 (T ) − Cov(T ))i,i = Var [Xi ].
X∈u T
Since the mean of T is µ, for all i, we have that VarX∈u T [Xi ] ≤ EX∈u T [kX − µk2∞ ]. We seek to bound the
RHS from above.
′
′
′ |µSQ
′ |, we
+ |E|µE . Since |S ′ | − |E| = |SP′ | + |SQ
Note that µ satisfies |S ′ |µ = |SP′ |µSP + |SQ
′
′
′
′ |(µSQ − µSP ) + |E|(µE − µ). Using that |S ′ | − |E| = (1 + O(ε))|S|,
have (|S ′ | − |E|)(µ − µSP ) = |SQ
′ | = (1 − α)|S| − O(ε), |E| ≤ O(ε)|S|, we have
|SQ
′
′
′
kµ − µSP k∞ ≤ (1 − α + O(ε))kµSQ − µSP k∞ + O(ε) .
Similarly,
′
′
′
kµ − µSQ k∞ ≤ (α + O(ε))kµSQ − µSP k∞ + O(ε) .
Since S is ε-good for Π, it follows that kµSP − pk∞ ≤ ε/d and kµSQ − qk∞ ≤ ε/d. Also,
′
k|SP |µSP − |SP′ |µSP k∞ ≤ |SP | − |SP′ | ≤ O(ε)|S| .
Thus,
′
′
kµSP − µSP k∞ ≤ kµSP − (|SP′ |/|SP |)µSP k∞ + (|SP | − |SP′ |)/|SP | ≤ O(ε)|S|/|SP | ≤ O(αε) .
Similarly, we show that
′
kµSQ − µSQ k∞ ≤ O((1 − α)ε).
Finally, kp − qk∞ ≤ δ. Thus, by the triangle inequality, we get
′
′
kµSQ − µSP k∞ ≤ O(αε) + ε/d + δ + ε/d + O((1 − α)ε) ≤ δ + O(ε) .
We have the following sequence of inequalities:
|S ′ | Var [Xi ] ≤ |S ′ | E [kX − µk2∞ ]
X∈u T
X∈u T
=
|SP′ |kµ
′
′
′
|kµ − µSQ k2∞
− µSP k2∞ + |SQ
+ |E| E [kX − µk2∞ ]
X∈u T
≤
(|SP′ |
2
′
′
′
+ |SQ
|)(kµSQ − µSP k∞ + O(ε))2 + |E|
≤ (δ + O(ε))|S ′ | .
Thus,
kCov0 (T ) − Cov(T )k2 = max(Cov0 (T ) − Cov(T ))i,i = max Var(Ti ) ≤ O(δ2 + ε).
i
i
73
′
|S ′ |
|S |
′ )k terms in (36). To analyze the first of
It remains to bound the |SP′ | kCov0 (SP′ )k2 + |SQ′ | kCov0 (SQ
2
these terms, note that Cov0 (P ) = 0. We have that
kCov0 (SP′ )k2 = kCov0 (SP′ ) − Cov(P ) + Diag( Var (Xi ))k2
X∼P
≤
kCov(SP′ )
− Cov(P )k2 + max(| Var ′ (Xi ) − Var (Xi )|) .
i
X∈u SP
X∼P
Noting that
| Var ′ (Xi ) − Var (Xi )| = eTi (Cov(SP′ ) − Cov(P ))ei ,
X∈u SP
X∼P
we have that
max(| Var ′ (Xi ) − Var (Xi )|) ≤ kCov(SP′ ) − Cov(P )k2 ,
i
X∼P
X∈u SP
and so
kCov0 (SP′ )k2 ≤ 2kCov(SP′ ) − Cov(P )k2 .
By the triangle inequality,
kCov(SP′ ) − Cov(P )k2 ≤ kCov(SP′ ) − Cov(SP )k2 + kCov(SP ) − Cov(P )k2 .
Note that since S is good, the (i, j)-th entry of Cov(SP ) − Cov(P ) has absolute value at most ε/d. Thus,
kCov(SP ) − Cov(P )k2 ≤ kCov(SP ) − Cov(P )kF ≤ ε ,
which gives
kCov0 (SP′ )k2 ≤ 2kCov(SP′ ) − Cov(SP )k2 + O(ε).
We have
kCov(SP′ )
− Cov(SP )k2 = sup
|v|2 =1
| Var ′ (v · X) − Var (v · X)| .
X∈u SP
X∈u SP
Since SP′ ⊆ SP ,
|SP′ | Var ′ (v · X) ≤ |SP |
X∈u SP
E
X∈u SP
′
[(v · X − µSP ]
′
≤ |SP |( Var (v · X) + kµSP − µSP k22 )
X∈u SP
≤ (1 + O(ε/α))|SP′ | · ( Var (v · X) + O(ε2 log(1/ε)/α2 )) .
X∈u SP
Thus,
| Var ′ (v · X) − Var (v · X)| ≤ O(ε/α) Var (v · X) + O(ε2 log(1/ε)/α2 ))
X∈u SP
X∈u SP
X∈u SP
≤ O(ε/α) Var (v · X) + O(ε2 log(1/ε)/α2 ))
X∼P
2
(since kCov(SP ) − Cov(P )k2 ≤ ε)
≤ O(ε/α) + O(ε log(1/ε)/α2 ))
≤ O(ε log(1/ε)/α) .
(since kCov(P )k2 ≤ 1)
Thus, we have that
kCov0 (SP′ )k2 ≤ 2 · O(ε log(1/ε)/α) + O(ε) ≤ O(ε log(1/ε)/α).
74
(since α ≥ ε)
Therefore, combining the above we have that
′
|SP |
kCov0 (SP′ )k2 = (α + O(ε))kCov0 (SP′ )k2 = O(ε log(1/ε)) .
|S ′ |
A similar argument shows
′ |
|SQ
|S ′ |
!
′
kCov0 (SQ
)k2 = O(ε log(1/ε)).
Combining this with the above gives that
kCov0 (S ′ ) − Cov(T )k2 = O(δ2 + ε log(1/ε)).
By the assumption on δ in Theorem 7.17, δ2 = Ω(ε log(1/ε)), and the proof is complete.
We are now prepared to show that the first return condition outputs a correct answer. We begin by
′
′
showing that vectors u with large inner products with µSP − µ or µSQ − µ correspond to large eigenvectors
of Cov0 (S ′ ).
Lemma 7.23. For u ∈ Rd , we have
′
′
α(u · (µSP − µ))2 + (1 − α)(u · (µSP − µ))2 ≤ 2uT Cov0 (S ′ )u + O(δ2 )kuk22 .
Proof. Using Lemma 7.22, we have uT Cov0 (S ′ )u = VarX∈u T (u · X) + O(δ2 )kuk22 . From the definition
of T it follows that
!
′
′ |
|SQ
′
|SP |
|E|
′
SP
2
(u · (µSQ − µ))2 + ′ Var (u · X)
(u · (µ − µ)) +
Var (u · X) ≥
′
′
X∈u T
|S |
|S |
|S | X∈u E
′
′
≥ (α − 2ε)(u · (µSP − µ))2 + (1 − α − 2ε)(u · (µSQ − µ))2
′
′
≥ α/2 · (u · (µSP − µ))2 + (1 − α)/2 · (u · (µSQ − µ))2 .
(since α, 1 − α ≥ 4ε)
Next, we show that, if there is only one large eigenvalue of Cov0 (S ′ ), the means in question are both
close to a given line.
√
√
Lemma 7.24. There are p̃, q̃ ∈ L such that kp − p̃k2 ≤ O(δ/ α) and kq − q̃k2 ≤ O(δ/ 1 − α).
Proof. Let p′ = µ + av ∗ , q ′ = µ + bv ∗ with a, b ∈ R be the closest points to p and q on the line µ + cv ∗ ,
for c ∈ R. Then, v ∗ · (p′ − p) = 0 and since v ∗ is the only eigenvector of the symmetric matrix Cov0 (S ′ )
with eigenvalue more than C(δ2 + ε log(1/ε)), we have that
p
(p′ − p)T Cov0 (S ′ )(p′ − p) ≤ C(δ2 + ε log(1/ε))kp′ − pk22 .
We thus obtain:
kp′ − pk42 = (p′ − p) · (p − µ)2
≤ 2(p′ − p) · (p − µSP )2 + 2(p′ − p) · (p − µSP )2
≤ O(ε2 log(1/ε)/α2 )kp′ − pk22 + (4/α) · (p′ − p)T Cov0 (S ′ )(p′ − p)T + O(δ2 /α)kp′ − pk22
≤ O((δ2 + ε log(1/ε))/α)kp′ − pk22
(since α ≥ ε)
≤ O(δ2 /α)kp′ − pk22 ,
75
√
where the second line uses Lemmas 7.21 and 7.23. We thus have that kp′ − pk2 ≤ O(δ/ α). Letting iδ be
the nearest integer multiple to a, we have that p̃ := µ + iδv ∗ has
√
kp − p̃k2 ≤ kp′ − pk2 + kp′ − p̃k2 ≤ O(δ/ α).
√
√
√
√
Note that we have kp − p′ k2 ≤ kp − µk2 ≤ dkp − µk∞ ≤ d. So, a ≤ d/δ. Thus, |i| ≤ 1 + d/δ.
If p̃ ∈
/ [c, 1 − c], then replacing any coordinates less than c with c and more than 1 − c with 1 − c can only
decrease the distance to p, since p ∈ [c, 1 − c]d .
√
Similarly, we show that there is a q̃ ∈ L such that kq − q̃k2 ≤ O(δ/ 1 − α), which completes the
proof.
Corollary 7.25. If the algorithm outputs a set of distributions in Step 6, then one of those distributions has
√
dTV (Π′ , Π) ≤ O(δ/ c).
Proof. There is a distribution in the set Π = α′ P ′ + (1 − α′ )Q′ , where |α − α′ | ≤ ε and the means of
√
√
P ′ and Q′ are p̃ and q̃ as in Lemmap
7.24. Then, we have dTV (P, P ′ ) ≤ kp − p̃k/ c ≤ O(δ/ αc) and
√
dTV (Q, Q′ ) ≤ kp − p̃k/ c ≤ O(δ/ (1 − α)c). Thus, we have
√
√
√
√
dTV (Π′ , Π) ≤ O(ε)+αdTV (P, P ′ )+(1−α)dTV (Q, Q′ ) ≤ O(ε)+O(( α+ 1 − α)δ/ c) ≤ O(δ/ c) .
Next, we analyze the second case of the algorithm. We must show that Step 8 will find an r and t. First,
′
′
we claim that there is a θ which makes r nearly perpendicular to µSp − µSQ .
Lemma 7.26. There exists a r = (cos θ)u∗ + (sin θ)v ∗ , with θ a multiple of δ2 /d, that has
√
′
′
|r · (µSP − µSQ )| ≤ δ2 / d.
′
′
∗
Proof. Let z = (µSP − µSQ ). If u∗ · z = 0, then θ = 0 suffices. Otherwise, we take θ ′ = cot−1 ( uv ∗·z
·z ).
2 /d to θ ′ . Note that | cos θ − cos θ ′ |, | sin θ − sin θ ′ | ≤ |θ − θ ′ | and
Then, let θ be the p
nearest multiple
of
δ
√
|u∗ · z|, |v ∗ · z| ≤ kzk2 ≤ d. Then, we have
|r · z| = |(cos θ)(u∗ · z) + (sin θ)(v ∗ · z)|
√
≤ |(cos θ ′ )(u∗ · z) + (sin θ ′ )(v ∗ · z)| + |θ − θ ′ | d
√
= | sin θ ′ ||u∗ · z + (cot θ ′ )(v ∗ · z)| + |θ − θ ′ | d
√
≤ 0 + δ2 / d .
′
′
We now need to show that for this r, Step 8 will find a t. For this r, r · µSP and r · µSQ are close. We
need to show that E contains many elements x whose r · x is far from these. We can express this in terms
of T :
√
′
′
Lemma 7.27. Let r be a unit vector in r ∈ hu∗ , v ∗ i with |r · (µSP − µSQ )| ≤ δ2 / d. Then, there is a t > 1
such that
′
Pr (r · (X − µSP ) > 2t) > 12 exp(−(t − 1)2 /4) + 3ε/d .
X∈u T
76
′
Proof. First, we wish to show that EX∈u E [(r · (X − µSP ))2 ] is large.
Since r ∈ span(u∗ , v ∗ ), |r T Cov0 (S ′ )r| ≥ Cδ2 . By Lemma 7.22, we have that
Var (r · X) = r T Cov(T )r ≥ r T Cov0 (S ′ )r − O(δ2 ) ≥ (C − O(1))δ2 ≥ (C/2)δ2 ,
X∈u T
for sufficiently large C and we also have that r T Cov0 (S ′ )r is positive.
We note that
r T Cov(T )r = Var(r · T )
′
′
= (|E|/|S ′ |) Var (r · X) + O(α)(r · (µ − µSP ))2 + O(1 − α)(r · (µ − µSQ ))2
X∈u E
′
+ (|E|/|S |)(r · (µ − µE ))2
′
E 2
= (|E|/|S |) Var (r · X) + (r · (µ − µ )) + O(δ2 ) .
X∈u E
Now,
(37)
′
′
E [(r · (X − µSP ))2 ] = Var (r · X) + (r · (µSP − µE ))2 .
X∈u E
X∈u E
We also have that
′
′
′
′
|(r · (µSP − µSQ )) + |E|(r · µE )
|S ′ |(r · µ) = (|S ′ | − |E|)(r · µSP ) + |SQ
′
= (|S ′ | − |E|)(r · µSP ) + |E|(r · µE ) + |S ′ |O(δ2 ) .
Thus,
′
(|S ′ | − |E|)(r · (µ − µE )) = (|S ′ | − |E|)(r · (µSP − µE )) + |S ′ |O(δ2 ) ,
or
′
(r · (µ − µE )) = (r · (µSP − µE )) + O(δ2 ) .
This implies that
′
(r · (µSP − µE ))2 ≥ (r · (µ − µE ))2 /2 − O(δ4 ) .
Substituting into (37), we have
′
′
(|E|/|S ′ |) E [(r · (X − µSP ))2 ] = (|E|/|S ′ |)[ Var [r · X] + (r · (µSP − µE ))2 ] − O(δ4 ) ≫ C/2δ2 .
X∈u E
X∈u E
Thus, for C sufficiently large,
′
E [(r · (X − µSP ))2 ] ≫ δ2 /ε.
X∈u E
Suppose for a contradiction that this lemma does not hold. Then, since E ⊂ T , we have
′
Pr r · (X − µSP ) > 2t) ≤ (|S ′ |/|E|)12 exp(−t2 /2) + 3ε/d .
X∈u E
Thus, we have
′
Pr (r · (X − µSP ) > t)) ≤ (|S ′ |/|E|)12 exp(−(t − 1)2 /4) + 3ε/d ,
X∈u E
77
and we can write
|S ′ |δ2 ≪ |E| E [(r.X − r.µSP )2 ]
X∈u E
√
d
= |E|
Z
≪ |E|
0
Z
1+
0
′
Pr (r · (X − µSP ) > t))tdt
X∈u E
√
Z
′
ln(|S |/|E|)/2
′
tdt + |S |
∞
1+
√
ln(|S ′ |/|E|)/2
2
exp(−(t − 1) /4)tdt +
≪ |E| ln(|S ′ |/|E|) + |E| + |S ′ |(|E|/|S|) + ε
Z
√
n
ε/dtdt
0
≤ |S ′ | · O(ε log(1/ε)) .
Since we assumed that δ2 ≥ Ω(ε log(1/ε), this is a contradiction.
′ are suitably concentrated about their
To get a similar result for S ′ , we first need to show that SP′ and SQ
means:
Lemma 7.28. If t ≥ 1,
(1 − |E|/|S ′ |)
If t ≥ 1 +
Proof.
p
Pr′
′
X∈u SP ∪SQ
′
r · (X − µSP ) > t ≤ (5/4) exp(−(t − 1)2 /2) + ε/5d.
2 ln 6/ε, this is strictly less than 2ε/3.
′
′
Pr ′ (r · (X − µSP ) ≤ t) ≤ (|SP |/|SP′ |) Pr (r · (X − µSP ) ≤ t)
X∈u SP
X∈u SP
′
SP
≤ (1 + O(ε)/α) · Pr (r · (X − µ ) ≤ t) + ε/(12d)
X∼P
′
SP
= (1 + O(ε)/α) · Pr r · (X − p) ≤ t − (r · (µ − p)) + ε/(6d)
X∼P
≤ (1 + O(ε)/α) · 2 exp(−(t − 1/2)2 /2) + ε/6d .
(using Lemma 7.21 and Hoeffding’s inequality)
Similarly,
′
Pr ′ (r · (X − µSQ ) ≤ t) ≤ (1 + O(ε)/(1 − α)) · (2 exp(−(t − 1/2)2 /2) + ε/6d).
X∼SQ
√
Since |r · (µSQ − µSP )| ≤ δ2 / d ≤ 1/2, we have
′
Pr ′ r · (X − µSQ ) ≤ t ≤ (1 + O(ε)/(1 − α)) · 2 exp(−(t − 1)2 /2) + ε/6d .
X∼SQ
78
′ |)/|S ′ | = |E|/|S ′ | ≥ 4ε/3, we have
Noting that 1 − (|SP′ | + |SQ
(1 − |E|/|S ′ |)
Pr′
′
X∈u SP ∪SQ
= (|SP′ |/|S ′ |) Pr ′
X∼SP
′
r · (X − µSP ) > t
′
′
′
|/|S ′ |) Pr ′ r · (X − µSP ) > t
r · (X − µSP ) > t + (|SQ
X∼SQ
= ((α + O(ε)) (1 + O(ε)/α) + (1 − α + O(ε)) (1 + O(ε)/(1 − α)) · 2 exp(−(t − 1)2 /2) + ε/6d
≤ (1 + O(ε)) · 2 exp(−(t − 1)2 /2) + ε/6d
≤ (5/2) exp(−(t − 1)2 /2) + ε/5d .
If t ≥ 1 +
p
2 ln 6/ε, this expression is (5/2)(ε/6) + ε/5d ≤ 2ε/3.
Now we can finally show that a t exists for this r, so Step 8 will succeed:
p
Lemma 7.29. There is a t ≥ 1 + 2 ln(9/ε) such that
Pr ′
Pr ′ (r · (X − Y ) > t) < 2ε > 12 exp(−(t − 1)2 /4) + 3ε/d.
X∈u S
Y ∈u S
Proof. By Lemma 7.27, there exists a t ≥ 1 such that
′
Pr r · (X − µSP ) > 2t) > 12 exp(−(t − 1)2 /4) + 3ε/d.
X∈u T
′
′
Using the definition of T , the points when x = µSP or x = µSQ do not contribute to this probability so all
′
points in T that satisfy r · (x − µSP ) > 2t come from E. Since E ⊂ S ′ and |S ′ | = |T |, we have
′
′
(38)
Pr ′ r · (X − µSP ) > 2t ≥ Pr r · (X − µSP ) > 2t > 12 exp(−(t − 1)2 /4) + 3ε/d.
X∈u T
X∈u S
′
2
S
Noting that |E|/|S ′ | ≤ 4ε/3, all except a 4ε/3 fraction
p of points x ∈ T have r · (x − µ P ) = O(δ ). So,
2
4ε/3 ≥ 12 exp(−(t − 1) /4). Therefore, t ≥ 1 + 2 ln(9/ε).
′
Thus, by Lemma 7.28, we have (1 − |E|/|S ′ |) PrX∈u SP′ ∪SQ′ r · (X − µSP ) > t < 2ε/3. Again,
using that |E|/|S ′ | ≤ 4ε/3, we have that
Pr
X∈u S ′
′
r · (X − µSP ) > t < 2ε .
′
Consequently, if x satisfies r ·(x−µSP ) > 2t, then it satisfies PrY ∈u S ′ (r · (x − Y ) ≤ t) < 2ε. Substituting
this condition into Equation (38) gives the lemma.
Again we need to show that any filter does not remove too many points of S. We need to show this for
′
′
an arbitrary r, not just one nearly parallel to µSP − µSQ .
p
Lemma 7.30. For any unit vector r ′ and t ≥ 2 ln(1/ε), we have
′
′
(1 − |E|/|S |)
Pr′ ′
Pr ′ r · (X − Y ) ≤ t < 2ε ≤ 3 exp(−t2 /4) + ε/4d .
X∈u SP ∪SQ
Y ∈u S
79
Proof. Using Hoeffding’s inequality, we have
|SP′ | Pr ′ (r · (p − X) > t/2) ≤ |SP | Pr |r ′ · (X − p)| > t/2
X∈u SP
X∈u SP
≤ |SP | Pr (|r ′ · (X − p)| > t/2) + ε/6d
X∼P
≤ |SP | 2 exp(−t2 /4) + ε/6d .
(39)
Every point x with |r ′ · (x − p)| ≤ t/2 has |r ′ · (x − y) ≤ t| for all y with |r ′ · (y − p)| ≤ t/2. Thus,
for x with |r ′ · (x − p)| ≤ t/2, we have
Pr (r ′ · (x − Y ) ≤ t) ≥ (|SP |/|S ′ |) − (|SP |/|S ′ |)(2 exp(−t2 /4) + ε/6d) .
Y ∈u S ′
p
When t ≥ 2 ln(1/ε), (|SP /|S ′ |)(2 exp(−t2 /4)+ε/6d) ≤ (1+2ε)·(2ε+ε/6d) ≤ 3ε. Also, (|SP |/|S ′ |) ≤
((α − ε/6)|S|)/(|S|(1 − 2ε)) ≤ α − 3ε ≤ 7ε. Thus, we have PrY ∈u S ′ (r · (x − Y ) ≤ t) ≥ 4ε > 2ε.
But inequality (39) gives a bound on the number of x in SP that do not satisfy this condition. That is,
′
′
|SP | Pr ′
Pr ′ r · (X − Y ) ≤ t < 2ε ≤ |SP | 2 exp(−t2 /4) + ε/6d .
X∈u SP
Y ∈u S
Similarly, every point x with |r ′ · (x − q)| ≤ t/2 has
Pr (r ′ · (x − y) ≤ t) > 2ε
y∈u S ′
′ | Pr
′
2
′ (r · (X − p) > t/2) ≤ (2 exp(−t /4) + ε/6d). Thus,
and |SQ
X∈u SQ
′
|SQ
|
Pr
′
X∈u SQ
Pr ′ (r · (X − Y ) ≤ t) < 2ε ≤ |SQ | 2 exp(−t2 /4) + ε/6d .
′
Y ∈u S
Summing these gives
(|SP′ |
′
+ |SQ
|)
Pr
′ ∪S ′
X∈u SP
Q
Pr ′ (r · (X − Y ) ≤ t) < 2ε ≤ |S| 2 exp(−t2 /4) + ε/6d .
′
Y ∈u S
Dividing by |S ′ | and noting that |S| ≤ (1 + 2ε)|S ′ | ≤ (3/2)|S ′ | completes the proof.
Now, we can show that the filter improves ∆(S, S ′′ ), and so that the algorithm is correct in the filter
case.
Claim 7.31. If we reach Step 9 and return S ′′ , then ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − 2ε/d.
′′ ∪ E ′′ , where E ′′ has disjoint support from S \ S ′′ and S \ S ′′ . Note
Proof. We can write S ′′ = SP′′ ∪ SQ
P
Q
P
Q
′′ ⊆ S ′ and E ′′ ⊆ E. We assume
that, since we have S ′′ ⊂ S ′ , we can define these sets so that SP′′ ⊆ SP′ , SQ
Q
that we do. Now we have that
∆(S, S ′ ) − ∆(S, S ′′ ) =
Therefore,
′
′′
∆(S, S ) − ∆(S, S ) =
′′ \ S ′ |
|E ′′ \ E ′ | − |SP′′ \ SP′ | − |SQ
Q
|S|
.
′′ \ S ′ |)
|S ′′ \ S ′ | − 2(|SP′′ \ SP′ | + |SQ
Q
|S|
80
.
p
In Step 8, we found a vector r and t ≥ 1 + 2 ln(1/ε) such that
Pr ′
Pr ′ (|r · (X − Y )| < t) < 2ε > 12 exp(−(t − 1)2 /4) + 3ε/d .
X∈u S
Y ∈u S
Then in Step 9, we remove at least a 12 exp(−t2 /4) + 3ε/d fraction of points. That is,
|S ′′ \ S ′ | ≥ (12 exp(−t2 /4) + 3ε/d)|S ′ | .
p
The fact that t ≥ 1 + 2 ln(1/ε) allows us to use Lemma 7.30, with r ′ = r, yielding that:
′
(1 − |E|/|S |)
Pr′ ′
Pr ′ (r · (X − Y ) ≤ t − 1) < 2ε ≤ 3 exp(−(t − 1)2 /4) + ε/(4d) .
Y ∈u S
X∈u SP ∪SQ
This implies that
′
(1 − |E|/|S |)
Pr
′ ∪S ′
X∈u SP
Q
Pr ′ (r · (X − Y ) < t) < 2ε ≤ 3 exp(−(t − 1)2 /4) + ε/(4d) .
Y ∈u S
Thus,
′′
′
|SP′′ \ SP′ | + |SQ
\ SQ
| ≤ (3 exp(−(t − 1)2 /4) + ε/4d)|S ′ | ,
and we have
∆(S, S ′ ) − ∆(S, S ′′ ) ≥
12 exp(−(t − 1)2 /4) + 3ε/d − 2 3 exp(−(t − 1)2 /4) + ε/4d |S ′ |/|S|
≥ 2ε/d ,
since |S ′ | ≥ |S|(1 − ∆(S, S ′ )) ≥ (1 − 2ε)|S| ≥ 5|S|/6.
8 Agnostically Learning a Gaussian, via Filters
8.1 Learning a Gaussian With Unknown Mean
In this section, we use our filter technique to give an agnostic learning algorithm for an unknown mean
Gaussian with known covariance matrix. More specifically, we prove:
Theorem 8.1. Let G be a Gaussian with identity covariance matrix and ε, τ > 0. Let S be a multiset of
Ω(d3 log(d/ετ ))4 /ε2 ) independent samples from G and S ′ be a multiset obtained by arbitrarily changing
′
an ε-fraction of the points in S. There exists a polynomial time
p algorithm that returns a Gaussian G so that
′
with probability at least 1 − τ we have dTV (G, G ) = O(ε log(1/ε)).
′
Our output hypothesis will be a Gaussian G′
Let µG , µG denote the mean vectors of G, G′ respectively. p
′
G
G
with identity covariance matrix, such that kµ − µ k2 = O(ε log(1/ε)). The desired bound on the total
variation distance follows from Corollary 2.13.
1 P
1 P
G
G T
Notation. We will denote by µS = |S|
X∈S X and Cov[S] = |S|
X∈S (X − µ )(X − µ ) the sample
mean and sample covariance matrix of the set S.
We start by defining a condition on the uncorrupted set of samples S, under which our algorithm will
succeed.
Definition 8.2. Let G be a Gaussian in d dimensions with identity covariance and ε, τ > 0. We say that a
multiset S of elements in Rd is (ε, τ )-good with respect to G if the following conditions are satisfied:
81
p
(i) For all x ∈ S we have kx − µG k2 ≤ O( d log(|S|/τ )).
(ii) For every affine function L : Rd → R we have | PrX∈u S (L(X) ≥ 0) − PrX∼G (L(X) ≥ 0)| ≤
ε/(d log(d/ετ )) .
(iii) We have that kµS − µG k2 ≤ ε.
(iv) We have that kCov[S] − Ik2 ≤ ε/d.
Next, we show that a sufficiently large set of independent samples from G is (ε, τ )-good (with respect
to G) with high probability. The proof of the following lemma is deferred to Section F.
Lemma 8.3. Let G be a Gaussian with identity covariance and ε, τ > 0. If the multiset S is obtained by
taking Ω(d3 log(d/ετ )/ε2 ) independent samples from G, it is (ε, τ )-good with respect to G with probability
at least 1 − τ.
We recall our definition of ∆(S, S ′ ) that quantifies the extent to which a multiset has been corrupted.
Definition 8.4. Given finite multisets S and S ′ we let ∆(S, S ′ ) be the size of the symmetric difference of S
and S ′ divided by the cardinality of S.
As in the convex program case, we will first use NAIVE P RUNE to remove points which are far from the
mean. Then, we iterate the algorithm whose performance guarantee is given by the following:
Proposition 8.5. Let G be a Gaussian with identity covariance, ε > 0 be sufficiently small and τ > 0. Let
′
′
′
S be an (ε, τ )-good
p set with respect to G. Let S be any multiset with ∆(S, S ) ≤ 2ε and for any x, y ∈ S ,
kx − yk2 ≤ O( d log(dε/τ )). There exists a polynomial time algorithm F ILTER -G AUSSIAN -U NKNOWN M EAN that, given S ′ and ε > 0, returns one of the following:
p
(i) A mean vector µ such that dTV (N (µ, I), G) = O(ε log(1/ε)).
(ii) A multiset S ′′ ⊆ S ′ such that ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − ε/(d log(d/ετ )).
We start by showing how Theorem 8.1 follows easily from Proposition 8.5.
Proof of Theorem 8.1. By the definition of ∆(S, S ′ ), since S ′ has been obtained from S by corrupting an
ε-fraction of the points in S, we have that ∆(S, S ′ ) ≤ 2ε. By Lemma 8.3, the set S of uncorrupted samples
is (ε, τ )-good with respect to G with probability at least 1 − τ. We
phenceforth condition on this event.
Since S is (ε, τ )-good, all x ∈ S have kx − µG k2 ≤ O( d log |S|/τ ). Thus, the NAIVE P RUNE
′′
′′
′
procedure does not remove from S ′ any member of S. Hence,
p its output, S , has ∆(S, S ) ≤ ∆(S, S ) and
′′
for any x ∈ S , there is ap
y ∈ S with kx − ykp
2 ≤ O( d log |S|/τ ). By the triangle inequality, for any
x, z ∈ S ′′ , kx − zk2 ≤ O( d log |S|/τ ) = O( d log(d/ετ )).
Then we iteratively apply the F ILTER -G AUSSIAN -U NKNOWN -M EAN procedure
of Proposition 8.5 unp
til it terminates returning a mean vector µ with dTV (N (µ, I), G) = O(ε log(1/ε)). We claim that we
need at most O(d log(d/ετ )) iterations for this to happen. Indeed, the sequence of iterations results in a
sequence of sets Si′ , so that ∆(S, Si′ ) ≤ ∆(S, S ′ ) − i · ε/(d log(d/ετ )). Thus if we do not terminate in the
first 2d log(d/ετ ) iterations, in the next iteration the algorithm terminates outputting the sample mean of
S.
82
8.1.1
Algorithm F ILTER -G AUSSIAN -U NKOWN -M EAN: Proof of Proposition 8.5
In this subsection, we describe the efficient algorithm establishing Proposition 8.5 and prove its correctness.
Our filter algorithm is very simple: We consider the empirical distribution defined by the (corrupted) sample
′
multiset S ′ . We calculate its mean vector µS and covariance matrix M . If the matrix Σ has no large
′
eigenvalues, we return µS . Otherwise, we use the eigenvector v ∗ corresponding to the maximum magnitude
′
eigenvalue of Σ and the mean vector µS to define a filter. Our efficient filtering procedure is presented in
detailed pseudocode below.
Algorithm 14 Filter algorithm for a Gaussian with unknown mean and identity covariance
1: procedure F ILTER -G AUSSIAN -U NKNOWN -M EAN (S ′ , ε, τ )
input: A multiset S ′ such that there exists an (ε, τ )-good S with ∆(S, S ′ ) ≤ 2ε
output: Multiset S ′′ or mean vector p′ satisfying Proposition 8.5
′
2:
Compute the sample mean µS = EX∈u S ′ [X] and the sample covariance matrix Σ , i.e., Σ =
′
′
(Σi,j )1≤i,j≤d with Σi,j = EX∈u S ′ [(Xi − µSi )(Xj − µSj )].
3:
Compute approximations for the largest absolute eigenvalue of Σ − I, λ∗ := kΣ − Ik2 , and the
associated unit eigenvector v ∗ .
′
4:
if kΣ − Ik2 ≤ O(ε log(1/ε)), then return µS .
p
5:
Let δ := 3 εkΣ − Ik2 . Find T > 0 such that
′
Pr (|v ∗ · (X − µS )| > T + δ) > 8 exp(−T 2 /2) + 8ε/(d log(dε/τ )).
X∈u S ′
′
return the multiset S ′′ = {x ∈ S ′ : |v ∗ · (x − µS )| ≤ T + δ}.
6:
The rest of this section is devoted to the analysis of our algorithm F ILTER -G AUSSIAN -U NKOWN M EAN.
8.1.2
Proof of Correctness of F ILTER -G AUSSIAN -U NKOWN -M EAN
By definition, there exist disjoint multisets L, E, of points in Rd , where L ⊂ S, such that S ′ = (S \ L) ∪ E.
′
With this notation, we can write ∆(S, S ′ ) = |L|+|E|
|S| . Our assumption ∆(S, S ) ≤ 2ε is equivalent to
|L| + |E| ≤ 2ε · |S|, and the definition of S ′ directly implies that (1 − 2ε)|S| ≤ |S ′ | ≤ (1 + 2ε)|S|.
Throughout the proof, we assume that ε is a sufficiently small constant.
′
We define µG , µS , µS , µL , and µE to be the means of G, S, S ′ , L, and E, respectively.
Our analysis will make essential use of the following matrices:
• MS ′ denotes EX∈u S ′ [(X − µG )(X − µG )T ],
• MS denotes EX∈u S [(X − µG )(X − µG )T ],
• ML denotes EX∈u L [(X − µG )(X − µG )T ], and
• ME denotes EX∈u E [(X − µG )(X − µG )T ].
Our analysis will hinge on proving the important claim that Σ − I is approximately (|E|/|S ′ |)ME . This
means two things for us. First, it means that if the positive errors align in some direction (causing ME to
have a large eigenvalue), there will be a large eigenvalue in Σ − I. Second, it says that any large eigenvalue
of Σ − I will correspond to an eigenvalue of ME , which will give an explicit direction in which many error
points are far from the mean.
83
We begin by noting that we have concentration bounds on G and therefore, on S due to its goodness.
Fact 8.6. Let w ∈ Rd be any unit vector, then for any T > 0, PrX∈u S (|w · (X − µG )| > T ) ≤
2 exp(−T 2 /2) + ε/(d log(d/ετ )), and PrX∼G (|w · (X − µG )| > T ) ≤ 2 exp(−T 2 /2).
Proof. The second line is a standard concentration result for Gaussians, and the former follows from it using
the goodness of S.
By using the above fact, we obtain the following simple claim:
Claim 8.7. Let w ∈ Rd be any unit vector, then for any T > 0,
′
′
Pr (|w · (X − µS )| > T + kµS − µG k2 ) ≤ 2 exp(−T 2 /2) + ε/(d log(d/ετ )).
X∈u S
and
′
′
Pr (|w · (X − µS )| > T + kµS − µG k2 ) ≤ 2 exp(−T 2 /2).
X∼G
′
′
Proof. This follows from Fact 8.6 upon noting that |w·(X−µS )| > T +kµS −µG k2 only if |w·(X−µG )| >
T.
We can also use this to prove concentration bounds for L. In particular:
Corollary 8.8. kML k2 = O(log(|L|/|S|)).
Proof. We need to bound v T ML v = EX∈u L [(v · (X − µG ))2 ] for all unit vectors v. This is
Z
∞
0
G
Pr (|v · (X − µ )| > T )2T dT =
X∈u L
Z
O(d log(d/ετ ))
Pr (|v · (X − µG )| > T )2T dT.
X∈u L
0
Our bound then follows from noting that, since L is a subset of S, we have that
G
G
Pr (|v · (X − µ )| > T ) ≤ min 1, (|S|/|L|) · Pr (|v · (X − µ )| > T ) ,
X∈u L
X∈u S
and using Claim 8.7.
Next, we prove a bound on MS ′ .
Lemma 8.9. We have that
MS ′ = I + (|E|/|S ′ |)ME + O(ε log(1/ε)) ,
where the O(ε log(1/ε)) term denotes a matrix of spectral norm O(ε log(1/ε)).
Proof. Notice that
|S ′ |MS ′ = |S|MS − |L|ML + |E|ME
= |S|(I + O(ε)) − O(ε log(1/ε)) + |E|ME ,
where the second line is by the goodness of S. Therefore, we have that
MS ′ = I + (|E|/|S ′ |)ME + O(ε log(1/ε)).
84
We get a similarly useful bound on the difference between the mean vectors:
Lemma 8.10. We have that
p
′
(µS − µG ) = (|E|/|S ′ |)(µE − µG ) + O(ε log(1/ε)) ,
p
p
where the O(ε log(1/ε)) denotes a vector with ℓ2 norm at most O(ε log(1/ε)).
Proof.
′
|S ′ |(µS − µG )) = |S|(µS − µG ) − |L|(µL − µG ) + |E|(µE − µG ).
S
G
L
G
L
G T
L
G
Since
p S is good kµ − µ k2 = O(ε). Since kML k ≫ (µ − µ )(µ − µ ) , we have that kµ − µ k2 ≤
O( log(|L|/|S|)). Therefore, we have that
p
′
(µS − µG ) = (|E|/|S ′ |)(µE − µG ) + O(ε log(1/ε)) ,
as desired.
From this we conclude that Σ − I is approximately proportional to ME .
Corollary 8.11.
Σ − I = (|E|/|S ′ |)ME + O(ε log(1/ε)) + O(|E|/|S|)2 kME k2 .
Proof. We have that
′
′
Σ − I = MS ′ − I − (µS − µG )(µS − µG )T .
By Lemmas 8.9 and 8.10, this is
(|E|/|S ′ |)ME + O(ε log(1/ε)) + O((|E|/|S|)2 kµE − µG k22 ) = (|E|/|S ′ |)ME + O(ε log(1/ε)) + O(|E|/|S|)2 kME k2 .
From this we can immediately see that if our algorithm returns on Step 4, then it returns a correct answer.
In particular, if λ∗ = kΣ − Ik2 = O(ε log(1/ε)), we have by Corollary 8.11 that
|E|/|S ′ |kME k2 ≤ λ∗ + O(ε log(1/ε)) + O(|E|/|S|)2 kME k2 .
Therefore,
kME k2 = O(ε log(1/ε)).
On the other hand, we have kME k2 ≥ kµE − µG k22 , and therefore by Lemma 8.10,
p
p
p
′
kµS − µG k2 ≤ (|E|/|S ′ |)kME k2 + O(ε log(1/ε)) = O(ε log(1/ε)).
Next, we need to show that if λ∗ > Cε log(1/ε) that a T as in Step 5 exists. To show this, we note again
by Corollary 8.11 that
λ∗ = (v ∗ )T (Σ − I)v ∗ = (|E|/|S ′ | + O(|E|/|S|)2 )(v ∗ )T ME v ∗ + O(ε log(1/ε)).
Therefore,
E [|v ∗ · (X − µG )|2 ] = (v ∗ )T ME v ∗ ≫ (|S ′ |/|E|)λ∗ .
X∈u E
85
On the other hand,
∗
G 2
E [|v · (X − µ )| ] =
X∈u E
Z
√
O(
d log(d/ετ ))
0
≤ |S ′ |/|E|
Z
√
O(
√
Pr (|v ∗ · (X − µG )| > T )2T dT
X∈u E
d log(d/ετ ))
Pr (|v ∗ · (X − µG )| > T )2T dT + λ∗ /100.
X∈u S ′
λ∗ /10
√
Therefore, it must be the case that for some value of T ′ > λ∗ /10 for which PrX∈u S ′ (|v ∗ · (X − µG )| >
T ′ ) > exp(−T ′2 /4) + 8ε/(d log(dε/τ )). Letting T = T ′ − δ yields the necessary T .
Finally, we need to show that if our algorithm returns an S ′′ in Step 6 that ∆(S, S ′′ ) < ∆(S, S ′ ) −
2ε/(d log(d/ετ )).
Claim 8.12. We have that ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − 2ε/(d log(d/ετ )) .
Proof. Recall that S ′ = (S \ L) ∪ E, with E and L disjoint multisets such that L ⊂ S. We can similarly
write S ′′ = (S \ L′ ) ∪ E ′ , with L′ ⊇ L and E ′ ⊂ E. Since
∆(S, S ′ ) − ∆(S, S ′′ ) =
|E \ E ′ | − |L′ \ L|
,
|S|
it suffices to show that |E \ E ′ | ≥ |L′ \ L| + ε|S|/d. Note that |L′ \ L| is the number of points rejected
by the filter that lie in S ∩ S ′ . Note that the fraction of elements of S that are removed to produce S ′′
(i.e., satisfy |v ∗ · (x − µS ′ )| > T + δ) is at most 2 exp(−T 2 /2) + ε/(d log(d/ετ )). Hence, it holds that
|L′ \ L| ≤ (2 exp(−T 2 /2) + ε/(d log(d/ετ )))|S|. On the other hand, Step 5 of the algorithm ensures that
the fraction of elements of S ′ that are rejected by the filter is at least 8 exp(−T 2 /2) + 8ε/(d log(d/ετ )).
Note that |E \ E ′ | is the number of points rejected by the filter that lie in S ′ \ S. Therefore, we can write:
|E \ E ′ | ≥ (8 exp(−T 2 /2) + 8ε/(d log(d/ετ )))|S ′ | − (2 exp(−T 2 /2) + ε/(d log(d/ετ )))|S|
≥ (8 exp(−T 2 /2) + 8ε/(d log(d/ετ )))|S|/2 − (2 exp(−T 2 /2) + ε/(d log(d/ετ )))|S|
≥ (2 exp(−T 2 /2) + 3ε/(d log(d/ετ )))|S|
≥ |L′ \ L| + 2ε|S|/(d log(d/ετ )) ,
where the second line uses the fact that |S ′ | ≥ |S|/2 and the last line uses the fact that |L′ \ L|/|S| ≤
(2 exp(−T 2 /2) + ε/(d log(d/ετ ))). Noting that log(d/ετ ) ≥ 1, this completes the proof of the claim.
8.2 Learning a Gaussian With Unknown Covariance
In this subsection, we use our filter technique to agnostically learn a Gaussian with zero mean vector and
unknown covariance. By combining the algorithms of the current and the previous subsections, as in our
convex programming approach (Section 4.5), we obtain a filter-based algorithm to agnostically learn an
arbitrary unknown Gaussian.
The main result of this subsection is the following theorem:
Theorem 8.13. Let G be a Gaussian in d dimensions with mean 0 and unknown covariance, and let ε, τ > 0.
Let S be a set of N independent samples from G for N = Ω((d log(d/ετ ))6 /ε2 ). Let S ′ be a set of N
elements of Rd obtained from S by adversarially changing at most an ε-fraction of the samples of S. There
exists an algorithm A that, given S ′ and ε, returns the parameters of a Gaussian distribution G′ so that with
probability at least 1 − τ over the samples from S, it holds dTV (G, G′ ) ≤ δ = O(ε log(1/ε)).
As in the previous subsection, we will need a condition on S under which our algorithm will succeed.
86
Definition 8.14. Let G be a d-dimensional Gaussian with mean 0 and covariance matrix Σ. Let S be a set
of N points of Rd , and ε, τ > 0. We say that S is (ε, τ )-good with respect to G if the following conditions
hold:
• For all x ∈ S we have that xT Σ−1 x = O(d log(N/τ )).
d
• For
p any even degree-2 polynomial p : R → R we have that |EX∼G [p(X)] − EX∈u S [p(X)]| ≤
ε EX∼G [p(X)2 ] , EX∼G [p2 (X)] − EX∈u S [p2 (X)] ≤ ε EX∼G [p(X)2 ] , and
Pr (p(X) ≥ 0) ≤ 2 Pr (p(X) ≥ 0) + ε/(d log(N/τ ))2 .
X∈u S
X∼G
Let us first note some basic properties of such polynomials on a normal distribution. The proof of this
lemma is deferred to Section F.
Lemma 8.15. For any even degree-2 polynomial p : Rd → R, we can write p(x) = (Σ−1/2 x)T P2 (Σ−1/2 x)+
p0 , for a d × d symmetric matrix P2 and p0 ∈ R. Then, for X ∼ G, we have
1. E[p(X)] = p0 + tr(P2 ),
2. Var[p(X)] = 2kP2 k2F and
3. For all T > 1, Pr(|p(X) − E[p(X)]| ≥ T ) ≤ 2e1/3−2T /3 Var[p(X)] .
4. For all δ > 0, Pr(|p(X)| ≤ δ2 ) ≤ O(δ).
We note that, if S is obtained by taking random samples from G, then S is good with high probability.
The proof of this lemma is also deferred to Section F.
Lemma 8.16. Let G be a d-dimensional Gaussian with mean 0 and let ε, τ > 0. Then, for Ω((d log(d/ετ ))6 /ε2 ) ≤
N ≤ poly(d/ετ ), we have that if S is obtained by taking N independent samples from G, then S is (ε, τ )good with respect to G with probability at least 1 − τ .
As in Definition 8.4, ∆(S, S ′ ) is the be the size of the symmetric difference of S and S ′ divided by |S|.
The basic thrust of our algorithm is as follows. By Lemma 8.16, with probability 1 − τ we have that S
is (ε, τ )-good with respect to G. The algorithm is then handed a new set S ′ so that ∆(S, S ′ ) ≤ 2ε|S|. The
algorithm will run in stages. In each stage the algorithm will either output G′ or will return a new set S ′′
so that ∆(S, S ′′ ) < ∆(S, S ′ ). In the latter case, the algorithm will recurse on S ′′ . We formalize this idea
below:
Proposition 8.17. There is an algorithm that given a set finite S ′ ⊂ Rd , so that there is a mean 0 Gaussian
G and a set S that is (ε, τ )-good with respect to G with ∆(S, S ′ ) ≤ 2ε|S|, runs in time poly(d log(1/τ )/ε)
and returns either the parameters of a Gaussian G′ with dTV (G, G′ ) ≤ O(ε log(1/ε)) or a subset S ′′ of Rd
with ∆(S, S ′′ ) < ∆(S, S ′ ).
Given Proposition 8.17, the proof of Theorem 8.13 is straightforward. By Lemma 8.16 the original set
S is (ε, τ )-good with respect to G with probability at least 1 − τ . Then, S ′ satisfies the hypotheses of Proposition 8.17. We then repeatedly iterate the algorithm from Proposition 8.17 until it outputs a distribution G′
close to G. This must eventually happen because at every step the distance between S and the set returned
by the algorithm decreases by at least 1.
87
8.2.1
Analysis of Filter-based Algorithm: Proof of Proposition 8.17
We now turn our attention to the proof of Proposition 8.17. We first define the matrix Σ′ to be EX∈S ′ [XX T ],
and let G′ be the mean 0 Gaussian with covariance matrix Σ′ . Our goal will be to either obtain proof that G′
is close to G or to devise a filter that allows us to clean up S ′ by removing some elements, most of which
are not in S. The idea here is the following: We know by Corollary 2.14 that G and G′ are close unless
I − Σ−1/2 Σ′ Σ−1/2 has large Frobenius norm. This happens if and only if there is some matrix M with
kM kF = 1 so that
tr(M Σ−1/2 Σ′ Σ−1/2 − M ) =
E [(Σ−1/2 X)T M (Σ−1/2 X) − tr(M )]
X∈u S ′
is far from 0. On the other hand, we know that the distribution of p(X) = (Σ−1/2 X)T M (Σ−1/2 X)−tr(M )
for X ∈u S is approximately that of p(G), which is a variance O(1) polynomial of Gaussians with mean
0. In order to substantially change the mean of this function, while only changing S at a few points, one
must have several points in S ′ for which p(X) is abnormally large. This in turn will imply that the variance
of p(X) for X from S ′ must be large. This phenomenon will be detectable as a large eigenvalue of the
matrix of fourth moments of X ∈ S ′ (thought of as a matrix over the space of second moments). If such a
large eigenvalue is detected, we will have a p with p(X) having large variance. By throwing away from S ′
elements for which |p| is too large, we will return a cleaner version of S ′ . The algorithm is as follows:
Algorithm 15 Filter algorithm for a Gaussian with unknown covariance matrix.
1: procedure F ILTER -G AUSSIAN -U NKNOWN -C OVARIANCE(S ′ , ε, τ )
input: A multiset S ′ such that there exists an (ε, τ )-good S with ∆(S, S ′ ) ≤ 2ε
output: Either a set S ′′ with ∆(S, S ′′ ) < ∆(S, S ′ ) or the parameters of a Gaussian G′ with dTV (G, G′ ) =
O(ε log(1/ε)).
Let C > 0 be a sufficiently large universal constant.
2:
Let Σ′ be the matrix EX∈u S ′ [XX T ] and let G′ be the mean 0 Gaussian with covariance matrix Σ′ .
3:
if there is any x ∈ S ′ so that xT (Σ′ )−1 x ≥ Cd log(|S ′ |/τ ) then
4:
return S ′′ = S ′ − {x : xT (Σ′ )−1 x ≥ Cd log(|S ′ |/τ )}.
5:
Let L be the space of even degree-2 polynomials p so that EX∼G′ [p(X)] = 0.
6:
Define two quadratic forms on L
(i) QG′ (p) = E[p2 (G′ )] ,
(ii) QS ′ (p) = EX∈u S ′ [p2 (X)] .
7:
8:
9:
10:
11:
Computing maxp∈L−{ok QS ′ (p)/QG′ (p) and the associated polynomial p∗ (x) normalized so that
QG′ (p∗ ) = 1 using F IND - MAX - POLY below.
if QS ′ (p∗ ) ≤ (1 + Cε log2 (1/ε))QG′ (p∗ ) then
return G′
Let µ be the median value of p∗ (X) over X ∈ S ′ .
Find a T ≥ C ′ so that
Pr (|p∗ (X) − µ| ≥ T ) ≥ 12 exp(−(T − 4/3) + 3ε/(d log(N/τ ))2
X∈u S ′
12:
return S ′′ = {X ∈ S ′ : |p∗ (X) − µ| < T }.
Thr function F IND - MAX - POLY uses similar notation to S EPARATION O RACLE U NKNOWN C OVARIANCE
so that F ILTER -G AUSSIAN -U NKNOWN -C OVARIANCE and S EPARATION O RACLE U NKNOWN C OVARIANCE
can be more easily compared.
88
Algorithm 16 Algorithm for maximizing QS ′ (p)/QG′ (p).
1: function F IND - MAX - POLY (S ′ , Σ′ )
input: A multiset S ′ and a Gaussian G′ = N (0, Σ′ )
output: The even degree-2 polynomial p∗ (x) with EX∼G′ [p∗ (X)] ≈ 0 and QG′ (p∗ ) ≈ 1 that approximately maximizes QS ′ (p∗ ) and this maximum λ∗ = QS ′ (p∗ ).
2:
Compute an approximate eigendecomposition of Σ′ and use it to compute Σ′−1/2
3:
Let x(1) , . . . , x(|S ′ |) be the elements of S ′ .
⊗2
.
4:
For i = 1, . . . , |S ′ |, let y(i) = Σ′−1/2 x(i) and z(i) = y(i)
′|
P
|S
T .
5:
Let TS ′ = I ♭ I ♭T + (1/|S ′ |) i=1 z(i) z(i)
∗
6:
Approximate the top eigenvalue λ and corresponding unit eigenvector v ∗ of TS ′ ..
7:
Let p∗ (x) = √12 ((Σ′−1/2 x)T v ∗♯ (Σ′−1/2 x) − tr(v ∗♯ ))
8:
return p∗ and λ∗ .
Let us first show that F IND - MAX - POLY is correct.
Claim 8.18. Algorithm F IND - MAX - POLY is correct and F ILTER -G AUSSIAN -U NKNOWN -C OVARIANCE
runs time poly(d log τ /ε).
Proof. First, assume that we can compute all eigenvalues and eigenvectors exactly. By Lemma 8.15 all even
polynomials with degree-2 that have EX∼G [p(X)] = 0√
can be written as p(x) = (Σ′−1/2 x)T P2 (Σ′−1/2 x) −
♯
tr(P2 ) for a symmetric matrix P2 . If we take P2 = v / 2 for a unit vector v such that v ♯ is symmetric, then
VarX∼G′ [p(X)] = 2kP2 kF = kv2 k = 1.
Note that since the covariance matrix of S ′ is Σ′ , we have
E [p(X)] = E ′ [(Σ′−1/2 X)T P2 (Σ′−1/2 X) − tr(P2 )]
X∼S ′
X∼S
= E ′ [tr((XX T )Σ′−1/2 P2 Σ′−1/2 )] − tr(P2 )
X∼S
= tr( E ′ [(XX T )]Σ′−1/2 P2 Σ′−1/2 ) − tr(P2 )
X∼S
′ ′−1/2
= tr(Σ Σ
P2 Σ′−1/2 ) − tr(P2 ) = 0 .
We let√T ′ be the multiset of y = Σ−1/2 x for x ∈ S ′ and U ′ the multiset of z = y ⊗2 for y in T ′ . Recall that
P2♭ = 2v. We thus have
QS ′ (p) :=
=
=
=
E [p(X)2 ] = E ′ [(Y T P2 Y − tr(P2 ))2 ]
X∈u S ′
T
2
Y ∼T
E [(Y P2 Y ) ] + tr(P2 )2 − 2tr(P2 ))2 ]
Y ∈u T ′
E [tr(((Y Y T )P2 )2 ] + tr(P2 I)2 − 0
Y ∈u T ′
E [2(Z T v)2 ] + 2(v T I ♭ )2
Z∈u U ′
E [2v T (ZZ T )v] + 2v T (I ♭ I ♭T )v
Z∈u U ′
T
= 2v TS ′ v .
Thus, the p(x) that maximizes QS ′ (p) is given by the unit vector v that maximizes v T TS ′ v subject to v ♯
being symmetric.
89
Let v ′ = v ♯T ♭ . Note that v T TS ′ v = v ′T TS ′ v ′ by symmetries of TS ′ . Thus, by linearity, v ′′ = v/2 + v ′ /2
also has v ′′T TS ′ v ′′ = v T TS ′ v. However, if v ♯ is not symmetric, v ′′ has kv ′′ k2 < 1. Thus, the unit vector
v ′′ /kv ′′ k2 achieves a higher value of the bilinear form. Consequently, v ∗♯ is symmetric.
Now we have that p∗ (x) that maximizes QS ′ (p) is given by the unit vector v that maximizes v T TS ′ v.
Since QG′ (p) := EX∼G′ [p(X)2 ] = 2kP2 kF = kvk2 = 1, this also maximizes QS ′ (p)/QG′ (p).
We note that we can achieve EX∼G′ [p∗ (X)] = O(ε2 ) and EX∼G′ [(p∗ (X))2 ] = 1 + O(ε2 ) in time
poly(ε/d) using standard algorithms to compute the eigen-decomposition of a symmetric matrix. This
suffices for the correctness of the remaining part of F ILTER -G AUSSIAN -U NKNOWN -C OVARIANCE The
other steps in F ILTER -G AUSSIAN -U NKNOWN -C OVARIANCE can be easily done in poly(|S ′ |d log(1τ )/ε)
time.
In order to analyze algorithm F ILTER -G AUSSIAN -U NKNOWN -C OVARIANCE, we note that we can write
S ′ = (S \ L) ∪ E where L = S \ S ′ and L = S ′ \ S. It is then the case that ∆(S, S ′ ) = (|L| + |E|)/|S|.
Since this is small we have that |L|, |E| = O(ε|S ′ |). We can also write Σ′ and ΣS\L ((|S| − |L|)/|S ′ |) +
ΣE (|E|/|S ′ |) = ΣS\L + O(ε)(ΣE − ΣS\L ), where ΣS\L = EX∈u S\L [XX T ], ΣE = EX∈u E [XX T ]. A
critical part of our analysis will be to note that ΣS\L is very close to Σ, and thus that either Σ′ is very close
to Σ or else ΣE is very large in some direction.
Lemma 8.19. We have that
kI − Σ−1/2 ΣS\L Σ−1/2 kF = O(ε log(1/ε).
Proof. Note that, since the matrix inner product is an inner product,
tr(M Σ−1/2 ΣS\L Σ−1/2 ) − tr(M ) .
kI − Σ−1/2 ΣS\L Σ−1/2 kF = sup
kM kF =1
We need to show that for any M with kM kF = 1 that tr(M Σ−1/2 ΣS\L Σ−1/2 ) − tr(M ) is small.
Since tr(M Σ−1/2 ΣS\L Σ−1/2 ) = tr(M T Σ−1/2 ΣS\L Σ−1/2 ) = tr( 12 (M + M T )Σ−1/2 ΣS\L Σ−1/2 ) and
1
k 2 (M + M T )kF ≤ 21 (kM kF + kM T kF ) = 1, we may assume WLOG that M is symmetric.
Consider such an M . We note that
tr(M Σ−1/2 ΣS\L Σ−1/2 ) =
E
[tr(M Σ−1/2 XX T Σ−1/2 )] =
X∈u S\L
E
X∈u S\L
[(Σ−1/2 X)T M (Σ−1/2 X)].
Let p(x) denote the quadratic polynomial
p(x) = (Σ−1/2 x)T M (Σ−1/2 x) − tr(M ).
By Lemma 8.15, EX∼G [p(X)] = 0 and VarX∼G [p(X)] = 2kM k2F = 2. p
Since S is (ε, τ )-good with respect to G, we have that EX∈S [p(X)] = ε EX∼G [p2 (X)] = O(ε).
Therefore, it suffices to show that the contribution from L is small. In particular, it will be enough to show
that
P
X∈L p(X)
= O(ε log(1/ε)).
|S|
This holds essentially because the distribution of p(X) for X ∈ S is close to that for p(G), which has
rapidly decaying tails. Therefore, throwing away an ε-fraction of the mass cannot change the value of the
90
mean by very much. In particular, we have that
P
Z ∞
|{x ∈ L : p(x) > t}|
x∈L p(x)
≤
dt
|S|
|S|
0
Z ∞
min(ε, Pr (p(X) > t))dt
≤
≤
Z
X∈u S
0
O(d log(N/τ ))
0
≤ε+
≤ε+
Z
0
Z
∞
min(ε, 2 Pr(p(G) > t)) + ε/(d log(N/τ ))2 dt
min(ε, 2 exp(−Ω(t − 1)))dt
C log(1/ε)
εdt +
Z
∞
2 exp(−Ω(t))dt
C log(1/ε)−1
0
≤ ε + O(ε log(1/ε)) + O(ε)
= O(ε log(1/ε)) ,
where C above is a sufficiently large constant. The third line uses that
max |p(x)| ≤ max kΣ−1/2 xk22 kM k2 + trM ≤ O(d log(N/τ )) · 1 +
x∈S
√
x∈S
d = O(d log(N/τ )) .
This completes the proof.
As a corollary of this we note that Σ′ cannot be too much smaller than Σ.
Corollary 8.20.
Σ′ ≥ (1 − O(ε log(1/ε)))Σ.
Proof. Lemma 8.19 implies that Σ−1/2 ΣS\L Σ1/2 has all eigenvalues in the range 1 ± O(ε log(1/ε). Therefore, ΣS\L ≥ (1+O(ε log(1/ε)))Σ. Our result now follows from noting that Σ′ = ΣS\L ((|S|−|L|)/|S ′ |)+
ΣE (|E|/|S ′ |), and ΣE = EX∈u E [XX T ] ≥ 0.
The first step in verifying correctness is to note that if our algorithm returns on Step 4 that it does so
correctly.
Claim 8.21. If our algorithm returns on Step 4, then ∆(S, S ′′ ) < ∆(S, S ′ ).
Proof. This is clearly true if we can show that all x removed have x 6∈ S. However, this follows because
(Σ′ )−1 ≤ 2Σ−1 , and therefore, by (ε, τ )-goodness, all x ∈ S satisfy
xT (Σ′ )−1 x ≤ 2xT Σ−1 x = O(d log(N/τ )) < Cd log(N/τ )
for C sufficiently large.
Next, we need to show that if our algorithm returns a G′ in Step 9 that dTV (G, G′ ) is small.
Claim 8.22. If our algorithm returns in Step 9, then dTV (G, G′ ) = O(ε log(1/ε)).
Proof. By Corollary 2.14, it suffices to show that
kI − Σ−1/2 Σ′ Σ−1/2 kF = O(ε log(1/ε)).
91
However, we note that
kI − Σ−1/2 Σ′ Σ−1/2 kF ≤ kI − Σ−1/2 ΣS\L Σ−1/2 kF + (|E|/|S ′ |)kI − Σ−1/2 ΣE Σ−1/2 kF
≤ O(ε log(1/ε)) + (|E|/|S ′ |)kI − Σ−1/2 ΣE Σ−1/2 kF .
Therefore, we will have an appropriate bound unless kI − Σ−1/2 ΣE Σ−1/2 kF = Ω(log(1/ε)).
Next, note that there is a matrix M with kM kF = 1 so that
kI − Σ−1/2 ΣE Σ−1/2 kF = tr(M Σ−1/2 ΣE Σ−1/2 − M ) =
E [(Σ−1/2 X)T M (Σ−1/2 X) − tr(M )].
X∈u E
Indeed we can take M = (I − Σ−1/2 ΣE Σ−1/2 )/kI − Σ−1/2 ΣE Σ−1/2 kF . Thus, there is a symmetric M
such that this holds.
Letting p(X) be the polynomial
p(X) = (Σ−1/2 X)T M (Σ−1/2 X) − tr(M ),
Using Lemma 8.15, EX∼G [p(X)] = 0 and VarX∼G [p(X)] = 2. Therefore, p ∈ L and QG′ (p) = 2. We
now compare this to the size of QS ′ (p). On the one hand, we note that using methodology similar to that
used in Lemma 8.19 we can show that EX∈u S\L [p2 (X)] is not much less than 2. In particular,
P
2
X∈L p (X)
2
2
E [p (X)] ≥
E [p (X)] −
.
X∈u S
|S|
X∈u S\L
On the one hand, we have that
E [p2 (X)] ≤ E[p2 (G)](1 + ε) = 2 + O(ε) ,
X∈u S
by assumption. On the other hand, we have that
P
Z ∞
2
|{X ∈ L : p2 (X) > t}|
X∈L p (X)
≤
dt
|S|
|S|
0
Z ∞
min(ε, Pr (p2 (X) > t))dt
≤
≤
Z
X∈u S
0
O(d log(N/τ ))2
0
≤ε+
≤ε+
Z
0
Z
∞
min(ε, 2 Pr(p2 (G) > t)) + ε/(d log(N/τ )2 )dt
√
min(ε, 2 exp(−Ω( t − 1)))dt
C log2 (1/ε)
εdt +
0
2
Z
∞
√
2 exp(−Ω( t))dt
C log2 (1/ε)−1
≤ ε + O(ε log (1/ε)) + O(ε log(1/ε))
= O(ε log2 (1/ε)).
Therefore, we have that EX∈u S\L [p2 (X)] = 2 + O(ε log2 (1/ε)). Since, by assumption QS ′ (p) ≤ 2 +
O(ε log2 (1/ε)), this implies that (|E|/|S ′ |) EX∈u E [p2 (X)] = O(ε log2 (1/ε)). By Cauchy-Schwartz, this
implies that
r
p
(|E|/|S ′ |) E [p(X)] ≤ (|E|/|S ′ |) (|E|/|S ′ |) E [p2 (X)] = O(ε log(1/ε)).
X∈u E
X∈u E
Thus,
(|E|/|S ′ |)kI − Σ−1/2 ΣE Σ−1/2 kF = O(ε log(1/ε)).
This shows that if the algorithm returns in this step, it does so correctly.
92
Next, we need to show that if the algorithm reaches Step 11 that such a T exists.
Claim 8.23. If the algorithm reaches Step 11, then there exists a T > 1 so that
Pr (|p(X) − µ| ≥ T ) ≥ 12 exp(−(T − 1)/3) + 3ε/(d log(N/τ ))2 .
X∈u S ′
Proof. Before we begin, we will need the following critical Lemma:
Lemma 8.24.
Var [p(X)] ≤ 1 + O(ε log(1/ε)).
X∼G
Proof. We note that since VarX∼G′ (p(G′ )) = QG′ (p) = 1, that we just need to show that the this variance
with respect to G instead of G′ is not too much larger. This will essentially be because the covariance matrix
of G cannot be much bigger than the covariance matrix of G′ by Corollary 8.20.
Using Lemma 8.15, we can write
p(x) = (Σ′−1/2 x)T P2 (Σ′−1/2 x) + p0 ,
where kP2 kF = 21 VarX∼G′ (p(G′ )) = 21 and p0 = µ + tr(P2 ). We can also express p(x) in terms of
G as p(x) = (Σ−1/2 x)T M (Σ−1/2 x) + p0 , and have VarX∼G [p(X)] = kM kF . Here, M is the matrix Σ1/2 Σ′−1/2 P2 Σ′−1/2 Σ1/2 . By Corollary 8.20, it holds Σ′ ≥ (1 − O(ε log(1/ε)))Σ. Consequently,
Σ1/2 Σ′−1/2 ≤ (1+O(ε log(1/ε)))I, and so kΣ1/2 Σ′−1/2 k2 ≤ 1+O(ε log(1/ε)). Similarly, kΣ′−1/2 Σ1/2 k2 ≤
1 + O(ε log(1/ε)).
We claim that
Pthen kABkF ≤ kAk2 kBkF . If Bj are the columns of B, then we
Pif A, B are matrices,
have kABk2F = j kABj k22 ≤ kAk22 j kBj k22 = (kAk2 kBkF )2 . Similarly for rows, we have kABkF ≤
kAkF kBk2 .
Thus, we have
Var [p(X)] = 2kM kF ≤ 2kΣ1/2 Σ′−1/2 k2 kP2 kF kΣ′−1/2 Σ1/2 k2 ≤ 1 + O(ε log(1/ε)) .
X∼G
Next, we need to consider µ. In particular, we note that by the similarity of S and S ′ , µ must be between
the 40 and 60 percentiles of values of p(X) for X ∈ S. However, since S is (ε, τ )-good, this must be
between the 30 and 70 percentiles of p(G). Therefore, by Cantelli’ s inequality,
q
(40)
|µ − E [p(X)]| ≤ 2 Var [p(X)] ≤ 3 .
X∼G
X∼G
We are now ready to proceed. Our argument will follow by noting that while QS ′ (p) is much larger than
expected, very little of this discrepancy can be due to points in S \ L. Therefore, the points of E must
provide a large contribution. Given that there are few points in E, much of this contribution must come
from there being many points near the tails, and this will guarantee that some valid threshold T exists.
Let µ̂ be EX∼G [p(X)]. In particular, we have that VarX∈u S ′ (p(X)) = QS ′ (p) ≥ 1 + Cε log2 (1/ε),
which means that
P
2
X∈S ′ |p(X) − µ̂|
≥ 1 + Cε log2 (1/ε).
|S ′ |
Now, because S is good, we know that
P
2
X∈S |p(X) − µ̂|
= E[|p(G) − µ̂|2 ](1 + O(ε)) = Var [p(X)](1 + O(ε)) ≤ 1 + O(ε log(1/ε)).
X∼G
|S|
93
Therefore, we have that
P
X∈S\L |p(X)
|S ′ |
− µ̂|2
≤ 1 + O(ε log(1/ε)).
Hence, for C sufficiently large, it must be the case that
X
|p(X) − µ̂|2 ≥ (C/2)ε log 2 (1/ε)|S ′ | ,
X∈E
and therefore,
X
X∈E
|p(X) − µ|2 ≥ (C/3)ε log 2 (1/ε)|S ′ | .
On the other hand, we have that
Z ∞
X
2
{X ∈ E : |p(X) − µ| > t}2tdt
|p(X) − µ| =
0
X∈E
≤
Z
C 1/4 log(1/ε)
0
O(tε|S ′ |)dt +
Z
C 1/4
{X ∈ E : |p(X) − µ| > t}2tdt
C 1/4 log(1/ε)
Z (d log(N/τ ))
≤ O(C 1/2 ε log2 (1/ε)|S ′ |) + |S ′ |
Therefore, we have that
Z (d log(N/τ ))
(d log(N/τ ))
Pr (|p(X) − µ| > t)2tdt .
′
C 1/4 log(1/ε) X∈u S
Pr (|p(X) − µ| > t)2tdt ≥ (C/4)ε log 2 (1/ε) .
′
log(1/ε) X∈u S
(41)
Assume for sake of contradiction that
Pr (|p(X) − µ| ≥ T ) < 12 exp(−(T − 4)/3) + 3ε/(d log(N/τ ))2 ,
X∈u S ′
for all T > 1. Plugging this into Equation (41), we find that
Z
(d log(N/τ ))
24 exp(−t/3)dt +
C 1/4
log(1/ε)−4
Z
(d log(N/τ ))
C 1/4
log(1/ε)
6tε/(d log(N/τ ))2 dt ≥ (C/4)ε log 2 (1/ε) ,
which yields a contradiction.
Finally, we need to verify that if our algorithm returns output in Step 12, that it is correct.
Claim 8.25. If the algorithm returns during Step 12, then ∆(S, S ′′ ) ≤ ∆(S, S ′ ) − ε/(d log(N/τ ))2 .
Proof. We note that it is sufficient to show that |E \ S ′′ | > |(S \ L) \ S ′′ |. In particular, it suffices to show
that
|{X ∈ E : |p(X) − µ| > T }| > |{X ∈ S \ L : |p(X) − µ| > T }|.
For this, it suffices to show that
|{X ∈ S ′ : |p(X) − µ| > T }| > 2|{X ∈ S \ L : |p(X) − µ| > T }|,
or that
|{X ∈ S ′ : |p(X) − µ| > T }| > 2|{X ∈ S : |p(X) − µ| > T }|.
94
By assumption, we have that
|{X ∈ S ′ : |p(X) − µ| > T }| > (12 exp(−(T − 4)/3) + 3ε/(d log(N/τ ))2 )|S ′ | .
On the other hand, we have that
|{X ∈ S : |p(X) − µ| > T }| ≤ |S|(2 Pr(|p(G) − µ| > T ) + ε/(d log(N/τ ))2 )
≤ 3|S ′ | Pr (|p(X) − µ| > T ) + 3ε/(d log(N/τ ))2 )|S ′ | .
X∼G
On the other hand by the last claim, p(X) for X ∼ G is an even degree-2 polynomial of a Gaussian with
variance at most 3/2. Therefore, by Lemma 8.15
Pr (|p(X) − E[p(X)]| ≥ T ) ≤ 2e1/3−2T /3 Var[p(X)] ≤ exp((T − 1)/3)) .
X∼G
However, from equation (40), E[p(X)] − µ ≤ 3 and so
Pr (|p(X) − µ| > T ) ≤ 2 exp(−(T − 4)/3).
X∼G
This completes our proof.
Acknowledgements
J.L. would like to thank Michael Cohen and Sam Hopkins for some very helpful discussions.
References
[ABL14]
P. Awasthi, M. F. Balcan, and P. M. Long. The power of localization for efficiently learning
linear separators with noise. In STOC 2014, pages 449–458, 2014.
[ADLS15] J. Acharya, I. Diakonikolas, J. Li, and L. Schmidt. Sample-optimal density estimation in
nearly-linear time. CoRR, abs/1506.00671, 2015.
[AGHK13] A. Anandkumar, R. Ge, D. Hsu, and S. Kakade. A tensor spectral approach to learning mixed
membership community models. In COLT 2013, pages 867–881, 2013.
[AGM12]
S. Arora, R. Ge, and A. Moitra. Learning topic models - going beyond SVD. In FOCS 2012,
pages 1–10, 2012.
[AGMS15] S. Arora, R. Ge, A. Moitra, and S. Sachdeva. Provable ICA with unknown gaussian noise, and
implications for gaussian mixtures and autoencoders. Algorithmica, 72(1):215–236, 2015.
[AHK12]
A. Anandkumar, D. Hsu, and S. Kakade. A method of moments for mixture models and Hidden
Markov Models. Journal of Machine Learning Research - Proceedings Track, 23:33.1–33.34,
2012.
[AJOS14]
J. Acharya, A. Jafarpour, A. Orlitsky, and A. T. Suresh. Sorting with adversarial comparators
and application to density estimation. In Proceedings of the 2014 IEEE International Symposium on Information Theory (ISIT), pages 1682–1686, 2014.
[AK01]
S. Arora and R. Kannan. Learning mixtures of arbitrary Gaussians. In Proceedings of the 33rd
Symposium on Theory of Computing, pages 247–257, 2001.
95
[BBBB72] R.E. Barlow, D.J. Bartholomew, J.M. Bremner, and H.D. Brunk. Statistical Inference under
Order Restrictions. Wiley, New York, 1972.
[BD15]
S. Balmand and A. Dalalyan. Convex programming approach to robust estimation of a multivariate gaussian model, 2015.
[Ber06]
T. Bernholt. Robust estimators are hard to compute. Technical report, University of Dortmund,
Germany, 2006.
[Bru09]
S. C. Brubaker. Robust PCA and clustering in noisy mixtures. In SODA 2009, pages 1078–
1087, 2009.
[BS10]
M. Belkin and K. Sinha. Polynomial learning of distribution families. In FOCS, pages 103–
112, 2010.
[BV08]
S. C. Brubaker and S. Vempala. Isotropic PCA and Affine-Invariant Clustering. In Proc. 49th
IEEE Symposium on Foundations of Computer Science, pages 551–560, 2008.
[CDSS13]
S. Chan, I. Diakonikolas, R. Servedio, and X. Sun. Learning mixtures of structured distributions over discrete domains. In SODA, pages 1380–1394, 2013.
[CDSS14a] S. Chan, I. Diakonikolas, R. Servedio, and X. Sun. Efficient density estimation via piecewise
polynomial approximation. In STOC, pages 604–613, 2014.
[CDSS14b] S. Chan, I. Diakonikolas, R. Servedio, and X. Sun. Near-optimal density estimation in nearlinear time using variable-width histograms. In NIPS, pages 1844–1852, 2014.
[CGG02]
M. Cryan, L. Goldberg, and P. Goldberg. Evolutionary trees can be learned in polynomial time
in the two state general Markov model. SIAM Journal on Computing, 31(2):375–397, 2002.
[CLM+ 16] M. B. Cohen, Y. T. Lee, G. L. Miller, J. W. Pachocki, and A. Sidford. Geometric median in
nearly linear time. In Proceedings of STOC’16, 2016.
[CLMW11] E. J. Candès, X. Li, Y. Ma, and J. Wright. Robust principal component analysis? J. ACM,
58(3):11, 2011.
[CW01]
A. Carbery and J. Wright. Distributional and Lq norm inequalities for polynomials over convex
bodies in Rn . Mathematical Research Letters, 8(3):233–248, 2001.
[Das99]
S. Dasgupta. Learning mixtures of Gaussians. In Proceedings of the 40th Annual Symposium
on Foundations of Computer Science, pages 634–644, 1999.
[DDKT16] C. Daskalakis, A. De, G. Kamath, and C. Tzamos. A size-free CLT for poisson multinomials
and its applications. In Proceedings of the 48th Annual ACM Symposium on the Theory of
Computing, STOC ’16, New York, NY, USA, 2016. ACM.
[DDO+ 13] C. Daskalakis, I. Diakonikolas, R. O’Donnell, R.A. Servedio, and L. Tan. Learning Sums of
Independent Integer Random Variables. In FOCS, pages 217–226, 2013.
[DDS12a]
C. Daskalakis, I. Diakonikolas, and R.A. Servedio. Learning k-modal distributions via testing.
In SODA, pages 1371–1385, 2012.
[DDS12b]
C. Daskalakis, I. Diakonikolas, and R.A. Servedio. Learning Poisson Binomial Distributions.
In STOC, pages 709–728, 2012.
96
[DDS15a]
C. Daskalakis, I. Diakonikolas, and R. A. Servedio. Learning poisson binomial distributions.
Algorithmica, 72(1):316–357, 2015.
[DDS15b]
A. De, I. Diakonikolas, and R. Servedio. Learning from satisfying assignments. In Proceedings
of the 26th Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2015, pages 478–
497, 2015.
[DG85]
L. Devroye and L. Györfi. Nonparametric Density Estimation: The L1 View. John Wiley &
Sons, 1985.
[DK14]
C. Daskalakis and G. Kamath. Faster and sample near-optimal algorithms for proper learning
mixtures of Gaussians. In Proceedings of the 27th Annual Conference on Learning Theory,
COLT ’14, pages 1183–1213, 2014.
[DKS15]
I. Diakonikolas, D. M. Kane, and A. Stewart. Optimal learning via the fourier transform for
sums of independent integer random variables. CoRR, abs/1505.00662, 2015.
[DKS16]
I. Diakonikolas, D. M. Kane, and A. Stewart. The fourier transform of poisson multinomial
distributions and its algorithmic applications. In Proceedings of STOC’16, 2016.
[DKT15]
C. Daskalakis, G. Kamath, and C. Tzamos. On the structure, covering, and learning of Poisson
multinomial distributions. In Proceedings of the 56th Annual IEEE Symposium on Foundations
of Computer Science, FOCS ’15, Washington, DC, USA, 2015. IEEE Computer Society.
[DL96]
L. Devroye and G. Lugosi. A universally acceptable smoothing factor for kernel density estimates. Annals of Statistics, 24:2499–2512, 1996.
[DL97]
L. Devroye and G. Lugosi. Nonasymptotic universal smoothing factors, kernel complexity and
Yatracos classes. Annals of Statistics, 25:2626–2637, 1997.
[DL01]
L. Devroye and G. Lugosi. Combinatorial methods in density estimation. Springer Series in
Statistics, Springer, 2001.
[FM99]
Y. Freund and Y. Mansour. Estimating a mixture of two product distributions. In Proceedings
of the 12th Annual COLT, pages 183–192, 1999.
[FOS05]
J. Feldman, R. O’Donnell, and R. Servedio. Learning mixtures of product distributions over
discrete domains. In Proc. 46th Symposium on Foundations of Computer Science (FOCS),
pages 501–510, 2005.
[GLS88]
M. Grötschel, L. Lovász, and A. Schrijver. Geometric Algorithms and Combinatorial Optimization, volume 2. Springer, 1988.
[Hau92]
D. Haussler. Decision theoretic generalizations of the PAC model for neural net and other
learning applications. Information and Computation, 100:78–150, 1992.
[HK13]
D. Hsu and S. M. Kakade. Learning mixtures of spherical gaussians: moment methods and
spectral decompositions. In Innovations in Theoretical Computer Science, ITCS ’13, pages
11–20, 2013.
[HM13]
M. Hardt and A. Moitra. Algorithms and hardness for robust subspace recovery. In COLT
2013, pages 354–375, 2013.
97
[HP15]
M. Hardt and E. Price. Sharp bounds for learning a mixture of two Gaussians. In Proceedings
of the 47th Annual ACM Symposium on the Theory of Computing, STOC ’15, pages 753–760,
New York, NY, USA, 2015. ACM.
[HR09]
P. J. Huber and E. M. Ronchetti. Robust statistics. Wiley New York, 2009.
[HRRS86] F. R. Hampel, E. M. Ronchetti, P. J. Rousseeuw, and W. A. Stahel. Robust statistics. The
approach based on influence functions. Wiley New York, 1986.
[Hub67]
P. J. Huber. The behavior of maximum likelihood estimates under nonstandard conditions.
In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability,
Volume 1: Statistics, pages 221–233. University of California Press, 1967.
[JP78]
D. S. Johnson and F. P. Preparata. The densest hemisphere problem. Theoretical Computer
Science, 6:93–107, 1978.
[KL93]
M. J. Kearns and M. Li. Learning in the presence of malicious errors. SIAM Journal on
Computing, 22(4):807–837, 1993.
[KLS09]
A. Klivans, P. Long, and R. Servedio. Learning halfspaces with malicious noise. Proc. 17th
Internat. Colloq. on Algorithms, Languages and Programming (ICALP), 2009.
[KMR+ 94] M. Kearns, Y. Mansour, D. Ron, R. Rubinfeld, R. Schapire, and L. Sellie. On the learnability
of discrete distributions. In Proc. 26th STOC, pages 273–282, 1994.
[KMV10]
A. T. Kalai, A. Moitra, and G. Valiant. Efficiently learning mixtures of two Gaussians. In
STOC, pages 553–562, 2010.
[KSS94]
M. Kearns, R. Schapire, and L. Sellie. Toward Efficient Agnostic Learning. Machine Learning,
17(2/3):115–141, 1994.
[LM00]
B. Laurent and P. Massart. Adaptive estimation of a quadratic functional by model selection.
Annals of Statistics, pages 1302–1338, 2000.
[LMTZ12] G. Lerman, M. B. McCoy, J. A. Tropp, and T. Zhang. Robust computation of linear models, or
how to find a needle in a haystack. CoRR, abs/1202.4044, 2012.
[LS15]
J. Li and L. Schmidt. A nearly optimal and agnostic algorithm for properly learning a mixture
of k gaussians, for any constant k. CoRR, abs/1506.01367, 2015.
[LT15]
P. L. Loh and X. L. Tan. High-dimensional robust precision matrix estimation: Cellwise corruption under -contamination. Available at arXiv:1509.07229, 2015.
[MR05]
E. Mossel and S. Roch. Learning nonsingular phylogenies and Hidden Markov Models. In
Proceedings of the 37th Annual Symposium on Theory of Computing (STOC), 2005.
[MV10]
A. Moitra and G. Valiant. Settling the polynomial learnability of mixtures of Gaussians. In
FOCS, pages 93–102, 2010.
[Sco92]
D.W. Scott. Multivariate Density Estimation: Theory, Practice and Visualization. Wiley, New
York, 1992.
[Ser01]
R. Servedio. Smooth boosting and learning with malicious noise. In Proceedings of the Fourteenth Annual Conference on Computational Learning Theory, pages 473–489, 2001.
98
[Ser03]
R. Servedio. Smooth boosting and learning with malicious noise. JMLR, 4:633–648, 2003.
[Sil86]
B. W. Silverman. Density Estimation. Chapman and Hall, London, 1986.
[SOAJ14]
A. T. Suresh, A. Orlitsky, J. Acharya, and A. Jafarpour. Near-optimal-sample estimators for
spherical gaussian mixtures. In NIPS, pages 1395–1403, 2014.
[Tuk75]
J.W. Tukey. Mathematics and picturing of data. In Proceedings of ICM, volume 6, pages
523–531, 1975.
[Val85]
L. Valiant. Learning disjunctions of conjunctions. In Proc. 9th IJCAI, pages 560–566, 1985.
[Ver10]
R. Vershynin. Introduction to the non-asymptotic analysis of random matrices, 2010.
[VW02]
S. Vempala and G. Wang. A spectral algorithm for learning mixtures of distributions. In FOCS,
pages 113–122, 2002.
[Whi82]
H. White. Maximum likelihood estimation of misspecified models. Econometrica, 50(1):1–25,
1982.
[Yat85]
Y. G. Yatracos. Rates of convergence of minimum distance estimators and Kolmogorov’s
entropy. Annals of Statistics, 13:768–774, 1985.
[ZL14]
T. Zhang and G. Lerman. A novel m-estimator for robust pca. J. Mach. Learn. Res., 15(1):749–
808, January 2014.
A
Deferred Proofs from Section 2
Proof of Lemma 2.15: We include the simple proof for completeness. By Kraft’s inequality, for any pair
of distributions, we have that d2TV (P, Q) ≤ 2H 2 (P, Q), where H(P, Q) denotes the Hellinger distance
between P, Q. Since P, Q are product measures, we have that
d
d
Y
Y
p
√
2
( pi qi + (1 − pi )(1 − qi )) .
(1 − H (Pi , Qi )) =
1 − H (P, Q) =
2
i=1
i=1
√
√
√
The elementary inequality 2 ab = a + b − ( a − b)2 , a, b > 0, gives that
√
Therefore,
pi q i +
p
(1 − pi )(1 − qi ) ≥ 1 −
d2TV (P, Q)
(pi − qi )2
= 1 − zi .
(pi + qi )(2 − pi − qi )
d
d
X
Y
zi ,
(1 − zi )) ≤ 2
≤ 2 · (1 −
i=1
where the last inequality follows from the union bound.
99
i=1
B Deferred Proofs from Section 4
This section contains deferred proofs of several concentration inequalities.
Proof of Lemma 4.5: Recall that for any J ⊆ [m], we let wJ ∈ Rm be the vector which is given by wiJ =
for i ∈ J and wiJ = 0 otherwise. By convexity, it suffices to show that
"
N
X
Pr ∃J : |J| = (1 − 2ε)N, and
i=1
wiJ Yi YiT − (1 − 2ε)I
2
1
|J|
#
≥ δ1 ≤ τ .
For any fixed wJ we have
N
X
i=1
wiJ Yi YiT − I =
=
X
1
Yi YiT − I
(1 − 2ε)N
i∈J
1
(1 − 2ε)N
N
X
i=1
Yi YiT −
Therefore, by the triangle inequality, we have
N
X
i=1
wiI Yi YiT
− (1 − 2ε)I
2
1
1
I −
1 − 2ε
(1 − 2ε)N
N
X
1
1
Yi YiT −
≤
I
(1 − 2ε)N
1 − 2ε
i=1
2
X
i6∈J
Yi YiT −
1
1 − 2ε
− 1 I .
X
1
1
T
−1 I
+
Yi Yi −
(1 − 2ε)N
1 − 2ε
i6∈J
By Lemma 4.3, the first term is at most δ1 with probability 1 − τ by setting
d + log(1/τ )
N =Ω
.
δ12
We now wish to show that with probability 1 − τ , there does not exist a J so that the second term is
more than δ1 , for our choice of N . By multiplying both sides by γ = (1 − 2ε)/ε, this event is equivalent to
the event that there is a J so that
1 X
Yi YiT − I
2εN
i6∈J
> γ1 δ1 .
2
By Lemma 4.3, for any fixed τ ′ , this happens with probability at most τ ′ if we choose
N ≥c
d + log(1/τ ′ )
,
γδ12
(42)
N
for some universal constant c. We wish to union bound over all J of size (1 − 2ε)N . There are (1−2ε)N
−1
N
such events. Thus, we take τ ′ = (1−2ε)N
τ . We now need to show that for our choice of N and τ ′ , (42)
n
is satisfied. By algebraic manipulations, and the fact that log αn
= O(nH(α)) for n sufficiently large,
where H(x) is the binary entropy function, the RHS of (42) simplifies to
c
d + log 1/τ + O(N H(2ε))
d + log 1/τ ′
=c
.
2
γδ1
γδ12
100
.
2
Thus, (42) is satisfied if
N
By our choice of δ1 , 1 −
H(2ε)
γδ12
H(2ε)
1−
γδ12
≥c
d + log 1/τ
.
γδ12
≥ 1/2, and hence, (42) is satisfied for our choice of m.
Proof of Lemma 4.9: Let U = V ΛV T be the diagonalization of U . Then by the properties of trace, we have
N
1 X
tr(Xi XiT U ) − tr(U )
N
=
i=1
N
1 X
tr(Xi XiT V ΛV T ) − tr(Λ)
N
i=1
=
N
1 X
tr(XiT V ΛV T Xi ) − tr(Λ) .
N
i=1
Let Yi = (Yi1, , . . . , Yid ) = V T Xi . Since V is an orthogonal matrix, and because a spherical Gaussian is
rotationally invariant, we have Yi ∼ N (0, I). By computation, the lemma is equivalent to the statement that
d
N X
X
λj
i=1 j=1
N
(Yij2 − 1) ≤ δ .
Since Yi ∼ N (0, I), we have that the Yij2 − 1 are all independent random variables with kYij2 − 1kψ1 ≤ c
for some universal constant c. Thus, we may invoke Bernstein’s unequality (Prop 5.16), where our weight
vector a ∈ RN d is given by aij = λj /N , and thus satisfies kak∞ = kU k/N and
kak22
d
N X
X
λ2j
=
N2
i=1 j=1
N
X
1
1
=
=
N2
N
i=1
Pd
= kU k2F = 1. Thus, Bernstein’s inequality gives us that
2
d
N X
X
λj 2
δ
δ
Pr
(Y − 1) > δ ≤ 2 exp −K min 2
,
N ij
c /N ckU k/N
i=1 j=1
≤ 2 exp −Ω min δ2 N, δN
since by assumption
2
i=1 λj
≤τ ,
for our choice of N , where the second line follows because the spectral norm of any matrix with Frobenius
norm 1 is at most 1.
Proof of Theorem 4.15: Let v = A♭ ∈ Ssym . We will show that
T
hv, M vi = 2v T Σ⊗2 v + v T Σ♭ Σ♭ v .
2
Since M is a symmetric operator on Rd , its quadratic form uniquely identifies it and this suffices to prove
the claim.
101
P
Since A is symmetric, it has a eigenvalue expansion A = di=1 λi ui uTi , which immediately implies
P
that v = di=1 λi ui ⊗ ui . Let X ∼ N (0, Σ). We compute the quadratic form:
hv, M vi =
=
d
X
i,j=1
d
X
i,j=1
=
d
X
i,j=1
=
d
X
i,j=1
=
d
X
i,j=1
λi λj hui ⊗ ui , E[(X ⊗ X)(X ⊗ X)T ]uj ⊗ uj i
λi λj E hui ⊗ ui , (X ⊗ X)(X ⊗ X)T uj ⊗ uj i
λi λj E hui , Xi2 huj , Xi2
λi λj E hB T ui , Y i2 hB T uj , Y i2
λi λj hB T ui , B T ui ihB T uj , B T uj i + 2hB T ui , B T uj i2 ,
where the last line follows by invoking Isserlis’s theorem. We now manage both sums individually. We have
d
X
i,j=1
λi λj hB T ui , B T ui ihB T uj , B T uj i =
d
X
=
d
X
λi uTi Σui
i=1
i=1
!2
!
2
♭
λi (ui ⊗ ui ) Σ
T
T
= v T Σ♭ Σ♭ v ,
and
d
X
i,j=1
λi λj hB T ui , B T uj i2 =
=
X
i,j
λi λj h(B T ui )⊗2 , (B T uj )⊗2 i
d
X
i,j=1
=
d
X
i,j=1
λi λj h(B T ⊗ B T )ui ⊗ ui , (B T ⊗ B T )uj ⊗ uj i
λi λj (ui ⊗ ui )Σ⊗2 (uj ⊗ uj )
= v T Σ⊗2 v .
B.1
Proof of Theorem 4.17
We will proceed in a couple of steps. First, we prove concentration of the empirical fourth moment, without
worrying about removing points. We will use the following theorem:
102
Theorem B.1 (Theorem 5.44 of Vershynin). Let A1 , . . . , AN ∈ Rd√be independent random variables with
identical second moment M = E[Ai ATi ]. Let R be so that kAi k ≤ R almost surely for all i. Then, for all
t ≥ 0, we have that with probability at least 1 − n · exp(−ct2 ),
N
1 X
Ai ATi − Σ
N
i=1
2
r
R
1/2
2
where γ = t
≤ max kΣk2 γ, γ
.
d
With some massaging, this allows us to show:
Lemma B.2. Fix ε, τ > 0. Let M4 denote the canonical flattening of the fourth moment tensor of N (0, I).
We have that with probability 1 − τ ,
s
N
N
1
3
X
d log τ log τ ε
1
+ε .
Zi ZiT − M4 ≤ O
N
N
i=1
2
In particular, if
then
1
N
PN
T
i=1 Zi Zi
d3 log2
ε2
e
N =Ω
− M4
2
1
τ
!
,
≤ O(ε).
We briefly remark that with somewhat more involved (but standard) methods, we believe we can
strengthen this bound to
2
d3
d + log2 1/τ
e
+
.
N =Ω
ε2
ε
However, for simplicity of exposition we only present the weaker bound.
2
Proof. Let D be the distribution over Rd which is given by X ⊗2 where X ∼ N (0, I) if kXk22 ≤
O(d log(N/(τ ε))), and 0 otherwise. That is, a sample from D is drawn by sampling X ⊗2 , and zeroing
it out if its norm is too large. Let I ⊆ [N ] be the set of indices so that kXi k22 ≤ O(d log τNε ). For each i ∈ I,
we have that Xi ∼ D. Let M4′ = EX∼D [(X ⊗2 )(X ⊗2 )T ]. Since we are only removing large points, it is
evident that M4′ M4 , and kM4 k = O(d2 ). By Theorem B.1, we have
s
1
m
d
log
log
1 X
τ
τε
Zi ZiT − M4′ ≤ max O(d)γ, γ 2 where γ = O
,
|I|
|I|
i∈I
2
except with probability 1 − O(τ ). By basic Gaussian concentration, we know that the probability that we
throw away Zi from this process is at most O(τ /m). Hence with probability 1−O(τ ), we have that I = [N ].
Union bounding over these two events and simplifying yields that
s
N
1
N
3
X
d
log
log
1
τ
τε
Zi ZiT − M4′ ≤ O
.
N
N
i=1
2
103
It now suffices to show that kM4 − M4′ k2 ≤ O(ε). We have
kM4 −
M4′ k2
T
T
= E[(X ⊗ X)(X ⊗ X) ] − E (X ⊗ X)(X ⊗ X)
i
h
≤ E k(X ⊗ X)(X ⊗ X)T k2 1{kXk2 ≥Ω(d log N )}
2
τε
h
i
4
= E kXk2 1{kXk2 ≥Ω(d log N )}
2
τε
Z ∞
2
2
≤
t Pr[kXk2 ≥ t]dt
Ω(d log τNε )
≤ O(ε) ,
kXk22
N
≤ O d log
τε
2
by basic Gaussian concentration.
We have now shown that spectral concentration holds when we don’t allow any subselection of the
points. To complete the proof of the theorem, we now need to use concentration of low-degree polynomials
to show that this still holds when we allow ourselves to remove points.
Proof of Theorem 4.17. We will first prove a couple of key claims. First, we shall show that there cannot be
too many i so that p2 (Xi ) is large:
Claim B.3. Fix τ, ε > 0. Let N ≥ Ω log(1/τ )/ε2 . There is a universal constant (independent of ε, τ ) c′
so that for all c > c′ , we have that with probability 1 − O(τ ),
1
{i : p2 (Xi ) ≥ c log2 (1/ε)} ≤ 2ε .
N
Proof. Let Ui = 1{p2 (Xi )≥c log2 (1/ε)} . Then, E[Ui ] = PrX∼N (0,I) [p2 (X) ≥ c log2 (1/ε)] ≤ O(ε). The
P
result then follows from a Hoeffding bound on N1 N
i=1 Ui .
We now show that if we remove all points so that p2 (Xi ) is too large, then this cannot change the
expectation by too much:
Claim B.4. Let c > 0 be sufficiently large. Fix τ, ε > 0, and let N ≥ Ω log(1/τ )/ε2 . Let Ai = p2 (Xi ) if
p2 (Xi ) ≤ c log2 (1/ε), and 0 otherwise. Let α = EX∼N (0,I) [p2 (X)]. Then
N
1 X
Ai − α ≤ O(c · ε log2 1/ε) .
N
i=1
Proof. Let X ∼ N (0, I), and let A = p2 (X) if p2 (X) ≤ c log2 1/ε, and 0 otherwise. Let A be distributed
as D. Then Ai ∼ D for all i. In particular, each Ai is a bounded random variable. Hence, by Hoeffding’s
inequality, by our choice of N , we have
N
1 X
Ai − α′ ≤ O(ε) ,
N
i=1
104
where α′ = EA∼D [p2 (A)]. It thus suffices to show that |α′ − α| < O(c · ε log2 1/ε). We have
i
h
|α′ − α| =
E
p2 (X)1{p2 (X)≥c log2 1/ε}
X∼N (0,I)
Z ∞
= √
t2 Pr[p(X) ≥ t]dt
c log 1/ε
Z ∞
≤ √
t2 exp(−Ω(t))dt
c log 1/ε
≤ O(c · ε log2 1/ε) .
Let us now prove the theorem. We first condition on the event that for all homogenous degree 2 polynomials p with kpk2 = 1, we have
N
1 X 2
(43)
p (Xi ) − 1 ≤ O(ε) .
N
i=1
We observe that every v = M ♭ ∈ S corresponds to a homogenous degree two polynomial p(X) = X T M X,
and moreover, kpk2 = 2kM kF . Then Lemma 7 asserts that by our choice of N , (43) holds with probability
1 − O(τ ). Moreover, this implies that to show the theorem, it suffices to show that for all homogenous
degree two polynomials p with kpk2 = 1, we have
N
X
i=1
wi p2 (Xi ) − 1 ≤ O(ε log2 (1/ε)) ,
for all w ∈ Sm,ε .
Moreover, let C be a (1/3)-net for the set of all homogenous degree 2 polynomials with kpk2 = 1. It is
2
well-known that we can take |C| = 2O(d ) . Let us also condition on the event that for all p ∈ C, we have
1
{i : p2 (Xi ) ≥ c log2 (1/ε)} ≤ 2ε, and
N
m
1 X
Ai − 1 ≤ O(ε log2 1/ε) ,
N
(44)
(45)
i=1
where Ai is defined as in Claim B.4 depends on p. By claims B.3 and B.4, by our choice of N , for any fixed
p, these hold with probability 1 − O(τ ′ ), where τ ′ = O(τ /|C|), so that by a union bound this event occurs
with probability 1 − O(τ ).
Fix any p ∈ C. Let J ∗ be the set of 2εN indices with largest (p2 (Xi ) − 1), and let J1∗ = {i : p2 (Xi ) ≥
c log2 (1/ε)}. Observe that J1∗ ⊆ J ∗ , by (44). Moreover, by (45), we have
1 X 2
(p (Xi ) − 1) ≤ O(ε log2 (1/ε)) .
N
∗
(46)
i6∈J1
Fix any I ⊆ [m] so that |I| = (1 − 2ε)N . Partition [N ] \ I into two sets J + ∪ J − where J + = {i 6∈ I :
105
p2 (Xi ) ≥ 1}, and J − = {i 6∈ I : p2 (Xi ) < 1}. We then have
1 X
1 X 2
1 X 2
(p (Xi ) − 1) ≤
(p2 (Xi ) − 1) −
(p (Xi ) − 1)
|I|
|I|
|I|
+
−
i∈I
i∈I∪J
≤
i∈J
N
1 X 2
1 X 2
(p (Xi ) − 1)
p (Xi ) − 1 + 2
|I|
|I|
−
i=1
i∈J
|J − |
≤ O(ε) +
|I|
≤ O(ε) .
(47)
where the penultimate line follows from (43) and since p2 (Xi ) ≥ 0 and hence |p2 (Xi ) − 1| ≤ 1 for all
i ∈ J − . Moreover, we have that
X
1
1 X 2
(p (Xi ) − 1) ≥
(p2 (Xi ) − 1)
|I|
(1 − 2ε)N
∗
i6∈J1
i∈I
≥ (1 − O(ε))
2
1 X 2
(p (Xi ) − 1)
N
∗
i6∈J1
≥ −O(ε log 1/ε) ,
2
1 P
2
by (46) and hence this along with (47) implies that |I|
i∈I (p (Xi ) − 1) ≤ O(ε log 1/ε). Since this
holds for all I with |I| = (1 − 2ε)N , by convexity this implies that for all w ∈ SN,ε , we have
X
i∈I
wi (p2 (Xi ) − 1) ≤ O(ε log2 1/ε) .
Finally, by the usual net argument, this implies that for all degree 2 p with kpk2 = 1, we have
X
i∈I
wi (p2 (Xi ) − 1) ≤ O(ε log2 1/ε) ,
which by the logic above, suffices to prove the theorem.
C
Deferred Proofs from Section 5
Proof of Theorem 5.8: The first two properties follow directly from (25). We now show the third property.
Suppose this does not happen, that is, there are j, j ′ so that ℓ = ℓ(j) = ℓ(j ′ ) so that kµj − µj ′ k22 ≥
Ω(dk log k/ε). That means that by (25) there is some sequence of clusters S1 , . . . , St so that Si ∩ Si+1 6= ∅
for each i, |Si | ≥ 4εN for each i, and moreover, there is a Xi ∈ S1 so that kXi − µ1 k22 ≤ O(d log k/ε)
and an Xi′ ∈ St so that kXi′ − µ2 k22 ≤ O(d log k/ε). But by (25), we know that each Si contains an
point Xi′′ so that kXi′′ − µri k22 ≤ O(d log k/ε). In particular, by the triangle inequality, this means that if
kµri − µri+1 k22 ≤ O(d log k/ε) for all i = 1, . . . , t − 1, and we can set µr1 = µj and µrt = µj ′ .
Construct an auxiliary graph on k vertices, where we put an edge between nodes ri and ri+1 . By the
above, there must be a path from j to j ′ in this graph. Since this graph has k nodes, there must be a path
of length at most k from j to j ′ ; moreover, by the above, we know that this implies that kµj − µj ′ k22 ≤
O(kd log k/ε).
106
Finally, the fourth property follows from the same argument as the proof of the third.
P
T
Proof of Lemma 5.9: Let C = N
i=1 wi (Xi − µ)(Xi − µ) − I. Let v be the top eigenvector of
N
X
i=1
wi (Xi − µ)(Xi − µ)T − I −
X
j∈[k]
αj (µj − µ)(µj − µ)T
Observe that by (26), we have
N
X
i=1
wi (Xi − µ)(Xi − µ)T
X
i∈G
wi (Xi − µ)(Xi − µ)T
wg (I + Q) − f (k, γ, δ1 )I
(1 − ε)(I + Q) − f (k, γ, δ1 )I ,
and so in particular
N
X
i=1
wi (Xi − µ)(Xi − µ)T − (I + Q) −ε(I + Q) − f (k, γ, δ1 )I .
Therefore, for any unit vector u ∈ Rd , we must have
!
N
X
c
wi (Xi − µ)(Xi − µ)T − (I + Q) u ≥ −εuT (I + Q)uT − f (k, γ, δ1 ) ≥ − h(k, γ, δ) .
uT
2
i=1
In particular, since v T
P
N
T − (I + Q) v ≥ ckh(k, γ, δ), we must have
w
(X
−
µ)(X
−
µ)
i
i
i=1 i
v
T
N
X
i=1
!
T
wi (Xi − µ)(Xi − µ) − (I + Q) v > 0,
and hence
v
T
N
X
i=1
!
T
wi (Xi − µ)(Xi − µ) − (I + Q) v ≥ ckh(k, γ, δ) .
Let U = [v, u1 , . . . , ud−1 ] be an d × k matrix with orthonormal columns, where the columns span the
set of vectors {(µj − µ) : j ∈ [k]} ∪ {v}. We note the rank of this set is at most k due to the definition of
µ.
Using the dual characterization of the Schatten top-k norm, we have that
kCkTk = max Tr(X T CX) ≥ Tr(U T CU ).
X∈Rd×k
107
Observe that since span(Q) ⊆ span(U ), we have
T
kCkTk ≥ Tr U CU = Tr U
N
X
T
i=1
! !
T
wi (Xi − µ)(Xi − µ) − (I + Q) U
X
= Tr U T (C − Q)U +
γj
+ kQkTk
j∈[k]
k−1
X
= v T (C − Q)v +
i=1
uTi (C − Q)ui +
X
γj
j∈[k]
X
c
≥ ckh(k, γ, δ) − (k − 1) h(k, γ, δ) +
γj
2
j∈[k]
X
c
≥ kh(k, γ, δ) +
γj ,
2
j∈[k]
as claimed.
Proof of Lemma 5.10: By Fact 4.2 and (27) we have k
inequality, we have
X
i∈E
wi (Xi − µ)
2
X
≥ k∆k2 −
i∈G
P
wi
i=G wg Xi
− µk ≤ k1/2 δ2 . Now, by the triangle
wi (Xi − µ) − wg µ
2
≥ Ω(k∆k2 ).
Using the fact that variance is nonnegative we have
!
!T
X wi
X wi
,
(Xi − µ)
(Xi − µ)
wb
wb
X wi
(Xi − µ)(Xi − µ)T
wb
i∈E
i∈E
i∈E
and therefore
X
i∈E
T
wi (Xi − µ)(Xi − µ)
2
≥Ω
k∆k22
wb
≥Ω
k∆k22
ε
.
On the other hand,
X
i∈G
wi (Xi − µ)(Xi − µ)T − I
2
≤
X
i∈G
wi (Xi − µ)(Xi − µ)T − wg I
2
+ wb ≤ f (k, γ, δ1 ) + wb .
where in the last inequality we have used Fact 4.2 and (26). Hence altogether this implies that
N
X
i=1
T
wi (Xi − µ)(Xi − µ) − I
2
≥Ω
k∆k22
ε
as claimed.
108
− wb − f (k, γ, δ1 ) ≥ Ω
k∆k22
ε
,
C.1 Proof of Theorem 5.12
Once more, let ∆ = µ − µ̂ and expand the formula for M :
N
X
i=1
wi Yi YiT − I =
=
=
N
X
i=1
N
X
i=1
N
X
i=1
wi (Xi − µ + ∆)(Xi − µ + ∆)T − I
T
wi (Xi − µ)(Xi − µ) − I +
N
X
i=1
T
wi (Xi − µ)∆ + ∆
N
X
i=1
wi (Xi − µ)T + ∆∆T
wi (Xi − µ)(Xi − µ)T − I − ∆∆T .
We start by proving completeness.
Claim C.1. Suppose that w = w∗ . Then kM kTk ≤
P
i∈[k] γj
+
ckh(k,γ,δ1 )
.
2
Proof. w∗ are the weights that are uniform on the uncorrupted points. Because E ≤ 2εN
p , we have that
∗
∗
w ∈ SN,ε . Using (26), this implies that w ∈ Cf (k,γ,δ1 ) . By Corollary 5.11, k∆k2 ≤ O(ε log 1/ε).
m
X
i=1
≤
wi∗ (Xi − µ)(Xi − µ)T − I − ∆∆T
N
X
i=1
wi∗ (Xi − µ)(Xi − µ)T − I −
≤ kf (k, γ, δ1 ) +
<
X
γj +
j∈[k]
X
X
j∈[k]
Tk
αj (µj − µ)(µj − µ)T
+
Tk
X
j∈[k]
αi (µj − µ)(µj − µ)T
2
γj + O(ε log 1/ε)
j∈[k]
ckh(k, γ, δ)
.
2
Claim C.2. Suppose that w 6∈ Cckh(k,γ,δ). Then kM kTk >
Proof. We split into two cases. In the first case, k∆k22 ≤
N
X
i=1
wi (Xi − µ)(Xi − µ)T − I
Tk
P
i∈[k] γj
ckh(k,γ,δ)
.
10
≥
X
+
ckh(k,γ,δ1 )
.
2
By Lemma 5.9, we have that
γj +
j∈[k]
3ckh(k, γ, δ)
.
4
By the triangle inequality,
kM kTk ≥
X
j∈[k]
γj +
X
3ckh(k, γ, δ)
ckh(k, γ, δ)
− k∆k22 ≥
γj +
,
4
2
i∈[k]
as desired.
109
+ k∆∆T k2
Tk
In the other case, k∆k22 ≥
M as follows:
M=
N
X
i=1
=
ckh(k,γ,δ)
.
10
Recall that Q =
P
j∈[k] αj (µj
− µ)(µj − µ)T from (22). Write
wi (Xi − µ)(Xi − µ)T − I − ∆∆T
X
wi (Xi − µ)(Xi − µ)T − wg I − wg Q
i∈G
!
+ wg Q +
X
i∈E
wi (Xi − µ)(Xi − µ)T − wb I − ∆∆T
Now taking the Schatten top-k norm of M , we have
!
X
X
wi (Xi − µ)(Xi − µ)T − wg I − wg Q + wg Q +
wi (Xi − µ)(Xi − µ)T − wb I − ∆∆T
i∈G
i∈E
≥ wg Q +
≥ wg Q +
≥
≥
≥
X
j∈[k]
X
X
i∈E
X
i∈E
j∈[k]
wi (Xi − µ)(Xi − µ)T
γj − 4εkγ +
γj + Ω
j∈[k]
X
wi (Xi − µ)(Xi − µ)T
γj +
k∆k22
ε
X
i∈E
Tk
Tk
X
−
i∈G
wi (Xi − µ)(Xi − µ)T − wg I − wg Q
Tk
Tk
− kwb Ik2 − ∆∆T
− kf (k, γ, δ1 ) − 4ε − k∆k22
wi (Xi − µ)(Xi − µ)T
Tk
− kf (k, γ, δ1 ) − 4ε − k∆k22
(48)
ckh(k, γ, δ)
.
2
The first inequality is the triangle inequality, the second is by (26) and Fact 4.2, the third is because the
summed matrices are positive semidefinite, the fourth follows from Lemma 5.10, and the fifth holds for all
c sufficiently large.
By construction, we have that ℓ(w) ≥ 0. It remains to show that ℓ(w∗ ) < 0.
1 X
(Xi − µ̂)(Xi − µ̂)T − I
|G|
i∈G
Tk
1 X
=
(Xi − µ + ∆)(Xi − µ + ∆)T − I
|G|
i∈G
≤
+
Tk
X
1 X
(Xi − µ)(Xi − µ)T − I −
αj (µj − µ)(µj − µ)T
|G|
i∈G
X
j∈[k]
j∈[k]
αj (µj − µ)(µj − m)T
≤ kf (k, γ, δ1 ) +
X
j∈[k]
+ 2k∆k2
i∈G
Tk
γj + 2k
1/2
1 X
(Xi − µ)
|G|
δ2 k∆k2 +
110
k∆k22
Tk
Tk
+ k∆k22
2
Therefore,
ℓ(w∗ ) ≤ kf (k, γ, δ) +
If k∆k22 ≤
ckh(k,γ,δ)
,
10
∗
X
j∈[k]
γj + 2k1/2 δk∆k2 + k∆k22 − Λ.
then
ℓ(w ) ≤
X
γj + kh(k, γ, δ) +
2kδ
j∈[k]
We wish to show that
2kδ
p
ch(k, γ, δ) ckh(k, γ, δ)
√
− Λ.
+
10
10
p
ch(k, γ, δ)
ckh(k, γ, δ)
√
≤
,
10
10
or equivalently, that
p
ch(k, γ, δ)
√
.
2 10
√
p
But this is true for c sufficiently large, as h(k, γ, δ) ≥ δ. Therefore,
δ≤
ℓ(w∗ ) ≤
X
γj +
j∈[k]
where the second inequality holds since Λ >
P
On the other hand, consider when k∆k22 ≥
Λ≥
(c + 5)kh(k, γ, δ)
− Λ ≤ 0,
5
ckh(k,γ,δ)
.
j∈[k] γj +
2
ckh(k,γ,δ)
. By (48), we
10
X
γj + Ω
j∈[k]
k∆k22
ε
know that
.
Then we know
∗
ℓ(w ) ≤ kf (k, γ, δ) + 2k
1/2
δk∆k2 +
The first and third terms are immediately dominated by Ω
k
1/2
δk∆k2 = O
k∆k22
k∆k22
ε
k∆k22
ε
−Ω
k∆k22
ε
.
, it remains to show that
.
Or equivalently, k1/2 δε = O (k∆k2 ) . This follows since
√
p
k∆k2 ≥ O( h(k, γ, δ)) ≥ O( kδ2 ) = O(k1/2 δε)
Therefore in this case as well, ℓ(w∗ ) < 0, as desired.
D
Deferred Proofs from Section 6
Proof of Lemma 6.22: By Lemma 6.6 applied with ε′ := ε3/2 /10d in place of ε, since we have Ω(d4 log(1/τ )/ε′2 )
samples from P, with probability at least 1 − τ, the set S is such that for all affine functions L, it holds
| PrX∈u S (L(X) ≥ 0) − PrX∼P (L(X) ≥ 0)| ≤ ε′ /d. We henceforth condition on this event.
111
Let CT be the event that all coordinates in T take their most common value. For a single coordinate i,
the probability that it does not take its most common value, min{pi , 1 − pi }, satisfies
min{pi , 1 − pi } = pi (1 − pi )/ max{pi , 1 − pi } ≤ 2pi (1 − pi ).
Thus, by a union bound, we have that PrP (CT ) ≥ 3/5. Let #T (x) be the number of coordinates of x
in T which do not have their most common value, and observe that #T (x) is an affine function of x.
Noting that for x ∈ {0, 1}d , we have that 1 − #T (x) > 0 if and only if CT holds for x, it follows that
| PrS (CT ) − PrP [CT ]| ≤ ε′ /d. Hence, we have that PrS (CT ) ≥ 1/2.
For any affine function L(x), let
LT (x) = L(x) − #T (x) · max L(y).
y∈{0,1}d
Note that for x ∈ {0, 1}d , we have that LT (x) > 0 if and only if L(x) > 0 and CT holds for x. Therefore,
we can write
PrX∈u S (LT (X) > 0) PrX∼P (LT (X) > 0)
−
PrX∈u S (CT )
PrX∼P (CT )
|PrX∈u S (LT (X) > 0) PrX∼P (CT ) − PrX∼P (LT (X) > 0) PrX∈u S (CT )|
=
PrX∈u S (CT ) PrX∼P (CT )
≤ (10/3) ·
Pr (LT (X) > 0) · Pr (CT ) − Pr (CT )
X∈u S
X∼P
X∈u S
− Pr (CT ) Pr (LT (X) > 0) − Pr (LT (X) > 0)
Pr (L(X) > 0) − Pr (L(X) > 0) =
X∈u S
X∼P
X∈u S
X∼P
′
X∈u S
3/2
≤ (10/3) · 2ε /d ≤ ε
2
/d .
This completes the proof of Lemma 6.22.
E Deferred Proofs from Section 7
Proof of Lemma 7.6: Let SP ⊆ S be the set of samples drawn from P and SQ ⊆ S be the set of samples
drawn from Q. Firstly, we note that by a Chernoff bound, ||SP |/|S| − α| ≤ O(ε/d2 ) with probability
1 − τ /3. Assuming this holds, it follows that |SP | ≥ (α/2)|S| ≥ (ε1/6 /2)|S| = Ω(d4 log(1/τ )/ε2 ).
Similarly, |SQ | ≥ (1 − α)|S|/2 ≥ Ω(d4 log(1/τ )/ε2 ). By Lemma 6.6 applied with ε′ := (c2 /4) · ε in place
of ε, since we have Ω((d4 + d2 log(τ ))/ε′2 ) samples, with probability 1 − τ /3, the set SP is ε′ -good for P ,
i.e., it satisfies that for all affine functions L, | PrX∈u SP (L(X) > 0) − PrX∼P (L(X) > 0)| ≤ ε′ /d. We
show that assuming S is ε′ -good, it is (ε, i) good for each 1 ≤ i ≤ d.
Note that PrX∼P [Xi = 1] = pi ≥ c and PrX∼P [Xi = 0] = 1 − pi ≥ c. Since | PrX∼P [Xi =
1] − PrX∈u SP [Xi = 1]| ≤ c2 ε/(4d), it follows that PrX∈u S [Xi = 1] ≥ c/2. For any affine function L,
define L(0) (x) := L(x) − xi (maxy |L(y)|) and L(1) (x) := L(x) − (1 − xi )(maxy |L(y)|). Then, we have
112
the following:
Pr (L(X) > 0) − Pr (L(X) > 0 | Xi = 1) =
X∼P
1
X∈u SP
=
PrX∈u S 1 L1 (X) > 0
P
PrX∈u S 1 (Xi > 0)
P
≤ (2/c2 )
2
Pr
1
X∈u SP
≤ (2/c ) pi
PrX∼P L1 (X) > 0
−
pi
L1 (X) > 0 pi − Pr
X∼P
Pr
1
X∈u SP
≤ 2/c2 · 2ε′ /d ≤ ε/d .
L (X) > 0 − Pr
L1 (X) > 0 Pr (Xi > 0)
1
X∼P
X∈u S1
L (X) > 0
1
!
− Pr
X∼P
!
L (X) > 0
1
Pr (Xi > 0) − pi
1
X∈u SP
Similarly, we obtain that
Pr (L(X) > 0) − Pr (L(X) > 0) ≤ ε/d .
X∈u S1
X∼Π
So, we have that SP is (ε, i) good for P for all 1 ≤ i ≤ d with probability 1 − τ /3 . Similarly, SQ is (ε, i)
good for Q for all 1 ≤ i ≤ d with probability 1 − τ /3. Thus, we have that ||SP |/|S| − α| ≤ ε/d2 , SP
is (ε, i) good for P and SQ is (ε, i) good for Q for all 1 ≤ i ≤ d with probability 1 − τ . That is, S is
(ε, i)-good for Π for all 1 ≤ i ≤ d with probability at least 1 − τ .
Proof of Lemma 7.19: Let SP ⊆ S be the set of samples drawn from P and SQ ⊆ S be the set of samples
drawn from Q. Firstly, we note that by a Chernoff bound, ||SP |/|S| − α| ≤ O(ε/d2 ) with probability at
least 1 − τ /3. Assuming this holds, |SP | ≥ (α/2)|S| ≥ δ|S| = Ω(d4 log(1/τ )/ε2 ). Similarly, |SQ | ≥
(1 − α)|S|/2 ≥ Ω(d4 log(1/τ )/ε2 ).
By Lemma 6.6 applied with ε′ := ε/6, since we have Ω(d4 log(1/τ )/ε′2 ) samples, with probability at
least 1 − τ /3, the set SP is ε-good for P . Similarly, with probability at least 1 − τ /3, the set SQ is ε-good
for Q. Thus, with probability 1 − τ , we have that
Q respectively.
F
|SP |
|S|
− α ≤ ε and that SP and SQ are ε-good for P and
Deferred Proofs from Section 8
Proof of Lemma 8.3: By Fact 4.8, the first property holds with probability at least 1 − τ /3. After translating
by µG , we note that (iii) and (iv) follow immediately from Lemmas 4.3 and 4.4 with probability 1 − τ /3. It
remans to show that, conditioned on (i), (ii) holds with probability at least
p 1 − τ /3.
To simplify some expressions, let δ := ε/(d log(d/ετ )) and R = C d log(|S|/τ ).
We need to show that, if f is the indicator function of any halfspace, with probability at least 1 − τ /4 it
holds
E [f (X)] − E [f (X)] ≤ δ.
X∼G
x∈u S
In fact, we will show that with probability at least 1 − τ /8 we have
E [f (X)] ≤ E [f (X)] + δ ,
x∈u S
X∼G
as the other direction will hold with a similar probability.
In order to do this, we will demonstrate a finite collection C of indicators of halfspaces so that for any
halfspace indicator f , there is a g ∈ C so that
113
!!
• g ≥ f on B(µG , R), and
• EX∼G [g(X)] ≤ EX∼G [f (X)] + δ/2.
We note that given such a C, by an application of the Chernoff and union bounds, we have that so long
as |S| ≫ log(|C|)/δ2 it holds
E [g(X)] − E [g(X)] < δ/2
X∼G
x∈u S
for all g ∈ C. This would imply that for all f in question, if g ∈ C is as guaranteed above, we have
E [f (X)] ≤ E [g(X)]
x∈u S
x∈u S
≤ E [g(X)] + δ/2
X∼G
≤ E [f (X)] + δ ,
X∼G
where the first line above holds since S is supported on B(µG , R).
It remains to construct C. We claim that it suffices to take C be the the set of indicators of halfspaces of
the form v · (x − µG ) ≥ T , where v is taken from a δ/(CR)-cover of the unit sphere and T is a multiple of
δ/C with |T | ≤ 2R. By standard results, v can be taken from a set of size O(δ/R)d and T from a set of size
O(R/δ), thus |C| = O(δ/R)d+1 . Hence, condition (ii) will be satisfied with appropriately high probability
so long as |S| ≫ d log(δ/R)/ε2 , which is sufficient for our purposes.
What remains to show is that the C described above is sufficient. Consider f , the indicator of the
halfspace w · (x − µG ) ≥ T for some unit vector w. If |T | ≥ R, f is constant on B(µG , R), and replacing
T by sgn(T )R does not change f on B(µG , R) and changes the expectation of f by a negligible amount.
Therefore, we can assume that |T | ≤ R. Let v be a unit vector with kv − wk2 ≤ (δ/(CR)) and T ′ ∈
[T − 2δ/C, T − δ/C] be a multiple of δ/C. We claim that the g ∈ C given by the halfspace v ·(x− µG ) ≥ T ′
suffices.
To verify the first property, note that if w · (x − µG ) ≥ T , then
v · (x − µG ) = w · (x − µG ) + (v − w) · (x − µG ) ≥ T − R(δ/(CR)) ≥ T ′ .
For the second property, we note that the expectation of f is an error function of T , while the expectation of g
is an error function of T ′ . Since the error function is Lipschitz, the difference is O(T −T ′ ) = O(δ/C) < δ/2
for C sufficiently large. This completes the proof.
Proof of Lemma
terms. Thus, we may write p(x) =
P
P 8.15: Note that an even polynomial has no degree-1
1
′
2+
′)
=
(P
p
x
p
x
x
+
p
.
Taking
(P
)
=
p
and
(P
o
2 i,i
i,i
2 )j,i = 2 pi,j , for i > j, gives that
2 i,j
i i,i i
i>j i,j i j
p(x) = xT P2′ x + p0 . Taking P2 = Σ1/2 P2′ Σ1/2 , we have p(x) = (Σ−1/2 x)T P2 (Σ−1/2 x) + p0 , for a d × d
symmetric matrix P2 and p0 ∈ R.
Let P2 = U T ΛU , where U is orthogonal and Λ is diagonal be an eigen-decomposition of the symmetric
−1/2 x)T P (U Σ−1/2 x). Let X ∼ G and Y = U Σ−1/2 X. Then, Y ∼ N (0, I)
matrix P2 . Then,
2
P p(x)2 = (U Σ
and p(X) = i λi Yi +p0 for independent Gaussians Yi . Thus, p(X) follows a generalized χ2 -distribution.
Thus, we have
#
"
X
X
2
λi = p0 + tr(P2 ) ,
E[p(X)] = E
λi Y i + p 0 = p 0 +
i
i
and
Var[p(X)] = Var
"
X
λi Yi2
#
+ p0 =
i
114
X
i
λ2i = kPF k2 .
P
Lemma F.1 (cf. Lemma 1 from [LM00]). Let Z = i ai Yi2 , where Yi are independent random variables
distributed as N (0, 1). Let a be the vector with coordinates ai . Then,
√
Pr(Z ≥ 2kak2 x + 2kak∞ x) ≤ exp(−x) .
We thus have:
sX
X
λi (Yi2 − 1) > 2 (
Pr
λ2i )t + 2(max λi )t ≤ e−t .
i
Noting that tr(P2 ) =
P
i λi ,
i
P
i
λ2i = kP2 kF and maxi λi = kP2 k2 ≤ kP2 k, for µp = E[p(X)] we have:
√
Pr(p(X) − µp > 2kP2 kF ( t + t)) ≤ e−t .
i
√
√
Noting that 2 a = 1 + a − (1 − a)2 ≤ 1 + a for a > 0, we have
Pr(p(X) − µp > kP2 kF (3t + 1)) ≤ e−t .
Applying this for −p(x) instead of p(x) and putting these together, we get
Pr(|p(X) − µp | > kP2 kF (3t + 1)) ≤ 2e−t .
Substituting t = T /3kP2 kF − 1/3, and 2kP2 k2F = VarX∼G (p(X)) gives:
Pr(|p(X) − E [p(X)]| ≥ T ) ≤ 2e1/3−2T /3 VarX∼G [p(X)] .
X∼G
The final property is a consequence of the following anti-concentration inequality:
Theorem F.2 ([CW01]). Let p : Rd → R be a degree-d polynomial. Then, for X ∼ N (0, I), we have
p
Pr(|p(X)| ≤ ε E[p(X)2 ] ≤ O(dε1/d ) .
This completes the proof.
Proof of Lemma 8.16: Note that we can reduce this to the case where Σ = I. We can consider the multiset T
of y with y = Σ−1/2 x for the corresponding x ∈ S. Then, T is a set of independent samples from N (0, I)
and note that if it is (ε, τ )good for N (0, I), then S is (ε, τ )-good for G. Thus, we assume that Σ = I.
By Fact 4.8, the first property holds with probability at least 1 − τ /4.
Let MS be the empirical second moment matrix of S, i.e., the d×d matrix with (MS )i,j = EX∈u S [Xi Xj ].
Then, by Lemma 4.3, with probability at least 1 − τ /4, kMS − Ik2 ≤ ε/4d3/2 .
Let TS be d2 × d2 matrix that is given by the canonical flattening of the empirical fourth moment tensor
of S. Then, by B.2, we have that with probability at least 1 − τ /4, kTS − M4 k2 ≤ ε/2, where M4 is the
flattened fourth moment tensor of G = N 0, I.
We claim that these bounds on MS and TS imply the the conditions about the expectation of even degree2 polynomials. Using Lemma 8.15, we can write p(x) = xT P2 x + p0 for a symmetric matrix P2 ∈ Rd×d
2
and p0 ∈ R. Recall that we write P2♭ ∈ Rd for the flattened out version of P2 . Then, in terms of these, we
can write
E [p(X)] = p0 + tr(P2T MS ) .
X∈u S
115
Lemma 8.15, gives that EX∼G [p(X)] = p0 + tr(P2 ). Thus, we have
| E [p(X)] − E [p(X)]| = |tr(P2T (MS − I))| ≤ kP2 kF kMS − IkF
X∈u S
X∼G
q
≤ Var [p(X)]/2 · dkMS − Ik2
(using Lemma 8.15)
X∼G
q
E [p(X)2 ] .
≤ε
X∼G
Again, in terms of P2 and p0 , we can write:
E [p(X)2 ] = p20 + 2p0 tr(P2T MS ) + P2♭T TS P2♭ ,
X∈u S
and
E [p(X)2 ] = p20 + 2p0 tr(P2 ) + P2♭T M4 P2♭ .
X∼G
We can take norms of the difference between these equations, giving
| E [p(X)2 ] − E [p(X)2 ]| ≤ 2p0 kP2 kF kMS − IkF + kP2 k2F kTS − M4 k2 .
X∈u S
X∼G
We need to express the RHS in terms of EX∼G [p(X)2 ] = VarX∼G [p(X)]+EX∼G [p(X)]2p
. Using Lemma 8.15,
2 = Var
2 ]/2 and |p | = | E
kP
k
[p(X)]/2
≤
E
[p(X)
[p(X)]
+
tr(P
)|
≤
EX∼G [p(X)2 ] +
X∼G
X∼G
0
X∼G
2
p
√2 F
√
2
dkP2 kF ≤ d EX∼G [p(X) ]. Thus, we have
| E [p(X)2 ] − E [p(X)2 ]| ≤ 2p0 kP2 kF kMS − IkF + kP2 k2F kTS − M4 k2
X∈u S
X∼G
√
≤ (2 dkMS − IkF + kTS − M4 k2 ) E [p(X)2 ]
X∼G
3/2
≤ (2d
kMS − Ik2 + kTS − M4 k2 ) E [p(X)2 ]
X∼G
2
≤ ε E [p(X) ] .
X∼G
For the last property, we need a finite set of degree-2 polynomials so that we can take a union bound.
Claim F.3. For any γ > 0, there exists a set C of degree-2 polynomials on Rd of size (γ/d log(N/τ )(d2 +
2
1))d +2 , such that for every degree-2 polynomial p(x) on Rd such no coefficient of p(x) has absolute value
bigger than 1, except possibly for the constant term which has absolute value at most O(d log(N/τ )), there
is a q(x) ∈ C such that for all x with kxk22 = O(d log(N/τ )), |p(x) − q(x)| ≤ γ.
Proof. Note that any degree 2 polynomial on Rd has at most d2 + 1 coefficients.
We take C to be the set of polynomials with coefficients that are multiples of γ/ log(N/τ )d(d2 + 1) with
absolute value no bigger than 1, and constant term a multiple of γ/ log(N/τ )d(d2 +1) with absolute value no
2
bigger than O(d log(N/τ )). Then, |C| ≤ (γ/d log(N/τ )(d2 + 1))d +2 . Now, given p(x), we take q(x) ∈ C
to be the polynomial whose coefficients are the corresponding coefficients of p(x) rounded towards zero to a
multiple of γ/ log(N/τ )d(d2 + 1). When xT Σ−1 x = O(d log(N/τ )), the absolute value of any monomial
of degree at most 2 is bounded from above by kxk2∞ ≤ O(d log(N/τ )). Thus, the absolute value of the
error between corresponding terms of p(x) and q(x) is at most γ/(d2 + 1) and so |p(x) − q(x)| ≤ γ.
Let δ = ε/C(d log(N/τ ))2 for some sufficiently high constant C. We apply the above with γ = δ2 . Let
2
be the set containing q(x), q(x)−δ2 and q(x)+δ2 for all q(x) ∈ C. Then |C ′ | ≤ poly((d log(N/τ ))/ε)d .
By Chernoff and union bound for N ≥ Ω((d log(d/ετ ))6 /ε2 ), then with probability at least 1 − τ /4,
for all q ∈ C ′ that
C′
Pr (q(X) ≥ 0) − Pr (q(X) ≥ 0) < ε/2(d log(N/τ ))2 .
X∈u S
X∼G
116
We now ahow that if for all x ∈ S we have that xT Σ−1 x = O(d log(N/τ )), this implies the last property.
Note that if p(x) satifies the three claims then so does αp(x) for any α > 0. Thus, we may assume without loss of generality that VarX∼G p(x) = 1. Note that if p0 = 0 and x ∈ S, then |p(x)| ≤
xT P2 x ≤ kxk22 kP2 k2 ≤ O(d log(N/τ )) · kP2 kF = O(d log(N/τ )), since by Lemma 8.15, 2kP2 kF =
VarX∼G p(x) = 1. Thus, if |p0 | ≥ Ω(d log(N/τ )) then PrX∈u S (p(X) ≥ 0) is 0 or 1. In this case, the
claim for the minimum value |p0 | can take and have this probabiltiy 0 or 1 respectively is stronger. We
can therefore assume that |p0 | = O(d log(N/τ )). Since kP2 kF = 1/2, the absolute value of each other
coefficient is less than 1.
Thus, there is a q(x) ∈ C with |p(x)−q(x)| ≤ δ2 for all x ∈ S. Note that this q(x) has VarX∼G [q(x)] =
2
2
Θ(1)
p and thus, by anti-concentration bounds, PrX∼G (|q(X)| ≤ δ ) ≤ O(δ). Also, PrX∈G [kXk2 ≥
Ω( d log(N/τ )] ≤ τ /N ≤ δ. Hence, we have:
Pr (p(x) ≥ 0) ≥ Pr (q(x) − δ2 ≥ 0)
X∈u S
X∈u S
≥ Pr (q(X) − δ2 ≥ 0) − Cδ/2
(since q(x) − δ2 ∈ C ′ )
X∼G
≥ Pr (q(X) + δ2 ≥ 0) − (C/2 + O(1))δ
X∼G
p
≥ Pr (p(x) ≥ 0) − (C/2 + O(1))δ − Pr [kXk2 ≥ Ω( d log(N/τ )]
X∼G
X∈G
≥ Pr (p(x) ≥ 0) − Cδ .
X∼G
Similarly PrX∈u S (p(x) ≥ 0) ≤ PrX∼G (p(x) ≥ 0) + Cδ. Thus, we have
Pr (p(X) ≥ 0) ≤ Pr (p(X) ≥ 0) + ε/(d log(N/τ ))2 ,
X∈u S
X∼G
which is stronger than the required bound.
By a union bound, all four of the assumptions we made that hold with probability at least 1 − τ /4, hold
with probability at least 1 − τ .
117
| 10math.ST
|
Under consideration for publication in Theory and Practice of Logic Programming
1
arXiv:cs/0110034v2 [cs.PL] 10 Jul 2003
Inference of termination conditions for numerical
loops in Prolog
ALEXANDER SEREBRENIK and DANNY DE SCHREYE
Department of Computer Science, K.U. Leuven
Celestijnenlaan 200A, B-3001, Heverlee, Belgium
(e-mail: {Alexander.Serebrenik,Danny.DeSchreye}@cs.kuleuven.ac.be)
Abstract
We present a new approach to termination analysis of numerical computations in logic programs.
Traditional approaches fail to analyse them due to non well-foundedness of the integers. We present
a technique that allows overcoming these difficulties. Our approach is based on transforming a program in a way that allows integrating and extending techniques originally developed for analysis of
numerical computations in the framework of query-mapping pairs with the well-known framework of
acceptability. Such an integration not only contributes to the understanding of termination behaviour
of numerical computations, but also allows us to perform a correct analysis of such computations
automatically, by extending previous work on a constraint-based approach to termination. Finally,
we discuss possible extensions of the technique, including incorporating general term orderings.
Keywords: termination analysis, numerical computation.
1 Introduction
One of the important aspects in verifying the correctness of logic programs (as well as
functional programs and term rewrite systems) is verification of termination. Due to the
declarative formulation of programs, the danger of non-termination may be increased. As
a result, termination analysis received considerable attention in logic programming (see
e.g. (Apt et al. 1994; Bossi et al. 2002; Bruynooghe et al. 2002; Codish and Taboch 1999;
Colussi et al. 1995; Decorte et al. 1999; Dershowitz et al. 2001; Genaim et al. 2002; Lindenstrauss and Sagiv 1997;
Mesnard 1996; Mesnard and Ruggieri 2003; Plümer 1991; Ruggieri 1997; Verbaeten et al. 2001)).
Numerical computations form an essential part of almost any real-world program. Clearly,
in order for a termination analyser to be of practical use it should contain a mechanism for
inferring termination of such computations. However, this topic attracted less attention of
the research community. In this paper we concentrate on automatic termination inference
for logic programs depending on numerical computations.
Dershowitz et al. (Dershowitz et al. 2001) showed that termination of general numerical computations, for instance on floating point numbers, may be counter-intuitive, i.e.,
the observed behaviour does not necessarily coincide with the theoretically expected one.
Moreover, as the following program shows, similar results can be obtained even if the
built-in predicates of the underlying language are restricted to include “greater than” and
multiplication only.
Example 1
2
Alexander Serebrenik and Danny De Schreye
Consider the following program, that given a positive number x results in a sequence of
calls p(0.25x ), p((0.25)2 x ), . . .
p(X ) ← X > 0, X 1 is X ∗ 0.25, p(X 1).
If we reason purely in terms of real numbers, we might expect that the computation started
by p(1.0) will be infinite. However, in practice the goal above terminates with respect to
this program, since there exists k , such that (0.25)k is small enough for the comparison
(0.25)k > 0 to fail.
✷
We discuss these issues in detail in (Serebrenik and De Schreye 2002). In the current paper
we avoid these complications by restricting to integer computations only.
Next, we illustrate the termination problem for integer computations with the following
example:
Example 2
Consider the following program:
p(X ) ← X < 7, X 1 is X + 1, p(X 1).
This program terminates for queries p(X ), for all integer values of X .
✷
Most of the existing automated approaches to termination analysis for logic programs (Codish and Taboch 1999;
Lindenstrauss and Sagiv 1997; Mesnard and Neumerkel 2001; Ohlebusch 2001) fail to prove
termination for such examples. The reason is that they are most often based on the notion
of a level mapping, that is, a function from the set of all possible atoms to the natural
numbers, which should decrease while traversing the rules. Usually level mappings are defined to depend on the structure of terms and to ignore constants, making the analysis of
Example 2 impossible.
Of course, this can be easily repaired, by considering level mappings that map each natural number to itself. In fact, the kernels of two termination analysers for logic programs,
namely cTI (Mesnard 1996; Mesnard and Neumerkel 2001) and TerminWeb (Codish and Taboch 1999),
rely on abstracting logic programs to CLP(N ) programs, and use the identity level mapping on N in the analysis of the abstract versions of the programs1.
Note however that this is insufficient for the analysis of Example 2. In fact, there remain
two problems. First, the program in Example 2 is defined on a (potentially negative) integer
argument. This means that we need a level mapping which is different from the identity
function.
Two approaches for solving this problem are possible. First, one can change the definition of the level mapping to map atoms to integers. However, integers are not well-founded.
To prove termination one should prove that the mapping is to some well-founded subset
of integers. In the example above (−∞, 7) forms such a subset with an ordering ≻, such
that x ≻ y if x < y, with respect to the usual ordering on integers. Continuing this line of
thought one might consider mapping atoms to more general well-founded domains. In fact,
already in the early days of program analysis (Floyd 1967; Katz and Manna 1975) general
well-founded domains were discussed. However, the growing importance of automatic termination analysers and requirements of robustness and efficiency stimulated researchers
1
We thank anonymous referees for pointing this link to related work out to us.
Inference of termination conditions
3
to look for more specific instances of well-founded domains, such as natural numbers in
logic programming and terms in term-rewriting systems.
The second approach, that we present in the paper, does not require changing the definition of level mapping. Indeed, the level mapping as required exists. It maps p(X ) to 7 − X
if X < 7 and to 0 otherwise. This level mapping decreases while traversing the rule, i.e.,
the size of p(X ), 7 − X for X < 7, is greater than the size of p(X 1), 6 − X for X < 7 and
0 for X ≥ 7, thus, proving termination.
A second problem with approaches based on the identity function, as the level mapping
used on CLP(N ), is that, even if the program in Example 2 would have been defined on
natural values of X only, they would still not be able to prove termination. The reason is
that the natural argument increases under the standard ordering of the natural numbers.
Such bounded increases (be it of structure-sizes or of numerical values) are not dealt with
by standard termination analysers. Note that the two approaches presented above also solve
this second problem.
The main contribution of this paper is that we provide a transformation - similar to multiple specialisation (Winsborough 1992) - that allows us to define level mappings of the form
illustrated in the second approach above in an automatic way. To do so, we incorporate
techniques of (Dershowitz et al. 2001), such as level mapping inference, in the framework
of the acceptability with respect to a set (De Schreye et al. 1992; Decorte and De Schreye 1998).
This integration provides not only a better understanding of termination behaviour of integer computations, but also the possibility to perform the analysis automatically as in
Decorte et al. (Decorte et al. 1999).
Moreover, we will also be somewhat more general than (Decorte et al. 1999), by studying the problem of termination inference, rather than termination verification. More precisely, we will be inferring conditions that, if imposed on the queries, will ensure that the
queries will terminate. Inference of termination conditions was studied in (Mesnard 1996;
Mesnard and Neumerkel 2001; Genaim and Codish 2001). Unlike termination conditions
inferred by these approaches, stated in terms of groundedness of arguments, our technique
produces conditions based on domains of the arguments, as shown in Example 3.
Example 3
Extend the program of Example 2 with the following clause:
p(X ) ← X > 7, X 1 is X + 1, p(X 1).
This extended program terminates for X ≤ 7 and this is the condition we will infer.
✷
The rest of the paper is organised as follows. After making some preliminary remarks,
we present in Section 3 our transformation—first by means of an example, then more formally. In Section 4 we discuss more practical issues and present the algorithm implementing the termination inference. Section 5 contains the results of an experimental evaluation
of the method. In Section 6 we discuss further extensions, such as proving termination of
programs depending on numerical computations as well as symbolic ones. We summarise
our contribution in Section 7, review related work and conclude.
4
Alexander Serebrenik and Danny De Schreye
2 Preliminaries
We follow the standard notation for terms and atoms. A query is a finite sequence of atoms.
Given an atom A, rel(A) denotes the predicate occurring in A. AtomP (TermP ) denotes
the set of all atoms (terms) that can be constructed from the language underlying P . The
extended Herbrand Base BPE (the extended Herbrand Universe UPE ) is the quotient set of
AtomP (TermP ) modulo the variant relation. An SLD-tree constructed using the left-toright selection rule of Prolog is called an LD-tree. A goal G LD-terminates for a program
P , if the LD-tree for (P , G) is finite.
The following definition is similar to Definition 6.30 (Apt 1997).
Definition 1
Let P be a program and p, q be predicates occurring in it. We say that
• p refers to q in P if there is a clause in P that uses p in its head and q in its body.
• p depends on q in P and write p ⊒ q, if (p, q) is in the transitive closure of the
relation refers to.
• p and q are mutually recursive and write p ≃ q, if p ⊒ q and q ⊒ p.
The only difference between our definition and the one by Apt (Apt 1997) is that we require
the relation ⊒ to be the transitive closure of the relation refers to, while (Apt 1997) requires
it to be transitive, reflexive closure. Using our definition we call a predicate p recursive if
p ≃ p holds.
We recall some basic notions, related to termination analysis. A level mapping is a function | · |: BPE → N , where N is the set of the naturals. Similarly, a norm is a function
k · k : UPE → N .
We study termination of programs with respect to sets of queries. The following notion
is one of the most basic notions in this framework.
Definition 2
Let P be a definite program and S be a set of atomic queries. The call set, Call(P , S ),
is the set of all atoms A from the extended Herbrand Base BPE , such that a variant of
A is a selected atom in some derivation for P ∪ {← Q }, for some Q ∈ S and under the
left-to-right selection rule.
The following definition (Serebrenik and De Schreye 2001) generalises the notion of acceptability with respect to a set (De Schreye et al. 1992; Decorte and De Schreye 1998) by
extending it to mutual recursion.
Definition 3
Let S be a set of atomic queries and P a definite program. P is acceptable with respect to
S if there exists a level mapping | · | such that
• for any A ∈ Call(P , S )
• for any clause A′ ← B1 , . . . , Bn in P , such that mgu(A, A′ ) = θ exists,
• for any atom Bi , such that rel(Bi ) ≃ rel(A) and for any computed answer substitution σ for ← (B1 , . . . , Bi −1 )θ holds that
| A | > | Bi θσ | ·
De Schreye et al. (De Schreye et al. 1992) characterise LD-termination in terms of acceptability.
Inference of termination conditions
5
Theorem 1 (cf. (De Schreye et al. 1992))
Let P be a definite program. P is acceptable with respect to a set of atomic queries S if
and only if P is LD-terminating for all queries in S .
We also need to introduce notions of rigidity and of interargument relations. Given a
norm k · k and a term t , Bossi et al. (Bossi et al. 1991) call t rigid with respect to k · k if
for any substitution σ, kt σk = kt k. Observe that ground terms are rigid with respect to all
norms. The notion of rigidity is obviously extensible to atoms and level mappings. Interargument relations have initially been studied by (Ullman and Van Gelder 1988; Plümer 1991;
Verschaetse and De Schreye 1991). In this paper we use the definition of (Decorte et al. 1999).
Definition 4
Let P be a definite program, p/n a predicate in P . An interargument relation for p/n is a
relation Rp ⊆ N n . Rp is a valid interargument relation for p/n with respect to a norm k · k
if and only if for every p(t1 , . . . , tn ) ∈ AtomP if P |= p(t1 , . . . , tn ) then (kt1 k, . . . , ktn k) ∈
Rp .
Combining the notions of rigidity, acceptability and interargument relations allows us to
reason on termination completely at the clause level.
Theorem 2 (rigid acceptability (cf. (Decorte et al. 1999)))
Let S be a set of atomic queries and P a definite program. Let k · k be a norm and, for each
predicate p in P , let Rp be a valid interargument relation for p with respect to k · k. If there
exists a level mapping | · | which is rigid on Call(P , S ) such that
• for any clause H ← B1 , . . . , Bn ∈ P , and
• for any atom Bi in its body such that rel(Bi ) ≃ rel(H ),
• for substitution θ such that the arguments of the atoms in (B1 , . . . , Bi −1 )θ all satisfy
their associated interargument relations RB1 , . . . , RBi−1 :
| H θ |>| Bi θ |
then P is acceptable with respect to S .
3 Methodology
In this section we introduce our methodology using a simple example. In the subsequent
sections, we formalise it and discuss different extensions.
Computing a query with respect to the following example results in a sequence of
calls with oscillating arguments like p(−2), p(4), p(−16), . . . and stops if the argument
is greater than 1000 or smaller than −1000. The treatment is done first on the intuitive
level.
Example 4
We are interested in proving termination of the set of queries {p(z ) | z is an integer} with
respect to the following program:
p(X ) ← X > 1, X < 1000, X 1 is − X ∗ X , p(X 1).
p(X ) ← X < −1, X > −1000, X 1 is X ∗ X , p(X 1).
6
Alexander Serebrenik and Danny De Schreye
The direct attempt to define the level mapping of p(X ) as X fails, since X can be positive
as well as negative. Thus, a more complex level mapping should be defined. We start with
some observations.
The first clause is applicable if 1 < X < 1000, the second one, if −1000 < X < −1.
Thus, termination of p(X ) for X ≤ −1000, −1 ≤ X ≤ 1 or X ≥ 1000 is trivial. Moreover,
if the first clause is applied and 1 < X < 1000 holds, then either −1000 < X 1 < −1 or
X 1 ≤ −1000 ∨ −1 ≤ X 1 ≤ 1 ∨ X 1 ≥ 1000 should hold. Similarly, if the second clause is
applied and −1000 < X < 1 holds, either 1 < X 1 < 1000 or X 1 ≤ −1000 ∨ −1 ≤ X 1 ≤
1 ∨ X 1 ≥ 1000 should hold.
We use this observation and split the domain of the argument of p, denoted p1 , in three
parts as following:
a
b
c
1 < p1 < 1000
−1000 < p1 < −1
p1 ≤ −1000 ∨ −1 ≤ p1 ≤ 1 ∨ p1 ≥ 1000
Next we replace the predicate p with three new predicates p a , p b and p c . We add conditions before the calls to p to ensure that p a is called if p(X ) is called and 1 < X < 1000
holds, p b is called if p(X ) is called and −1000 < X < −1 holds and p c is called if p(X )
is called and X ≤ −1000 ∨ −1 ≤ X ≤ 1 ∨ X ≥ 1000 holds. The following program is
obtained:
p a (X ) ← X > 1, X < 1000, X 1 is − X ∗ X ,
−1000 < X 1, X 1 < −1, p b (X 1).
a
p (X ) ← X > 1, X < 1000, X 1 is − X ∗ X ,
(X 1 ≤ −1000; (−1 ≤ X 1, X 1 ≤ 1); X 1 ≥ 1000), p c (X 1).
b
p (X ) ← X < −1, X > −1000, X 1 is X ∗ X ,
1 < X 1, X 1 < 1000, p a (X 1).
p b (X ) ← X < −1, X > −1000, X 1 is X ∗ X ,
(X 1 ≤ −1000; (−1 ≤ X 1, X 1 ≤ 1); X 1 ≥ 1000), p c (X 1).
Observe that the transformation we performed is a form of multiple specialisation, wellknown in the context of abstract interpretation (Winsborough 1992).
Now we define three different level mappings, one for atoms of p a , another one for
atoms of p b and the last one for atoms of p c . Let
1000 − n if 1 < n < 1000
| p a (n) | =
0
otherwise
1000
+
n
if
−1000 < n < −1
| p b (n) | =
0
otherwise
| p c (n) | = 0
We verify acceptability of the transformed program with respect to {p a (n) | 1 < n <
1000} ∪ {p b (n) | −1000 < n < −1} via the specified level mappings. This implies termination of the transformed program with respect to these queries, and thus, termination
of the original program with respect to {p(z ) | z is an integer}.
For the sake of brevity we discuss only queries of the form p a (n) for 1 < n < 1000.
Heads of the first and the second clauses can be unified with this query, however, the
Inference of termination conditions
7
second clause does not contain calls to predicates mutually recursive with p a and the only
such atom in the first clause is p b (m), where m = −n 2 . Then, | p a (n) | > | p b (m) | should
hold, i.e., 1000 − n > 1000 + m, that is 1000 − n > 1000 − n 2 (n > 1 and m = −n 2 ), which
is true for n > 1.
For queries of the form p b (n), the acceptability condition is reduced to 1000 + n >
1000 − n 2 which is true for n < −1.
✷
The intuitive presentation above hints at the main issues to be discussed in the following sections: how the cases such as those above can be extracted from the program, and
how given the extracted cases, the program should be transformed. Before discussing the
answers to these questions we present some basic notions.
3.1 Basic notions
In this section we formally introduce some notions that further analysis will be based on.
Recall that the aim of our analysis is to find, given a predicate and a query, a sufficient
condition for termination of this query with respect to this program. Thus, we need to
define a notion of a termination condition. We start with a number of auxiliary definitions.
Given a predicate p, pi denotes the i-th argument of p and is called argument position
denominator.
Definition 5
Let P be a program, S be a set of queries. An argument position i of a predicate p is called
integer argument position, if for every p(t1 , . . . , tn ) ∈ Call(P , S ), ti is an integer.
Argument position denominators corresponding to integer argument positions will be called
integer argument position denominators.
An integer inequality is an atom of one of the following forms Exp1 > Exp2, Exp1 <
Exp2,Exp1 ≥ Exp2 or Exp1 ≤ Exp2, where Exp1 and Exp2 are constructed from integers,
variables and the four operations of arithmetics on integers. A symbolic inequality over
the arguments of a predicate p is constructed similarly to an integer inequality. However,
instead of variables, integer argument positions denominators are used.
Example 5
X > 0 and Y ≤ X + 5 are integer inequalities. Given a predicate p of arity 3, having only
integer argument positions, p1 > 0 and p2 ≤ p1 + p3 are symbolic inequalities over the
arguments of p.
✷
Disjunctions of conjunctions based on integer inequalities are called integer conditions.
Similarly, disjunctions of conjunctions based on symbolic inequalities over the arguments
of the same predicate are called symbolic conditions over the integer arguments of this
predicate.
Definition 6
Let p(t1 , . . . , tn ) be an atom and let cp be a symbolic condition over the arguments of p.
An instance of the condition with respect to an atom, denoted cp (p(t1 , . . . , tn )), is obtained
by replacing the argument positions denominators with the corresponding arguments, i.e.,
pi with ti .
8
Alexander Serebrenik and Danny De Schreye
Example 6
Let p(X , Y , 5) be an atom and let cp be a symbolic condition (p1 > 0) ∧ (p2 ≤ p1 + p3 ).
Then, cp (p(X , Y , 5)) is (X > 0) ∧ (Y ≤ X + 5).
✷
Now we are ready to define termination condition formally.
Definition 7
Let P be a program, and Q be an atomic query. A symbolic condition crel(Q ) is a termination condition for Q if given that crel(Q ) (Q ) holds, Q left-terminates with respect to
P.
For any integer z a termination condition for p(z ) with respect to Examples 2 and 4
is true, i.e., for any integer z , p(z ) terminates with respect to these programs. Clearly,
more than one termination condition is possible for a given query with respect to a given
program. For example, termination conditions for p(5) with respect to Example 3, are
among others, true, p1 ≤ 7 and p1 > 0. Analogously, false, p1 ≤ 7, p1 < 10 are termination
conditions for p(11) with respect to Example 3. It should also be noted that a disjunction
of two termination conditions is always a termination condition.
Similarly to Theorems 1 and 2 we would like to consider termination with respect to
sets of atomic queries. Therefore we extend the notion of termination condition to a set
of queries. This, however, is meaningful only if all the queries of the set have the same
predicate. We call such a set single predicate set of atomic queries. For a single predicate
set of atomic queries S , rel(S ) denotes the predicate of the queries of the set.
Definition 8
Let P be a program, and S be a single predicate set of atomic queries. A symbolic condition
crel(S ) is a termination condition for S if crel(S ) is a termination condition for all Q ∈ S .
From the discussion above it follows that a termination condition for S = {p(z ) | z is an
integer} with respect to Examples 2 and 4 is true. This is not the case for Example 3, since
termination is observed only for some queries of S , namely p(z ), such that z ≤ 7. Thus,
p1 ≤ 7 is a termination condition for S with respect to Example 3.
We discuss now inferring what values integer arguments can take during traversal of the
rules, i.e., the “case analysis” performed in Example 4. It provides already the underlying
intuition—calls of the predicate p c are identical to the calls of the predicate p, where c
holds for its arguments. More formally, we define a notion of a set of adornments. Later
we specify when it is guard-tuned and we show how such a guard-tuned set of adornments
can be constructed.
Definition 9
Let p be a predicate. The set Ap = {c1 , . . . , cn } of symbolic conditions over the integer
W
arguments of p is called set of adornments for p if ni=1 ci = true and for all i, j such that
1 ≤ i < j ≤ n, ci ∧ cj = false.
A set of adornments partitions the domain for (some of) the integer variables of the
predicate. Similarly to Example 4, in the examples to come, elements of a set of adornments
are denoted a, b, c, . . .
Inference of termination conditions
9
Example 7
Example 4, continued. The following are examples of sets of adornments:
{a, b, c}
where a is 1 < p1 < 1000, b is − 1000 < p1 < −1
and c is p1 ≤ −1000 ∨ −1 ≤ p1 ≤ 1 ∨ p1 ≥ 1000.
and
{d, e} where d is p1 ≤ 100 and e is p1 > 100.
✷
In the next section we are going to present a transformation, related to the multiple
specialisation technique. To define it formally we introduce the following definition:
Definition 10
Let H ← B1 , . . . , Bn be a rule. B1 , . . . , Bi , is called an integer prefix of the rule, if for all
j , 1 ≤ j ≤ i ≤ n, Bj is an integer inequality and the only variables in its arguments are
variables of H . B1 , . . . , Bi is called the maximal integer prefix of the rule, if it is an integer
prefix and B1 , . . . , Bi , Bi +1 is not an integer prefix.
Since an integer prefix constrains only variables appearing in the head of a clause, there
exists a symbolic condition over the arguments of the predicate of the head, such that the
integer prefix is its instance with respect to the head. In general, this symbolic condition is
not necessarily unique.
Example 8
Consider the following program: p(X , Y , Y ) ← Y > 5· The only integer prefix of this rule
is Y > 5. There are two symbolic conditions over the arguments of p, p2 > 5 and p3 > 5,
such that Y > 5 is their instance with respect to p(X , Y , Y ).
✷
In order to guarantee the uniqueness of such symbolic conditions we require integer argument positions in the heads of the rules to be occupied by distinct variables. For the sake
of simplicity we assume all argument positions in the heads of the rules to be occupied by
distinct variables. Apt et al. (Apt et al. 1994) call such a rule homogeneous. Analogously, a
logic program is called homogeneous if all its clauses are homogeneous. Programs can be
easily rewritten to a homogeneous form (see (Apt et al. 1994)). In the following we assume
that all programs are homogeneous.
3.2 Program transformation
The next question that should be answered is how the program should be transformed given
a set of adornments. After this transformation p c (X1 , . . . , Xn ) will behave with respect
to the transformed program exactly as p(X1 , . . . , Xn ) does, for all calls that satisfy the
condition c. Intuitively, we replace each call to the predicate p in the original program by
a number of possible calls in the transformed one.
S
Given a program P and a set of possible adornments A = p ∈P Ap , the transformation
is performed in a number of steps. Below we use Example 2 as a running example to
illustrate the different steps. Recall that it consists of only one clause
p(X ) ← X < 7, X 1 is X + 1, p(X 1).
10
Alexander Serebrenik and Danny De Schreye
As set of adornments we use
Ap = {a, b}, where a is p1 < 7 and b is p1 ≥ 7.
1. For each clause r in P and for each call p(t1 , . . . , tn ) to a recursive predicate p
W
occurring in r add c ∈Ap c(p(t1 , . . . , tn )) before p(t1 , . . . , tn ). By Definition 9 the
disjunction is true, thus, the transformed program is equivalent to the original one.
In the example, the clause is transformed to
p(X ) ← X < 7, X 1 is X + 1, (X 1 < 7 ; X 1 ≥ 7), p(X 1).
2. For each clause, such that the head of the clause, say p(t1 , . . . , tn ), has a recursive
W
predicate p, add c ∈Ap c(p(t1 , . . . , tn )) as the first subgoal in its body. As for the
previous step, the introduced call is equivalent to true, so that the transformation is
obviously correct. In the example, we obtain:
p(X ) ← (X < 7 ; X ≥ 7), X < 7,
X 1 is X + 1, (X 1 < 7 ; X 1 ≥ 7), p(X 1).
3. Next, moving to an alternative procedural interpretation of disjunction, for each
clause in which we introduced a disjunction in one of the previous two steps, and for
W
each such introduced disjunction c ∈Ap c(p(t1 , . . . , tn )) we split these disjunctions,
introducing a separate clause for each disjunct. Thus, we apply the transformation
H
← B1 , . . . , (A1 ; . . . ; Ak ), . . . , Bn .
H
← B1 , . . . , A1 , . . . , Bn .
H
← B1 , . . . , A2 , . . . , Bn .
..
.
H
← B1 , . . . , Ak , . . . , Bn .
to
to each disjunction introduced in steps 1 and 2.
For our running example, we obtain four clauses:
p(X ) ← X
p(X ) ← X
p(X ) ← X
p(X ) ← X
< 7, X
< 7, X
≥ 7, X
≥ 7, X
< 7, X 1 is X + 1, X 1 < 7, p(X 1).
< 7, X 1 is X + 1, X 1 ≥ 7, p(X 1).
< 7, X 1 is X + 1, X 1 < 7, p(X 1).
< 7, X 1 is X + 1, X 1 ≥ 7, p(X 1).
Note that, although this transformation is logically correct, it is not correct for Prolog programs with non-logical features. For instance, in the presence of “cut”, it may
produce a different computed answer set. Also, in the context of “read” or “write”
calls, the procedural behaviour may become very different. However, for purely logical programs with integer computations, both the declarative semantics and the computed answer semantics are preserved. Likewise, the termination properties are also
preserved. Indeed, the transformation described can be seen as a repeated unfolding
of ; using the following clauses:
; (X , Y ) ← X .
; (X , Y ) ← Y .
Inference of termination conditions
11
It is well-known that unfolding cannot introduce infinite derivations (Bossi and Cocco 1994).
On the other hand, an infinite derivation of the original program can be easily mimicked by the transformed program.
From here on we will restrict our attention to purely logical programs, augmented
with integer arithmetic. To prepare the next step in the transformation, note that, in
the program resulting from step 3, for each rule r and for each recursive predicate p:
• if a call p(t1 , . . . , tn ) occurs in r , then it is immediately preceded by some c(p(t1 ,
. . . , tn )),
• if an atom p(t1 , . . . , tn ) occurs as the head of r , then it is immediately followed
by some c(p(t1 , . . . , tn )).
Moreover, since the elements Ap partition the domain (see Definition 9), conjuncts
like ci (p(t1 , . . . , tn )), p(t1 , . . . , tn ) and cj (p(t1 , . . . , tn )), p(t1 , . . . , tn ) for i 6= j , are
mutually exclusive, as well as the initial parts of the rules, like
p(t1 , . . . , tn ) ← ci (p(t1 , . . . , tn )) and p(t1 , . . . , tn ) ← cj (p(t1 , . . . , tn )),
i 6= j . This means that we can now safely rename the different cases apart.
4. Replace each occurrence of c(p(t1 , . . . , tn )), p(t1 , . . . , tn ) in the body of the clause
with c(p(t1 , . . . , tn )), p c (t1 , . . . , tn ) and each occurrence of a rule
p(t1 , . . . , tn ) ← c(p(t1 , . . . , tn )), B1 , . . . , Bn
with a corresponding rule
p c (t1 , . . . , tn ) ← c(p(t1 , . . . , tn )), B1 , . . . , Bn .
In our example we get:
p a (X ) ← X
p a (X ) ← X
p b (X ) ← X
p b (X ) ← X
< 7, X
< 7, X
≥ 7, X
≥ 7, X
< 7, X 1 is X
< 7, X 1 is X
< 7, X 1 is X
< 7, X 1 is X
+ 1, X 1 < 7, p a (X 1).
+ 1, X 1 ≥ 7, p b (X 1).
+ 1, X 1 < 7, p a (X 1).
+ 1, X 1 ≥ 7, p b (X 1).
Because of the arguments presented above, the renaming is obviously correct, in the
sense that the LD-trees that exist for the given program and for the renamed program
are identical, except for the names of the predicates and for a number of failing 1step derivations (due to entering clauses that fail in their guard in the given program).
As a result, both the semantics (up to renaming) and the termination behaviour of
the program are preserved.
5. Remove all rules with a maximal integer prefix which is inconsistent, and remove
from the bodies of the remaining clauses all subgoals that are preceded by an inconsistent conjunction of inequalities. In the example, both rules defining p b are
eliminated and we obtain:
p a (X ) ← X < 7, X < 7, X 1 is X + 1, X 1 < 7, p a (X 1).
p a (X ) ← X < 7, X < 7, X 1 is X + 1, X 1 ≥ 7, p b (X 1).
Performing this step requires verifying the consistency of a set of constraints, a task
12
Alexander Serebrenik and Danny De Schreye
that might be computationally expensive. Depending on the constraints the implementation of our technique is supposed to deal with, the programmer can either
opt for more restricted but potentially faster solvers, such as linear rational solver
(Holzbaur 1995), or for more powerful but potentially slower ones, such as mixed
integer programminging solver (ILOG 2001).
6. Replace each rule
p c (t1 , . . . , tn ) ← c(p(t1 , . . . , tn )), B1 , . . . , Bn
by a rule
p c (t1 , . . . , tn ) ← B1 , . . . , Bn ·
In the example we obtain:
p a (X ) ← X < 7, X 1 is X + 1, X 1 < 7, p a (X 1)·
p a (X ) ← X < 7, X 1 is X + 1, X 1 ≥ 7, p b (X 1)·
which is the adorned program, P A (P {a,b} in our case). Note that this last step is only
correct if we also transform the set of original queries. Namely, given a single prediS
cate set of original atomic queries S for P and a set of adornments A = p ∈P Ap , the
corresponding set of queries considered for P A is S A = {c1 (Q ) ∧ Q c1 , . . . , cn (Q ) ∧
Q cn | Q ∈ S , {c1 , . . . , cn } = Arel(Q ) }, where Q c denotes p c (t1 , . . . , tn ) if Q is
p(t1 , . . . , tn ). In our running example the set of queries is {z < 7 ∧ p a (z ), z ≥ 7 ∧
p b (z ) | z is an integer}.
Before stating our results formally we illustrate the transformation by a second example.
Example 9
Example 4, continued. With the first set of adornments from Example 7 we obtain P {a,b,c} :
p a (X ) ← X > 1, X < 1000, X 1 is − X ∗ X ,
−1000 < X 1, X 1 < −1, p b (X 1).
a
p (X ) ← X > 1, X < 1000, X 1 is − X ∗ X ,
(X 1 ≤ −1000; (−1 ≤ X 1, X 1 ≤ 1); X 1 ≥ 1000), p c (X 1).
b
p (X ) ← X < −1, X > −1000, X 1 is X ∗ X ,
1 < X 1, X 1 < 1000, p a (X 1).
b
p (X ) ← X < −1, X > −1000, X 1 is X ∗ X ,
(X 1 ≤ −1000; (−1 ≤ X 1, X 1 ≤ 1); X 1 ≥ 1000), p c (X 1).
The set of queries to be considered is
{
z > 1 ∧ z < 1000 ∧ p a (z ),
z > −1000 ∧ z < −1 ∧ p b (z ),
(z ≤ −1000 ∨ (z ≥ −1 ∧ z ≤ 1) ∨ z ≥ 1000) ∧ p c (z )
|
}
z is an integer
Inference of termination conditions
13
If the second set of adornments is used, the program P {d,e} is obtained:
p d (X ) ← X
p e (X ) ← X
p d (X ) ← X
p d (X ) ← X
> 1, X < 1000, X 1 is − X ∗ X , X 1 ≤ 100, p d (X 1).
> 1, X < 1000, X 1 is − X ∗ X , X 1 ≤ 100, p d (X 1).
< −1, X > −1000, X 1 is X ∗ X , X 1 ≤ 100, p d (X 1).
< −1, X > −1000, X 1 is X ∗ X , X 1 > 100, p e (X 1).
Analogously, the following is the set of the corresponding queries
{z ≤ 100 ∧ p d (z ), z > 100 ∧ p e (z ) | z is an integer}
✷
Formally, the following lemma holds:
Lemma 1
Let P be a definite pure logical program with integer computations, let Q be an atomic
query, let A = ∪Ap be a set of adornments and let c be an adornment in Arel(Q ) . Let P A be
a program obtained as described above with respect to A . Then, c is a termination condition
for Q with respect to P if and only if P A LD-terminates with respect to c(Q ) ∧ Q c .
Proof
The construction of P A implies that the LD-tree of c(Q ) ∧ Q c with respect to P A is
isomorphic to the LD-tree of c(Q ) ∧ Q with respect to P , implying the theorem.
Again, in practice we do not prove termination of a single query, but of a single predicate set of queries. Furthermore, recalling that a disjunction of termination conditions is
a termination condition itself we can generalise our lemma to disjunctions of adornments.
Taking these two considerations into account, the following theorem holds.
Theorem 3
Let P be a definite pure logical program with integer computations, let S be a single
predicate set of atomic queries, let A = ∪Ap be a set of adornments and let c1 , . . . , cn be
adornments in Arel(S ) . Let P A be a program obtained as described above with respect to
A . Then, c1 ∨ . . . ∨ cn is a termination condition for S with respect to P if and only if P A
LD-terminates with respect to {c1 (Q ) ∧ Q c1 , . . . , cn (Q ) ∧ Q cn | Q ∈ S }.
Proof
Immediately from Lemma 1 and the preceding observations.
The goal of the transformation presented is, given a program and a partition of the domain, to generate a program having separate clauses for each one of the cases. Clearly,
this may (and usually will) increase the number of clauses. Each clause can be replaced
by maximum c n +1 new clauses, were c is a number of adornments and n is a number of
recursive body subgoals. Thus, the size of the transformed program doesn’t exceed
r × c n +1 ,
(1)
where r is the size of the original one. This may seem a problematically large increase,
however, the number of recursive body atoms (depending on numerical arguments) in numerical programs is usually small.
14
Alexander Serebrenik and Danny De Schreye
Since the transformation preserves termination, acceptability of the transformed program implies termination of the original program. In the next section we will see that having separate clauses for different cases allows us to define less sophisticated level-mappings
for proving termination. Such level-mappings can be constructed automatically, and thus,
play a key role in automation of the approach.
4 Generating adornments, level mappings and termination constraints
In the previous section we have shown the transformation that allows reasoning on termination of the numerical computations in the framework of acceptability with respect to
a set of queries. In this section we discuss how adornments can be generated, how level
mappings can be proposed and which termination conditions finally turn up.
4.1 Guard-tuned sets of adornments
In Example 7 we have seen two different sets of adornments. Both of them are valid according to Definition 9. However, recalling P {a,b,c} and P {d,e} as shown in Example 9,
we conclude that {a, b, c} is in some sense preferable to {d, e}. Observe that P {d,e} does
not only have two mutually recursive predicates, as P {a,b,c} does, but also self-loops on
one of the predicates. To distinguish between “better” and “worse” sets of adornments we
define guard-tuned sets of adornments.
Intuitively, a set of adornments of a predicate p is guard-tuned if it is based on “subcases”
of maximal integer prefixes.
Definition 11
Let P be a homogeneous program, let p be a predicate in P . A set of adornments Ap
is called guard-tuned if for every A ∈ Ap and for every rule r ∈ P , defining p, with the
symbolic condition c corresponding to its maximal integer prefix, either c ∧ A = false or
c ∧ A = A holds.
Example 10
The first set of adornments, presented in Example 7, is guard-tuned while the second one
is not guard-tuned.
✷
Examples 7 and 10 suggest the following way of constructing a guard-tuned set of adornments. First, we collect the symbolic conditions, corresponding to the maximal integer
prefixes of the rules defining a predicate p (we denote this set Cp ). Let Cp be {c1 , . . . , cn }.
Then we define Ap to be the set of all conjunctions ∧ni=1 di , where di is either ci or ¬ci .
Computing Ap might be exponential in the number of elements of Cp , i.e., in the number
of maximal integer prefixes. The number of integer prefixes is bounded by the number of
clauses. Thus, recalling (1), the upper bound on the size of the transformed program is
r × 2r (n +1) , i.e., it is exponential in the number of clauses r and in a number of recursive
subgoals n. However, again our experience suggests that numerical parts of real-world
programs are usually relatively small and depend on one or two different integer prefixes.
Analogously, clauses having more than two recursive body subgoals are highly exceptional. Therefore, we conclude that in practice the size of the transformed program is not
problematic.
Inference of termination conditions
15
We claim that the constructed set Ap is always a guard-tuned set of adornments. Before
stating this formally, consider the following example.
Example 11
Consider the following program.
r (X ) ← X > 5.
r (X ) ← X > 10, r (X ).
Then, Cr = {r1 > 5, r1 > 10}. The following conjunctions can be constructed from the
elements of Cp and their negations: {r1 > 5 ∧ r1 > 10, r1 > 5 ∧ ¬(r1 > 10), ¬(r1 > 5) ∧
r1 > 10, ¬(r1 > 5) ∧ ¬(r1 > 10)}. After simplifying and removing inconsistencies Ar =
{r1 > 10, r1 > 5 ∧ r1 ≤ 10, r1 ≤ 5}.
✷
Lemma 2
Let P be a program, p be a predicate in P and Ap be constructed as described. Then Ap is
a guard-tuned set of adornments.
Proof
The proof is done by checking the definitions.
1. Let a1 , a2 ∈ Ap and a1 6= a2 . Then, there exists ci ∈ Cp , such that a1 = d1 ∧ . . . ∧ ci ∧
. . . ∧ dn and a2 = d1 ∧ . . . ∧ ¬ci ∧ . . . ∧ dn . Thus, a1 ∧ a2 = false.
2. By definition of Ap , ∨ai ∈Ap ai = true. Thus, Ap is a set of adornments.
3. Let a ∈ Ap be an adornment and let c be a symbolic condition corresponding to the
maximal integer prefix of a rule. By definition of Cp , c ∈ Cp . Thus, either c is one
of the conjuncts of a or ¬c is one of the conjuncts of a. In the first case, c ∧ a = a.
In the second case c ∧ a = c ∧ (¬c) = false. Therefore, Ap is a guard-tuned set of
adornments.
From here on we assume that all sets of adornments are guard-tuned.
4.2 How to define a level mapping.
One of the questions that should be answered is how the level mappings should be generated automatically. Clearly, one cannot expect automatically defined level mappings to be
powerful enough to prove termination of all terminating examples. In general we cannot
hope but for a good guess.
The problem with level mappings is that they should reflect changes on possibly negative arguments and remain non-negative at the same time. We also like to remain in the
framework of level mappings on atoms defined as linear combinations of sizes of their
arguments (Bossi et al. 1994). We solve this problem by defining different level mappings
for different adorned versions of the predicate. The major observation underlying the technique presented in this subsection is that if p1 > p2 appears in the adornment of a recursive
clause, then for each call to this adorned predicate p1 − p2 will be positive, and thus, can
be used for defining a level mapping. On the other hand, p1 < p2 can always be interpreted
as p2 > p1 . These observations form a basis for definition of a primitive level mapping.
16
Alexander Serebrenik and Danny De Schreye
Definition 12
Let p c be an adorned predicate. The primitive level mapping, | · |pr , is defined as
• if c is E1 ρ E2 , where E1 and E2 are expressions and ρ is either > or ≥ then
(E1 − E2 )(t1 , . . . , tn ) if E1 (t1 , . . . , tn ) ρ E2 (t1 , . . . , tn )
c
pr
| p (t1 , . . . , tn ) | =
0
otherwise
• if c is E1 ρ E2 , where E1 and E2 are expressions and ρ is either < or ≤ then
(E2 − E1 )(t1 , . . . , tn ) if E1 (t1 , . . . , tn ) ρ E2 (t1 , . . . , tn )
| p c (t1 , . . . , tn ) |pr =
0
otherwise
• otherwise,
| p c (t1 , . . . , tn ) |pr = 0·
If more than one conjunct appears in the adornment, the level mapping is defined as a
linear combination of primitive level mappings corresponding to the conjuncts.
Definition 13
Let p c1 ∧...∧cn be an adorned predicate such that each ci is E1i ρi E2i for some expressions
E1i and E2i and ρi is either > or ≥. Let wc1 , . . . , wcn be natural numbers. Then, a level
mapping | · | satisfying
| p c1 ∧...∧cn (t1 , . . . , tn ) | = ∑ wci | p ci (t1 , . . . , tn ) |pr ,
i
is called a natural level mapping.
Example 12
The level mappings used in Example 4 are natural level mappings such that wp1 >1 =
wp1 <−1 = 0, wp1 <1000 = wp1 >−1000 = 1. We have seen that these level mappings are powerful enough to prove termination.
✷
The definition of natural level mapping implies that if c is a disjunction, it is ignored.
The reason for doing so is that disjunctions are introduced only as negations of symbolic
constraints corresponding to maximal integer prefixes of the rules. Thus, they signify that
some rule cannot be applied, and can be ignored.
Example 13
Example 4, continued. Recalling that c denotes p1 ≤ −1000 ∨ −1 ≤ p1 ≤ 1 ∨ p1 ≥ 1000
the following holds for any integer n, | p c (n) |pr = 0.
✷
Of course, if the original program already contains disjunctions of numerical constraints,
then we transform it in a preprocessing to eliminate the disjunctions.
As the following example illustrates, natural level mappings gain their power from the
fact that the set of adornments used is guard-tuned.
Inference of termination conditions
17
Example 14
In Example 7 we have seen two different sets of adornments. We have seen in Example 4
that if a guard-tuned set of adornments is chosen the natural level mapping is powerful
enough to prove acceptability and, thus, termination. If a non guard-tuned set of adornments is chosen, the second program of Example 9 is obtained. Then, a following natural
level mapping is defined (for some natural numbers wp1 ≤100 and wp1 >100 ):
100 − X if X ≤ 100
| p d (X ) | = wp1 ≤100 ∗
0
otherwise
X
−
100
if
X > 100
| p e (X ) | = wp1 >100 ∗
0
otherwise
Consider the following clause.
p d (X ) ← X > 1, X < 1000, X 1 is − X ∗ X , X 1 ≤ 100, p d (X 1).
In order to prove acceptability we have to show that the size of the call to p d (X ) is greater
than the size of the corresponding call to p d (X 1). If the first argument x at the call to
p d (X ) is greater than 1 and less than 10, the acceptability decrease requires wp1 ≤100 (100 −
x ) > wp1 ≤100 (100 + x 2), contradicting x > 1 and wp1 ≤100 being a natural number. Thus,
acceptability cannot be proved with natural level mappings.
✷
The approach of (Decorte et al. 1999) defines symbolic counterparts of the level mappings and infers the values of the coefficients by solving a system of constraints. Intuitively,
instead of considering wci ’s as given coefficients, they are regarded as variables. More formally, similarly to (Decorte et al. 1999), we introduce the following notion.
Definition 14
Let p c1 ∧...∧cn be an adorned predicate. A symbolic counterpart of a natural level mapping
is an expression:
| p c1 ∧...∧cn (t , . . . , t ) |s = W | p ci (t , . . . , t ) |pr ,
1
n
∑
ci
1
n
i
where the Wci ’s are symbols, associated to a predicate p c1 ∧...∧cn .
The intuition behind the symbolic counterpart of a natural level mapping is that natural
level mappings are instances of it. Therefore, we also require Wc ≥ 0 to hold for any
constraint c.
Example 15
Example 4, continued. Recalling that a stands for 1 < p1 < 1000, a symbolic counterpart
of a natural level mapping for p a (n) is Wp1 >1 (n − 1) + Wp1<1000(1000 − n).
✷
In order to verify the rigid acceptability condition (Theorem 2) interargument relations
may be required as well. Interargument relations are usually represented as saying that a
weighted sum of sizes of some arguments (with respect to a given norm) is greater or equal
to a weighted sum of sizes of other arguments (see e.g. (Plümer 1990)). In the numerical
case these sizes should be replaced with expressions as used in Definition 12. Observe
that for simpler examples no interargument relations are needed. Symbolic counterparts
of norms and interargument relations can be defined analogously to Definition 14. In the
next subsection we discuss how the symbolic counterparts are used to infer termination
conditions.
18
Alexander Serebrenik and Danny De Schreye
4.3 Inferring termination constraints
In this section, we combine the steps studied so far into an algorithm that infers termination
conditions. The program transformation, described in Section 3.2, implies that a trivial
termination condition can be computed as a disjunction of the adornments corresponding
to the predicates that can be completely unfolded, i.e., to the predicates that do not depend
directly or indirectly on recursive predicates. More formally we can draw the following
corollary from Theorem 3:
Corollary 1
Let P be a program, let S be a single predicate set of atomic queries and let A be a set of
adornments for rel(S ). Let A = {c | c ∈ A , for all q such that rel(S )c ⊒ q : q is not recursive
W
inP A }. Then c ∈A c is a termination condition for S with respect to P .
Proof
By definition of A, for all Q ∈ S , P A LD-terminates with respect to {c(Q ) ∧ Q c | c ∈ A}.
W
Thus, by Theorem 3, c ∈A c is a termination condition for S with respect to P .
Example 16
The termination condition constructed according to Corollary 1 for Example 4 is c, i.e.,
(p1 ≤ −1000) ∨ (−1 ≤ p1 ≤ 1) ∨ (p1 ≥ 1000).
✷
In general, the termination condition is constructed as a disjunction of two conditions:
cond1 for non-recursive cases, according to Corollary 1, and cond2 , for recursive cases.
The later condition is initialised to be ¬cond1 and further refined by adding constraints
obtained from the rigid acceptability condition, as in (Decorte et al. 1999)2 . The algorithm
is sketched in Figure 1.
Termination inference is inspired by the constraints-based approach of Decorte et al. (Decorte et al. 1999).
Similarly to their work we start by constructing symbolic counterparts of the level mappings (Definition 14) and interargument relations, and construct conditions following from
rigid acceptability (Theorem 2) and validity of interargument relations. Unlike their work
in our case no rigidity constraints are needed (since integer arguments are ground and obviously rigid) and norms are fixed. Thus, the constraints system turns out to be simpler and
better suited for automation. Finally, the conditions constructed are solved with respect
to the symbolic variables (Wci ’s). In Example 17 below we are going to see that rigid
acceptability will be implied by the following constraint (Σ):
Wq1 >q2 (X − Y ) > Wq1 >q2 ((X − Y ) − Y ),
that is Wq1 >q2 Y > 0 should hold. In both approaches Wq1 >q2 ≥ 0 is required to hold. At
this point the approach of (Decorte et al. 1999), interpreting Y as a norm of arguments
(i.e., Y is a natural number), will conclude Wq1 >q2 > 0. In our case, we do not know
a priori that Y is a natural number. Therefore, we would infer from Σ that Y > 0 and
Wq1 >q2 > 0.
In general, given the system of constraints inferred by (Decorte et al. 1999), we distinguish between the following cases:
2
Any other technique proving termination and able to provide some constraint that, if satisfied, implies termination can be used instead of (Decorte et al. 1999).
Inference of termination conditions
19
• There is no solution. We report cond1 as a termination condition (Corollary 1). Observe that when the algorithm reports the termination condition to be false, it suspects
the possibility of non-termination.
• There is a solution for any values of integer variables. Namely, there are natural level
mappings and interargument relations that prove termination of the program for any
values of integer variables. Termination condition in this case is, thus, true.
• There is a solution for some values of integer variables. In other words, the solution
constrains integer variables appearing in the clauses. Two cases can be distinguished:
— Integer variables constrained appear in the heads of the clauses. Then, constraints
on these variables can be regarded as constraints on the arguments of the queries
posed. In this case termination can be shown if these constraints are fulfilled.
— Integer variables constrained do not appear in the heads of the clauses. In this
case our methodology is too weak to obtain some information implying termination of the queries. The best we can do is to report termination for cond1 .
Let P be a homogeneous program, let S be a single predicate set of atomic queries and let q
be rel(S ).
1. For each p ≃ q construct a guard-tuned set Ap . (Section 4.1)
S
2. Adorn P with respect to q and p≃q Ap . (Section 3.2)
3. Let A = {c | c ∈ Aq , for all p such that q c ⊒ p : p is not recursive in P A }.
W
W
Let cond1 = c∈A c. Let cond2 = c∈Aq ,c6∈A c.
4. Let S ′ be {c(Q) ∧ Q c | c ∈ Aq , c 6∈ A, c(Q) ∧ Q c ∈ S A }.
5. Define the symbolic counterparts of level mappings and interargument relations. (Section 4.2)
6. Let Σ be a set of constraints on the symbolic variables, following from rigid acceptability of S ′ with respect to P A and validity of interargument relations.
7. Solve Σ with respect to the symbolic variables.
(a) Solution of Σ doesn’t produce extra constraints on variables.
Report termination for true.
(b) Solution of Σ produces extra constraints on integer variables, appearing in the
heads of the clauses.
Conjunct these constraints to termination condition cond2 .
Report termination for cond1 ∨ cond2 .
(c) There is no solution or integer variables, constrained by the solution of Σ, do not
appear in the heads of the clauses
Report termination for cond1 .
Fig. 1. Termination Inference Algorithm
Example 17
Consider the following program.
q(X , Y ) ← X > Y , Z is X − Y , q(Z , Y )·
We look for integer values of X and Y such that q(X , Y ) terminates. First, the algorithm
20
Alexander Serebrenik and Danny De Schreye
infers adornments. In our case {a, b} are inferred, such that a denotes q1 > q2 and b denotes
q1 ≤ q2 .
The adorned version of this program is
q a (X , Y ) ← X > Y , Z is X − Y , Z > Y , q a (Z , Y )·
q a (X , Y ) ← X > Y , Z is X − Y , Z ≤ Y , q b (Z , Y )·
The corresponding set of queries is
{x > y ∧ q a (x , y), x ≤ y ∧ q b (x , y) | x , y are integers}·
There is no clause defining q b . By Corollary 1, b, i.e., q1 ≤ q2 is a termination condition.
This is the one we denoted cond1 . The termination condition for q a , denoted cond2 , is
initialised to be q1 > q2 . The symbolic counterpart of a natural level mapping is
X − Y if X > Y
| q a (X , Y ) | = Wq1 >q2 ∗
0
otherwise
The set of constraints Σ implied by rigid acceptability is:
Wq1 >q2 (X − Y ) > Wq1 >q2 ((X − Y ) − Y ),
(2)
that is Wq1 >q2 Y > 0 should hold. Since Wq1 >q2 ≥ 0, Y > 0 and Wq1 >q2 > 0 should hold.
Variable Y appears in the head of the clause, i.e., Y > 0 can be viewed as a constraint
on the query. We update cond2 to be (q1 > q2 ) ∧ (q2 > 0) and report termination for q1 ≤
q2 ∨ (q1 > q2 ∧ q2 > 0).
✷
Formally the following theorem holds.
Theorem 4
Let P be a homogeneous pure logical program with integer computation, let S be a single
predicate set of atomic queries and let Algo be the algorithm presented in Figure 1. Then
the following holds:
• Algo(P , S ) terminates;
• Let c be a symbolic condition returned by Algo(P , S ). Then c is a termination condition for S .
Proof
• Termination of Algo(P , S ) follows from termination of its steps. Termination of
steps 1 and 2 follows from the presentation of these transformations in Sections 4.1
and 3.2, respectively. Termination of steps 3–7 is obvious.
• Partial correctness follows from the correctness of transformations and the corresponding result of (Decorte et al. 1999). Correctness of step 1 is established by Lemma 2,
of step 2 by Theorem 3. For step 4 observe that termination for queries in S A \S ′ is
obvious by choice of A. Correctness of steps 6 and 7 follows from the corresponding
result of (Decorte et al. 1999).
In Example 17 the termination condition inferred by our algorithm was optimal, i.e., any
other termination condition implies it. However, undecidability of the termination problem
implies that no automatic tool can always guarantee optimality.
Inference of termination conditions
21
Example 18
Consider the following program.
q(X , Y ) ← X > Y , Z is X − Y , Y 1 is Y + 1, q(Z , Y 1)·
We would like to study termination of this program with respect to {q(z1 , z2 ) | z1 , z2 are
integers}. Our algorithm infers the following termination condition: q1 ≤ q2 ∨ (q1 > q2 ∧
q2 ≥ 0). This is a correct termination condition, but it is not optimal as q(z1 , z2 ) terminates,
in fact, for all values of z1 and z2 , i.e., the optimal termination condition is true.
✷
5 Experimental evaluation
The algorithm presented in Figure 1 was integrated in the system implementing the constraintbased approach of (Decorte et al. 1999). As a preliminary step of our analysis, given a
program and a set of atomic queries, the call set has to be computed. To do so, the type inference technique of Janssens and Bruynooghe (Janssens and Bruynooghe 1992) was used.
We opted for a very simple type inference technique that provides us only with information
whether some argument is integer or not. More refined analysis can be used. For instance,
the technique presented in (Janssens et al. 1994) would have allowed us to know whether
some numerical argument belongs to a certain interval. Alternatively, the integer intervals
domain fo Cousot and Cousot (Cousot and Cousot 1976; Cousot and Cousot 1977) might
have been used.
We have tested our system on a number of examples. First, we considered examples
from two textbooks chapters dedicated to programming with arithmetic, namely, Chapter
8 of Sterling and Shapiro (Sterling and Shapiro 1994) and Chapter 9 of Apt (Apt 1997).
These results are summarised in Tables 1 and 2, respectively. We can prove termination
of all the examples presented for all possible values of the integer arguments, that is, the
termination condition inferred is true. Next, we’ve collected a number of programs from
different sources. Table 3 presents timings and results for these programs. Again, termination of almost all programs can be shown for all possible values of the integer arguments.
We believe that the reason for this is that most textbooks authors prefer to write programs
ensuring termination. Finally, Table 4 demonstrates some of the termination conditions inferred by our system. We can summarise our results by saying that the system turned out
to be powerful enough to analyse correctly a broad spectrum of programs, while the time
spent on the analysis never exceeded 0.20 seconds. In fact, for 90% of the programs results
were obtained in 0.10 seconds or less.
The core part of the implementation was done in SICStus Prolog (SICS 2002), type inference of Janssens and Bruynooghe (Janssens and Bruynooghe 1992) was implemented in
MasterProLog (IT Masters 2000). Tests were performed on SUN SPARC Ultra-60, model
2360. The SPECint 95 and SPECfp 95 ratings for this machine are 16.10 and 29.50, respectively.
In Tables 1–4 the following abbreviations are used:
• Ref: reference to the program;
• Name: name of the program;
• Queries: single predicate set of atomic queries of interest, where the arguments are
denoted
22
Alexander Serebrenik and Danny De Schreye
Table 1. Examples of Sterling and Shapiro
Ref
Queries
Time
8.1
8.2
8.3
8.4
8.5
8.6a
8.6b
8.7a
8.7b
8.8
8.9
8.10
8.11
8.12
greatest common divisor(i , i , v )
factorial(i , v )
factorial(i , v )
factorial(i , v )
between(i , i , v )
sumlist(li , v )
sumlist(li , v )
inner product(li , li , v )
inner product(li , li , v )
area(lp, v )
maxlist(li , v )
length(v , li )
length(li , v )
range(i , i , v )
0.03
0.02
0.03
0.03
0.03
0.00
0.00
0.00
0.01
0.03
0.02
0.01
0.01
0.03
Table 2. Examples of Apt
—
—
—
—
—
—
Name
Queries
Time
between
delete
factorial
in tree
insert
length1
maximum
ordered
quicksort
quicksort acc
quicksort dl
search tree
tree minimum
between(i , i , v )
delete (i , i , v )
fact(i , v )
in tree(i , t)
insert(i , t, v )
length(li , v )
maximum(li , v )
ordered(li )
qs(li , v )
qs acc(li , v )
qs dl(li , v )
is search tree(t)
minimum(t, v )
0.02
0.04
0.01
0.01
0.01
0.00
0.00
0.01
0.10
0.10
0.13
0.06
0.01
c, if the argument is a character;
i, if the argument is an integer;
li, if the argument is a list of integers;
lp, if the argument is a list of pairs of integers;
t , if the argument is a binary tree, containing integers;
v , if the argument is a variable;
• Time: time (in seconds) needed to analyse the example;
Inference of termination conditions
23
Table 3. Various examples
Name
Ref
Queries
Time
T
dldf
exp
fib
fib
forwardfib
money
(Bratko 1986)
(Coelho and Cotta 1988)
(Bueno et al. 1994)
(O’Keefe 1990)
(Bratko 1986)
(Bueno et al. 1994)
0.03
0.07
0.16
0.05
0.02
0.20
T
N+
T
T∗
T
T
oscillate
p32
p33
p34
primes
pythag
r
triangle
Example 4
(Hett 2001)
(Hett 2001)
(Hett 2001)
(Clocksin and Mellish 1981)
(Clocksin and Mellish 1981)
(Dershowitz et al. 2001)
(McDonald and Yazdani 1990)
depthfirst2(c, v , i )
exp(i , i , v )
fib(i , v )
fib(i , v )
fib3(i , v )
money(v , v , v , v ,
v, v, v, v)
p(i )
gcd(i , i , v )
coprime(i , i )
totient phi(i , v )
primes(i , v )
pythag(v , v , v )
r(i , v )
triangle(i , v )
0.07
0.03
0.05
0.14
0.08
0.05
0.01
0.03
T
T
T
T
T
N+
T
N+
Table 4. Examples of inferring termination conditions
Name
Ref
Queries
Time
Condition
q
q
gcd
Example 17
Example 18
(Bratko 1986)
q(i , i )
q(i , i )
gcd(i ,i ,v )
0.04
0.05
0.10
q1 ≤ q2 ∨ (q1 > q2 ∧ q2 > 0)
q1 ≤ q2 ∨ (q1 > q2 ∧ q2 ≥ 0)
q1 = q2 ∨ (q1 > q2 ∧ q2 ≥ 1)
• T: termination behaviour, see further.
One of the less expected results was finding non-terminating examples in Prolog textbooks. The first one, due to Coelho and Cotta (Coelho and Cotta 1988), should compute
an nth power of a number.
exp(X , 0, 1)·
exp(X , Y , Z ) ← even(Y ), R is Y /2, P is X ∗ X , exp(P , R, Z )·
exp(X , Y , Z ) ← T is Y − 1, exp(X , T , Z 1), Z is Z 1 ∗ X ·
even(Y ) ← R is Y mod 2, R = 0·
The termination condition inferred by our system is false and indeed, this is the only termination condition possible, since for any goal G the LD-tree of this program and G is
infinite. This fact is denoted in Table 3 as N+. Similarly, the fact that for any goal G the
LD-tree of the program and G is finite and our system is powerful enough to discover this
is denoted as T.
24
Alexander Serebrenik and Danny De Schreye
McDonald and Yazdani suggest the following exercise in their book (McDonald and Yazdani 1990):
write a predicate triangle which finds the number of balls in a triangle of base N . For example, for N = 4 the number of balls is 4 + 3 + 2 + 1 = 10. The next program is the solution
provided by the authors:
triangle(1, 1)·
triangle(N , S ) ← M is N − 1, triangle(M , R), S is M + R·
Once more, the termination condition inferred by our system is false, and it is the only
possible one.
O’Keefe (O’Keefe 1990) suggested a more efficient way of calculating Fibonacci numbers performing O (n) work each time it is called, unlike the version of (Bueno et al. 1994)
that performs an exponential amount of work each time.
fib(1, X ) ←!, X = 1·
fib(2, X ) ←!, X = 1·
fib(N , X ) ← N > 2, fib(2, N , 1, 1, X )·
fib(N , N , X 2, , X ) ←!, X = X 2·
fib(N0 , N , X 2, X 1, X ) ← N 1 is N0 + 1,
X 3 is X 2 + X 1, fib(N1 , N , X 3, X 2, X )·
Termination of goals of the form fib(i,v) with respect to this example depends on the cut
in the first clause of fib/5. If it is removed and if we add at the beginning of the second
clause N0 6= N termination can be proved. This fact is denoted T∗ in Table 3. Note that
this replacement does not affect complexity of the computation. Observe also that a more
declarative way to write the program might be to add N0 < N instead of N0 6= N . However,
while the latter condition can be inferred automatically from the program, it is not clear
whether this is also the case for the former one.
6 Further extensions
In this section we discuss possible extensions of the algorithm presented in Section 4. First
of all, we discuss integrating termination analysis of numerical and symbolic computations,
and then show how our results can be used to improve existing termination analyses of
symbolic computations, such as (Mesnard 1996; Codish and Taboch 1999).
6.1 Integrating numerical and symbolic computation
In the real-world programs numerical computations are sometimes interleaved with symbolic ones, as illustrated by the following example collecting leaves of a tree with a variable
branching factor, being a common data structure in natural language processing (Pollard and Sag 1994).
Example 19
collect(X , [X |L], L) ← atomic(X ) ·
Inference of termination conditions
collect(T , L0, L) ← compound(T ), functor(T , , A),
25
(3)
process(T , 0, A, L0, L) ·
process( , A, A, L, L) ·
process(T , I , A, L0, L2) ← I < A, I 1 is I + 1, arg(I 1, T , Arg),
(4)
collect(Arg, L0, L1), process(T , I 1, A, L1, L2)·
To prove termination of {collect(t, v, [])}, where t is a tree and v is a variable, three decreases should be shown: between a call to collect and a call to process in (3), between a
call to process and a call to collect in (4) and between two calls to process in (4) . The first
two can be shown only by a symbolic level mapping, the third one—only by the numerical
approach.
✷
Thus, our goal is to combine the existing symbolic approaches with the numerical one
presented so far. One of the possible ways to do so is to combine two level mappings,
| · |1 and | · |2 by mapping each atom A ∈ BPE either to a natural number | A |1 or to a
pair of natural numbers (| A |1 , | A |2 ) and prove termination by establishing decreases via
orderings on (N ∪ N 2 ) as suggested in (Serebrenik and De Schreye 2001).
Example 20
Example 19, continued. Define ϕ : BPE → (N ∪ N 2 ) as follows: ϕ(collect(t , l 0, l )) = kt k,
ϕ(process(t , i, a, l 0, l )) = (kt k, a − i) where k · k is a term-size norm. The decreases are
satisfied with respect to >, such that A1 > A2 if and only if ϕ(A1 ) ≻ ϕ(A2 ), where ≻ is
defined as: n ≻ m, if n >N m, n ≻ (n, m), if true, (n, m1 ) ≻ (n, m2 ), if m1 >N m2 and
(n1 , m) ≻ n2 , if n1 >N n2 and >N is the usual order on the naturals.
✷
This integrated approach allows one to analyse correctly examples such as ground, unify,
numbervars (Sterling and Shapiro 1994) and Example 6.12 in (Dershowitz et al. 2001).
6.2 Termination of symbolic computations—revised
A number of modern approaches to termination analysis of logic programs (Codish and Taboch 1999;
Mesnard 1996; Mesnard et al. 2002) abstract a program to CLP(N) and then infer termination of the original program from the corresponding property of the abstract one. However,
as mentioned in the introduction, techniques used to prove termination of the numerical
program are often restricted to the identity function as the level-mapping.
Example 21
Consider the following example:
p(X ) ← append(X , , [ , , , , , , , ]), p([ |X ]) ·
append([], L, L) ·
append([H |X ], Y , [H |Z ]) ← append(X , Y , Z )·
Using the list-length norm, defined as
1 + kt ′k if t = [h|t ′ ]
kt k =
0
otherwise
26
Alexander Serebrenik and Danny De Schreye
the following CLP(N )-abstraction can be computed:
p(X ) ← append(X , , 7), p(1 + X ) ·
append(0, L, L) ·
append(1 + X , Y , 1 + Z ) ← append(X , Y , Z )·
Computing a model for the abstraction of append and transforming the clause for p as
described by Mesnard (Mesnard 1996) the following program is obtained:
p(X ) ← X ≤ 7, p(1 + X )·
Termination of this program cannot be shown by the identity function as a level-mapping.
Thus, non-termination will be suspected.
✷
Our approach is able to bridge the gap and provide the correct analysis of Example 21.
Since our results have been stated for numerical computations and not for CLP(N ) minor
changes in the abstraction process are required. Instead of replacing a term t in an atom a
with the size of t , a fresh variable V is introduced. Then, we add a goal V is size(t ) before
a (if a is a body subgoal) or after a (if it is a head of the clause). Next, we replace t in a
by V , and proceed with the transformation of (Mesnard 1996).
Example 22
Example 21, continued. Applying the abstraction technique above with respect to the listlength norm the following program is obtained:
p(X ) ← append(X , , 7), X 1 is X + 1, p(X 1) ·
append(0, L, L) ·
append(X 1, Y , Z 1) ← X 1 is X + 1, Z 1 is Z + 1, append(X , Y , Z )·
After computing the models this program is transformed to:
p(X ) ← X ≤ 7, X 1 is X + 1, p(X 1)·
Termination of p(n) with respect to this program can be shown by our approach for any
integer number n. This implies termination of p(t ) with respect to the original one for any
list of finite length t .
✷
To summarise this discussion, we believe that integrating our technique for proving termination of numerical computations with CLP(N ) abstracting methodologies of (Codish and Taboch 1999;
Mesnard 1996; Mesnard et al. 2002) will significantly extend the class of logic programs
that can be analysed automatically.
7 Conclusion
We have presented an approach to verification of termination for logic programs with integer computations. This functionality is lacking in current available termination analysers
for Prolog, such as cTI (Mesnard 1996; Mesnard and Neumerkel 2001), TerminWeb (Codish and Taboch 1999),
and TermiLog (Lindenstrauss and Sagiv 1997; Lindenstrauss et al. 1997). The main contribution of this work is threefold. First, from the theoretical perspective, our study improves
Inference of termination conditions
27
the understanding of termination of numerical computations, situates them in the wellknown framework of acceptability and allows integration with the existing approaches to
termination of symbolic computations. Moreover, our technique can be used to strengthen
the existing techniques for proving termination of symbolic computations.
Second, unlike the majority of works on termination analysis for logic programs concerned with termination verification, we go further and do inference, i.e., we infer conditions on integer arguments of the queries that imply termination. To perform the inference
task we apply a methodology inspired by the constraints based approach (Decorte et al. 1999),
i.e., we start by symbolic counterparts of level mappings and interargument relations and
infer constraints on the integer arguments from rigid acceptability condition and validity
of interargument relations.
Finally, the methodology presented has been integrated in the automatic termination
analyser of (Decorte et al. 1999). It was shown that our approach is robust enough to prove
termination for a wide range of numerical examples, including gcd and mod (Dershowitz et al. 2001)
all examples appearing in Chapter 8 of (Sterling and Shapiro 1994) and those appearing
in (Apt 1997).
Termination of numerical computations was studied by a number of authors (Apt 1997;
Apt et al. 1994; Dershowitz et al. 2001). Apt et al. (Apt et al. 1994) provided a declarative
semantics, so called Θ-semantics, for Prolog programs with first-order built-in predicates,
including arithmetic operations. In this framework the property of strong termination, i.e.,
finiteness of all LD-trees for all possible goals, was completely characterised based on
an appropriately tuned notion of acceptability. This approach provides important theoretical results, but seems to be difficult to integrate in automatic tools. In (Apt 1997) it is
claimed that an unchanged acceptability condition can be applied to programs in pure Prolog with arithmetic by defining the level mappings on ground atoms with the arithmetic
relation to be zero. This approach ignores the actual computation, and thus, its applicability is restricted to programs using arithmetic but whose termination behaviour is not
dependent on their arithmetic part, such as quicksort. Moreover, there are many programs
that terminate only for some queries, such as Example 17. Alternatively, Dershowitz et
al. (Dershowitz et al. 2001) extended the query-mapping pairs formalism of (Lindenstrauss and Sagiv 1997)
to deal with numerical computations. However, this approach inherited the disadvantages
of (Lindenstrauss and Sagiv 1997), such as high computational price, inherent to this approach due to repetitive fixpoint computations. Moreover, since our approach gains its
power from the underlying framework of (Decorte et al. 1999), it allows one to prove termination of some examples that cannot be analysed correctly by (Dershowitz et al. 2001),
similar to confused delete (Decorte et al. 1999).
More research has been done on termination analysis for constraint logic programming (Colussi et al. 1995; Mesnard 1996; Ruggieri 1997). Since numerical computations
in Prolog should be written in a way that allows a system to verify their satisfiability we can
see numerical computations of Prolog as an ideal constraint system. Thus, all the results
obtained for ideal constraints systems can be applied. Unfortunately, the research was either oriented towards theoretical characterisations (Ruggieri 1997) or restricted to domains
isomorphic to N (Mesnard 1996), such as trees and terms.
In a contrast to the approach of (Dershowitz et al. 2001) that was restricted to verifying
28
Alexander Serebrenik and Danny De Schreye
termination, we presented a methodology for inferring termination conditions. It is not
clear whether and how (Dershowitz et al. 2001) can be extended to infer such conditions.
Numerical computations have been also analysed in the early works on termination
analysis for imperative languages (Floyd 1967; Katz and Manna 1975), considering, as we
have already pointed out in Section 1, general well-founded domains. However, our approach to automation differs significantly from these works. Traditionally, the verification
community considered automatic generation of invariants (Bjørner et al. 1997), while automatic generation of ranking functions (level mappings, in the logic programming parlance)
just started to emerge (Colón and Sipma 2001; Colón and Sipma 2002). The inherent restriction of the latter results is that ranking functions have to be linear. Moreover, in order
to perform the analysis of larger programs, such as mergesort, in a reasonable amount of
time, authors further restricted the ranking functions to depend on one variable only. Unlike
these results, our approach doesn’t suffer from such limitations.
The idea of splitting a predicate into cases was first mentioned by Ullman and Van
Gelder (Ullman and Van Gelder 1988), where existence has been assumed of a preprocessor that transformed a set of clauses to the new set, in which every subgoal unifies with all
of the rules for its predicate symbol. However, neither in this paper, nor in the subsequent
one ((Sohn and Van Gelder 1991)) the methodology proposed was presented formally. To
the best of our knowledge the first formal presentation of splitting in the framework of
termination analysis is due to Lindenstrauss et al. (Lindenstrauss et al. 1998). Unlike these
results, a numerical and not a symbolic domain was considered in the current paper.
The termination condition inferred for Example 17 is optimal, i.e., it is implied by any
other termination condition. Clearly, undecidability of the termination problem implies
that no automatic tool can always guarantee optimality of the condition inferred. However, verifying if the condition inferred is optimal seems to be an interesting question, related to looping analysis (Bol 1991; De Schreye et al. 1990; Shen 1997; Shen et al. 2001;
Skordev 1997). So far, in the context of logic programming, optimality of termination conditions inferred has been studied by Mesnard et al. (Mesnard et al. 2002) only for symbolic
computations.
8 Acknowledgement
Alexander Serebrenik is supported by GOA: “LP + : a second generation logic programming language”. We are very grateful to Gerda Janssens and Vincent Englebert for making
their type analysis systems available for us. Many useful suggestions and helpful comments were proposed to us by anonymous referees—we are much obliged to their careful
reading.
References
A PT, K. R. 1997. From Logic Programming to Prolog. Prentice-Hall International Series in Computer Science. Prentice Hall.
A PT, K. R., M ARCHIORI , E., AND PALAMIDESSI , C. 1994. A declarative approach for first-order
built-in’s in Prolog. Applicable Algebra in Engineering, Communication and Computation 5, 3/4,
159–191.
Inference of termination conditions
29
B JØRNER , N., B ROWNE , A., AND M ANNA , Z. 1997. Automatic generation of invariants and intermediate assertions. Theoretical Computer Science 173, 1 (February), 49–87.
B OL , R. N. 1991. Loop checking in logic programming. Ph.D. thesis, Universiteit van Amsterdam.
B OSSI , A. AND C OCCO , N. 1994. Preserving universal temination through unfold/fold. In Algebraic
and Logic Programming, G. Levi and M. Rodrı́guez-Artalejo, Eds. Lecture Notes in Computer
Science, vol. 850. Springer Verlag, 269–286.
B OSSI , A., C OCCO , N., E TALLE , S., AND ROSSI , S. 2002. On modular termination proofs of
general logic programs. Theory and Practice of Logic Programming 2, 3, 263–291.
B OSSI , A., C OCCO , N., AND FABRIS , M. 1991. Proving Termination of Logic Programs by Exploiting Term Properties. In Proceedings of CCPSD-TAPSOFT’91. Lecture Notes in Computer
Science, vol. 494. Springer Verlag, 153–180.
B OSSI , A., C OCCO , N., AND FABRIS , M. 1994. Norms on terms and their use in proving universal
termination of a logic program. Theoretical Computer Science 124, 2 (February), 297–328.
B RATKO , I. 1986. Prolog programming for Artificial Intelligence. Addison-Wesley.
B RUYNOOGHE , M., C ODISH , M., G ENAIM , S., AND VANHOOF, W. 2002. Reuse of results in
termination analysis of typed logic programs. In Static Analysis, 9th International Symposium,
M. V. Hermenegildo and G. Puebla, Eds. Lecture Notes in Computer Science, vol. 2477. Springer
Verlag, 477–492.
B UENO , F., G ARC ÍA DE LA BANDA , M. J., AND H ERMENEGILDO , M. V. 1994. Effectiveness
of global analysis in strict independence-based automatic parallelization. In Logic Programming,
Proceedings of the 1994 International Symposium, M. Bruynooghe, Ed. MIT Press, 320–336.
C LOCKSIN , W. F. AND M ELLISH , C. S. 1981. Programming in Prolog. Springer Verlag.
C ODISH , M. AND TABOCH , C. 1999. A semantic basis for termination analysis of logic programs.
Journal of Logic Programming 41, 1, 103–123.
C OELHO , H. AND C OTTA , J. C. 1988. Prolog by example. Springer Verlag.
C OL ÓN , M. A. AND S IPMA , H. B. 2001. Synthesis of linear ranking functions. In Tools and Algorithms for the Construction and Analysis of Systems, 7th International Conference, T. Margaria
and W. Yi, Eds. Lecture Notes in Computer Science, vol. 2031. Springer Verlag, 67–81.
C OL ÓN , M. A. AND S IPMA , H. B. 2002. Practical methods for proving program termination. In Computer Aided Verification, 14th International Conference, E. Brinksma and K. Guldstrand Larsen, Eds. Lecture Notes in Computer Science, vol. 2404. Springer Verlag, 442–454.
C OLUSSI , L., M ARCHIORI , E., AND M ARCHIORI , M. 1995. On termination of constraint logic
programs. In Principles and Practice of Constraint Programming - CP’95,, U. Montanari and
F. Rossi, Eds. Lecture Notes in Computer Science, vol. 976. Springer Verlag, 431–448.
C OUSOT, P. AND C OUSOT, R. 1976. Static determination of dynamic properties of programs. In
Proceedings of the Second International Symposium on Programming. Dunod, Paris, France, 106–
130.
C OUSOT, P. AND C OUSOT, R. 1977. Abstract interpretation: a unified lattice model for static analysis of programs by construction or approximation of fixpoints. In Conference Record of the Fourth
Annual ACM SIGPLAN-SIGACT Symposium on Principles of Programming Languages. ACM
Press, New York, NY, Los Angeles, California, 238–252.
D E S CHREYE , D., V ERSCHAETSE , K., AND B RUYNOOGHE , M. 1990. A practical technique for
detecting non-terminating queries for a restricted class of Horn clauses, using directed, weighted
graphs. In Logic Programming, Proceedings of the Seventh International Conference, D. H. Warren and P. Szeredi, Eds. MIT Press, 649–663.
D E S CHREYE , D., V ERSCHAETSE , K., AND B RUYNOOGHE , M. 1992. A framework for analyzing
the termination of definite logic programs with respect to call patterns. In Proceedings of the
International Conference on Fifth Generation Computer Systems., I. Staff, Ed. IOS Press, 481–
488.
30
Alexander Serebrenik and Danny De Schreye
D ECORTE , S. AND D E S CHREYE , D. 1998. Termination analysis: some practical properties of the
norm and level mapping space. In Proceedings of the 1998 Joint International Conference and
Symposium on Logic Programming, J. Jaffar, Ed. MIT Press, 235–249.
D ECORTE , S., D E S CHREYE , D., AND VANDECASTEELE , H. 1999. Constraint-based termination
analysis of logic programs. ACM TOPLAS 21, 6 (November), 1137–1195.
D ERSHOWITZ , N., L INDENSTRAUSS , N., S AGIV, Y., AND S EREBRENIK , A. 2001. A general
framework for automatic termination analysis of logic programs. Applicable Algebra in Engineering, Communication and Computing 12, 1-2, 117–156.
F LOYD , R. W. 1967. Assigning meanings to programs. In Mathematical Aspects of Computer
Science, J. Schwartz, Ed. American Mathematical Society, 19–32. Proceedings of Symposiumsia
in Applied Mathematics; v. 19.
G ENAIM , S. AND C ODISH , M. 2001. Inferring termination conditions for logic programs using
backwards analysis. In Logic for Programming, Artificial Intelligence, and Reasoning, 8th International Conferencerence, Proceedings, R. Nieuwenhuis and A. Voronkov, Eds. Lecture Notes in
Computer Science, vol. 2250. Springer Verlag, 685–694.
G ENAIM , S., C ODISH , M., G ALLAGHER , J., AND L AGOON , V. 2002. Combining norms to prove
termination. In Third International Workshop on Verification, Model Checking and Abstract Interpretation, A. Cortesi, Ed. Lecture Notes in Computer Science, vol. 2294. Springer Verlag, 126–
138.
H ETT, W. 2001.
P-99: Ninety-nine Prolog problems.
http://www.hta-bi.bfh.ch/˜hew/informatik3/prolog/p-99/.
Available
at
H OLZBAUR , C. 1995. OFAI CLP(Q,R) Manual. Tech. Rep. TR-95-09, Austrian Research Institute
for Artificial Intelligence, Vienna.
ILOG. 2001. ILOG Solver 5.1 User’s Manual. ILOG s.a. http://www.ilog.com.
IT M ASTERS. 2000.
MasterProLog
http://www.itmasters.com/.
Programming
Environment.
Available
at
JANSSENS , G. AND B RUYNOOGHE , M. 1992. Deriving descriptions of possible values of program
variables by means of abstract interpretation. Journal of Logic Programming 13, 2&3 (July),
205–258.
JANSSENS , G., B RUYNOOGHE , M., AND E NGLEBERT, V. 1994. Abstracting numerical values in
CLP(H, N). In Programming Language Implementation and Logic Programming, 6th International
Symposiumsium, PLILP’94, M. V. Hermenegildo and J. Penjam, Eds. Lecture Notes in Computer
Science, vol. 844. Springer Verlag, 400–414.
K ATZ , S. AND M ANNA , Z. 1975. A closer look at termination. Acta Informatica 5, 333–352.
L INDENSTRAUSS , N. AND S AGIV, Y. 1997. Automatic termination analysis of logic programs.
In Proceedings of the Fourteenth International Conference on Logic Programming, L. Naish, Ed.
MIT Press, 63–77.
L INDENSTRAUSS , N., S AGIV, Y., AND S EREBRENIK , A. 1997. TermiLog : A system for checking termination of queries to logic programs. In Computer Aided Verification, 9th International
Conference, O. Grumberg, Ed. Lecture Notes in Computer Science, vol. 1254. Springer Verlag,
63–77.
L INDENSTRAUSS , N., S AGIV, Y., AND S EREBRENIK , A. 1998. Unfolding the mystery of mergesort . In Proceedings of the 7th International Workshop on Logic Program Synthesis and Transformation, N. Fuchs, Ed. Lecture Notes in Computer Science, vol. 1463. Springer Verlag.
M C D ONALD , C. AND YAZDANI , M. 1990. Prolog programming: a tutorial introduction. Artificial
Intelligence Texts. Blackwell Scientific Publications.
M ESNARD , F. 1996. Inferring left-terminating classes of queries for constraint logic programs. In
Proceedings of the 1996 Joint International Conference and Syposium on Logic Programming,
M. Maher, Ed. The MIT Press, Cambridge, MA, USA, 7–21.
Inference of termination conditions
31
M ESNARD , F. AND N EUMERKEL , U. 2001. Applying static analysis techniques for inferring termination conditions of logic programs. In Static Analysis, 8th International Symposium, SAS 2001,
P. Cousot, Ed. Lecture Notes in Computer Science, vol. 2126. Springer Verlag, 93–110.
M ESNARD , F., PAYET, E., AND N EUMERKEL , U. 2002. Dectecting optimal termination conditions
of logic programs. In Static Analysis, 8th International Symposium, SAS 2002, P. Cousot, Ed.
Lecture Notes in Computer Science, vol. 2477. Springer Verlag, 509–527.
M ESNARD , F. AND RUGGIERI , S. 2003. On proving left termination of constraint logic programs.
ACM Transaction on Computational Logic 4, 2, 207–259.
O HLEBUSCH , E. 2001. Automatic termination proofs of logic programs via rewrite systems. Applicable Algebra in Engineering, Communication and Computing 12, 1-2, 73–116.
O’K EEFE , R. A. 1990. The Craft of Prolog. MIT Press, Cambridge, MA, USA.
P L ÜMER , L. 1990. Termination proofs for logic programs based on predicate inequalities. In Proceedings of ICLP’90. MIT Press, 634–648.
P L ÜMER , L. 1991. Automatic termination proofs for Prolog programs operating on nonground
terms. In International Logic Programming Symposium. MIT Press.
P OLLARD , C. AND S AG , I. A. 1994. Head-driven Phrase Structure Grammar. The University of
Chicago Press.
RUGGIERI , S. 1997. Termination of constraint logic programs. In Automata, Languages and Programming, 24th International Colloquium, ICALP’97, P. Degano, R. Gorrieri, and A. MarchettiSpaccamela, Eds. Lecture Notes in Computer Science, vol. 1256. Springer Verlag, 838–848.
S EREBRENIK , A. AND D E S CHREYE , D. 2001. Non-transformational termination analysis of logic
programs, based on general term-orderings. In Logic Based Program Synthesis and Transformation 10th International Workshop, Selected Papers, K.-K. Lau, Ed. Lecture Notes in Computer
Science, vol. 2042. Springer Verlag, 69–85.
S EREBRENIK , A. AND D E S CHREYE , D. 2002. On termination of logic programs with floating
point computations. In 9th International Static Analysis Symposium, M. V. Hermenegildo and
G. Puebla, Eds. Lecture Notes in Computer Science, vol. 2477. Springer Verlag, 151–164.
S HEN , Y.-D. 1997. An extended variant of atoms loop check for positive logic programs. New
Generation Computing 15, 2, 187–204.
S HEN , Y.-D., Y UAN , L.-Y., AND YOU , J.-H. 2001. Loop checks for logic programs with functions.
Theoretical Computer Science 266, 1–2, 441–461.
SICS. 2002. SICStus User Manual. Version 3.10.0. Swedish Institute of Computer Science.
S KORDEV, D. 1997. An abstract approach to some loop detection problems. Fundamenta Informaticae 31, 2, 195–212.
S OHN , K. AND VAN G ELDER , A. 1991. Termination detection in logic programs using argument
sizes. In Proceedings of the Tenth ACM SIGACT-SIGART-SIGMOD Symposium on Principles
of Database Systems. ACM Press, 216–226.
S TERLING , L. AND S HAPIRO , E. 1994. The Art of Prolog. The MIT Press, Cambridge, MA, USA.
U LLMAN , J. D. AND VAN G ELDER , A. 1988. Efficient tests for top-down termination of logical
rules. Journal of the ACM 35, 2 (April), 345–373.
V ERBAETEN , S., S AGONAS , K., AND D E S CHREYE , D. 2001. Termination proofs for logic programs with tabling. ACM Transactions on Computational Logic 2, 1, 57–92.
V ERSCHAETSE , K. AND D E S CHREYE , D. 1991. Deriving termination proofs for logic programs,
using abstract procedures. In Logic Programming, Proceedings of the Eigth International Conference, K. Furukawa, Ed. MIT Press, 301–315.
W INSBOROUGH , W. 1992. Multiple specialization using minimal-function graph semantics. Journal
of Logic Programming 13, 2/3, 259–290.
| 2cs.AI
|
1
Adaptive Channel Prediction, Beamforming and
Scheduling Design for 5G V2I Network
arXiv:1707.07068v1 [cs.IT] 21 Jul 2017
Tadilo Endeshaw Bogale+ , Xianbin Wang++ and Long Bao Le+
Institute National de la Recherche Scientifique (INRS)
Université du Québec, Montréal, Canada+
Western University, London, Canada++
Email: {tadilo.bogale, long.le}@emt.inrs.ca, [email protected]
Abstract— One of the important use-cases of 5G network is
the vehicle to infrastructure (V2I) communication which requires
accurate understanding about its dynamic propagation environment. As 5G base stations (BSs) tend to have multiple antennas,
they will likely employ beamforming to steer their radiation
pattern to the desired vehicle equipment (VE). Furthermore, since
most wireless standards employ an OFDM system, each VE may
use one or more sub-carriers. To this end, this paper proposes
a joint design of adaptive channel prediction, beamforming and
scheduling for 5G V2I communications. The channel prediction
algorithm is designed without the training signal and channel
impulse response (CIR) model. In this regard, first we utilize the
well known adaptive recursive least squares (RLS) technique for
predicting the next block CIR from the past and current block
received signals (a block may have one or more OFDM symbols).
Then, we jointly design the beamforming and VE scheduling for
each sub-carrier to maximize the uplink channel average sum
rate by utilizing the predicted CIR. The beamforming problem
is formulated as a Rayleigh quotient optimization where its global
optimal solution is guaranteed. And, the VE scheduling design is
formulated as an integer programming problem which is solved
by employing a greedy search. The superiority of the proposed
channel prediction and scheduling algorithms over those of the
existing ones is demonstrated via numerical simulations.
Index Terms— Channel prediction, Beamforming, Scheduling,
Recursive least squares (RLS), Adaptive filtering.
I. I NTRODUCTION
With the proliferation of mobile internet and smart devices,
it is clear that data traffic over wireless networks will continue
to grow exponentially in foreseeable future. Driven by many
emerging use cases and industry applications, intensive efforts
have been undertaken by both academia and industry to design
the next generation 5G network. One of the important sectors
that uses the 5G network infrastructure is the automotive
industry which is currently undergoing dramatic technological
transformations, as more and more vehicles are connected to
the Internet and with each other [1], [2]. The development of
effective vehicular communication systems in highly mobile
environment requires accurate understanding about the propagation channel. Furthermore, as 5G base stations (BSs) tend to
have large scale antennas, BSs will likely employ beamforming technology to steer their radiation pattern to the desired
vehicle equipment (VE). In addition, the current vehicle to
vehicle (V2V) communication standard (i.e., IEEE 802.11p)
employs a multicarrier system where each of the VEs may use
one or more sub-carriers. Thus, efficient and integrated design
of channel impulse response (CIR) prediction, beamforming
and scheduling is crucial for the successful realization of the
future vehicle to infrastructure (V2I) networks.
A number of efforts have been made to estimate the CIR
of vehicular communication systems. In [3], least square and
minimum mean square error (MMSE) estimation techniques
are discussed and their performances in different vehicular AdHoc network (VANET) scenarios are thoroughly investigated.
In [4], a cross-layer channel estimator for the 802.11p system
where known data are inserted at the higher layers is proposed.
The Wiener filter approach to estimate the channel transfer
function is discussed in [5]. The filter coefficients are evaluated
in IEEE 802.11p systems. In [6], channel estimation and
tracking algorithms exploiting the transmitted symbols, time
domain truncation, decision-directed feedback, pilot information and V2V channel characteristics is presented. In [2],
location aware channel state information (CSI) estimation and
beamforming is proposed for millimeter wave vehicular communications. The authors of [7] propose a channel prediction
based scheduling strategy for a single antenna system in flat
fading channels for data dissemination in VANETs. Along this
line, a scheduling method that employs successive interference
cancellation (SIC) for V2V systems is discussed in [8].
Measurement results show that the coherence time of vehicular channels is sometimes less than 1 ms due to the high
mobility of vehicles in highway scenarios [9]. Furthermore,
as the scattering environment can rapidly be changed due to
mobility, it is not trivial to find a general CIR statistical model
for vehicular channels [7], [10]. These two factors create a
significant challenge to learn the instantaneous CIR and its
statistical model using the conventional training based CIR
estimation approach. This is because the estimated CIR will
be likely outdated when it is used to transmit data in the
subsequent time slots (this is validated in Section IV). One way
of addressing this challenge is by just increasing the training
overhead. However, such an approach will dramatically reduce
the network throughput especially for a multiuser setup as the
training overhead increases with the number of VEs.
For these reasons, the current paper proposes adaptive
channel prediction, beamforming and scheduling design that
does not utilize the training signal and CIR statistical model.
In this regard, first we apply the well known adaptive recursive
least squares (RLS) technique (which does not require the
channel statistics) for predicting the next block CIR from the
past and current block received signals (a block may have one
or more orthogonal frequency division multiplexing (OFDM)
symbols) [11], [12]. Then, we jointly design the beamforming
and VE scheduling for each sub-carrier to maximize the uplink
channel average sum rate by utilizing the predicted CIR. The
receive beamforming problem is formulated as a Rayleigh
quotient problem where its global optimality is guaranteed.
And, the VE scheduling design is formulated as a non-convex
integer programming problem where we propose the greedy
search based algorithm [13]. The superiority of the proposed
channel prediction and scheduling algorithms over the existing
ones (i.e., algorithms that employ outdated CIR with a greedy
scheduling and predicted CIR with a random VE selection) is
demonstrated by numerical simulations.
This paper is organized as follows. Section II discusses the
system and channel models. In Section III, the proposed CIR
prediction, and joint beamforming and scheduling designs are
presented. Computer simulation results are provided in Section
IV. Finally, conclusions are drawn in Section V.
II. S YSTEM AND C HANNEL M ODEL
We consider a multiuser system with a transmission scheme
having a sampling period Ts and a channel having a maximum
delay spread Td . It is assumed that the BS and each VE
are equipped with N antennas and 1 antenna, respectively.
For these settings, the number of multipath CIR taps L is
approximated as L = TTds [14]. The multipath coefficients
between the kth VE to the nth BS antenna is denoted as [15]
h̄kn = [h̄kn1 , h̄kn2 , · · · , h̄knL ]T .
(1)
The analytical model for H̄k = [h̄k1 , h̄k2 , · · · , h̄kN ] may vary
from one standard to another. In this paper, we assume that
each element of h̄kn is independent of others. However, h̄knl
is correlated over time. We further assume that the spatial
correlation matrix of all multipath components are the same.
Under these assumptions, H̄k can be expressed as [16]
√
√
(2)
H̄k = Pk H̃k Rk
where each entry of H̃k is an independent and identically
distributed (i.i.d) zero mean circularly symmetric complex
Gaussian (ZMCSCG) random variable with unit variance,
Rk ∈ C N ×N is a positive semidefinite spatial covariance
2
2
2
2
matrix, and Pk = diag(gk1
, gk1
, · · · , gkL
) ∈ <L×L with gkl
being the channel gain corresponding to the lth path.
In most wireless standards (e.g., 802.11p), OFDM based
transmission is commonly adopted. For such a transmission,
the channel coefficient of each sub-carrier has practical importance. In fact, this coefficient can be obtained by linearly
combining the multipath tap coefficients (1). To this end, the
channel between the kth VE to the BS’s nth antenna in subcarrier s can be expressed as [14], [17]
hkns = fsH h̄kn
−j 2π
M s
−j 2π
M 2s
(3)
−j 2π
M (L−1)s
= [1, e
,e
,··· ,e
] with M
where
being the fast Fourier transform (FFT) size of the OFDM.
Under this model, one can consider the CIR prediction,
beamforming and scheduling designs either for the uplink or
fsH
downlink channels. Here, we consider the uplink channel, and
the sth sub-carrier kth VE signal to interference plus noise
ratio (SINR) for this channel can be expressed as1
γks = PK
i6=k
H
Pks |wks
hks |2
H h |2 + σ 2 |w |2
Pis |wks
is
ks
(4)
where hks = [hk1s , · · · , hkN s ]T , Pks (wks ) is the kth VE sth
sub-carrier transmitted power (receive beamforming vector)
and σ 2 is the noise power at the BS. We assume that H̄k
is constant at least for Sc OFDM symbols where Sc is related
to the channel coherence time2 . The ith block CIR for the kth
VE H̄k [i] can thus be given as [18]
√
√
H̄k [i] = Pk [i]H̃k [i] Rk [i].
(5)
III. CIR PREDICTION , BEAMFORMING AND SCHEDULING
DESIGN
This section discusses the proposed CIR prediction, joint
beamforming and scheduling design algorithms.
A. CIR prediction design
As stated in the introduction section, the paper assumes that
the CIR statistical model is not known.
√ proposed
√ Thus, the
approach may need to jointly predict Pk [i + 1], Rk [i + 1]
and H̃k [i+1] from the past and
√ current block
√ received signals.
Since the time scale at which Pk [i] and Rk [i] are varied is
much larger than that
√ of H̃k [i] [19],
√ [20], we predict H̃k [i+1]
by assuming that Pk [j] and
are constant for j =
√ Rk [j] √
i+1, i, · · · , i−Cb +1, where Pk [i] and Rk [i] are computed
from the Cb most recent block received signals and Cb is the
design parameter. In the following, we discuss the proposed
algorithm to predict H̃k [i + 1].
For better mathematical tractability, we denote the Eigenvalue decomposition of Rk as (we ignore here the time index
i since it is clear from the context)
svd(Rk ) , Uk Dk UH
k
(6)
where Dk is a Qk × Qk sized diagonal matrix containing the
non-zero eigenvalues of Rk arranged in decreasing order. As
Qk ≤ N , we can represent the low dimensional equivalent
CIR matrix of H̃k ∈ C L×Qk using (6) as
q
q
˜ = P−1 H̄ U D−1 .
H̃
(7)
k
k k
k
k
˜ [i] is independent, the CIR prediction
Since each element of H̃
k
˜ [i + 1] separately. We utilize
can be done for each entry of H̃
k
˜ [i+1] from H̃
˜ [i], · · · , H̃
˜ [i−
the RLS algorithm to predict H̃
k
k
k
Sb + 1] where Sb is the design parameter. The RLS algorithm
is explained as follows [12], [21].
˜ [i] as f [i] (i.e., ignoring k and
Denote an element of H̃
k
the entry indexes). The main idea of the RLS algorithm is to
predict f [i + 1] from f [i] , [f [i], · · · , f [i − Sb + 1]]T as
fˆ[i + 1] = q[i]T f [i].
(8)
1 The methodology used in this paper can be applied for the downlink
channel.
2 In the following, we use the word block to denote S consecutive OFDM
c
symbols.
And the predictor q[n] is calculated to optimize
n
X
min
q[n]
j=n−Sb +1
λn−j |f [j + 1] − q[n]T f [j]|2
(9)
where 0 < λ ≤ 1 is a forgetting factor that accounts for a
possible non stationarity of the input samples f [j]. After doing
some straightforward steps, one can compute q[n] recursively
as [12], [21]
q[n] = q[n − 1] + τ [n − 1]e[n]
(10)
where e[n] = f [n] − fˆ[n],
τ [n] =
Z[n − 1]f [n]
λ + f [n]H Z[n − 1]f [n]
(11)
and
matrix Z is the inverse of the sample covariance matrix
P the
n−j
λ
f [j]f [j]H which can be calculated recursively as
j
1
(I − τ [n]f [n]H )Z[n − 1].
(12)
λ
One can observe from (10) that q[n] depends on q[n − 1]
and τ [n − 1]. This recursive behavior requires appropriate
parameter initializations. In the current paper, we employ
λ = 0.999, Z[0] = δI and q[0] = 0 with δ = 100 [12]3 .
˜ [i + 1] are predicted, the next
Once all the elements of H̃
k
task is to perform the joint beamforming and scheduling design
which is presented in the next subsection.
Z[n] =
B. Joint beamforming and VE scheduling design
Once the CIR of each VE is predicted, the next task is to
schedule VEs to utilize one or more sub-carriers. In practice,
a scheduler is designed to optimize some performance criteria.
In the current paper, we consider the maximization of the total
sum rate achieved at each sub-carrier corresponding to the first
OFDM symbol of the (i + 1)th block under a per VE power
constraint which can be mathematically formulated as4
max
Ks ,Ks ,wks
Ks
X
k=1
p
log(1 + γks
), Pks ≤ Pmax , k ∈ Ks
(13)
p
where γks
(wks ) is the kth VE sth sub-carrier SINR (receive beamforming vector), Pmax is the maximum transmitted
power at each sub-carrier of a VE and Ks is the scheduled set
p
of VEs at sub-carrier s. Note that since γks
is computed using
the predicted CIR, this SINR could be different from the true
γks achieved at the first OFDM symbol of the (i + 1)th block.
The VE scheduling problem (13) is an integer programming
non-convex problem which prohibits us from getting the global
optimal solution. Here we propose to solve the above problem
by utilizing greedy scheduling technique where we sequentially increase the number of VEs (i.e., Ks ) while ensuring a
non-decreasing sum rate [13]. In fact, for fixed Ks , the global
3 These initializations may not necessarily yield the optimal performance.
More information about the optimal selection of λ can be found in [22].
4 Note that some applications require a minimum γ p to ensure a certain
ks
service quality. And considering the rate maximization problem by taking into
p
account the minimum γks is left for future research.
optimal solution that maximizes (4) (which also maximizes
(13)) can be obtained by applying the generalized Rayleigh
quotient approach [23]. The proposed joint beamforming and
VE scheduling algorithm is summarized in Algorithm I.
Algorithm I: Joint beamforming and scheduling algorithm.
Input: The number of VEs to be scheduled (Kt ).
Initialization: Set Kt = {1, 2, · · · , Kt }, Ks = ∅, Ks = 0
and fs (Ws ) = 0, where fs (Ws ) is the objective function
of (13) with Ks VEs and Ws = [w1s , w2s , · · · , wKs s ].
while Ks ≤ Kt do
• Set u = ∅.
for j ∈ Kt and j 6∈ Ks do
• Compute fs (W̃s ) with K̃s = Ks ∪ {j} VEs, where
W̃s is the receive beamforming for the VEs in K̃s .
if fs (W̃s ) > fs (Ws ) then
• Set u = j and fs (Ws ) = fs (W̃s ).
end if
end for
if u = ∅ then
• Break.
else
• Set Ks = Ks ∪ {u} and Ks = Ks + 1.
end if
end while
Practical Issues: Up to now, we have provided the framework for the joint adaptive channel prediction, beamforming
and scheduling algorithms. In fact, these algorithms require the
predicted CIR which is computed by employing H̄k [j], j ≤
i, ∀k. However, since more than one VEs could be scheduled to
use a given sub-carrier, H̄k [j] may be observed with additional
noise and the CIR of other VEs. In a general setup, the
received signal at the sth sub-carrier and jth OFDM symbol
can be expressed as
ys [j] = fsH H̄k [j]xks [j] +
K
s −1
X
fsH H̄i [j]xis [j] + ns [j] (14)
i=1
where xks [j] is the transmitted signal from the kth VE at the
sth sub-carrier of jth OFDM symbol, and ns [.] is the received
noise at the sth sub-carrier where each of its entries is i.i.d
ZMCSCG random variable with variance σ 2 .
One can examine (14) for two cases: Ks = 1 and Ks > 1.
In the case of Ks = 1, if the kth VE utilizes L sub-carriers
FL = [fs1 , fs2 , · · · , fsL ] that ensures FL FH
L ≈ I and xks , ∀s
are known perfectly, H̄k [j] can be estimated as [15]
ˆ [j] = H̄ [j] + n̄[j]
H̄
k
k
(15)
where the entries of n̄[.] are i.i.d ZMCSCG random variables
each with variance σ 2 . When Ks > 1, one can achieve the CIR
expression similar to (15) if there exists a set of sub-carriers
Ok such that xks , s ∈ Ok is orthogonal to xis , s ∈ Ok , ∀i 6= k.
In the current paper, we assume that the transmitted symbols xks [j], ∀k, j are recovered perfectly at the BS from the
received signal ys [j] (i.e., the predicted CSI corresponding
to the jth OFDM symbol has sufficient quality to recover
xks [j], ∀k, j without any error) and the number of OFDM
symbols per block Sc is selected such that the observed
samples corresponding to H̄k [j] is as given in (15) for all
40
MMSE
Proposed RLS
Outdated CIR
5
35
Average sum SE (bps/Hz)
Mean square error (MSE)
5.5
4.5
4
3.5
Perfect CIR
MMSE
Proposed RLS
Outdated CIR
Random scheduling
30
25
20
15
10
5
0
3
0
50
100
150
200
250
300
350
1
400
Fig. 1.
Comparison of MSEs achieved by different CIR predictors.
14
Average SE (bps/Hz)
10
8
6
4
2
0
2.5
5
7.5
10
12.5
15
17.5
VEs. In addition, the spatial covariance matrix Rk of (6) is
computed by taking into account the effect of n̄[.] like in [24].
Thus, in practice, the proposed algorithms require training
signals just at the beginning of the transmission duration or
when the BS is not able to reliably decode the transmitted
signals xks [j] which occurs quiet rarely.
IV. S IMULATION R ESULTS
This section presents simulation results. We consider a
typical scenario of IEEE 802.11p standard where the relative
powers of gkl , ∀k, l are [0 dB, −6.3 dB, −25.1 dB, −22.7 dB]
(i.e., VTV - Expressway Oncoming without Wall, 300 m - 400
m scenario) [25], [26]. The spatial channel covariance matrices for this environment is characterized for the commonly
adopted uniform linear array (ULA) model as
Rk =
5
6
7
8
Fig. 3.
Effect of the maximum number of scheduled VEs Kmax .
A. CIR prediction
Average SE achieved with single user scheduling.
Ck
X
4
20
SNR (dB)
Fig. 2.
3
where αkij = J0 (2πfk |i − j|) with J0 (.) as the zeroth-order
Bessel function and fk is the maximum Doppler frequency
normalized by the sampling rate T1s [12].
Furthermore, we set L = 4, M = 64, N = 8, Ts = 166µs,
Cb = Sb = 8, the carrier frequency as 5.6 GHz, the speed of
each VE as 90 km/hr, Rk = R, ∀k with Ck = 8 and θki , ∀i
are selected randomly. The signal to noise ratio (SNR) for
and is controlled by
each sub-carrier is defined as γ = Pmax
σ2
varying Pmax while keeping σ 2 constant. All of the plots are
generated for s = 4 by averaging 10000 channel realizations5 .
Perfect CIR
MMSE
Proposed RLS
Outdated CIR
Random scheduling
12
2
Maximum number of scheduled VEs (K max )
Block index (i)
a(θki )a(θki )H
(16)
i=1
where Ck is the number of clusters in the environment corresponding to the kth VE CIR, θki is the angle of arrival (AoA) for the ith cluster, and a(θ) =
2π
2π
√1 [1, ej λ̃ d sin (θ) , · · · , ej(N −1) λ̃ d sin (θ) ]T with λ̃ as the
N
transmission wave length and d = 0.5λ̃ is the antenna spacing
[24]. We assume the well known Jake’s model to capture the
channel correlation between H̃k [i] and H̃k [j] as
E{(H̃k [i])m,n (H̃k [j])H
m,n }
= αkij
(17)
This simulation discusses the accuracy of the predicted CIR
coefficients and convergence of the proposed RLS algorithm
when γ is set to γ = 20 dB. We compare the mean square error
(MSE) values between the true and predicted CIR coefficients
for three algorithms in Fig. 1: the MMSE predictor which
requires the CIR statistics, the proposed RLS predictor and
the design that utilizes outdated CIR. This figure shows that
employing the outdated CIR suffers from high prediction
error. Furthermore, the MSE achieved by the proposed RLS
approach decreases with i and achieves closer to that of the
MMSE approach in around 200 blocks. This confirms that the
proposed RLS predictor converges to the optimal estimator6 .
B. Joint beamforming and scheduling
This subsection presents the performance of the proposed
joint beamforming and scheduling design. In this regard, we
set the total number of VEs as Kt = 24 and Ks ≤ Kmax . We
compare the average spectrum efficiency (SE) (i.e., average
rate per Hertz) achieved by five approaches as shown in Fig.
2 when Kmax = 1: perfect CIR, MMSE predicted CIR,
proposed RLS predicted CIR and outdated CIR, all with the
proposed scheduling, and proposed RLS predicted CIR with
random scheduling. One can observe from this figure that
the average SE achieved by the proposed CIR prediction and
scheduling approach is higher than those of random scheduling
and outdated CIR ones, and its performance is closer to the
5 Note
that similar average performance is observed for other sub-carriers.
that the proposed RLS predictor cannot perform better than the
MMSE approach as the latter assumes perfect CIR statistical knowledge.
6 Note
18
Average sum SE (bps/Hz)
16
Perfect CIR
MMSE
Proposed RLS
Outdated CIR
Random scheduling
14
12
10
8
6
4
2
4
8
12
16
20
24
Total number of scheduled VEs (K t)
Fig. 4.
Effect of the total number of scheduled VEs Kt .
average SE achieved by the MMSE predictor approach which
utilizes the CIR statistics. The best average SE is achieved
when the CIR is known perfectly which is expected.
Next we examine the effect of Kmax on the performance
of the approaches when γ is set to 10 dB as shown in Fig.
3. One can observe from this figure that the sum SE degrades
with Kmax for all approaches except the perfect CIR case.
The potential reason for this is that when Kmax > 1, the
scheduling algorithm likely supports more than one VEs and
each VE will experience undesired interference which occurs
due to the channel prediction error and worsens with Kmax
7
. Lastly, we evaluate the effect of Kt on the performance of
the approaches when Smax = 2 and γ = 10 dB as shown in
Fig. 4. As can be seen from this figure, the average sum SE
achieved by all approaches increase with Kt up to a certain
limit which fits with the phenomena observed in [13].
V. C ONCLUSIONS
This paper proposes the joint design of adaptive channel prediction, beamforming and scheduling for 5G V2I network. The
channel prediction algorithm is designed without assuming the
training and CIR model. In this regard, first we utilize the well
known RLS technique, which does not require the channel
statistics, for predicting the next block CIR. Then, we jointly
design the beamforming and sub-carrier scheduling for the
uplink communication to maximize the average sum rate by
utilizing the predicted CIR. The receive beamforming problem
is formulated as a Rayleigh quotient problem where its global
optimal solution is guaranteed. And, the VE scheduling design
problem is formulated as an integer programming problem
where we propose the greedy based algorithm. Simulation
results have confirmed the superiority of the proposed channel
prediction and scheduling algorithm over the existing ones in
terms of CIR accuracy and average rate.
R EFERENCES
[1] “5G automotive vision,” 2015, ERTICO - ITS EUROPE, 5G-PPP.
[2] N. Garcia, H. Wymeersch, E. G. Strm, and D. Slock, “Location-aided
mm-wave channel estimation for vehicular communication,” in Proc.
IEEE International Workshop on Signal Processing Advances in Wireless
Communications (SPAWC), July 2016, pp. 1–5.
7 Recall that the proposed greedy algorithm schedules the VEs based on the
predicted CIRs which may be different from the true CIRs.
[3] A. M. S. Abdelgader, S. Feng, and L. Wu, “On channel estimation in
vehicular networks,” IET Communications, vol. 11, no. 1, pp. 142–149,
2017.
[4] K. K. Nagalapur, F. Brnnstrm, and E. G. Strm, “On channel estimation
for 802.11p in highly time-varying vehicular channels,” in Proc. IEEE
International Conference on Communications (ICC), June 2014, pp.
5659–5664.
[5] J. Nuckelt, M. Schack, and T. Kurner, “Performance evaluation of
Wiener filter designs for channel estimation in vehicular environments,”
in Proc. IEEE Vehicular Technology Conference (VTC Fall), Sept 2011,
pp. 1–5.
[6] M. M. Awad, K. G. Seddik, and A. Elezabi, “Channel estimation and
tracking algorithms for harsh vehicle to vehicle environments,” in Proc.
IEEE Vehicular Technology Conference (VTC-Fall), Sept 2015, pp. 1–5.
[7] F. Zeng, R. Zhang, X. Cheng, and L. Yang, “Channel prediction based
scheduling for data dissemination in VANETs,” IEEE Commun. Letters,
vol. PP, no. 99, pp. 1–1, 2017.
[8] Y. Watanabe, K. Sato, and T. Fujii, “A scheduling method for V2V
networks using successive interference cancellation,” in IEEE Vehicular
Networking Conference (VNC), Dec. 2016, pp. 1–2.
[9] R. C. Daniels and R. W. Heath, “Link adaptation with position/motion
information in vehicle-to-vehicle networks,” IEEE Tran. Wireless
Commun., vol. 11, no. 2, pp. 505–509, Feb. 2012.
[10] L. Azpilicueta, C. Vargas-Rosales, and F. Falcone, “Intelligent vehicle
communication: Deterministic propagation prediction in transportation
systems,” IEEE Vehic. Techno. Mag., vol. 11, no. 3, pp. 29–37, Sept.
2016.
[11] J. Akhtman and L. Hanzo, “Channel impulse response tap prediction
for time-varying wireless channels,” IEEE Tran. Vehic. Tech., vol. 56,
no. 5, pp. 2767–2769, Sept 2007.
[12] D. Schafhuber and G. Matz, “MMSE and adaptive prediction of timevarying channels for OFDM systems,” IEEE Tran. Wireless Commun.,
vol. 4, no. 2, pp. 593–602, March 2005.
[13] T. E. Bogale, L. B. Le, and A. Haghighat, “User scheduling for massive
MIMO OFDMA systems with hybrid analog-digital beamforming,” in
Proc. IEEE Conference on Communications (ICC), Jun. 2015.
[14] D. Tse and P. Viswanath, Fundamentals of Wireless Communication,
Cambridge University Press, 2005.
[15] T. E. Bogale, L. B. Le, X. Wang, and L. Vandendorpe, “Pilot
contamination in wideband massive MIMO system: Number of cells
vs multipath,” in Proc. IEEE Global Telecommun. Conf., 2015.
[16] F. Talebi and T. Pratt, “Channel sounding and parameter estimation for a
wideband correlation-based MIMO model,” IEEE Trans. Veh. Technol.,
pp. 1 – 9, 2015.
[17] H. Zarrinkoub, Understanding LTE with MATLAB: From Mathematical
Modeling to Simulation and Prototyping, John W&S, Ltd, 2014.
[18] X. Hou, S. Li, C. Yin, and G. Yue, “Two-dimensional recursive least
square adaptive channel estimation for OFDM systems,” in Proc.
International Conference on Wireless Communications, Networking and
Mobile Computing., Sept 2005, vol. 1, pp. 232–236.
[19] M. R. Akdeniz, Y. Liu, M. K. Samimi, S. Sun, S. Rangan, T. S.
Rappaport, and E. Erkip, “Millimeter wave channel modeling and
cellular capacity evaluation,” IEEE J. Select. Areas in Commun., vol.
32, no. 6, pp. 1164 – 1179, Jun. 2014.
[20] M. C. Vanderveen, A-J. V. Veen, and A. Paulraj, “Estimation of
multipath parameters in wireless communications,” IEEE Trans. Signal
Process., vol. 46, no. 3, pp. 682 – 690, Mar. 1998.
[21] S. Haykin, Adaptive Filter Theory, PrenticeHall, Englewood Cliffs, NJ,
USA, 1991.
[22] S. Van Vaerenbergh, I. Santamara, and M. Lzaro-Gredilla, “Estimation
of the forgetting factor in kernel recursive least squares,” in Proc. IEEE
International Workshop on Machine Learning for Signal Processing,
Sept. 2012, pp. 1–6.
[23] T. E. Bogale and L. Vandendorpe, “Max-Min SNR signal energy based
spectrum sensing algorithms for cognitive radio networks with noise
variance uncertainty,” IEEE Trans. Wireless Commun., vol. 30, no. 1,
pp. 280 – 290, Jan. 2014.
[24] T. E. Bogale, L. B. Le, and X. Wang, “Hybrid analog-digital channel
estimation and beamforming: Training-throughput tradeoff,” IEEE
Trans. Commun., Dec. 2015.
[25] L. Bernado, N. Czink, T. Zemen, and P. Belanovic, “Physical layer simulation results for IEEE 802.11p using vehicular non-stationary channel
model,” in Proc. IEEE International Conference on Communications
Workshops, May 2010, pp. 1–5.
[26] M. A. Ingram, G. Acosta, and L. Dong, “Small-scale fading models for
the IEEE 802.11p (WAVE/DSRC) standard,” 2006, Georgia Institute of
Technology (Technical Report).
| 7cs.IT
|
arXiv:1802.07789v1 [cs.CV] 21 Feb 2018
Semantic Segmentation Refinement by Monte Carlo Region Growing of High
Confidence Detections
Philipe A. Dias
Henry Medeiros
[email protected]
[email protected]
EECE, Marquette University
Milwaukee, WI, USA
Abstract
tasks aim at pixel-wise classification of known objects or of
different instances of these objects. These subproblems aim
to achieve increasingly complex goals and have therefore
been addressed with different levels of success.
In recent years, the combination of Deep Convolutional
Neural Networks (DCNN) and increasingly larger publicly
available datasets has led to substantial improvements in
the state-of-the-art of image classification tasks [15, 12].
Similar advances became then possible for object detection
with the advent of region proposal networks (RPN) [23].
For segmentation tasks, however, the performance of typical DCNNs architectures is limited by the typical downsampling (pooling) intrinsic to such networks. Downsampling is employed to learn hierarchical features (from lowlevel to high-level features), but it ultimately leads to imprecise, coarse segmentation in scenarios that require pixelwise predictions. Different strategies with varying levels
of complexity have been proposed to address this limitation, including atrous convolutions [3], upsampling strategies [18, 24, 17], and post-processing based on superpixels
[20] or conditional random fields (CRFs) [3, 14]. Although
these methods represent significant steps towards solving
this problem, the segmentation they produce tend not to be
finely aligned with the boundaries of the objects.
In this paper, we propose the Region Growing Refinement (RGR) algorithm, a post-processing module that operates on the principle of appearance-based region growing
to refine the predictions generated by a DCNN for semantic
segmentation. Based on the classification scores available
from the detector, our method first divides the image into
three regions: high confidence background, high confidence
object, and uncertainty region. The pixels within the uncertainty region, which are the ones that tend to be misclassified by DCNN-based methods, are then labeled by means of
region growing. We apply Monte Carlo sampling to select
initial seeds from the high confidence regions, which are
then grown using a similarity criterion based on a distance
metric computed in the 5-D space of spatial and color coordinates. Our model has the advantage of not requiring any
Despite recent improvements using fully convolutional
networks, in general, the segmentation produced by most
state-of-the-art semantic segmentation methods does not
show satisfactory adherence to the object boundaries. We
propose a method to refine the segmentation results generated by such deep learning models. Our method takes as
input the confidence scores generated by a pixel-dense segmentation network and re-labels pixels with low confidence
levels. The re-labeling approach employs a region growing mechanism that aggregates these pixels to neighboring
areas with high confidence scores and similar appearance.
In order to correct the labels of pixels that were incorrectly
classified with high confidence level by the semantic segmentation algorithm, we generate multiple region growing
steps through a Monte Carlo sampling of the seeds of the
regions. Our method improves the accuracy of a state-ofthe-art fully convolutional semantic segmentation approach
on the publicly available COCO and PASCAL datasets, and
it shows significantly better results on select sequence of the
finely-annotated DAVIS dataset.
1. Introduction
Image understanding is a primary goal of computer vision, which is crucial to several applications ranging from
activity recognition to object tracking, autonomous navigation, security and surveillance, among many others. One of
the most important aspects of image understanding is the
determination of the objects present in the image. Such determination can be carried out at different levels of granularity, which correspond to four well-known subproblems
of image understanding: image classification, object detection, semantic segmentation, and instance segmentation.
While the objective of image classification is to label the
scene at the image-level and object detection attempts to localize the bounding boxes of known objects, segmentation
1
dataset-specific parameter optimization, working in a fully
unsupervised manner.
We employ the Fully Convolutional Instance-aware Semantic Segmentation (FCIS) algorithm [16], the winner
of the MS COCO 2016 Semantic Segmentation Challenge
[17], as a predictor module as well as the baseline for
our performance evaluation. When evaluating our results
on the MS COCO dataset, we observed that the available
ground truth annotations contain non-negligible inaccuracies in terms of adherence to the actual boundaries of the
objects. To assess the effects of this problem and achieve a
more accurate evaluation of the benefits of our approach, we
also assess its efficacy on the PASCAL VOC2012 [6] validation set as well as on selected video sequences from the
DAVIS (Densely Annotated VIdeo Segmentation) dataset
[22]. Compared to the first two datasets, annotations provided for the DAVIS dataset are more fine-grained, with
tighter adherence to actual object real boundaries. As a
result, in addition to relatively small increases in segmentation accuracy for the MS COCO dataset (+1.5% in AP )
and the PASCAL validation sets (+0.5% in mAP %), we
report significantly better quantitative results for the DAVIS
sequences (+3.2% in IoU %), which more realistically reflect the segmentation improvement observed through qualitative (visual) inspection.
2. Related Work
The relatively recent development of DCNNs trained on
large publicly available datasets represented an inflection
point in image understanding research. Compared to the period when models based on hand-engineered features (e.g.
HOG [5], SIFT [19]) were the norm, currently the state-ofthe-art in object recognition tasks is being improved at a
dramatically faster pace. Between 2011 and 2015, the top5 classification error on the ImageNet dataset was reduced
from 25.8% to 3.57% [12].
However, traditional CNN-based methods that are effective for image classification and object detection present
limitations for segmentation tasks. The combination of
max-pooling and striding operations constitute a standard
strategy for learning hierarchical features, which are a determining factor in the success of deep learning models. They
explore the transition invariance favored by image-level labeling tasks, i.e., the fact that only the presence of an object
matters, not its location. Such premise, however, does not
hold for pixel-dense classification tasks, which require precise object localization. As pointed out by Chen et al. in [3],
such spatial insensitivity and image downsampling inherent
to conventional DCNNs are two major hindering factors for
their performance in image segmentation. A particularly
noticeable effect is that these methods generate coarse segmentation masks, with limited boundary adherence.
One particularly successful strategy investigated for seg-
mentation tasks is the concept of fully convolutional networks (FCNs) [18]. In FCNs, the traditionally fully connected layers at the end of conventional CNNs are replaced
by convolutional layers, which upsample the feature maps
to generate a pixel-dense prediction. Different upsampling
or decoding approaches have been proposed, including the
use of deconvolution layers [21, 18], and encoder-decoder
architectures with skip-layer connections [24, 13, 11, 17].
While these approaches aim at recovering the information
lost due to downsampling, other methods focus on reducing
the downsampling rate. The Deeplab architecture [3] is one
such method in which the concept of dilated (or atrous) convolutions is successfully employed to reduce the downsampling rate from 1/32 to 1/8. Such strategy, however, performs convolutions on high-resolution feature maps, requiring substantially more computational resources than other
approaches.
A variation of the semantic segmentation problem is instance segmentation, which requires detecting and segmenting individual object instances. Coarse segmentations are a
significant challenge for this type of task, since neighboring
instances of objects of the same class are frequently merged
into a single segmentation. Dai et al. in [4] introduce the
concept of instance-sensitive FCNs, in which an FCN is designed to compute score maps that determine the likelihood
that a pixel belongs to a relative position (e.g.right-upper
corner) of an object instance. FCIS [16] is an extension
of this approach, which achieved the state-of-the-art performance and won the COCO 2016 segmentation competition.
In addition to adjustments in CNN architectures, multiple works have investigated the use of low-level information strategies to construct models for object detection and
semantic segmentation. Following the success of CNNs on
image classification tasks, Girschick et al. in [9] introduced
the model known as RCNN that performs object detection
by evaluating multiple region proposals. In RCNN, the region proposals are generated using Selective Search [26],
which consists on merging a set of small regions [8] based
on different similarity measures, an approach closely related to the concept of superpixels.
Superpixel are perceptually meaningful clusters of pixels, which are grouped based on color similarity and other
low-level properties. Stutz et al. [25] provide a review of
existing superpixel approaches, with a comprehensive evaluation of 28 state-of-the-art algorithms. One of the most
widely-used method is the Simple Linear Iterative Clustering (SLIC) algorithm [2], a k-means-based method that
groups pixels into clusters according to their proximity in
both spatial and color spaces. Semantic segmentation models using superpixels traditionally employ them as a preprocessing step. The image is first divided into superpixels,
followed by a prediction step in which each element is individually evaluated using engineered hierarchical features
[10] or DCNNs [7, 20].
In addition to the generation of region proposals or presegmentation elements, local-appearance information has
been also employed in post-processing steps to improve
segmentations obtained with deep CNN models. One such
post-processing approach is to model the problem as a CRF
defined over neighboring pixels or small patches, which
however might lead to excessive smoothing of boundaries
due to the lack of long-range connections [14]. Krähenbühl
et al. in [14] introduces an efficient algorithm for fully connected CRFs containing pairwise potentials that associate
all pairs of pixels in a image. This algorithm is successfully
exploited by the Deeplab model [3] to produce fine-grained
segmentations. However, this refinement model contains
parameters that have to be optimized in a supervised manner when applied to different datasets.
3. Proposed Approach
The method we propose for refinement of segmentation
boundaries is a generic post-processing module that can be
coupled to the output of any CNN or similar model for semantic segmentation. Our RGR algorithm consists of four
main steps: 1) identification of low and high confidence
classification regions; 2) Monte Carlo sampling of initial
seeds; 3) region growing; and, 4) majority voting and final
classification. The operations that comprise these steps are
described in detail below. In our description, we make reference to Figure 1 and Algorithm 1, which lists the operations
performed by our method for each proposed detection in an
image. If multiple detections are present, the algorithm is
executed once for each detection.
1) Step 1 - Thresholding the image into three regions:
Conventional models typically infer a final classification by
pixel-wise thresholding the scores obtained for each class.
In such case, the selection of an adequate threshold represents a trade-off between false positives and false negatives.
Instead of relying on intermediate threshold values, our refinement module directly exploits the available classification scores to differentiate between three regions: a) high
confidence foreground (or object) RF ; b) high confidence
background RB ; and, c) uncertainty region RU . As defined
in Eq. 1, these regions are identified using two high confidence thresholds τF and τB
RF = {pj |M (pj ) > τF } ,
RU = {pj |τB < M (pj ) < τF } ,
(1)
RB = {pj |M (pj ) < τB } ,
where pj is the j-th pixel in the input image I, and M (pj )
is its corresponding detection score computed according to
the detection confidence map M . High confidence regions
correspond to areas where pixels present scores near the extremes of the likelihood range. For normalized scoremaps,
values lower than a threshold τB ∼ 0.0 identify pixels in
the high confidence background, while values larger than
τF ∼ 1.0 indicate high confidence foreground elements.
In order to recover possible false negatives, morphological
thickening is performed on the background RB boundaries.
The fourth leftmost image in Figure 1 illustrates the division
of the image into the three regions RF , RU , and RB .
2) Step 2 - Monte Carlo sampling of initial seeds: Inspired by the notion of pixel affinity employed by many algorithms for superpixel segmentation, our method applies
a region growing approach to classify pixels within the uncertainty zone. More specifically, we build on the simple
non-iterative clustering (SNIC) algorithm [1].
The first difference between RGR and SNIC is the selection of initial seeds. SNIC selects initial seeds on a regular
grid over the whole image. In contrast, our algorithm selects seeds by Monte Carlo sampling the high confidence
regions. Such an adaptation guarantees that pixels composing the uncertainty region will be labeled according to high
confidence information.
Our Monte Carlo approach consists of ns iterations. In
each iteration i, a set S (i) of initial seeds is defined by uniformly sampling the high-confidence area RH = RB ∪ RF .
Let ph represents pixels within the region RH , where the index h = 1, . . . , |RH |. Uniform sampling of seeds can thus
be performed index-wise according to h ∼ U (1, |RH |).
We determine the number of seeds to be sampled as a
function of the sampling domain area (i.e. |RH |) and the desired average spacing between samples. Such a spacing between seeds is necessary to ensure the availability of paths
through which all initial centroids can propagate through
the uncertainty region. By drawing multiple sets of seeds,
we reduce the impact of possible high confidence false positives on the final segmentation.
3) Step 3 - Region Growing: Initial centroids are grown
according to an adaptation of the SNIC algorithm. The
dashed block at the bottom of Figure 1 illustrates the sequence of operations performed by RGR for region growing. Like its predecessor SLIC [2], SNIC measures the similarity between a pixel and a centroid as their distance in
a five-dimensional space of color and spatial coordinates.
Let the spatial position of a pixel be represented by a vector x = [x y]T , while its color is expressed in the CIELAB
color-space by c = [l a b]T . The distance between a pixel j
and a centroid k is given by
s
kxj − xk k22
kcj − ck k22
+
,
(2)
dj,k =
θs
θm
where θs and θm as normalizing factors for spatial and color
distances, respectively.
Aiming at lower computational times, we restrict the region growing to a Region of Interest (RoI) around the uncertain zone RU . Based on the spatial distance to RU , the
Figure 1. Best viewed in color. Diagram illustrating the sequence of tasks performed by the proposed RGR model for segmentation
refinement. Each task and its corresponding output (shown above the arrows) are described in Algorithm 1.
background RB can be split into two regions: far background Rf B and near background RnB . Since the far background is unlikely to influence the clustering of the uncertain region, Rf B can be ignored during the region growing
computation. Hence, we define the RoI for region growing
as
RoI = RnB ∪ RF ∪ RU .
(3)
The region growing implementation relies on a priority
queue, which is constantly populated with nodes that correspond to unlabeled pixels 4 or 8-connected to a region being
grown. This queue is sorted according to the similarity between the candidate pixel and the growing region, given by
the metric defined in Eq. 2. While the queue is not empty,
each iteration consists of: 1) popping the first element of
the queue, which corresponds to the unlabeled pixel that is
mostly similar to a neighboring centroid k; 2) annexing this
(i)
pixel to the respective cluster ψk ; 3) updating the region
centroid; and, 4) populating the queue with neighbors of
this pixel that are still unlabeled.
Compared to the original SNIC algorithm, we add a con-
straint in order to reduce the incidence of false detections.
A node is only pushed into the queue if its distance to the
corresponding centroid is smaller than a certain value dMax .
In this way, we ensure that an unlabeled pixel within RU
will only inherit information from high confidence pixels
that are sufficiently similar to it. This creates the possibility
of “orphan” elements, i.e., pixels for which no node is ever
created. Such pixels are therefore classified as background.
For each set of initial seeds Si the region growing process
generates as output a cluster map Ψ(i) that associates each
(i)
pixel to a respective cluster ψk .
4) Step 4 - Majority voting and final classification: Following the region growing process, RGR conducts a majority voting procedure to ultimately classify each generated
cluster into foreground or background. As expressed in Eq.
4, a pixel pj contributes a positive vote for foreground classification if its original prediction score M (pj ) is larger than
a threshold τ0 . We compute the ratio of positive votes across
(i)
all pixels pj ∈ ψk , generating a refined likelihood map
Algorithm 1 RGR refinement algorithm
Input: Image I, confidence map M
Output: Refined semantic segmentation Ŷ of image I
1: Threshold M into three regions: background RB , foreground RF and uncertain zone RU
2: Define a Region of Interest (RoI) according to Eq. 3
3: for i = 1 to ns do
4:
Form a set S (i) of initial seeds by uniformly sampling from RB ∪ RF
5:
Generate a set of clusters Ψ(i) by performing region
growing using the SNIC algorithm with S (i) as input
(i)
For each generated cluster ψk ∈ Ψ(i) , compute
confidence map M̂ (i) according to Eq. 4
7: end for
8: Compute the pixel-wise average M of M̂ (i) , i =
1, . . . , n
9: Generate Ŷ by pixel-wise thresholding M
6:
M̂ (i) for each set of clusters according to
Y = {pj |M (pj ) > τ0 } ,
M̂
(i)
(pj ) =
|Y |
(i)
|ψk |
(4)
4.1. COCO 2016 segmentation dataset
(5)
Based on the COCO official guidelines, we evaluate the
performance obtained by FCIS and FCIS+RGR on the validation set composed of 40k images. Table 4 summarizes
the performance of both methods according to the standard
COCO metrics including AP (averaged over 0.5 : 0.95
IoU thresholds), AP50 , AP75 APS , APM APL (AP at different scales). While increasing the number of true positive detections (+0.3% in AR), for all scenarios the naive
FCIS+SPPX approach decreases the AP by introducing a
larger number of false positives into the final detection.
Our refinement model, RGR, on the other hand, increases the baseline FCIS overall performance by 1.8%.
Compared to the variation of +0.5% in AP50 , the increase
of 2.8% in AP75 demonstrates that RGR is particularly successful for cases where the detections obtained from the input CNN are accurately located.
In addition, Figure 2 presents the average precision obtained for each of the 80 COCO categories. Due to its region
growing strategy based on local affinity characteristics, it is
natural that RGR refinement is especially effective for objects with more uniform appearance (e.g. airplane, frisbee,
baseball bat). Nevertheless, these results also demonstrate
the robustness of the refinement provided by RGR, since
none of the object categories observe a noticeable decrease
in average precision.
A closer inspection by means of a qualitative analysis
also shows that the metrics obtained might be misleading.
Despite the extensive efforts and systematic approaches described in [16] to create such a huge dataset, for some in-
The likelihood maps M̂(i) obtained from the ns Monte
Carlo samplings are P
averaged to generate a final pixel dense
ns
score map M = n1s i=1
M (i) . Finally, each pixel is classified into foreground if more than 50% of its average votes
correspond to positives. Otherwise, the region is labeled as
background. Mathematically, this is equivalent to
Ŷ (pj ) = 1M(pj )>0.5 .
to 1.5 × τ0 , hence 0.6. As for the background detection, we
configure the high confidence lower threshold as τB = 0.0.
We opted for the publicly available FCIS model [16]
trained on the MS COCO 2016 dataset, which represents
the state-of-the-art approach on semantic segmentation, as
the input for our refinement algorithm as well as the baseline comparison approach. We denote the combination of
FCIS and the refinement module RGR as FCIS+RGR.
Since RGR works in a region growing manner that is inspired by the concept of superpixel segmentation, our performance analysis also includes FCIS+SPPX, a naive refinement method that consists of performing majority voting within superpixels generated with SNIC.
We also considered refinement using the CRF approach
proposed in [14]. We evaluated the efficacy of this refinement technique on both PASCAL validation set and
DAVIS selected sequences. In order to evaluate generalization capability, these experiments were performed without
any dataset-specific fine-tuning of parameters. Decrease in
average precision values were observed for both datasets,
revealing that parameters optimization is crucial for this
method to be efficient.
(6)
4. Experiments
To the best of our knowledge, the COCO dataset is the
largest publicly available dataset for semantic/instance segmentation, with a total of 2.5 million labeled instances in
328, 000 images [17]. However, as discussed in Section
4.1, COCO ground truth annotations contain imprecisions
intrinsic from its labeling procedure. Hence, with the purpose of minimizing the influence of dataset-specific errors,
we assess the performance of FCIS and FCIS+RGR on: i)
the COCO 2016 validation set; ii) the PASCAL VOC 2012
dataset; and, iii) selected video sequences of the DAVIS
dataset.
For all our experiments, we defined the thresholds as
follows. First, τ0 corresponds to the original detector optimal threshold. For FCIS this value is 0.4, as reported in
[16]. To identify high confidence foreground, we empirically selected a high confidence threshold τF corresponding
FCIS
FCIS + SPPX
FCIS + RGR
AP (%)
35.1
34.9
36.9
AP50 (%)
60.1
59.0
60.6
COCO 2016 (val)
AP75 (%) APS (%)
36.5
9.8
36.4
9.3
39.3
11.4
APM (%)
38.4
38.2
40.7
APL (%)
59.8
59.6
60.5
PASCAL 2012 (val)
mAP (%)
70.6
70.9
71.1
DAVIS
IoU (%)
71.2
72.8
74.4
Table 1. Comparison between results obtained by FCIS+RGR and FCIS on COCO 2016 validation set.
stances the segmentation annotations provided as ground
truth lack accurate adherence to real object boundaries. As
consequence, a visual inspection reveals that improvements
obtained through RGR refinement are not reflected in the final metrics for a significant number of images. Figure 3 provides multiple examples of the imprecisions in ground truth
annotations, collected from different object classes. While
for the airplane example the segmentation obtained using
RGR is clearly better in qualitative terms, its overlap (IoU )
with the ground truth is 7.7% lower than the one associated
with the FCIS output.
4.2. PASCAL VOC 2012 segmentation dataset
Table 4 also contains a summary of the results obtained
by FCIS, FCIS+SPPX and our method FCIS+RGR. The
segmentation refinement generated using RGR provides final mAP slightly better than both FCIS (+0.5%) and the
refined version using naive superpixel-wise majority voting.
However, as illustrated in Figure 4, a qualitative inspection
reveals that such a small difference in performance does not
properly reflect the higher boundary adherence provided by
the refinements produced by RGR. Although the PASCAL
annotations are more pixel-accurate than the ones available
for COCO, they contain 5 pixels-wide border areas that are
not consider neither object nor background, ultimately limiting the evaluation of boundary adherence.
4.3. DAVIS 2016 selected video sequences
Given the aforementioned inaccuracies in the annotations available for the COCO and PASCAL datasets, we
additionally report quantitative results on selected video sequences of the DAVIS dataset. The DAVIS dataset for
Video Object Segmentation is composed by high quality
video sequences [22], with pixel-accurate ground truth segmentation for each frame. From its 50 original video sequences, we selected a total of 24 sequences that contain
Figure 2. Mean average precision values obtained by FCIS and FCIS+RGR for each of the 80 COCO categories.
Figure 3. Examples of detections on the COCO dataset. From left to right: original image, ground truth, FCIS detection, FCIS+RGR.
Ground truth annotations for images in the second and third rows are also examples of annotation with poor boundary adherence.
Figure 4. Examples of detections on the PASCAL VOC 2012
dataset. From left to right: original image, ground truth, FCIS
detection, FCIS+RGR.
target objects whose corresponding classes are contained
within the 80 objects categories composing the COCO
dataset. Moreover, these sequences encompass 13 different COCO classes, including classes for which FCIS+RGR
did not provide significant performance improvements on
the COCO evaluation (e.g. person, bicycle, boat and bear).
As summarized in Table 4, segmentations obtained using FCIS+RGR have an average IoU 3.2% higher than the
original FCIS ones, while 1.6% better than the results obtained using a naive superpixel-wise majority voting for refinement. Official metrics for performance evaluation on
the DAVIS also include a contour accuracy metric F, which
corresponds to a contour F-measure computed based on
contour-based precision and recall. Segmentations refined
using RGR yielded an increase of 2.8% in F, confirming
its ability to improve boundary adherence.
Figure 6 presents the results obtained for each video
Figure 5. Mean average precision (mAP ) obtained by FCIS and
FCIS+RGR for each of the PASCAL VOC 2012 classes (val).
the robustness against high confidence false positives, errors might be propagated if the score maps collected from
the detector module (CNN) contain regions with high concentration of false positives. An example of such case is
found in Figure 7. Given the high confidence of the false
positives generated by the FCIS model for internal parts
of the bicycle wheels, RGR is therefore unable to correct
these mistakes and hence these regions remain incorrectly
segmented.
5. Conclusion
Figure 6. Mean IoU (Jaccard) obtained by FCIS and FCIS+RGR
for each of the selected DAVIS 2016 video sequences.
sequence, with FCIS+RGR providing improvements in a
range from 1.1% to 6.9% depending on the sequence.
The fact that larger quantitative improvements are obtained for the DAVIS sequences corroborates our argument
that the annotations available for the COCO and PASCAL
datasets provide limited information regarding boundary accuracy/adherence of segmentation methods.
We have presented RGR, an effective post-processing
algorithm for segmentation refinement. Traditionally, the
final classification step of existing methods for semantic
segmentation consists of thresholding score maps obtained
from DCNNs. Based on the concepts of Monte Carlo sampling and region growing, our algorithm achieves an adaptive thresholding by exploiting high confidence detections
and low-level image information. Our results demonstrate
the efficacy of RGR refinement, with precision increases on
three different segmentation datasets.
Moreover, we highlight limitations of existing datasets in
terms of boundary adherence of available ground truth annotations. This is an intrinsic limitation of annotations procedures that rely on approximating segmentations as polygons. While the quality of the segmentation is proportional
to the number of vertices selected by the user, on the other
hand selecting more points increases the labeling time per
object. As future work, we consider exploiting RGR for
improving annotations available for existing datasets or designing an alternative annotation tool. Finally, we hypothesize that the performance of existing DCNNs for semantic segmentation could be improved by training with RGR
as an additional step before computing losses. Such an arrangement could lead to a detector module that optimally interacts with RGR, identifying keypoints for the refinement
process.
References
Figure 7. Examples of detections on the DAVIS dataset. From
left to right: original image, ground truth, FCIS detection,
FCIS+RGR.
4.4. Failure cases
As demonstrated by the experimental results, the quality
of the refinement obtained using RGR depends on the accuracy of the detector model used as input, especially in terms
of localization. Although Monte Carlo sampling improves
[1] R. Achanta and S. Sabine. Superpixels and Polygons using
Simple Non-Iterative Clustering. In CVPR ’17, pages 4651–
4660, 2017. 3
[2] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, and
S. Süsstrunk. SLIC superpixels compared to state-of-the-art
superpixel methods. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 34(11):2274–2281, 2012. 2, 3
[3] L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and
A. L. Yuille. Deeplab: Semantic image segmentation with
deep convolutional nets, atrous convolution, and fully connected crfs. arXiv preprint arXiv:1606.00915, 2016. 1, 2,
3
[4] J. Dai, K. He, Y. Li, S. Ren, and J. Sun. Instance-sensitive
fully convolutional networks. In European Conference on
Computer Vision, pages 534–549. Springer, 2016. 2
[5] N. Dalal and B. Triggs. Histograms of oriented gradients for
human detection. In Computer Vision and Pattern Recognition, 2005. CVPR 2005. IEEE Computer Society Conference
on, volume 1, pages 886–893. IEEE, 2005. 2
[6] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn,
and A. Zisserman. The PASCAL Visual Object Classes
Challenge 2012 (VOC2012) Results. http://www.pascalnetwork.org/challenges/VOC/voc2012/workshop/index.html.
2
[7] C. Farabet, C. Couprie, L. Najman, and Y. LeCun. Learning
hierarchical features for scene labeling. IEEE transactions
on pattern analysis and machine intelligence, 35(8):1915–
1929, 2013. 3
[8] P. F. Felzenszwalb and D. P. Huttenlocher. Efficient graphbased image segmentation. International journal of computer vision, 59(2):167–181, 2004. 2
[9] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic
segmentation. Proceedings of the IEEE Computer Society
Conference on Computer Vision and Pattern Recognition,
pages 580–587, 2014. 2
[10] S. Gould, J. Rodgers, D. Cohen, G. Elidan, and D. Koller.
Multi-class segmentation with relative location prior. International Journal of Computer Vision, 80(3):300–316, 2008.
3
[11] B. Hariharan, P. Arbeláez, R. Girshick, and J. Malik. Hypercolumns for object segmentation and fine-grained localization. In Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, volume
07-12-June, pages 447–456, 2015. 2
[12] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages
770–778, 2016. 1, 2
[13] A. Kendall, V. Badrinarayanan, and R. Cipolla. Bayesian
SegNet: Model Uncertainty in Deep Convolutional EncoderDecoder Architectures for Scene Understanding. arXiv
preprint arXiv:1511.02680, 2015. 2
[14] P. Krähenbühl and V. Koltun. Efficient Inference in Fully
Connected CRFs with Gaussian Edge Potentials. In Advances in neural information processing systems, pages 109–
117, 2011. 1, 3, 5
[15] A. Krizhevsky, I. Sutskever, and G. E. Hinton. ImageNet
Classification with Deep Convolutional Neural Networks.
Advances In Neural Information Processing Systems, pages
1–9, 2012. 1
[16] Y. Li, H. Qi, J. Dai, X. Ji, and Y. Wei. Fully Convolutional Instance-aware Semantic Segmentation. arXiv
preprint arXiv:1611.07709, 2016. 2, 5
[17] G. Lin, A. Milan, C. Shen, and I. Reid. RefineNet: MultiPath Refinement Networks for High-Resolution Semantic
Segmentation. arXiv preprint arXiv:1611.06612, 2016. 1,
2, 5
[18] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional
networks for semantic segmentation. In Proceedings of the
IEEE Computer Society Conference on Computer Vision and
Pattern Recognition, volume 07-12-June, pages 3431–3440,
2015. 1, 2
[19] D. G. Lowe. Distinctive Image Features from Scale-Invariant
Keypoints. International Journal of Computer Vision, 2004.
2
[20] M. Mostajabi, P. Yadollahpour, and G. Shakhnarovich. Feedforward semantic segmentation with zoom-out features. In
Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, pages 3376–3385, 2015. 1, 3
[21] H. Noh, S. Hong, and B. Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE
International Conference on Computer Vision, volume 2015
Inter, pages 1520–1528, 2015. 2
[22] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. V. Gool,
M. Gross, and A. Sorkine-Hornung. A Benchmark Dataset
and Evaluation Methodology for Video Object Segmentation. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 724–732, 2016. 2, 6
[23] S. Ren, K. He, R. Girshick, and J. Sun. Faster R-CNN:
Towards Real-Time Object Detection with Region Proposal
Networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(6):1137–1149, 2017. 1
[24] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and
Computer-Assisted Intervention, pages 234–241. Springer,
2015. 1, 2
[25] D. Stutz, A. Hermans, and B. Leibe. Superpixels: An evaluation of the state-of-the-art. Computer Vision and Image
Understanding, 2017. 2
[26] J. R. R. Uijlings, K. E. A. Van De Sande, T. Gevers, and
A. W. M. Smeulders. Selective search for object recognition.
International Journal of Computer Vision, 104(2):154–171,
2013. 2
| 1cs.CV
|
arXiv:1605.03733v1 [cs.SY] 12 May 2016
Kernel-based system identification
from noisy and incomplete input-output
data
Riccardo S. Risuleo, Giulio Bottegal and Håkan Hjalmarsson
∗
May 13, 2016
Abstract
In this contribution, we propose a kernel-based method for the identification of linear systems from noisy and incomplete input-output datasets.
We model the impulse response of the system as a Gaussian process whose
covariance matrix is given by the recently introduced stable spline kernel.
We adopt an empirical Bayes approach to estimate the posterior distribution of the impulse response given the data. The noiseless and missing
data samples, together with the kernel hyperparameters, are estimated
maximizing the joint marginal likelihood of the input and output measurements. To compute the marginal-likelihood maximizer, we build a
solution scheme based on the Expectation-Maximization method. Simulations on a benchmark dataset show the effectiveness of the method.
1
Introduction
Common formulations of system identification problems postulate the perfect
knowledge of the input signal feeding the unknown system [1]. In many applications however, the input signal is available only in a noisy version, giving raise to
a setup usually referred to as an errors-in-variables (EIV) model [2]. Static EIV
models have been subject of extensive studies in the statistical literature since
the beginning of the last century [3]; later the system identification community
has become interested in dynamical EIV models [4, 5, 2].
Identification of EIV systems is a challenging task; even in the linear case,
standard least-squares yields biased estimates, due to the presence of noise in
the regressors. Therefore, lots of efforts have been devoted to the development
of ad-hoc methods for EIV systems. Bias eliminating least-squares (BELS) have
∗ R. S. Risuleo, G. Bottegal and H. Hjalmarsson are with the ACCESS Linnaeus
Center, School of Electrical Engineering, KTH Royal Institute of Technology, Sweden
(risuleo;bottegal;[email protected]). This work was supported by the European Research Council under the advanced grant LEARN, contract 267381 and by the Swedish Research Council under contract 621-2009-4017.
1
been introduced in [6] to correct the bias of standard least-squares. Another
method for the identification of EIV systems has been obtained by generalizing
the so-called Frisch scheme, originally developed for static EIV models [3], to
dynamic models. Interestingly, the dynamic Frisch scheme provides a unique
identified model under mild conditions [7]. This is in contrast with the static
case, where in general many models are compatible with the observed data.
The accuracy of the Frisch scheme for dynamic system identification has been
extensively studied in the literature [8]; the method has been recently extended
to more general noise setups [9, 10, 11]. Other EIV identification methods rely
upon maximum-likelihood criteria. Both time-domain [12, 13] and frequencydomain [14] approaches have been developed in the past; for a survey and a
comparison of the maximum likelihood methods see [15].
In this paper, we consider a more general dynamic EIV setup. Specifically,
we assume that some of the samples may be missing, for instance due to lossy
transmission channels or sensor malfunction. Therefore, our task is to jointly
identify the system and reconstruct the missing input-output values. Some
techniques to deal with this problem have been proposed in the past, both
in time and frequency domains [16, 17, 18, 19, 20]. Recently, regularization
techniques based on nuclear norm have been proposed for system identification
with missing data [21, 22].
The method described in this paper to deal with EIV models with missing
data relies on a regularized kernel-based approach. Interpreting kernel-based
regularization as a Gaussian regression problem [23], we model the unknown
impulse response of the system as a zero-mean Gaussian random vector. The
covariance matrix is given by the recently introduced stable spline kernel [24, 25],
which penalizes non-exponentially stable systems. According to the empirical
Bayes (EB) paradigm [26], we obtain an impulse response estimator as a function of the noiseless input and the kernel hyperparameters. These quantities are
estimated maximizing the joint marginal likelihood (marginal likelihood) of the
noisy inputs and outputs. We devise an iterative algorithm to solve the marginal
likelihood maximization problem, based on the EM method. We briefly address
the problem of identifiability. We test the proposed approach with numerical
simulations.
In this paper we consider the case of both missing input and missing output samples, as well as noisy data; as compared to [27] where a kernel-based
approach is adopted for the case of noiseless input and missing output samples.
The paper is organized as follows. In Section 2 we formulate the problem
the identification of EIV models with missing data. In Section 3 we show how
to estimate the system impulse response. In Section 4 we solve the marginal
likelihood problem that yields the missing samples and the kernel hyperparameters. In Section 5 we discuss some pitfalls in the model. In Section 6 we validate
our method on a benchmark dataset. In Section 7 we discuss our results and
conclude the paper.
2
1.1
Notation
b
We denote by “{ak }” a sequence of scalars ak indexed by k; “{ak }k=a ” is the
b
set of ak with k ranging from a to b. Given {ak }k=a , “a” indicates the column
vector of the stacked scalars and “ai ” indicates the ith element of said vector.
If a ∈ RN is a column vector, “Tm×n (·)” indicates the Toeplitz operator that
associates to the vector a the m × n matrix A such that
(
ai−j+1 , i ≥ j, i − j + 1 ≤ N,
(1)
A i,j =
0,
otherwise.
The symbol “δi,j ” denotes the Kronecker delta and “⊗” is the standard Kronecker
product between matrices. The symbol “ ∼
=” indicates equality up to an additive
constant.
2
Problem formulation
We consider the problem of identifying a dynamic system from noisy samples
of input and output. Figure 1 shows a schematic representation of the setup
under study. The system is strictly causal, asymptotically stable, and linear
wt
ηt
+
vt
S
ut
εt
+
yt
Figure 1: Block scheme of the system setup.
time-invariant. The input-output relation can be represented as
vt =
∞
X
gk wt−k ,
(2)
k=1
where the gk is the (unknown) impulse response of the system. The objective is
to reconstruct the samples of the impulse response from N samples of the input
ut and output yt . These samples are measurements of the true system input wt
and output vt , corrupted by sensor noises
ut = wt + ηt ,
y t = vt + ε t .
(3)
The noise sequences ηt and εt are assumed mutually independent, Gaussian and
white, with unknown variance σu2 and σy2 , respectively. The ratio γ = σy2 /σu2 is
assumed known, in order to guarantee identifiability (see e.g. [5]). We suppose
that the system is at rest prior to the collection of the measurements, that is
wk = 0, vk = 0, for all k < 1.
3
We assume also that some of the samples are not available; see the following
example.
Example 1. We have run a system for N = 9 time instants collecting the
following measurements:
u1
y1
×
u3
u4
×
×
u7
y2
×
y4
y5
y6
×
u9 ,
× y9 .
×
N
u
u
such that 1 ≤ tiu < ti+1
≤ N,
Define the set of natural numbers {tuk }k=1
y Ny
and utui is an available measurement. In a similar fashion define {tk }k=1 . These
sets indicate the Nu and Ny time instants at which we have available sensor
measurements of the input and output respectively. We define the available
measurement vectors ũ ∈ Nu and ỹ ∈ Ny such that
R
R
ũk = utuk ,
k = 1, . . . , Nu ,
ỹk = ytyk ,
k = 1, . . . , Ny .
P
(4)
Py as the Nu ×N respectively Ny ×N
Furthermore, define the operators u and
matrices defined by
,
u i,j = δi,tu
j
P
Py i,j = δi,t
y
j
.
(5)
By construction, these matrices are right semi-orthogonal :
PuPTu = IN
u
Py PTy = IN ;
,
(6)
y
they have full row rank and they represent the mappings betweev the complete
data and the available data:
ũ =
Pu u ,
ỹ =
Py y,
(7)
where u and y are vectors of all the stacked values of all (available and not)
measurements of input and output.
Example 1 (continued). The times of available input measurements are:
{tuk } = {1, 3, 4, 7, 9},
and the 5 × 9 matrix
(8)
Nu = 5,
Pu is
Pu
1
0
=
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
1
We can now formally define the problems of interest in this paper.
4
(9)
Problem 1. [System Identification] Given Nu ordered samples of the input
process ut , collected at times tuk , and Ny ordered samples of the output process yt ,
collected at times tyk , estimate the first n samples of the system impulse response
n
{gk }k=1 .
We also consider the problem of reconstructing the missing samples:
Problem 2. [Input smoothing] Given Nu ordered samples of the input process ut , collected at times tuk , and Ny ordered samples of the output process yt ,
collected at times tyk , estimate the sample wk , k ∈ , k ≤ N .
N
Problem 3. [Output smoothing] Given Nu ordered samples of the input process ut , collected at times tuk , and Ny ordered samples of the output process yt ,
collected at times tyk , estimate the sample vk , k ∈ , k ≤ N .
N
By adopting a kernel-based approach, we introduce a nonparametric model
for the impulse response that allows us to solve the three proposed problems
with a single algorithm based on a marginal likelihood approach. We will first
see how to solve Problem 1, using an EB approach.
Remark 1. If tuk = tyk = k for k = 1, . . . , N , then ũ = u and ỹ = y, and
Problem 1 corresponds to the standard dynamic EIV setup (see [2] for a survey).
3
Kernel-based linear system identification
We first focus on the problem of identifying g. For a given wt , Problem 1
becomes a linear regression problem: collecting wt into the column vector w,
we can construct the N × n Toeplitz matrix W = TN ×n (w); using this matrix
we can write the convolution (2) as the matrix product
(10)
v = W g,
and we can formulate the regression problem in the available output data:
ỹ =
P
Py W g + ε̃.
(11)
In this equation ε̃ = y ε are the samples of the noise that correspond to the
available samples of the output. From the semi-orthogonality of y , we have
that
Eε̃ = 0, Eε̃ε̃T = σy2 INy .
(12)
P
Adopting a kernel-based approach [25], we model g as a Gaussian random vector,
with covariance matrix given by a kernel function suitable for linear system
identification. In particular, we use the first-order stable-spline kernel [24], so
that
g ∼ N (0, λKβ ),
Kβ i,j := β max(i,j) .
(13)
The quantity λ > 0 is a scaling factor, while β ∈ (0, 1) is a shaping parameter
that regulates the exponential decay of the realizations from (13). These two
parameters are usually referred to as hyperparameters.
5
By postulating (13) a Gaussian prior for the impulse response, we can derive
the joint distribution of the measurements ỹ and g as
Σỹ ΣTgỹ
ỹ
0
∼N
,
(14)
g
0
Σgỹ λKβ
where
Py W Kβ W T PTy + σy2IÑ ,
Σgỹ = λKβ PTy .
Σỹ = λ
y
(15)
From (14) we can calculate the posterior distribution of the unknown impulse
response parameters given the available data ỹ (See, e.g., [12, App. B.7]):
(16)
g|ỹ ∼ N (m, P )
where
1 T
W
P =
σy2
1
m = 2PWT
σy
P Py W + (λKβ )
−1
T
y
P
−1
(17)
T
y ỹ.
With the posterior distribution, we find the (Bayesian) minimum variance estimate of g as the posterior mean m. From (17), we see that the posterior mean
depends on the quantities λ, β and w, as well as on the output noise covariance σy2 . All these parameters are unknown and need to be estimated from the
data. Using an EB approach, we estimate the parameters by replacing them
with their maximum marginal likelihood estimates λ̂, β̂ and ŵ (and σ̂y2 , but
this needs a special treatment: see Section 5). In the next section we focus on
the problem of finding the maximizers of the marginal likelihood. Solving the
marginal likelihood problem is also the key to solve Problem 2 and Problem 3.
4
Kernel-based input and output smoothing
Input smoothing and hyperparameter selection
Consider the measurement model (3). We can write it as a regression in the
smoothed input wt by observing that
v = W g = Gw
(18)
where G = TN ×N (g). Considering also the unavailability of some data, we
obtain the linear regression model in the available measurements
ỹ
ε̃
yG
=
w+
(19)
ũ
η̃
u
P
P
6
P
where η̃ = u η is white noise of variance σu2 . Under the Bayesian prior assumption (13), the available observation vector and the impulse response parameter
are jointly Gaussian, with a log-likelihood given by
L(ỹ, ũ, g; w, λ, β) =
log p(ỹ|g; w) + log p(ũ; w) + log p(g; λ, β) ,
(20)
where
P
1
log p(ỹ|g; w) ∼
= − 2 kỹ − y Gwk2
2σy
1
log p(ũ|w) ∼
= − 2 kũ − u wk2
2σu
1
1
−1
log p(g; λ, β) = − g T (λKβ ) g − log det (λKβ )
2
2
P
(21)
Since g is not available, we can interpret it as a latent variable, which we estimate
using the expectation maximization (EM) method. The term “EM method”
refers to a class of algorithms used to solve maximum likelihood problems with
latent variables. In these methods, an iterative algorithm is built by alternating
between estimating the likelihood, and updating the likelihood parameters using
the estimated likelihood.
The estimated likelihood is created in the expectation step by taking the
conditional expectation of the log-likelihood with respect to the posterior distribution of the latent variables given the available data, for some estimate of
the parameters:
Q θ; θ(k) = E L(ỹ, ũ, g; θ)
(22)
where θ = [wT , λ, β] are the parameters to be estimated, and the expectation
is taken with respect of the distribution (16), with the vector θ replaced with
an estimate θ(k) . By construction this function is such that
L ỹ, ũ; θ − L ỹ, ũ; θ(k) ≥ Q θ; θ(k) ) − Q θ; θ(k) .
(23)
where L ỹ, ũ; θ is the marginal likelihood of the available data. In the subsequent maximization step, the parameter update is chosen as the maximum of
Q(θ; θ(k) ), so that
Q θ(k+1) ; θ(k) − Q θ; θ(k) > 0
(24)
and consequently the marginal likelihood, in the updated parameters, is increased as well. By iterating the expectation and maximization steps, from any
initialization of the parameters, we obtain a sequence of estimates of the parameters that converge to a local maximizer of the marginal likelihood of the
available data (for a complete look on the EM algorithm [28]).
In the case at hand, we can compute the expectation of the joint log-
7
likelihood (20) in closed form:
P
1 T (k)
1
T
(k)
∼
E log p(ỹ|g; w) = 2 ỹ y M w − w A w ,
σy
2
1 T T
1
T
∼
E log p(ũ|w) = 2 ũ u w − w u u w ,
σu
2
n
o
−1 (k)
1
E log p(g; λ, β) = − trace λKβ
(P + m(k) m(k) T )
2
1
− log det (λKβ ) .
2
P
PP
All the expectations are taken with respect to the posterior distribution (16),
with λ, β, and w replaced by their estimates λ(k) , β (k) and w(k) . In these expressions m(k) and P (k) are the posterior mean and covariance as expressed in (17),
with the unknown parameters replaced by their estimates; M (k) = TN ×N (m(k) )
is the Toeplitz matrix of the posterior mean and A(k) is the posterior second
T
T
moment of the matrix GT Ty = TN ×N (g)
y , that is
A(k) = E GT Ty y G
h
i
(25)
= RT (P (k) + m(k) m(k) T ⊗ Ty y R .
P
P
PP
PP
The N n × N matrix R is defined as
RT = I N S
S2
···
Sn
(26)
where S is the N × N upward shift operator
S i,j = δi,j−1 .
(27)
To find the updated parameter values, we maximize the conditional expectations with respect to the parameters:
w(k+1) = arg max E log p(ỹ|g; w) + E log p(ũ|w) ,
(28)
λ(k+1) , β (k+1) = arg max E log p(g; λ, β) .
(29)
w
λ,β
The cost function in (28) is quadratic in the decision variable, so the maximum is available in closed form as
−1
w(k+1) = A(k) + γ Tu u
M (k) T Ty ỹ + γ Tu ũ ,
PP
P
P
where γ = σy2 /σu2 . The optimization (29) can be solved in closed form with
respect to λ:
n
o
1
(30)
λ∗ (β) = trace Kβ−1 P (k) + m(k) m(k) T .
n
With this, the update of β is given by
β (k+1) = arg max n log λ∗ (β) + log det Kβ ,
(31)
β∈(0, 1)
8
which can be solved with scalar optimization methods, or grid search. Once we
have β (k+1) , we also have the update for λ:
λ(k+1) = λ∗ β (k+1) .
(32)
Appealing to theory of the EM-method, we have the following result:
Theorem 1. The sequences {w(k) }, {λ(k) }, and {β (k) } generated by the iterations (28) and (29) are such that:
L(ỹ, ũ; w(k+1) , λ(k+1) , β (k+1) ) > L(ỹ, ũ; w(k) , λ(k) , β (k) ).
(33)
where L(ỹ, ũ; w, λ, β) is the marginal likelihood of the data; and
L(ỹ, ũ; w(k) , λ(k) , β (k) ) → L∗
(34)
as k → ∞, where L∗ is a local extremum of L(ỹ, ũ; w, λ, β).
Proof. By construction, the iterations (28) and (29) are iterations in an EM
algorithm. The E-step function (22), seen as the function of two variables
Q(x; y) is continuous in x and y. So the sequence generated satisfy the conditions
in Theorem 1 in [29] and (33) and (34) follow.
Interestingly, except for pathological cases, the EM-method is guaranteed
to converge to a local maximum of the marginal likelihood (see [28, Ch. 3] for
details).
Corollary 1. If the sequences {w(k) }, {λ(k) }, and {β (k) } generated by the iterations (28) and (29) are such that:
kw(k+1) − w(k) k22 + kβ (k+1) − β (k) k22 + kλ(k+1) − λ(k) k22 → 0,
as k → ∞, then they converge to a stationary point of L(ỹ, ũ; w, λ, β).
Proof. Follows directly from Theorem 6 in [29].
Remark 2. Theorem 1 gives a natural stopping criterion for the EM algorithm.
When the increase in the likelihood between two iterates is below a certain threshold, (approximate) convergence to a maximimum is safely guaranteed. Corollary 1 further guarantees the convergence of the parameters to a local maximizer
when the change between iterations is infinitesimal.
4.1
Kernel-based output smoothing
To solve Problem 3 we first observe that, since the output noise is white, the
output smoothing problem is a simulation problem, and the smoothed output
signal is given by the convolution W g. After solving Problem 1 and Problem 2,
we can find an estimate of the smoothed output signal v̂ by plugging in the
estimates ŵ and ĝ in the convolution, obtaining
v̂ = Ŵ ĝ.
9
(35)
5
Some remarks on identifiability
It is well known (see, e.g. [30], [8]) that, in general, errors-in-variables problems
are not identifiable. Different models may explain the same observed data and
therefore it is impossible to assess the validity of a certain model from the data.
In the case of Gaussian noise, where only second moments carry information
about the distributions, any attempt to identify the noise variances, the system,
and the input samples is bound to fail. In our EM framework, this follows from
the shape of the likelihood (20): for instance leaving free both σu2 and w, we can
choose w = ũ and L(ỹ, ũ, g; w, λ, β) can be made arbitrarily large by choosing
a small enough σu2 . Various additional assumptions can be posed to circumvent
the non-identifiability issue, see [2]. In our setup, if we know the ratio γ = σy2 /σu2
we can estimate the unknown variances by adding the following equation to the
iterations of the EM method:
σy2 (k+1) =
Py M (k)w(k+1) + w(k+1) T A(k)w(k+1)
ỹ T ỹ − 2ỹ T
T
+γ
T
ũ ũ − 2ũ
Pu w
Nu + Ny
+ w(k+1) T
Nu + Ny
(k+1)
PTu Puw(k+1) ,
(36)
σu2 (k+1) = σy2 (k+1) /γ.
In the case of missing data we have other identifiability problems, in addition
to the ones inherited from errors-in-variables. The possibility of multiple models
explaining the available data is linked to aliasing, as the missing data can be
seen as data decimation [31]. In order to have a unique solution to the likelihood
problem
maximize log p(ỹ|g; w) + log p(ũ; w) + log p(g; λ, β),
(37)
w, λ, β
where g is the true impulse response, we need that the symmetric matrix
1 T
G
σy2
PTy Py G+ σ12 PTu Pu
(38)
u
is invertible. This is the case as long as the effect of every missing input sample
is visible at least once in the output:
Proposition 1. There is a unique solution to (37) if and only if for every
missing input sample time τu ∈
/ {tui }, there is a k ∈ {0 . . . n} and a τy ∈ {tyi }
such that gk 6= 0 and k + τ u = τy .
P
Proof. Matrix (38) is invertible iff there is no α such that y Gα
P = 0 and
α = 0 means that α can be written as i∈{t
αi ei
u α = 0. The condition
/ u
k}
N
where ei are vectors in the canonical
basis
of
and
a
are
scalars.
The
i
P
i
condition y Gα 6= 0 translates into i∈{t
a
S
g
=
6
0,
where
S
is
defined
u
/ k} i y
in (27). This concludes the proof.
P
P
R
P
P
g
10
6
Simulations
To evaluate the performance of the proposed method, we perform a set of Monte
Carlo (MC) simulations. In the MC simulations, we identify the impulse responses of 500 systems from the dataset D1 described in [32]. For each system
in the dataset, we generate N = 210 input and output samples; the input is
Gaussian white noise with variance equal to 1. The output measurements are
affected by Gaussian white noise of variance equal 0.1, namely 10% of the noiseless output variance. The variance of the noise affecting the input varies with
the experiment.
We use the iterative method presented in Section 4 to estimate the first n =
100 samples of the impulse response. The noise variance is updated iteratively
with (36). The iterations are initialized at w(0) = u, λ(0) = 10, β (0) = 0.6. The
noise variances σy2 and σu2 are initialized, respectively, at the sample variance of
2 (0)
the least squares residuals and at σy /γ. The iterations are stopped when the
relative change of the parameter updates is below 1%.
We evaluate the goodness of fit using the standard score
fit(a, aref ) = 1 −
ka − aref k2
,
karef − mean(aref )k2
(39)
where aref is a true value and a its estimate. We calculate the median fit of the
estimated impulse responses, inputs, and outputs over the dataset.
We consider two different scenarios. In the first scenario, we corrupt the
dataset with increasing fractions of missing samples. In the second scenario, we
corrupt the dataset with input noises of increasing variance. Table 1 gives a
summary of the experimental conditions.
Scenario A: Missing data
The input noise is Gaussian white noise with variance 0.1 (10% of the input
signal variance). Before performing the identification, we randomly select and
remove a fraction of the available data: in Exp. 1, we remove from 0% to 50%
of the input samples, in 10% increments; in Exp. 2, we remove from 0% to 50%
of the output samples, in 10% increments; in Exp. 3, we remove equal fractions
of input and output, between 0% and 25%, in 5% increments. The results are
plotted in Figure 2. Interestingly, a large fraction of missing input samples has
A (Exp. 1)
A (Exp. 2)
A (Exp. 3)
B
σy2
σu2
missing input
missing output
10%
10%
10%
10%
10%
10%
10%
0% ÷ 100%
0% ÷ 50%
0%
0% ÷ 25%
0%
0%
0% ÷ 50%
0% ÷ 25%
0%
Table 1: Experimental conditions in the simulation scenarios
11
severe effect on the performance, whereas a large fraction of missing output
samples has a milder effect on the identification performance. In Exp. 1 and
Exp. 2, the model has always resulted identifiable, whereas in Exp. 3 a number
of systems were non-identifiable. The results are collected in Table 2.
Scenario A
fit(ĝ, g)
0.8
0.7
0.6
fit(ŵ, w)
0.5
0%
0.8
10%
15%
20%
25%
30%
35%
40%
45%
50%
5%
10%
15%
20%
25%
30%
35%
40%
45%
50%
20% 25% 30% 35%
Total missing samples
40%
45%
50%
0.7
0.6
0.5
0%
0.8
fit(v̂, v)
5%
0.7
0.6
0.5
0%
Exp. 1
Exp. 2
Exp. 3
5%
10%
15%
Figure 2: Plot of the median fit of the impulse response (top), the smoothed
input (middle) and the smoothed output (bottom) over 500 MC runs, for increasing fractions of missing samples; In Exp.1 we remove input samples, in
Exp.2 we remove output samples, in Exp.3 we remove input and output samples.
12
Scenario B: Input noise
The input noise is Gaussian white noise. We consider values of the input noise
variance between 0 (no noise) and 1 (same variance as the input), in increments
of 0.2. The results are plotted in Figure 3. In this scenario, we compare the
performance of the proposed method with a kernel-based identification method
that does not account for input noise. We estimate the impulse response using
the posterior mean m from (17), with λ, β and σy2 estimated trough marginal
likelihood, with all instances of w replaced by the noisy measurements u.
Scenario B
fit(ĝ, g)
0.8
0.6
0.4
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Input noise variance [a.u.]
0.8
0.9
1
Figure 3: Plot of the median fit of the impulse response estimate over 500 MC
runs, for increasing values of the input noise variance. We compare the proposed
estimator (blue) with performance of an estimator that does not account for
input noise (black).
7
Conclusions
In this paper we have presented a nonparametric kernel-based method for the
identification of the impulse response of a linear systems in the presence of noisy
and missing input-output data. The method relies upon a Gaussian regression
framework, where the impulse response of the system is modeled as a Gaussian process with a suitable covariance matrix. Using an EB approach, we find
the minimum mean-squared estimate of the impulse response. This estimate
depends on the unknown noiseless input, as well as on the kernel hyperparameters and the noise variances. These quantities are estimated from the marginal
Unsolvable problems
0
3
5
16
27
40
Total missing samples
0%
10%
20%
30%
40%
50%
Table 2: Number of non-identifiable systems in Exp. 3 (out of 500)
13
likelihood of the data, obtained integrating out the impulse response. We have
devised an iterative scheme that solves the marginal likelihood maximization in
simple updates, and we have discussed the convergence properties of the algorithm. We have tested the method on a data bank of linear systems, where we
have analyzed the degradation in performance for increasing amounts of missing
data, and increasing noise variance on the input measurements. We have briefly
addressed the question of identifiability; and simulations seem to validate our
theoretical results, however, this aspect needs further study.
References
[1] L. Ljung, System Identification, Theory for the User. Prentice Hall, 1999.
[2] T. Söderström, “Errors-in-variables methods in system identification,” Automatica, vol. 43, no. 6, pp. 939–958, 2007.
[3] R. Frisch, “Statistical confluence analysis by means of complete regression
systems (university institute of economics, oslo, 1934, pp. 5–8),” in The
Foundations of Econometric Analysis, pp. 271–273, Cambridge University
Press, 1934.
[4] B. D. Anderson, “Identification of scalar errors-in-variables models with
dynamics,” Automatica, vol. 21, no. 6, pp. 709–716, 1985.
[5] K. V. Fernando and H. Nicholson, “Identification of linear systems with
input and output noise: the koopmans-levin method,” IEE Proc. D. Control
Theory Appl., vol. 132, no. 1, pp. 30–36, 1985.
[6] W.-X. Zheng and C.-B. Feng, “Unbiased parameter estimation of linear
systems in the presence of input and output noise,” Int. J. Adapt. Control
Signal Process., vol. 3, no. 3, pp. 231–251, 1989.
[7] S. Beghelli, R. Guidorzi, and U. Soverini, “The frisch scheme in dynamic
system identification,” Automatica, vol. 26, no. 1, pp. 171–176, 1990.
[8] T. Söderström, U. Soverini, and K. Mahata, “Perspectives on errors-invariables estimation for dynamic systems,” Signal Process., vol. 82, no. 8,
pp. 1139–1154, 2002.
[9] D. Fan and G. Luo, “Frisch scheme identification for errors-in-variables
systems,” in Proc. IEEE Int. Conf. Cogn. Infor. (ICCI), pp. 794 – 799,
2010.
[10] L. Ning, T. T. Georgiou, A. Tannenbaum, and S. P. Boyd, “Linear models
based on noisy data and the frisch scheme,” SIAM Rev., vol. 57, no. 2,
pp. 167–197, 2015.
14
[11] E. Zhang and R. Pintelon, “Errors-in-variables identification of dynamic
systems in general cases,” in Proc. IFAC Symp. System Identification
(SYSID), vol. 48, pp. 309–313, 2015.
[12] T. Söderström, “Identification of stochastic linear systems in presence of
input noise,” Automatica, vol. 17, no. 5, pp. 713–725, 1981.
[13] R. Diversi, R. Guidorzi, and U. Soverini, “Maximum likelihood identification of noisy input–output models,” Automatica, vol. 43, no. 3, pp. 464–472,
2007.
[14] J. Schoukens, R. Pintelon, G. Vandersteen, and P. Guillaume, “Frequencydomain system identification using non-parametric noise models estimated
from a small number of data sets,” Automatica, vol. 33, no. 6, pp. 1073–
1086, 1997.
[15] T. Söderström, M. Hong, J. Schoukens, and R. Pintelon, “Accuracy analysis of time domain maximum likelihood method and sample maximum
likelihood method for errors-in-variables and output error identification,”
Automatica, vol. 46, no. 4, pp. 721–727, 2010.
[16] A. Isaksson, “Identification of arx-models subject to missing data,” IEEE
Trans. Autom. Control, vol. 38, no. 5, pp. 813–819, 1993.
[17] R. Pintelon and J. Schoukens, “Frequency domain system identification
with missing data,” IEEE Trans. Autom. Control, vol. 45, no. 2, pp. 364–
369, 2000.
[18] R. Wallin and A. Hansson, “Maximum likelihood estimation of linear SISO
models subject to missing output data and missing input data,” Int. J.
Control, pp. 1–11, 2014.
[19] I. Markovsky and R. Pintelon, “Identification of linear time-invariant systems from multiple experiments,” IEEE Trans. Signal Process., vol. 63,
no. 13, pp. 3549–3554, 2015.
[20] E. Zhang, R. Pintelon, and J. Schoukens, “Errors-in-variables identification
of dynamic systems excited by arbitrary non-white input,” Automatica,
vol. 49, no. 10, pp. 3032–3041, 2013.
[21] Z. Liu, A. Hansson, and L. Vandenberghe, “Nuclear norm system identification with missing inputs and outputs,” Syst. Control Lett., vol. 62, no. 8,
pp. 605–612, 2013.
[22] I. Markovsky and K. Usevich, “Structured low-rank approximation with
missing data,” SIAM J. Matrix Anal. & Appl., vol. 34, no. 2, pp. 814–830,
2013.
[23] C. Williams and C. Rasmussen, Gaussian processes for machine learning.
2006.
15
[24] G. Pillonetto and G. De Nicolao, “A new kernel-based approach for linear
system identification,” Automatica, vol. 46, no. 1, pp. 81–93, 2010.
[25] G. Pillonetto, F. Dinuzzo, T. Chen, G. D. Nicolao, and L. Ljung, “Kernel
methods in system identification, machine learning and function estimation:
A survey,” Automatica, vol. 50, pp. 657–682, mar 2014.
[26] J. Maritz and T. Lwin, Empirical bayes methods. Chapman and Hall London, 1989.
[27] G. Pillonetto and A. Chiuso, “A Bayesian learning approach to linear system identification with missing data,” in Proc. IEEE Conf. Decis. Control
(CDC), pp. 4698–4703, 2009.
[28] G. McLachlan and T. Krishnan, The EM algorithm and extensions, vol. 382.
John Wiley and Sons, 2007.
[29] C. F. J. Wu, “On the convergence properties of the EM algorithm,” Ann.
Statist., vol. 11, pp. 95–103, mar 1983.
[30] G. Bottegal, G. Picci, and S. Pinzoni, “On the identifiability of errorsin-variables models with white measurement errors,” Automatica, vol. 47,
no. 3, pp. 545–551, 2011.
[31] R. Wallin and A. Isaksson, “Multiple optima in identification of arx models
subject to missing data,” EURASIP J. Adv. Signal Process., vol. 2002,
no. 1, pp. 1–8, 2002.
[32] T. Chen, L. Ljung, M. Andersen, A. Chiuso, F. Carli, and G. Pillonetto,
“Sparse multiple kernels for impulse response estimation with majorization minimization algorithms,” in Proc. IEEE Conf. Decis. Control (CDC),
pp. 1500–1505, 2012.
16
| 3cs.SY
|
Under review as a conference paper at ICLR 2017
R ENDER GAN: G ENERATING R EALISTIC L ABELED
DATA
Leon Sixt, Benjamin Wild, & Tim Landgraf
Fachbereich Mathematik und Informatik
Freie Universität Berlin
Berlin, Germany
arXiv:1611.01331v5 [cs.NE] 12 Jan 2017
{leon.sixt, benjamin.wild, tim.landgraf}@fu-berlin.de
A BSTRACT
Deep Convolutional Neuronal Networks (DCNNs) are showing remarkable performance on many computer vision tasks. Due to their large parameter space,
they require many labeled samples when trained in a supervised setting. The costs
of annotating data manually can render the use of DCNNs infeasible. We present a
novel framework called RenderGAN that can generate large amounts of realistic,
labeled images by combining a 3D model and the Generative Adversarial Network
framework. In our approach, image augmentations (e.g. lighting, background, and
detail) are learned from unlabeled data such that the generated images are strikingly realistic while preserving the labels known from the 3D model. We apply
the RenderGAN framework to generate images of barcode-like markers that are
attached to honeybees. Training a DCNN on data generated by the RenderGAN
yields considerably better performance than training it on various baselines.
1
I NTRODUCTION
When an image is taken from a real world scene, many factors determine the final appearance: background, lighting, object shape, position and orientation of the object, the noise of the camera sensor,
and more. In computer vision, high-level information such as class, shape, or pose is reconstructed
from raw image data. Most real-world applications require the reconstruction to be invariant to
noise, background, and lighting changes.
In recent years, deep convolutional neural networks (DCNNs) advanced to the state of the art in many
computer vision tasks (Krizhevsky et al., 2012; He et al., 2015; Razavian et al., 2014). More training
data usually increases the performance of DCNNs. While image data is mostly abundant, labels for
supervised training must often be created manually – a time-consuming and tedious activity. For
complex annotations such as human joint angles, camera viewpoint or image segmentation, the
costs of labeling can be prohibitive.
In this paper, we propose a method to drastically reduce the costs of labeling such that we can train
a model to predict even complex sets of labels. We present a generative model that can sample
from the joint distribution of labels and data. The training procedure of our model does not require
any manual labeling. We show that the generated data is of high quality and can be used to train a
model in a supervised setting, i.e. a model that maps from real samples to labels, without using any
manually labeled samples.
We propose two modifications to the recently introduced GAN framework (Goodfellow et al., 2014).
First, a simple 3D model is embedded into the generator network to produce samples from corresponding input labels. Second, the generator learns to add missing image characteristics to the
model output using a number of parameterized augmentation functions. In the adversarial training
we leverage large amounts of unlabeled image data to learn the particular form of blur, lighting,
background and image detail. By constraining the augmentation functions we ensure that the resulting image still represents the given set of labels. The resulting images are hard to distinguish from
real samples and can be used to train a DCNN to predict the labels from real input data.
1
Under review as a conference paper at ICLR 2017
10
9
8
7
11 0
1
12 2
13 3
4
6 5
(b) Tagged bees in the hive
(a) Tag structure
Figure 1: (a) The tag represents a unique binary code (cell 0 to 11) and encodes the orientation with
the semicircles 12 and 13. The red arrow points towards the head of the bee. This tag encodes the
id 100110100010. (b) Cutout from a high-resolution image.
The RenderGAN framework was developed to solve the scarcity of labeled data in the BeesBook
project (Wario et al., 2015) in which we analyze the social behavior of honeybees. A barcode-like
marker is attached to the honeybees’ backs for identification (see Fig. 1). Annotating this data is
tedious, and therefore only a limited amount of labeled data exists. A 3D model (see the upper row
of Fig. 2) generates a simple image of the tag based on position, orientation, and bit configuration.
The RenderGAN then learns from unlabeled data to add lighting, background, and image details.
Training a DCNN on data generated by the RenderGAN yields considerably better performance
compared to various baselines. We furthermore include a previously used computer vision pipeline
in the evaluation. The networks’ detections are used as feature to track the honeybees over time.
When we use detections from the DCNN instead of the computer vision pipeline, the accuracy of
assigning the true id increases from 55% to 96%.
Our contributions are as follows. We present an extension of the GAN framework that allows to sample from the joint distribution of data and labels. The generated samples are nearly indistinguishable
from real data for a human observer and can be used to train a DCNN end-to-end to classify real
samples. In a real-world use case, our approach significantly outperforms several baselines. Our
approach requires no manual labeling. The simple 3D model is the only form of supervision.
2
R ELATED WORK
There exists multiple approaches to reduce the costs associated with labeling.
A common approach to deal with limited amount of labels is data augmentation (Goodfellow et al.,
2016, Chapter 7.4). Translation, noise, and other deformations can often be applied without changing the labels, thereby effectively increasing the number of training samples and reducing overfitting.
DCNNs learn a hierarchy of features – many of which are applicable to related domains (Yosinski
et al., 2014). Therefore, a common technique is to pre-train a model on a larger dataset such as
ImageNet (Deng et al., 2009) and then fine-tune its parameters to the task at hand (Girshick et al.,
2014; Long et al., 2015; Razavian et al., 2014). This technique only works in cases where a large
enough related dataset exists. Furthermore, labeling enough data for fine-tuning might still be costly.
If a basic model of the data exists (for example, a 3D model of the human body), it can be used to
generate labeled data. Peng et al. (2015) generated training data for a DCNN with 3D-CAD models.
Figure 2: First row: Images from the 3D model without augmentation. Below: Corresponding
images from the RenderGAN. Last row: Real images of bee’s tags.
2
Under review as a conference paper at ICLR 2017
real data
D
noise
G
0-1
fake
Figure 3: Topology of a GAN. The discriminator network D is trained to distinguish between ”fake”
and real data. The generator network G receives a random vector as input. G is optimized to
maximize the chance of the discriminator making a mistake.
Su et al. (2015) used 3D-CAD models from large online repositories to generate large amounts of
training images for the viewpoint estimation task on the PASCAL 3D+ dataset (Xiang et al., 2014).
Massa et al. (2015) are matching natural images to 3D-CAD models with features extracted from
a DCNN. Richter et al. (2016) and Ros et al. (2016) used 3D game engines to collect labeled data
for image segmentation. However, the explicit modeling of the image acquisition physics (scene
lighting, reflections, lense distortions, sensor noise, etc.) is cumbersome and might still not be able
to fully reproduce the particularities of the imaging process such as unstructured background or
object specific noise. Training a DCNN on generated data that misses certain features will result in
overfitting and poor performance on the real data.
Generative Adversarial Networks (GAN) (see Fig. 3) can learn to generate high-quality samples
(Goodfellow et al., 2014), i.e. sample from the data distribution p(x). Denton et al. (2015) synthesized images with a GAN on the CIFAR dataset (Krizhevsky, 2009), which were hard for humans to
distinguish from real images. While a GAN implicitly learns a meaningful latent embedding of the
data (Radford et al., 2015), there is no simple relationship between the latent dimensions and the labels of interest. Therefore, high-level information can’t be inferred from generated samples. cGANs
are an extension of GANs to sample from a conditional distribution given some labels, i.e. p(x|l).
However, training cGANs requires a labeled dataset. Springenberg (2015) showed that GANs can
be used in a semi-supervised setting but restricted their analysis to categorical labels. Wang &
Gupta (2016) trained two separate GANs, one to model the object normals and another one for the
texture conditioned on the normals. As they rely on conditional GANs, they need large amounts
of labeled data. Chen et al. (2016) used an information theoretic to disentangle the representation.
They decomposed the representation into a structured and unstructured part. And successfully related on a qualitative level the structured part to high-level concepts such as camera viewpoint or
hair style. However, explicitly controlling the relationship between the latent space and generated
samples without using labeled data is an open problem, i.e. sampling from p(x, l) without requiring
labels for training.
3
R ENDER GAN
Most supervised learning tasks can be modeled as a regression problem, i.e. approximating a function fˆ : Rn 7→ L that maps from data space R to label space L. We consider fˆ to be the best
available function on this particular task. Analogous to ground truth data, one could call fˆ the
ground truth function.
In the RenderGAN framework, we aim to solve the inverse problem to this regression task: generate
data given the labels. This is achieved by embedding a simple 3D model into the generator of a
GAN. The samples generated by the simple model must correspond to the given labels but may lack
many factors of the real data such as background or lightning. Through a cascade of augmentation
functions, the generator can adapt the images from the 3D model to match the real data.
We formalize image augmentation as a function φ(x, d), which modifies the image x based on the
augmentation parameter d (a tensor of any rank). The augmentation must preserve the labels of the
image x. Therefore, it must hold for all images x and all augmentations parameters d:
fˆ (φ(x, d)) = fˆ(x)
3
(1)
Under review as a conference paper at ICLR 2017
real data
D
M
Gl
noise
...
φ0
G0
φk
0-1
fake
Gk
G
Figure 4: The generator G cannot directly produce samples. Instead, G has to predict parameters
Gl for the 3D model M . The image generated by M is then modified through the augmentation
functions φi parameterized by Gi to match the real data.
The augmentation function must furthermore be differentiable w.r.t. x and d as the gradient will be
back-propagated through φ to the generator. Image augmentations such as lighting, surrounding,
and noise do preserve the labels and fit this definition. We will provide appropriate definitions of φ
for the mentioned augmentations in the following section.
If appropriate augmentation functions are found that can model the missing factors and are differentiable, we can use the GAN framework to find parameters that result in realistic output images.
Multiple augmentation functions can be combined to perform a more complex augmentation. Here,
we will consider multiple augmentation functions applied sequentially, i.e. we have k augmentation
functions φi and k corresponding outputs Gi from the generator. The output of the previous augmentation function is the input to the next one. Thus, we can write the generator given some labels
l as:
g(z, l) = φk (φk−1 (. . . φ0 (M (l), G0 (z)) . . . , Gk−1 (z)), Gk (z))
(2)
where M is the 3D model. We can furthermore learn the label distribution with the generator. As the
discriminator loss must be backpropagated through the 3D model M, it must be differentiable. This
can be achieved by emulating the 3D model with a neural network (Dosovitskiy et al., 2015). The
resulting generator g(z) can be written as (see Fig. 4 for a visual interpretation):
g(z) = φk (φk−1 (. . . φ0 (M (Gl (z)), G0 (z)) . . . , Gk−1 (z)), Gk (z))
(3)
As any differentiable function approximator can be employed in the GAN framework, the theoretical
properties still hold. The training is carried out as in the conventional GAN framework. In a real
application, the augmentation functions might restrict the generator from converging to the data
distribution.
If the training converges, we can collect generated realistic data with g(z) and the high-level information captured in the 3D model with Gl (z). We can now train a supervised learning algorithm on
the labeled generated data (Gl (z) , g (z)) and solve the regression task of approximating fˆ without
depending on manual labels.
4
A PPLICATION TO THE B EES B OOK PROJECT
In the BeesBook project, we aim to understand the complex social behavior of honey bees. For
identification, a tag with a binary code is attached to the back of the bees.
The orientations in space, position, and bits of the tags are required to track the bees over time.
Decoding this information is not trivial: the bees are oriented in any direction in space. The tag
might be partially occluded. Moreover, spotlights on the tag can sometimes even confuse humans.
A previously used computer vision pipeline did not perform reliably. Although we invested a substantial amount of time on labeling, a DCNN trained on this data did not perform significantly better
(see Sec. 3). We therefore wanted to synthesize labeled images which are realistic enough to train
an improved decoder network.
Following the idea outlined in section 3, we created a simple 3D model of a bee marker. The 3D
model comprises a mesh which represents the structure of the marker and a simple camera model to
project the mesh to the image plane. The model is parameterized by its position, its pitch, yaw and
roll, and its ID. Given a parameter set, we obtain a marker image, a background segmentation mask
4
Under review as a conference paper at ICLR 2017
outputs:
zbits
3D model
φblur
φlighting
φbg
φdetail
0.90
zoffset
G
Figure 5: Augmentation functions of the RenderGAN applied to the BeesBook project. The arrows
from G to the augmentation functions φ depict the inputs to the augmentation functions. The generator provides the position and orientations to the 3D model, whereas the bits are sampled uniformly.
On top, the output of each stage is shown. The output of φdetail is forwarded to the discriminator.
and a depth map. The generated images lack many important factors: blur, lighting, background,
and image detail (see Fig. 2). A DCNN trained on this data does not generalize well (see Sec. 5).
Over the last years we collected a large amount of unlabeled image data. We successfully augmented
the 3D model using this dataset, as described below.
We trained a neural network to emulate the 3D model. Its outputs are indistinguishable from the
images of the 3D model. The discriminator error can now be backpropagated through the 3D model
which allows the generator to also learn the distributions of positions and orientations of the bee
marker. The IDs are sampled uniformly during training. The weights of the 3D model network are
fixed during the RenderGAN training.
We apply different augmentation functions that account for blur, lighting, background, and image
detail. The output of the 3D model and of each augmentation function is of shape (64, 64) and in
the range [−1, 1]. In Fig. 5, the structure of the generator is shown.
Blurriness: The 3D model produces hard edges, but the images of the real tags show a broad range
of blur. The generator produces a scalar α ∈ [0, 1] per image that controls the blur.
φblur (x, α) = (1 − α) (x − bσ (x)) + bσ (x)
(4)
where bσ (x) = x ∗ kσ denotes convolving the image x with a Gaussian kernel kσ of scale σ. The
implementation of the blur function is inspired by Laplacian pyramids (Burt & Adelson, 1983).
As required for augmentation functions, the labels are preserved, because we limit the maximum
amount of blur by picking σ = 2. φblur is also differentiable w.r.t the inputs α and x.
Lighting of the tag: The images from the 3D model are binary. In real images, tags exhibit different
shades of gray. We model the lighting by a smooth scaling and shifting of the pixel intensities. The
generator provides three outputs for the lighting: scaling of black parts sb , scaling of white parts
sw and a shift t. All outputs have the same dimensions as the image x. An important invariant is
that the black bits of the tag must stay darker than the white bits. Otherwise, a bit could flip, and
the label would change. By restricting the scaling sw and sb to be between 0.10 and 1, we ensure
that this invariant holds. The lighting is locally corrolated and should cause smooth changes in the
image. Hence, Gaussian blur b(x) is applied to sb , sw , and t.
φlighting (x, sw , sb , t) = x · b(sw ) · W (x) + x · b(sb ) · (1 − W (x)) + b(t)
(5)
The segmentation mask W (x) is one for white parts and zero for the black part of the image. As the
intensity of the input is distributed around -1 and 1, we can use thresholding to differentiate between
black and white parts.
Background: The background augmentation can change the background pixels arbitrarily. A segmentation mask Bx marks the background pixels of the image x which are replaced by the pixels
from the generated image d.
φbg (x, d) = x · (1 − Bx ) + d · Bx
5
(6)
Under review as a conference paper at ICLR 2017
The 3D model provides the segmentation mask. As φbg can only change background pixels, the
labels remain unchanged.
Details: In this stage, the generator can add small details to the whole image including the tag. The
output of the generator d is passed through a high-pass filter to ensure that the added details are
small enough not to flip a bit. Furthermore, d is restricted to be in [−2, 2] to make sure the generator
cannot avoid the highpass filter by producing huge values. With the range [−2, 2], the generator has
the possibility to change black pixels to white, which is needed to model spotlights.
φdetail (x, d) = x + highpass(d)
(7)
The high-pass is implemented by taking the difference between the image and a blurred version of
the image (σ = 3.5). As the spotlights on the tags are only a little smaller than the bits, we increase
its slope after the cutoff frequency by repeating the high-pass filter three times.
The image augmentations are applied in the order as listed above: φdetail ◦ φbackground ◦ φlighting ◦
φblur . Please note that there exist parameters to the augmentation functions that could change the
labels. As long as it is guaranteed that such augmentations will result in unrealistic looking images,
the generator network will learn to avoid them. For example, even though the detail augmentation
could be used to add high-frequency noise to obscure the tag, this artifact would be detected by the
discriminator.
Architecture of the generator: The generator network has to produce outputs for each augmentation function. We will outline only the most important parts. See our code available online for all the
details of the networks1 . The generator starts with a small network consisting of dense layers, which
predicts the parameters for the 3D model (position, orientations). The output of another dense layer
is reshaped and used as starting block for a chain of convolution and upsampling layers. We found
it advantageous to merge a depth map of the 3D model into the generator as especially the lighting depends on the orientation of the tag in space. The input to the blur augmentation is predicted
by reducing an intermediate convolutional feature map to a single scalar. An additional network is
branched off to predict the input to the lighting augmentation. For the background generation, the
output of the lighting network is merged back into the main generator network together with the
actual image from the 3D model.
For the discriminator architecture, we mostly rely on the architecture given by Radford et al. (2015),
but doubled the number of convolutional layers and added a final dense layer. This change improved
the quality of the generated images.
Clip layer: Some of the augmentation parameters have to be restricted to a range of values to ensure
that the labels remain valid. The training did not converge when using functions like tanh or sigmoid
due to vanishing gradients. We are using a combination of clipping and activity regularization
to keep the output in a given interval [a, b]. If the input x is out of bounds, it is clipped and a
regularization loss r depending on the distance between x and the appropriate bound is added.
γ||x − a||1 if x < a
r(x) = 0
(8)
if a ≤ x ≤ b
γ||x − b||1 if x > b
f (x) = min(max(a, x), b)
(9)
With the scalar γ, the weight of the loss can be adapted. For us γ = 15 worked well. If γ is chosen
too small, the regularization loss might not be big enough to move the output of the previous layer
towards the interval [a, b]. During training, we observe that the loss decreases to a small value but
never vanishes.
Training: We train generator and discriminator as in the normal GAN setting. We use 2.4M unlabeled images of tags to train the RenderGAN. We use Adam (Kingma & Ba, 2014) as an optimizer
with a starting learning rate of 0.0002, which we reduce in epoch 200, 250, and 300 by a factor of
0.25. In Fig. 6b the training loss of the GAN is shown. The GAN does not converge to the point
where the discriminator can no longer separate generated from real samples. The augmentation
functions might restrict the generator too much such that it cannot model certain properties. Nevertheless, it is hard for a human to distinguish the generated from real images. In some cases, the
1
https://github.com/berleon/deepdecoder
6
Under review as a conference paper at ICLR 2017
16
real
fake
D
G
Loss
12
8
4
0.0
0.2
0.4
0.6
0.8
0
1.0
50 100 150 200 250 300 350
Epoch
(a)
(b)
Figure 6: (a) Histogram of the discriminator scores on fake and real samples. (b) Losses of the
generator (G) and discriminator (D).
Figure 7: Random points in the z-space given the tag parameters
generator creates unrealistic high-frequencies artifacts. The discriminator unfailingly assigns a low
score to theses images. We can therefore discard them for the training of the supervised algorithm.
More generated images are shown in Appendix A. In Fig. 7, we show random points in the latent
space, while fixing the tag parameters. The generator indeed learned to model the various lighting
conditions, noise intensities, and backgrounds.
5
R ESULTS
We constructed the RenderGAN to generate labeled data. But does a DCNN trained with the RenderGAN data perform better than one trained on the limited amounts of real data? And are learned
augmentations indeed needed or do simple hand-designed augmentation achieve the same result?
The following paragraphs describe the different datasets used in the evaluation. We focus on the
performance of a DCNN on the generated data. Thus, we do not compare our method to conventional GANs as those do not provide labels and are generally hard to evaluate.
Data from the RenderGAN: We generate 5 million tags with the RenderGAN framework. Due to
the abundance, one training sample is only used twice during training. It is not further augmented.
Real Data: The labels of the real data are extracted from ground truth data that was originally
collected to evaluate bee trajectories. This ground truth data contains the path and id of each bee
over multiple consecutive frames. Data from five different time spans was annotated – in total 66K
tags. As the data is correlated in time (same ids, similar lighting conditions), we assign the data from
one time span completely to either the train or test set. The data from three time spans forms the train
set (40K). The test set (26K) contains data from the remaining two time spans. The ground truth data
lacks the orientation of the tags, which is therefore omitted for this evaluation. Due to the smaller
Table 1: Datasets created with learned representations and hand-designed augmentations
Name
HM 3D
HM LI
HM BG
Learned
3D model
3D model, blur, lighting
3D model, blur, lighting, background
7
Hand-Designed
blur, lighting, background, noise, spotlights
background, noise, spotlights
noise, spotlights
Under review as a conference paper at ICLR 2017
Table 2: Comparison of the mean Hamming distance (MHD) on the different data sets. More
samples of the training data can be found in Appendix D.
Data
MHD
Real
0.956
HM 3D
0.820
HM LI
0.491
HM BG
0.505
RenderGAN
0.424
RenderGAN
+ Real
CV
Training Data
0.416
1.08
size of the real training set, we augment it with random translation, rotation, shear transformation,
histogram scaling, and noise (see Appendix C for exact parameters).
RenderGAN + Real: We also train a DCNN on generated and real data which is mixed at a 50:50
ratio.
Handmade augmentations: We tried to emulate the augmentations learned by the RenderGAN by
hand. For example, we generate the background by an image pyramid where the pixel intensities
are drawn randomly. We model all effects, i.e. blur, lighting, background, noise and spotlights (see
Appendix B for details on their implementation). We apply the handmade augmentation to different
learned representations of the RenderGAN, e.g. we use the learned lighting representation and add
the remaining effects such as background and noise with handmade augmentations (HM LI). See
Table 1 for the different combinations of learned representations and hand designed augmentations.
Computer vision pipeline CV : The previously used computer vision pipeline (Wario et al., 2015) is
based on manual feature extraction. For example, a modified Hough transformation to find ellipses.
The MHD obtained by this model is only a rough estimate given that the computer vision pipeline
had to be evaluated and fine-tuned on the same data set due to label scarcity.
Training setup: An epoch consists of 1000 batches á 128 samples. We use early stopping to select
the best parameters of the networks. For the training with generated data, we use the real training
set as the validation set. When training on real data, the test set is also used for validation. We could
alternatively reduce the real training set further and form an extra validation set, but this would harm
the performance of the DCNN trained on the real data. We use the 34-layer ResNet architecture
(He et al., 2015) but start with 16 feature maps instead of 64. The DCNNs are evaluated on the
mean Hamming distance (MHD) i.e. the expected value of bits decoded wrong. Human experts can
decode tags with a MHD of around 0.23.
Results: In Table 2, we present the results of the evaluation. The training losses of the networks
are plotted in Fig. 8. The model trained with the data generated by the RenderGAN has an MHD of
0.424. The performance can furthermore be slightly improved by combining the generated with real
data. The small gap in performance when adding real data is a further indicator of the quality of the
generated samples.
If we use predictions from this DCNN instead of the computer vision pipeline, the accuracy of the
tracking improves from 55% of the ids assigned correctly to 96%. At this quality, it is possible to
analyze the social behavior of the honeybees reliably.
Compared to the handmade augmentations (HM 3D), data from the RenderGAN leads to considerably better performance. The large gap in performance between the HM 3D and HM LI data
highlights the importance of the learned lighting augmentation.
8
Under review as a conference paper at ICLR 2017
0.5
Loss
0.4
0.3
0.2
0.1
0.0
0
10
20
30
40
50
60
70
80
Real train
Real val
HM 3D train
HM 3D val
HM LI train
HM LI val
HM BG train
HM BG val
RenderGAN train
RenderGAN val
RenderGAN + Real train
RenderGAN + Real val
Epoch
Figure 8: Training and validation losses of DCNNs trained on different data sets. As some data sets
are missing the orientation of the tags, only the loss of the bits are plotted. Binary crossentropy is
used as loss for the bits. The train and validation loss of each dataset have the same color.
6
D ISCUSSION
We proposed a novel extension to the GAN framework that is capable of rendering samples from
a basic 3D model more realistic. Compared to computer graphics pipelines, the RenderGAN can
learn complex effects from unlabeled data that would be otherwise hard to model with explicit rules.
Contrary to conventional GANs, the generator provides explicit information about the synthesized
images, which can be used as labels for a supervised algorithm. The training of the RenderGAN
requires no labels.
We showed an application of the RenderGAN framework to the BeesBook project, in which the
generator adds blur, lighting, background, and details to images from a basic 3D model. The generated data looks strikingly real and includes fine details such as spotlights, compression artifacts, and
sensor noise.
In contrast to previous work that applied 3D models to produce training samples for DCNNs (Su
et al., 2015; Richter et al., 2016; Ros et al., 2016), we were able to train a DCNN from scratch with
only generated data that still generalizes to unseen real data.
While some work is required to adapt the RenderGAN to a specific domain, once set up, arbitrary
amounts of labeled data can be acquired cheaply, even if the data distribution changes. For example,
if the tag design changes to include more bits, small adaptions to the 3D model’s source code and
eventually the hyperparameters of the augmentation functions would be sufficient. However, if we
had labeled the data by hand, then we would have to annotate data again.
While the proposed augmentations represent common image characteristics, a disadvantage of the
RenderGAN framework is that these augmentation functions must be carefully customized for the
application at hand to ensure that high-level information is preserved. Furthermore, a suitable 3D
model must be available.
7
F UTURE W ORK
For future work, it would be interesting to see the RenderGAN framework used on other tasks where
basic 3D models exist e.g. human faces, pose estimation, or viewpoint prediction. In this context,
one could come up with different augmentation functions e.g. colorization, affine transformations,
or diffeomorphism. The RenderGAN could be especially valuable to domains where pre-trained
models are not available or when the annotations are very complex. Another direction of future work
might be to extend the RenderGAN framework to other fields. For example, in speech synthesis,
one could use an existing software as a basic model and improve the realism of the output with a
similar approach as in the RenderGAN framework.
9
Under review as a conference paper at ICLR 2017
R EFERENCES
Peter Burt and Edward Adelson. The laplacian pyramid as a compact image code. IEEE Transactions on communications, 31(4):532–540, 1983.
Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets.
arXiv preprint arXiv:1606.03657, 2016.
Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale
hierarchical image database. In Computer Vision and Pattern Recognition, 2009. CVPR 2009.
IEEE Conference on, pp. 248–255. IEEE, 2009.
Emily L Denton, Soumith Chintala, Rob Fergus, et al. Deep generative image models using a
laplacian pyramid of adversarial networks. In Advances in neural information processing systems,
pp. 1486–1494, 2015.
Alexey Dosovitskiy, Jost Tobias Springenberg, and Thomas Brox. Learning to generate chairs with
convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and
Pattern Recognition, pp. 1538–1546, 2015.
Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pp. 580–587, 2014.
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems, pp. 2672–2680, 2014.
Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. Book in preparation for MIT
Press, 2016. URL http://www.deeplearningbook.org.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. arXiv preprint arXiv:1512.03385, 2015.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Alex Krizhevsky. Learning Multiple Layers of Features from Tiny Images, 2009. ISSN 1098-6596.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pp. 1097–1105,
2012.
Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully Convolutional Networks for Semantic
Segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
pp. 3431–3440, 2015. ISSN 10636919. doi: 10.1109/CVPR.2015.7298965.
Francisco Massa, Bryan Russell, and Mathieu Aubry. Deep exemplar 2d-3d detection by adapting
from real to rendered views. arXiv preprint arXiv:1512.02497, 2015.
Xingchao Peng, Baochen Sun, Karim Ali, and Kate Saenko. Learning Deep Object Detectors from
3D Models. Iccv, 2015. doi: 10.1109/ICCV.2015.151.
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks, 2015.
Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, and Stefan Carlsson. Cnn features offthe-shelf: an astounding baseline for recognition. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition Workshops, pp. 806–813, 2014.
Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth
from computer games. In European Conference on Computer Vision, pp. 102–118. Springer,
2016.
10
Under review as a conference paper at ICLR 2017
German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The
synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes.
In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 3234–
3243, 2016.
Jost Tobias Springenberg. Unsupervised and semi-supervised learning with categorical generative
adversarial networks. arXiv preprint arXiv:1511.06390, 2015.
Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Proceedings of the IEEE
International Conference on Computer Vision, pp. 2686–2694, 2015.
Xiaolong Wang and Abhinav Gupta. Generative image modeling using style and structure adversarial networks. arXiv preprint arXiv:1603.05631, 2016.
Fernando Wario, Benjamin Wild, Margaret Jane Couvillon, Raúl Rojas, and Tim Landgraf. Automatic methods for long-term tracking and the detection and decoding of communication dances
in honeybees. Frontiers in Ecology and Evolution, 3:103, 2015.
Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond pascal: A benchmark for 3d object
detection in the wild. In IEEE Winter Conference on Applications of Computer Vision, pp. 75–82.
IEEE, 2014.
Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep
neural networks? In Advances in neural information processing systems, pp. 3320–3328, 2014.
11
Under review as a conference paper at ICLR 2017
Appendices
A
G ENERATED I MAGES
(a) Generated images
(b) Real images
Figure 9: Continuum visualization on the basis of the discriminator score: Most realistc scored samples top left corner to least realistc bottom right corner. Images with artifacts are scored unrealistic
and are not used for training.
12
Under review as a conference paper at ICLR 2017
Figure 10: Images generated with the generator given a fixed bit configuration
Figure 11: Correspondence of generated images and 3D model
13
Under review as a conference paper at ICLR 2017
B
H ANDMADE AUGMENTATIONS
We constructed augmentations for blur, lighting, background, noise and spotlights manually. For
synthesizing lighting, background and noise, we use image pyramids, i.e. a set of images L0 , . . . , L6
of size (2i × 2i ) for 0 ≤ i ≤ 6. Each level Li in the pyramid is weighted by a scalar ωi . Each pixel
of the different level Li is drawn from N (0, 1). The generated image I6 is given by:
I0 = ω0 L0
Ii = ωi Li + upscale(Ii−1 )
(10)
(11)
, where upscale doubles the image dimensions. The pyramid enables us to generate random images
while controlling their frequency domain by weighting the pyramid levels appropriately.
• Blur: Gaussian blur with randomly sampled scale.
• Lighting: Similar as in the RenderGAN. Here, the scaling of the white and black parts
and shifting is constructed with image pyramids.
• Background: image pyramids with the lower levels weight more.
• Noise: image pyramids with only the last two layer.
• Spotlights: overlay with possible multiple 2D Gauss function with a random position on
the tag and random covariance.
We selected all parameters manually by comparing the generated to real images. However, using
slightly more unrealistic images resulted in better performance of the DCNN trained with the HM
3D data. The parameters of the handmade augmentations can be found online in our source code
repository.
C
AUGMENTATIONS OF THE REAL DATA
We scale and shift the pixel intensities randomly, i.e. sI +t, where I is the image and s, t are scalars.
The noise is sampled for each pixel from N (0, ), where ∼ max(0, N (µn , σn )) is drawn for each
image separately. Different affine transformations (rotation, scale, translation, and shear) are used.
Table 3: Parameters of the augmentation of the real data
Name
Intensity Scale (s)
Intensity Shift (t)
Noise Mean (µn )
Noise Std (σn )
Rotation
Scale
Shear
Translation
14
Distribution
unif(0.9, 1.1)
unif(-0.2, 0.2)
0.04
0.03
unif(0, 2π)
unif(0.7, 1.1)
unif(-0.3, 0.3)
unif(-4, 4)
Under review as a conference paper at ICLR 2017
D
T RAINING SAMPLES
(a) Real trainings samples
(b) HM 3D training samples
(c) HM LI training samples
(d) HM BG training samples
(e) RenderGAN
Figure 12: Training samples from the different datasets.
15
| 9cs.NE
|
arXiv:1607.04030v3 [math.GT] 26 Feb 2018
THE EXTENDED MAPPING CLASS GROUP CAN BE
GENERATED BY TWO TORSIONS
XIAOMING DU
Abstract. Let Sg be the closed oriented surface of genus g and let
Mod± (Sg ) be the extended mapping class group of Sg . When the genus
is at least 5, we prove that Mod± (Sg ) can be generated by two torsion
elements. One of these generators is of order 2, and the other one is of
order 4g + 2.
1. Introduction
Let Sg be the closed oriented surface of genus g. The extended mapping
class group Mod± (Sg ) is defined as Homeo± (Sg )/Homeo0 (Sg ), the group of
homotopy classes of homeomorphisms (including orientation-preserving ones
and orientation-reversing ones) of Sg , and the mapping class group Mod(Sg )
is defined by Homeo+ (Sg )/Homeo0 (Sg ), the group of orientation-preserving
homotopy classes of homeomorphisms of Sg .
For Mod(Sg ), Dehn and Lickorish found independently Dehn twist generating sets of Mod(Sg ) [3, 11]. Humphries reduced the number of Dehn twist
generators to the lowest bound [5]. Wajnryb in [17] found that the minimal
number of the generators (not only Dehn twist generators) for the mapping
class groups is 2.
We are also interested in the torsion generating set. McCarthy and Papadopoulos in [15] proved that Mod(Sg ) can be generated by infinitely many
elements of order 2 when g ≥ 3. Luo showed that Mod(Sg ) is generated by
12g + 6 elements of order 2 when g ≥ 3 [12]. Brendle and Farb reduced
the number of the involution generators to 6 [1] and prove Mod(Sg ) can be
generated by three torsion elements. Kassabov reduced the number of the
involution generators to 4 [8] for g ≥ 7. Korkmaz in [9] proved Mod(Sg )
can be generated by two torsion elements of order 4g + 2. Monden in [14]
proved Mod(Sg ) can be generated by 3 torsion elements of order 3.
2010 Mathematics Subject Classification. 57N05, 57M20, 20F38.
Key words and phrases. mapping class group, generator, torsion.
This research is supported by NSFC (Grant No. 11401219). The author wish to thank
BICMR for its hospitality and thank Boju Jiang who asked me about the problem of
finding lower bounds of the orders and the numbers of torsion generators for the mapping
class groups. The author also wish to thank Dingsen Yan who teach me how to draw
pictures with the metapost package of the latex.
1
2
XIAOMING DU
For the extended mapping class group Mod± (Sg ), Dehn-Nielsen-Baer theorem [4] says that Mod± (Sg ) is isomorphic to Out(π1 (Sg )), the outer automorphism group of π1 (Sg ). Ivanov proved that Mod± (Sg ) is the automorphism group of the curve complex [6]. Brock and Margalit showed that
Mod± (Sg ) is the automorphism group of the pants complex and it is also
the isometry group of the Teichmüller space under the Weil-Petersson metric
([2],[13]). Korkmaz showed that Mod± (Sg ) can be generated by 2 elements,
one of which is a Dehn twist [9]. Stukow in [16] proved Mod± (Sg ) is generated by 3 elements of order 2.
It is an open problem if Mod± (Sg ) can be generated by two torsion elements (See [10], Problem 5.3). In this paper, under the condition that the
genus of the surface is at least 5, we answer this question affirmatively:
Theorem 1.1. For g ≥ 5, the extended mapping class group Mod± (Sg ) can
be generated by 2 torsion elements. One of these generators is of order 2
and the other one is of order 4g + 2.
2. Preliminaries
Notations:
(a) We use the convention of functional notation, namely, elements of the
mapping class group are applied right to left, i.e. the composition F G means
that G is applied first.
(b) A Dehn twist means a right-hand Dehn twist.
(c) We denote the curves by lower case letters a, b, c, d (possibly with subscripts) and the Dehn twists about them by the corresponding capital letters
A, B, C, D. Notationally we do not distinguish a diffeomorphism/curve and
its isotopy class.
We recall the following results (see, for instance, section 3.3, 5.1, 7.5 of
[4]):
Lemma 2.1. For any ϕ ∈ Mod(Sg ) and any isotopy classes a, b of simple
closed curves in Sg satisfying ϕ(a) = b, we have:
B = ϕ A ϕ−1 .
Lemma 2.2. For any ϕ ∈ Mod± (Sg ) \ Mod(Sg ) and any isotopy classes a, b
of simple closed curves in Sg satisfying ϕ(a) = b, we have:
B −1 = ϕ A ϕ−1 .
Lemma 2.3. Let a, b be two simple closed curves on Sg . If a is disjoint
from b, then
AB = BA.
THE EXTENDED MAPPING CLASS GROUP CAN BE GENERATED BY TWO TORSIONS
3
c
y
x
d
a
b
z
Figure 1
Lemma 2.4 (Lantern relation). Let a, b, c, d, x, y, z be the curves showed in
Figure 1 on a genus zero surface with four boundaries. Then
ABCD = XY Z.
In other words, since a, b, c are disjoint from x, y, z, we have
D = (XA−1 )(Y B −1 )(ZC −1 ).
The elements having the form of U V −1 for some disjoint simple closed
curves u, v play a crucial role in our proof. The proof of the main result
relies on Humphries’ theorem:
Theorem 2.5 (Humphries). Let a1 , a2 , . . . a2g , b be the curves as on the lefthand side of Figure 2. Then the mapping class group Mod(Sg ) is generated
by Ai ’s and B.
a1
b
a3
a4
a2
a3
a4
a5
a2 a1
a 2g
b
σ
a6
a 2g
a 2g
a1
a2
a4
a3
Figure 2
Consider the genus g surface as a regular (4g+2)-gon whose corresponding
opposite sides are glued together, as indicated on the right-hand side of
Figure 2. We redraw the curves of Humphries’ generating set on the righthand side of Figure 2 as follow. The set of curves ai ’s on the left-hand side
of Figure 2 is a chain of simple closed curves and fills the surface. We find
a chain of simple closed curves which also fills the surface on the right-hand
side of Figure 2, identify them with ai ’s. To see the corresponding curve of
b, notice that b intersects a4 once and is disjoint from other ai ’s. Then find
such a curve on the right-hand side of Figure 2.
Look at the (4g+2)-gon. There is a natural 2π/(4g+2) rotation preserving
the gluing way of the (4g + 2)-gon. This rotation induces a period map σ
4
XIAOMING DU
of the genus g surface. Moreover, σ(ai ) = ai+1 for 1 ≤ i ≤ 2g − 1. Take
a0 = σ(a2g ). Then we have a1 = σ(a0 ). Under modulo 2g + 1, we have
ak = σ k (a0 ) for all integer k and ai+1 = σ(ai ) still holds for all i’s. Similarly,
since σ is of order 4g+2, under modulo 4g+2, we take b0 = b and bk = σ k (b).
Then we have σ(bj ) = bj+1 for all j’s.
To see back the image of bj ’s on the left-hand side of Figure 2, since the
set of ak ’s form a chain of simple closed curves filling the surface, we only
need to calculate the geometric intersection number i(bj , ak ) for each k on
the right-hand side of Figure 2, then find a curve with the same geometric
intersection numbers with ak ’s as on the left-hand side of Figure 2. Figure
3 shows this way to think of the images of bj ’s. Figure 2 and Figure 3 will
be used in the proof of Theorem 3.1 to verify the disjointness between some
ai and bj .
b2
b1
b2g-4
b2g+2 b2g+4
b4g-4 b4g-2
b2g+3
b4g-3
b2g-5 b2g-3
b3
b4g
b2g-2
b4g-1
b2g-1
b0
b2g
b4g+1
b2g+1
Figure 3
3. The main result and the proof
Let τ be the symmetry of the (4g + 2)-gon as on the left-hand side of
Figure 4. Then τ induces an orientation-reversing homeomorphism of the
genus g surface. The fixed point set of τ forms a non-separating curve. A
part of such a non-separating curve is along the axis of the symmetry. We
still denote the homotopy class of such a homeomorphism as τ . Since τ
preserves the curve b, it is easy to check that (τ ◦ B)2 = Id.
THE EXTENDED MAPPING CLASS GROUP CAN BE GENERATED BY TWO TORSIONS
5
E
´
´
´
Figure 4
Theorem 3.1. Let σ be the element of order 4g + 2 and τ ◦B be the element
of order 2 as we described above. Then for g ≥ 5 we have Mod± (Sg ) =
hσ, τ ◦ Bi.
Remark 3.2. This generating set can also apply to Mod± (Sg,1 ), i.e., the
extended mapping class group of the surface with one marked point. This is
because the element σ and τ ◦ B fix the center point of the 4g + 2-gon.
Remark 3.3. From Theorem 3.1 we can also deduce that Mod± (Sg ) can
be generated by 3 symmetries when g ≥ 5. The reason is that σ is like the
rotation of the (4g + 2)-gon. In the dihedral group of the (4g + 2)-gon, the
rotation is the product of two reflections (see the right-hand side of Figure
4). Such reflections of (4g + 2)-gon induce orientation reversing order 2
maps τ, τ2 on the surface. Hence Mod± (Sg ) = hτ ◦ B, τ, τ2 i. So we can get
another proof of the result in [16] under the condition g ≥ 5. Our generating
set is different from the generating set in [16]. The fixed-point sets of two
of the symmetries in [16] are separating curves. The fixed-point sets of our
symmetries are non-separating curves.
Proof of Theorem 3.1. Denote the subgroup generated by σ and τ ◦ B as G.
We prove that G = Mod± (Sg ) in four steps:
Step 1. Under modulo 4g + 2, the following two conditions are equivalent:
(1) integers i, k satisfy k ∈ {4, 5, 6, . . . , 4g − 2} \ {2g − 2, 2g, 2g + 2, 2g + 4};
−1
and
(2) bi is disjoint from bi+k . Under such conditions we have Bi Bi+k
−1
Bi Bi+k are in G.
Step 2. Under modulo 2g + 1 for m and modulo 4g + 2 for n, the following
two conditions are equivalent: (1) integers m, n satisfy m 6∈ {n, n + 4}; (2)
am is disjoint from bn . Under such conditions we have Am Bn−1 and A−1
m Bn
are in G.
Step 3. Using the lantern relation, prove that for all k, Ak ∈ G.
Step 4. Finally, G = Mod± (Sg ).
The proof of Step 1:
It is obvious that b0 is disjoint from bk if and only if k ∈ {4, 5, 6, . . . , 4g −
2} \ {2g − 2, 2g, 2g + 2, 2g + 4} (see Figure 2 and Figure 3). So do bi and
6
XIAOMING DU
bi+k . Under such conditions, we first prove that B0 Bk−1 ∈ G. Consider
the element σ k (τ ◦ B) σ k (τ ◦ B). In the dihedral subgroup of Mod± (Sg ),
σ k τ σ k τ is the identity. After adding the Dehn twist B, since bk is disjoint
from b0 , we can easily check that σ k (τ ◦ B) σ k (τ ◦ B) = B0 Bk−1 . For every
−1
integer i, conjugate B0 Bk−1 by σ i . Then we have Bi Bi+k
is in G. The
−1
commutativity of Bi and Bi+k promise Bi Bi+k is in G.
The proof of Step 2:
It is obvious that the disjointness between ai and b0 is equivalent to i 6∈
{0, 4}. So am is disjoint from bn is equivalent to m 6∈ {n, n + 4}. From
these conditions, we also see that ai is disjoint from b0 if and only if a4−i is
disjoint from b0 .
By the conjugacy relation, for some i, k, m, n satisfying the conditions in
Step 1 and Step 2, if there exists ϕ ∈ G and a pair of disjoint curves (bi , bi+k )
−1
such that ϕ : (bi , bi+k ) 7→ (am , bn ), since Bi Bi+k
is in G, Am Bn−1 is also in
G. We need to find such a ϕ ∈ G.
The proofs in odd genus case and in even genus case are slightly different.
Suppose first that the genus g is odd. See Figure 5.
-1
-1
(B2g+2 B2g-4 ) (B2g+2 B2g-3 )
b2g+2
b4g+1
c1
b2g+2
-1
-1
... (B2g-3 B-1
(B2g-3 B 6 ) (B2g-3 B-1
2g-8) (B2g-3 B 2g-7)
7)
c2
c1 b2g-3
-1
B0 B2g-3
-1
(B3 B 2g-3 ) (B2 B-1
2g-3 )
a3
b2g-3
c3
b2g-3
b2g-3
Figure 5
−1
By Step 1, B2g+2 B4g+1
is in G. If g ≥ 5, b2g+2 is disjoint from b2g−3 , b2g−4 .
−1
−1
−1
Both B2g+2 B2g−4 and B2g+2 B2g−3
are in G. The element (B2g+2 B2g−4
) ·
−1
(B2g+2 B2g−3 ) maps the pair of curves (b2g+2 , b4g+1 ) to the pair of curves
−1
(b2g+2 , c1 ) as in Figure 5 (2). Hence B2g+2 C1−1 is in G. Now B2g−3 B2g+2
is
−1
also in G. So B2g−3 C1 is in G.
The curve b2g−3 is disjoint from b2g−7 , b2g−8 , . . . , b7 , b6 . The elements
−1
−1
(B2g−3 B6−1 ), (B2g−3 B7−1 ) . . . (B2g−3 B2g−8
), (B2g−3 B2g−7
) are in G. Then
THE EXTENDED MAPPING CLASS GROUP CAN BE GENERATED BY TWO TORSIONS
7
−1
−1
their product (B2g−3 B6−1 ) · (B2g−3 B7−1 ) . . . (B2g−3 B2g−8
) · (B2g−3 B2g−7
)
maps the pair of curves (b2g−3 , c1 ) to the pair of curves (b2g−3 , c2 ) as in
Figure 5 (4). We have B2g−3 C2−1 is in G.
The curve b2g−3 is disjoint from b3 and b2 . The elements (B2g−3 B2−1 ),
−1
−1
(B2g−3 B3−1 ) are in G. Their product (B3 B2g−3
) · (B2 B2g−3
) maps the pair of
curves (b2g−3 , c2 ) to the pair of curves (b2g−3 , a3 ). So we get that B2g−3 A−1
3
−1
is in G. Its inverse A3 B2g−3
is also in G.
Suppose now that g is even. See Figure 6.
-1
-1
(B 2g-3 B4-1)(B 2g-3 B5-1) ... (B 2g-3 B2g-8
)(B 2g-3 B2g-7)
c2
c1 b2g-3
b2g-3
-1
B0 B2g-3
a4
b2g-3
Figure 6
In this case, the proof of B2g−3 C1−1 ∈ G is the same as the odd genus case.
−1
−1
The element (B2g−3 B4−1 ) · (B2g−3 B5−1 ) . . . (B2g−3 B2g−8
) · (B2g−3 B2g−7
)
maps the pair of curves (b2g−3 , c1 ) to the pair of curves (b2g−3 , c2 ) as in
−1
Figure 6 (2). The element B0 B2g−3
maps the pair of curves (b2g−3 , c2 ) to
the pair of curves (b2g−3 , a4 ). So we get that B2g−3 A−1
4 is in G. Its inverse
−1
A4 B2g−3 is also in G.
The element σ 2g+5 maps (a3 , b2g−3 ) to (a2g+8 , b0 ) when g is odd and maps
(a4 , b2g−3 ) to (a2g+9 , b0 ) when g is even. We have Ai B0−1 = B0−1 Ai ∈ G for
some i where ai is disjoint from b0 .
Notice that τ σ k τ maps ai to ai−k , hence Ai−k = (τ σ k τ ) Ai (τ σ k τ )−1 .
So as long as ai−k is disjoint from b0 , we have
(τ ◦ B0 ) σ k (τ ◦ B0 ) (Ai B0−1 ) σ k = (B0−1 ◦ τ ) σ k (τ ◦ Ai ) σ k
= B0−1 (τ σ k τ ) Ai σ k = B0−1 (Ai−k τ σ k τ ) σ k = B0−1 Ai−k = Ai−k B0−1 .
Taking all possible k and conjugating by σ n , we have all Am Bn−1 ∈ G for
m 6∈ {n, n + 4}. The commutativity of Am and Bn promises A−1
m Bn is in G.
The proof of Step 3:
See Figure 7. There is a natural lantern lying on the surface, bounded
by a1 , a3 , a5 and f . By the lantern relation, we have B0 B2 E = A1 A3 A5 F ,
−1
−1 ), where e and f are the curves showed in
or A1 = (B0 A−1
3 )(B2 A5 )(EF
−1
−1
Figure 7. The elements B0 A−1
3 and B2 A5 are the inverses of A3 B0 and
8
XIAOMING DU
A5 B2−1 respectively. By the result of step 2, they are in G. We only need to
prove EF −1 is also in G.
b
a1
a1
f
b
a3
f
e
a5
a3
a5
b g-
b g-
Figure 7
When g ≥ 5, b2g−3 is disjoint from b0 , . . . , b3 , a1 , . . . , a6 , e, f . Notice
−1
−1
EF −1 = (EB2g−3
)(B2g−3 F −1 ). We verify EB2g−3
and B2g−3 F −1 are in G.
−1
−1
−1
By the previous steps, The element (B2g−3 B3 ) · (A6 B2g−3
) · (A5 B2g−3
)
−1
· (A4 B2g−3
) is in G and maps the pair of curves (b2g−3 , b0 ) to (b2g−3 , f ).
−1
Notice B2g−3 B0−1 is in G. So B2g−3 F −1 is in G. The element (A2 B2g−3
)
−1
−1
−1
· (A1 B2g−3 ) · (A4 B2g−3 ) · (B1 B2g−3 ) maps the pair of curves (b2g−3 , a5 )
−1 is in G. We now get
to (b2g−3 , e). Notice B2g−3 A−1
5 is in G. So B2g−3 E
−1
−1
−1
EF = (EB2g−3 )(B2g−3 F ) is in G.
The proof of Step 4:
The fact that the elements A1 and A1 B0−1 are in G means that B0 ∈ G.
Now all the curves ai ’s are in the same orbit of σ. So do bj ’s. So all the Ai ’s
and Bj ’s are in G. These include Humphries’ generators of Mod(Sg ). So
G contains Mod(Sg ). Finally, one of the generator τ ◦ B is an orientationreversing mapping class. Hence G = Mod± (Sg ).
Remark 3.4. The idea of Step 2 and Step 3 in the above proof is based
on the method in [9] to prove Mod(Sg ) is generated by two torsions of order
4g + 2.
References
[1] T. E. Brendle and B. Farb. Every mapping class group is generated by 3
torsion elements and by 6 involutions. J. Algebra 278 (2004), 187–198.
[2] J. Brock and D. Margalit. Weil-Petersson isometries via the pants complex.
Proc. Amer. Math. Soc., 135(3):795–803 (electronic), 2007.
[3] M. Dehn. Papers on group theory and topology (Springer-Verlag, New York,
1987) (Die Gruppe der Abbildungsklassen, Acta Math. Vol. 69 (1938), 135–
206).
[4] B. Farb and D. Marglait. A Primer on Mapping Class Groups. Princeton
Math. Ser., (Princeton University Press, 2012), 623–658.
[5] S. P. Humphries. Generators for the mapping class group. Topology of lowdimensional manifolds. Proc. Second Sussex Conf. Chelwood Gate 1977 Lecture Notes in Math. 722 (Springer, 1979), 44–47.
[6] N. V. Ivanov. Automorphism of complexes of curves and of Teichmuller spaces.
Internat. Math. Res. Notices, (14):651–666, 1997.
THE EXTENDED MAPPING CLASS GROUP CAN BE GENERATED BY TWO TORSIONS
9
[7] D. Johnson. The structure of Torelli group I: A finite set of generators for I.
Ann. of Math. 118 (1983), 423–442.
[8] M. Kassabov. Generating Mapping Class Groups by Involutions.
arXiv:math.GT/0311455 v1 25 Nov 2003.
[9] M. Korkmaz. Generating the surface mapping class group by two elements.
Trans. Amer. Math. Soc. 357 (2005), 3299–3310.
[10] M. Korkmaz. Minimal generating sets for the mapping class group of a surface.
Handbook of Teichmüller theory. Volume III, 441–463, IRMA Lect. Math.
Theor. Phys., 17, Eur. Math. Soc., Zürich, 2012.
[11] W. B. R. Lickorish. A finite set of generators for the homeotopy group of a
2-manifold. Proc. Camb. Phils. Soc. 60 (1964), 769–778.
[12] F. Luo. Torsion Elements in the Mapping Class Group of a Surface.
arXiv:math.GT/0004048 v1 8 Apr 2000.
[13] D. Margalit. Automorphisms of the pants complex. Duke Math. J.,
121(3):457–479, 2004.
[14] N. Monden. Generating the mapping class group by torsion elements of small
order. Mathematical Proceedings of the Cambridge Philosophical Society, 154
(2013), pp 41–62, doi:10.1017/S0305004112000357.
[15] J. D. McCarthy and A. Papadopoulos. Involutions in surface mapping class
groups.Enseign. Math.(2) 33 (1987), no. 3–4, 275–290. 456.
[16] M. Stukow. The extended mapping class group is generated by 3 symmetries.
C. R. Math. Acad. Sci. Paris 338 (2004), no. 5, 403–406.
[17] B. Wajnryb. Mapping class group of a surface is generated by two elements.
Topology, 35(2), 377–383, 1996.
South China University of Technology, Guangzhou 510640, P.R.China
E-mail address: [email protected]
| 4math.GR
|
STUDY OF JOINT MMSE CONSENSUS AND RELAY SELECTION ALGORITHMS
FOR DISTRIBUTED BEAMFORMING
arXiv:1707.00953v1 [cs.IT] 28 May 2017
Hang Ruan * and Rodrigo C. de Lamare*#
*Department of Electronics, The University of York, England, YO10 5BB
#
CETUC, Pontifical Catholic University of Rio de Janeiro, Brazil
Emails: [email protected], [email protected] ∗
ABSTRACT
This work presents joint minimum mean-square error (MMSE)
consensus algorithm and relay selection algorithms for distributed
beamforming. We propose joint MMSE consensus relay and selection schemes with a total power constraint and local communications
among the relays for a network with cooperating sensors. We also
devise greedy relay selection algorithms based on the MMSE consensus approach that optimize the network performance. Simulation
results show that the proposed scheme and algorithms outperform
existing techniques for distributed beamforming.
Index Terms— Distributed beamforming, relay selection, consensus algorithms.
1. INTRODUCTION
Distributed beamforming has been widely investigated in wireless
communications in recent years [1, 2, 3, 15]. It is key for situations
in which the channels between the sources and the destination have
poor quality so that devices cannot communicate directly and the
destination relies on relays that receive and forward the signals [2].
The work in [3] formulates an optimization problem that maximizes
the output signal-to-interference-plus-noise ratio (SINR) under the
individual relay power constraints. The approach in [7] proposes an
MMSE consensus cooperative relay networking scheme to exchange
data among all the relays under a total power constraint, which limits the total power of all relays regardless of the power allocation.
While local communications among the relays are enabled, the ability to mitigate fading effects in wireless channels of the network can
be improved [8]. Further earlier works in [12] and [13] explored local communications, while avoiding network centralized processing,
which is not desirable and always comes along with the use of total
power constraints [7].
However, in most scenarios relays are either not ideally distributed in terms of locations or the channels involved with some of
the relays have poor quality. Possible solutions can be categorized in
two approaches. One is to adaptively adjust the power of each relay
according to the qualities of its associated channels, known as adaptive power control or power allocation. Some power control methods
based on channel magnitude and relative analysis has been studied
in [6, 14]. An alternative solution is to use relay selection, which
selects a number of relays according to a criterion of interest while
discarding the remaining relays. In [4], multi-relay selections algorithm have been developed to maximize the secondary receiver in
a two-hop cognitive relay network. In [8], several optimum singlerelay selection schemes and a multi-relay selection scheme using
∗ This
work was supported in part by The University of York
relay ordering based on maximizing the output SNR under individual relay power constraints are developed and discussed. The work
in [9] proposed a low-cost greedy search method for the uplink of
cooperative direct sequence code-division multiple access systems,
which approaches the performance of an exhaustive search. Other
approaches include the use of subspace techniques [16, 17, 18, 19,
20, 21, 22, 135, 37, 25, 26, 27, 28, 29, 30, 31, 137, 33, 34, 35, 36,
47, 37, 38, 39, 40, 41, 42, 43, 44, 49, 46, 48, 49, 50, 51, 52, 53, 54,
55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 196, 68, 69, 70, 71].
and large sensor arrays [197, 198, 74, 75, 76, 77, 78, 79, 80, 81, 82,
83, 84, 85, 86, 87, 88, 99, 89, 90, 91, 92, 93, 95, 96, 97, 98, 100,
101],[174, 161, 179, 176, 164, 107, 108, 138, 163, 131, 112, 138,
114, 193, 116, 117, 133].
In this work, we propose a joint minimum mean-square error
(MMSE) consensus and relay selection approach and develop iterative greedy search based relay selection algorithms for distributed
beamforming. In the first proposed algorithm we aim to find largest
minimum value of a cost function regarding the desired signal by
enabling only one relay at once, testing through all the relays one
by one, then we select all the disabled relays when the network
has the largest MMSE (LMMSE) value in each iteration, namely,
the SMMSE consensus greedy (LMMSEC-G) relay selection algorithm. In the second proposed approach, we aim to find the smallest
minimum value of the same cost function, by disabling only one relay at once, testing through all the relays one by one, then preserve
all the selected relays when the network has the smallest MMSE
(SMMSE) estimate in each iteration, namely, the the SMMSE consensus greedy (SMMSEC-G) relay selection algorithm. We compare the proposed greedy relay selection algorithms to the exhaustive search and the scenario without relay selection, showing their
excellent output SINR performance which is close to the exhaustive
search approach.
This paper is organized as follows. In Section 2, the system
model is introduced. In Section 3, the joint MMSE consensus and
relay selection scheme is presented. In Section 4, the greedy relay
selection algorithms are proposed. In Section 5, simulations are presented and discussed. Finally, conclusions are drawn in Section 6.
2. SYSTEM MODEL
We consider a wireless communication network consisting of K signal sources (one desired signal with the others as interferers), M distributed single-antenna relays and a destination. It is assumed that
the quality of the channels between the signal sources and the destination is very poor so that direct communications is not possible and
their links are negligible. The M relays receive information transmitted by the signal sources and then retransmit to the destination as
a beamforming procedure, in which a two-step amplify-and-forward
(AF) protocol (as shown in Fig. 1) is considered as required for cooperative communications.
M relays
K signal sources
destination
where γ is the distance based path loss, L is the known path loss at
the destination, d is the distance of interest relative to the destination and ρ is the path loss exponent, which can vary due to different
environments and is typically set within 2 to 5, with a lower value
representing a clear and uncluttered environment which has a slow
attenuation and a higher value describing a cluttered and highly attenuating environment. Shadow fading describes the phenomenon
where objects can obstruct the propagation of the signal attenuating
the signal further, and can be modeled as a random variable with
probability distribution given by
β = 10(
noise n
g
F
noise ν
Fig. 1. System model.
In the first step, the sources transmit the signals to the relays as
x = Fs + ν,
(1)
where s = [s1 , s2 , · · · , sK ] ∈ C1×K are
√ signal sources with zero
mean, [.]T denotes the transpose, sk = Pk s, E[|s|2 ] = 1, Pk is
the transmit power of the kth signal source, k = 1, 2, · · · , K, s is the
information symbol. Without loss of generality we can assume s1 as
the desired signal while the others are treated as interferers. F =
[f1 , f2 , · · · , fK ] ∈ CM ×K is the channel matrix between the signal sources and the relays, fk = [f1,k , f2,k , · · · , fM,k ]T ∈ CM ×1 ,
fm,k denotes the channel between the mth relay and the kth source
(m = 1, 2, · · · , M , k = 1, 2, · · · , K). ν = [ν1 , ν2 , · · · , νM ]T ∈
CM ×1 is the complex Gaussian noise vector at the relays and σν2
is the noise variance at each relay (νm ˜ CN (0, σν2 )). The vector
x ∈ CM ×1 represents the received data at the relays. In the second step, the relays transmit y ∈ CM ×1 which is an amplified and
phase-steered version of x, which can be written as
y = Wx,
(2)
where W = diag[w1 , w2 , · · · , wM ] ∈ CM ×M is a diagonal matrix
whose diagonal entries denote the beamforming weights. The signal
received at the destination is given by
z = gT y + n,
,
(5)
where β is the shadowing parameter, N (0, 1) means the Gaussian
distribution with zero mean and unit variance, σs is the shadowing
spread in dB. The shadowing spread reflects the severity of the attenuation caused by shadowing, and is typically given between 0dB to
9dB. The channels modeled with both path-loss and shadowing are
described by
F = γβF0 ,
(6)
g = γβg0 ,
(7)
where F0 and g0 denote the Rayleigh distributed channels without
path-loss and shadowing [10, 11].
3. PROPOSED JOINT MMSE CONSENSUS AND RELAY
SELECTION
In this section, we detail the proposed joint MMSE consensus and
relay selection scheme for distributed beamforming using an alternating optimization approach in which the relay selection is followed
by MMSE consensus of beamformers. We assume at the mth relay
the MMSE estimate of the desired signal ŝm,1 can be found as
ŝm,1 = φm xm ,
(8)
where
φm = arg minφ E[|s1 − φxm |2 ] = PK
k=1
For convenience we define s̃m,1 =
relay weight as
ŝm,1
E[|ŝm,1 |2 ]
∗
fm,1
P1
|fm,k |2 Pk + σn2
.
and the normalized
wm φm
, so that the total transmission
E[|ŝ
|2 ]
PMm,1
PM
2
2
E[|w
m s̃m,1 | ] =
m=1
m=1 |wm | .
power can
be expressed as
Therefore,
the MMSE consensus optimization associated with a fixed set of relays under a total power constraint is given by
(3)
where z is a scalar, g = [g1 , g2 , · · · , gM ]T ∈ CM ×1 is the complex
Gaussian channel vector between the relays and the destination, n
(n ˜ CN (0, σn2 ), σn2 = σν2 ) is the noise at the destination and z is the
received signal at the destination.
Note that both F and g are modeled as Rayleigh distributed (i.e.,
both the real and imaginary coefficients of the channel parameters
have Gaussian distribution). Using the Rayleigh distribution for the
channels, we also consider distance based large-scale channel propagation effects that include distance-based fading (or path loss) and
shadowing. Distance-based fading represents how a signal is attenuated as a function of the distance and can be highly affected by the
environment. [10, 11] An exponential based path loss model can be
described by
√
L
γ= √ ,
(4)
dρ
σs N (0,1)
)
10
∗
wm
= arg min
wm
M
X
m=1
E[|s1 − gm wm s̃m,1 |2 ]
s.t.
M
X
m=1
(9)
2
|wm | ≤ PT ,
where PT is the maximum allowable total transmit power of all relays. The relay selection problem for the MMSE consensus can be
described as an optimization problem using a total relay transmit
power constraint described by
Sopt = arg min M M SE(S)
α,w
s.t.
M
X
m=1
αm |wm |2 ≤ PT ,
αm ∈ {0, 1}, m = 1, 2, · · · , M
(10)
PM
2
where M M SE(S) =
m=1 αm E[|s1 − gm wm s̃m,1 | ], w =
T
M ×1
[w1 , w2 , · · · , wM ] ∈ C
, Sopt and S are the optimum relay set
of size Mopt , (1 ≤ Mopt ≤ M ) and the original relay set of size M ,
respectively. The vector α = [α1 , α2 , · · · , αM ]T ∈ RM ×1 , αm
(m = 1, · · · , M ) is the relay cooperation parameter vector which
determines if the mth relay will cooperate. The solution of (10) regarding wm indicates the following relationship:
∗
wm
=
∗
αm gm
λ + αm |gm |2
s
|fm,1 |2 P12
,
2
2
k=1 |fm,k | Pk + σn
Table 1. LMMSEC-G relay selection algorithm
Initialize Sopt = S(0), α = 1, w̃m = 1,
λm (0) = 1, τm,q (0) = 1 and compute M M SEo = M M SE(0).
for i = 1, · · · , M − Mmin
step 1:
compute τm,q (i) = τm,q (i − 1) + µτ (um − uq ),
and τm,q (i) = τm,q (i − 1) + µτ (um − uq ).
step 2:
compute the consensus weight using (14).
step 3:
solve the optimization problem (17) and obtain S̄(i).
compute S(i) using S(i − 1)-S̄(i).
compute M M SE(i) using S(i).
compare M M SE(i) to M M SE(i)(i − 1),
if M M SE(i) < M M SE(i − 1)
update Sopt = S(i) and M M SEo = M M SE(i).
update α(i).
else
keep Sopt = S(i − 1) and M M SEo = M M SE(i − 1).
break.
end if.
end for.
(11)
PK
where λ is the Langrange multiplier, which can be determined by
enabling the local communication of the relays with an MMSE consensus approach. To this end, we employ an auxiliary beamforming
weight vector w̃m = [w̃1,m , · · · , w̃M,m ]T and consider the following joint optimization problem:
min αm E[|s1 − gm wm s̃m,1 |2 ]
w̃m ,α
(12)
s.t. ||α||1 ||w̃m ||2 ≤ PT , w̃m = w,
αm ∈ {0, 1}, m = 1, 2, · · · , M
where w = diag(W) ∈ CM ×1 . It is supposed that the mth relay
is connected to a subset of relays denoted by Mm . The second
constraint in (12) can be replaced by w̃m = w̃q , q ∈ Mm so that
(12) is reformulated as
min
2
w̃m ,α
4.1. LMMSEC-G relay selection algorithm
2
We develop the LMMSEC-G algorithm to obtain the solution of
(12), which depends on the relay selection parameter vector α to
find the optimum α. The LMMSEC-G works in an iterative way
and discards only the worst relay to find the optimal relay set in
each iteration. Additionally, the parameter Mmin can be introduced
to restrict the minimum number of relays that must be used. The
LMMSEC-G algorithm finds the largest MMSE at each iteration and
selects the complementary relays, which is described as follows
αm E[|s1 − gm wm s̃m,1 | ] + λm (i)(||α||1 ||w̃m || − PT )
X T
+
τm,q (w̃m − w̃q ),
q∈Mm
(13)
where λm (i) and τm,q are Lagrange multipliers.The proposed algorithmic solution relies on the alternating optimization associated
with relay selection, computation of the optimal weights and Lagrange multipliers at the mth relay as
w̃t,m
r
∗
|fm,1 |2 P12
αm gm
PK
(
2 −
λm (i)+αm |gm |2
|fm,k |2 Pk +σn
k=1
= if t =Pm
αm q∈M τm,q;t
m
−
,
2λm (i)
if t 6= m
P
S̄(i) = arg max M M SE(S(i − 1))
α(i)
q∈Mm τm,q;m
2
(14)
where τm,q;t denotes the tth element of τm,q . The Lagrange multipliers are updated as follows
λm (i) = |λm (i − 1) + µλ (||w̃m ||2 − PT )|,
(15)
τm,q (i) = τm,q (i − 1) + µτ (um − uq ),
(16)
where µλ and µτ are step sizes with small positive values, um =
[|w1,m |, · · · , |wM,m |]T and i is the time index [7].
4. PROPOSED GREEDY RELAY SELECTION
ALGORITHMS
In this section, we detail the algorithms that perform relay selection, develop the LMMSEC-G and SMMSEC-G relay selection algorithms and review the exhaustive search.
),
s.t. ||w̃m (i)||2 =≤ PT , w̃m (i) = w(i),
αm (i) ∈ {0, 1}, m = 1, 2, · · · , M
M − i ≥ Mmin ,
(17)
where S̄(i) denotes the complement of the set S(i) from set S(i−1).
The optimization problem compares all the MMSE values assuming
that only one different single relay is enabled while the others are
disabled. LMMSEC-G cancels the relay with largest MMSE value
from set S(i − 1) and evaluates the MMSE performance of the remaining relays, which is solved only once in each iteration. If the
MMSE in the current iteration is smaller than that in the previous
iteration (i.e. M M SE(i) < M M SE(i − 1)), then the selection
process continues; if M M SE(i) ≥ M M SE(i − 1), we cancel the
selection of the current iteration and keep the relay set S(i − 1) and
M M SE(i − 1). The LMMSEC-G algorithm is shown in Table. 1.
4.2. SMMSEC-G relay selection algorithm
The proposed SMMSEC-G algorithm is an alternative way to find
the solution of (12) regarding α, which aims to find the smallest
MMSE from the remaining relays after disabling a single relay each
time. It is an improved greedy search based method which also
Table 2. SMMSECG
Initialize Sopt = S(0), α = 1, w̃m = 1,
λm (0) = 1, τm,q (0) = 1 and compute M M SEo = M M SE(0).
for i = 1, · · · , M − Mmin
step 1:
compute τm,q (i) = τm,q (i − 1) + µτ (um − uq ),
and τm,q (i) = τm,q (i − 1) + µτ (um − uq ).
step 2:
compute the consensus weight using (14).
step 3:
solve the optimization problem (18) and obtain S(i).
compute M M SE(i) using S(i).
compare M M SE(i) to M M SE(i − 1),
if M M SE(i) < M M SE(i − 1)
update Sopt = S(i) and M M SEo = M M SE(i).
update α(i).
else
keep Sopt = S(i − 1) and M M SEo = M M SE(i − 1).
break.
end if.
end for.
by varying the input SNR or the total number of relays in the network. The parameters used include: number of signal sources K =
3, the path loss exponent ρ = 2, the power path loss from signals to the destination L = 10dB, shadowing spread σs = 3dB,
PT = 1dBW, Mmin = 1. For the local communication between the
relays, we set Mm = {m+1} and M = 1, µλ = µτ = 0.001. 100
repetitions are executed for each of the studied methods. In Fig. 2,
we fixed the total number of relays M = 5 and interference-to-noise
ratio (INR) at 10dB and evaluate the SINR versus SNR performance
of the joint MMSE consensus and relay selection approaches and the
existing techniques. Both the greedy search based methods, namely,
LMMSEC-G and SMMSEC-G, increase the SINR performance as
compared with the case without any relay selection and approach the
exhaustive search especially at low SNRs. Fig. 3 illustrates that with
a fixed SNR(0dB) and INR(0dB) how the output SINR varies when
the total number of relays in the network increases. It is clear that
using more relays enhance the overall network performance and the
SMMSEC-G method performs very close to the exhaustive search.
The proposed techniques could also be evaluated in terms of BER
performance [90, 91, 92, 93, 94, 174, 163, 131, 138, 114, 193, 116,
117, 133, 199].
10
5
0
−5
SINR (dB)
works in iterations but with higher complexity and much better performance. We also consider Mmin as a restriction to the minimum
number of relays that must be used. Before the first iteration all
relays are considered (i.e., S(0) = S). Consequently, we solve
the following problem once for each iteration in order to cancel the
relay with worst performance from set S(i − 1) and evaluate the
M M SE(i) at time instant i:
−20
no relay selection
exhaustive search
LMMSEC−G
SLMMSEC−G
−30
−35
α(i)
s.t. (M − i)||w̃m (i)|| ≤ PT , w̃m (i) = w(i),
αm (i) ∈ {0, 1}, m = 1, 2, · · · , M
M − i ≥ Mmin ,
−15
−25
S(i) = arg min M M SE(S(i − 1))
2
−10
−40
−10
(18)
−5
0
5
SNR (dB)
10
15
20
Fig. 2. SINR performance versus SNR
If the MMSE in the current iteration is lower than that in the previous
iteration (i.e. M M SE(i) < M M SE(i − 1)), then the selection
process continues; if M M SE(i) ≥ M M SE(i − 1), we cancel the
selection of the current iteration and keep the relay set S(i − 1) and
M M SE(i − 1). The SMMSEC-G algorithm is shown in Table. 2.
−10
−12
−14
4.3. Exhaustive Search
SINR (dB)
In an exhaustive search procedure, we test every possible combinations among all the relays, which means the change of status if a
relay is chosen or not will contribute to a different possible combination. To obtain the global optimum solution, we need to run the
consensus algorithm once without iterations. Also, we can predefine Mf ix as the required selected number of relays as an additional
requirement. However, the complexity can be extremely high depending on the number of relays.
−16
−18
−20
−22
no relay selection
exhaustive search
LMMSEC−G
SMMSEC−G
−24
−26
−28
−30
1
2
3
4
5
6
7
8
M
5. SIMULATIONS
In the simulations we focus on the output SINR performance comparison of the proposed LMMSEC-G and SMMSEC-G algorithms
Fig. 3. SINR performance versus M
9
10
6. CONCLUSION
We have proposed a joint MMSE consensus and relay selection approach and developed efficient algorithms for distributed beamforming. We have proposed the LMMSEC-G and SMMSEC-G greedy
optimization algorithms based on the MMSE criterion with known
network quantities and relay selection strategies, which determines
if a relay should cooperate or not in the network. The LMMSEC-G
and SMMSEC-G algorithms have shown excellent performance and
outperformed previously reported techniques.
7. REFERENCES
[1] R. Mudumbai, D.R. Brown III, U. Madhow, and H.V. Poor, “Distributed Transmit Beamforming Challenges and Recent Progress,”
IEEE Communications Magazine, Vol. 47, Issue. 2, pp. 102-110, 2009.
[2] A. B. Gershman, N. D. Sidiropoulos, S. Shahbazpanahi, M. Bengtsson,
and B. Ottersten, “Convex Optimization-Based Beamforming,” IEEE
Signal Processing Magazine, Vol. 27, Issue. 3, pp. 62-75, May 2010.
[3] J. Uher, T. A. Wysocki, and B. J. Wysocki, “Review of Distributed
Beamforming,” Journal of Telecommunications and Information Technology, 2011.
[4] J. Xu, H. Zhang, D. Yuan, Q. Jin, and C. Wang, “Novel Multiple Relay Selection Schemes in Two-Hop Cognitive Relay Networks,” Third
International Conference on Communications and Mobile Computing,
pp 207-310, April 2011.
[5] L. Landau, “Advanced Robust Adaptive Beamforming for Wireless
Networks,” Master Thesis, Technische Universitat Ilmenau Fakultat
fur Elektrotechnik und Informationstechnik, May 2011.
[6] Y. Jing and H. Jafarkhani, “Network Beamforming Using Relays With
Perfect Channel Information,” IEEE Transactions on Information Theory, VOL. 55, NO. 6, pp. 2499-2517, June 2009.
[7] J. Choi, “Distributed Beamforming Using a Consensus Algorithm for
Cooperative Relay Networks,” IEEE Communication Letters, VOL.
15, Issue 4, pp. 368-370, April 2011.
[8] Y. Jing, and H. Jafarkhani, “Single and Multiple Relay Selection
Schemes and their Achievable Diversity Orders,” IEEE Transactions
on Wireless Communications, VOL. 8, NO. 3, pp. 1414-1423, March
2009.
[9] J. Gu and R. C. de Lamare, “Joint PIC and Relay Selection Based
on Greedy Techniques for Cooperative DS-CDMA Systems,” Proc. of
International Conference on Speech Acoustics and Signal Processing,
pp 2754-2758, 4-9 May 2014.
[10] T. Hesketh, R, C. de Lamare and S. Wales, “Joint maximum likelihood
detection and link selection for cooperative MIMO relay systems,” IET
Communications, Vol. 8, Issue 14, pp 2489-2499, Sep 2014.
[11] T. J. Hesketh, “Detection and Resource Allocation Algorithms for Cooperative MIMO Relay Systems,” Ph.D. Thesis, University of York,
Electronics, Communications Research Group, Feb 2014.
[12] B. Johannson, C. M. Carretti, and M. Johansson, “On Distributed Optimization Using peer-to-peer Communications in Wireless Sensor Networks,” Proc. IEEE SECON, pp. 497-505, 2008.
[16] L. L. Scharf and D. W. Tufts, “Rank reduction for modeling
stationary signals,” IEEE Transactions on Acoustics, Speech
and Signal Processing, vol. ASSP-35, pp. 350-355, March
1987.
[17] A. M. Haimovich and Y. Bar-Ness, “An eigenanalysis interference canceler,” IEEE Trans. on Signal Processing, vol. 39, pp.
76-84, Jan. 1991.
[18] D. A. Pados and S. N. Batalama ”Joint space-time auxiliary
vector filtering for DS/CDMA systems with antenna arrays”
IEEE Transactions on Communications, vol. 47, no. 9, pp.
1406 - 1415, 1999.
[19] J. S. Goldstein, I. S. Reed and L. L. Scharf ”A multistage representation of the Wiener filter based on orthogonal projections”
IEEE Transactions on Information Theory, vol. 44, no. 7, 1998.
[20] Y. Hua, M. Nikpour and P. Stoica, ”Optimal reduced rank estimation and filtering,” IEEE Transactions on Signal Processing,
pp. 457-469, Vol. 49, No. 3, March 2001.
[21] M. L. Honig and J. S. Goldstein, “Adaptive reduced-rank interference suppression based on the multistage Wiener filter,”
IEEE Transactions on Communications, vol. 50, no. 6, June
2002.
[22] E. L. Santos and M. D. Zoltowski, “On Low Rank MVDR
Beamforming using the Conjugate Gradient Algorithm”, Proc.
IEEE International Conference on Acoustics, Speech and Signal Processing, 2004.
[23] Q. Haoli and S.N. Batalama, “Data record-based criteria for the
selection of an auxiliary vector estimator of the MMSE/MVDR
filter”, IEEE Transactions on Communications, vol. 51, no. 10,
Oct. 2003, pp. 1700 - 1708.
[24] R. C. de Lamare and R. Sampaio-Neto, “Reduced-Rank Adaptive Filtering Based on Joint Iterative Optimization of Adaptive
Filters”, IEEE Signal Processing Letters, Vol. 14, no. 12, December 2007.
[25] Z. Xu and M.K. Tsatsanis, “Blind adaptive algorithms for minimum variance CDMA receivers,” IEEE Trans. Communications, vol. 49, No. 1, January 2001.
[26] R. C. de Lamare and R. Sampaio-Neto, “Low-Complexity
Variable Step-Size Mechanisms for Stochastic Gradient Algorithms in Minimum Variance CDMA Receivers”, IEEE Trans.
Signal Processing, vol. 54, pp. 2302 - 2317, June 2006.
[27] C. Xu, G. Feng and K. S. Kwak, “A Modified Constrained Constant Modulus Approach to Blind Adaptive Multiuser Detection,” IEEE Trans. Communications, vol. 49, No. 9, 2001.
[28] Z. Xu and P. Liu, “Code-Constrained Blind Detection of
CDMA Signals in Multipath Channels,” IEEE Sig. Proc. Letters, vol. 9, No. 12, December 2002.
[13] M. G. Rabbat, R. D. Nowak, and J. A. Bucklew, “eneralized Consensus
Computation in Networked Systems with Erasure Links,” Proc. IEEE
SPAWC, pp. 1088-1092, 2005.
[29] R. C. de Lamare and R. Sampaio Neto, ”Blind Adaptive CodeConstrained Constant Modulus Algorithms for CDMA Interference Suppression in Multipath Channels”, IEEE Communications Letters, vol 9. no. 4, April, 2005.
[14] S. Schedler and V. Kuehn, “Resource Allocation for Distributed Beamforming with Multiple Relays and Individual Power Constraints,”
Proc. International Sympositum Wireless Communications Systems
(ISWCS), pp 1-5, Aug 2014.
[30] L. Landau, R. C. de Lamare and M. Haardt, “Robust adaptive
beamforming algorithms using the constrained constant modulus criterion,” IET Signal Processing, vol.8, no.5, pp.447-457,
July 2014.
[15] O. T. Demir, T. E. Tuncer, “Optimum Discrete Distributed Beamforming for Single Group Multi Casting Relay Networks with Relay Selection,” ICASSP 2015, pp 2524-2528, April 2015.
[31] R. C. de Lamare, “Adaptive Reduced-Rank LCMV Beamforming Algorithms Based on Joint Iterative Optimisation of Filters”, Electronics Letters, vol. 44, no. 9, 2008.
[32] R. C. de Lamare and R. Sampaio-Neto, “Adaptive ReducedRank Processing Based on Joint and Iterative Interpolation,
Decimation and Filtering”, IEEE Transactions on Signal Processing, vol. 57, no. 7, July 2009, pp. 2503 - 2514.
[33] R. C. de Lamare and Raimundo Sampaio-Neto, “Reducedrank Interference Suppression for DS-CDMA based on Interpolated FIR Filters”, IEEE Communications Letters, vol. 9, no.
3, March 2005.
[34] R. C. de Lamare and R. Sampaio-Neto, “Adaptive ReducedRank MMSE Filtering with Interpolated FIR Filters and Adaptive Interpolators”, IEEE Signal Processing Letters, vol. 12,
no. 3, March, 2005.
[35] R. C. de Lamare and R. Sampaio-Neto, “Adaptive Interference Suppression for DS-CDMA Systems based on Interpolated FIR Filters with Adaptive Interpolators in Multipath
Channels”, IEEE Trans. Vehicular Technology, Vol. 56, no. 6,
September 2007.
[36] R. C. de Lamare, “Adaptive Reduced-Rank LCMV Beamforming Algorithms Based on Joint Iterative Optimisation of Filters,” Electronics Letters, 2008.
[37] R. C. de Lamare and R. Sampaio-Neto, “Reduced-rank adaptive filtering based on joint iterative optimization of adaptive
filters”, IEEE Signal Process. Lett., vol. 14, no. 12, pp. 980983, Dec. 2007.
[38] R. C. de Lamare, M. Haardt, and R. Sampaio-Neto, “Blind
Adaptive Constrained Reduced-Rank Parameter Estimation
based on Constant Modulus Design for CDMA Interference
Suppression”, IEEE Transactions on Signal Processing, June
2008.
[39] M. Yukawa, R. C. de Lamare and R. Sampaio-Neto, “Efficient
Acoustic Echo Cancellation With Reduced-Rank Adaptive Filtering Based on Selective Decimation and Adaptive Interpolation,” IEEE Transactions on Audio, Speech, and Language
Processing, vol.16, no. 4, pp. 696-710, May 2008.
[40] R. C. de Lamare and R. Sampaio-Neto, “Reduced-rank spacetime adaptive interference suppression with joint iterative least
squares algorithms for spread-spectrum systems,” IEEE Trans.
Vehi. Technol., vol. 59, no. 3, pp. 1217-1228, Mar. 2010.
[41] R. C. de Lamare and R. Sampaio-Neto, “Adaptive reducedrank equalization algorithms based on alternating optimization design techniques for MIMO systems,” IEEE Trans. Vehi.
Technol., vol. 60, no. 6, pp. 2482-2494, Jul. 2011.
[42] R. C. de Lamare, L. Wang, and R. Fa, “Adaptive reduced-rank
LCMV beamforming algorithms based on joint iterative optimization of filters: Design and analysis,” Signal Processing,
vol. 90, no. 2, pp. 640-652, Feb. 2010.
[43] R. Fa, R. C. de Lamare, and L. Wang, “Reduced-Rank STAP
Schemes for Airborne Radar Based on Switched Joint Interpolation, Decimation and Filtering Algorithm,” IEEE Transactions on Signal Processing, vol.58, no.8, Aug. 2010, pp.41824194.
[44] L. Wang and R. C. de Lamare, ”Low-Complexity Adaptive
Step Size Constrained Constant Modulus SG Algorithms for
Blind Adaptive Beamforming”, Signal Processing, vol. 89, no.
12, December 2009, pp. 2503-2513.
[45] L. Wang and R. C. de Lamare, “Adaptive Constrained Constant
Modulus Algorithm Based on Auxiliary Vector Filtering for
Beamforming,” IEEE Transactions on Signal Processing, vol.
58, no. 10, pp. 5408-5413, Oct. 2010.
[46] L. Wang, R. C. de Lamare, M. Yukawa, ”Adaptive ReducedRank Constrained Constant Modulus Algorithms Based on
Joint Iterative Optimization of Filters for Beamforming,” IEEE
Transactions on Signal Processing, vol.58, no.6, June 2010,
pp.2983-2997.
[47] L. Qiu, Y. Cai, R. C. de Lamare and M. Zhao, “Reduced-Rank
DOA Estimation Algorithms Based on Alternating Low-Rank
Decomposition,” IEEE Signal Processing Letters, vol. 23, no.
5, pp. 565-569, May 2016.
[48] L. Wang, R. C. de Lamare and M. Yukawa, “Adaptive reducedrank constrained constant modulus algorithms based on joint
iterative optimization of filters for beamforming”, IEEE Transactions on Signal Processing, vol.58, no. 6, pp. 2983-2997,
June 2010.
[49] L. Wang and R. C. de Lamare, “Adaptive constrained constant modulus algorithm based on auxiliary vector filtering for
beamforming”, IEEE Transactions on Signal Processing, vol.
58, no. 10, pp. 5408-5413, October 2010.
[50] R. Fa and R. C. de Lamare, “Reduced-Rank STAP Algorithms
using Joint Iterative Optimization of Filters,” IEEE Transactions on Aerospace and Electronic Systems, vol.47, no.3,
pp.1668-1684, July 2011.
[51] Z. Yang, R. C. de Lamare and X. Li, “L1-Regularized STAP
Algorithms With a Generalized Sidelobe Canceler Architecture for Airborne Radar,” IEEE Transactions on Signal Processing, vol.60, no.2, pp.674-686, Feb. 2012.
[52] Z. Yang, R. C. de Lamare and X. Li, “Sparsity-aware spacetime adaptive processing algorithms with L1-norm regularisation for airborne radar”, IET signal processing, vol. 6, no. 5,
pp. 413-423, 2012.
[53] Neto, F.G.A.; Nascimento, V.H.; Zakharov, Y.V.; de Lamare,
R.C., ”Adaptive re-weighting homotopy for sparse beamforming,” in Signal Processing Conference (EUSIPCO), 2014 Proceedings of the 22nd European , vol., no., pp.1287-1291, 1-5
Sept. 2014
[54] Almeida Neto, F.G.; de Lamare, R.C.; Nascimento, V.H.; Zakharov, Y.V.,“Adaptive reweighting homotopy algorithms applied to beamforming,” IEEE Transactions on Aerospace and
Electronic Systems, vol.51, no.3, pp.1902-1915, July 2015.
[55] L. Wang, R. C. de Lamare and M. Haardt, “Direction finding algorithms based on joint iterative subspace optimization,” IEEE Transactions on Aerospace and Electronic Systems, vol.50, no.4, pp.2541-2553, October 2014.
[56] S. D. Somasundaram, N. H. Parsons, P. Li and R. C.
de Lamare, “Reduced-dimension robust capon beamforming using Krylov-subspace techniques,” IEEE Transactions on
Aerospace and Electronic Systems, vol.51, no.1, pp.270-289,
January 2015.
[57] H. Ruan and R. C. de Lamare, “Robust adaptive beamforming
using a low-complexity shrinkage-based mismatch estimation
algorithm,” IEEE Signal Process. Lett., vol. 21 no. 1 pp. 60-64,
Nov. 2013.
[58] H. Ruan and R. C. de Lamare, “Robust Adaptive Beamforming
Based on Low-Rank and Cross-Correlation Techniques,” IEEE
Transactions on Signal Processing, vol. 64, no. 15, pp. 39193932, Aug. 2016.
[59] S. Xu and R.C de Lamare, , Distributed conjugate gradient
strategies for distributed estimation over sensor networks, Sensor Signal Processing for Defense SSPD, September 2012.
[60] S. Xu, R. C. de Lamare, H. V. Poor, “Distributed Estimation
Over Sensor Networks Based on Distributed Conjugate Gradient Strategies”, IET Signal Processing, 2016 (to appear).
[74] M. Costa, ”Writing on dirty paper,” IEEE Trans. Inform. Theory, vol. 29, no. 3, pp. 439-441, May 1983.
[61] S. Xu, R. C. de Lamare and H. V. Poor, Distributed Compressed Estimation Based on Compressive Sensing, IEEE Signal Processing letters, vol. 22, no. 9, September 2014.
[75] R. C. de Lamare and A. Alcaim, ”Strategies to improve the
performance of very low bit rate speech coders and application
to a 1.2 kb/s codec” IEE Proceedings- Vision, image and signal
processing, vol. 152, no. 1, February, 2005.
[62] S. Xu, R. C. de Lamare and H. V. Poor, “Distributed reducedrank estimation based on joint iterative optimization in sensor networks,” in Proceedings of the 22nd European Signal
Processing Conference (EUSIPCO), pp.2360-2364, 1-5, Sept.
2014
[76] P. Clarke and R. C. de Lamare, ”Joint Transmit Diversity Optimization and Relay Selection for Multi-Relay Cooperative
MIMO Systems Using Discrete Stochastic Algorithms,” IEEE
Communications Letters, vol.15, no.10, pp.1035-1037, October 2011.
[63] S. Xu, R. C. de Lamare and H. V. Poor, “Adaptive link selection
strategies for distributed estimation in diffusion wireless networks,” in Proc. IEEE International Conference onAcoustics,
Speech and Signal Processing (ICASSP), , vol., no., pp.54025405, 26-31 May 2013.
[77] P. Clarke and R. C. de Lamare, ”Transmit Diversity and Relay
Selection Algorithms for Multirelay Cooperative MIMO Systems” IEEE Transactions on Vehicular Technology, vol.61, no.
3, pp. 1084-1098, October 2011.
[64] S. Xu, R. C. de Lamare and H. V. Poor, “Dynamic topology adaptation for distributed estimation in smart grids,” in
Computational Advances in Multi-Sensor Adaptive Processing
(CAMSAP), 2013 IEEE 5th International Workshop on , vol.,
no., pp.420-423, 15-18 Dec. 2013.
[65] S. Xu, R. C. de Lamare and H. V. Poor, “Adaptive Link Selection Algorithms for Distributed Estimation”, EURASIP Journal on Advances in Signal Processing, 2015.
[66] T. G. Miller, S. Xu, R. C. de Lamare and H. V. Poor, “Distributed Spectrum Estimation Based on Alternating Mixed
Discrete-Continuous Adaptation,” IEEE Signal Processing
Letters, vol. 23, no. 4, pp. 551-555, April 2016.
[67] N. Song, R. C. de Lamare, M. Haardt, and M. Wolf, “Adaptive
Widely Linear Reduced-Rank Interference Suppression based
on the Multi-Stage Wiener Filter,” IEEE Transactions on Signal Processing, vol. 60, no. 8, 2012.
[68] N. Song, W. U. Alokozai, R. C. de Lamare and M. Haardt,
“Adaptive Widely Linear Reduced-Rank Beamforming Based
on Joint Iterative Optimization,” IEEE Signal Processing Letters, vol.21, no.3, pp. 265-269, March 2014.
[78] Y. Cai, R. C. de Lamare, and R. Fa, “Switched Interleaving
Techniques with Limited Feedback for Interference Mitigation
in DS-CDMA Systems,” IEEE Transactions on Communications, vol.59, no.7, pp.1946-1956, July 2011.
[79] Y. Cai, R. C. de Lamare, D. Le Ruyet, “Transmit Processing
Techniques Based on Switched Interleaving and Limited Feedback for Interference Mitigation in Multiantenna MC-CDMA
Systems,” IEEE Transactions on Vehicular Technology, vol.60,
no.4, pp.1559-1570, May 2011.
[80] T. Wang, R. C. de Lamare, and P. D. Mitchell, “LowComplexity Set-Membership Channel Estimation for Cooperative Wireless Sensor Networks,” IEEE Transactions on Vehicular Technology, vol.60, no.6, pp.2594-2607, July 2011.
[81] T. Wang, R. C. de Lamare and A. Schmeink, ”Joint linear receiver design and power allocation using alternating optimization algorithms for wireless sensor networks,” IEEE Trans. on
Vehi. Tech., vol. 61, pp. 4129-4141, 2012.
[82] R. C. de Lamare, “Joint iterative power allocation and linear interference suppression algorithms for cooperative DS-CDMA
networks”, IET Communications, vol. 6, no. 13 , 2012, pp.
1930-1942.
[69] R.C. de Lamare, R. Sampaio-Neto and M. Haardt, ”Blind
Adaptive Constrained Constant-Modulus Reduced-Rank Interference Suppression Algorithms Based on Interpolation and
Switched Decimation,” IEEE Trans. on Signal Processing,
vol.59, no.2, pp.681-695, Feb. 2011.
[83] T. Peng, R. C. de Lamare and A. Schmeink, “Adaptive Distributed Space-Time Coding Based on Adjustable Code Matrices for Cooperative MIMO Relaying Systems”, IEEE Transactions on Communications, vol. 61, no. 7, July 2013.
[70] Y. Cai, R. C. de Lamare, “Adaptive Linear Minimum BER
Reduced-Rank Interference Suppression Algorithms Based on
Joint and Iterative Optimization of Filters,” IEEE Communications Letters, vol.17, no.4, pp.633-636, April 2013.
[84] T. Peng and R. C. de Lamare, “Adaptive Buffer-Aided Distributed Space-Time Coding for Cooperative Wireless Networks,” IEEE Transactions on Communications, vol. 64, no.
5, pp. 1888-1900, May 2016.
[71] R. C. de Lamare and R. Sampaio-Neto, “Sparsity-Aware
Adaptive Algorithms Based on Alternating Optimization and
Shrinkage,” IEEE Signal Processing Letters, vol.21, no.2,
pp.225,229, Feb. 2014.
[85] J. Gu, R. C. de Lamare and M. Huemer, “Buffer-Aided
Physical-Layer Network Coding with Optimal Linear Code
Designs for Cooperative Networks,” IEEE Transactions on
Communications, 2017.
[72] R. C. de Lamare, “Massive MIMO Systems: Signal Processing Challenges and Future Trends”, Radio Science Bulletin,
December 2013.
[86] K. Zu, R. C. de Lamare, “Low-Complexity Lattice ReductionAided Regularized Block Diagonalization for MU-MIMO Systems”, IEEE. Communications Letters, Vol. 16, No. 6, June
2012, pp. 925-928.
[73] W. Zhang, H. Ren, C. Pan, M. Chen, R. C. de Lamare, B. Du
and J. Dai, “Large-Scale Antenna Systems With UL/DL Hardware Mismatch: Achievable Rates Analysis and Calibration”,
IEEE Trans. Commun., vol.63, no.4, pp. 1216-1229, April
2015.
[87] K. Zu, R. C. de Lamare, “Low-Complexity Lattice ReductionAided Regularized Block Diagonalization for MU-MIMO Systems”, IEEE. Communications Letters, Vol. 16, No. 6, June
2012.
[88] K. Zu, R. C. de Lamare and M. Haart, “Generalized design
of low-complexity block diagonalization type precoding algorithms for multiuser MIMO systems”, IEEE Trans. Communications, 2013.
[89] M. Tomlinson, ”New automatic equaliser employing modulo
arithmetic,” Electronic Letters, vol. 7, Mar. 1971.
[90] C. T. Healy and R. C. de Lamare, “Decoder-optimised progressive edge growth algorithms for the design of LDPC codes with
low error floors”, IEEE Communications Letters, vol. 16, no.
6, June 2012, pp. 889-892.
[91] A. G. D. Uchoa, C. T. Healy, R. C. de Lamare, R. D. Souza,
“LDPC codes based on progressive edge growth techniques for
block fading channels”, Proc. 8th International Symposium on
Wireless Communication Systems (ISWCS), 2011, pp. 392-396.
[92] A. G. D. Uchoa, C. T. Healy, R. C. de Lamare, R. D.
Souza, “Generalised Quasi-Cyclic LDPC codes based on progressive edge growth techniques for block fading channels”,
Proc. International Symposium Wireless Communication Systems (ISWCS), 2012, pp. 974-978.
[93] A. G. D. Uchoa, C. T. Healy, R. C. de Lamare, R. D. Souza,
“Design of LDPC Codes Based on Progressive Edge Growth
Techniques for Block Fading Channels”, IEEE Communications Letters, vol. 15, no. 11, November 2011, pp. 1221-1223.
[94] C. T. Healy and R. C. de Lamare, “Design of LDPC Codes
Based on Multipath EMD Strategies for Progressive Edge
Growth,” IEEE Transactions on Communications, vol. 64, no.
8, pp. 3208-3219, Aug. 2016.
[95] H. Harashima and H. Miyakawa, ”Matched-transmission technique for channels with intersymbol interference,” IEEE Trans.
Commun., vol. 20, Aug. 1972.
[96] K. Zu, R. C. de Lamare and M. Haardt, “Multi-branch
tomlinson-harashima precoding for single-user MIMO systems,” in Smart Antennas (WSA), 2012 International ITG
Workshop on , vol., no., pp.36-40, 7-8 March 2012.
[97] K. Zu, R. C. de Lamare and M. Haardt, “Multi-Branch
Tomlinson-Harashima Precoding Design for MU-MIMO Systems: Theory and Algorithms,” IEEE Transactions on Communications, vol.62, no.3, pp.939,951, March 2014.
[98] L. Zhang, Y. Cai, R. C. de Lamare and M. Zhao, “Robust Multibranch TomlinsonHarashima Precoding Design in
Amplify-and-Forward MIMO Relay Systems,” IEEE Transactions on Communications, vol.62, no.10, pp.3476,3490, Oct.
2014.
[99] W. Zhang et al., “Widely Linear Precoding for Large-Scale
MIMO with IQI: Algorithms and Performance Analysis,”
IEEE Transactions on Wireless Communications, vol. 16, no.
5, pp. 3298-3312, May 2017.
[100] B. Hochwald, C. Peel and A. Swindlehurst, ”A vectorperturbation technique for near capacity multiantenna multiuser communication - Part II: Perturbation,” IEEE Trans.
Commun., vol. 53, no. 3, Mar. 2005.
[101] C. B. Chae, S. Shim and R. W. Heath, ”Block diagonalized
vector perturbation for multiuser MIMO systems,” IEEE Trans.
Wireless Commun., vol. 7, no. 11, pp. 4051 - 4057, Nov. 2008.
[102] R. C. de Lamare, R. Sampaio-Neto, “Adaptive MBER decision feedback multiuser receivers in frequency selective fading
channels”, IEEE Communications Letters, vol. 7, no. 2, Feb.
2003, pp. 73 - 75.
[103] A. Rontogiannis, V. Kekatos, and K. Berberidis,” A SquareRoot Adaptive V-BLAST Algorithm for Fast Time-Varying
MIMO Channels,” IEEE Signal Processing Letters, Vol. 13,
No. 5, pp. 265-268, May 2006.
[104] R. C. de Lamare, R. Sampaio-Neto, A. Hjorungnes, “Joint
iterative interference cancellation and parameter estimation for
CDMA systems”, IEEE Communications Letters, vol. 11, no.
12, December 2007, pp. 916 - 918.
[105] Y. Cai and R. C. de Lamare, ”Adaptive Space-Time Decision Feedback Detectors with Multiple Feedback Cancellation”, IEEE Transactions on Vehicular Technology, vol. 58, no.
8, October 2009, pp. 4129 - 4140.
[106] J. W. Choi, A. C. Singer, J Lee, N. I. Cho, “Improved linear
soft-input soft-output detection via soft feedback successive interference cancellation,” IEEE Trans. Commun., vol.58, no.3,
pp.986-996, March 2010.
[107] R. C. de Lamare and R. Sampaio-Neto, “Blind adaptive
MIMO receivers for space-time block-coded DS-CDMA systems in multipath channels using the constant modulus criterion,” IEEE Transactions on Communications, vol.58, no.1,
pp.21-27, January 2010.
[108] R. Fa, R. C. de Lamare, “Multi-Branch Successive Interference Cancellation for MIMO Spatial Multiplexing Systems”,
IET Communications, vol. 5, no. 4, pp. 484 - 494, March 2011.
[109] R.C. de Lamare and R. Sampaio-Neto, “Adaptive reducedrank equalization algorithms based on alternating optimization
design techniques for MIMO systems,” IEEE Trans. Veh. Technol., vol. 60, no. 6, pp. 2482-2494, July 2011.
[110] P. Li, R. C. de Lamare and R. Fa, “Multiple Feedback Successive Interference Cancellation Detection for Multiuser MIMO
Systems,” IEEE Transactions on Wireless Communications,
vol. 10, no. 8, pp. 2434 - 2439, August 2011.
[111] R.C. de Lamare, R. Sampaio-Neto, “Minimum mean-squared
error iterative successive parallel arbitrated decision feedback
detectors for DS-CDMA systems,” IEEE Trans. Commun., vol.
56, no. 5, May 2008, pp. 778-789.
[112] R.C. de Lamare, R. Sampaio-Neto, “Minimum mean-squared
error iterative successive parallel arbitrated decision feedback
detectors for DS-CDMA systems,” IEEE Trans. Commun., vol.
56, no. 5, May 2008.
[113] R.C. de Lamare and R. Sampaio-Neto, “Adaptive reducedrank equalization algorithms based on alternating optimization
design techniques for MIMO systems,” IEEE Trans. Veh. Technol., vol. 60, no. 6, pp. 2482-2494, July 2011.
[114] P. Li, R. C. de Lamare and J. Liu, “Adaptive Decision Feedback Detection with Parallel Interference Cancellation and
Constellation Constraints for Multiuser MIMO systems”, IET
Communications, vol.7, 2012, pp. 538-547.
[115] J. Liu, R. C. de Lamare, “Low-Latency Reweighted Belief
Propagation Decoding for LDPC Codes,” IEEE Communications Letters, vol. 16, no. 10, pp. 1660-1663, October 2012.
[116] P. Li and R. C. de Lamare, Distributed Iterative Detection
With Reduced Message Passing for Networked MIMO Cellular Systems, IEEE Transactions on Vehicular Technology,
vol.63, no.6, pp. 2947-2954, July 2014.
[117] A. G. D. Uchoa, C. T. Healy and R. C. de Lamare, “Iterative Detection and Decoding Algorithms For MIMO Systems
in Block-Fading Channels Using LDPC Codes,” IEEE Transactions on Vehicular Technology, 2015.
[118] R. C. de Lamare, ”Adaptive and Iterative Multi-Branch
MMSE Decision Feedback Detection Algorithms for MultiAntenna Systems”, IEEE Trans. Wireless Commun., vol. 14,
no. 10, October 2013.
[119] R. Combes, Z. Altman and E. Altman, “Interference coordination in wireless networks: A flow-level perspective,” Proceedings IEEE INFOCOM 2013, vol., no., pp.2841,2849, 1419 April 2013.
[120] R. Aggarwal, C. E. Koksal, and P. Schniter, “On the design
of large scale wireless systems”, IEEE J. Sel. Areas Commun,
vol. 31, no. 2, pp. 215-225, Feb. 2013.
[121] C. Shepard, H. Yu, N. Anand, L. E. Li, T. L. Marzetta, R.
Yang, and L. Zhong, Argos: Practical many-antenna base stations, in ACM Int. Conf.Mobile Computing and Networking
(MobiCom), Istanbul, Turkey, Aug. 2012.
[122] J. Hoydis, C. Hoek, T. Wild, and S. ten Brink, Channel measurements for large antenna arrays, in IEEE International Symposium on Wireless Communication Systems (ISWCS), Paris,
France, Aug. 2012.
[123] X. Gao, F. Tufvesson, O. Edfors, and F. Rusek, Measured
propagation characteristics for very-large MIMO at 2.6 GHz,
in Proc. of the 46th Annual Asilomar Conference on Signals,
Systems, and Computers,, Pacific Grove, California, USA,
Nov. 2012.
[124] J. Arnau, B. Devillers, C. Mosquera, A. Prez-Neira, “Performance study of multiuser interference mitigation schemes
for hybrid broadband multibeam satellite architectures”,
EURASIP Journal on Wireless Communications and Networking, 2012:132 (5 April 2012).
[125] J. Jose, A. Ashikhmin, T. L. Marzetta, S. Vishwanath, “Pilot Contamination and Precoding in Multi-Cell TDD Systems,”
IEEE Transactions on Wireless Communications, vol.10, no.8,
pp. 2640-2651, August 2011.
[126] H. Q. Ngo, E. G. Larsson, and T. L. Marzetta, “Energy and
spectral efficiency of very large multiuser MIMO systems”,
IEEE Trans. Commun., vol. 61, no. 4, pp. 1436-1449, Apr.
2013.
[127] H. Yang and T. L. Marzetta, Performance of conjugate
and zero-forcing beamforming in large-scale antenna systems,
IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. 172179, Feb.
2013.
[128] A. Ashikhmin and T. L. Marzetta, Pilot contamination precoding in multi-cell large scale antenna systems, in IEEE International Symposium on Information Theory (ISIT), Cambridge, MA, Jul. 2012.
[129] J. Zhang, X. Yuan, and L. Ping, Hermitian precoding for distributed MIMO systems with individual channel state information, IEEE J. Sel. Areas Commun., vol. 31, no. 2, pp. 241250,
Feb. 2013.
[130] K. Vardhan, S. Mohammed, A. Chockalingam and B. Rajan,
“A low-complexity detector for large MIMO systems and multicarrier CDMA systems”, IEEE J. Sel. Commun., vol. 26, no.
3, pp. 473-485, Apr. 2008.
[131] R.C. de Lamare, R. Sampaio-Neto, Minimum mean-squared
error iterative successive parallel arbitrated decision feedback
detectors for DS-CDMA systems, IEEE Trans. Commun., vol.
56, no. 5, May 2008, pp. 778-789.
[132] P. Li and R. D. Murch, Multiple Output Selection-LAS Algorithm in Large MIMO Systems, IEEE Commun. Lett., vol. 14,
no. 5, pp. 399-401, May 2010.
[133] R. C. de Lamare, ”Adaptive and Iterative Multi-Branch
MMSE Decision Feedback Detection Algorithms for MultiAntenna Systems”, IEEE Transactions on Wireless Communications, vol. 14, no. 10, October 2013.
[134] Q. Zhou and X. Ma, Element-Based Lattice Reduction Algorithms for Large MIMO Detection, IEEE J. Sel. Areas Commun, vol. 31, no. 2, pp. 274-286, Feb. 2013.
[135] H. Qian and S. N. Batalama, Data-record-based criteria
for the selection of an auxiliary vector estimator of the
MMSE/MVDR filter, IEEE Trans. Commun., vol. 51, no. 10,
pp. 17001708, Oct. 2003.
[136] Y. Sun, V. Tripathi, and M. L. Honig, Adaptive, iterative,
reducedrank (turbo) equalization, IEEE Trans. Wireless Commun., vol. 4, no. 6, pp. 27892800, Nov. 2005.
[137] R. C. de Lamare and R. Sampaio-Neto, Adaptive reducedrank processing based on joint and iterative interpolation, decimation, and filtering, IEEE Trans. Signal Process., vol. 57, no.
7, July 2009, pp. 2503-2514.
[138] R.C. de Lamare and R. Sampaio-Neto, Adaptive reduced-rank
equalization algorithms based on alternating optimization design techniques for MIMO systems, IEEE Trans. Veh. Technol., vol. 60, no. 6, pp. 2482-2494, July 2011.
[139] Z. Tu and R.S. Blum, Multiuser diversity for a dirty paper
approach, IEEE Commun. Lett., vol. 7, no. 8, pp. 370372, Aug.
2003.
[140] G. Dimic and N.D. Sidiropoulos, On downlink beamforming
with greedy user selection: Performance analysis and a simple
new algorithm, IEEE Trans. Signal Processing, vol. 53, no. 10,
pp. 38573868, Oct. 2005.
[141] Z. Shen, J.G. Andrews, R.W. Heath Jr., and B.L. Evans, Low
complexity user selection algorithms for multiuser MIMO systems with block diagonalization, IEEE Trans. Signal Processing, vol. 54, no. 9, pp. 36583663, Sept. 2006.
[142] N. Dao and Y. Sun, “User-selection algorithms for multiuser
precoding,” IEEE Trans. Veh. Technol., vol. 59, no. 7, Sep.
2010.
[143] P. Clarke and R. C. de Lamare, ”Transmit Diversity and Relay
Selection Algorithms for Multi-relay Cooperative MIMO Systems”, IEEE Transactions on Vehicular Technology, vol. 61 ,
no. 3, March 2012, pp. 1084 - 1098.
[144] G. Caire and S. Shamai (Shitz), “On the achievable throughput of a multiantenna Gaussian broadcast channel,” IEEE
Trans. Inform. Theory, vol. 49, no. 7, pp. 16911706, July 2003.
[145] M. Joham, W. Utschick, and J. Nossek, “Linear transmit processing in MIMO communications systems,” IEEE Trans. Signal Processing, vol. 53, no. 8, pp. 27002712, Aug. 2005.
[146] Q. Spencer, A.L. Swindlehurst, and M. Haardt, “Zero-forcing
methods for downlink spatial multiplexing in multiuser MIMO
channels,” IEEE Trans. Signal Processing, vol. 52, no. 2, pp.
462471, Feb. 2004.
[147] V. Stankovic and M. Haardt, “Generalized design of multiuser MIMO precoding matrices,” IEEE Trans. Wireless Commun., vol. 7, no. 3, pp. 953-961, Mar. 2008.
[148] K. Zu, R. C. de Lamare and M. Haardt, “Generalized design
of low- complexity block diagonalization type precoding algorithms for multiuser MIMO systems”, IEEE Trans. Communications, 2013.
[164] J. H. Choi, H. Y. Yu, Y. H. Lee, Adaptive MIMO decision
feedback equalization for receivers with time-varying channels, IEEE Trans. Signal Proc., 2005, 53, no. 11, pp. 42954303.
[149] C. Windpassinger, R. Fischer, T. Vencel and J. Huber, “Precoding in multiantenna and multiuser communications,” IEEE
Trans. Wireless Commun., vol. 3, no. 4, Jul. 2004.
[165] C. Windpassinger, L. Lampe, R.F.H. Fischer, T.A Hehn, “A
performance study of MIMO detectors,” IEEE Transactions
on Wireless Communications, vol. 5, no. 8, August 2006, pp.
2004-2008.
[150] C.B. Peel, B.M. Hochwald, and A.L. Swindlehurst, A vectorperturbation technique for near capacity multiantenna multiuser communicationpart I: channel inversion and regularization, IEEE Trans. Commun., vol. 53, no. 1, pp. 195202, Jan.
2005.
[151] M. Biguesh, A.B. Gershman, “Training-based MIMO channel estimation: a study of estimator tradeoffs and optimal training signals”, IEEE Transactions on Signal Processing, vol. 54
no. 3, March 2006.
[152] S. Verdu, Multiuser Detection, Cambridge, 1998.
[153] E. Viterbo and J. Boutros, “A universal lattice code decoder
for fading channels”, IEEE Trans. on Inf. Theory, vol. 45, no.
5, pp.16391642, July 1999.
[154] M. O. Damen, H. E. Gamal, and G. Caire, On maximum likelihood detection and the search for the closest lattice point,
IEEE Trans. Inform. Theory, vol. 49, pp. 23892402, Oct. 2003.
[155] Z. Guo and P. Nilsson, Algorithm and Implementation of the
K-Best Sphere Decoding for MIMO Detection, IEEE Journal on Selected Areas in Communications, vol. 24, no. 3, pp.
491503, March 2006.
[156] C. Studer, A. Burg, and H. Bolcskei, Soft-output sphere decoding: algorithms and VLSI implementation, IEEE J. Sel. Areas Commun., vol. 26, pp. 290-300, Feb. 2008.
[166] Y. H. Gan, C. Ling, and W. H. Mow, Complex lattice reduction algorithm for low-complexity full-diversity MIMO detection, IEEE Trans. Signal Processing, vol. 56, no. 7, July 2009.
[167] K. J. Kim, J. Yue, R. A. Iltis, and J. D. Gibson, “A QRDM/Kalman filter-based detection and channel estimation algorithm for MIMO-OFDM systems”, IEEE Trans. Wireless Communications, vol. 4,pp. 710-721, March 2005.
[168] Y. Jia, C. M. Vithanage, C. Andrieu, and R. J. Piechocki,
Probabilistic data association for symbol detection in MIMO
systems, Electron. Lett., vol. 42, no. 1, pp. 3840, Jan. 2006.
[169] S. Yang, T. Lv, R. Maunder, and L. Hanzo, ”Unified BitBased Probabilistic Data Association Aided MIMO Detection
for High-Order QAM Constellations”, IEEE Transactions on
Vehicular Technology, vol. 60, no. 3, pp. 981-991, 2011.
[170] M. K. Varanasi, “Decision feedback multiuser detection: A
systematic approach,” IEEE Trans. on Inf. Theory, vol. 45, pp.
219-240, January 1999.
[171] J. F. Rler and J. B. Huber, ”Iterative soft decision interference cancellation receivers for DS-CDMA downlink employing 4QAM and 16QAM,” in Proc. 36th Asilomar Conf. Signal,
Systems and Computers, Pacific Grove, CA, Nov. 2002.
[157] B. Shim and I. Kang, On further reduction of complexity in
tree pruning based sphere search, IEEE Trans. Commun., vol.
58, no. 2, pp. 417422, Feb. 2010.
[172] J. Luo, K. R. Pattipati, P. K. Willet and F. Hasegawa, “Optimal
User Ordering and Time Labeling for Ideal Decision Feedback
Detection in Asynchronous CDMA”, IEEE Trans. on Communications, vol. 51, no. 11, November, 2003.
[158] A. Duel-Hallen, “Equalizers for Multiple Input Multiple Output Channels and PAM Systems with Cyclostationary Input
Sequences,” IEEE J. Select. Areas Commun., vol. 10, pp. 630639, April, 1992.
[173] G. Woodward, R. Ratasuk, M. L. Honig and P. Rapajic, “Minimum Mean-Squared Error Multiuser Decision-Feedback Detectors for DS-CDMA,” IEEE Trans. on Communications, vol.
50, no. 12, December, 2002.
[159] G. D. Golden, C. J. Foschini, R. A. Valenzuela and P. W. Wolniansky, “Detection algorithm and initial laboratory results using V-BLAST space-time communication architecture”, Electronics Letters, vol. 35, No.1, January 1999.
[174] R.C. de Lamare, R. Sampaio-Neto, “Adaptive MBER decision feedback multiuser receivers in frequency selective fading
channels”, IEEE Communications Letters, vol. 7, no. 2, Feb.
2003, pp. 73 - 75.
[160] J. Benesty, Y. Huang, and J. Chen, A fast recursive algorithm
for optimum sequential signal detection in a BLAST system,
IEEE Trans. Signal Processing, vol. 51, pp. 17221730, July
2003.
[175] F. Cao, J. Li, and J. Yang, ”On the relation between PDA and
MMSE-ISDIC,” IEEE Signal Processing Letters, vol. 14, no.
9, Sep. 2007.
[161] A. Rontogiannis, V. Kekatos, and K. Berberidis,” A SquareRoot Adaptive V-BLAST Algorithm for Fast Time-Varying
MIMO Channels,” IEEE Signal Processing Letters, Vol. 13,
No. 5, pp. 265-268, May 2006.
[176] Y. Cai and R. C. de Lamare, ”Adaptive Space-Time Decision Feedback Detectors with Multiple Feedback Cancellation”, IEEE Transactions on Vehicular Technology, vol. 58, no.
8, October 2009, pp. 4129 - 4140.
[162] R. Fa, R. C. de Lamare, “Multi-Branch Successive Interference Cancellation for MIMO Spatial Multiplexing Systems”,
IET Communications, vol. 5, no. 4, pp. 484 - 494, March 2011.
[177] P. Li and R. C. de Lamare, ”Adaptive Decision-Feedback Detection With Constellation Constraints for MIMO Systems”,
IEEE Transactions on Vehicular Technology, vol. 61, no. 2,
853-859, 2012.
[163] P. Li, R. C. de Lamare and R. Fa, “Multiple Feedback Successive Interference Cancellation Detection for Multiuser MIMO
Systems,” IEEE Transactions on Wireless Communications,
vol. 10, no. 8, pp. 2434 - 2439, August 2011.
[178] M. Reuter, J.C. Allen, J. R. Zeidler, R. C. North, “Mitigating error propagation effects in a decision feedback equalizer”, IEEE Transactions on Communications, vol. 49, no. 11,
November 2001, pp. 2028 - 2041.
[179] R.C. de Lamare, R. Sampaio-Neto, A. Hjorungnes, “Joint iterative interference cancellation and parameter estimation for
cdma systems”, IEEE Communications Letters, vol. 11, no. 12,
December 2007, pp. 916 - 918.
[195] T. Adali, P. J. Schreier and L. L. Scharf, “Complex-valued
signal processing: the proper way to deal with impropriety”,
IEEE Trans. Signal Processing, vol. 59, no. 11, pp. 51015125 ,
November 2011.
[180] S. Haykin, Adaptive Filter Theory, 4th ed. Englewood Cliffs,
NJ: Prentice- Hall, 2002.
[196] N. Song, R. C. de Lamare, M. Haardt, and M. Wolf, “Adaptive
Widely Linear Reduced-Rank Interference Suppression based
on the Multi-Stage Wiener Filter,” IEEE Transactions on Signal Processing, vol. 60, no. 8, 2012.
[181] C. Berrou and A. Glavieux, “Near optimum error-correcting
coding and decoding: Turbo codes,” IEEE Trans. Commun.,
vol. 44, Oct. 1996.
[182] C. Douillard et al., “Iterative correction of intersymbol interference: Turbo equalization,” European Trans. Telecommun.,
vol. 6, no. 5, pp. 507511, Sept.Oct. 1995.
[183] X. Wang and H. V. Poor, “Iterative (turbo) soft interference
cancellation and decoding for coded CDMA,” IEEE Trans.
Commun., vol. 47, pp. 10461061, July 1999.
[184] M. Tuchler, A. Singer, and R. Koetter, ”Minimum mean
square error equalization using a priori information,” IEEE
Trans. Signal Processing, vol. 50, pp. 673-683, Mar. 2002.
[185] B. Hochwald and S. ten Brink, Achieving near-capacity on a
mutliple- antenna channel, IEEE Trans. Commun., vol. 51, pp.
389-399, Mar. 2003.
[186] J. Hou, P. H. Siegel, L. B. Milstein, “Design of multiinput multi-output systems based on low-density Parity-check
codes,” IEEE Transactions on Communications, vol. 53, no. 4,
pp. 601- 611, April 2005.
[187] H. Lee, B. Lee, and I. Lee, “Iterative detection and decoding
with an improved V-BLAST for MIMO-OFDM Systems, IEEE
J. Sel. Areas Commun., vol. 24, pp. 504-513, Mar. 2006.
[188] J. Wu, H.-N. Lee, “Performance Analysis for LDPC-Coded
Modulation in MIMO Multiple-Access Systems,” IEEE Transactions on Communications, vol. 55, no. 7, pp. 1417-1426, July
2007
[189] X. Yuan, Q. Guo, X. Wang, and Li Ping, ”Evolution analysis
of low-cost iterative equalization in coded linear systems with
cyclic prefixes,” IEEE J. Select. Areas Commun. (JSAC), vol.
26, no. 2, pp. 301-310, Feb. 2008.
[190] J. W. Choi, A. C. Singer, J Lee, N. I. Cho, Improved linear
soft-input soft-output detection via soft feedback successive interference cancellation, IEEE Trans. Commun., vol.58, no.3,
pp.986-996, March 2010.
[191] M. J. Wainwright, T. S. Jaakkola, and A.S. Willsky, “A new
class of upper bounds on the log partition function,” IEEE
Trans. Information Theory, vol. 51, no. 7, pp. 2313 - 2335,
July 2005.
[192] H. Wymeersch, F. Penna and V. Savic, “Uniformly
Reweighted Belief Propagation for Estimation and Detection in
Wireless Networks,” IEEE Trans. Wireless Communications,
vol. PP, No. 99, pp. 1-9, Feb. 2012.
[193] J. Liu, R. C. de Lamare, “Low-Latency Reweighted Belief
Propagation Decoding for LDPC Codes,” IEEE Communications Letters, vol. 16, no. 10, pp. 1660-1663, October 2012.
[194] A. Hakkarainen, J. Werner, K. R. Dandekar and M. Valkama,
“Widely-linear beamforming and RF impairment suppression
in massive antenna arrays,” Journal of Communications and
Networks, vol.15, no.4, pp.383,397, Aug. 2013
[197] R. C. de Lamare, ”Massive MIMO Systems: Signal Processing Challenges and Future Trends”, URSI Radio Science Bulletin, 2013.
[198] W. Zhang, H. Ren, C. Pan, M. Chen, R. C. de Lamare, B.
Du and J. Dai, “Large-Scale Antenna Systems With UL/DL
Hardware Mismatch: Achievable Rates Analysis and Calibration”, IEEE Transactions on Communications, vol.63, no.4, pp.
1216-1229, April 2015.
[199] J. Gu and R. C. de Lamare, “Buffer-aided distributed spacetime coding schemes and algorithms for cooperative DSCDMA systems”, EURASIP Journal on Wireless Communications and Networking, December 2016.
[200] S. Vishwanath, N. Jindal and A. J. Goldsmith, “On the capacity of multiple input multiple output broadcast channels,”
in Proc. IEEE International Conf. Commun. (ICC), New York,
USA, Apr. 2002, pp. 1444-1450.
| 7cs.IT
|
Beyond Markov Logic: Efficient Mining of Prediction Rules in
Large Graphs
arXiv:1802.03638v2 [cs.DB] 13 Feb 2018
Tommaso Soru
André Valdestilhas
AKSW, University of Leipzig
Leipzig, Germany 04109
[email protected]
AKSW, University of Leipzig
Leipzig, Germany 04109
[email protected]
Edgard Marx
Axel-Cyrille Ngonga Ngomo
Leipzig University of Applied Sciences
Leipzig, Germany 04277
[email protected]
ABSTRACT
Graph representations of large knowledge bases may comprise billions of edges. Usually built upon human-generated ontologies,
several knowledge bases do not feature declared ontological rules
and are far from being complete. Current rule mining approaches
rely on schemata or store the graph in-memory, which can be unfeasible for large graphs. In this paper, we introduce HornConcerto,
an algorithm to discover Horn clauses in large graphs without the
need of a schema. Using a standard fact-based confidence score,
we can mine close Horn rules having an arbitrary body size. We
show that our method can outperform existing approaches in terms
of runtime and memory consumption and mine high-quality rules
for the link prediction task, achieving state-of-the-art results on a
widely-used benchmark. Moreover, we find that rules alone can perform inference significantly faster than embedding-based methods
and achieve accuracies on link prediction comparable to resourcedemanding approaches such as Markov Logic Networks.
CCS CONCEPTS
•Computing methodologies → Knowledge representation and
reasoning; •Information systems → Association rules;
KEYWORDS
Rule Mining, Knowledge Bases, SPARQL, RDF
ACM Reference format:
Tommaso Soru, André Valdestilhas, Edgard Marx, and Axel-Cyrille Ngonga
Ngomo. . Beyond Markov Logic: Efficient Mining of Prediction Rules in
Large Graphs. In Proceedings of Submitted, 2018, KDD, 7 pages.
DOI:
1
INTRODUCTION
The number of published triple-based datasets has been exponentially increasing in the last years. Knowledge Bases such as DBpedia [13], UniProt [6], and LinkedGeoData [25] comprise more than
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
Submitted, 2018
© 2018 Copyright held by the owner/author(s). .
DOI:
DICE, Paderborn University
Paderborn, Germany 33098
[email protected]
one billion triples each. While these knowledge bases span manifold domains, they are usually far from being complete. Several
approaches have been devised to tackle this problem, known as
link prediction or knowledge base completion. Most of these methods
belong to the categories of translation-based approaches, tensorfactorization algorithms, neural networks, or general statisticallearning algorithms, some of which involving the use of first-order
rules. Beyond their utility in inference engines, rules help to understand insights in data, which is not a trivial task on large datasets.
In particular, rules have shown to be more explainable than factorization models, as they can be easily understood by humans as well
as machines.
Rules have been specifically used in different environments, e.g.
reasoners, rule-based systems, and Markov Logic Networks (MLNs).
Non-probabilistic reasoners utilize declared schemata describing a
definition or a restriction, which imply – for their nature – a crisp
reasoning. For instance, with the restriction ∃spouse v Person
(“All entities who have a spouse are persons.”), one can infer that
an entity is a person with a confidence value of 1.0, if they have
a spouse. On the other hand, probabilistic reasoners assign a real
confidence value between 0 and 1 to each statement, which can thus
exist within some probability degree [27]. However, the growth of
published structured data has made rule mining and probabilistic
inference impracticable on certain datasets. In order to optimize
the computation, existing methods load data in-memory or rely
on schemata while loading data into in-disk databases. Unfortunately, classes and properties in ontologies are as complete as their
instantiations. For instance, in DBpedia 2016-04, among 63, 764
properties, only 3.5% and 3.8% of them are provided with a domain
and range, respectively.1 Moreover, loading the largest available
datasets in-memory is not feasible on a machine with average resources.
In this paper, we introduce an algorithm to the efficient discovery
of Horn rules in graphs dubbed HornConcerto.Using a standard
fact-based confidence score, we can mine closed Horn rules having
an arbitrary body size. Our work brings the following contributions:
(1) We outperform existing rule mining approaches in terms of
runtime and memory consumption, while scaling to very
large knowledge bases.
(2) We mine high-quality rules for link prediction.
1 Retrieved
on August 5th, 2017.
Submitted, KDD, 2018
Tommaso Soru, André Valdestilhas, Edgard Marx, and Axel-Cyrille Ngonga Ngomo
(3) We find that rules alone can perform inference significantly
faster than embedding-based methods and achieve accuracies on link prediction comparable to resource-demanding
approaches such as Markov Logic Networks.
This paper is arranged as follows. The next section presents the
related work, while Section 3 introduces the preliminary concepts.
The HornConcerto approach is described in Section 4. Section 5
shows the evaluation results, which are discussed in Section 6. In
Section 7, we conclude.
2
RELATED WORK
This work is mainly related with Inductive Logic Programming (ILP)
systems to mine rules in datasets where statements are composed by
triples (or triplets). The main differences among these approaches
are in data storage type and the confidence score used for rules.
Warmr is a framework which combines ILP with association-rule
mining, performing a breadth-first search to find frequent patterns.
It uses a standard confidence function [7]. Aleph implements several measures for rule quality and has the peculiarity to generate
random negative examples [17]. With this respect, Sherlock, Amie,
and Amie+ use only declared statements as counterexamples, thus
embracing the open-world assumption (OWA) [8, 23]. While Amie
uses an in-memory database to discover rules, Ontological Pathfinding (OP) – state-of-the-art rule miner in terms of performances –
utilizes an external DBMS [4].
Among the frameworks based on MLNs, the majority do not provide the possibility of discovering rules. NetKit-SRL [16], ProbCog [10],
Tuffy [19], Markov theBeast [22], and RockIt [20] all need rules as
input in order to proceed with the weight learning, grounding, and
inference phases. Alchemy [12] instead performs the construction
of rules in a combinatorial way; however, this method is highly
inefficient, especially on datasets having an elevated number of
distinct properties. Besides link prediction and knowledge expansion [5], probabilistic inference systems showed to be effective in
fields such as question answering [24], instance matching [14], and
class expression learning [3].
3 PRELIMINARIES
3.1 Horn clauses
In the context of first-order logic, an atom is a formula which contains a predicate and a list of undefined arguments and cannot be
divided into sub-formulas. For instance, the atom A := a(x 1 , x 2 )
features a binary predicate a on the free variables x 1 , x 2 . The typical structure of first-order rules expects an implication sign, an
atom H as the head (i.e., the implication target), and a set of atoms
B 1 , . . . , Bn as the body. The body can be composed by one or more
conjunction literals, i.e. H ⇐= B 1 ∧ . . . ∧ Bn . An expression can be
rewritten in disjunctive normal form (DNF) if it can be expressed
as a chain of inclusive disjunctions. Turning the rule above to the
equivalent DNF, we have H ∨ ¬B 1 ∨ . . . ∨ ¬Bn .
Definition 3.1. Horn clause. A Horn clause (or rule) is a disjunction of literals with at most one positive, i.e. non-negated,
literal.
Ó
→
−
→
−
Therefore, let B := ni=1 Bi , any first-order rule H ⇐= B is also
a Horn clause.
3.2
Rule scoring
Most rule mining algorithms measure and rank rules according to
some score. The most used measure for rule quality relies on the
number of occurrences in data and is called standard confidence.
Definition 3.2. Standard confidence. The support of a Horn
clause estimates the probability of the rule head to hold true, given
the body:
→
−
P(H ∩ B )
→
−
P(H | B ) =
(1)
→
−
P( B )
→
−
The standard confidence of a rule R := H ⇐ B is defined as:
→
−
{H ∧ B }
c(R) =
(2)
→
−
{B }
The numerator in Equation (2) is also called support. For the sake
of simplicity, we will name the denominator “body support”.
→
−
For instance, for H := h(x, y) and B = B 1 := b1 (x, y), the standard confidence is a rate between two absolute numbers, i.e. (i) the
number of occurrences found in graph G such that b1 (x, y) ∧h(x, y)
holds over (ii) the number of occurrences such that b1 (x, y) holds.
We now formally define the concept of triple in a graph.
Definition 3.3. Triple. Given a directed labelled multi-graph
G = (V , E, l) with labelling function l : V ∪ E → U , where U is the
set of all possible labels, a triple (s, p, o) := (s̄, p̄, ō) belongs to the
graph G iff s, o ∈ V ∧ e = (s, o) ∈ E ∧ l(s) = s̄ ∧ l(e) = p̄ ∧ l(o) = ō.
Algorithms such as Amie and Sherlock introduced a score which
can deal with incomplete data [8, 23]. This score is referred to with
the name partial completeness assumption (PCA) score.
Definition 3.4. PCA score. A Horn clause can be associated to
a score called PCA score. The rationale behind the PCA is that
a missing relationship should not be treated as counterexample;
instead, only triples (s, p, o 1 ) can be treated as counterexamples for
a triple (s, p, o 2 ) with o 1 , o 2 . Being H̄ any relationship h(·, ·), the
→
−
PCA score of a rule R := H ⇐ B is defined as:
→
−
{H ∧ B }
pca(R) =
(3)
→
−
{H̄ ∧ B }
4
HORN CONCERTO
In this section, the core of the approach is presented. HornConcerto is a rule mining algorithm designed with scalability in mind.
Moreover, we adopted a few constraints in order to let the computation focus on the more informative sections of the datasets
first. HornConcerto is a complete algorithm, meaning that if
run without constraints, it can detect all Horn rules of a given
type. However, when dealing with large datasets, these constraints
are necessary to reduce runtime and resource consumption. The
constraints are defined as such:
(1) consider only rules having a confidence value of at least
θ ∈ [0, 1];
(2) consider only rules having one of the top P properties in
its body.
→
−
(3) consider only top Ti properties (i ≥ 2) as part of a | B | = i.
Beyond Markov Logic: Efficient Mining of Prediction Rules in Large Graphs
4.1
Algorithm
The HornConcerto algorithm is shown in Algorithm 1. First, we
retrieve the top P properties sorted by their occurrence counts. The
occurrence counts are the body support values for rules having a
body size of 1. Then, we mine the rules and compute their standard
confidence (see Definition 3.2). For each top-property q̄, we count
the occurrences of p(x, y) and q̄(x, y) to happen together. We can
then find a map m 1 : Π → N such that:
m 1 (p) = |{(x, y) ∈ V 2 : (x, p, y), (x, q̄, y) ∈ G ∧ p , q̄}|∀p ∈ Π
(4)
where Π = {l(e) : e ∈ E}, i.e. the set of properties in the graph.
The confidence of a rule is thus the ratio between their support
in Equation (4) and body support. Intuitively, changing (x, y) into
(y, x) in the rule body, we can discover rules of type p(x, y) ⇐
q(y, x), i.e. inverse properties.
Algorithm 1: HornConcerto
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
Data: target graph G, θ , P, T
Result: set P of mined rules and confidence values
retrieve the top P properties;
for each rule type (in parallel) do
initialize cache;
for each top property in P do
if rile body size ≤ 1 then
Find matching digons and compute respective
rule support;
Store matching digons into set R;
else
4 Find matching triangles and compute respective
rule support;
Store matching triangles into set R;
for each rule in R do
Compute body support (i.e., find
#adjacencies of body properties);
end
end
for each rule in R sorted by descend confidence do
if confidence ≥ θ then
Add rule and its confidence to output set P;
else
break;
end
end
end
end
We then extend this concept to rules having a body size of 2.
These cases involving 3 nodes and 3 edges can be viewed as cliques
(n)
of length 3 (or triangles). The aim is to compute a map m 2 : Π 2 →
N which returns the triangles involving properties (p, r ) given q̄.
4 m 2 (p, r ) = |{(x, y, z) ∈ V 3 : (x, p, y), (z, r , y), (x, q̄, z) ∈ G}|∀p, r ∈ Π
(5)
(n)
Submitted, KDD, 2018
Consequently, rules of type p(x, y) ⇐ q̄(x, z) ∧ r (z, y) can be
mined computing the absolute number of adjacencies among q̄ and
r ; such values are the body support values for rules having a body
size of 2.
m 2 (p, r ) = |{(x, y, z) ∈ V 3 : (x, q̄, z), (z, r , y) ∈ G}|∀p, r ∈ Π
(6)
In the cases above, since free variables always appear at least
twice in a rule, the rules are called closed. Closed rules have been
preferred over open rules by existing algorithms, as they express
propositions found in natural language [4, 8]. Proceeding by induction, the maps above can be extended to mine closed rules having
an arbitrary size.
In our evaluation, we mine six different types of Horn clauses,
of which two are composed by two predicates (i.e., digons) and four
by three predicates (i.e., triangles):
(1) p(x, y) ⇐= q(x, y)
(2) p(x, y) ⇐= q(y, x)
(3) p(x, y) ⇐= q(x, z) ∧ r (z, y)
(4) p(x, y) ⇐= q(x, z) ∧ r (y, z)
(5) p(x, y) ⇐= q(z, x) ∧ r (z, y)
(6) p(x, y) ⇐= q(z, x) ∧ r (y, z)
(d )
4.2
Implementation and complexity analysis
HornConcerto was implemented2 in Python 2.7 and the Sparql
query language3 , using Numpy and the Joblib library for multithread processing. A Docker version of Virtuoso OpenSource 7.20
was used as triple store and query engine. These architectural
choices were motivated by the high availability of large knowledge
bases in the so-called Web of Data [1]. We therefore made HornConcerto compliant with the Resource Description Framework
(RDF) and the Sparql query language.
The computational complexity of our implementation excluding
constraints is linear with respect to the number of properties in
the graph. [21] proved Sparql patterns to be Pspace-complete.
However, the time complexity relies on the Sparql engine which
computes the queries. The number of graph patterns in queries
grows linearly with the size of the rule body. Since our queries
contain only AND and FILTER operators, they can be computed in a
→
−
→
−
time O(| B | · |E|), where B is the body size and E is the total number
of triples [21]. We provided HornConcerto with optimizations
based on caching, pruning, and parallel processing. As queries for
adjacencies might be repeated more than once, we introduced a
caching system which stores the values and avoids unnecessary
queries. Moreover, as rules are sorted by confidence value in descending order, the algorithm can easily prune search spaces. We
implemented a parallel-processing version which executes each
rule type on a different thread.
5
EVALUATION
In our setup, we evaluated the rule mining algorithms on four
different measures: (1) execution runtime, (2) memory consumption
(RAM and disk), (3) number of rules discovered, and (4) quality of
rules for link prediction.
2 The
source code is available online at the address https://github.com/mommi84/
horn-concerto/.
3 https://www.w3.org/TR/sparql11-query/
Submitted, KDD, 2018
Tommaso Soru, André Valdestilhas, Edgard Marx, and Axel-Cyrille Ngonga Ngomo
The datasets used are described in Table 1. D1-D2 are part of
a benchmark for link prediction [2]. Both datasets are divided
into three parts, i.e. training, validation, and test set. The sizes
shown in the table are the union of the training and validation sets,
while the test sets feature 5, 000 and 59, 071 triples respectively.
Datasets D3-D4 are instead excerpts of the DBpedia knowledge
graph [13]. D3-D4, unlike D1-D2 which were manually annotated,
are naturally incomplete and hence, the closed world assumption
does not apply. We compared HornConcerto with two state-ofthe-art rule mining approaches, i.e. OP and Amie+. We ran all
three algorithms using their default settings; for HornConcerto,
θ = 0.001, P = 200,T = 10. For a fair comparison in the runtime
and memory evaluation, we set all approaches to utilize the same
confidence thresholds. Finally, while HornConcerto and Amie+
expect only a triple dataset as input, OP expects an optional schema
containing domains and ranges for each property. We thus extracted
the values of domain and range for every property featuring this
information and added them to the OP input. The experiments were
carried out on a 8-core machine with 32 GB of RAM and running
Ubuntu 16.04.
Table 1: Datasets used in the evaluation.
Source
D1
D2
D3
D4
5.1
# triples # nodes # prop.
WordNet (WN18)
146,442
40,943
Freebase (FB15k)
533,144
14,951
DBpedia 2.0 Person
7,035,092 2,299,075
DBpedia EN 2016-04 397,831,457 5,174,547
18
1,345
10,341
63,764
Rule mining results
The results of the comparative evaluation on runtime, rules discovered, and memory consumption can be seen in Table 2. Only
HornConcerto could complete the computation on both datasets,
as Amie+ threw an out-of-memory error due to the large size of D4
and OP started but did not finish to fill out its DBMS tables within
5 days. HornConcerto discovered less rules (i.e., 1,401) on D4,
which can be considered a superset of D3. This is explained by the
fact that the confidence score depends on certain ratios in data;
increasing the dataset section does not necessarily lead to the same
ratio values.
To assess the scalability of our method, we executed it using
different numbers of CPUs. We evaluated how parallelism and the
use of CPUs affect the runtime of the rule mining process. Figure 1a
shows that the runtime remains almost the same until 4 CPUs, drops
at around 6, and then stabilizes again after 6 CPUs. The reason is
because we parallelize the computation on rule types, which are 6
in total.
5.2
Link prediction results
We compared HornConcerto with Amie+ and other state-of-theart approaches on the link prediction task using datasets D1 and
D2. D3 and D4 are not suited for link prediction, since there the
closed world assumption does not apply, meaning that a missing
edge is not necessarily false. As previously mentioned, the OP
algorithm relies on schemata to build rules from data. Being D1-D2
both schema-free datasets, all executions of OP returned 0 rules
discovered, forcing us to abort its evaluation on link prediction.
Then, we compared their performance with the aforementioned
TransE [2] and the following: HolE, a state-of-the-art algorithm
for link prediction based on holographic embeddings and neural
networks by [18]; ComplEx, a latent-factorization method by [26];
Analogy by [15], an approach to link prediction using analogies
among nodes embedded in a vector space; an ensemble learning
approach dubbed DistMult, by [11].
In order to evaluate the rules discovered by HornConcerto
and Amie+ on link prediction, we used four different settings. The
first three settings are based on the intuition which states that
the probability of a missing link is proportional to the confidence
scores of the rules involving its predicate. We chose three different
functions to compute this probability, i.e. average (Avg), maximum
(Max), and product of complements (Prod). These functions take
the vector of the confidence scores as input. The fourth setting
relies on MLNs, instead. We thus plugged the input datasets and
the rules discovered by HornConcerto and Amie+ to Mandolin4 ,
a framework based on MLNs.
Dataset D1-D2 are split into training, validation, and test set.
For each triple in the test set, corrupted triples are generated by
altering subjects and objects. Then, the probability of the triple
is computed using the functions described above and MLNs. The
mean reciprocal rank is obtained by inverting the average rank (by
descending probability) of a correct triple among the corrupted
ones. Hits@N measures the percentage of cases where a correct
triple ranks in the top-N triples. Link prediction results are shown
in Table 3. Results show that HornConcerto achieves the best
Hits@3 on D2 and state-of-the-art accuracies on dataset D1, where
the differences among the two rule mining approaches are relatively
small.
In Figure 2a, we plotted the number of relevant rules (i.e., rules
meeting the confidence threshold θ = 0.001) discovered over time
on dataset D3. We launched 5 different settings, varying the number
of top properties (P = 50, 100, 200) and the number of adjacencies
(T = 5, 10, 20). The configuration hP,T i which yielded the maximum number of rules is h200, 10i, while h50, 10i produced less
than one third of rules of the former. Increasing P or T trivially
leads to more rules discovered. As can be seen, all the curves have
a sudden braking at around t = 200s, which is a consequence of
having smaller (and thus simpler) rules be mined first.
In order to understand the effects of hyperparameters θ , P, and
T on the accuracy, we ran HornConcerto on D1 in a grid pattern. We computed the Hits@10 accuracy on D1, varying θ =
{0.0001, ..., 0.1} and P = {10, ..., 500}. Figure 2b shows that choosing a smaller θ of one order of magnitude has little to no influence
on the Hits@10 value; the same applies to T , which was set to 100
in figure. On the other hand, experiments with P ≥ 50 achieved a
score increment of +6% w.r.t. P = 10.
6
DISCUSSION
As pointed out in [9], implementing the PCA score in Sparql is
inefficient. In fact, we showed that the choice of using the PCA score
4 https://github.com/AKSW/Mandolin
Beyond Markov Logic: Efficient Mining of Prediction Rules in Large Graphs
Submitted, KDD, 2018
Table 2: Results on all datasets. In the disk memory consumption, the dataset itself is not included.
Dataset Approach
D1
D2
Rules RAM (GB) Disk (GB)
20 sec
151
0.1
< 0.1
----- No Schema Available ----12 sec
365
1
0.1
Amie+
OP
HornConcerto
2 h 45 min 45,427
0.1
< 0.1
----- No Schema Available ----2 h 01 min 17,585
8
1
Amie+
OP
HornConcerto
D3
> 5 days > 6,000
> 5 days
N/A
59 min
3,125
32
5
32
< 0.1
> 1, 000
1
Amie+
----- Out Of Memory ----OP
> 5 days
N/A
2
> 1, 000
HornConcerto 11 h 34 min
1,401
32
5.6
D4
Time in minutes
Runtime
Amie+
OP
HornConcerto
hits1
hits3
hits10
96
100
94
92
80
90
60
2
88
4
6
8
number of CPUs
0
0.5
Gibbs iter.
(a)
1
·109
(b)
Figure 1: Effects of parallelization on runtime (left) and Gibbs sampling iterations on accuracy (right) on dataset D3.
Table 3: Link prediction results on D1 and D2. Hits@N values are in %. For each measure, the best result is highlighted in bold.
D1 (WN18)
D2 (FB15k)
MRR Hits@1 Hits@3 Hits@10 MRR Hits@1 Hits@3 Hits@10
TransE
HolE
ComplEx
Analogy
DistMult E
0.495
0.938
0.941
0.942
0.790
11.3
93.0
93.6
93.9
78.4
88.8
94.5
94.5
94.4
N/A
94.3
94.9
94.7
94.7
95.0
0.463
0.524
0.692
0.725
0.837
29.7
40.2
59.9
64.6
79.7
57.8
61.3
75.9
78.5
N/A
74.9
73.9
84.0
85.4
90.4
Amie+ +Avg
Amie+ +Max
Amie+ +Prod
Amie+ +MLNs
0.961
0.970
0.927
0.892
95.9
96.8
91.9
89.2
96.2
97.1
92.9
94.3
96.5
97.1
94.6
96.0
0.333
0.352
0.336
0.307
29.8
32.2
30.6
30.7
33.4
35.3
33.7
36.7
39.5
39.8
38.7
39.8
HornConcerto +Avg
HornConcerto +Max
HornConcerto +Prod
HornConcerto +MLNs
0.963
0.971
0.941
0.904
95.9
96.9
91.8
90.4
96.6
97.3
96.4
94.2
97.0
97.4
97.2
95.9
0.479
0.810
0.508
0.224
41.2
79.2
44.7
22.4
50.6
81.9
52.7
31.4
60.3
83.6
61.9
37.2
instead of the standard confidence, even in systems not based on
Sparql query engines, can lead to scalability issues. Moreover, our
evaluation suggests that approaches which rely on schemata as the
only mean to achieve a good optimization still struggle with large
datasets. The concept of having to provide a property domain-range
schema falls into contradiction with the need for rules, since this
environment would more likely occur when no or few ontological
insights are available.
Submitted, KDD, 2018
P=50, T=10
P=100, T=10
P=200, T=10
P=100, T=5
P=100, T=20
7000
6000
relevant rules
Tommaso Soru, André Valdestilhas, Edgard Marx, and Axel-Cyrille Ngonga Ngomo
5000
4000
3000
2000
1000
0
149
599 1049 1499 1949 2399 2849
Time (seconds)
(b) Hits@10 accuracy
(a) Relevant rules
Figure 2
The link prediction results we obtained were the most interesting ones. Firstly, HornConcerto was able to perform better than
Amie+, in contrast with previous research where PCA was showed
being more effective than standard confidence on the extraction of
higher-quality rules. More interestingly, we achieved the highest
accuracy on Hits@N and MRR on D1 and Hits@3 on D2 just by using a simple maximization of the rule confidence scores. The reason
why dataset D2 was harder to learn might be that rules having a
longer body are needed to describe more complex relationships. The
fact that embedding-based approaches perform well could confirm
this hypothesis. However, these approaches are extremely slow at
generating links; as their models compute a probability value for a
given triple, yielding new links has a complexity of O(|V | 2 · |Π|). Finally, we surprisingly found that, independently on the rule mining
algorithm, rules alone can achieve higher accuracies on link prediction than cumbersome and resource-demanding approaches such
as Markov Logic Networks. Even after one billion Gibbs sampling
iterations (see Figure 1b), the links predicted through MLNs were
not as good as the ones found by our three simple functions. This
can pave the way to an investigation on the existence of boundaries
and theoretical limits that a rule-based prediction algorithm cannot
overcome, due to mathematical constraints.
7
CONCLUSION
In this paper, we presented HornConcerto, an algorithm for mining rules in large directed labelled graphs. Our Sparql-based implementation has shown an unprecedented ability to scale on large
datasets. In the rule mining task, HornConcerto achieved stateof-the-art performances in execution runtime and disk memory
consumption. While these accomplishments were fulfilled, our
algorithm managed to mine high-quality rules in datasets having
hundreds of millions of triples without the need of a schema and
achieve state-of-the-art link predictions on a widely-used benchmark. Our findings suggest that (i) rule-based algorithms are still
competitive on link prediction and (ii) rules alone can achieve
accuracies comparable to cumbersome and resource-demanding
approaches such as Markov Logic Networks.
REFERENCES
[1] Christian Bizer, Tom Heath, and Tim Berners-Lee. 2009. Linked data-the story
so far. Semantic services, interoperability and web applications: emerging concepts
(2009).
[2] Antoine Bordes, Nicolas Usunier, Alberto Garcia-Duran, Jason Weston, and
Oksana Yakhnenko. 2013. Translating embeddings for modeling multi-relational
data. In Advances in Neural Information Processing Systems.
[3] Lorenz Bühmann, Daniel Fleischhacker, Jens Lehmann, Andre Melo, and Johanna Völker. 2014. Inductive Lexical Learning of Class Expressions. In
Knowledge Engineering and Knowledge Management. https://doi.org/10.1007/
978-3-319-13704-9 4
[4] Yang Chen, Sean Goldberg, Daisy Zhe Wang, and Soumitra Siddharth Johri. 2016.
Ontological Pathfinding. In Proceedings of the 2016 International Conference on
Management of Data. ACM, 835–846.
[5] Yang Chen and Daisy Zhe Wang. 2014. Knowledge expansion over probabilistic
knowledge bases. In Proceedings of the 2014 ACM SIGMOD international conference
on Management of data. ACM.
[6] The UniProt Consortium. 2017. UniProt: the universal protein knowledgebase. Nucleic Acids Research 45, Database-Issue (2017), D158–D169. http:
//dblp.uni-trier.de/db/journals/nar/nar45.html#Consortium17a
[7] Luc Dehaspe and Hannu Toivonen. 2001. Discovery of relational association
rules. In Relational data mining. Springer, 189–212.
[8] Luis Galárraga, Christina Teflioudi, Katja Hose, and Fabian M Suchanek. 2015.
Fast rule mining in ontological knowledge bases with AMIE+. The VLDB Journal
24, 6 (2015).
[9] Luis Antonio Galárraga, Christina Teflioudi, Katja Hose, and Fabian Suchanek.
2013. AMIE: association rule mining under incomplete evidence in ontological
knowledge bases. In WWW. ACM.
[10] Dominik Jain and Michael Beetz. 2010. Soft evidential update via markov chain
monte carlo inference. In Annual Conference on Artificial Intelligence. Springer.
[11] Rudolf Kadlec, Ondrej Bajgar, and Jan Kleindienst. 2017. Knowledge Base Completion: Baselines Strike Back. arXiv preprint arXiv:1705.10744 (2017).
[12] Stanley Kok, Marc Sumner, Matthew Richardson, Parag Singla, Hoifung Poon,
Daniel Lowd, Jue Wang, and Pedro Domingos. 2009. The Alchemy system for
statistical relational {AI }. Technical Report. Department of Computer Science
and Engineering, University of Washington.
[13] Jens Lehmann, Chris Bizer, Georgi Kobilarov, Sören Auer, Christian Becker,
Richard Cyganiak, and Sebastian Hellmann. 2009. DBpedia - A Crystallization
Point for the Web of Data. Journal of Web Semantics 7, 3 (2009), 154–165. https:
//doi.org/doi:10.1016/j.websem.2009.07.002
[14] Luı́s Leitão, Pável Calado, and Melanie Weis. 2007. Structure-based inference of
XML similarity for fuzzy duplicate detection. In CIKM. ACM.
[15] Hanxiao Liu, Yuexin Wu, and Yiming Yang. 2017. Analogical Inference for
Multi-Relational Embeddings. arXiv preprint arXiv:1705.02426 (2017).
Beyond Markov Logic: Efficient Mining of Prediction Rules in Large Graphs
[16] Sofus A Macskassy and Foster Provost. 2005. Netkit-srl: A toolkit for network
learning and inference. In Proceeding of the NAACSOS Conference.
[17] Stephen Muggleton. 1995. Inverse entailment and Progol. New generation
computing 13, 3 (1995), 245–286.
[18] Maximilian Nickel, Lorenzo Rosasco, and Tomaso Poggio. 2015. Holographic
Embeddings of Knowledge Graphs. AAAI (2015).
[19] Feng Niu, Christopher Ré, AnHai Doan, and Jude Shavlik. 2011. Tuffy: Scaling
up statistical inference in markov logic networks using an rdbms. VLDB 4, 6
(2011).
[20] Jan Noessner, Mathias Niepert, and Heiner Stuckenschmidt. 2013. Rockit: Exploiting parallelism and symmetry for map inference in statistical relational
models. AAAI (2013).
[21] Jorge Pérez, Marcelo Arenas, and Claudio Gutierrez. 2009. Semantics and complexity of SPARQL. ACM Transactions on Database Systems (TODS) 34, 3 (2009),
16.
[22] Sebastian Riedel. 2008. Improving the accuracy and Efficiency of MAP Inference
for Markov Logic. In Proceedings of the 24th Annual Conference on Uncertainty in
AI (UAI ’08). 468–475.
[23] Stefan Schoenmackers, Oren Etzioni, Daniel S Weld, and Jesse Davis. 2010.
Learning first-order horn clauses from web text. In EMNLP.
[24] Saeedeh Shekarpour, Edgard Marx, Axel-Cyrille Ngonga Ngomo, and Sören Auer.
2014. SINA: Semantic Interpretation of User Queries for Question Answering on
Interlinked Data. Web Semantics 1 (2014), –.
[25] Claus Stadler, Jens Lehmann, Konrad Höffner, and Sören Auer. 2012. LinkedGeoData: A Core for a Web of Spatial Open Data. Semantic Web Journal 3, 4 (2012),
333–354. http://jens-lehmann.org/files/2012/linkedgeodata2.pdf
[26] Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume
Bouchard. 2016. Complex embeddings for simple link prediction. In International
Conference on Machine Learning. 2071–2080.
[27] Beat Wuthrich. 1995. Probabilistic knowledge bases. IEEE Transactions on
Knowledge and Data Engineering 7, 5 (1995), 691–698.
Submitted, KDD, 2018
| 2cs.AI
|
Disentangled Representations via Synergy Minimization
arXiv:1710.03839v1 [cs.LG] 10 Oct 2017
Greg Ver Steeg1 , Rob Brekelmans1 , Hrayr Harutyunyan2 , and Aram Galstyan1
Abstract— Scientists often seek simplified representations of
complex systems to facilitate prediction and understanding.
If the factors comprising a representation allow us to make
accurate predictions about our system, but obscuring any
subset of the factors destroys our ability to make predictions,
we say that the representation exhibits informational synergy.
We argue that synergy is an undesirable feature in learned
representations and that explicitly minimizing synergy can help
disentangle the true factors of variation in underlying data.
We explore different ways of quantifying synergy, deriving new
closed-form expressions in some cases, and then show how to
modify learning to produce representations that are minimally
synergistic. We introduce a benchmark task to disentangle
separate characters from images of words. We demonstrate
that Minimally Synergistic (MinSyn) representations correctly
disentangle characters while methods relying on statistical
independence fail.
I. I NTRODUCTION
Instead of the colors “blue” and “green”, imagine a
language with the concepts of “grue”, which describes
something that is green during the daytime and blue at night,
and “bleen”, which is blue by day and green by night [1].
The Grue language looks unnecessarily complex to us, as we
would have to describe a blue-eyed person as bleen-eyed by
day and grue-eyed at night. However, one can imagine that the
Grue-speaking people are familiar with two types of jellyfish:
a poisonous, grue-colored species and a non-poisonous, bleencolored species. They would find our language inconvenient
because of cumbersome warnings to avoid “green during the
day and blue at night”-jellyfish. Which set of concepts is
simpler ultimately depends on the types of observations one
is trying to characterize and this can be formalized with the
concept of informational synergy.
Synergy is colloquially defined as a situation where the
“whole is more than the sum of its parts.” To predict the
current visual appearance of a green object from a description
in the Grue language requires knowing whether the object
is grue or bleen and whether it is currently night or day.
Knowing either fact alone imparts no information. This
mirrors the canonical example of synergy, the XOR gate
on binary random variables with X = Z1 ⊕ Z2 , where either
input, Zj , alone is uninformative about X while both together
perfectly determine X.
The focus of this work is to explore whether synergy can be
useful in the context of unsupervised learning. Just as different
languages can exhibit more or less synergy while being
equally expressive, latent variable models can exhibit more or
1
University
Institute
of
Southern
California,
Information
Sciences
[email protected], [email protected],
[email protected]
2 Yerevan State University, [email protected]
less synergy while being equally predictive. We thus introduce
the principle of minimum synergy (MinSyn) for representation
learning. We expect minimally synergistic representations to
be more interpretable since each learned latent variable is
encouraged to be individually informative about predicted
observations. Disentangling factors of variation is often cited
as a goal for unsupervised learning [2], but pinning down this
hazy concept has proven difficult. Statistical independence of
latent factors is often used as a proxy for disentangling [3],
[4], but we demonstrate how independence-based approaches
can fail to recover true structure while an approach based on
synergy minimization succeeds.
After reviewing various attempts to define a well-behaved
information-theoretic measure of synergy [5], [6], [7], [8],
we identify a candidate measure which is well suited to
our unsupervised representation learning problem and derive
concrete formulations for the Gaussian and binary cases. We
introduce an intuitive benchmark task where we train a model
on images of handwritten words and check whether factors
are learned that correspond to individual characters. MinSyn
learning correctly disentangles the characters while other
methods fail. We conclude with a discussion of open questions
for using synergy to improve representation learning.
II. Q UANTIFYING S YNERGY
While there is no consensus on the correct way to measure
synergistic information, we can understand how a group of
random variables interact to predict a target variable using
the Partial Information Decomposition framework [8]. For a
given set of inputs Z1:m = {Z1 , ..., Zj , ...Zm }, we seek to
understand how the mutual information I(Z1:m : X),with
some target X, is distributed across predictors and subsets
of predictors. We seek to identify information, or reduction
in uncertainty about the target X [9], as unique, if it can be
achieved only by conditioning on a single given predictor,
redundant, if it is available in more than one predictor, or
synergistic, if it appears only from considering interactions
among two or more predictors. Much of the literature to
date on information decomposition has focused on finding
measures that maintain a number of technical properties
such as symmetry, monotonicity, and positivity, while also
matching intuition across a variety of canonical examples
(see [5] for a detailed discussion).
We now briefly introduce several candidate synergy measures and explore their properties before eventually deciding
to use Correlational Importance, or CI synergy, to guide
representation learning in Section II-B. To compare measures
and build intuition, Sec. II-C describes a simple example of
synergy for Gaussian variables.
A. Candidate Synergy Measures
Among proposed measures of synergy, Whole Minus Sum
synergy, arising in the neuroscience literature [10], directly
formalizes the intuition of information in the whole minus
information in the sum of the parts:
WMS(Z1:m ; X) ≡ I(Z1:m ; X) −
m
X
I(Zj ; X)
(1)
j=1
This measure is maximized when each individual Zj contains
no information about X, but taken together they contain full
information about X. However, this quantity can be negative
when there is redundant information that is counted multiple
times in the sum over singleton predictors. Thus, by observing
WMS synergy equal to zero, we cannot determine whether
no synergy is present or if there is a mixture of synergy and
redundancy [8].
To address these concerns, [5] instead consider synergy
as the “whole minus union of the parts,” where the union
information is taken with respect to a variational distribution
which solves the following optimization problem.
U (Z1:m ; X) =
min Iq (Z1:m : X)
q(z1:m ,x)
(2)
s.t. q(zj , x) = p(zj , x), ∀zj , x
B. Correlational Importance (CI) Synergy
We now consider Correlational Importance [11], [12], or CI
synergy, as a promising and tractable candidate measure for
use in representation learning. CI was originally presented to
characterize the relationship between an experimental stimulus
and observed neural responses resulting from the brain’s noisy
encoding procedure. The measure reflects the extent to which
the neural code stores information about a given stimulus
in individual neurons versus in correlations among multiple
neurons.
More concretely, [11], [12] calculate CI synergy as the
KL divergence between the true posterior of stimulus given
responses and a distribution pCI (x|z1:m ) which ignores
correlations in the encoding process:
C(Z1:m ; X) = DKL p(x|z1:m ) pCI (x|z1:m )
(4)
Q
p(x) j p(zj |x)
(5)
pCI (x|z1:m ) ≡
pCI (z1:m )
X
Y
pCI (z1:m ) ≡
p(x)
p(zj |x)
x
We use a subscript to denote mutual information with respect
to the variational distribution, q. The marginal constraints on q
ensure that the information contained in individual predictors
is unchanged. The minimization squeezes out any information
contained in higher order interactions among the Zj . With
this measure of union information, GK synergy is defined
as the information contained in Z that is not present in the
union of information from individual Zj ’s, as follows.
S(Z1:m ; X) ≡ I(Z1:m : X) − U (Z1:m ; X)
overly simplistic. Therefore, despite the theoretical appeal of
GK synergy, for the purpose of optimal representation learning
we are motivated to explore more practical measures.
(3)
While this measure performs nicely for several lowdimensional exemplar problems and fits many theoretical
desiderata [5], its applicability is limited by the intractable
optimization in Eq. 2 over all distributions consistent with
the marginals, for which we have no closed form solution.
To make GK synergy tractable for learning problems, we
considered a simpler scenario where we restrict distributions
to be Gaussian. In that case, we were able to derive the first
closed form expression for GK synergy, which is shown in
Appendix A. The synergy-minimizing distribution obtained
ends up looking quite simple, qGK (x|z1:m ) ∼ N (ρ∗ z∗ , 1 −
ρ2∗ ), where Z∗ denotes the predictor most correlated with X
and ρ∗ measures the Pearson correlation between X and the
predictor Z∗ . In other words, if we have multiple predicted
variables, Xi , synergy is minimized if each Xi depends on
only a single latent factor, Zparent(i) . An immediate corollary
provides an interesting theoretical connection between synergy
minimization and representation learning: Gaussian latent
tree models minimize (Gaussian) GK synergy of latent factors,
Z1:m , with respect to observed variables, Xi .
Latent tree models are difficult to learn, and the constraint
of having only one parent for each observed variable can be
j
CI synergy was shown by [12] and [13] to provide an upper
bound on the information loss to using the CI decoder
rather than the true maximum likelihood decoder [9] in
communicating over a noisy channel with input X and output
Z1:m .
A sufficient condition for CI to achieve a global minimum
of zero is that the encoding distribution follows theQ
conditional
independence assumptions and p(z1:m |x) =
j p(zj |x).
As shown in [14], this condition is not necessary and the
minimum can also be attained for non-factorized distributions
if, over the entire support of X:
p(z1:m ) Y
p(z1:m |x) =
p(zj |x)
pCI (z1:m ) j
We will explore a learning framework under the condition of
minimal CI synergy in Section III.
C. An Example of Gaussian Synergy
Given the simple linear structure of a multivariate Gaussian,
it may be surprising that these models can exhibit informational synergy. For a three dimensional joint distribution over
Z1 , Z2 , and X, what types of correlation structures will lead
to Z1 and Z2 being synergistic with respect to X?
Mutual information is invariant under scaling and translation of the marginals, so w.l.o.g. we take each variable to
have a mean of zero and a standard deviation of one. We use
brackets to indicate expectation values, defining ρj = Zj X
and Σ1,2 = Z1 Z2 . The marginal distributions between X
and each Zj are thus fully specified as
1 ρj
p(zj , x) ∼ N ~0,
ρj 1
����������� (����)
���
I(Z:X)
U(Z;X)
S(Z;X)
C(Z;X)
���
���
���
(a)
���
-���
�
(b)
���� ���
����
�
��
(a)
(b)
(c)
Fig. 1. For normal distributions over Z1 , Z2 , X, it is possible for Z1 , Z2 to contain synergistic information about X. The ellipsoids represent level sets
of normal distributions with covariance specified in Eq. 6 along with the marginal projections for each pair of variables. Between (a) and (b), the value of
Σ1,2 changes while ρ1 , ρ2 are unchanged. While (a) is synergistic, (b) exhibits zero GK synergy. (c) We show how mutual information, union information,
GK synergy, and CI synergy change over the entire range of Σ1,2 .
Further, the set of joint distributions compatible with
these marginal relationships can be written as q(z1 , z2 , x) ∼
N (~0, Σ), where the covariance matrix has the form
1
Σ = Σ1,2
ρ1
Σ1,2
1
ρ2
ρ1
ρ2
1
(6)
D. Independence versus Synergy
Statistical independence can be quantified with total correlation [15], a multivariate generalization of mutual information.
Dependence among latent variables Z1:m = {Z1 , Z2 , ..., Zm }
is characterized:
m
X
T C(Z1:m ) =
H(Zj ) − H(Z1:m )
(7)
j=1
For this to specify a valid covariance matrix, it must be
positive semidefinite which implies the constraints ρ2j ≤
1, Σ21,2 ≤
p1, and the less trivial constraint that Σ1,2 is between
ρ1 ρ2 ± (1 − ρ21 )(1 − ρ22 ).
If we choose the X-marginals to be ρ1 = 1/2 and ρ2 = 3/4,
the space of joint distributions is then parametrized only
by the correlation between Z1 and Z2 , Σ1,2 . We visualize
the joint induced by two different choices of Σ1,2 in Figure
1: a synergistic case with Σ1,2 = −.10 in 1(a), and a nonsynergistic case with Σ1,2 = .67 in 1(b). In both cases, the
ellipsoids represent the level sets of the distribution, with the
green representing the joint, and the vertical axis representing
the target X.
For the synergistic case in Figure 1(a), observe that,
specifying the value of one of the predictors, Z1 for example,
leaves a plane of possible distributions for the other two
variables. The joint distribution is a tilted plate, so the value
of X, or the coordinate of intersection with this plane on
the vertical axis, is not specified by Z1 alone. On the other
hand, knowledge of both predictors gives a tight estimate of
X since the plate is thin, and we thus expect any candidate
measure to give non-zero synergy.
In the non-synergistic case of Figure 1(b), the ellipsoid for
the joint distribution is fatter and more symmetric, meaning
that our uncertainty about X is similar whether conditioning
on a single predictor or both together. Indeed, there is zero
GK synergy in this case. We also plot how GK synergy and
CI synergy vary with Σ1,2 in Figure 1(c). While areas of
low and high synergy are qualitatively similar, the synergyminimizing distribution, which is the main interest of this
paper, differs for the two synergy measures.
This quantity is zero if and only if all Zj are independent, which can also be seen by a simple rearrangement of the entropy
terms to obtain a KL divergence
Q
DKL (p(z1:m )|| j p(zj )). Minimizing dependence between
latent variables is a commonly proposed method for achieving
disentangled representations [3], [4], [16], [17].
Representations comprised of statistically independent
latent factors differ from ones with minimal synergy. Consider
the classic example of the XOR gate, where we take Z1 , Z2 to
be iid random variables taking binary values with probability
half, and X = Z1 ⊕Z2 . By definition Z1 , Z2 are independent,
meaning T C(Z1:2 ) = 0, but all candidate measures for the
synergy of Z1:2 with respect to the prediction target, X,
take on a maximal value of 1 bit [5]. Differences in learned
representations are illustrated in Sec. IV.
III. M IN S YN L EARNING
For representation learning, we must shift from considering
a one-dimensional target X to a high-dimensional set of
input features X1:n = {X1 , ..., Xi , ..., Xn }. We consider
a standard auto-encoder architecture [18], with encoding
functions Zj = fj (X1:n ) transforming the data into some
latent factor space, and decoding functions X̄i = gi (Z1:m )
that approximately recover observations from latent representations. The functions can be parametrized, for instance,
as neural networks. Common training objectives for autoencoders minimize a reconstruction loss between the observed
variables, Xi , and the reconstructed ones, X̄i [19]. While the
optimal decoder with respect to the posterior probability of
Xi given Z1:m would be gi (z) = Ep (Xi |Z1:m = z), this is
typically intractable. Thus, we replace the expectation over
p with a variational distribution q, which we will take to
minimize the synergy among the latent factors Z1:m with
respect to each predicted variable, Xi .
We frame the synergy minimization problem over variational distributions q(xi , z1:m ) which approximate the true
distribution, p, as follows.
arg
min
Synergyq (Z1:m ; Xi )
s.t. ∀j
q(xi , zj ) = p(xi , zj )
q(xi |z1:m )
(8)
The marginal constraint ensures that the information that each
individual latent factor contains about a predicted variable is
unchanged, while the information contained jointly among
latent factors is tuned to minimize a synergy measure.
For mathematical tractability, we choose to minimize CI
synergy. Solving the optimization in Eq. 8 using the CI
synergy measure in Eq. 4, we see that the global minimizer
of CI = 0 is achieved for the distribution q(xi |z1:m ) =
pCI (xi |z1:m ) which satisfies the constraints of our optimization problem and depends only on marginal relationships
between Xi , Zj .
Given this framework for MinSyn learning, we just need
to obtain an explicit form for the decoder which minimizes
CI synergy. We provide expressions for MinSyn decoders for
the Gaussian and binary cases below.
A. CI Synergy for Gaussian Decoders
Assume the relationships between each latent variable Zj
and Xi are Gaussian. For simplicity and w.l.o.g. we give
expressions below for variables that have been standardized
to have zero mean and unit variance. The form of the CI
minimizing distribution, pCI (xi |z1:m ), is given in Eq. 5. The
distribution depends on the correlations ρij ≡ Xi Zj .
X ρij
1
1
(9)
pCI (xi |z1:m ) = N
zj ,
1 + Ri j 1 − ρ2ij
1 + Ri
P ρ2ij
where Ri =
j 1−ρ2ij . Now the CI synergy in Eq. 4 is
just defined as the KL divergence between two Gaussian
distributions, the actual distribution p(xi |z1:m ) and the
distribution assuming conditional independence above.
B. CI Synergy for Binary Decoders
In Sec. IV we will also show results where outputs are
considered as binary variables, so we give an expression for
a MinSyn decoder where all variables are binary. As above,
this takes the form of pCI (xi = 1|z1:m ), which implies the
following form for the decoder distribution.
X
pCI (xi = 1|z1:m ) = sig(
wi,j zj + bi )
(10)
j
where the sigmoid function, often used in neural networks,
is defined sig(v) = 1/(1 + e−v ) and w, b denote weights
and biases respectively [19]. Instead of explicitly training
w, b as in a typical neural network, these parameters are now
set according to the joint statistics of the observations and
encoded states as follows.
p(Xi = 1) X
p(Zj = 0|Xi = 1)
bi = log
+
log
p(Xi = 0)
p(Z
j = 0|Xi = 0)
j
wi,j = log
p(Zj = 1|Xi = 1)p(Zj = 0|Xi = 0)
p(Zj = 0|Xi = 1)p(Zj = 1|Xi = 0)
In both the binary and Gaussian cases, our MinSyn decoder
depends on the statistics of the pairwise marginals. It would
be time consuming to compute these for the entire dataset
after each gradient update. Therefore, we follow a similar
strategy to that used in Batch Normalization [20]. During
training, we calculate statistics only on the batch of samples
currently being used for a gradient update. At the same time,
we update a moving average of these statistics. During test
time, we used the learned average.
IV. R ESULTS
A. EMNIST Handwritten Words Dataset
We begin by introducing a new dataset and benchmark task
for disentangling natural factors of variation. A common task
in representation learning is to construct low-dimensional
latent factor representations of handwritten characters (like
the famous MNIST digits) that can be used to reconstruct
the original images. We extend this problem to consider
combinations of handwritten letters into words. What is the
most natural (disentangled) way to represent this type of data?
One way to represent this space is to include a latent factor for
each possible word. However, the number of possible words is
large so representing the data in this way could be inefficient.
Moreover, if new words are introduced, we will need to
add new factors and retrain the model. A more succinct
representation would learn about the existence of characters
and represent words as combinations of characters. The
number of factors required is then just the number required
to represent a character times the number of characters in the
word. Therefore, we hypothesize that character disentangled
representations will be better able to reconstruct new words
that were not seen in training.
We consider representations that recover a characterlevel representation from raw image data to have correctly
disentangled the natural factors of variation. It is common in
computer vision to build in assumptions of spatial locality
using convolutional layers and pooling. Since characters are
trivially distinguished through spatial localization, we can
only fairly use this benchmark to test disentangling on domainagnostic methods. Knowledge of the spatial layout of pixels
cannot be used.
ICA and latent factor generative models typically assume
independent latent factors. For this example, it is clear that
the character-level factors are not independent since not all
combinations of characters result in valid English words.
For instance if you see a word that starts with “t”, the
second letter is much more likely to be “h” than any other
character due to the ubiquity of the word “the” in the English
language. Therefore, we hypothesize that disentangling based
{a,b,d,f,l,p,s,t} x {a,e,i,n,o,r,s,u} x {d,e,m,n,r,s,t,y}
English words used in training (149 valid of 512 possible)
aid,aim,air,and,ans,ant,any,are,arm,arr,art,ass,bad,ban,bar
bas,bat,bay,bed,bee,bes,bet,bey,bid,bin,bis,bit,bod,bot,boy
brr,bud,bum,bun,bur,bus,but,buy,dad,dam,day,den,did,die,dim
din,dis,doe,don,dos,dot,dry,dud,due,dun,fad,fan,far,fas,fat
fay,fed,fee,fem,fen,fer,fey,fie,fin,fir,fit,foe,for,fry,fun
fur,fut,lad,lam,las,lat,lay,led,lee,let,lid,lie,lit,lot,pad
pan,par,pas,pat,pay,pee,pen,per,pet,pie,pin,pis,pit,pod,pot
pry,pun,pus,put,sad,sat,say,see,sen,set,sin,sir,sis,sit,sod
son,sos,sot,soy,sud,sue,sum,sun,tad,tam,tan,tar,tat,tee,ten
tie,tin,tis,tit,toe,tom,ton,tor,tot,toy,try,tum,tun,tut
Fig. 2.
Characters and words used to construct the training set.
on independence will perform poorly for this benchmark task.
For the current experiments, we construct a simple dataset
of handwritten word images composed of letters from the
EMNIST dataset [21] (see Figure 4 for several examples).
We first observe the 300 most commonly appearing English
words, and consider the 8 letters that appear most frequently
in the first, second, and third positions of these words. Any
of the 512 combinations of these letters which form valid
English words are included as a training sample, leading to a
dataset of 149 word images (Fig. 2). We construct a test set
using all remaining combinations of these letters, noting that
reconstruction performance on this set should be hampered
if our representations memorize words instead of learning
individual letters. In this experiment, we use identical image
samples each time a letter appears in a word to remove
variation in capitalization and writing style, an assumption
that would match a typewritten text scenario. This variation
could be included for added challenge in future experiments.
Code constructing this dataset can be found at [22], and can
be used as a template for exploring novel ways of sampling
words and letters. Just as in the Grue example, the ideal
disentangled representation may vary depending on details
of the relevant task.
B. Disentangled Representations from Handwritten Words
We now apply several latent factor learning methods to the
dataset of handwritten words generated as described in the
previous section. For each method, we use 9 latent factors
under the intuition that if a disentangled representation uses
3 latent factors to describe each character position, then these
three factors will be capable of distinguishing 23 = 8 different
characters in that position.
We considered several approaches including two methods
based on independence and two methods based on minimizing
synergy. The independence-based methods we used were
FastICA [3] and InfoMax [16], implemented using [23] and
[24], respectively. The MinSyn implementations are based on
the derivation of MinSyn decoders in the Gaussian and binary
cases in Sec. III. We also compare to an auto-encoder with
no synergy constraint. Finally, a common trick to improve
robustness of standard auto-encoders is to add Gaussian noise
to the inputs in training, a so-called denoising auto-encoder.
Implementations are described in detail in Sec. B.
1) Disentangled latent factors: In Figure 3, we visualize
decoder maps learned by different methods. Each method
has a decoder map for each of 9 latent factors, presented
in no particular order. Our expectation for disentangled
ICA Latent Factors
Fig. 3. Decoder weights for each of 9 latent factors learned by i) MinSyn
Binary, ii) FastICA, and iii) an auto-encoder without synergy constraints
representations is that each latent factor should correspond to
only a single character, and therefore the associated decoder
map will have nonzero weight in only one character position.
To quantify the overlap in decoder feature maps, we define
the average character concentration (ACC) score and report
it in Table IV-B.2. For a given factor j and pixel i, we have
decoder weights, wi,j . We denote the set of pixels that belong
to character k as Sk . We P
define the concentration
of weights
P 2
2
in character k as Cj,k = i∈Sk wi,j
/ i wi,j
. Finally, ACC
is defined as P
the P
average entropy across m latent factors,
ACC = 1/m j k −Cj,k log Cj,k . ACC would be zero if
each latent factor were concentrated on an individual character
position.
Figure 3 shows that standard auto-encoders produce very
entangled representations and looking at the ACC scores
in Table IV-B.2 shows that denoising does not help. As
hypothesized, the dependence between characters induced
by sampling English words causes entanglement of the ICA
sources, with learned features spanning multiple positions.
InfoMax fares even worse than ICA both visually (not shown)
and according to ACC score. While statistical independence
at first glance seems to be a natural way of quantifying disentanglement, this benchmark exemplifies how this intuition
can fail. In contrast, MinSyn learns qualitatively different
maps that both visually disentangle the characters and have
lower ACC scores than any other methods.
2) Reconstructing training and test data: One argument
in favor of character disentangled representations is that
Training data (English words) and reconstructions
Test data (not words) and reconstructions
Fig. 4. Train/test data shown in top row, with MinSyn reconstruction below.
they should in principle be equally capable of representing
any words comprised of the same characters, even if those
words were not present during training. Recall that we
trained on a set of valid three-letter English words. Using
the same characters, we can construct test “words” that
are not in the English dictionary and test the ability of
our latent factor representations to reconstruct them. We
visualize some words from the training and test set along
with their reconstructions using MinSyn in Fig. 4. In Table IVB.2, we compare the average L2 loss on training and
test data across methods. Minimizing synergy leads to the
best character disentanglement according to ACC and, as
hypothesized, the most disentangled method also produces the
best reconstruction results on both training and test data. Note
that ICA and InfoMax use PCA to reduce to 9 components,
before producing an orthogonal transformation into a basis
of independent components, and therefore the reconstruction
scores are identical even though the ACC’s differ. We also
found that MinSyn constrained representations were more
robust to different types of out-of-domain noise than their
unconstrained counterparts, but we omit those results here
due to space.
Method
Train loss
Test loss
ACC
MinSyn (Binary)
MinSyn (Gaussian)
ICA
InfoMax
PCA
Auto-encoder
Denoising
12.94
58.68
49.83
49.83
49.83
35.98
38.26
48.12
91.26
85.23
85.23
85.23
73.67
76.40
0.558
0.670
0.824
0.849
0.950
1.07
1.08
V. R ELATED WORK
Decomposing information in terms of unique, synergistic,
and redundant components is an active area of research
[25], [8], [5], with reviews of proposed measures appearing
in [6] and [7]. While synergy is rarely mentioned with regard
to representation learning, one motivation for exploring CI
synergy was a new approach to latent factor models demonstrating state-of-the-art results for clustering and covariance
estimation [26]; the approach combines total correlation
explanation [27], [28] with a condition that can be recognized
in the context of this paper as a CI synergy constraint.
Similarly to the notion of synergy, disentanglement in
representation learning has resisted a unique characterization,
with statistical independence [3], [16] and sparsity via L1
regularization representing the most common approaches
for achieving this goal [29], [2]. Another successful idea to
reduce the fragility of learned representations involves adding
noise to the latent factors (e.g. dropout). One can speculate
that dropout reduces synergies in representations since it
is not possible combine information from all latent factors
(some fraction of which will be dropped out) to produce a
given prediction. A recent approach [17] generalizes dropout
with learned multiplicative noise to approximate the Information Bottleneck objective [30]. Disentangled and minimally
sufficient P
representations are encouraged by minimizing the
objective j I(X1:n : Zj ) + T C(Z1:j ). Finally, other efforts
attempt to guide Generative Adversarial Networks toward
learning disentangled representations by maximizing mutual
information with latent codes that incorporate prior knowledge
about the data [31].
A line of work which superficially contradicts this one is the
“Integrated Information Theory of Consciousness” (IIT) [32].
This line of work proposes that what makes consciousness
(which we might call a learned neural representation) unique is
precisely the degree to which the whole is more than the sum
of the parts. A synergy measure is then introduced (integrated
information), and consciousness is posited to maximize
this synergy measure. While it appears that IIT proposes
representations that maximize synergy while we propose
representations that minimize synergy, these approaches
actually address different ideas. IIT is concerned with the
“internal synergy” of the brain as a system. In contrast, this
work always focuses on the synergy with respect to a given
prediction. It is possible that a system might have high
“internal synergy” as defined in IIT while having low synergy
with respect to particular predictions, hence IIT and MinSyn
are not directly comparable.
VI. C ONCLUSION
The absurdity of the Grue language mirrors a common
dysfunction in machine learning. Many factors (grue/bleen
and night/day) have to be combined to describe simple
observations (visual appearance of an object) and therefore
individual factors (like grue) do not seem very meaningful
to us. Concepts like blue/green seem more natural and we
suggest that informational synergy provides a way to quantify
this intuition. This presents an opportunity to leverage the
long history of academic interest in synergy measures to
quantify and eliminate undesirable and unnecessary synergies in representation learning, hopefully leading to more
interpretable models.
This work provides a first step in exploring the promise
of reducing synergies in learning, however a number of
challenges remain. Decades of research on synergy measures
have not produced a consensus so finding the best measure for
learning is an open problem. Because this field has remained
theoretical, even popular synergy measures are difficult to estimate with little effort expended on tractable approximations.
Existing work focuses on identifying synergies in observed
data, while we suggest a fundamental shift in focus to using
show their objective values coincide. For the dual, let k =
arg maxi ρi . Set all vi = ui = 0, except vk = ρk , uk = ρ2k .
This satisfies the constraints and gives an objective of ρ2k . For
the primal program, we propose the following feasible point.
In ρ
M̄ =
+B
(13)
ρ> ρ2k
synergy as a principle for designing representations. Finally,
while we minimized synergy in this work, it could be that
reducing synergies with a tunable regularization strength
would be more effective and that some amount of synergy is
desirable or even necessary in some cases.
Despite many open questions, we showed that at least
one measure, CI synergy, is tractable and straightforward to
apply to representation learning. We showed that MinSyn
led to the learning of qualitatively different representations.
We presented a benchmark task for disentangling factors
of variation in data and and found that MinSyn learning
succeeded where other approaches failed. While this proof
of concept is encouraging, the difficulties we highlighted
indicate ample room for improvement.
The symmetric matrix B is defined so that ∀i = 1, . . . , k −
1, k + 1, . . . , n, Bi,k = ρi /ρk , or zero otherwise. Inspection
of the matrix reveals that it is positive semidefinite and the
value of the objective is ρ2k .
Finally, we can write the expressions for union information
and GK synergy in terms of the correlation, ρk , with the
largest magnitude.
A PPENDIX
U (Z1:n ; X) = I(Zk ; X) = 1/2 log(1 − ρ2k )
(14)
1 − ρ> Σ−1 ρ
S(Z1:n ; X) = 1/2 log
1 − ρ2k
(15)
A. Derivation of GK Synergy for the Gaussian Case
a) GK synergy for Gaussians: We have seen that
measuring GK synergy depends on a difficult optimization
problem to find the union information. However, if we restrict
ourselves to Gaussian distributions, we are able to obtain an
analytic expression for GK synergy. We begin by recasting the
optimization in Eq. 2 after restricting ourselves to Gaussian
distributions. The scaled and centered Gaussian distributions
are completely characterized by a covariance matrix, M ,
which must be positive semi-definite. We take the sub-matrix
that is the covariance only for Z1:n to be Σ. The correlations
between Zi and X are ρi . The formula for mutual information
is expressed in terms of the determinants of these matrices,
with marginal constraints imposed on the covariance.
U (Z1:n ; X) = min
M 0
s.t.
det Σ
det M
Mii = 1, Mi,n+1 = ρi , ∀i
1/2 log
(11)
Using the Schur complement, we can re-write the objective
as −1/2 log(1 − ρ> Σ−1 ρ) which is equivalent to minimizing
ρ> Σ−1 ρ. We rewrite the minimization involving a matrix
inverse using a trick. We construct a matrix, M̄ , which is the
same as M but with M̄n+1,n+1 = t.
min t
M̄ 0
s.t.
M̄ii = 1, M̄i,n+1 = ρi , i = 1, . . . , n
(12)
Looking at the Schur complement for the condition M̄ 0,
we see that minimizing t is equivalent to minimizing ρ> Σ−1 ρ.
Now that we have a canonical SDP we can use standard
methods to construct a dual program. Feasible points of the
dual provide a lower bound on the solution to our primal
problem, Eq. 12, and vice versa [33]. If we find a feasible
point for the dual and primal whose objective value coincides,
then this point is the optimum for both programs. The dual
is as follows.
u1 . . . 0 −v1
X
min
2vi ρi − ui s.t. ...
un −vn 0
u,v∈Rn
i
−v1 −vn
1
To prove the optimum of the primal problem, we give a
feasible point for both the dual and primal problems and
If we now consider optimizing this quantity over some
variational distribution q matching the pairwise marginals,
qGK (z1:n , x) = arg min S(Z1:m ; X) s.t.
q(z1:n ,x)
p(zi , x) = q(zi , x)
we see that U (Z1:n ; X) and I(Zi ; X) are both constant for
given marginals (since they depend only on ρ). Therefore, GK
synergy reduces to minimizing mutual information subject to
the marginal constraints. While we cannot solve this problem
in general, we can again turn to the special case of Gaussian
distributions for a solution. This optimization is exactly the
one we had to solve to calculate the union information in
Eq. 11. We found and proved the optimal point using SDPs.
The solution can be expressed in terms of the covariance
matrix over X, Σ, which can be read from Eq. 13. For
simplicity and w.l.o.g., we re-order the ρi so that ρ1 has the
largest magnitude. In that case, qGK (z1:n ) is defined by its
covariance matrix as follows.
1
ρ2 /ρ1 ... ρn /ρ1
ρ2 /ρ1
1
0
0
Σ= .
(16)
.
..
..
0
0
ρn /ρ1
0
0
1
As a sanity check, we can use Eq. 16 to calculate that the
value of the minimum of S(Z1 , Z2 ; X) that appears in the
plot in Fig. 1 should be Σ1,2 = ρ2 /ρ1 = 1/2/3/4 = 2/3. A
matrix in this special form is called an arrowhead matrix.
The inverse can be found using standard identities1 and is a
diagonal matrix plus a rank one matrix. This inverse is needed
to complete our analysis since for Gaussians, qGK (x|z) ∼
N (ρ> Σ−1 z, 1 − ρ> Σ−1 ρ).
qGK (X|Z1:n = z) ∼ N ρ1 z1 , 1 − ρ21 ,
(17)
After all this work, we come to the trivial solution that
the least synergistic relationship according to MI synergy
measures is to take the best predictor and throw the others
1 If more than one ρ have the same, maximal magnitude, the inverse does
i
not exist and the space of solutions is degenerate.
away! This surprising conclusion holds for Gaussians, but
not in the more general case, where no analytic form for GK
synergy is known. In the Gaussian case, if we have many
variables to predict, Xi , then minimizing GK synergy implies
that each should depend on only a single latent factor. This is
equivalent to a Gaussian latent tree model. The implications
for learning for minimizing GK synergy in the general case
remain an open question.
B. Implementation Details
All models were implemented using TensorFlow [34] and
optimized using Adam [35] with a learning rate 0.001.
1) MinSyn Gaussian: We consider an auto-encoder with
learned encoder weights and biases
w, b, and a softplus
P
activation function zj = softplus( i wj,i xi +bi ), and decoder
weightsPand biases u, d according to the linear decoder
x̄i =
j ui,j zj + dj , trained to minimize the expected
squared reconstruction error, E[|X −X̄|22 ]. The only difference
between our implementation and a standard one is that instead
of learning the decoder weights as separate parameters, they
ρij
1
are set according to Eq. 9: ui,j = 1+R
2 .
i 1−ρij
2) MinSyn Binary: In the binary case, we re-scale pixel
grayscale values into the range [0, 1] and interpret them as
probabilities for a binary random variable to be on or off. The
encoded state, Z1:m is constructed as a sigmoid of a linear
function of the inputs. The decoder is specified similarly,
except with weights specified as in Eq. 10. We use the cross
entropy loss as is typical for binary random variables.
3) Auto-encoders: For comparison, we also consider a
fully-connected auto-encoder where the decoder weights are
trained rather than fixed according to the CI-minimizing
distributions above. We use a softplus non-linearity in the
encoder and a sigmoid activation in the decoder, training to
minimize cross entropy loss. For the denoising auto-encoder,
we add Gaussian noise with σ = .25 to each input feature.
ACKNOWLEDGMENT
We thank Jimmy Foulds, Virgil Griffith, and David Kale for
comments on this project. The authors acknowledge support
from from DARPA grants W911NF-16-1-0575 and FA875017-C-0106.
R EFERENCES
[1] N. Goodman, Fact, fiction and forecast. JSTOR, 1955, vol. 74.
[2] Y. Bengio, A. Courville, and P. Vincent, “Representation learning:
A review and new perspectives,” Pattern Analysis and Machine
Intelligence, vol. 35, no. 8, pp. 1798–1828, 2013.
[3] A. Hyvärinen and E. Oja, “Independent component analysis: algorithms
and applications,” Neural networks, vol. 13, no. 4, pp. 411–430, 2000.
[4] L. Dinh, D. Krueger, and Y. Bengio, “Nice: Non-linear independent
components estimation,” arXiv preprint arXiv:1410.8516, 2014.
[5] V. Griffith and C. Koch, “Quantifying synergistic mutual information,”
in Guided Self-Organization: Inception. Springer, 2014, pp. 159–190.
[6] E. Olbrich, N. Bertschinger, and J. Rauh, “Information decomposition
and synergy,” Entropy, vol. 17, no. 5, pp. 3501–3517, 2015.
[7] N. Timme, W. Alford, B. Flecker, and J. M. Beggs, “Synergy,
redundancy, and multivariate information measures: an experimentalist’s
perspective,” Journal of computational neuroscience, vol. 36, no. 2,
pp. 119–140, 2014.
[8] P. Williams and R. Beer, “Nonnegative decomposition of multivariate
information,” arXiv:1004.2515, 2010.
[9] T. M. Cover and J. A. Thomas, Elements of information theory. WileyInterscience, 2006.
[10] E. Schneidman, W. Bialek, and M. J. Berry, “Synergy, redundancy,
and independence in population codes,” the Journal of Neuroscience,
vol. 23, no. 37, pp. 11 539–11 553, 2003.
[11] S. Nirenberg, S. M. Carcieri, A. L. Jacobs, and P. E. Latham, “Retinal
ganglion cells act largely as independent encoders,” Nature, vol. 411,
no. 6838, pp. 698–701, 2001.
[12] P. E. Latham and S. Nirenberg, “Synergy, redundancy, and independence
in population codes, revisited,” The Journal of neuroscience, vol. 25,
no. 21, pp. 5195–5206, 2005.
[13] N. Merhav, G. Kaplan, A. Lapidoth, and S. S. Shitz, “On information
rates for mismatched decoders,” IEEE Transactions on Information
Theory, vol. 40, no. 6, pp. 1953–1967, 1994.
[14] S.-i. Amari and H. Nakahara, “Correlation and independence in the
neural code,” Neural computation, vol. 18, no. 6, pp. 1259–1267, 2006.
[15] S. Watanabe, “Information theoretical analysis of multivariate correlation,” IBM Journal of research and development, vol. 4, no. 1, pp.
66–82, 1960.
[16] A. J. Bell and T. J. Sejnowski, “An information-maximization approach
to blind separation and blind deconvolution,” Neural computation, vol. 7,
no. 6, pp. 1129–1159, 1995.
[17] A. Achille and S. Soatto, “Information dropout: learning optimal
representations through noise,” arXiv preprint arXiv:1611.01353, 2016.
[18] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol, “Extracting
and composing robust features with denoising autoencoders,” in
Proceedings of the 25th international conference on Machine learning.
ACM, 2008, pp. 1096–1103.
[19] I. Goodfellow, Y. Bengio, and A. Courville, Deep learning. MIT
Press, 2016.
[20] S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep
network training by reducing internal covariate shift,” arXiv preprint
arXiv:1502.03167, 2015.
[21] G. Cohen, S. Afshar, J. Tapson, and A. van Schaik, “Emnist: an extension of mnist to handwritten letters,” arXiv preprint arXiv:1702.05373,
2017.
[22] R. Brekelmans, “Emnist handwritten character disentanglement
benchmark,” software available from https://github.com/brekelma/
character-disentanglement-emnist.
[23] F. Pedregosa, et al., “Scikit-learn: Machine learning in Python,” Journal
of Machine Learning Research, vol. 12, pp. 2825–2830, 2011.
[24] A. Gramfort, M. Luessi, E. Larson, D. A. Engemann, D. Strohmeier,
C. Brodbeck, L. Parkkonen, and M. S. Hmlinen, “{MNE} software
for processing {MEG} and {EEG} data,” NeuroImage, vol. 86, pp.
446 – 460, 2014. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/S1053811913010501
[25] R. G. James, J. Emenheiser, and J. P. Crutchfield, “Unique information
via dependency constraints,” arXiv:1709.06653, 2017.
[26] G. Ver Steeg and A. Galstyan, “Low complexity gaussian latent factor
models and a blessing of dimensionality,” arXiv:1706.03353, 2017.
[27] G. Ver Steeg and A. Galstyan, “Discovering structure in highdimensional data through correlation explanation,” in Advances in
Neural Information Processing Systems (NIPS), 2014.
[28] G. Ver Steeg and A. Galstyan, “Maximally informative hierarchical
representations of high-dimensional data,” in Artificial Intelligence and
Statistics (AISTATS), 2015.
[29] R. Tibshirani, “Regression shrinkage and selection via the lasso,”
Journal of the Royal Statistical Society. Series B (Methodological), pp.
267–288, 1996.
[30] N. Tishby, F. C. Pereira, and W. Bialek, “The information bottleneck
method,” arXiv preprint physics/0004057, 2000.
[31] X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever, and
P. Abbeel, “Infogan: Interpretable representation learning by information maximizing generative adversarial nets,” in Advances in Neural
Information Processing Systems, 2016, pp. 2172–2180.
[32] M. Oizumi, L. Albantakis, and G. Tononi, “From the phenomenology
to the mechanisms of consciousness: integrated information theory 3.0,”
2014.
[33] L. Vandenberghe and S. Boyd, “Semidefinite programming,” SIAM
review, vol. 38, no. 1, pp. 49–95, 1996.
[34] M. Abadi, et al., “TensorFlow: Large-scale machine learning on
heterogeneous systems,” 2015, software available from tensorflow.org.
[Online]. Available: http://tensorflow.org/
[35] D. Kingma and J. Ba, “Adam: A method for stochastic optimization,”
arXiv preprint arXiv:1412.6980, 2014.
| 7cs.IT
|
Representation Learning using Event-based STDP
Amirhossein Tavanaeia , Timothée Masquelierb , Anthony Maidaa
arXiv:1706.06699v3 [cs.NE] 9 Mar 2018
a
The School of Computing and Informatics
University of Louisiana at Lafayette, Lafayette, LA 70504, USA
Email: {tavanaei,maida}@louisiana.edu
b
CERCO UMR 5549,
CNRS-Université de Toulouse 3, F-31300, France
Email: [email protected]
Abstract
Although representation learning methods developed within the framework of traditional neural
networks are relatively mature, developing a spiking representation model remains a challenging problem. This paper proposes an event-based method to train a feedforward spiking neural
network (SNN) layer for extracting visual features. The method introduces a novel spike-timingdependent plasticity (STDP) learning rule and a threshold adjustment rule both derived from
a vector quantization-like objective function subject to a sparsity constraint. The STDP rule
is obtained by the gradient of a vector quantization criterion that is converted to spike-based,
spatio-temporally local update rules in a spiking network of leaky, integrate-and-fire (LIF) neurons. Independence and sparsity of the model are achieved by the threshold adjustment rule and
by a softmax function implementing inhibition in the representation layer consisting of WTAthresholded spiking neurons. Together, these mechanisms implement a form of spike-based,
competitive learning. Two sets of experiments are performed on the MNIST and natural image
datasets. The results demonstrate a sparse spiking visual representation model with low reconstruction loss comparable with state-of-the-art visual coding approaches, yet our rule is local in
both time and space, thus biologically plausible and hardware friendly.
Keywords: Representation Learning, spiking neural networks, quantization, STDP,
bio-inspired model
1. Introduction
Unsupervised learning approaches using neural networks have frequently been used to extract features from visual inputs [1, 2]. Single layer networks using distributed representations
or autoencoder networks [3, 4] have offered effective representation platforms. However, the robust, high level, and efficient representation that is obtained by networks in the brain is still not
fully understood [5, 6, 7, 8, 9, 10, 11]. Understanding the brain’s functionality in representation
learning can be accomplished by studying spike activity [12] and bio-inspired spiking neural
Preprint submitted to Neural Networks
March 12, 2018
networks (SNNs) [13, 14, 15]. SNNs provide a biologically plausible architecture, high computational power, and an efficient neural implementation [16, 17, 18]. The main challenge is to
develop a spiking representation learning model that encodes input spike trains to uncorrelated,
sparse, output spike trains using spatio-temporally local learning rules.
In this study, we seek to develop representation learning in a network of spiking neurons to
address this challenge. Our contribution determines novel spatio-temporally local learning rules
embedded in a single layer SNN to code independent features of visual stimuli received as spike
trains. Synaptic weights in the proposed model are adjusted based on a novel spike-timingdependent plasticity (STDP) rule which achieves spatio-temporal locality.
Nonlinear Hebbian learning has played a key role in the development of a unified unsupervised learning approach to represent receptive fields [19]. Földiák [20], influenced by Barlow
[21], was one of the early designers of sparse, weakly distributed representations having low redundancy. Földiák’s model introduced a set of three learning rules (Hebbian, anti-Hebbian, and
homeostatic) to work in concert to achieve these representations. Zylberberg et al. [22] showed
that Földiák’s plasticity rules, in a spiking platform, could be derived from the constraints of reconstructive accuracy, sparsity, and decorrelation. Furthermore, the acquired receptive fields of
the representation cells in their model (named SAILnet) qualitatively matched those in primate
visual cortex. The representation kernels determining the synaptic weight sets have been successfully utilized by our recent study [23] for a spiking convolutional neural network to extract
primary visual features of the MNIST dataset. Additionally, the learning rules only used information which was locally available at the relevant synapse. Although SAILnet utilized spiking
neurons in the representation layer and the plasticity rules were spatially local, the learning rules
were not temporally local. The SAILnet plasticity rules use spike counts accumulated over the
duration of a stimulus presentation interval. Since the SAILnet rules do not use spike times, the
question of training the spiking representation network using a spatio-temporally local, spikebased approach like spike-timing-dependent plasticity (STDP) [24], which needs neural spike
times, remains unresolved. Later work, [25], extends [22] to use both excitatory and inhibitory
neurons (obeying Dale’s law), but the learning rules still use temporal windows of varying duration to estimate spike rates, rather than the timing of spike events. Our work seeks to develop
a learning rule which matches this performance but remains local in both time and space.
In another line of research based on cost functions, Olshausen and Field [26] and Bell and
Sejnowski [27] showed that the constraints of reconstructive fidelity and sparseness, when applied
to natural images, could account for many of the qualitative receptive field (RF) properties of
primary visual cortex (area 17, V1). These works were agnostic about the possible learning
mechanisms used in visual cortex to achieve these representations. Following [26], Rehn and
Sommer [28] developed the sparse-set coding (SSC) network which minimizes the number of
active neurons instead of the average activity measure. Later, Olshausen et al. [29] introduced an
L1 -norm minimization criterion embedded in a highly overcomplete neural framework. Although
these models offer great insight into what might be computed when receptive fields are acquired,
they do not offer insight into details of the learning rules used to achieve these representations.
Early works that proposed a learning mechanism to explain the emergence of orientation
2
selectivity in visual cortex are those of von der Malsburg [30] and Bienenstock et al. [31]. A
state-of-the-art model is that of Masquelier [32]. This model blends strong biological detail with
signal processing analysis and simulation to establish a proof-of-concept demonstration of the
original Hubel and Wiesel [33] feedforward model of orientation selectivity. A key feature of that
model, relevant to the present paper, is the use of STDP to account for RF acquisition. STDP
is the most popular learning rule in SNNs in which the synaptic weights are adapted according
to the relative pre- and postsynaptic spike times [24, 34]. Different variations of STDP have
shown successful visual feature extraction in layer-wise training of SNNs [35, 36, 37, 38]. In a
similar vein, Burbank [39] has also proposed an STDP-based autoencoder. This autoencoder
uses a mirrored pair of Hebbian and anti-Hebbian STDP rules. Its goal is to account for the
emergence of symmetric, but physically separate, connections for encoding weights (W ) and
decoding weights (W T ).
Another component playing a key role in representing uncorrelated visual features in a bioinspired SNN pertains to the inhibition circuits embedded within a layer. For instance, Savin
et al. [40] developed an independent component analysis (ICA) computation within an SNN
using STDP and synaptic scaling in which independent neural activities in the representation
layer were controlled by lateral inhibition. Lateral inhibition established a winner(s)-take-all
(WTA) neural circuit to maintain the independence and sparsity of the neural representation
layer. More recent work [41] has combined a layer of unsupervised STDP with an explicit layer
of non-learning inhibitory neurons. The inhibitory neurons impose a WTA discipline. Their
representations were tested on the handwritten MNIST dataset and have been shown to be
effective for recognition of such digits. The acquired representations tended to resemble MNIST
prototypes, although their reconstructive properties were not directly studied. [42] also studied
a spiking network with stochastic neurons that performs MNIST classification and acquires
MNIST prototype representations. Their architecture is a 3-layer network where the hidden
layer uses a soft WTA to implement inhibition. Since there is no functional need to introduce
an explicit inhibitory layer if there is no learning, our work uses a softmax function [43, 44] to
achieve WTA inhibition. In our work, the standard softmax is adapted to a spiking network.
Our acquired representations, when trained on the MNIST dataset, acquires representations
resembling V1-like receptive fields, in contrast to the MNIST prototypes of the research described
above.
Other works related to spike-based clustering and vector quantization are the evolving SNNs
(eSNNs and deSNNs) of [45, 46, 47, 48, 49] which acquire representations via a recruitment
learning paradigm [50] where neurons are recruited to participate in the representation of the
new pattern (based on similarity or dissimilarity to preexisting representations). In the deSNN
framework, if a new online pattern is sufficiently similar to an already represented pattern, the
representations are merged to form a cluster. This later work uses a number of bio-plausible
mechanisms, including spiking neurons, rank-order coding [51], a variant of STDP, and dynamic
synapses [52].
The present research proposes event-based, STDP-type rules embedded in a single layer SNN
for spatial feature coding. Specifically, this paper proposes a novel STDP-based representation
3
learning method in the spirit of [32, 39, 22]. Its learning rules are local in time and space
and implement an approximation to clustering-based, vector quantization [53] using the SNN
while controlling the sparseness and independence of visual codes. Local in time means that
the information to modify the synapse is recent, say within at most a couple of membrane time
constant of the postsynaptic spike that triggers the STDP. By local in space, we mean that
the information used to modify the synaptic weight is, in principle, available at the presynaptic
terminal and the postsynaptic cell membrane. Our derivation uses a continuous-time formulation
and takes the limit as the length of the stimulus presentation interval tends to one time step.
This leads to STDP-type learning rules, although they differ from the classic rules found in [32]
and [34]. In this sense, the rules and resulting visual coding model are novel. Independence
and sparsity are also maintained by an implicit inhibition and a new threshold adjustment rule
implementing a WTA circuit.
2. Background
Földiák [54] developed a feedforward network with anti-Hebbian interconnections for visual
feature extraction. The Hebbian rule in his model, shown in Eq. 1, is inspired from Oja’s learning
rule [55] that extracts the largest principal component from an input sequence,
∆wji ∝ (yj xi − wji yj2 )
(1)
X
(2)
yj =
xi wji
i
where, wji is the weight associated with the synapse connecting input (presynaptic) neuron i and
representation (postsynaptic) unit j. xi and yj are input and linear output respectively. Over
repeated trials, the term yj xi increases the weight when the input and the output are correlated.
The second term (−wji yj2 ) maintains the learning stability [54]. With respect to binary (or
spiking) units, a more appropriate assumption was made by Földiák [20]. He modified the
previous feedforward network by incorporating non-linear threshold units in the representation
layer. The units are binary neurons with a threshold of 0.5 in which yj ∈ {0, 1} (Note: yj2 = yj ).
Thus, the Hebbian rule in Eq. 1 is simplified to
∆wji ∝ yj (xi − wji ).
(
P
1,
i xi wji > 0.5
yj =
0, otherwise
(3)
(4)
The weight change rules defined in Eqs. 1 and 3 are based on the input and output correlation.
Another interpretation for Eq. 3 can be explained in terms of vector quantization (or clustering
in a WTA circuit) [56, 57] in which the weights connected to each output neuron represent particular clusters (centroids). The weight change is also affected by the output neuron activation,
yj . In this paper, we utilize the vector quantization concept to define an objective function. The
4
objective function can be adapted to develop a spiking visual representation model equipped
with a temporally local learning rule while still maintaining sparsity and independence. Our
motivation is to use event-based, STDP-type learning rules. This requires the learning to be
temporally local, specifically using spike times between pre- and postsynaptic neurons.
3. Spiking Visual Representation
The proposed model adopts a constrained optimization approach to develop learning rules
that are synaptically local. The spiking representation model is a single layer SNN shown in
Fig. 1. The representation layer recodes a p × p image patch (p × p spike trains) using D spike
trains generated by neurons, zj , in the representation layer.
We derive plasticity rules that operate over a stimulus presentation interval T (non-local in
time) and then take the limit as T tends to one local time step to derive event-based rules. In
the case of a linear unit, yj , the objective function to be minimized is shown below. It uses both
the vector quantization criterion and a regularizer that prefers small weight values.
X
2
F (xi , wji ) = yj (xi − wji )2 + yj λwji
,
yj =
xi wji
(5)
i
The variables xi , yj , wji ≥ 0 denote: normalized input pixel intensities in the range [0, 1], the
linear output activation, and the excitatory synaptic weight, respectively. The first component
shows a vector quantization criterion that is scaled by the output neuron’s activity, yj . The
yj scales the weight update rule according to the neuron’s response to the input pattern (xi ).
The second component (regularizer) is also scaled by the output neuron’s activity to control the
weight decay criterion (e.g. if yj = 0, wji does not undergo learning). We assume that the input
and output values can be converted to the spike counts over T ms. The hyperparameter λ ≥ 0
controls the model’s relative preference for smaller weights. As λ → 0, the objective function
emphasizes the vector quantization criterion. In contrast, as λ → ∞ the vector quantization
component is eliminated and the minimum of the objective function is obtained when the wji ’s
→ 0.
In response to a stimulus presentation, a subset of spiking neurons in the representation
layer is activated to code the input. To represent the stimuli by uncorrelated codes, the neurons
should be activated independently and sparsely. That is, the representation layer demands a
WTA neural implementation. This criterion can be achieved by a soft constraint such that
X
X
g(x) =
zj ≤ 1 ⇒ 1 −
zj ≥ 0 .
(6)
j
j
where, zj shows the binary state of unit j after the T ms presentation interval such that zj = 1
if unit j fires at least once. Also, the firing status of a neuron can be controlled by its threshold,
θ. Therefore, this constraint can be addressed by a threshold adjustment rule.
5
Figure 1: Spiking representation network. p × p image patch encoded by D spike trains in the representation
layer. W shows the synaptic weight sets corresponding to the D kernels.
The goal is to minimize the objective function (Eq. 5) while maintaining the constraint
(Eq. 6). This can be achieved by using a Lagrangian function
X
2
L(xi , yj , z; wji , α) = yj (xi − wji )2 + yj λwji
− α(1 −
zj )
(7)
|
{z
}
j
Objective Function
|
{z
}
Constraint
where, α is a Lagrange multiplier. Minimizing the first component of Eq. 7 results in a coding
module that represents the input by a new feature vector which can cluster the data via the
synaptic weights. Minimizing the second component supports the sparsity and independence
of the representation to finally (as a special case) end with a winner-take-all network in which
exactly one neuron fires upon stimulus presentation. This matter is accomplished by adapting
the neuron’s threshold, θ = −α. The optimum of the Lagrangian function can be obtained by
gradient descent on its derivatives
∂L
= −2yj (xi − wji ) + 2yj λwji
∂wji
(8)
X
∂L
∂L
=−
=−
zj − 1
∂θ
∂α
(9)
j
From gradient descent on Eq. 8 (reversing the sign on the derivative), we obtain
∆wji ∝ yj (xi − wji ) − yj λwji
(10)
However, the information needed in Eq. 10 is not yet temporally local. xi denotes the rescaled
pixel intensity and does not represent the input spike train. To re-encode a pixel intensity,
xi , to a spike train, Gi , we use uniformly distributed spikes (however, each spike train has a
different random lag) with rate according to the normalized pixel intensity in the range [0, 1].
The maximum number of spikes (for a completely white pixel) over a T = 40 ms interval is
6
40. Additionally, yj is a positive value (approximated by spike count) denoting the neuron’s
activation in response to a stimulus presentation and is not available at synapse, wji . The value
yj can be reexpressed as Hj representing the output spike train of neuron j. Spike trains Gi
and Hj are formulated by the sum of Dirac functions as shown in Eq. 11. Gi (t) and Hj (t) are
either 0 or 1 for a given t.
X
X
δ(t − tf )
(11)
δ(t − tf )
,
Hj (t) =
Gi (t) =
tf ∈Rjf
tf ∈Sif
Sif and Rjf are the sets of presynaptic and postsynaptic spike times. After coding xi and yj by
spike trains Gi and Hj respectively, we propose a local, STDP learning rule following Eq. 10.
When, xi and yj are coded by spike trains over T ms, the synaptic change in continuous time is
given by
Z T
Z T
Z T
1
0
0
0
0
Hj (t0 )dt0
(12)
∆wji ∝
Hj (t )dt
Gi (t )dt − wji − λwji
K
0
0
0
K is a normalizer denoting the maximum number of presynaptic spikes over the T ms interval.
Over a short time period (t ∈ [t0 , t0 + γ), γ < 1 ms, so that K = 1), the weight adjustment at
time t is calculated by
∆wji (t) ∝ rj (t) si (t) − wji (t) − λwji (t)rj (t)
(13)
rj (t) shows the firing status of neuron j at time t (rj (t) ∈ {0, 1}). si (t) specifies the presence of
a presynaptic spike emitted from neuron i at time interval (t − , t]. In our experiments = 1
ms. The synaptic weight is changed only when a postsynaptic spike occurs (rj (t) = 1). Finally,
the learning rule is formulated (upon firing of output neuron j) as follows
∆wji (t) ∝ si (t) − wji (t)(1 + λ) .
(14)
Where, wji ≥ 0. This learning rule is applied to wji when postsynaptic neuron j fires. The
weight change is related to the presynaptic spike times received by the postsynaptic neurons.
This scenario resembles spike-timing-dependent plasticity (STDP). In this STDP rule (Eq. 14),
the current synaptic weight affects the magnitude of the weight change. For instance, if λ = 0
and wji ∈ [0, 1] (it will be proved in Eq. 19), the smaller weights undergo larger LTP and LTD;
and vice versa. It also represents a form of nearest-neighbor spike interaction [58].
The second adaptation rule is the threshold learning rule. Eq. 9 is used to implement
a learning rule for adjusting the threshold, θ. The threshold learning rule shown in Eq. 15
provides an independent and sparse feature representation. The threshold is the same for all D
neurons in the representation layer.
∆θ ∝
D
X
zj − 1
(15)
j=1
In this section, the theory of the proposed spiking representation learning algorithm was
explained. The next section will describe the SNN architecture and the learning rules derived
from Eqs. 14 and 15.
7
4. Network Architecture and Learning
4.1. Neuron Model
The network architecture is shown in Fig. 1 consisting of p2 and D neurons in the input
and representation layers respectively. Stimuli are converted to spike trains over T ms for both
layers. At a given time step, a neuron in the representation layer is allowed to fire only if
its firing criterion is met. The firing criterion records the neuron’s score in a winners-take-all
competition. The WTA score at time step t, given the entire set of incoming weights, W , into
the representation layer, is given by
P
exp( i wji ζi (t))
P
WTAscore j (t; W ) = P
(16)
k exp( i wki ζi (t))
ζi (t) =
X
e−
(t−tf )
τ
(17)
tf
where, ζi (t) is the excitatory postsynaptic potential (EPSP) generated by input neuron i and
the tf s are the recent spike times of unit i during a small interval (t − ν, t], where ν is 4 ms. The
decay time constant, τ , is set to 0.5 ms.
In our network, the softmax value governs the time at which STDP occurs. If WTAscore of
a neuron is greater than the adaptive threshold, θ, STDP is triggered and a spike is emitted.
The softmax phenomologically implements mutual inhibition among the representation neurons
to develop a winners-take-all circuit [59, 44] in the representation layer. The neurons in the
representation layer are purely excitatory and there is no explicit lateral inhibition between
them other than that implicitly implemented by the softmax. When softmax inhibition is
imposed within the representation layer, the network implements a form of competitive learning
by virtue of STDP being triggered by the firing of postsynaptic neurons. Only neurons that
“win the competition” are allowed to learn.
4.2. Learning Rules
The synaptic weight change shown in Eq. 14 defines an STDP rule where the current synaptic
weight influences the magnitude of the change. STDP events are triggered upon postsynaptic
firing yielding two cases corresponding to whether the presynaptic neuron fired within the (t−, t]
time interval. Eq. 18 shows the final STDP rule derived from Eq. 14. The weights fall in the
range [0, 1] and are initialized randomly by sampling from the uniform distribution.
(
a · 1 − wji (1 + λ) , if si = 1
∆wji =
(18)
a · − wji (1 + λ) , if si = 0
a is the learning rate. If λ = 0, the first and second adaptation cases increase and decrease
the synaptic weight respectively (LTP and LTD). If λ → ∞, then both cases are negative and
decrease the weights down to the minimum value (wji = 0). Our experiments study the model’s
8
performance using different λ values over a broad range [0, 10−4 , . . . , 104 ]. Results are shown in
Fig. 2c.
The weight adjustment, at equilibrium, reveals a probabilistic interpretation as follows
E[∆wji ] = 0 ↔
(19)
a · P (si = 1|rj = 1)(1 − wji (1 + λ))−
a · P (si = 0|rj = 1)wji (1 + λ) =
a · P (si = 1|rj = 1)(1 − wji (1 + λ))−
a · (1 − P (si = 1|rj = 1))wji (1 + λ) = 0 ↔
(1 + λ)wji = P (si = 1|rj = 1)
(20)
Therefore, the synaptic weight converges to the (1 + λ) scaled probability of presynaptic spike
occurrence given postsynaptic spike (LTP probability). From Eq. 20, the weights fall in the
1
range (0, 1+λ
) so that the first case refers to LTP (∆wji ≥ 0) and the second one refers to LTD
(∆wji ≤ 0), at the equilibrium point.
To show that the STDP rule (Eq. 18) is consistent with the learning rule in Eq. 10, we
rewrite the non-local rule with learning rate, a, as follows
non−local
∆wji
= a · yj xi − wji − λwji
(21)
As stated earlier, this rule is temporally non-local and shows the weight change over a T ms
interval. In contrast, the STDP rule is temporally local, applying the weight change at one time
step when the postsynaptic neuron fires. To make Eq. 21 and Eq. 18 (which is derived from
Eq. 14) comparable with each other, we consider a time interval with only one postsynaptic
spike where rj = 1. Specifically, we break the T ms interval into subintervals whose boundaries
are determined by the event of a postsynaptic spike. It is sufficient to analyze an arbitrary
subinterval. Therefore, Eq. 21 at time t simplifies to
non−local
∆wji
= a xi − wji − λwji
(22)
Following Eq. 19 for calculating the expected weight change using the proposed STDP rule,
where rj = 1, we find that
E[∆wji ] = a P (si = 1) − wji − λwji
(23)
Where, P (si = 1) is the firing probability of presynaptic neuron i. Also, we generated the
presynaptic spike trains using the normalized pixel intensities in the range [0, 1] with different
random lags. Thus, this probability value is the same as the normalized pixel intensity, xi , as
firing rate. Therefore,
non−local
(24)
E[∆wji ] = a xi − wji − λwji = ∆wji
9
which matches the weight change shown in Eq. 22. This shows that the proposed STDP rule
is consistent with the non-local rule. Additionally, the STDP weight change is an unbiased
estimation for the non-local (non-spike based) learning rule. Over a short time period, the
proposed learning rule is also an unbiased estimation for the Hebbian rule of Földiák [20] (Eq. 3).
For the threshold adaptation, following Eq. 15, the threshold learning rule can be written as
∆θ = b mz − 1
(25)
where, b is the learning rate. mz is the number of neurons in the representation layer firing in
T ms. This rule adjusts the threshold such that only one neuron fires in response to a stimulus.
This criterion provides a framework to extract independent features in a sparse representation.
As we used softmax-based neurons in the representation layer, the initial threshold value, θinit ,
should be in the following range:
1
< θinit << 0.5
(26)
D
Where, D is the number of neurons in the representation layer. The upper-bound of 0.5 allows
more than one neuron to be active at the initial training steps to capture visual features (θinit <<
0.5). On the other hand, the initial threshold should be big enough to stop high synchronization
1
at the beginning (θinit > D
). According to the minimum number of neurons we used in the
experiments (D = 8, 1/D = 0.125), the initial threshold was set to 0.15.
5. Evaluation Metrics
We use the following metrics to judge the quality of the representation acquired in Fig. 1.
5.1. Reconstructed image
We use reconstructed image to qualitatively assess the extent that the representation layer
captures the information contained in the image patches. The representation filter set, W =
{w1 , w2 , ..., wD }, is a p2 × D weight matrix coding an image patch (p2 input spike trains) to
a vector of D postsynaptic spike trains. To reconstruct the image patch from the coded spike
trains, the reconstruction filter set, W rec ≡ W T , is used to build p2 spike trains. For this
purpose, neurons in the input layer receive spike trains from the neurons in the representation
layer via the transposed synaptic weight matrix (like an autoencoder).
5.2. Reconstruction loss
To report the reconstruction loss, we use the correlation measure (Pearson correlation) and
the root mean square (RMS) error between the normalized original, ym , and reconstructed, ŷm ,
patches as shown in Eqs. 27 and 28 respectively. A patch stands for p2 spike rates, y.
M
1 X
Corr Recon Loss =
1 − Cor(ym , ŷm )
M
m=1
10
(27)
RMS =
1
M
M
X
m=1
v
u
p2
u1 X
t
(yi,m − ŷi,m )2 .
p2
(28)
i=1
M is the number of patches extracted from the image.
5.3. Sparsity
To calculate the sparsity, we use average activity and breadth tuning measures. The average
activity specifies the density of spikes released from neurons in the representation layer over T
time steps given in Eq. 29.
1 XX
Sparsity =
rj (t) .
(29)
D·T
t
j
The breadth tuning measure introduced by Rolls and Tovee [60] specifies the density of neural
layer activity (Eq. 30) calculated by the ratio of mean, µ, and standard deviation, σ, of spike
frequencies in the representation layer upon presenting a stimulus. The breadth tuning measures
the neural selectivity such that the sparse code distribution concentrates near zero with a heavy
tail [61]. For a neural layer where most of the neurons fire, the activity distribution is more
uniformly spread and Breadth Tuning is greater than 0.5. In contrast, in a sparse code where
most of the neurons do not fire, the distribution is peaked at zero and Breadth Tuning is less
than 0.5.
1
σ
Breadth Tuning = 2
, C=
(30)
C +1
µ
6. Experiments and Results
We ran two experiments using the MNIST [62] and the natural image [26] datasets to evaluate
the proposed local representation learning rules embedded in the single-layer SNN. For both
datasets, the intensities of the gray-scale images were normalized to fall in the range of [0, 1]
yielding possible spike rates to generate uniformly distributed spike trains for the input layer
over T = 40 ms. The learning rates for STDP learning, a, (Eq. 18) and for threshold adjustment,
b, (Eq. 25) were set to 0.0005 and 0.0001, respectively1 . We ran a number of experiments with
different learning rates and found that changing a and b in the range [0.00005, 0.001] did not
change the model’s performance significantly. Additionally, as the threshold adjustment rule is
not modulated by the current threshold value, we chose a smaller learning rate (0.0001) for it
to avoid possible threshold instability.
1
The maximum number of postsynaptic spikes is 40 and the maximum number of patches sampled from an
1
MNIST digit is 25. Our simple strategy for setting the learning rates is: a, b < 25×40
= 0.001.
11
6.1. Experiment 1: MNIST dataset
Experiments were run using 5 × 5 patches sampled from 28 × 28 MNIST digits. We used
a random subset of the MNIST digits divided into 15,000 training and 1000 testing images for
learning and evaluating the model. The SNN architecture consists of 25 (5 × 5 image patch)
neurons in the input layer and D = {2i , i = 3 . . . 7} neurons in the representation layer. These
variations of the network architecture (different D values) determine under-complete to overcomplete representations. Trained filters, after 1 through 15,000 iterations, for the network with
32 neurons in the representation layer are shown in Fig. 2a. After 1000 training iterations, the
kernels start becoming selective to specific visual patterns (orientations). The filters shown in
this image tend to be orientation selective and extract different visual features.
Fig. 2c shows the RMS reconstruction loss and other statistical characteristics (max, min,
mean) of the trained weights versus the log regularizer hyperparameter (log10 λ). For λ ≤ 0.1,
the RMS loss values reach a near optimal uniformly low plateau2 . For this reason, λ was set
to zero for further experiments. Additionally, Fig. 2c shows that the maximum and minimum
synaptic weights after training are 1/(1 + λ) and 0 respectively as predicted by Eq. 20.
The three performance measures from the Section 5 were used to assess the model. These
were the reconstructed images, the reconstruction loss, and the sparsity. The reconstructed
images of randomly selected digits 0 through 9, acquired by the SNN with D = 32 neurons
in the representation layer, are shown in Fig. 2b. The reconstructed maps show high quality
images comparable with the original images. The reconstruction loss measures for the SNNs
with D = 8 through 128 filters appear in Figs. 3a and 3b. The SNNs with D = 16 and 32 show
the lowest reconstruction loss after training. The sparsity measures reported by the average
sparsity and the breadth tuning are shown in Fig. 3c and 3d. The sparsity measures also show
better performance for the networks with D = {16, 32, 64} filters. The average sparsity value
of 0.09 shows that only 9% of the neurons were active on each trial. The breadth tuning value
of 0.23 indicates the sparse stimulus representation. Figs. 3e and 3f depict the summary of the
model’s performance after training for D = {8, . . . , 128} and D = 32 kernels, respectively.
6.2. Experiment 2: Natural images
This experiment evaluates representations acquired from 16 × 16 natural image patches [26].
Fig. 4a shows the trained representation filters for the SNNs with 8, 16, 32, 64, and 128 neurons
in the representation layer. For instance, where D = 32, except for the filters marked with
dotted circles, the other filters have low correlation with each other. For visual assessments,
Fig. 4b shows four natural images and their reconstructed maps. Performance of the proposed
model in terms of the reconstruction loss and sparsity measures on natural images is shown
in Fig. 4c. The lowest reconstruction loss belongs to the networks with {16, 32, 64} neurons
in the representation layer. The small number of neurons (D = 8) is not able to capture the
2
The average RMS reconstruction loss values for the SNNs with λ ≤ 0.1 was reported 0.167±0.001 (Ninety-five
percent confidence intervals of the RMS loss values (standard error of the mean; n = 5) were calculated).
12
(a)
(b)
(c)
Figure 2: (a): D = 32 trained filters after 1, 10, ..., 15000 iterations. The red-blue spectrum denotes the
maximum-minimum synaptic weights. (b): Reconstructed images based on overlapped and non-overlapped 5 × 5
patches. The overlapped patches are selected by 5 × 5 windows sliding over the image with a stride of 1. The
non-overlapped patches slide over the image with a stride of 5. (c): RMS reconstruction loss and synaptic weight
ranges for the SNN with D = 32 filters versus log10 regularizer hyperparameter, λ. λ = 0 is approximated by
10−31 .
13
Figure 3: (a)-(d): Model performance trends on MNIST after 1 through 15000 training iterations in terms of
(a): correlation-based reconstruction loss, (b): RMS reconstruction loss, (c): average sparsity, and (d): breadth
tuning.(e): The model’s performance after training. (f ): The evaluation measures for the trained visual representation model with D = 32 kernels. Error bars show standard error of the mean.
14
(a)
(b)
(c)
Figure 4: Model’s performance on the natural image patches. (a): Representation filters after training the
SNNs with D = {8, 16, 32, 64, 128} spiking neurons in the representation layer. (b): Original (first row) and
reconstructed (second row) image sections (D = 32). (c): Reconstruction loss and sparsity measures of the
models with 8 through 512 filters.
visual codes. On the other hand, using too many neurons (D > 128) increases reconstruction
loss because a number of neurons cannot be involved in the learning process due to the WTA
constraint.
6.3. Comparisons
The proposed spiking representation learning method shows better performance than the
traditional K-means clustering [63] and the restricted Boltzman machine (RBM) [64, 65] while
introducing local learning in time and space. We implemented these two methods, as traditional
quantization-like representation learning examples, using the same training/testing images. The
K-means and RBM approaches were applied to the normalized pixel intensities of image patches
(not spike trains). Thus, these methods are not temporally local. Table 1 shows this comparison
in terms of reconstruction loss (correlation and RMS). Our model outperforms the RBM and
K-means methods except for the two cases (natural images) in which the RBM shows slightly
better performance. Fig. 5 shows the trained filters obtained by K-means, RBMs, and our
model based on the MNIST and natural image patches. K-means, similar to our model, detects
different visual orientations for the MNIST and natural image patches, but the filters are highly
correlated. The RBM did not perform well for the MNIST dataset but it successfully learned
15
Figure 5: D = 16, 32, and 64 representation filters trained on the MNIST and natural images datasets using
K-means, RBM, and our STDP.
representative visual filters for the natural image patches where D = 64. These trained filters
(Fig. 5) confirm the reconstruction loss variations reported in Table 1.
Table 2 compares our results with the only (to the best of our knowledge) spike-based representation learning models. The correlation-based reconstruction loss on MNIST and natural
images (0.2 and 0.4) shows improvement over the existing spiking autoencoder using mirrored
STDP (0.2 and 0.65) proposed by Burbank [39]. The sparse representation introduced by King
et al [25], which is a modified version of the SAILnet algorithm [22], reported an RMS reconstruction loss around 0.74 that is calculated based on the spike rates normalized to unit standard
deviation (let’s say zRMS). Our model compared favorably with their model with zRMS=0.67.
However, our model did not scale well to a larger number of neurons when D ≥ 128 in the
representation layer. The problem appears
to stem from the threshold adjustment rule (Eq. 25).
If we change the rule to ∆θ = b mz − q , where q is a proportion of D, the representation layer
would be more active and a large number of filters can be trained to reduce the reconstruction
16
Table 1: Reconstruction loss (correlation and RMS) obtained by K-means and RBM in comparison with our
method.
Rec. Loss
D
K-means
RBM
Our STDP
MNIST Corr.
16
32
64
0.22
0.23
0.26
0.49
0.49
0.40
0.20 0.20 0.24
MNIST RMS
16
32
64
0.18
0.21
0.26
0.27
0.27
0.26
0.17 0.17 0.21
Natural Corr.
16
32
64
0.45 0.52
0.57
0.92
0.41
0.44
0.49
0.40 0.47
Natural RMS
16
32
64
0.31
0.36
0.40
0.47
0.27 0.26
0.24 0.27 0.40
Table 2: The reconstruction loss values reported by Burbank [39] in terms of correlation loss and King et al. [25]
in terms of zRMS in comparison with our results.
Dataset
Natural images
MNIST
Burbank [39]
Corr: 0.65
Corr: 0.2
King et al. [25]
zRMS: 0.74
-
Our Model
Corr: 0.4, zRMS: 0.67
Corr: 0.2
loss.
7. Discussion and Conclusion
This paper derived a novel STDP-based representation learning method to be embedded in
an SNN and evaluated the acquired representations in two experiments to establish the method’s
initial viability. The derived rules were extremely simple, yet the evaluated reconstruction loss
was extremely low. The simplicity of the rules (resulting from the constraint of temporal locality)
makes them attractive for hardware implementation.
The learning rules were derived by constrained optimization incorporating a vector quantizationlike objective function with regularization and a sparsity constraint. The learning rules included
spatio-temporally local STDP-type weight adaptation and a threshold adjustment rule. The
STDP rule at equilibrium showed a probabilistic interpretation of the synaptic weights scaled
by the regularizer hyperparameter. In addition to the threshold adaptation rule, the WTAthresholded neurons in the representation layer implemented inhibition (by a novel temporal,
spiking softmax function) to represent sparse and independent visual features. Softmax is a
standard way to implement a winners-take-all (WTA) circuit and to implement mutual inhibition without using explicit inhibitory neurons in the representation layer [44, p. 181], [43, p.
238].
The experimental results showed high performance of the proposed model in comparison
with spiking and non-spiking approaches. Our model almost outperformed the traditional KMeans and RBM models in representation learning and training of the orientation selective
kernels. Also, our method showed better performance (in terms of reconstruction loss) than the
state-of-the-art spiking representation learning approaches used by [39] (spiking autoencoder)
and [22, 25] (sparse representation).
To obtain the spatio-temporally local learning rules embedded in the SNN, we started from
a non-spiking quantization criterion inspired from [20]. Then, we developed novel rules to im17
plement an STDP based representation learning and a threshold adjustment rule for spiking
platforms. The spike-based platform and spatio-temporally local learning rules lead the main
difference between our study and well-known, traditional representation learning methods introduced in the literature. Existing spiking representation learning methods in the literature suffer
from limitations such as violating Dale’s law [22], synapses that can change sign [22, 25], low
performance in terms of reconstruction loss [39], and non-spiking input signals [22, 25]. In this
study we proposed an STDP learning rule which updates the synaptic weights falling within
the range [0, 1]. The SNN architecture consists of excitatory neurons and an implicit inhibition occurring in the representation layer. The implicit inhibition is analogous to a separate
inhibitory layer balancing neural activities in the representation neural layer where Dale’s law
is maintained. Furthermore, the proposed SNN implements spiking neurons in both the input
and representation layers and the neurons only communicate through temporal spike trains.
To the best of our knowledge, our approach is the only high performance (in terms of
reconstruction loss) representation learning model implemented on SNNs. There are several
studies in the literature developing SNNs equipped with bio-inspired STDP for unsupervised
feature extraction through single or multi-layer spike-based architectures. The most recent
works of [41, 42, 38], and [23] have utilized these features to classify MNIST digits. Although
these networks introduce novel spiking network architectures for feature representation, they do
not offer a pure representation learning approach with low reconstruction loss.
Although our proposed spiking representation learning was successful for reconstruction,
there is a limitation that the spike rate of the presynaptic neurons is higher than biological spiking
neurons. Our future work seeks to reduce this spike rate to be more biologically plausible. Using
more presynaptic neurons presenting mutual exclusive intensity bands would be a starting point.
Additionally, it is a matter of future work to determine how well the acquired representations
from our STDP algorithm perform in a pattern recognition context. It can also be tested in
future work whether our acquired representations are stackable to afford the ability for multilayer, STDP-based learning.
References
[1] M. Bhand, R. Mudur, B. Suresh, A. Saxe, A. Y. Ng, Unsupervised learning models of
primary cortical receptive fields and receptive field plasticity, in: Advances in neural information processing systems, 2011, pp. 1971–1979.
[2] H. Lee, C. Ekanadham, A. Y. Ng, Sparse deep belief net model for visual area V2, in:
Advances in neural information processing systems, 2008, pp. 873–880.
[3] A. Coates, A. Ng, H. Lee, An analysis of single-layer networks in unsupervised feature
learning, in: Proceedings of the fourteenth international conference on artificial intelligence
and statistics, 2011, pp. 215–223.
18
[4] Y. Bengio, A. Courville, P. Vincent, Representation learning: A review and new perspectives, IEEE transactions on pattern analysis and machine intelligence 35 (8) (2013) 1798–
1828.
[5] R. Q. Quiroga, L. Reddy, G. Kreiman, C. Koch, I. Fried, Invariant visual representation by
single neurons in the human brain, Nature 435 (7045) (2005) 1102–1107.
[6] N. K. Logothetis, D. L. Sheinberg, Visual object recognition, Annual review of neuroscience
19 (1) (1996) 577–621.
[7] M. Riesenhuber, T. Poggio, Neural mechanisms of object recognition, Current opinion in
neurobiology 12 (2) (2002) 162–168.
[8] M. P. Young, S. Yamane, Sparse population coding of faces in the inferotemporal cortex,
Science 256 (5061) (1992) 1327–1331.
[9] S. M. Landi, W. A. Freiwald, Two areas for familiar face recognition in the primate brain,
Science 357 (6351) (2017) 591–595.
[10] B. A. Wandell, Foundations of vision., Sinauer Associates, 1995.
[11] Y. Frégnac, J. Fournier, F. Gérard-Mercier, C. Monier, M. Pananceau, P. Carelli, X. Troncoso, The visual brain: Computing through multiscale complexity, in: Micro-, Meso-and
Macro-Dynamics of the Brain, Springer, 2016, pp. 43–57.
[12] M. W. Self, J. C. Peters, J. K. Possel, J. Reithler, R. Goebel, P. Ris, D. Jeurissen, L. Reddy,
S. Claus, J. C. Baayen, et al., The effects of context and attention on spiking activity in
human early visual cortex, PLoS biology 14 (3) (2016) e1002420.
[13] W. Maass, Networks of spiking neurons: the third generation of neural network models,
Neural Networks 10 (9) (1997) 1659–1671.
[14] E. M. Izhikevich, Which model to use for cortical spiking neurons?, IEEE transactions on
neural networks 15 (5) (2004) 1063–1070.
[15] S. Ghosh-Dastidar, H. Adeli, Spiking neural networks, International Journal of Neural
Systems 19 (04) (2009) 295–308.
[16] W. Maass, To spike or not to spike: that is the question, Proceedings of the IEEE 103 (12)
(2015) 2219–2224.
[17] W. Maass, On the computational power of noisy spiking neurons, in: Advances in neural
information processing systems, 1996, pp. 211–217.
[18] D. Neil, M. Pfeiffer, S.-C. Liu, Learning to be efficient: Algorithms for training low-latency,
low-compute deep spiking neural networks, in: Proceedings of the 31st Annual ACM Symposium on Applied Computing, ACM, 2016, pp. 293–298.
19
[19] C. S. Brito, W. Gerstner, Nonlinear Hebbian learning as a unifying principle in receptive
field formation, PLoS computational biology 12 (9) (2016) e1005070.
[20] P. Földiák, Forming sparse representations by local anti-Hebbian learning, Biological Cybernetics 64 (2) (1990) 165–170.
[21] H. B. Barlow, Unsupervised learning, Neural Computation 1 (3) (1989) 295–311.
[22] J. Zylberberg, J. T. Murphy, M. R. DeWeese, A sparse coding model with synaptically local
plasticity and spiking neurons can account for the diverse shapes of V1 simple cell receptive
fields, PLoS Comput Biol 7 (10) (2011) e1002250.
[23] A. Tavanaei, A. S. Maida, Multi-layer unsupervised learning in a spiking convolutional
neural network, in: Neural Networks (IJCNN), 2017 International Joint Conference on,
IEEE, 2017, pp. 2023–2030.
[24] H. Markram, W. Gerstner, P. J. Sjöström, Spike-timing-dependent plasticity: a comprehensive overview, Frontiers in synaptic neuroscience 4 (2012) 2.
[25] P. D. King, J. Zylberberg, M. R. DeWeese, Inhibitory interneurons decorrelate excitatory
cells to drive sparse code formation in a spiking model of V1, The Journal of Neuroscience
33 (13) (2013) 5475–5485.
[26] B. A. Olshausen, D. J. Field, Emergence of simple-cell receptive field properties by learning
a sparse code for natural images, Nature 381 (6583) (1996) 607–609.
[27] A. J. Bell, T. J. Sejnowski, The “independent components” of natural scenes are edge filters,
Vision Research 37 (23) (1997) 3327–3338.
[28] M. Rehn, F. T. Sommer, A network that uses few active neurones to code visual input
predicts the diverse shapes of cortical receptive fields, Journal of computational neuroscience
22 (2) (2007) 135–146.
[29] B. A. Olshausen, C. F. Cadieu, D. K. Warland, Learning real and complex overcomplete
representations from the statistics of natural images, in: Proc SPIE, Vol. 7446, 2009, pp.
74460S–1.
[30] C. von der Malsburg, Self-organization of orientation sensitive cells in the striate cortex,
Kybernetik 14 (1973) 85–100.
[31] E. L. Bienenstock, L. N. Cooper, P. W. Munro, Theory for the development of neuron
selectivity: Orientation specifity and binocular interaction in visual cortex, Journal of Neuroscience 2 (1) (1892) 32–48.
20
[32] T. Masquelier, Relative spike time coding and STDP-based orientation selectivity in the
early visual system in natural continuous and saccadic vision: a computational model,
Journal of Computational Neuroscience 32 (3) (2012) 425–441.
[33] D. H. Hubel, T. N. Wiesel, Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex, The Journal of Physiology 160 (1) (1962) 106–154.
[34] N. Caporale, Y. Dan, Spike timing-dependent plasticity: a Hebbian learning rule, Annu.
Rev. Neurosci. 31 (2008) 25–46.
[35] T. Masquelier, S. J. Thorpe, Unsupervised learning of visual features through spike timing
dependent plasticity, PLoS computational biology 3 (2) (2007) e31.
[36] S. R. Kheradpisheh, M. Ganjtabesh, T. Masquelier, Bio-inspired unsupervised learning of
visual features leads to robust invariant object recognition, Neurocomputing 205 (2016)
382–392.
[37] A. Tavanaei, T. Masquelier, A. S. Maida, Acquisition of visual features through probabilistic
spike-timing-dependent plasticity, in: Neural Networks (IJCNN), 2016 International Joint
Conference on, IEEE, 2016, pp. 307–314.
[38] S. R. Kheradpisheh, M. Ganjtabesh, S. J. Thorpe, T. Masquelier, STDP-based spiking deep
convolutional neural networks for object recognition, Neural Networks 99 (2017) 56–67.
[39] K. S. Burbank, Mirrored STDP implements autoencoder learning in a network of spiking
neurons, PLoS Comput Biol 11 (12) (2015) e1004566.
[40] C. Savin, P. Joshi, J. Triesch, Independent component analysis in spiking neurons, PLoS
Comput Biol 6 (4) (2010) e1000757.
[41] P. U. Diehl, M. Cook, Unsupervised learning of digit recognition using spike-timingdependent plasticity, Frontiers in computational neuroscience 9 (2015) 99.
[42] A. Shrestha, K. Ahmed, Y. Wang, Q. Qiu, Stable spike-timing dependent plasticity rule
for multilayer unsupervised and supervised learning, in: Neural Networks (IJCNN), 2017
International Joint Conference on, IEEE, 2017, pp. 1999–2006.
[43] C. M. Bishop, Neural networks for pattern recognition, Oxford university press, 1995.
[44] I. Goodfellow, Y. Bengio, A. Courville, Deep learning, MIT press, 2016.
[45] S. G. Wysoski, L. Benuskova, N. Kasabov, Evolving spiking neural networks for audiovisual
information processing, Neural Networks 23 (7) (2010) 819–835.
[46] S. G. Wysoski, L. Benuskova, N. Kasabov, Fast and adaptive network of spiking neurons
for multi-view visual pattern recognition, Neurocomputing 71 (13-15) (2008) 2563–2575.
21
[47] S. Schliebs, N. Kasabov, Evolving spiking neural network a survey, Evolving Systems 4 (2)
(2013) 87–98.
[48] N. Kasabov, K. Dhoble, N. Nuntalid, G. Indiveri, Dynamic evolving spiking neural networks
for on-line spatio-and spectro-temporal pattern recognition, Neural Networks 41 (2013)
188–201.
[49] S. Soltic, N. Kasabov, Knowledge extraction from evolving spiking neural networks with
rank order population coding, International Journal of Neural Systems 20 (06) (2010) 437–
445.
[50] S. Grossberg, Adaptive resonance theory: How a brain learns to consciously attend, learn,
and recognize a changing world, Neural Networks 37 (2012) 1–47.
[51] S. Thorpe, J. Gautrais, Rank order coding, Computational Neuroscience 113 (1998) 113–
119.
[52] W. Maass, H. Markram, Synapses as dynamic memory buffers, Neural Networks 15 (2)
(2002) 155–161.
[53] A. Coates, A. Y. Ng, Learning feature representations with k-means, in: Neural networks:
Tricks of the trade, Springer, 2012, pp. 561–580.
[54] P. Földiák, Adaptive network for optimal linear feature extraction, in: Neural Networks
(IJCNN), 1989 International Joint Conference on, Vol. 1, IEEE, 1989, pp. 401–405.
[55] E. Oja, Simplified neuron model as a principal component analyzer, Journal of mathematical
biology 15 (3) (1982) 267–273.
[56] B. Hammer, T. Villmann, Generalized relevance learning vector quantization, Neural Networks 15 (8) (2002) 1059–1068.
[57] P. Schneider, M. Biehl, B. Hammer, Distance learning in discriminative vector quantization,
Neural computation 21 (10) (2009) 2942–2969.
[58] J. Sjöström, W. Gerstner, Spike-timing dependent plasticity, Spike-timing dependent plasticity 35 (0) (2010) 0–0.
[59] A. Tavanaei, A. S. Maida, Bio-inspired spiking convolutional neural network using layerwise sparse coding and STDP learning, arXiv preprint arXiv:1611.03000 (2016) 1–16.
[60] E. T. Rolls, M. J. Tovee, Sparseness of the neuronal representation of stimuli in the primate
temporal visual cortex, Journal of Neurophysiology 73 (2) (1995) 713–726.
[61] P. Foldiak, D. Endres, Sparse coding, Scholarpedia 3 (1) (2008) 2984.
22
[62] Y. LeCun, C. Cortes, C. J. Burges, The MNIST database, URL http://yann. lecun.
com/exdb/mnist.
[63] C. M. Bishop, Pattern recognition and machine learning, Springer, 2006.
[64] G. Hinton, A practical guide to training restricted Boltzmann machines, Momentum 9 (1)
(2010) 926.
[65] N. Le Roux, Y. Bengio, Representational power of restricted Boltzmann machines and deep
belief networks, Neural computation 20 (6) (2008) 1631–1649.
23
| 9cs.NE
|
Adversarial Examples Detection in Deep Networks with Convolutional Filter
Statistics
arXiv:1612.07767v2 [cs.CV] 26 Oct 2017
Xin Li, Fuxin Li
School of Electrical Engineering and Computer Science
Oregon State University
[email protected], [email protected]
Abstract
Δ𝐼
Giant Panda (99.32% confidence)
=
+0.03
Deep learning has greatly improved visual recognition
in recent years. However, recent research has shown that
there exist many adversarial examples that can negatively
impact the performance of such an architecture. This paper
focuses on detecting those adversarial examples by analyzing whether they come from the same distribution as the
normal examples. Instead of directly training a deep neural
network to detect adversarials, a much simpler approach
was proposed based on statistics on outputs from convolutional layers. A cascade classifier was designed to efficiently detect adversarials. Furthermore, trained from one
particular adversarial generating mechanism, the resulting
classifier can successfully detect adversarials from a completely different mechanism as well. The resulting classifier
is non-subdifferentiable, hence creates a difficulty for adversaries to attack by using the gradient of the classifier.
After detecting adversarial examples, we show that many of
them can be recovered by simply performing a small average filter on the image. Those findings should lead to more
insights about the classification mechanisms in deep convolutional neural networks.
𝐼
Δ𝐼
+0.03
Shark (93.89% confidence)
Goldfish (95.15% confidence)
=
Figure 1. An optimization algorithm finds adversarial examples
which, with almost negligible perturbations to human eyes, completely distort the prediction result of a deep neural network [28].
Such algorithms have been found to be universal to different deep
networks. This paper studies their properties and seeks a defense.
tonomous driving, robotics, or any automatic identification
(face, iris, speech, etc.). If the result of the network can
be hacked at the will of a hacker, wrong authentications
and other devastating effects would be unavoidable. Therefore, there are ample reasons to believe that it is important
to identify whether an example comes from a normal or an
adversarial distribution. A reliable procedure can prevent
robots from behaving in undesirable manners because of the
false perceptions it made about the environment.
The understanding of whether an example belongs to the
training distribution has deep roots in statistical machine
learning. The i.i.d. assumption was commonly used in
learning theory, so that the testing examples were assumed
to be drawn independently from the same distribution of the
training examples. This is because machine learning is only
good at performing interpolation, where some training examples surround a testing example. Extrapolation is known
to be difficult, since it is extremely difficult to estimate data
labels or statistics if the data is extremely different from any
known or learned observations. Many current approaches
deal with adversarial examples by adding them back to the
training set and re-train. However in their experiments, new
adversarials can almost always be found from the re-trained
classifier. This is because that the space of extrapolation is
significantly larger than the area a machine learning algo-
1. Introduction
Recent advances in deep learning have greatly improved
the capability to recognize visual objects [13, 26, 7]. Stateof-the-art neural networks perform better than human on
difficult, large-scale image classification tasks. However, an
interesting discovery has been that those networks, albeit resistant to overfitting, would have completely failed if some
of the pixels in the image were perturbed via an adversarial
optimization algorithm [28, 4] . An image indistinguishable from the original for a human observer could lead to
significantly different results from a deep network(Fig. 1).
Those adversarial examples are dangerous if a deep network is utilized in any crucial real application, be it au1
rithm can interpolate, and the ways to find vulnerabilities of
a deep learning system are almost endless.
A more conservative approach is to refrain from making
a prediction if the system does not feel comfortable about
it. Such an approach seeks to build a wall to fence all testing examples in the extrapolation area out of the predictor,
and only predict in the small interpolation area. Work such
as [16] provides basic theoretical frameworks of classification with an abstain option.
Although these concepts are well-known, the difficulties
lie in the high-dimensional spaces that are routinely used
in machine learning and especially deep learning. Is it even
possible to define interpolation vs. extrapolation in a 4, 000dimensional or 40, 000-dimensional space? It looks like
almost everything is extrapolation since the data is inherently sparse in such a high-dimensional space [9, 6], a phenomenon well-known as the curse of dimensionality. The
enforcement of the i.i.d. assumption seems impossible in
such a high-dimensional space, because the inverse problem
of estimating the joint distribution requires an exponential
number of examples to be solved efficiently. Some recent
work on generative adversarial networks proposes using a
deep network to train this discriminative classifier [3, 22],
where a generative approach is required to generate those
samples, but it is largely confined to unsupervised settings
and may not be applicable for every domain convolutional
networks (CNNs) have been applied to.
In this work we propose a discriminative approach to
identify adversarial examples, which trains on simple features and can approach good accuracy with limited training
examples. The main difference between our approach and
previous outlier detection/adversarial detection algorithms
(e.g. [2]) is that their approaches usually treat deep learning
as a black box and only works at the final output layer, while
we believe that the learned filters in the intermediate layers
efficiently reduce the dimensionality and are useful for detecting adversarial examples. We make a number of empirical visualizations that show how the adversarial examples
change the prediction of a deep network. From those intuitions, we extract simple statistics from convolutional filter
outputs of various layers in the CNN. A cascade classifier is
proposed that utilizes features from many layers to discriminate between normal and adversarial examples.
Experiments show that our features from convolutional
filter output statistics can separate between normal and adversarial examples very well. Trained with one particular adversarial generation method, it is robust enough to
generalize to adversarials produced from another generation approach [20] without any special adaptation or additional training. Those confidence estimates may improve
the safety of applying these deep networks, and hopefully
provide insights for further research on self-aware learning.
As a simple extension, the results from visualizations of the
features prompted us to perform an average filter on corrupted images, and found out that many correct predictions
can be recovered from this simple filtering.
2. Deep Convolutional Neural Networks
A deep convolutional neural network consists of
many convolutional layers which are connected to spatially/temporally adjacent nodes in the next layer:
Zm+1 = [T (W1 ∗ Zm ), T (W2 ∗ Zm ), . . . , T (Wk ∗ Zm )]
(1)
where Zm is the input features at layer m, W1 , . . . WK
are filters that could be much smaller than the size of Zm
(e.g. 3 × 3, 5 × 5, 7 × 7), ∗ is the convolution operator, and
T is a nonlinear transformation function such as the rectified linear unit (ReLU) T (x) = max(0, x). Other commonly used layers in a CNN include max-pooling layers,
or other normalization layers [13] such as batch normalization layers [10]. Most deep networks adopt similar principles while adding more structural complexity in the system
such as more layers and smaller filters in each layer [26],
multi-layered network within each layer [27], residual network [7], etc. A convolutional neural network makes sense
in structured data because it naturally exploits the locality
structure in data. In an image, pixels that are located close
to each other are naturally more correlated than pixels that
are far away [17]. The same holds for temporal data (video,
speech) where objects (frames, utterances) that are temporally close can be assumed to be more correlated.
3. Understanding the Trained Deep Classifier
Under Adversarial Optimization
3.1. Adversarial Optimization
The famous result that deep networks can be broken easily [28] is an important motivation of this work. The idea
is to start from an existing example (image) and optimize
to obtain an example that will be classified to another category while being close to the original example. Namely, the
following optimization problem is solved:
min ckrk1 + L(fθ (x0 + r, y))
r
s.t.
x0 + r ∈ [0, 1]d
(2)
where x0 is a known example and y is an arbitrary category label, d is the input dimensionality. c is a parameter that can be tuned for trading off between proximity to
the original example x0 and the classification loss on the
other category y. It has been shown, to the astound of many,
that one can choose an r with very small norm while completely change the output of the algorithm (e.g. Fig. 1),
this can even be done universally for almost all networks,
datasets and categories [28, 4]. Besides, adversarials trained
from one network may even fool a related one trained from
the same dataset [18]. This has led many people to question whether deep networks are really learning the “proper”
rules for classifying those images.
3.2. Adversarial Behavior
In order to gain a deeper understanding of the behavior of
a deep network and illustrate the difference between adversarial and normal example distributions, we utilize spectral
analysis. As a starting point, we perform principal component analysis (PCA) [11] at the 14-th layer of a VGG
network trained on the ImageNet dataset (the first fullyconnected layer). The rationale behind using PCA is that
each deep learning layer is a nonlinear activation function
on a linear transformation, hence a large part of the learning process lies within the linear transformation, for which
PCA is a standard tool to analyze.
A linear PCA is performed on the entire collection of
50, 000 images from the ImageNet validation set, as well as
4, 000 adversarials collected using the approach in (2), starting from random images in the collection. The result shows
very interesting findings (Fig. 2) and sheds more light on
the internal mechanics of those adversarial examples. In
Fig. 2(a), we show the PCA projection onto the first two
eigenvectors. This cannot separate normal and adversarial
examples, as one could possibly imagine. The adversarial
examples seem to exactly belong to the same distribution
as normal ones. However, it does seem that the adversarial examples reside mostly in the center while the normal
examples occupy a bigger chunk of space.
Interestingly, as we move to the tail of the PCA projection space, the picture starts to change significantly. In
Fig. 2(b), we can see that there are a significant amount of
adversarial examples that has extremely large values w.r.t.
to the normal examples in the tail of the distribution. We
chose to print the projection on the 3, 547-th and 3, 844-th
eigenvector, but similar distributions can be found all over
the tail. As one can see, at such a far end on the tail, the
projections of normal examples are very similar to random
samples under a Gaussian distribution. An explanation for
that could be that under these “uninformative” directions,
most of the weighted features are nearly independent w.r.t.
each other, hence the distribution of their sum is similar to
Gaussian, according to the central limit theorem1 . However, although normal examples behave similarly to a Gaussian, some adversarial examples are having projections with
a deviation as large as 5 or 10 times the standard deviation,
which are extremely unlikely to occur under a Gaussian distribution.
1 Note
this is without a ReLU transformation. ReLU would destroy the
negative part of the data distribution so that it no longer looks like a Gaussian. However, some tail effects can be observed even in the distribution
after ReLU.
Fig. 2(c) and Fig. 2(d) show that there are two distinct
phenomena:
• The extremal values and standard deviations on the
projections onto the first 500 − 700 eigenvectors are
decidedly lower in adversarial examples than in normal ones.
• The extremal values and standard deviations on the
projections onto the last 1, 000 − 1, 500 eigenvectors
are decidedly higher in the adversarial examples than
the normal ones.
It is interesting to reflect about the causes and consequences of those properties. One deciding property is that
there is a strong regularization effect in adversarial examples on almost all the informative directions. Hence, the
predictions in adversarial examples are lower than those
in normal examples, rather than the confidence values may
have indicated (Fig. 1). In Fig. 3, we show the number of
categories with a prediction higher than a threshold, before
the final softmax transformation
exp(fi (x))
pi (x) = P
i exp(fi (x))
(3)
that converts raw predictions fi (x) into probabilities. The
result shows that normal examples have on average one category with a raw prediction value more than 20, however
adversarial examples have only 0.01 category with raw predictions more than 20. The reason that those adversarial examples appear more confident after softmax is because that
the predictions on all the other categories are regularized
even more. Hence the normalization component of softmax
has decided that the single prediction, although much less
strong, should be assigned a probability of more than 90%.
We note that this issue was also pointed out by [2] in a different manner and they proposed a solution in the OpenMax
classifier, which we compare against in the experiments.
But besides that, it seems that such extremal and standard deviation statistics are evident features that could help
discriminating normal and adversarial examples. Unfortunately, they only occur as a statistic from a large sample, as
any single point in Fig. 2(a) looks similar to a single point
in the normal distribution. We have tried to utilize the tail
distributions (Fig. 2(b)) to create a classifier which easily
achieved more than 99% accuracy separating adversarials from normals, however we subsequently found out that
since the tail almost do not contribute to the classification,
knowing this defense, the adversarial example can deliberately remove their footprints on the tail distributions.
This leads us to think about an approach that would turn
a single image into a distribution, so that we can use statistics as detectors for adversarial examples. An image is a
distribution of pixels. Especially, the output of each filter
from each convolutional layer is an image which could be
1
0
-1
-2
8
4
2
0
-2
-3
-4
-4
-6
-6
-4
-2
0
10
6
2
8
6
4
-10
-5
Eigenvector Number 1
0
5
Eigenvector Number 3547
(a)
2
0
500
1000 1500 2000 2500 3000 3500 4000 4500
Eigenvector Number
(b)
Normalized Standard Deviation
1.6
Normal examples
Adversarial Examples
Extremal Value
2
Eigenvector Number vs. Extremal Value
12
Normal examples
Adversarial Examples
10
Eigenvector Number 3984
Eigenvector Number 2
Tail Distribution
12
Normal examples
Adversarial Examples
3
(c)
Normalized Standard Deviation
Head Distribution
4
Normal examples
Adversarial Examples
1.4
1.2
1
0.8
0.6
0
500
1000 1500 2000 2500 3000 3500 4000 4500
Eigenvector Number
(d)
Figure 2. Blue indicates normal examples and red/orange indicate adversarial examples. Projections are normalized by dividing the standard
deviation of all normal examples projected on the corresponding dimension. (a) The projection of the data at layer 14 onto the 2 most
prominent directions; Adversarial example cannot be identified from normal ones. (b) Projection of the same data to the 3, 547-th and
3, 844-th PCA projections, some adversarial examples are having significantly higher deviation to the mean; (c) The absolute normalized
extremal value in the projection to each eigenvector; (d) The average normalized standard deviation of normal and adversarial examples on
each projection. Standard deviations of normal examples stand at 1 because of the normalization.
treated as a distribution where the samples are the pixels.
Therefore, in the following section we aim to build a classifier based on collecting statistics from such distributions.
4. Identifying Adversarial Examples
4.1. Feature Collection
Suppose the output at a convolutional layer m is an
W × H × K tensor, where W and H represent the width
and height of the image at that stage (smaller than original after max-pooling), and K represents the number of
convolutional filters. Such a tensor can be considered as a
K-channel image where each pixel has a K-dimensional
feature. We consider the feature on every pixel to be a
random vector drawn from the distribution Dm of convolutional pixel outputs, a K-dimensional distribution.
Average Number of Categories per
Example with Prediction Above Threshold
Normal examples
Adversarial Examples
102
Average Number of Categories
Above Threshold
Average Number of Categories
Average Number of Categories per Example
with Prediction Larger than Threshold
100
10
-2
10-4
0
10
20
30
40
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
50
0.1
Network Prediction Threshold before SoftMax
(a)
Normal examples
Adversarial Examples
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Network Prediction Threshold After SoftMax
(b)
Figure 3. Average number of categories per example with predictions higher than a threshold. (a) Before softmax; (b) After softmax. As one can see, in normal examples, there are on average
about 1 category with a prediction score of more than 20 (before
softmax), while with adversarial ones, only 1% examples have a
category with a prediction score more than 20. However, since
prediction values on all categories have dropped, after softmax adversarial examples obtain much higher likelihood on one category.
on each of the K-dimensional features. Normalized PCA
coefficients are collected via Algorithm 1. Extremal and
percentile statistics are straightforward to understand.
The features we collect are non-subdifferentiable, hence
essentially preventing adversaries to use gradient-based attacks to counter the classifier. Although we are interested in a generative adversarial network-type adversary
which would learn to avoid our detector, such adversaries
would have to resort to derivative-free optimization methods, which currently do not scale to the size of a realistic image. The best derivative-free approach we have tried scales
up to several hundreds of variables. The genetic algorithm
in [20] scales better, but as we will soon show, their lowlevel feature statistics are so different from natural images,
making them very easy to be detected, even without training
on any data from their adversarial generation algorithm.
Algorithm 1 PCA Statistics Extraction
1: INPUT: Image I, layer m.
2: For all normal images in a training set, compute their
CNN filter output of layer m to form an example matrix
Zm .
3: Compute the mean e and PCA projection matrix W of
Zm .
4: Compute the standard deviation s on each dimension in
the PCA projection W> (Zm − e1> ).
5: For each image I, project its CNN filter output of layer
m ZmI using PCA: zmI = W> (ZmI − e1> ), and
normalize them by dividing the standard deviation s on
each respective dimension.
1
6: Collect the statistic for each image as xI = n
kzmI k1 ,
where L1 norm is the vector L1 norm. The resulting
statistic is K-dimensional.
The list of statistics we collect is:
• Normalized PCA coefficients
4.2. Classifier Cascade
• Minimal and Maximal values
[29] proposed a famous strategy for face detection by using a cascaded boosting classifier composed by a sequence
• 25-th, 50-th and 75-th percentile values
of base classifiers. A cascade classifier is ideal when it is
easy to identify many of the examples from a category but
some important cases can be difficult. In Fig. 4, SC N
represents the classifier at each stage. X is the input of the
cascade classifier. The negatives in a cascade classifier from
each stage will be outputted directly, while the positives will
go to the next stage.
In our case, the normal category is much easier to detect
than the adversarial category (see e.g. Fig. 6). In our initial
experiments with VGGNet, we found that more than 80%
of normal examples can be determined from the first convolutional layer with 100% precision. Therefore, we constructed a cascade classifier based on convolutional layers:
the first stage works with features collected from the outputs
of the first convolutional layer, the second with the second
layer, etc. (Fig. 4). The base classifiers will not solely consider statistics from their own stage, instead, after one stage
of training, the remaining positive examples will be concatenated to the corresponding features on the next stage.
Figure 4. A cascade classifier is defined on each of the convolutional layers in a convolutional network (SC i represents the i-th
convolutional layer)
The operations that are represented by Fig. 4 can also be
summarized as Algorithm 2.
Algorithm 2 Training Process of a cascade of Classifier
1: Npool ← Normal example pool, Ptrain ← Training set
of Np perturbed examples, L ← Number Of convolutional layers
2: while current layer ≤ L Or Npool 6= ∅ do
3:
Draw Np sized subset Pnormal from Npool
4:
T ← Pnormal ∪ Ptrain
5:
Train SVM on T
6:
Predict SVM on Npool , eliminate those predicted as
normal above a threshold (described in text)
7: end while
The overall false positive rate ofQa K stage cascade clasK
sifier can be represented as: F = i=1 fi , where fi is the
false positive rate at each layer. And similarly the true
QKpositive rate can be represented in the same form: T = i=1 ti
where ti is the true positive rate at each stage. In order to
maximize recall, we maintain a high true positive rate and
select a classification threshold which corresponds to a high
true positive rate (97% in AlexNet and 98% in VGG).
5. Related Work
Szegedy et al. [28] proposes the adversarial optimization formulation in eq. (2). [4] proposes an explanation
of the adversarial mechanism, and proposed a simpler adversarial optimization mechanism that only corrupts based
on the signs of gradient of the network. The fact that such
examples can be generated so easily with the gradient sign
method shows that adversarial examples come from attacking the magnifying effect coming from the linearities in the
network. [20] proposes another mechanism to generate adversarials using evolutionary optimization. The result of
these do not resemble natural images but still can be classified by deep networks with high confidence(Fig. 5). [19]
proposes another efficient approach. [23] proposes an approach to generate adversarials that match the convolutional
filter outputs as well as perturbing the data. [25, 8] propose
approaches to sample adversaries or minimax optimization
for making learning more robust. While most of the work
are done on standard benchmarks such as MNIST, CIFAR
and ImageNet, [14] is an interesting work on projecting the
adversaries in physical world.
Recently, there have been a lot of focus on training adversarial generation networks to create Generative Adversarial Networks (GANs) [3, 22, 32, 24]. These networks
play a two-player game where a generator network aims to
generate adversarials that will not be correctly classified by
another discriminator network, and the goal is to generate
images more and more similar to natural images. It has been
shown that these networks generate images that resemble
natural images. However, this generative approach is different from our goal, where we aim to create discriminative
networks that discriminates from images that are already indistinguishable from natural images (e.g. Fig.1).
Mechanisms for detecting and countering adversarial examples have also been proposed [5, 21]. [18] proposes to
use the foveation mechanism to alleviate adversarial examples when it is already known to be adversarial, but did not
attempt to detect adversarials. The open-set deep networks
proposed by [2] seek to alleviate concerns from a soft-max
classification by creating an abstain option. The universum classifier [31] is similar but with more theoretical arguments.
Self-aware learning (classification with an abstain option) had been proposed in e.g. [12, 16]. It is relevant to
robust learning (e.g. [15]), however robust learning usually
seek to directly optimize the minimax loss under adversarial
conditions, instead of outputting an abstain option. [1, 30]
also focuses on classification with an abstain option.
6. Experiments
Our algorithm is tested on 2 approaches to generate adversaries. The main one is data generated using the L-BFGS
were much easier to distinguish, even though our classifier was trained only on L-BFGS adversarials instead of EA
ones (Table 3). We only need the first three convolutional
layers to reach an overall 97.34% classification accuracy.
We believe the reason is that our features capture natural
image statistics, and because EA-adversarials look so unnatural, their statistics are vastly different than natural images (see Sec. 6.3 for more discussions).
Table 1. Classification Result with AlexNet for Normal vs.
LBFGS-adversarials
Figure 5. Evolved Adversarial Images Unrecognizable to Humans
algorithm by [28]. We generated 4, 000 adversarials from
a random subset of the ILSVRC-2012 validation set (total
of 50, 000 images). In order to test the out-of-sample generalization capability, we included another dataset, which
includes 5, 000 EA-adversarial images generated using the
algorithm in [20]. These datasets are tested with 2 different network structures, VGG-16 [26] and AlexNet [13].
All input images had been normalized and re-shaped into
224 × 224 for VGG-16 model and 227 × 227 for AlexNet.
The MatConvNet toolbox is used for the deep networks.
Features are extracted from the statistics of interest, mentioned in section 4.1, in the lower 9 convolutional layers
from the VGG-16 network, and in all 5 convolutional layers
from AlexNet.
All the classifiers were trained with a subset of 2, 000
L-BFGS adversarials and a randomly selected subset of
45, 000 normal images in the ILSVRC-2012 dataset. Testing was done on a random held-out dataset with the other
5, 000 normal images and 2, 000 L-BFGS adversarials for
the L-BFGS experiments, and on a randomly selected 5, 000
normal images and 5, 000 EA-adversarials for the EA experiments. The training and testing sets do not overlap in
any means: for all the testing LBFGS adversarials, the corresponding normal images were not seen in the training set.
EA experiments are only conducted with AlexNet since we
found out that EA-adversarials did not succeed fooling the
VGG-16 network. We have also tried some other fast adversarial generation approaches such as [4], but they generally do not generate adversarials that fool classifiers as completely (with more than 90% confidence) or as universally
(on 100% of the images) as the L-BFGS algorithm.
Additional experiment results using DeepFool and on
ResNet can be found in the supplementary material.
6.1. Single-Layer Results
The first experiments we did were by extracting statistics
on each single convolutional layer output, instead of using
the cascade classifier. From Tables 1 and 2, one can see that
the statistics from a single layer have some discriminative
power to distinguish normal examples from adversarials,
but are not extremely effective. However, EA-adversarials
Network Layer
Accuracy
Network Layer
Accuracy
2nd
57.5 ± 0.7
5th
74.9 ± 0.9
3rd
67.3 ± 0.7
6th
78.95 ± 0.6
4th
70.9 ± 0.6
Table 2. Classification Result with VGG-16 for Normal vs.
LBFGS-Adersarials
Network Layer
Accuracy
Network Layer
Accuracy
Network Layer
Accuracy
2nd
72.1 ± 0.7
5th
81.4 ± 0.9
8th
74.2 ± 0.7
3rd
84.1 ± 0.7
6th
74.3 ± 0.6
9th
71.2 ± 0.7
4th
80.3 ± 0.6
7th
73.9 ± 0.6
10th
74.3 ± 0.8
Table 3. Classification Result for Normal vs. EA-Adversarials
Layer
Accuracy
2nd
93.45 ± 0.69
3rd
98.3 ± 0.73
4th
97.9 ± 0.57
6.2. Experiment for LBFGS-Adversarials Detection
Next we test the cascade classifier on both AlexNet and
VGG-16. The parameter C is set to 0.005. On AlexNet,
the average accuracy of the cascade classifier reaches 83.4%
over 20 random trials, and the AUC (area-under-curve) metric is 90.7%. We compared against the recently published
OpenMax method [2]. To learn the Weibull distribution
required for OpenMax, the EVT was applied on the same
training set as the algorithm. Figure 6(a) shows the results,
where we were able to outperform OpenMax by over 9% in
area-under-curve (AUC) and 11% in terms of accuracy.
In VGG-16, the results were even better. The accuracy
of the classifier was on average 90.665% over 20 random
trials. Fig.6(b) shows the ROC curve. We believe the fact
that VGG has a lot more layers than AlexNet helps setting
more constraints on the layer statistics, and is subsequently
helpful for detecting adversarial examples.
Finally, the cascade classifier was tested on EAadversarials. We obtained more than 96% accuracy with 0
false positive rate, with a final accuracy of 97.3% and AUC
of 98.2% (Fig.6(c)). In other words, our algorithm is rarely
fooled by EA-adversarials, even without training on them.
(a)
(b)
(c)
Figure 6. (a) Comparison Between OpenMax detection Methods and Cascade Classifier: The blue curve represents the performace of
OpenMax Method, and green curve represents the perfornace for Cascade Classifier.(b) Overall ROC Performance Curve of Cascade
Classifier Trained on VGG-16 Network. (c) Overall ROC of data generated from EA-adversarials dataset on AlexNet.
6.3. Visualization of Statistics
Our experiment results show that EA-adversarials are
easy to detect with our detector. To gain more insight
into this result, we made a few comparisons between the
statistics of interest extracted from normal images, LBFGSadversarials and EA-adversarials.
We visualized the average of the statistics that are used
for the detection task from the first layer of the AlexNet
on all its dimensions. As can be seen in Fig.7(a), the difference on the PCA projection statistics on extracted from EAadversarials and that of the normal images is very dramatic.
Meanwhile, compared to the EA-adversarials, the statistics
from LBFGS-adversarial have much less difference from
the normal data and the difference does not change very
much across different dimensions.
From Fig. 7(b), one can see that LBFGS-adversarials
have smaller extremal values than normal images. This
might imply that the LBFGS optimization worked to diminish strong signals from the original image by introducing small pixel perturbations, and that helped our classifiers
separating them from normal images. From Fig. 7(c), we
see the EA-adversarials evidently differ from normal images. Those results illustrate why EA-adversarials are easier
to detect. We suspect it would be easy to reach 100% accuracy, had we actually trained on some EA-adversarials. The
capability to generalize to EA-adversarials without training
on them showed the general capability of our cascade classifiers to capture natural image statistics and distinguish natural images from unnatural ones.
7. Discussions
7.1. Self-Aware Learning with an Abstain Option
The framework of self-aware learning [16, 2, 31] considers the case where the learning algorithm has an abstain
option of saying “I don’t know”, instead of always making
an actual prediction. We define a framework that is slightly
different than [16], avoiding the requirement in some frameworks of never making a mistake.
We assume that the training input is drawn i.i.d. from a
distribution P (x, y), where x is the input and y is the output. Assume that the testing input is drawn from a mixture
distribution between P (x, y) and Q(x, y):
Pm = ΩP (x, y) + (1 − Ω)Q(x, y)
(4)
, where Ω ∈ {0, 1} is an unknown mixture weight, and
Q(x, y) is an adversarial distribution. Assume that we have
a classifier that includes a function f (x), and a boolean
strategy ai between predict and abstain that can be
chosen for each individual xi . Assume that the expected error from our classifier on the adversarial distribution is eq
(which could be assumed, if no other prior is present, as
the random guessing error of C−1
C for a C-class classification problem). Further assume that abstaining always incur
a fixed cost ea . As long as ea < eq , abstaining would be
better than predicting on the example drawn from the adversarial distribution, however, ea should be set sufficiently
large so that the classifier would still make predictions when
confident, instead of abstaining everything.
For each testing input, the testing of the self-aware classifier is then trying to optimize mina EPm La (x, y) where
P (yi 6= f (xi )), if ai = predict,
(xi , yi ) ∼ P (x, y)
eq
if ai = predict
La (xi , yi ) =
, (xi , yi ) ∼ Q(x, y)
ea
if ai = abstain
(5)
hence the classifier needs to select between making a prediction using its function f (x) and risk paying eq versus
abstaining. It is easy to derive the optimal strategy:
ai = predict, if P (Ω = 1|xi )P (yi 6= f (xi ))
(6)
+P (Ω = 0|xi )eq < ea
ai = abstain,
otherwise
(7)
(a)
(b)
(c)
Figure 7. (a) PCA Projection Comparison; (b) Maximum Feature Map Extremal Value Comparison; (c) Median Value Comparison
Our approach can be seen as estimating P (Ω = 1|xi ) in
this framework. Experiments about the effect of such selfaware learning is shown in the supplementary material. We
eagerly hope to apply it in realistic applications in future
work.
7.2. Image Recovery
Insights from [4] indicate that the adversarial mechanism
is very specifically attacking vulnerable gradients starting
from the first convolutional layer. Insights from the previous experiments also suggest that LBFGS-adversarials
work to diminish filter responses from the first convolutional layer. Therefore a natural idea would be to destroy
the adversarial effects in the first convolutional layer to try
to recover the original image. We tried a very simple approach: applying a small (e.g. 3 × 3) average filter on the
adversarial image before using the CNN to classify it. The
positive and negative adverse gradients will average out in
this approach, and make the masked activations from the
normal images more prominent. In Table 4 we illustrate
such recovery results: after using a 3 × 3 average filter on
identified adversarial examples, the classification accuracy
improved from almost 0% to 73.0%, showcasing the effectiveness of this simple average filter.
Table 4. Recovery Results. Simply using a 3 × 3 average filter
we can recover a large proportion of adversarial examples after
detecting them using the algorithm described previously. More
complex cancellation approaches such as foveation in [18] that
utilizes cropping can achieve better results.
Approach
Original Image (Non-corrupted)
3 × 3 Average Filter
5 × 5 Average Filter
Foveation (Object Crop MP) [18]
Top-5 Accuracy
(Recovered Images)
86.5%
73.0%
68.0%
82.6%
Those results show that we can both detect and recover
from adversarial examples with high accuracy. But the main
reason we performed this (overly simplistic) experiment is
to show how simple it might be to cancel out some adversarial perturbations. Importantly, this result indicates that current deep convolutional networks are too locally focused:
these are corruptions that can be cancelled out by a simple
3 × 3 average filter, however they can adversely impact the
entire result of the deep network. For human with a large
receptive field, they will not even care about what happens
within a 3 × 3 area. Therefore, we believe that future deep
learning approaches should focus on enlarging the receptive
field in order to reduce the chance of being fooled by adversarial examples. Another potential direction is to research
classification approaches that do not require a softmax-type
normalization, in order to avoid regularizing attacks such as
the ones used in the adversarial optimization in (2).
8. Conclusion
This paper proposes an approach that detects adversarial examples using simple statistics on convolutional layer
outputs. A cascade classifier was designed based on simple
statistics on filter outputs from each layer. And it was capable of detecting more than 85% of the adversarial examples.
Experiments showed that our cascade classifier significantly
outperforms state-of-the-art on detecting adversarial examples. Experiment also showed transfer learning capabilities of our classifier, since the classifier we trained with LBFGS adversarials are capable of detecting EA-adversarials
as well. Insights drawn from these experiments lead us to
perform simple 3 × 3 average filter to corrupted images,
which successfully recovered most of them. In the future,
we would like to explore GAN-type generative adversarial
networks from the current results, with multiple rounds of
adversarial detection and counter-detection.
Acknowledgements
This paper was supported by Future of Life grants 2015143880 and 2016-158701.
References
[1] A. Balsubramani. Learning to abstain from binary prediction. arXiv preprint arXiv:1602.08151, 2016.
[2] A. Bendale and T. E. Boult. Towards open set deep networks.
In IEEE Conference on Computer Vision and Pattern Recognition, 2016.
[3] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu,
D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In Advances in Neural Information
Processing Systems, pages 2672–2680, 2014.
[4] I. J. Goodfellow, J. Shlens, and C. Szegedy. Explaining and harnessing adversarial examples. arXiv preprint
arXiv:1412.6572, 2014.
[5] S. Gu and L. Rigazio. Towards deep neural network architectures robust to adversarial examples. arXiv preprint
arXiv:1412.5068, 2014.
[6] T. Hastie, R. Tibshirani, and J. Friedman. The Elements of
Statistical Learning. Springer-Verlag, New York, 2001.
[7] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning
for image recognition. In IEEE Conference on Computer
Vision and Pattern Recognition, 2016.
[8] R. Huang, B. Xu, D. Schuurmans, and C. Szepesvári. Learning with a strong adversary. In International Conference on
Learning Representations, 2016.
[9] P. Indyk and R. Motwani. Approximate nearest neighbors:
towards removing the curse of dimensionality. In Proceedings of the thirtieth annual ACM symposium on Theory of
computing, pages 604–613, 1998.
[10] S. Ioffe and C. Szegedy. Batch normalization: Accelerating
deep network training by reducing internal covariate shift.
arXiv preprint arXiv:1502.03167, 2015.
[11] I. Jolliffe. Principle Component Analysis. Springer-Verlag,
1986.
[12] R. Kleinberg, A. Niculescu-Mizil, and Y. Sharma. Regret
bounds for sleeping experts and bandits. Machine learning,
80(2-3):245–272, 2010.
[13] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet
classification with deep convolutional neural networks. In
Advances in Neural Information Processing Systems, pages
1097–1105, 2012.
[14] A. Kurakin, I. Goodfellow, and S. Bengio. Adversarial examples in the physical world. arXiv preprint arXiv:1607.02533,
2016.
[15] G. R. Lanckriet, L. E. Ghaoui, C. Bhattacharyya, and M. I.
Jordan. A robust minimax approach to classification. Journal
of Machine Learning Research, 3:555–582, 2003.
[16] L. Li, M. L. Littman, T. J. Walsh, and A. L. Strehl. Knows
what it knows: a framework for self-aware learning. Machine
learning, 82(3):399–443, 2011.
[17] X. Li, F. Li, X. Fern, and R. Raich. Filter shaping for convolutional networks. In International Conference on Learning
Representations, 2017.
[18] Y. Luo, X. Boix, G. Roig, T. A. Poggio, and Q. Zhao.
Foveation-based mechanisms alleviate adversarial examples.
arXiv preprint arXiv:1511.06292v3, 2016.
[19] S. Moosavi-Dezfooli, A. Fawzi, and P. Frossard. Deepfool:
a simple and accurate method to fool deep neural networks.
CoRR, abs/1511.04599, 2015.
[20] A. Nguyen, J. Yosinski, and J. Clune. Deep neural networks
are easily fooled: High confidence predictions for unrecognizable images. In IEEE Conference on Computer Vision
and Pattern Recognition, 2015.
[21] N. Papernot, P. McDaniel, X. Wu, S. Jha, and A. Swami. Distillation as a defense to adversarial perturbations against deep
neural networks. arXiv preprint arXiv:1511.04508, 2015.
[22] A. Radford, L. Metz, and S. Chintala. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.
[23] S. Sabour, Y. Cao, F. Faghri, and D. J. Fleet. Adversarial
manipulation of deep representations. In International Conference on Learning Representations, 2016.
[24] T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen. Improved techniques for training gans.
arXiv preprint arXiv:1606.03498, 2016.
[25] U. Shaham, Y. Yamada, and S. Negahban. Understanding adversarial training: Increasing local stability of neural nets through robust optimization.
arXiv preprint
arXiv:1511.05432, 2015.
[26] K. Simonyan and A. Zisserman. Very deep convolutional
networks for large-scale image recognition. arXiv preprint
arXiv:1409.1556, 2014.
[27] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed,
D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich.
Going deeper with convolutions. arXiv:1409.4842, 2014.
[28] C. Szegedy, W. Zaremba, I. Sutskever, J. Bruna, D. Erhan,
I. Goodfellow, and R. Fergus. Intriguing properties of neural
networks. arXiv preprint arXiv:1312.6199, 2013.
[29] P. Viola and M. J. Jones. Robust real-time face detection.
International journal of computer vision, 57(2):137–154,
2004.
[30] Y. Wiener and R. El-Yaniv. Agnostic selective classification. In Advances in Neural Information Processing Systems,
pages 1665–1673, 2011.
[31] X. Zhang and Y. LeCun. Universum prescription: Regularization using unlabeled data. In AAAI Conference on Artificial Intelligence, 2017.
[32] J. Zhao, M. Mathieu, and Y. LeCun. Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126,
2016.
Supplementary Material
1. Results on DeepFool
For this experiment we used 5000 adversarial images
generated with the DeepFool algorithm. We collected a
training set of images using the adversarials, with an equal
number of real images drawn from the ILSVRC2012 validation set. For the Deep- Fool adversarials we used the implementation given in the Foolbox algorithm library. Given
this data, we chose to use the ResNet 50 architecture as the
backbone CNN for our experiments. Because we use ImageNet data, we preprocessed our training set accordingly:
all images were reshaped to 224x224x3, the channels were
modified to BGR ordering, and the channel-wise mean was
subtracted from each sample. We then performed a forward
pass as usual on all images, performing the classifier cascade at each layer. Given the output of each convolutional
layer, we extracted features that would characterize images
as being either from a real or adversarial distribution. For
each output we extracted PCA coefficients, extremal values,
and values within the 25th, 50th and 75th percentile to form
a new training sample. We then used an SVM to learn the
real statistics from the adversarials. The real examples were
given a ground truth label of 1, and the adversarials were
given a label of 0. We performed a parameter search over
kernel type and C value. In all our experiments a C of 0.005
and a linear kernel performed the best. We then tested on
2000 real images and 2000 DeepFool adversarials. In our
experiments with ResNet, we only performed the cascade
for the first three layers before we found nearly all of the
adversarial images.
The result can be seen in Fig. 1. The algorithm maintained more than 90% AUC, showing that DeepFool did not
fundamentally change the type of adversarials.
2. Results on Self-Aware Learning
We implemented self aware learning using ResNet 50,
the goal being to use the softmax probabilities to learn parameters that would cause the network to more intelligently
classify inputs. The network should classify an image if it
was sure the image was real, or abstain if either the network
was not sufficiently confident, or if the image was adversarial, as described in Sec. 7.1. To test the presented algorithm,
we use 2, 000 real images drawn from ILSVRC2012 validation set, and 2, 000 adversarial images from the testing set
of the previous experiment, generated using the DeepFool
algorithm. We tested the self aware learning algorithm with
a high eq = 10. This worked well enough that the network
chose to abstain or classify, rather than incur a high penalty
for guessing incorrectly. We observed that for each testing image, our estimation of the source distribution resulted
in ea between 2 and 8. We then varied ea between these
values to see if there was a threshold at which we could
abstain from all adversarials, retaining predictions for only
real examples. We were also interested in thresholds that
maximized the true positive rate (prediction of real examples) while abstaining from as many adversarials as possible. We found the lower thresholds resulted in the abstaining from predicting on all adversarials, but it also abstained
from many (but not all) real examples. Higher thresholds
resulted in many more real predictions retained, but some
also some adversarials made it through. High thresholds
would finally result in the network not abstaining at all.
The results can be seen in Fig. 2. It can be seen that besides abstaining adversarial examples, the system also abstains from predicting on some normal examples that the
classifier is not confident on. Hence, with a high abstain
ratio the prediction accuracy on normal examples is also
higher.
3. Images Classified Correctly and Incorrectly
In this section we show some images classified correctly and incorrectly from the algorithm. Unfortunately we
are not quite able to observe any particular visible trends,
maybe due to the subtlety of adversarial images.
Figure 1. ROC Curve for Detecting DeepFool adversarials in ResNet-50, the algorithm maintained more than 90% in AUC
Figure 2. Self-aware learning results. In a mixture of half real and half adversarial examples, the classification accuracy of discarding
nothing falls a little under 50%, with more examples abstained, the accuracy improves significantly. The accuracy of retained normal
examples (blue curve) also improves when more examples are abstained, as the abstained examples also include normal examples that are
not predicted confidently.
(a)
(b)
(c)
(d)
Figure 3. Some of Misclassification on L-BFGS images by Our Classifier. (a) and (b) are from normal dataset. (c) and (d) are from
LBFGS-Adversarial dataset, which is misclassified to category n02408429(water buffalo) and n01518878(ostrich, Struthio camelus).
(a)
(b)
(c)
(d)
Figure 4. Some of Correctly Classified on L-BFGS images by Our Classifier. (a) and (b) are from normal dataset. (c) and (d) are from
LBFGS-Adversarial dataset, which is misclassified to category n04209133(shower cap) and n02328150(Angora).
(a)
(b)
(c)
(d)
Figure 5. Some of Misclassfied EA images by Our Classifier. From left to right, they are misclssified to category n03220513 (dome),
n01749939 (green mamba), n04118776 (rule, ruler) and n03935335 piggy (bank, penny bank)
Figure 6. Some of Correctly Classified EA images by Our Classifier. From left to right they are misclassified to n06874185 (traffic light,
traffic signal, stoplight), n03443371 (goblet), n04522168 (vase) and n03742115 (medicine chest, medicine cabinet)
(a)
(b)
(c)
(d)
Figure 7. Images Misclassified by OpenSet Method but Correctly Classified by Our Classifier. (c) and (d) are from LBFGS-Adversarial
dataset, which is misclassified to category n02133161(American black bear) and n02328150(Agona).
| 1cs.CV
|
Learning One-hidden-layer Neural Networks with Landscape Design
Rong Ge
Jason D. Lee
Tengyu Ma
arXiv:1711.00501v2 [cs.LG] 3 Nov 2017
November 6, 2017
Abstract
We consider the problem of learning a one-hidden-layer neural network: we assume the input
x ∈ Rd is from Gaussian distribution and the label y = a⊤ σ(Bx) + ξ, where a is a nonnegative
vector in Rm with m ≤ d, B ∈ Rm×d is a full-rank weight matrix, and ξ is a noise vector. We first
give an analytic formula for the population risk of the standard squared loss and demonstrate
that it implicitly attempts to decompose a sequence of low-rank tensors simultaneously.
Inspired by the formula, we design a non-convex objective function G(·) whose landscape is
guaranteed to have the following properties:
1. All local minima of G are also global minima.
2. All global minima of G correspond to the ground truth parameters.
3. The value and gradient of G can be estimated using samples.
With these properties, stochastic gradient descent on G provably converges to the global minimum and learn the ground-truth parameters. We also prove finite sample complexity result and
validate the results by simulations.
1
Introduction
Scalable optimization has been playing crucial roles in the success of deep learning, which has
immense applications in artificial intelligence. Remarkably, optimization issues are often addressed
through designing new models that make the resulting training objective functions easier to be
optimized. For example, over-parameterization [LSSS14], batch-normalization [IS15], and residual
networks [HZRS16a, HZRS16b] are often considered as ways to improve the optimization landscape
of the resulting objective functions.
How do we design models and objective functions that allow efficient optimization with guarantees? Towards understanding this question in a principled way, this paper studies learning neural
networks with one hidden layer. Roughly speaking, we will show that when the input is from
Gaussian distribution and under certain simplifying assumptions on the weights, we can design an
objective function G(·), such that
[a] all local minima of G(·) are global minima
[b] all the global minima are the desired solutions, namely, the ground-truth parameters (up to
permutation and some fixed transformation).
We note that designing such objective functions is challenging because 1) the natural ℓ2 loss
objective does have bad local minimum, and 2) due to the permutation invariance1 , the objective
1
Permuting the rows of B ⋆ and the coordinates of a⋆ correspondingly preserves the functionality of the network.
1
function inherently has to contain an exponential number of isolated local minima.
1.1
Setup and known issues with proper learning
We aim to learn a neural network with a one-hidden-layer using a non-convex objective function.
We assume input x comes from Gaussian distribution and the label y comes from the model
y = a⋆ ⊤ σ(B ⋆ x) + ξ
(1.1)
where a⋆ ∈ Rd , B ⋆ ∼ Rm×d are the ground-truth parameters, σ(·) is a element-wise non-linear
function, and ξ is a noise vector with zero mean. Here we can without loss of generality assume x
comes from spherical Gaussian distribution N (0, Idd×d ). 2
For technical reasons, we will further assume m ≤ d and that a⋆ has non-negative entries.
The most natural learning objective is perhaps the ℓ2 loss function, given the additive noise.
Concretely, we can parameterize with training parameters a ∈ Rd , B ∼ Rm×d of the same dimension
as a⋆ and B ⋆ correspondingly,
ŷ = a⊤ σ(Bx) ,
(1.2)
and then use stochastic gradient descent to optimize the ℓ2 loss function. When we have enough
training examples, we are effectively minimizing the following population risk with stochastic updates,
f (a, B) = E kŷ − yk2 .
(1.3)
However, empirically stochastic gradient descent cannot converge to the ground-truth parameters in the synthetic setting above when σ(x) = ReLU(x) = max{x, 0}, even if we have access
to an infinite number of samples, and B ⋆ is a orthogonal matrix. Such empirical results have been
reported in [LSSS14] previously, and we also provide our version in Figure 1 of Section 6. This
is consistent with observations and theory that over-parameterization is crucial for training neural
networks successfully [LSSS14, HMR16, SC16].
These empirical findings suggest that the population risk f (a, B) has spurious local minima
with inferior error compared to that of the global minimum. This phenomenon occurs even if we
assume we know a⋆ or a⋆ = 1 is merely just the all one’s vector. Empirically, such landscape issues
seem to be alleviated by over-parameterization. By contrast, our method described in the next
section does not require over-parameterization and might be suitable for applications that demand
the recovery of the true parameters.
1.2
Our contributions
Towards learning with the same number of training parameters as the ground-truth model, we first
study the landscape of the population risk f (·) and give an analytic formula for it — as an explicit
function of the ground-truth parameter and training parameter with the randomness of the data
being marginalized out. The formula in equation (2.3) shows that f (·) is implicitly attempting to
solve simultaneously a finite number of low-rank tensor decomposition problems with commonly
shared components.
2
This is because if x ∼ N (0, Σ), then we can whiten the data by taking x′ = Σ−1/2 x and define B ⋆′ = BΣ1/2 . We
note that B ⋆′ x′ = Bx and therefore we main the functionality of the model.
2
Inspired by the formula, we design a new training model whose associated loss function — named
and formally defined in equation (2.6) — corresponds to the loss function for decomposing a
matrix (2-nd order tensor) and a 4-th order tensor (Theorem 2.2). Empirically, stochastic gradient
descent on f ′ learns the network as shown in experiment section (Section 6).
Despite the empirical success of f ′ , we still lack a provable guarantee on the landscape of f ′ .
The second contribution of the paper is to design a more sophisticated objective function G(·)
whose landscape is provably nice — all the local minima of G(·) are proven to be global, and they
correspond to the permutation of the true parameters. See Theorem 2.3.
Moreover, the value and the gradient of G can be estimated using samples, and there are no
constraints in the optimization. These allow us to use straightforward stochastic gradient descent
(see guarantees in [GHJY15, JGN+ 17]) to optimize G(·) and converge to a local minimum, which
is also a global minimum (Corollary 2.4).
Finally, we also prove a finite-sample complexity result. We will show that with a polynomial
number of samples, the empirical version of G share almost the same landscape properties as G
itself (Theorem 2.7). Therefore, we can also use an empirical version of G as a surrogate in the
optimization.
f′
1.3
Related work
The work of Arora et al. [ABGM14] is one of the early results on provable algorithms for learning
deep neural networks, where the authors give an algorithm for learning deep generative models with
sparse weights. Livni et al. [LSSS14], Zhang et al. [ZLJ16, ZLWJ17], and Daniely et al. [DFS16]
study the learnability of special cases of neural networks using ideas from kernel methods. Janzamin
et al. [JSA15] give a polynomial-time algorithm for learning one-hidden-layer neural networks with
twice-differential activation function and known input distributions, using the ideas from tensor
decompositions.
A series of recent papers study the theoretical properties of non-convex optimization algorithms
for one-hidden-layer neural networks. Brutzkus and Globerson [BG17] and Tian [Tia17] analyze the
landscape of the population risk for one-hidden-layer neural networks with Gaussian inputs under
the assumption that the weights vector associated to each hidden variable (that is, the filters) have
disjoint supports. Li and Yuan [LY17] prove that stochastic gradient descent recovers the groundtruth parameters when the parameters are known to be close to the identity matrix. Zhang et
al. [ZPS17] studies the optimization landscape of learning one-hidden-layer neural networks with
a specific activation function, and they design a specific objective function that can recover a
single column of the weight matrix. Zhong et al. [ZSJ+ 17] studies the convergence of non-convex
optimization from a good initializer that is produced by tensor methods. Our algorithm works for a
large family of activation functions (including ReLU) and any full-rank weight matrix. To our best
knowledge, we give the first global convergence result for gradient-based methods for our general
setting.3
The optimization landscape properties have also been investigated on simplified neural networks
models. Kawaguchi [Kaw16] shows that the landscape of deep neural nets does not have bad local
minima but has degenerate saddle points. Hardt and Ma [HM17] show that re-parametrization using
identity connection as in residual networks [HZRS16a] can remove the degenerate saddle points in
3
The work of [JSA15, ZSJ+ 17] are closely related, but they require tensor decomposition as the algorithm/initialization.
3
the optimization landscape of deep linear residual networks. Soudry and Carmon [SC16] showed
that an over-parameterized neural network does not have bad differentiable local minimum. Hardt
et al. [HMR16] analyze the power of over-parameterization in a linear recurrent network (which is
equivalent to a linear dynamical system.)
The optimization landscape has also been analyzed for other machine learning problems, including SVD/PCA phase retrieval/synchronization, orthogonal tensor decomposition, dictionary
learning, matrix completion, matrix sensing [BH89, SJ13, GHJY15, SQW15, BBV16, GLM16,
BNS16, GJZ17]. Our analysis techniques build upon that for tensor decomposition in [GHJY15]
— we add two additional regularization terms to deal with spurious local minimum caused by the
weights a⋆ and to remove the constraints.
1.4
Notations:
We use N, R to denote the set of natural numbers and real numbers respectively. We use k·k to
denote the Euclidean norm of a vector and spectral norm of a matrix. We use k·kF to denote
the Frobenius/Euclidean norm of a matrix or high-order tensor. For a vector x, let kxk0 denotes
its infinity norm and for a matrix A, let |A|0 be a shorthand for kvec(A)k0 where vec(A) is the
vectorization of A. For a vector x, let |x|2nd denotes the second largest absolute values of the entries
for x. We note that |·|2nd is not a norm.
We use A ⊗ B to denote the Kronecker product of A and B, and A⊗k is a shorthand for
A ⊗ · · · ⊗ A where A appears k times. For vectors a ⊗ b and a⊗k denote the tensor product. We use
λmax (·), λmin (·) to denote the largest and smallest eigenvalues of a square matrix. Similarly, σmax (·)
and σmin (·) are used to denote the largest and smallest singular values. We denote the identity
matrix in dimension d × d by Idd×d , or Id when the dimension is clear from the context.
In the analysis, we rely on many properties of Hermite polynomials. We use hj to denote the jth normalized Hermite polynomial. These polynomials form an orthonormal basis. See Section 4.1
for an introduction of Hermite polynomials.
We will define other notations when we first use them.
2
2.1
Main Results
Connecting ℓ2 Population Risk with Tensor Decomposition
We first show that a natural ℓ2 loss for the one-hidden-layer neural network can be interpreted as
simultaneously decomposing tensors of different orders.
A straightforward approach of learning the model (1.1) is to parameterize the prediction by
ŷ = a⊤ σ(Bx) ,
(2.1)
where a ∈ Rd , B ∼ Rm×d are the training parameters. Naturally, we can use ℓ2 as the empirical
loss, which means the population risk is
f (a, B) = E kŷ − yk2 .
(2.2)
Throughout the paper, we use b⋆1 ⊤ , . . . , b⋆m ⊤ to denote the row vectors of B ⋆ and similarly for B.
4
⋆⊤
b1
b⊤
1
..
..
⋆
That is, we have B = . and B = . . Let ai and a⋆i ’s be the coordinates of a and a⋆
b⊤
b⋆m ⊤
m
respectively.
We give the following analytic formula for the population risk defined above.
Theorem 2.1. Assume vectors bi , b⋆i ’s are unit vectors. Then, the population risk f defined in
equation (2.2) satisfies that
f (a, B) =
X
σ̂k2
k∈N
X
a⋆i b⋆i ⊗k
i∈[m]
−
X
2
ai b⊗k
i
i∈[m]
+ const .
(2.3)
F
where σ̂k is the k-th Hermite coefficient of the function σ. See section 4.1 for a short introduction
of Hermite polynomial basis. 4
Connection to tensor decomposition: We see from equation (2.3) that the population risk of f is
essentially an average of infinite number of loss functions for tensor decomposition. For a fixed
k ∈ N, we have that the k-th summand in equation (2.3) is equal to (up to the scaling factor σ̂k2 )
fk , Tk −
X
2
ai b⊗k
i
i∈[m]
.
(2.4)
F
P
where Tk = i∈[m] a⋆i b⋆i ⊗k is a k-th order tensor in (Rd )⊗k . We note that the objective fk naturally
⊗k
attempts to decompose the k-order rank-m tensor Tk into m rank-1 components a1 b⊗k
i , . . . , am bm .
The proof of Theorem 2.1 follows from using techniques in Hermite Fourier analysis, which is
deferred to Section 4.2.
Issues with optimizing f :. It turns out that optimizing the population risk using stochastic
gradient descent is empirically difficult. Figure 1 shows that in a synthetic setting where the noise
is zero, the test error empirically doesn’t converge to zero for sufficiently long time with various
learning rate schemes, even if we are using fresh samples in iteration. This suggests that the
landscape of the population risk has some spurious local minimum that is not a global minimum.
See Section 6 for more details on the experiment setup.
An empirical fix:. Inspired by the connection to tensor decomposition objective described earlier
in the subsection, we can design a new objective function that takes exactly the same form as the
tensor decomposition objective function f2 + f4 . Concretely, let’s define
ŷ ′ = a⊤ γ(Bx)
(2.5)
where γ = σˆ2 h2 + σ̂4 h4 and h2 (t) = √12 (t2 − 1) and h4 (t) = √124 (t4 − 6t2 + 3) are the 2nd and 4th
normalized probabilists’ Hermite polynomials [Wik17b]. We abuse the notation slightly by using
4
When σ = ReLU , we have that σ̂0 =
σ̂n = 0.
√1 ,
2π
σ̂1 =
1
.
2
For n ≥ 2 and even, σ̂n =
5
((n−3)!!)2
√
.
2πn!
For n ≥ 2 and odd,
the same notation to denote the its element-wise application on a vector. Now for each example
we use kŷ ′ − yk2 as loss function. The corresponding population risk is
f ′ (a, B) = E kŷ ′ − yk2 .
(2.6)
Now by an extension of Theorem 2.1, we have that the new population risk is equal to the
σ̂22 f2 + σ̂42 f4 .
Theorem 2.2. Let f ′ be defined as in equation (2.6) and f2 and f4 be defined in equation (2.4).
Assume bi , b⋆i ’s are unit vectors. Then, we have
f ′ = σ̂22 f2 + σ̂42 f4 + const
(2.7)
It turns out stochastic gradient descent on the objective f ′ (a, B) (with projection to the set of
matrices B with row norm 1) converges empirically to the ground truth (a⋆ , B ⋆ ) or one of its
equivalent permutations. (See Figure 2.) However, we don’t know of any existing work for
analyzing the landscape of the objective f ′ (or fk for any k ≥ 3). We conjecture that the landscape
of f ′ doesn’t have any spurious local minimum under certain mild assumptions on (a⋆ , B ⋆ ). Despite
recent attempts on other loss functions for tensor decomposition [GM17], we believe that analyzing
f ′ is technically challenging and its resolution will be potentially enlightening for the understanding
landscape of loss function with permutation invariance. See Section 6 for more experimental results.
2.2
Landscape design for orthogonal B ⋆
The population risk defined in equation (2.6) — though works empirically for randomly generated
ground-truth (a⋆ , B ⋆ ) — doesn’t have any theoretical guarantees. It’s also possible that when
(a⋆ , B ⋆ ) are chosen adversarially or from a different distribution, SGD no longer converges to the
ground-truth.
To solve this problem, we design another objective function G(·), such that the optimizer of
G(·) still corresponds to the ground-truth, and G() has provably nice landscape — all local minima
of G() are global minima.
In this subsection, for simplicity, we work with the case when B ⋆ is an orthogonal matrix and
state our main result. The discussion of the general case is deferred to the end of this Section and
Section A.
We define our objective function G(B) as
X
X
G(B) , sign(σ̂4 ) E y ·
φ(bj , bk , x) − µ sign(σ̂4 ) E y ·
ϕ(bj , x)
j,k∈[d],j6=k
+λ
j∈[d]
m
X
i=1
(kbi k2 − 1)2
(2.8)
where ϕ(·, ·) is defined as
ϕ(v, x) =
1
1
1
kvk4 − (v ⊤ x)2 kvk2 + (v ⊤ x)4 .
8
4
24
6
(2.9)
and φ(·, ·, ·) is defined as
1
1
1
kvk2 kwk2 + hv, wi2 − kwk2 (v ⊤ x)2 − kvk2 (w⊤ x)2
2
2
2
1
+ 2(v ⊤ x)(w⊤ x)v ⊤ w + (v ⊤ x)2 (w⊤ x)2 .
2
φ(v, w, x) =
(2.10)
The rationale behind of the choices of φ and ϕ will only be clearer and relevant in later sections.
For now, the only relevant property of them is that both are smooth functions whose derivatives
are easily computable.
We remark that we can sample G(·) using the samples straightforwardly — it’s defined as an
average of functions of examples and the parameters. We also note that only parameter B appears
in the loss function. We will infer the value of a⋆ using straightforward linear regression after we
get the (approximately) accurate value of B ⋆ .
Due to technical reasons, our method only works for the case when a⋆i > 0 for every i. We
will assume this throughout the rest of the paper. The general case is left for future work. Let
a⋆max = max a⋆i , a⋆min = min a⋆i , and κ⋆ = max a⋆i / min a⋆i . Our result will depend on the value of
κ⋆ . Essentially we treat κ⋆ as an absolute constant that doesn’t scale in dimension. The following
theorem characterizes the properties of the landscape of G(·).
Theorem 2.3. Let c be a sufficiently small universal constant (e.g. c = 0.01 suffices) and suppose
the activation function σ satisfies σ̂4 6= 0. Assume µ ≤ c/κ⋆ , λ ≥ c−1 a⋆max , and B ⋆ is an orthogonal
matrix. The function G(·) defined as in equation (2.8) satisfies that
1. A matrix B is a local minimum of G if and only if B can be written as B = DP B ⋆ where
P is a permutation matrix and D is a diagonal matrix with Dii ∈ {±1 ± O(µa⋆max /λ)}.5
Furthermore, this means that all local minima of G are also global.
2. Any saddle point B has a strictly negative curvature in the sense that λmin (∇2 G(B)) ≥ −τ0
where τ0 = c min{µa⋆min /(κ⋆ d), λ}
3. Suppose B is an approximate local minimum in the sense that B satisfies
k∇G(B)k ≤ ε and λmin (∇2 G(B)) ≥ −τ0
Then B can be written as B = P DB ⋆ + EB ⋆ where P is a permutation matrix, D is a
diagonal matrix satisfying the same bound as in bullet 1, and |E|∞ ≤ O(ε/(σ̂4 a⋆min )).
As a direct consequence, B is Od (ε)-close to a global minimum in Euclidean distance, where
Od (·) hides polynomial dependency on d and other parameters.
The theorem above implies that we can learn B ⋆ (up to permutation of rows and sign-flip) if
we take λ to be sufficiently large and optimize G(·) using stochastic gradient descent. In this case,
the diagonal matrix D in bullet 1 is sufficiently close to identity (up to sign flip) and therefore a
local minimum B is close to B ⋆ up to permutation of rows and sign flip. The sign of each b⋆i can
be recovered easily after we recover a (see Lemma 2.5 below.)
Stochastic gradient descent converges to a local minimum [GHJY15] (under the additional
property as established in bullet 2 above), which is also a global minimum for the function G(·).
5
More precisely, |Dii | =
q
1
√
1−µ|σ̂4 |a⋆
/( 6λ)
i
7
We will prove the theorem in Section 5 as a direct corollary of Theorem 5.1. The technical bullet
2 and 3 of the theorem is to ensure that we can use stochastic gradient descent to converge to a
local minimum as stated below.6
Corollary 2.4. In the setting of Theorem 2.3, we can use stochastic gradient descent to optimize
function G(·) (with fresh samples at each iteration) and converge to an approximate global minimum
B that is ε-close to a global minimum in time poly(d, 1/ε).
After approximately recovering the matrix B ⋆ , we can also recover the coefficient a⋆ easily. Note
that fixing B, we can fit a using simply linear regression. For the ease of analysis, we analyze a
slightly different algorithm. The lemma below is proved in Section B.
Lemma 2.5. Given a matrix B whose rows have unit norm, and are δ-close to B ⋆ in Euclidean
distance up to permutation and sign flip with δ ≤ 1/(2κ⋆ ). Then, we can give estimates a, B ′ (using
e.g., Algorithm 1) such that there exists a permutation P where ka − P a⋆ k∞ ≤ δa⋆max and B ′ is
row-wise δ-close to P B ⋆ .
The key step towards analyzing objective function G(B) is the following theorem that gives an
analytic formula for G(·).
Theorem 2.6. The function G(·) satisfies
X
√
a⋆i
G(B) = 2 6|σ̂4 | ·
i∈[d]
+λ
X
|σ̂4 |µ X ⋆ ⋆
hb⋆i , bj i2 hb⋆i , bk i2 − √
ai hbi , bj i4
6
j,k∈[d],j6=k
i,j∈[d]
m
X
i=1
(kbi k2 − 1)2
(2.11)
Theorem 2.6 is proved in Section 4. We will motivate our design choices with a brief overview in
Section 3 and formally analyze the landscape of G in Section 5 (see Theorem 5.1).
Finite sample complexity bounds. : Extending Theorem 2.3, we can characterize the landb which implies that stochastic gradient on G
b also converges approxiscape of the empirical risk G,
mately to the ground-truth parameters with polynomial number of samples.
Theorem 2.7. In the setting of Theorem 2.3, suppose we use N empirical samples to approximate
b There exists a fixed polynomial poly(d, 1/ε) such that if N ≥
G and obtain empirical risk G.
b very similar properties to that of G.
poly(d, 1/ε), then with high probability the landscape of G
b
Precisely, if B is an approximate local minimum in the sense that λmin (∇2 G(B))
≥ −τ0 /2 and
⋆
⋆
b
k∇G(B)k ≤ ε/2, then B can be written as B = DP B + EB where P is a permutation matrix,
D is a diagonal matrix and |E|∞ ≤ O(ε/(σ̂4 a⋆min )).
All of the results above assume that B ⋆ is orthogonal. Since the local minimum are preserved by
linear transformation of the input space, these results can be extended to the general case when B ⋆
is not orthogonal but full rank (with some additional technicality) or the case when the dimension
is larger than the number of neurons (m < d). See Section A for details.
6
In the most general setting, converging to a local minimum of a non-convex function is NP-hard.
8
3
Overview: Landscape Design and Analysis
In this section, we present a general overview of ideas behind the design of objective function G(·).
Inspired by the formula (2.3), in Section 3.1, we envision a family of possible objective functions
for which we have unbiased estimators via samples. In Section 3.2, we pick a specific function that
feeds our needs: a) it has no spurious local minimum; b) the global minimum corresponds to the
ground-truth parameters.
3.1
Which objective can be estimated by samples?
Recall that in equation (2.2) of Theorem 2.1 we give an analytic formula for the straightforward
population risk f . Although the population risk f doesn’t perform well empirically, the lesson that
we learn from it help us design better objective functions. One of the key fact that leads to the
proof of Theorem 2.1 is that for any continuous and bounded function γ, we have that
h
i X
X
⊤
a⋆j hb⋆j , bi ik .
γ̂k σ̂k
E y · γ(bi x) =
k∈N
j∈[d]
Here σ̂k and γ̂k are the k-th Hermite coefficient of the function σ and γ. That is, letting hk the k-th
normalized probabilists’ Hermite polynomials [Wik17b] and h·, ·i be the standard inner product
between functions, we have σ̂k = hhk , σi.
Note that γ can be chosen arbitrarily to extract different terms. For example, by choosing
γ = hk , we obtain that
h
i
X
⊤
a⋆j hb⋆j , bi ik .
(3.1)
E y · hk (bi x) = σ̂k
j∈[d]
That is, we can always access functions forms that involves weighted sum of the powers of hb⋆i , bi i,
as in RHS of equation (3.1).
Using a bit more technical tools in Fourier analysis (see details in Section 4), we claim that
most of the symmetric polynomials over variables hb⋆i , bj i can be estimated by samples:
Claim 3.1 (informal). For an arbitrary polynomial p() over a single variable, there exits a corresponding function φp such that
X X
p
p(hb⋆j , bi i)
(3.2)
a⋆j
E [y · φ (B, x)] =
i
j
Moreover, for an any polynomial q(·, ·) over two variables, there exists corresponding φq such that
X X
q
a⋆j
q(hb⋆j , bi i, hb⋆k , bi i)
(3.3)
E [y · φ (B, x)] =
j
i,k
We will not prove these two general claims. Instead, we only focus on the formulas in Theorem 4.5
and Theorem 4.6, which are two special cases of the claims above.
Motivated by Claim 4.3, in the next subsection, we will pick an objective function which has
no spurious local minimum among those functional forms on the right-hand sides of equation (3.2)
and (3.3).
9
3.2
Which objective has no spurious local minima?
As discussed briefly in the introduction, one of the technical difficulties to design and analyze
objective functions for neural networks comes from the permutation invariance — if a matrix B
is a good solution, then any permutation of the rows of B still gives an equally good solution
(if we also permute the coefficients in a accordingly). We only know of a very limited number
of objective functions that guarantee to enjoy permutation invariance and have no spurious local
minima [GHJY15].
We start by considering the objective function used in [GHJY15],
XX
min P (B) =
hb⋆i , bj i2 hb⋆i , bk i2
i
j6=k
s.t. ∀i ∈ [d], kbi k = 1
(3.4)
Note that here we overload the notation by using b⋆i ’s to denote a set of fixed vectors that we wanted
to recover and using bi ’s to denote the variables. Careful readers may notice that P (B) doesn’t
fall into the family of functions that we described in the previous section (that is, RHS equation
of (3.2) and (3.3)), because it lacks the weighting a⋆i ’s. We will fix this issue later in the subsection.
Before that we first summarize the nice properties of the landscape
⋆ ⊤ of P (B).
(b1 )
..
⋆
For the simplicity of the discussion, let’s assume B = . forms an orthonormal matrix in
(b⋆d )⊤
the rest of the subsection. Then, any permutation and sign-flip of the rows of B ⋆ leads to a global
minimum of P (·) — when B = SQB ⋆ with a permutation matrix Q and a sign matrix S (diagonal
with ±1), we have that P (B) = 0 because one of hb⋆i , bj i2 and hb⋆i , bk i2 has to be zero for all i, j, k 7 ).
It turns out that these permutations/sign-flips of B ⋆ are also the only local minima8 of function
P (·). To see this, notice that P (B) is a degree-2 polynomial of B. Thus if we pick an index s and
fix every row except for bs , then P (B) is a quadratic function over unit vector bs – reduces to an
smallest eigenvector problem. Eigenvector problems are known to have no spurious local minimum.
Thus the corresponding function (w.r.t bs ) has no spurious local minimum. It turns out the same
property still holds when we treat all the rows as variables and add the row-wise norm constraints
(see proof in [GHJY15]).
However, there are two issues with using objective function P (B). The obvious one is that it
doesn’t involve the coefficients a⋆i ’s and thus doesn’t fall into the forms of equation (3.3). Optimistically, we would hope that for nonnegative a⋆i ’s the weighted version of P below would also
enjoy the similar landscape property
X X
hb⋆i , bj i2 hb⋆i , bk i2
P ′ (B) =
a⋆i
i
j6=k
When a⋆i ’s are positive, indeed the global minimum of P ′ are still just all the permutations of the
B ⋆ .9 However, when max a⋆i > 2 min a⋆i , we found that P ′ starts to have spurious local minima .
7
Note that B ⋆ is orthogonal, and j 6= k
We note that since there are constraints here, by local minimum we mean the local minimum on the manifold
defined by the constraints.
9
This is the main reason why we require a⋆ ≥ 0.
8
10
It seems that spurious local minimum often occurs when a row of B is a linear combination of a
smaller number of rows of B ⋆ . See Section D for a concrete example.
To remove such spurious local minima, we add a regularization term below that pushes each
row of B to be close to one of the rows of B ⋆ ,
X X
hb⋆i , bj i4
(3.5)
R(B) = −µ
a⋆i
i
j
We see that for each fixed j, the part in R(B) that involves bj has the form
X
X
−µ
a⋆i hb⋆i , bj i4 = −µh
a⋆i b⋆i ⊗4 , b⊗4
j i
i
(3.6)
i
P
This is commonly used objective function for decomposing tensor i a⋆i b⋆i ⊗4 . It’s known that for
orthogonal b⋆i ’s, the only local minima are ±b⋆1 , . . . , ±b⋆d [GHJY15]. Therefore, intuitively R(B)
pushes each of the bi ’s towards one of the b⋆i ’s. 10 Choosing µ to be small enough, it turns out that
P ′ (B) + R(B) doesn’t have any spurious local minimum as we will show in Section 5.
Another issue with the choice of P ′ (B) + R(B) is that we are still having a constraint minimization problem. Such row-wise norm constraints only make sense when the ground-truth B ⋆ is
orthogonal and thus has unit row norm. A straightforward generalization of P (B) to non-orthogonal
case requires some special constraints that also depend on the covariance matrix B ⋆ B ⋆⊤ , which
in turn requires a specialized procedure to estimate. Instead, we move the constraints into the
objective function by considering adding another regularization term that approximately enforces
the constraints.
It turns out the following regularizer suffices for the orthogonal case,
X
S(B) = λ
(kbi k2 − 1)2 .
(3.7)
i
Moreover, we can extend this easily to the non-orthogonal case (see Section A) without estimating
any statistics of B ⋆ in advance. We note that S(B) is not the Lagrangian multiplier and it does
change the global minima slightly. We will take λ to be large enough so that kbi k has to be close
to 1. As a summary, we finally use the unconstrained objective
min G(B) , P ′ (B) + R(B) + S(B)
Since R(B) and S(B) are degree-4 polynomials of B, the analysis of G(B) is much more delicate,
and we cannot use much linear algebra as we could for P ′ (B). See Section 5 for details.
Finally we note that a feature of this objective G(·) is that it only takes B as variables. We
will estimate the value of a⋆ after we recover the value of B. (see Section B).
4
Analytic Formula for Population Risks
4.1
Basics on Hermite Polynomials
In this section, we briefly review Hermite polynomials and Fourier analysis on Gaussian space.
Let Hj be the probabilists’ Hermite polynomial [Wik17b], and let hj = √1j! Hj be the normalized
10
However, note that R(B) by itself doesn’t work because it does not prevent the solutions where all the bi ’s are
equal to the same b⋆j .
11
Hermite polynomials. The normalized Hermite polynomial forms a complete orthonormal basis in
2
the function space L2 (R, e−x /2 ) in the following sense11 . For two functions f, g that map R to R,
define the inner product hf, gi with respect to the Gaussian measure as
hf, gi =
[f (x)g(x)] .
E
x∼N (0,1)
The polynomials h0 , . . . , hm , . . . are orthogonal to each other under this inner product:
hhi , hj i = δij .
Here δij = 1 if i = j and otherwise δij = 0. Given a function σ ∈ L2 (R, e−x
Hermite coefficient of σ be defined as
2 /2
) , let the k-th
σ̂k = hσ, hk i .
Since h0 , . . . , hm , . . . , forms a complete orthonormal basis, we have the expansion that
X
σ(x) =
σ̂k hk (x) .
k∈N
We will leverage several other nice properties of the Hermite polynomials in our proofs. The
following claim connects the Hermite polynomial to the coefficients of Taylor expansion of a certain
exponential function. It can also serve as a definition of Hermite polynomials.
Claim 4.1 ([O’D14, Equation 11.8]). We have that for t, z ∈ R,
∞
X 1
1
Hk (z)tk .
exp(tz − t2 ) =
2
k!
k=0
The following Claims shows that the expectation E [hn (x)hm (y)] can be computed easily when
x, y are (correlated) Gaussian random variables.
Claim 4.2 ([O’D14, Section 11.2]). Let (x, y) be ρ-correlated standard normal variables (that is,
both x,y have marginal distribution N (0, 1) and E[xy] = ρ). Then,
n
E [hm (x)hn (y)] = ρ δmn .
As a direct corollary, we can compute Ex∼N (0,Idd×d ) σ(u⊤ x)γ(v ⊤ x) by expanding in the Hermite
basis and applying the Claim above.
Claim 4.3. Let σ, γ be two functions from R to R such that σ 2 , γ 2 ∈ L2 (R, e−x
unit vectors u, v ∈ Rd , we have that
h
i X
σ(u⊤ x)γ(v ⊤ x) =
σ̂i γ̂i hu, vii .
E
x∼N (0,Idd×d )
11
2
We denote by L (R, e
−x2 /2
2 /2
). Then, for any
i∈N
2
) the weighted L2 space, namely, L (R, e−x
12
2
/2
n R
o
2
∞
) , f : −∞ f (x)2 e−x /2 dx < ∞
Proof of Claim 4.3. Let s = u⊤ x and t = v ⊤ x. Then s, t are two spherical standard normal random
variables that are hu, vi-correlated, and we have that
h
i
⊤
⊤
σ(u
x)γ(v
x)
= E [σ(s)γ(t)] .
E
x∼N (0,Idd×d )
We expand σ(s) and γ(t) in the Fourier basis and obtain that
X
X
σ̂i hi (s)
γ̂j hj (t)
E [σ(s)γ(t)] = E
=
X
i∈N
j∈N
σ̂i γ̂j E [hi (s)hj (t)]
i,j
=
X
i
4.2
σ̂i γ̂i hu, vii
(by Claim 4.2)
Analytic Formula for population risk f and f ′
In this section we prove Theorem 2.1 and Theorem 2.2, which both follow from the following more
general Theorem.
Theorem 4.4. Let γ, σ ∈ L2 (R, e−x
Define the population risk fγ as
2 /2
), and ŷ = a⊤ γ(Bx) with parameter a ∈ Rℓ and B ∈ Rℓ×d .
fγ (a, B) = E ky − ŷk2 .
⊤
⋆⊤
b1
b1
..
..
⋆
Suppose B = . and B = . and bi ’s and b⋆i ’s have unit ℓ2 norm. Then,
b⊤
b⋆m ⊤
ℓ
f (a, B) =
X
k∈N
σ̂k
X
i∈[m]
a⋆i b⋆i ⊗k − γ̂k
X
2
ai b⊗k
i
i∈[ℓ]
+ const,
F
where σ̂k , γ̂k are the k-th Hermite coefficients of the function σ and γ respectively.
We can see that Theorem 2.1 follows from choosing γ = σ and Theorem 2.2 follows from
choosing γ = σ̂2 h2 + σ̂4 h4 . The key intuition here is that we can decompose σ into a weighted
combination of Hermite polynomials, and each Hermite polynomial influence the population risk
more or less independently (because they are orthogonal polynomials with respect to the Gaussian
measure).
Proof of Theorem 4.4. We have
i
h
2
fγ = E kŷ − yk2 = E a⋆⊤ σ(B ⋆ x) − a⊤ γ(Bx)
13
X
= E
i∈[m]
X
=
i∈[m],j∈[ℓ]
−2
i∈[m],j∈[ℓ]
=
X
a⋆i aj
X
σ̂k
X
σ̂k2 hb⋆i , b⋆j ik +
X
k∈N
i∈[m],j∈[ℓ]
k∈N
4.3
i
⊤
⋆
⋆⊤
x)γ(b
x)
a
a
σ(b
E i j
j
i
k∈N
i,j∈[m]
−2
ai γ(b⊤
i x)
h
a⋆i a⋆j
X
i∈[ℓ]
2
h
i
⋆ ⋆
⋆⊤
⋆⊤
E ai aj σ(bi x)σ(bj x) +
X
X
=
a⋆i σ(b⋆i ⊤ x) −
X
i∈[m],j∈[ℓ]
ai aj
X
k∈N
i,j∈[ℓ]
h
i
⊤
⊤
E ai aj γ(bi x)γ(bj x)
γ̂k2 hbi , bj ik
σ̂k γ̂k hb⋆i , bj ik
a⋆i b⋆i ⊗k
i∈[m]
X
X
− γ̂k
X
(by Claim 4.3)
2
ai b⊗k
i
i∈[ℓ]
.
F
Analytic Formula for population risk G
In this section we show that the population risk G(·) (defined as in equation (2.8)) has the following
analytical formula:
X
X
√
G(B) = 2 6|σˆ4 | ·
hb⋆i , bj i2 hb⋆i , bk i2
a⋆i
i∈[d]
j,k∈[d],j6=k
m
X
|σ̂4 |µ X ⋆ ⋆
(kbi k2 − 1)2 .
ai hbi , bj i4 + λ
− √
6 i,j∈[d]
i=1
The formula will be crucial for the analysis of the landscape of G(·) in Section 5. The formula
follows straightforwardly from the following two theorems and the definition (2.8).
Theorem 4.5. Let φ(·, ·, ·) be defined as in equation (2.10), we have that
X
X
X
√
hb⋆i , bj i2 hb⋆i , bk i2 .
a⋆i
φ(bj , bk , x) = 2 6σ̂ ·
E y ·
i∈[d]
j,k∈[d],j6=k
j,k∈[d],j6=k
Theorem 4.6. Let ϕ(·, ·) be defined as in equation (2.9), then we have that
X
σ̂4 X ⋆ ⋆
ai hbi , bj i4 .
ϕ(bj , x) = √
E y ·
6
i,j∈[d]
j∈[d]
In the rest of the section we prove Theorem 4.5 and 4.6.
We start with a simple but fundamental lemma. Essentially all the result in this section follows
from expanding the two sides of equation (4.1) below.
14
Lemma 4.7. Let u, v ∈ Rd be two fixed vectors and x ∼ N (0, Idd×d ). Then, for any s, t ∈ R,
1
1
2 2
⊤
2 2
⊤
(4.1)
exp(hu, vist) = E exp(u xt − kuk t ) exp(v xs − kvk s ) .
2
2
Proof. Using the fact that E exp(v ⊤ x) = exp( 21 kvk2 ), we have that,
1
1
2 2
⊤
2 2
⊤
E exp(u xt − kuk t ) exp(v xs − kvk s )
2
2
h
i
1
1
= E exp((tu + sv)⊤ x) exp(− kuk2 t2 − kvk2 s2 )
2
2
1
1
1
2
2
2 2
= exp( ktu + svk − kukt − kvk s )
(by the formula E exp(v ⊤ x) = exp( 21 kvk2 ))
2
2
2
= exp(hu, vist) .
Next we extend some of the results in the previous section to the setting with different scaling
(such as when v in Claim 4.3 is no longer a unit vector.)
Lemma 4.8. Let u be a fixed unit vector and v be an arbitrary vector in Rd . Let ϕ(v, x) =
1 ⊤ 2
1
1
4
2
⊤ 4
8 kvk − 4 (v x) kvk + 24 (v x) .
h
i
(4.2)
hu, vi4 δ4,k = E Hk (u⊤ x)ϕ(v, x)
As a sanity check, we can verify that when v is a unit vector, ϕ(v, x) =
Lemma reduces to a special case of Claim 4.2.
√
24h4 (v ⊤ x) and th
Proof.
Let A, B be formal power series in variable
s, t defined as A = exp(hu, vist) and B =
⊤ xt − 1 kuk2 t2 ) exp(v ⊤ xs − 1 kvk2 s2 ) . We refer the readers to [Wik17a] for more backexp(u
E
2
2
grounds of power series. For casual readers, one can just think of A as B as two power series
obtained by expanding the exp(·) via Taylor expansion. For a formal power series A in variable x,
let [xα ]A to denote coefficient in front of the monomial xα . By Lemma 4.7, we have that A = B,
and thus
[s4 tk ]A = [s4 tk ]B ,
(4.3)
which implies that
1
1 2
1
4
4
⊤
2 2
k
⊤
hu, vi δ4,k = E [t ] exp(u xt − t ) · [s ] exp(v xs − kvk s )
24
2
2
1
=E
Hk (u⊤ x)ϕ(v, x) .
k!
(4.4)
where the last line is by the fact that ϕ(v, x) = [s4 ] exp(v ⊤ xs − 12 kvk2 s2 ). This can be verified by
T
applying Claim 4.1 with t = skvk and z = vkvkx , and noting that H4 (x) = x4 − 6x2 + 3.
Now we are ready prove Theorem 4.6 using Lemma 4.8.
15
P
⊤
Proof of Theorem 4.6. Using the fact that σ(v ⊤ x) = ∞
k=0 σ̂k hk (v x), we have that
h
i
X
X
⋆⊤
⋆
σ(b
x)ϕ(b
,
x)
=
a
y
·
ϕ(b
,
x)
E
j
j
i
i E
j
i,j∈[d]
X
=
i,j∈[d]
=
a⋆i
∞
h
i
X
⋆⊤
σ̂
h
(b
x)ϕ(b
,
x)
E k k i
j
k
h
i
σ̂4 X ⋆ ⋆
ai hbi , bj i4
a⋆i E σ̂4 h4 (b⋆i ⊤ x)ϕ(bj , x) = √
6
i,j∈[d]
i,j∈[d]
X
(by Lemma 4.8 and hj =
√1 Hj )
j
Towards proving Theorem 4.5, we start with the following Lemma. Inspired by the proofs above,
we design a function
φ(v, w, x) such that we can estimate hu, vi2 hu, wi2 by taking expectation of
E σ(u⊤ x)φ(v, w, x) .
Lemma 4.9. Let a be a fixed unit vector in Rd and v, w two fixed vectors in Rd . Let ϕ(·, ·) be
defined as in Lemma 4.8. Define φ(v, w, x) as
φ(v, w, x) = ϕ(v + w, x) + ϕ(v − w, x) − 2ϕ(v, x) − 2ϕ(w, x)
1
1
1
= kvk2 kwk2 + hv, wi2 − kwk2 (v ⊤ x)2 − kvk2 (w⊤ x)2
2
2
2
1 ⊤ 2 ⊤ 2
⊤
⊤
⊤
− 2(v x)(w x)v w + (v x) (w x) .
2
(4.5)
(4.6)
Then, we have that
h
i
√
2
2
⊤
E σ(u x)φ(v, w, x) = 2 6σ̂4 hu, vi hu, wi .
Proof. Using the fact that hu, v + wi2 + hu, v − wi4 − 2hu, vi2 − 2hu, wi4 = 12hu, vi2 hu, wi2 and
Lemma 4.8, we have that
h
i
12hu, vi2 hu, wi2 δ4,k = E Hk (u⊤ x)(ϕ(v + w, x) + ϕ(v − w, x) − 2ϕ(v, x) − 2ϕ(w, x))
h
i
(4.7)
= E Hk (u⊤ x)φ(v, w, x) .
Using the fact that σ(u⊤ x) =
h
P∞
k=0 σ̂k hk (u
⊤ x),
we conclude that
∞
i X
h
i
σ̂k E hk (u⊤ x)φ(v, w, x)
E σ(u x)φ(v, w, x) =
⊤
k=0
i
σ̂4 h
= √ E H4 (u⊤ x)φ(v, w, x)
6
√
= 2 6σ̂4 hu, vi2 hu, wi2
16
(by Lemma 4.8 and hj =
√1 Hj )
j
(by Lemma 4.8 again)
Now we are ready to prove Theorem 4.5 by using Lemma 4.9 for every summand.
Proof of Theorem 4.5. We have that
i
X
X X h
⋆⊤
φ(bj , bk , x) =
a⋆i
E y ·
E σ(bi x)φ(bj , bk , x)
i
j,k
j,k
X X
√
hb⋆i , bj i2 hb⋆i , bk i2 .
a⋆i
= 2 6σ̂4
i
5
(by Lemma 4.9)
j,k
Landscape of Population Risk G(·)
In this section we prove Theorem 2.3. Since the landscape property is invariant with respect to
rotations of parameters, without loss of generality we assume B ⋆ is the identity matrix Id throughout
this section. (See Section A for a precise statement for the invariance.) Recall that by Theorem 2.6,
the population risk G(·) in the case of B ⋆ = Id is equal to
X X
√
2
2 ⊤
G(B) = 2 6|σ̂4 |
a⋆i
(b⊤
j ei ) (bk ei )
i
|σ̂4 |µ
− √
6
d
X
i=1
a⋆i
j6=k
d
X
4
(b⊤
j ei )
+λ
d
X
j=1
j=1
2
kbj k2 − 1 .
(5.1)
In the rest of section we work with the formula above for G(·) instead of the original definition.
In fact, for future reference, we study a more general version of the function G. For nonnegative
vectors α, β and nonnegative number µ, let Gα,β,µ be defined as
Gα,β,µ (B) =
d
X
i=1
αi
X
j6=k
2 ⊤
2
(b⊤
j ei ) (bk ei )
−µ
d
X
i=1
βi
d
X
j=1
4
(b⊤
j ei )
d
X
(kbj k2 − 1)2
+λ
(5.2)
j=1
Here ei denotes the i-th natural basis vector. We see that G is sub-case of Gα,β,µ and we prove the
following extension of Theorem 2.3. Let αmax = maxi αi and αmin = mini αi .
Theorem 5.1. Let κα = αmax /αmin and c be a sufficiently small universal constant (e.g. c = 10−2
suffices). Suppose µ ≤ cαmin /βmax and λ ≥ 4 max(µβmax , αmax ). Then, the function Gα,β,µ (B)
defined as in equation (5.2) satisfies that
1. A matrix B is a local minimum of Gα,β,µ if and only if B can be written
= DP where
n q as B o
1
P is a permutation matrix and D is a diagonal matrix with Dii ∈ ± 1−µβ
.
i /λ
2. Any saddle point B has strictly negative curvature in the sense that λmin (∇2 Gα,β,µ (B)) ≤ −τ0
2 /β
where τ0 = c min{µβmin /(κα d), µβmin
max , λ}
3. Suppose B is an approximate local minimum in the sense that B satisfies
k∇g(B)k ≤ ε and λmin (∇2 g(B)) ≥ −τ0
17
Then B can be written as B = DP + E where P is a permutation matrix, D is a diagonal
matrix with the entries satisfying
ε
1
ε
1
18dε2
2
−
≤
D
≤
1
+
1
−
ii
2
i
i
2λ
2λ
βmin
1 − µβ
1 − µβ
λ
λ
and E is an error matrix satisfying
|E|∞ ≤ 3ε/βmin .
As a direct consequence, B is Od (ε)-close to a global minimum in Euclidean distance, where
Od (·) hides polynomial dependency on d and other parameters.
Here we recall that |E|∞ denotes the largest √
entries in the matrix E. Theorem
2.3 follows straight√
forwardly from Theorem 5.1 by setting α = 2 6|σ̂4 |a⋆ and β = |σ̂4 |a⋆ / 6. In the rest of the section
we prove Theorem 5.1.
Note that our
variable
B is a matrix of dimension d × d and we use bi to denote the rows of
⊤
b1
B, that is, B = ... . Naturally, towards analyzing the properties of a local minimum B, the first
b⊤
d
step is that we pick a row bs of B and treat only bs as variables and others rows as fixed. We
will show that local optimality of bs will imply that bs is equal to one of the basis vector ej up to
some scaling factor. This step is done in Section 5.1. Then in Section 5.2 we show that the local
optimality of all the variables in B implies that each of the rows of B corresponds to different basis
vector, which implies that B is a permutation matrix (up to scaling of the rows).
5.1
Step 1: Analysis of Local Optimality of a Single Row
Suppose we fix b1 , · · · , bs−1 , bs+1 , · · · , bd , and optimize only over bs , we obtain the objective h of
the following form:
hα,β,λ (x) =
d
X
i=1
αi x2i −
d
X
i=1
2
βi x4i + λ kxk2 − 1
(5.3)
P
2
⋆
We can see that setting αi = a⋆i k6=s (b⊤
k ei ) , βi = ai , and x = bs gives us the original objective
G(B). In this subsection, we will work with h(·) and analyze the properties of the local minima of
h(·).
The following lemma shows that a local minimum x of the objective h(·) must be a scaling of
a basis vector. Recall that |x|2nd denotes the second largest absolute value of the entries of x. The
lemma deals generally an approximate local minimum, though we suggest casual readers simply
think of ε, τ = 0 in the lemma.
d
Lemma 5.2. Let h(·) be defined in equation
p (5.3) with non-negative vectors α and β in R . Suppose parameters ε, τ ≥ 0 satisfy that ε ≤ τ 3 /βmin . If some point x satisfies k∇h(x)k ≤ ε and
λmin (∇2 h(x)) ≥ −τ , then we have
r
τ
|x|2nd ≤
.
βmin
18
Proof. Without loss of generality, we can take ε =
gradient and Hessian of function h(·) are
p
1/3
τ 3 /βmin which means τ = ε2/3 βmin . The
∇h(x) = 2 diag(α)x − 4 diag(β)x⊙3 + γx
∇2 h(x) = 2 diag(α) − 12 diag(β ⊙ x⊙2 ) + γId + 8λxx⊤ .
(5.4)
where γ , 4λ(kxk2 − 1).
Let S = {i : |xi | ≥ δ} be the indices of the coordinates that are significantly away from zero,
1/3
ε
. Since k∇h(x)k ≤ ε, we have that |∇h(x)i | ≤ ε for every i ∈ [d], which implies
where δ = βmin
that
∀i ∈ [d], 2αi xi + γxi − 4βi x3i ≤ ε
(5.5)
which further implies that
∀i ∈ S, 2αi + γ − 4βi x2i ≤
ε
δ
(5.6)
If |S| = 1, then we are done because |x|2nd ≤ δ. Next we prove that |S| ≥ 2. For the sake of
contradiction, we assume that |S| ≥ 2. Moreover, WLOG, we assume that |x|1 ≥ |x|2 are the two
largest entries of |x| in absolute values. p
p
We take v ∈ Rd such that v1 = −x2 / x21 + x22 , and v2 = x1 / x21 + x22 , and vj = 0 for j ≥ 2.
Then we have that v ⊤ x = 0 and kvk = 1. We evaluate the quadratic form and have that
v ⊤ ∇2 h(x)v = v ⊤ (2 diag(α) + γId )v − 12v ⊤ diag(β ⊙ x⊙2 )v
(since v ⊤ x = 0)
= (2α1 + γ)v12 + (2α2 + γ)v22 − 12β1 v12 x21 − 12β2 v22 x22
ε
(by equation (5.6) and kvk = 1)
≤ −8β1 v12 x21 − 8β2 v22 x22 +
δ
ε
x2 x2
≤ −8(β1 + β2 ) 2 2 1 2 +
x1 + x2 δ
ε
≤ −8βmin x22 + .
δ
Recall that δ =
ε
βmin
1/3
. Then we conclude that
1/3
v ⊤ ∇2 h(x)v ≤ −6βmin ε2/3 = −6τ.
1/3
This contradicts with the assumption that λmin (∇2 h(x)) ≥ −βmin ε2/3 = τ and that kvk = 1.
Therefore we have |S| = 1 and
1/3 r
p
ε
τ
≤
( using ε ≤ τ 3 /βmin )
|x|2nd ≤ δ =
βmin
βmin
For future reference, we can also show that for a sufficiently strong regularization term (sufficiently
large λ), the norm of a local minimum x should be bounded from below and above by 1/2 and 2.
This are rather coarse bounds that suffice for our purpose in this subsection. In Section 5.2 we will
show that all the rows of a local minimum B of G have norm close to 1.
19
Lemma 5.3. In the setting of Lemma 5.2,
1. Suppose in addition that λ ≥ 4 max(βmax , τ ) and ε ≤ 0.1βmin d−3/2 , then
kxk2 ≤ 2 .
2. Let i⋆ = arg maxi |xi |. In addition to the previous conditions in bullet 1, assume that λ ≥ 4αi⋆ .
Then,
1
kxk2 ≥ .
2
We remark that we have to state the conditions for the upperbounds and lowerbounds separately
since they will be used with these different conditions.
Proof. Let S = {i : |xi | ≥ δ} be the indices of the coordinates that are significantly away from zero,
1/3
ε
where δ = βmin
. We first show that kxk2 ≤ 2. We divide into two cases:
1 S is empty. Since ε ≤ 0.1βmin d−3/2 , then δ ≤
√
√2 .
d
We conclude that kxk2 ≤ 2.
2 S is non-empty. For i ∈ S, recall equation (5.6) which implies that
ε
+ 4βi x2i
δ
ε
≤ + 4βmax kxk2
δ
ε
βmax
2
kxk ≤
+
kxk2 + 1
4λδ
λ
4λ(kxk2 − 1) ≤
1/3
Since λ ≥ 4βmax , so λ ≥ βmin ε2/3 ≥
kxk2 ≤ 2.
3ε
4δ ,
and thus from the display above we have that
Next we show that kxk2 ≥ 21 . Again we divide into two cases:
1. S is empty. For the sake of contradiction, assume that kxk2 ≤ 12 , then γ ≤ −2λ. We show
that there is sufficient negative curvature. Recall that
∇2 h(x) = 2 diag(α) − 12 diag(β ⊙ x⊙2 ) + γI + 8λxx⊤
2 diag(α) − 12 diag(β ⊙ x⊙2 ) − 2λI + 8λxx⊤
Choose index j ⋆ so that αj ⋆ = αmin , then
2
2
2
e⊤
j ⋆ ∇ h(x)ej ⋆ = 2αmin − 12βj ⋆ xj ⋆ − 2λ + 8λxj ⋆
≤ 2αmin + 8λδ2 − 2l
≤ 2αmin − λ(2 − 8δ2 )
4
≤ 2αmin − λ
3
5
≤ − λ ≤ −3τ
6
(by δ2 ≤
1
12 )
( by λ ≥ 4 max{αmin , τ })
This contradicts with the fact that λmin (∇2 h(x)) ≥ −τ . Thus when S is empty, kxk2 ≥ 21 .
20
2. S is non-empty. Recall that i⋆ = arg maxi |xi |, and by definition i⋆ ∈ S. Using Equation (5.6)
γ ≥ −2αi⋆ −
ε
δ
which implies that
kxk2 ≥ 1 −
ε
αi⋆
−
.
λ
4λδ
1/3
Since λ ≥ 4αi⋆ , and λ ≥ βmin ε2/3 ≥ δε , we conclude that kxk2 ≥ 1/2.
We have shown that a local minimum x of h should be a scaling of the basis vector ei⋆ . The
following lemma strengthens the result by demonstrating that not all basis vector can be a local
minimum — the corresponding coefficient αi⋆ has to be reasonably small for ei⋆ being a local
minimum. The key intuition here is that if αi⋆ is very large compared to other entries of α, then if
we move locally the mass of ei⋆ from entry i⋆ to some other index j, the objective function will be
likely to decrease because αj x2j is likely to be smaller than αi⋆ x2i⋆ . (Indeed, we will show that such
movement will cause a second-order decrease of the objective function in the proof.)
Lemma 5.4. In the setting of Lemma 5.2, let p
i⋆ = arg maxi |xi |.
2
λmin (∇ h(x)) > −τ for 0 ≤ τ ≤ 0.1βmin /d and ε ≤ τ 3 /βmin , then
If k∇h(x)k ≤ ε, and
αi⋆ ≤ αmin + 2ε + 2τ + 4βi⋆ .
Proof. For the ease of notation, assume WLOG that i⋆ = 1. Let δ = (τ /βmin )1/2 . By the assumptions, we have that δ ≤ √16d . By Lemma 5.2, we have kxk2 ≥ 12 , which implies that
x21 ≥ kxk2 − (d − 1) |x|22nd ≥
Define v = −
1+
x2k
x21
xk
x1
1
1
− d |x|22nd ≥ 1 − dδ2 ≥ .
2
3
(5.7)
e1 + ek . Since x1 is the largest entry of x, we can verify that 1 ≤ kvk2 =
≤ 2. By the assumption, we have that
v ⊤ ∇2 h(x)v ≥ −τ kvk2 ≥ −4τ .
(5.8)
On the other hand, recall the form of Hessian (equation (5.4)), by straightforward algebraic
manipulation, we have that
v ⊤ ∇2 h(x)v = v ⊤ 2 diag(α) − 12 diag(β ⊙ x⊙2 ) + γId + 8λxx⊤ . v
xk
xk
(by v ⊤ x = 0)
= 2α1 ( )2 + 2αk − 12β1 x2k − 12βk x2k + γ( )2 + γ
x1
x1
xk
ε
≤ (2α1 + γ)( )2 − 12(β1 + βk )x2k + 2αk + (4β1 x21 − 2α1 +
) (by equation (5.6))
x1
|x1 |
21
≤ (4β1 x21 +
ε
xk
ε
)( )2 − 12(β1 + βk )x2k + 2αk + (4β1 x21 − 2α1 +
)
|x1 | x1
|x1 |
(by equation (5.6))
= −8β1 x2k − 12βk x2k + 4β1 x21 + 2αk − 2α1 + 4ε
≤ 2αk − 2α1 + 4ε + 8β1 .
(by |xk | ≤ |x1 | and |x1 |2 ≥ 1/3)
(by kxk ≤ 2 using Lemma 5.2)
Combining equation (5.8) and the equation above gives
α1 ≤ αk + 2ε + 2τ + 4β1 .
Since k is arbitrary we complete the proof.
The previous lemma implies that it’s very likely that the local minimum x can be written as
x = xi⋆ ei⋆ and the index i⋆ is also likely to be the argmin of α. The following technical lemma
shows that when this indeed happens, then we can strengthen Lemma 5.2 in terms of the error
bound’s dependency on ε and τ . In Lemma 5.2, we have that |x|2nd is bounded by a function of τ .
Here we strengthen the bound to be a function that only depends on ε. Thus as long as τ be small
enough so that we can apply Lemma 5.2 and Lemma 5.4 to meet the condition of the lemma below,
then we get an error bound that goes to zero as ε goes to zero. This translates to the error bound
in bullet 3 of Theorem 5.1 where the bound on E only depends on ε. For casual readers we suggest
to skip this Lemma since its precise functionality will only be clearer in the proof of Theorem 5.1.
Lemma 5.5. In the setting of Lemma 5.2, in addition we assume that i = argmink |αk | and that x
can be written as x = xi ei + x−i satisfying
√ p
kx−i k∞ ≤ 0.1 min{1/ d, βmin /(βmax )} .
Then, we can strengthen the bound to
kx−i k∞ ≤
3ε
.
βmin
Proof. WLOG, let i = 1. Let xj be the second largest entry of x in absolute value. Define
v1 = 4β1 x21 − 2α1 − γ, and similarly vj = 4βj x2j − 2αj − γ. Since k∇h(x)k ≤ ε, by equation (5.6), we
have that |v1 | ≤ |xε1 | and |v2 | = |xεj | . Subtracting 4β1 x21 = 2α1 + γ + v1 and 4βj x2j = 2αj + γ + vj ,
we obtain,
4β1 x21 = 4βj x2j − 2(αj − α1 ) + v1 − vj
≤ 4βj x2j + (v1 − vj )
Since kxk2 ≥ 12 , then x21 ≥
1
2
− dδ2 ≥ 31 . Since |xj | ≤ δ,
4βj x2j ≤ 4βmax δ2
Combining the above two displays,
(v1 − vj ) ≥ 4β1 x21 − 4βj δ2
22
(since αj − α1 ≥ 0)
≥ 4β1 x21 − 4βmax δ2
2
4
≥ β1 − βmin
3
3
2
≥ βmin
3
Since |v1 | ≤
ε
|x1 |
and |v2 | =
( using x21 ≥
1
3
and δ ≤
p
βmin /(6βmax ))
(5.9)
ε
|x2 | ,
2
ε
2
≥ βmin ,
|xj |
3
(5.10)
ε
.
and re-arranging gives |xj | ≤ 3 βmin
5.2
Local Optimality of All the Variables
In this section we prove Theorem 5.1. Results in Subsection 5.1 have established that if B is a local
minimum, then each row bs of B has to be a scaling of a basis vector. In this section we show that
these basis vectors need to be distinct from each other. The following proposition summaries such
a claim (with a weak error analysis).
Proposition 5.6. In the setting of Theorem 5.1, suppose B satisfies
k∇g(B)k ≤ ε and λmin (∇2 g(B)) ≥ −τ
for parameters τ, ε satisfying 0 ≤ τ ≤ c min{µβmin /(κα d), λ} and ε ≤ c min{αmin ,
Then, the matrix B can be written as
p
τ 3 /βmin }.
B = DP + E ,
where D is diagonal such that ∀i, |Dii | ∈ [1/4, 2], and P is a permutation matrix, and |E|∞ ≤ δ
1/2
.
with δ = µβτmin
As alluded before, in the proof we will first apply the results in Section 5.1 to show that when B
is a local minimum, each row bs has a unique large entry. Then we will show that the largest entries
of each row sit on different columns. The key intuition behind the proof is that if two rows, say
row s, t, have their large entries on the same column, then it means that there exists a column—
say column k — that doesn’t contain largest entry of any row. Then either row s or t will violate
Lemma 5.4. Or in other words, either row s or t can move their mass into the column k to decrease
the function value. This contradicts the assumption that B is a local minimum.
Proof. As pointed in the paragraph below equation (5.3), when we restrict our attention to a
particular row of B and fix the rest of the rows the function Gα,β,µ reduces to the function h(·) in
equation (5.3) so that we can apply lemmas in Section 5.1.
P
2
Concretely, fix an index s ∈ [d] and let x = bs . For all i ∈ [d], let ᾱi = αi j6=s (b⊤
j ei ) , and
β̄i = µβi . Then we have that
Gα,β,µ (B) =
d
X
i=1
ᾱi x2i −
X
23
i
2
β̄i x4i + λ kxk2 − 1
(5.11)
We view the function above as h(x). Now we apply Lemma 5.2 (by replacing α, β in Lemma 5.2
2
by ᾱ, β̄). The assumption that λmin (∇2 gα,β,µ (B)) ≥ −τ implies that λmin
p (∇ h(x)) ≥ −τ since
2
2
3
∇ h(x) is a submatrix of ∇ g(B). Moreover, k∇h(x)k ≤ k∇G(B)k ≤ ε ≤ τ /(µβmin )
Hence by Lemma 5.2, we have that the second largest entry of |bs | satisfies
∀s, |bs |2nd ≤ δ.
(5.12)
1/2
where δ , µβτmin
for the ease of notation. We can check that δ ≤ 4√1κ d by the assumption.
α
Therefore, we have essentially shown that each row of B has only one single large entry, since the
second largest entry is at most δ.
Next we show that each row of B has largest entries on distinct columns. For each row j ∈ [d],
let ij = arg maxi |e⊤
i bj | be the index of the largest entry of bj . We will show that i1 , . . . , id are
distinct.
For the sake of contradiction, suppose they are not distinct, that is, there are two distinct rows
s, t that have the same largest entries on column l, that is, we assume that is = it = l. This implies
that {i1 , . . . , id } =
6 [d] and let k ∈ [d] be the index such that k ∈
/ {i1 , . . . , id }. We note that by the
1/2
1
≤ 4√1κ d ≤ 4√
assumption δ = µβτmin
. We first bound from above ᾱk
d
α
ᾱk = αk
X
j6=s
2
2
(b⊤
j ek ) ≤ αk dδ ≤
1
αmin .
16
⊤
Assume in addition without loss of generality that |b⊤
s el | ≤ |bt el |. Let
X
2
(b⊤
zl ,
j el )
(by δ ≤
√1
)
4 κα d
(5.13)
j6=s
be the sum of squares of the entries on the column l without entry b⊤
j el , and that ᾱl = αl zl . We
first prove that zl ≥ 1/3.
For the sake of contradiction, assume zl < 1/3. Then we have that ᾱl = αl z ≤ 31 αl . This implies
that λ ≥ 4 max{ᾱl , τ }, and since l is the index of the largest column of bs we can invoke Lemma 5.3
and conclude that kbs k2 ≥ 1/2. This further implies that
√
2
2
2
(b⊤
(by δ ≤ 1/(4 d))
s el ) ≥ kbs k − d |bs |2nd ≥ 1/2 − dδ ≥ 1/3
⊤
Since we have assumed that |b⊤
s el | ≤ |bt el |. Then we obtain that
2
⊤ 2
zl ≥ |b⊤
t el | ≥ |bs el | ≥ 1/3 ,
which contradicts the assumption. Therefore, we conclude that zl ≥ 1/3. Then we are ready to
bound ᾱl from below:
1
ᾱl = αl zl ≥ αl .
3
The display above and Equation (5.13) implies that
ᾱl − ᾱk ≥
24
1
αmin .
4
(5.14)
Note that l is the largest entry in absolute value in the vector bs . We will apply Lemma 5.4.
We fix every
row of B except bs and consider the objective as a function of bs only. Again let
P
2
ᾱi = αi j6=s (b⊤
j ei ) , and β̄i = µβi and we have the equation (5.11). (Note that now ᾱ depends on
the choice of s which we fixed.) Lemma 5.4 gives us that
ᾱl ≤ ᾱk + 2ε + 2τ + 4β¯ℓ .
Since ε ≤
1
50 αmin ,
τ≤
1
50 αmin
and β̄l = µβl ≤
1
50 αmin ,
we obtain that
1
ᾱl ≤ ᾱk + αmin
5
(5.15)
which contradicts equation (5.14). Thus we have established that i1 , . . . , id are distinct.
Finally, let Q be the matrix that only contain the largest entries (in absolute value) of each
columns of B. Since i1 , . . . , id are distinct, we have that Q contains exactly one entry per row
and per column. Therefore Q can be written as DP where P is a permutation matrix and D is a
diagonal matrix. Moreover, we have that kbs k2∞ ≥ kbs k2 − d |bs |22nd ≥ 1/4 and kbs k2 ≤ 2. Therefore,
the largest entry of each row has absolute value between 1/4 and 2. Therefore |D|ii ∈ [1/4, 2]. Let
E = B − P D. Then we have that |E|∞ ≤ maxs |bs |2nd ≤ δ,which completes the proof.
Applying Lemma 5.5, we can further strengthen Proposition 5.6 with better error bounds and
better control of the largest entries of each column.
Proposition 5.7 (Strengthen of Proposition 5.6). In the setting of Proposition 5.6. Suppose in
2 /β
addition that τ satisfies τ ≤ cµβmin
max . Then, the matrix B can be written as
B = DP + E ,
where P is a permutation matrix, D is diagonal such that
1
ε
ε
1
18dε2
2
≤
|D
|
≤
1
+
−
1
−
∀i ∈ [d],
ii
2
i
i
2λ
2λ
βmin
1 − µβ
1 − µβ
λ
λ
and
|E|∞ ≤
3ε
.
βmin
1/2
. Now we use Lemma 5.5 to
Proof. By Proposition 5.6, we know that |E|∞ ≤ δ = µβτmin
strength the error bound.
As we have done in the proof of Proposition 5.6, we again fix an arbitrary sP
∈ [d] and all the
2
rows except bs and view Gα,β,µ as a function of bs . For all i ∈ [d], let ᾱi = αi j6=s (b⊤
j ei ) , and
β̄i = µβi and view Gα,β,µ as a function of the form h(x) with α, β replaced by ᾱ, β̄, namely,
h(x) =
X
k
ᾱk x2k −
X
k
2
β̄k x4k + λ kxk2 − 1 + const
25
We will verify the condition of Lemma 5.5. Let i be the index of the largest entry in absolute
value of the vector bs . Since we have shown that the largest entry in each row sits on different
columns, and the second largest entry is always less than δ, we have that,
ᾱi = αi
X
1
2
2
αmin .
(b⊤
j ei ) ≤ αi dδ ≤
16
(by δ ≤
j6=s
√1
)
4 κα d
For any k 6= i, we know that the column k contains some entry (k, jk ) which is the largest entry of
some row, and we also have that jk 6= s since the largest entry of row s is on column i. Therefore,
we have that
X
2
⊤
2
2
2
ᾱk = αk
(b⊤
j ek ) ≥ αl (bjk ek ) ≥ αl (kbk k − dδ )
j6=s
√
(by δ ≤ 1/(4 d))
1
≥ αl
3
Therefore, α¯k ≥ ᾱi for any k 6=√i and
p thus i = argmink |ᾱk |. By the fact that |E|∞ ≤ δ, we have
that kx−i k∞ ≤ δ ≤ 0.1 min{1/ d, βmin /(βmax )}. Now we are ready to apply Lemma 5.5 and
. Applying the argument for every row s gives |E|∞ ≤ β3ε
.
obtain that |bs |2nd ≤ β3ε
min
min
Finally, we give the bound for the entires in D. Let v be a short hand for ∇h(bs ) which is equal
to the s-th column of ∇G(B). Since B is an ε-approximate stationary point, then we have that
kvk ≤ ε and by straightforward calculation of the gradient, we have
vi = 2ᾱi xi −
4µβi x3i
d
X
x2j − 1)xi .
+ 4λ(
j=1
Since xi 6= 0, dividing by xi gives,
d
X
vi
x2j − 1) −
0 = 2ᾱi −
+ 4λ(
xi
j=1
X
vi
= (4λ − 4µβi )x2i + 4λ
x2j + 2ᾱi − 4λ −
xi
4µβi x2i
j6=i
Rearranging the equation above gives,
X
1
vi
4λ − 2ᾱi − 4λ
x2j −
4λ − 4µβi
xi
j6=i
X
ᾱ
1
v
i
1 − i −
=
x2j −
µβi
2λ
4λx
i
1− λ
j6=i
x2i =
To upper bound x2i , we note that |vi | < ε, ᾱi > 0, and
P
2
j6=i xj
2µβi + ε
1
ε
1+
≤1+
x2i ≤
µβi
2λ
λ
1− λ
26
> 0, so
(since λ ≥ 4µβi )
0.5
0.6
0.5
Parameter Error
Objective Value
0.4
0.3
0.2
0.1
0.4
0.3
0.2
0.1
0
0
1
2
3
4
Iterations
5
×10
0
4
1
2
3
Iterations
4
5
×10
4
Figure 1: Data are generated by a network with ReLU activation without noise. The training
model uses the same architecture. Left: the estimated population risk doesn’t converge to zero.
Right: the parameter error using the surrogate in equation (6.1).
P
For the lower bound of x2i , we note that |E|∞ ≤ δ = β3ε
implies j6=i x2j ≤ dδ2 . Moreover, we
min
have proved that each rows has largest entry at differentPcolumns. Also note that the largest entry
of row bs is on column i. Therefore, we have ᾱi = αi j6=s (bTj ei )2 ≤ αmax dδ2 . Using these two
, we have
estimates and δ = β3ε
min
x2i ≥
=
1
1−
µβi
λ
1−
µβi
λ
1
ε
αmax
+ 1)dδ2 −
)
2λ
2λ
18dε2
ε
1− 2 −
2λ
βmin
(1 − (
Finally we are ready to prove Theorem 5.1 by applying Proposition 5.6.
Proof of Theorem 5.1. By setting ε = 0, τ = 0 in Proposition 5.6, we have that any local minimum
B satisfies that B = DP where P is a permutation matrix and D is a diagonal and the precise
diagonal entries of D. It can be verified that all these points have the same function value, so that
they are all global minimizers.
Towards proving the second bullet, we note that a saddle point B satisfies that ∇G(B) = 0. We
will prove that λmin (∇2 G(B)) ≤ −τ0 . For the sake of contradiction, suppose λmin (∇n2 G(B))
q ≥ −τo0 .
1
,
Then setting ε = 0 and τ = τ0 in Propostion 5.7, we have that B = DP and Dii = ± 1−µβ
i /λ
which by bullet 1 implies that B is a local minimum. This contradicts the assumption that B is a
saddle point.
The 3rd bullet is a just a rephrasing of Proposition 5.7.
6
Simulation
In this section, we provide simple simulation results that verify that minimizing G(B) with SGD
recovers a permutation of B ⋆ ; however, minimizing Equation (2.2) with SGD results in finding
27
0.8
0.06
0.6
Parameter Error
Objective Value
0.08
0.04
0.02
0
0.4
0.2
0
0.5
1
1.5
2
2.5
Iterations
3
×10
0
1
5
2
3
Iterations
×10
5
Figure 2: The labels are generated from a network with ReLU activation. We learn with σ̂2 h2 + σ̂4 h4
activation. Left: the test loss subtracted by the theoretical global minimum value. Right: the error
in parameter space measured by equation (6.1)
spurious local minima. Based on the formula for the population risk in Equation (2.3), we also
verified empirically the conjecture that SGD would successfully recover B ⋆ using the activation
functions γ(z) = σ̂2 h2 (z) + σ̂4 h4 (z),12 even if the data were generated via a model with ReLU
activation. (See Section 2.1 for the rationale behind such conjectures.)
For all of our experiments, we chose B ⋆ = Idd×d with dimension d = 50 and a⋆ = 1 for simplicity,
and the data is generated from a one-hidden-layer network with ReLU activation without noise.
We use stochastic gradient descent with fresh samples at each iteration, and we plot the (expected)
population error (that is, the error on a fresh batch of examples).
1
0.2
0.8
Parameter Error
Objective Value
0.15
0.1
0.05
0
0.6
0.4
0.2
-0.05
-0.1
0
0
2
4
6
Iterations
8
10
0
4
2
4
Iterations
6
8
10
4
Figure 3: Learning with objective function G(·). Left: the test loss. Right: the error in parameter
space measured by equation (6.1).
To test whether SGD converges to a matrix B which is equivalent to B ⋆ up to permutation of
rows, we use a surrogate error metric to evaluate whether B ⋆ −1 B is close to a permutation matrix.
12
We also observed that using γ(z) =
mental results here.
1
|z|
2
also works but due to the space limitation we don’t report the experi-
28
Given a matrix Q with row norm 1, let
e(Q) = min{1 − min max |Qij |, 1 − min max |Qij |}.
i
j
j
i
(6.1)
√
Then we have that if e(Q) ≤ ε for some ε < 1/3, then it implies that Q is 2ε-close to a permutation
matrix in infinity norm. On the other direction, we know that if e(Q) > ε, then Q is not ε-close
to any permutation matrix in infinity norm. The latter statement also holds when Q doesn’t have
row norm 1.
Figure 1 shows that without over-parameterization, using ReLU as an activation function, SGD
doesn’t converge to zero test error and the ground-truth parameters. We decreased step-size by a
factor of 4 every 5000 number of iterations after the error plateaus at 10000 iterations. For the final
5000 iterations, the step-size is less than 10−9 , so we can be confident that the non-zero objective
value is not due to the variance of SGD. We see that none of the five runs of SGD converged to a
global minimum.
Figure 2 shows that using σ̂2 h2 + σ̂4 h4 as the activation function, SGD with projection to the
set of matrices B with row norm 1 converges to the ground-truth parameters. We also plot the
loss function which converges the value of a global minimum. (We subtracted the constant term in
equation (2.7) so that the global minimum has loss 0.)
Figure 3 shows that using our objective function G(B), the iterate converges to the ground
truth matrix B ⋆ . The fact that the parameter error goes up and down is not surprising, because
the algorithm first gets close to a saddle point and then breaks ties and converges to a one of the
global minima.
Finally we note that using the loss function G(·) seems to require significantly larger batch (and
sample complexity) to reduce the variance in the gradients estimation. We used batch size 262144
in the experiment for G(·). However, in contrast, for the σ̂2 h2 + σˆ4 h4 we used batch size 8192 and
for relu we used batch size 256.
7
Conclusion
In this paper we first give an analytic formula for the population risk of the standard ℓ2 loss, which
empirically may converge to a spurious local minimum. We then design a novel population loss
that is guaranteed to have no spurious local minimum.
Designing objective functions with well-behaved landscape is an intriguing and fruitful direction. We hope that our techniques can be useful for characterizing and designing the optimization
landscape for other settings.
We conjecture that the objective αf2 + βf4 has no spurious local minimum when α, β are
reasonable constants and the ground-truth parameters are in general position13 . We provided
empirical evidence to support the conjecture.
Our results assume that the input distribution is Gaussian. Extending them to other input
distributions is a very interesting open problem.
13
See equation (2.4) for the definition of fk and Theorem 2.2 for how to access αf2 + βf4 in the setting of onehidden-layer neural nets.
29
References
[ABGM14] Sanjeev Arora, Aditya Bhaskara, Rong Ge, and Tengyu Ma. Provable bounds for
learning some deep representations. In International Conference on Machine Learning,
pages 584–592, 2014.
[BBV16]
Afonso S Bandeira, Nicolas Boumal, and Vladislav Voroninski. On the low-rank approach for semidefinite programs arising in synchronization and community detection.
arXiv preprint arXiv:1602.04426, 2016.
[BG17]
Alon Brutzkus and Amir Globerson. Globally optimal gradient descent for a convnet
with gaussian inputs. arXiv preprint arXiv:1702.07966, 2017.
[BH89]
Pierre Baldi and Kurt Hornik. Neural networks and principal component analysis:
Learning from examples without local minima. Neural networks, 2(1):53–58, 1989.
[BNS16]
Srinadh Bhojanapalli, Behnam Neyshabur, and Nati Srebro. Global optimality of local
search for low rank matrix recovery. In Advances in Neural Information Processing
Systems, pages 3873–3881, 2016.
[DFS16]
Amit Daniely, Roy Frostig, and Yoram Singer. Toward deeper understanding of neural
networks: The power of initialization and a dual view on expressivity. In Advances In
Neural Information Processing Systems, pages 2253–2261, 2016.
[GHJY15] Rong Ge, Furong Huang, Chi Jin, and Yang Yuan. Escaping from saddle pointsonline
stochastic gradient for tensor decomposition. In Conference on Learning Theory, pages
797–842, 2015.
[GJZ17]
Rong Ge, Chi Jin, and Yi Zheng. No spurious local minima in nonconvex low rank
problems: A unified geometric analysis. arXiv preprint arXiv:1704.00708, 2017.
[GLM16]
Rong Ge, Jason D. Lee, and Tengyu Ma. Matrix completion has no spurious local
minimum. Advances in Neural Information Processing Systems (NIPS), 2016.
[GM17]
R. Ge and T. Ma. On the Optimization Landscape of Tensor Decompositions. ArXiv
e-prints, June 2017.
[HM17]
Moritz Hardt and Tengyu Ma. Identity matters in deep learning. In 5th International
Conference on Learning Representations (ICLR 2017), 2017.
[HMR16]
Moritz Hardt, Tengyu Ma, and Benjamin Recht. Gradient descent learns linear dynamical systems. CoRR, abs/1609.05191, 2016.
[HZRS16a] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning
for image recognition. In Proceedings of the IEEE conference on computer vision and
pattern recognition, pages 770–778, 2016.
[HZRS16b] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in
deep residual networks. In European Conference on Computer Vision, pages 630–645.
Springer, 2016.
30
[IS15]
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network
training by reducing internal covariate shift. In International Conference on Machine
Learning, pages 448–456, 2015.
[JGN+ 17]
Chi Jin, Rong Ge, Praneeth Netrapalli, Sham M Kakade, and Michael I Jordan. How
to escape saddle points efficiently. arXiv preprint arXiv:1703.00887, 2017.
[JSA15]
Majid Janzamin, Hanie Sedghi, and Anima Anandkumar. Beating the perils of nonconvexity: Guaranteed training of neural networks using tensor methods. arXiv
preprint arXiv:1506.08473, 2015.
[Kaw16]
Kenji Kawaguchi. Deep learning without poor local minima. In Advances in Neural
Information Processing Systems, pages 586–594, 2016.
[LSSS14]
Roi Livni, Shai Shalev-Shwartz, and Ohad Shamir. On the computational efficiency
of training neural networks. In Advances in Neural Information Processing Systems,
pages 855–863, 2014.
[LY17]
Yuanzhi Li and Yang Yuan. Convergence analysis of two-layer neural networks with
relu activation. arXiv preprint arXiv:1705.09886, 2017.
[MBM16]
Song Mei, Yu Bai, and Andrea Montanari. The landscape of empirical risk for nonconvex losses. arXiv preprint arXiv:1607.06534, 2016.
[O’D14]
Ryan O’Donnell. Analysis of boolean functions. Cambridge University Press, 2014.
[SC16]
Daniel Soudry and Yair Carmon. No bad local minima: Data independent training error
guarantees for multilayer neural networks. arXiv preprint arXiv:1605.08361, 2016.
[SJ13]
Nathan Srebro and Tommi Jaakkola. Weighted low-rank approximations. In ICML,
2013.
[SQW15]
Ju Sun, Qing Qu, and John Wright. When are nonconvex problems not scary? arXiv
preprint arXiv:1510.06096, 2015.
[Ste90]
Gilbert W Stewart. Matrix perturbation theory. 1990.
[Tia17]
Yuandong Tian. An analytical formula of population gradient for two-layered relu
network and its applications in convergence and critical point analysis. arXiv preprint
arXiv:1703.00560, 2017.
[Wik17a]
Wikipedia. Formal power series — wikipedia, the free encyclopedia, 2017. [Online;
accessed 20-September-2017 ].
[Wik17b]
Wikipedia. Hermite polynomials — wikipedia, the free encyclopedia, 2017. [Online;
accessed 1-September-2017 ].
[ZLJ16]
Yuchen Zhang, Jason D Lee, and Michael I Jordan. l1-regularized neural networks
are improperly learnable in polynomial time. In International Conference on Machine
Learning, pages 993–1001, 2016.
31
[ZLWJ17]
Yuchen Zhang, Jason Lee, Martin Wainwright, and Michael Jordan. On the learnability
of fully-connected neural networks. In Artificial Intelligence and Statistics, pages 83–91,
2017.
[ZPS17]
Qiuyi Zhang, Rina Panigrahy, and Sushant Sachdeva. Electron-proton dynamics in
deep learning. CoRR, abs/1702.00458, 2017.
[ZSJ+ 17]
Kai Zhong, Zhao Song, Prateek Jain, Peter L Bartlett, and Inderjit S Dhillon. Recovery
guarantees for one-hidden-layer neural networks. arXiv preprint arXiv:1706.03175,
2017.
A
Handling Non-Orthogonal Weights
In this section, we first show that when the weight vectors {b⋆i }′ s are not orthonormal, the local
optimum of a slight variant of G(B) still allow us to recover B ⋆ . The main observation is that the
set of local minima are preserved (in a certain sense) by linear transformation of the variables. We
design an objective function F (B) that is equivalent to G(B) up to a linear transformation. This
allows us to use Theorem 2.3 as a black box to characterize all the local minima of F .
A.1
Local Minimum after a Linear Transformation
Given a function f (y), we say function g(·) is a linear transformation of f (·) if there is a matrix W
such that g(x) = f (W x). If W has full rank, the local minima of f are closely related to the local
minima of g.
We recall some standard notation in calculus first. We use ∇f (t) to denote the gradient of f
evaluated at t. For example, ∇f (W x) is a shorthand for ∂f∂y(y) |y=W x , and similarly ∇2 f (W x) is
∂ 2 f (y)
|
.
(∂y)2 y=W x
The following theorem then connects the gradients and Hessians of f (W x) and g(x). Essentially,
it shows that the set of local minima and saddle points have a 1-1 mapping between f and g, and the
corresponding norms/eigenvalues only differ multiplicatively by quantities related to the spectrum
of W .
Theorem A.1. Let W ∈ Rd×m (d ≥ m) be a full rank matrix. Suppose g : Rm → R and f : Rd → R
are twice-differentiable functions such that g(x) = f (W x) for any x ∈ Rm . Then, for all x ∈ Rm ,
the following three properties hold:
1. σmin (W )k∇f (W x)k ≤ k∇g(x)k ≤ σmax (W )k∇f (W x)k.
2. If λmin (∇2 g(x)) < 0, then
σmax (W )2 λmin (∇2 f (W x)) ≤ λmin (∇2 g(x)) ≤ σmin (W )2 λmin (∇2 f (W x)).
3. The point x satisfies the first and second order optimality condition for g iff y = W x also
satisfy the first and second order optimality condition for f .
32
Proof. The proof follows from the relationship between the gradients of g and the gradients of f .
By basic calculus, we have
∇g(x) =
∂f (y)
∂f (W x)
= W⊤
∂x
∂y
y=W x
= W ⊤ ∇f (W x)
which immediately implies bullet 1. Similarly, we can compute the second order derivative:
∇2 g(x) = W ⊤ [∇2 f (W x)]W.
To simplify notation, let A = ∇2 f (W x). Let x = arg minkxk=1 x⊤ W ⊤ AW x, and y = (W x)/kW xk.
Therefore
λmin (A) ≤ y ⊤ Ay ≤ λmin (W ⊤ AW )/kW xk2 ≤ λmin (W ⊤ AW )/kW k2 .
On the other hand, let y be the unit vector that minimizes y ⊤ Ay, we know y is in column span
of W because f is only defined on the row span, so there must exist a unit vector x such that
W x = λy where λ ≥ σmin (W ). For this x we have λmin (W ⊤ AW ) ≤ x⊤ W ⊤ AW x = λ2 λmin (A) ≤
2 (W )λ
σmin
min (A). This finishes the proof for 2.
Finally, notice that W is full rank, so ∇g(x) = W ⊤ ∇f (W x) = 0 iff ∇f (W x) = 0. Also,
∇2 g(x) = W ⊤ [∇2 f (W x)]W 0 iff ∇2 f (W x) 0.
A.2
Objective for Non-Orthogonal Weights
Now we will design a new objective function that can be linearly transformed to the orthonormal
case. The main idea is to view the rows of B ⋆ as the new basis that we work on (which is not
necessarily orthogonal). Note that this is already the case for the first two terms of the objective
function G(B), we change the objective function as follows: More concretely, we define
X
X
√
Fα,µ,λ (B) = 2 6σ̂ ·
hb⋆i , bj i2 hb⋆i , bk i2
αi
i∈[d]
j,k∈[d]
m
m
X
X
σ̂4 µ X
αi hbj , b⋆i i2 − 1)2 − 1)2 .
− √
((
αi hb⋆i , bj i4 + λ
6 i,j∈[d]
j=1 i=1
Note
change in the objective is the regularizer for the norm of bj . It is now replaced
Pmthat the only
⋆
2
by (( i=1 αi hbj , bi i − 1)2 − 1)2 , which tries to ensure the “norm” of bj in the basis defined by row
of B ⋆ to be 1. The objective function that we will optimize corresponds to choosing αi = a⋆i .
Similar as before, this function can be computed as expectations
Fa⋆ ,µ,λ (B) = E y ·
X
φ(bj , bk , x) − µ E y ·
j,k∈[d],j6=k
m
X
+ λE(x,y),(x′ ,y′ ) [
i=1
X
j∈[d]
y · φ2 (bi , x) · y ′ · φ2 (bi , x′ )],
where (x′ , y ′ ) is an independent sample, and φ2 (v, x) = (v ⊤ x)2 − kvk2 .
33
ϕ(bj , x)
(A.1)
Intuitively, if we can find a linear transformation that makes {b⋆i }’s orthonormal, that will
reduce the problem
to the orthonormal case. This is in fact the whitening matrix:
P
⋆ ⋆ ⋆ ⊤
⋆
Let M = m
a
i=1 i bi (bi ) be the weighted covariance matrix of bi ’s. Suppose
pthe SVD of M is
⊤
−1/2
⊤
U DU and letpW = U D
. We apply the transformation W to the vectors a⋆i bi ’s and obtain
that oi = W ⊤ a⋆i b⋆i . We can verify that oi ’s are orthogonal vectors because
X
⊤
oi o⊤
i = W M W = Id
(A.2)
i∈[m]
For notational convenience, let’s extend the definition of the G(B) in equation by using the
putting the relevant information in the subscript
Gα,β,λ,o (B) =
√
+λ
6σ̂ ·
m
X
i=1
X
i∈[d]
a⋆i
X
σ̂4 µ X ⋆
hoi , b̄j i2 hoi , b̄k i2 − √
ai hoi , b̄j i4 .
6
j,k∈[d],j6=k
i,j∈[d]
(kb̄i k2 − 1)2
(That is, the index o denotes the ground-truth solution with respect to which G is defined.)
The next Theorem shows that we can rotate the objective function F properly so that it matches
the objective G with a ground-truth vector oi ’s.
Theorem A.2. Let W be defined as above, and let 1/a⋆ be the vector whose i-th entry is 1/a⋆i .
Then, we have that
G1/a⋆ ,µ,λ,oi (B) = Fa⋆ ,µ,λ (BW ⊤ )).
Note this can be interpreted as a linear transformation as in vector format BW ⊤ is equal to B ·
(W ⊤ ⊗ Idd×d ).
⊤
b̄1
..
Proof. The equality can be obtained by straightforward calculation. We note that since B = . ,
b̄⊤
m
(W ⊤
the rows of B ·
⊗ Idd×d ) are W b̄1 , . . . , W b̄m .
Therefore, we have that
Fa⋆ ,µ,λ (B · (W ⊤ ⊗ Idd×d ))
X
X
√
= 2 6σ̂ ·
hb⋆i , W b̄j i2 hb⋆i , W b̄k i2
a⋆i
i∈[d]
j6=k∈[d]
m
m X
X
σ̂4 µ X ⋆ ⋆
4
a⋆i hW b̄j , b⋆i i2 − 1)2 .
(
ai hbi , W b̄j i + λ
− √
6 i,j∈[d]
j=1 i=1
X 1 X p
√
p
= 2 6σ̂ ·
h a⋆i W ⊤ b⋆i , b̄j i2 h a⋆i W ⊤ b⋆i , b̄k i2
⋆
ai
i∈[d]
j,k∈[d]
m
m X
X
p
σ̂4 µ X 1 p ⋆ ⊤ ⋆
4
hb̄j , a⋆i W ⊤ b⋆i i2 − 1)2 .
(
a
W
b
,
b̄
i
+
λ
− √
h
i j
i
⋆
6 i,j∈[d] ai
j=1 i=1
34
(A.3)
X 1 X
√
= 2 6σ̂ ·
hoi , b̄j i2 hoi , b̄k i2
a⋆i
i∈[d]
j,k∈[d]
m
m X
X
σ̂4 µ X 1
4
( hb̄j , oi i2 − 1)2 .
hoi , b̄j i + λ
− √
a⋆i
6
i,j∈[d]
(by the definition of oi ’s)
j=1 i=1
From Theorem 2.3 we can immediately get the following Corollary (note that the only difference
is that the coefficients now are 1/a⋆i instead of a⋆i ). Recall a⋆max = maxi a⋆i and a⋆min = mini a⋆min ,
we have
Corollary A.3. Let κa = a⋆max /a⋆min . Let c be a sufficiently small universal constant (e.g. c =
0.01 suffices). Assume µ ≤ c/κa and λ ≥ (ca⋆min )−1 . The function G1/a⋆ ,µ,λ,oi (·) defined as in
Theorem A.2 satisfies that
1. A matrix B is a local minimum of G if and only if B can be written as B = P DO where O
is a matrix whose rows are oi ’s, P is a permutation matrix and D is a diagonal matrix with
Dii ∈ {±1 ± O(µ/λa⋆min )}.
2. Any saddle point B has a strictly negative curvature in the sense that λmin (∇2 G(B)) ≥ −τ0
where τ0 = c min{µ/(κa a⋆max d), λ}
3. Suppose B is an approximate local minimum in the sense that B satisfies
k∇g(B)k ≤ ε and λmin (∇2 g(B)) ≥ −τ0
Then B can be written as B = P DO + E where P is a permutation matrix, D is a diagonal
matrix and |E|∞ ≤ O(εa⋆max /σ̂4 ).
Finally, we can combine the theorem above
5.1 to give a guarantee for optimizing
p ⋆ and Theorem⋆ ⊤
F . Let Γ be a diagonal matrix with Γii = ai . Let M = B Γ2 B ⋆ and κ(M ) = kM k/σmin (M ).
Theorem A.4. Let c be a sufficiently small universal constant (e.g. c = 0.01 suffices). Let
κa = a⋆max /a⋆min . Assume µ ≤ c/κa and λ ≥ 1/(c · a⋆min ). The function F (·) defined as in
Theorem A.2 satisfies that
−⊤ = P DΓB ⋆ where P is
1. A matrix B is a local minimum of F if and only if B satisfy
p ⋆ B
a permutation matrix, Γ is a diagonal matrix with Γii = ai , and D is a diagonal matrix
with Dii ∈ {±1 ± O(µ/λa⋆min )}. Furthermore, this means that all local minima of F are also
global.
2. Any saddle point B has a strictly negative curvature in the sense that λmin (∇2 F (B)) ≥ −τ0
where τ0 = c min{µ/(κa da⋆max ), λ}σmin (M ).
3. Suppose B is an approximate local minimum in the sense that B satisfies
k∇F (B)k ≤ ε and λmin (∇2 F (B)) ≥ −τ0
Then B can be √written as B −⊤ = P DΓB ⋆ + E√where Γ, D, P are as in 1, the error term
kEk ≤ O(εa⋆max md · κ(M )1/2 /σ̂4 ) (when εa⋆max md · κ(M )1/2 /σ̂4 < c).
35
Proof. Note that we can immediately apply Theorem 2.3 to G1/a⋆ ,µ,λ,oi (B) to characterize all its
local minima. See Corollary A.3.
Next we will transform the properties for local minima of G (stated in Corollary A.3) to F
using Theorem A.1. First we note that the transformation matrix W and M are closely related:
W W ⊤ = M, σmin (W )2 = 1/kM k, kW k2 = 1/σmin (M ).
(A.4)
This is because according to the definition of W , the SVD of M is M = U DU ⊤ and W = U D −1/2 ,
so W W ⊤ = U D −1 U ⊤ = M −1 . The claims of the singular values follow immediately from the SVD
of M and W .
As a result, all local minimum of F are of the form BW ⊤ where B is a local minimum of G.
For B = BW ⊤ , the gradient and Hessian of F (B) and G(B) are also related by Theorem A.1.
Let us first prove 1. By Corollary A.3, we know every local minimum of G is of the form
B = P DO. According to the definition of O in Theorem A.2, we know each row vector oi is
equal to W ⊤ (a⋆i )1/2 b⋆i , therefore O = ΓB ⋆ W . As a result, all local minima of G are of the form
B = P DΓB ⋆ W . By Theorem A.1 and Theorem A.2, we know all local minima of F must be of the
form B = BW ⊤ = P DΓB ⋆ W W ⊤ = P DΓB ⋆ M −1 .
Now we try to compute B −⊤ . To do that observe that [ΓB ⋆ ]M −1 [ΓB ⋆ ]⊤ = I. Therefore
[ΓB ⋆ M −1 ]−⊤ = ΓB ⋆ , and for any local minimum B, we have
B −⊤ = (P DΓB ⋆ M −1 )−⊤ = P −⊤ D −⊤ (ΓB ⋆ M −1 )−⊤
= P −⊤ D −⊤ ΓB ⋆ .
Note that P ⊤ is still a permutation matrix, and D −⊤ is still a matrix whose diagonal entries are
{±1 ± O(µ/λa⋆min )},p
so this is exactly the form we stated in 1. More concretely, the rows of B −⊤
are permutations of a⋆i b⋆i .
For bullet 2, it follows immediately from Property 2 in Theorem A.1. Note that by property 2,
λmin (∇2 F (BW ⊤ )) ≤
λmin (∇2 G(B))
= λmin (∇2 G(B))σmin (M ).
kW k2
Finally we will prove 3. Let B = BW −⊤ , so that G(B) = F (B). We will prove properties of B
using the properties of B from Corollary A.3.
First we observe that by Theorem A.1,
λmin (∇2 G(B)) ≥ kW k2 λmin (∇2 F (B)) ≥ −c min{µ/(κa da⋆max , λ}.
Therefore the second order condition for Claim 3 in Corollary A.3 is satisfied. Now when
k∇F (B)k ≤ ε, we have k∇G(B)k ≤ εkW k = ε/σmin (M )1/2 . By Corollary A.3, we know B
can be expressed as P DO + E ′ where D is the diagonal matrix, P is a permutation matrix and
|E ′ |∞ ≤ O(εa⋆max /(σ̂4 σmin (M )1/2 )). We will apply perturbation Theorem A.9 for matrix inversion.
Since σmin (P DO) ≥ 1/2, we know when kE ′ k ≤ 1/4,
√
k(P DO + E ′ )−1 − (P DO)−1 k ≤ 8 2kE ′ k.
√
√
Here kEk is bounded by kEkF ≤ md|E ′ |∞ ≤ O(εa⋆max md/(σ̂4 σmin (M )1/2 )), which is smaller
than 1/4 when ε is small enough.
36
The corresponding point in F is B = BW ⊤ , and in 1 we have already proved (P DOW ⊤ )−⊤ is
of the form we want, therefore we can define E = B −⊤ − (P DOW ⊤ )−⊤ = (B − P DO)−⊤ W −1 , and
√
kEk = kW −1 kk(P DO + E ′ )−1 − (P DO)−1 k = O(εa⋆max md · κ(M )1/2 /σ̂4 ).
This finishes the proof.
A.3
Handle Undercomplete Case
The objective function F can handle the case when the weights b⋆i ’s are not orthogonal, but still
requires the number of components m to be equal to the number of dimensions d. In this section
we show how to use similar ideas for the case when the number of components is smaller than the
dimension (m < d).
Note that all the terms in F (B) only depends on the inner-products hbj , b⋆i i. Let S be the span
of {b⋆i }’s and PS be the projection matrix to this subspace, it is easy to see that F (B) satisfies
F (B) = F (BPS ).
That is, the previous objective function only depends on the projection of B in the space S. Using
similar argument as Theorem A.4, it is not hard to show the only local optimum in S satisfies the
same conditions, and allow us to recover B ⋆ . However, without modifying the objective, the local
optimum of F (B) can have arbitrary components in the orthogonal subspace S ⊥ .
In order to prevent the components from S ⊥ , we add an additional ℓ2 regularizer: define Fα,µ,λ,δ
as follows:
δ
Fα,µ,λ,δ (B) = Fα,µ,λ (B) + kBk2F
2
(A.5)
Intuitively, since the first term Fα,µ,λ (B) only cares about the projection BPS , minimizing kBk2F
will remove the components in the orthogonal subspace of S. We will choose δ carefully to make
sure that the additional term does not change the local optima of Fα,µ,λ (B) by too much, while
still ensuring a small projection on S ⊥ .
In this case we will consider pseudo-inverse instead of inverse. In particular, for a m × d matrix
B, define its pseudo-inverse B † to be the matrix such that BB † = Idm×m and B † B is the projection
to the row span
P of B.
⋆ ⋆ ⋆ ⊤
Let M = m
i=1 ai bi (bi ) , κ(M ) = kM k/σm (M ).
Theorem A.5. For any desired accuracy ε0 , we can choose parameters ε, δ, τ0 , µ, λ, such that for
the objective function Fa⋆ ,µ,λ,δ (B), for any B such that
k∇F(B)k ≤ ε,
∇2 F(B) ≥ −τ0 /2,
p
we have [B † ]⊤ = B ⋆ DΓP + E where Γ is a diagonal matrix with entries a⋆i , D is a diagonal
matrix with entries close to 1, P is a permutation matrix and kEk ≤ ε0 .
To choose the parameters, let c be a sufficiently small universal constant (e.g. c = 0.01 suf⋆
fices). Assume µ ≤ c/κ⋆ and λ ≥ 1/(c · a⋆min ). Let τ0 = c min{µ/(κda
max ), λ}σmin (M ). Let
p
cσ̂
ε
1/2
4
0
√
kM
k,
cε
δσ
δ ≤ min{ ⋆
,
τ
/2},
and
ε
=
min{λσ
(M
)
,
cδ/
0
min (M )}.
0
min
1/2
amax ·m dκ
(M )
37
We first show that if the gradient is small, then the point cannot have a large component in
S ⊥.
Lemma A.6. If k∇Fa⋆ ,µ,λ,δ (B)k ≤ ε, then kPS ⊥ BkF ≤ ε/δ.
Proof. Since Fa⋆ ,µ,λ (B) only depends BPS , we know ∇Fa⋆ ,µ,λ (B)PS ⊥ = 0. Therefore ε ≥
k∇Fa⋆ ,µ,λ,δ (B)PS ⊥ kF = k(δB)PS ⊥ kF = δkPS ⊥ BkF , and we have kBPS ⊥ kF ≤ ε/δ as desired.
Next we show that if the gradient of Fa⋆ ,µ,λ,δ (B) is small, and δ is also small, then the gradient
of Fa⋆ ,µ,λ (B) can be bounded.
Lemma A.7. In the setting of Theorem A.5, if k∇Fa⋆ ,µ,λ,δ (B)k ≤ ε ≤ λσmin (M )1/2 , then we have
p
k∇Fa⋆ ,µ,λ (B)k ≤ ε + δ 2m/σmin (M ).
Towards proving Lemma A.7, we first bound the norm of B by the following claim:
Claim A.8. If k∇Fa⋆ ,µ,λ,δ (B)k ≤ λσmin (M )1/2 , then each row bi must satisfy b⊤
i M bi ≤ 2.
Proof. We prove by contradiction. Assume towards contradiction that there is a column bi such
that b⊤
i M bi ≥ 2. We consider the quantity,
∂Fa⋆ ,µ,λ,δ
(B), bi i.
∂bi
√
P
P
⋆
2 ⋆
2
Note that Fa⋆ ,µ,λ,δ has 4 terms:
(1) 2 6σ̂ ·
i∈[d] αi
j,k∈[d] hbi , bj i hbi , bk i , (2)
P
P
P
m
δ
2
⋆ 2
2
2
− σ̂√46µ i,j∈[d] αi hb⋆i , bj i4 , (3) λ m
j=1 (( i=1 αi hbj , bi i − 1) − 1) , (4) 2 kBkF .
Among these 4 terms, the first, third and forth terms all contribute positively to this innerproduct
(because when bi is moved to (1 − ε)bi all those terms clearly decrease). Term 2
σ̂√4 µ P
− 6 i,j∈[m] a⋆i hb⋆i , bj i4 contribute negatively. Therefore we can ignore terms 1 and 4:
h
h
X
∂Fa⋆ ,µ,λ,δ
∂
σ̂4 µ X ⋆ ⋆
ai hbi , bj i4 + λ(
a⋆i hb⋆i , bj i2 − 1)2 ], bi i.
(B), bi i ≥ h
[− √
∂bi
∂bi
6
i∈[m]
i∈[m]
P
P
1
⋆ ⋆
4
⋆ 2 ⋆
4
2 ⋆
Let b⊤
i M bi = C ≥ 2, we know
i∈[m] ai hbi , bj i ≤ a⋆min
i∈[m] (ai ) hbi , bj i ≤ C /amin . Therefore,
4σ̂4 µ X ⋆ ⋆
4σ̂4 µ C 2
σ̂4 µ X ⋆ ⋆
∂
ai hbi , bj i4 ], bi i = − √
ai hbi , bj i4 ≥ − √ · ⋆
[− √
h
∂bi
6
6
6 amin
i∈[m]
i∈[m]
On the other hand,
X
∂
⊤
h
[λ(
a⋆i hb⋆i , bj i2 − 1)2 , bi i = 4λ(b⊤
i M bi − 1)(bi M bi ) = 4λC(C − 1).
∂bi
i∈[m]
By the choice of λ, µ, we can see that the negative term is negligible, and we know
∂Fa⋆ ,µ,λ,δ
(B), bi i ≥ 2λC(C − 1)
∂bi
p
C/σmin (M ). Therefore the norm of the gradient is at least
Since b⊤
i M bi = C, we√have kbi k ≤
2λC(C − 1)/kbi | ≥ 2 2λσmin (M ), this contradicts with the assumption. The norm of the rows
must all be bounded.
h
38
⊤
2
Proof of Lemma A.7. We havep
that
Pmbi M bi2 ≤ 2pimplies kbi k ≤ 2/σmin (M ). The norm of the whole
2m/σmin (M ), so by triangle inequality we have
matrix is bounded bykBkF ≤
i=1 kbi k ≤
p
k∇Fa⋆ ,µ,λ (B)k ≤ k∇Fa⋆ ,µ,λ,δ (B)k + δkBkF ≤ ε + δ 2m/σmin (M ).
Finally we are ready to prove Theorem A.5.
Proof of Theorem A.5. We will separate B into two components BS = BPS and B⊥ = BPS ⊥ .
We will first show that BS is close to the desirable solution. To do that wepwill use Theorem A.4
14 . By the choice of ε, δ, we know from Lemma A.7 that k∇F ⋆
a ,µ,λ (BS )k ≤ 2δ 2m/σmin (M ). Also,
∇2 Fa⋆ ,µ,λ (BS ) ≥ ∇2 Fa⋆ ,µ,λ (B) − δ ≥ −τ0 . Therefore we know BS must be of the form
[BS† ]⊤ = P DΓB ⋆ + E1 ,
where kE1 k < ε0 /2. Also at the same time from the proof of Theorem A.4 we know BS =
′ )W ⊤ where P DO+E ′ has singular values close to 1. Therefore σ
(P DO+E
min (BS ) ≥ σmin (W )/2 =
p
1/2 kM k.
By Lemma A.6 we know kB⊥ kF ≤ ε/δ. We apply inverse matrix perturbation (Theorem A.9)
again, using B = BS + B⊥ , therefore we know
B † = BS† + E2 ,
2 (B )) ≤ ε /2.
where kE2 k ≤ O(kB⊥ kF /σmin
S
0
Combining these two perturbations we know
[B † ]⊤ = P DΓB ⋆ + E1 + E2⊤ ,
and the error term E1 + E2⊤ has spectral norm at most ε0 .
A.4
Toolbox: Matrix Perturbation
In the proof we used the following theorem for the perturbation of matrices.
Theorem A.9 (Stewart and Sun [Ste90]). Consider the perturbation of a matrix A: if B =
A + E,then we have
√
kB † − A† k ≤ 2kA† kkB † kkEk.
As a corollary, if kEk ≤ σmin (A)/2, then we have
√
kB † − A† k ≤ 2 2σmin (A)−2 kEk.
B
Recovering the Linear Layer
We will show in this section that if we have are given a δ-approximation of B ⋆ , then it is easy to
recover a⋆ . The key observation here is that the correlation between the hb⋆i , xi and the output y
is exactly proportional to a⋆i . We also note that there could be multiple other ways to recover a⋆ ,
e.g., using linear regression with the σ(Bx) as input and the y as output. We chose this algorithm
mostly because of the ease of analysis.
14
If we restrict all the vectors to the subspace S, we can still apply Theorem A.4 as long as we replace all inverses
with pseudo-inverses.
39
Algorithm 1 Recovering a⋆
Input: A matrix B with unit row norms that is row-wise δ-close to B ⋆ in Euclidean distance.
b
b means the empirical average. Set ai ← |a′ | and bi ←
Return: Let a′i = 2E[yhx,
bi i] where E
i
′
bi sgn(ai )
Lemma (Restatement of Lemma 2.5). Given a matrix B whose rows are δ-close to B ⋆ in Euclidean
distance up to permutation and sign flip with δ ≤ 1/(2κ⋆ ). Then, we can give estimates a, B ′ (using
e.g., Algorithm 1) such that there exists a permutation P where ka − P a⋆ k∞ ≤ δa⋆max and B ′ is
row-wise δ-close to P B ⋆ .
To see why this simple algorithm works for recovering a⋆ , we need the following simple claim.
Claim B.1. For any vector v we have
m
1X ⋆ ⋆
E[yhx, vi] =
ai hbi , vi.
2
i=1
The proof of this claim follows immediately from the property of Hermite polynomials. Now
we are ready to prove the corollary.
Proof. Without loss of generality we assume B is close to a sign flip of B ⋆ . The unknown permutation does not change the proof.
Since bi is δ close to Bi⋆ , let u be the vector where uj = hb⋆j , bi − b⋆i i, we have
a′i
=
m
X
i=1
a⋆i hb⋆i , bi i
=
m
X
i=1
a⋆i (hb⋆i , b⋆i i + hb⋆i , bi − b⋆i i) = a⋆i + ha⋆i , ui ∈ a⋆i ± a⋆max δ.
Therefore a′i is always positive, ai is in the desirable range and kBi′ − Bi⋆ k ≤ δ.
Similarly, if −bi is δ close to Bi⋆ , we have a′i ∈ −a⋆i ± a⋆max δ, and the conclusion still holds.
For the settings considered in Section A, the vectors b⋆i are not necessarily orthogonal. In this
case we use the following algorithm:
Algorithm 2 Recovering a⋆ for general case
Input: A matrix B with unit row norms, and B is δ-close to B ⋆ in spectral norm up to permutation
and sign flip.
b
b means the empirical average.
Let ui = 2E[yhx,
bi i] where E
′
⊤
−1
Let a = (BB ) u.
Return: Set ai ← |a′i | and bi ← bi sgn(a′i )
Lemma B.2. Given a matrix B whose rows have unit norm, and kB − SP B ⋆ k ≤ δ for some
σ2 (B)
√ , we can give
permutation matrix P and diagonal matrix S with ±1 entries on diagonals.If 4√min
2κ⋆ m
estimates a, B ′ (using e.g., Algorithm 2) such that ka − P a⋆ k ≤
40
√
√
2 2a⋆max m
−2
σmin
(B)
· δ and kB ′ − P B ⋆ k ≤ δ.
Proof. We again use Claim B.1: in this case we know the vector u satisfies u = B(B ⋆ )⊤ a⋆ . As a
result, for the vector a′ , we have
a′ = (BB ⊤ )−1 (B(B ⋆ )⊤ )a⋆ = (B † )⊤ (B ⋆ )⊤ a⋆ = (B ⋆ B † )⊤ a⋆ .
By assumption we know B = SP B ⋆ + E where kEk ≤ δ. By the perturbation of matrix
inverse (Theorem
A.9), we know if kEk ≤ δ ≤ σmin (B)/2, then B † = (B ⋆ )† P −1 S −1 + E ′ where
√
kE ′ k ≤ 2 2σmin (B)−2 δ. Therefore
a′ = (P −1 S −1 + E ′ )⊤ a⋆ = S −⊤ P −⊤ a⋆ + (E ′ )⊤ a⋆ = SP a + (E ′ )⊤ a⋆ .
(Here the last equality is because for both permutation matrix P and sign flip matrix S, P −⊤ = P
and S −⊤ = S.) Therefore, coordinates of a′ are permutation and sign flips of a⋆ , up to an error
term (E ′ )⊤ a⋆ .
√
σ2 (B)
⋆
√ , we know k(E ′ )⊤ a⋆ k ≤ kE ′ ka⋆
When δ ≤ 4√min
max m ≤ amin /2, therefore the signs are
2κ⋆ m
all recovered correctly. After fixing the sign, we have ka − P ak ≤ k(E ′ )⊤ a⋆ k ≤
kB ′
C
−
P B ⋆k
≤ δ.
√
√
2 2δa⋆max m
,
−2
σmin (B)
and
Sample Complexity
In this section we will show that our algorithm only requires polynomially many samples to find
the desired solution. Note that we did not try to optimize the polynomial dependency.
Theorem C.1 (Theorem 2.7 Restated). In the setting of Theorem 2.3, suppose we use N empirical
b There exists a fixed polynomial such that if
samples to approximate G and obtain function G.
⋆
⋆
b
N ≥ poly(d, amax /amin , 1/ε), with high probability for any point B with λmin (∇2 G(B))
≥ −τ0 /2
b
and k∇G(B)k
≤ ε/2, then B can be written as B = DP + E where P is a permutation matrix, D
is a diagonal matrix and |E|∞ ≤ O(ε/(σ̂4 a⋆min )).
In order to bound the sample complexity, we will prove a uniform convergence result: we
b are point-wise close to
show that with polynomially many samples, the gradient and Hessian of G
b must also be an
the gradient and Hessian of G, therefore any approximate local minimum of G
approximate local minimum of G.
However, there are two technical issues in showing the uniform convergence result. The first
b are very
issue is that when the norm of B is very large, both the gradient and Hessian of G and G
large and we cannot hope for good concentration. We deal with this issue by showing when B has
b
a large norm, the empirical gradient ∇G(B)
must also have large norm, and therefore it can never
be an approximate local minimum (we do this later in Lemma C.5). The second issue is that our
objective function involves high-degree polynomials over Gaussian variables x, y, and is therefore
not sub-Gaussian or sub-exponential. We use a standard truncation argument to show that the
function does not change by too much if we restrict to the event that the Gaussian variables have
bounded norm.
Lemma C.2. Suppose P ′ (B) + R(B) = E(x,y) [f (x, y, B)] where f is a polynomial of degree at most
5 in x, y and at most 4 in B. Also assume that the sum of absolute values of coefficients is bounded
by Γ. For any ε ≤ Γ/2, let R = Cd log(a⋆max Γ/ε) for a large enough constant C, let F be the event
41
that kxk2 ≤ R, and let Gtrunc = E(x,y) [f (x, y, B)1F ]. For any B such that kbi k ≤ 2 for all rows, we
have
k∇G(B) − ∇Gtrunc (B)k ≤ ε,
and
k∇2 G(B) − ∇2 Gtrunc (B)k ≤ ε,
Proof. By standard χ2 concentration bounds, for large enough C and any z > R, the probability
that kxk2 ≥ z is at most exp(−10z).
By simple calculation, it is easy to check that k∇B f (x, y, B)k ≤ 4Γd1.5 a⋆max kxk5 , and
2
k∇B f (x, y, B)k ≤ 12Γd2 a⋆max kxk5 . We know k∇G(B) − ∇Gtrunc (B)k = kE[∇B f (x, y, B)(1 −
1F )k. The expectation between kxk2 ∈ [2i R, 2i+1 R], for i = 0, 1, 2, ..., is always bounded by
4Γd1.5 a⋆max k2i+1 Rk5 exp(−2i R) < ε/2i+1 . Therefore
k∇G(B) − ∇Gtrunc (B)k ≤
∞
X
i=0
ε/2i+1 ≤ ε.
The bound for the Hessian follows from the same argument.
Finally, we combine this truncation with a result of [MBM16] that proves universal convergence
of gradient and Hessian. For completeness here we state a version of their theorem with bounded
gradient/Hessian:
Theorem C.3 (Theorem 1 in [MBM16]). Let f (θ) be a function from Rp → R and fˆ be its
empirical version. If the norm of the gradient and Hessian of a function is always bounded by τ ,
for variables in a ball of radius r in p dimensions, there exists a universal constant C0 such that
for C = C0 max{log rτ /δ, 1}, the following hold:
(a) The sample gradient converges to the population gradient. Namely if N ≥ Cp log p we have
r
Cp log n
] ≥ 1 − δ.
Pr[ sup k∇f (θ) − ∇fˆθ k ≤ τ
n
kθk≤r
(b) The sample Hessian converges to the empirical Hessian. Namely if N ≥ Cp log p we have
r
Cp log n
Pr[ sup k∇f (θ) − ∇fˆθ k ≤ τ
] ≥ 1 − δ.
n
kθk≤r
As an immediate corollary of this theorem and Lemma C.2, we have
Corollary C.4. In the setting of Theorem 2.7, for every B whose rows have norm at most 2, we
have with high probability,
b
k∇G(B) − ∇G(B)k
≤ ε/2,
and
b
k∇2 G(B) − ∇2 G(B)k
≤ τ0 /2.
42
Proof. On the other hand, for all such matrices B, by Lemma C.2 we know the gradient and Hessian
of G is close to the gradient and Hessian of Gtrunc .
k∇G(B) − ∇Gtrunc (B)k ≤ ε/4,
and
k∇2 G(B) − ∇2 Gtrunc (B)k ≤ τ0 /4.
Now, the gradient and Hessian for individual samples for estimating Gtrunc are bounded by some
b are close to those
poly(d, 1/ε), therefore by Theorem C.3 we know the gradient and Hessian of G
of Gtrunc . When N ≥ poly(d, 1/ε) for a large enough polynomial, we have with high probability,
for all B with all rows kbi k ≤ 2,
b
k∇Gtrunc (B) − ∇G(B)k
≤ ε/4,
and
b
k∇2 Gtrunc (B) − ∇2 G(B)k
≤ τ0 /4.
The corollary then follows from triangle inequality.
Finally we handle the case when B has a row with large norm. We will show that in this case
b
∇G(B)
must also be large, so B cannot be an approximate local minimum.
Lemma C.5. If bi is the row with largest norm and kbi k ≥ 2, then when N ≥ poly(d, a⋆max /a⋆min )
b
for some fixed polynomial, we have with high probability h∇G(B),
bi i ≥ cλkbi k4 for some universal
constant c > 0.
Proof. The proof of this Lemma is very similar to Claimh A.8.
Note that by i equaP
b
tion (2.8) there are three terms in G(B):(1) sign(σ̂4 )Ê y · j,k∈[d],j6=k φ(bj , bk , x) , (2)
h P
i
P
2
2
−µ sign(σ̂4 )Ê y · j∈[d] ϕ(bj , x) , (3) λ m
i=1 (kbi k − 1) . Here Ê is the empirical average over
the samples.
Note that the first two terms are homogeneous degree 4 polynomials over B, and the third term
does not depend on the sample. By argument similar to Corollary C.4, we know for any B where
bi has the largest row norm, with the number of samples we choose the gradient of the first two
terms is ca⋆min kbi k3 close to the gradient of their expectations, where c < 0.01 is a small constant.
√
By Theorem 2.6, we know the expectation of the first two terms are equal to A1(B) = 2 6|σ̂4 | ·
P
P
|σ̂4 |µ P
⋆
2 ⋆
2
⋆
⋆ ⋆
4
√
j,k∈[d],j6=k hbi , bj i hbi , bk i and A2(B) = −
i∈[d] ai
i,j∈[d] ai hbi , bj i . Here the gradient of
6
the first term always have positive correlation with bi , so we can ignore it. For the second term,
we know the gradient
|σ̂4 |µ X ⋆ ⋆
∂
aj hbj , bi i3 b⋆j .
[A2(B)] = − √
∂bi
6 j
Taking the inner-product with bi , and use the fact that b⋆i form an orthonormal basis, we know
h
∂
[A2(B)], bi i ≥ −ca⋆min kbi k4 .
∂bi
On the other hand, when kbi k ≥ 2, we have for the third term
h
∂
[λ(kbi k2 − 1)2 ], bi i ≥ λ(kbi k − 1)4 ≥ λkbi k4 /16.
∂bi
43
Since λ is larger than a⋆max , we know the negative contribution from A2 and the difference between
b
the empirical version and G are both negligible. Therefore we have h∇G(B),
bi i ≥ cλkbi k4 as
desired.
Now we are ready to prove Theorem 2.7:
b
Proof. By Lemma C.5, any point B with ∇G(B)
≤ ε must have kbi k ≤ 2 for all i. Now by
Corollary C.4, we know the point B we have must satisfy
k∇G(B)k ≤ ε; ∇2 G(B) −τ0 Id.
By point 3 in Theorem 2.3, this implies the guarantee on B.
D
Spurious Local minimum for function P ′
In this section we give an example where the function P ′ does have spurious local minimum.
In this example, d = 4, and the true vectors are the standard basis vectors b⋆i = ei . We will set
⋆
a1 = 1, and a⋆2 = a⋆3 = a⋆4 = 2 + δ (where δ > 0 is an arbitrary positive constant).
The √spurious local minimum that we consider is b1 = b2 = e1 = b⋆1 , b3 = e2 = b⋆2 , b4 =
√
2
2
2 e3 + 2 e4 . That is,
1 0 0
0
1 0 0
0
.
B=
0 1 0
0
√
√
0 0 22 22
The objective P ′ (B) = 1 and the only non-zero term is a⋆1 hb⋆1 , b1 i2 hb⋆1 , b2 i2 . In order to improve
the objective locally, we need to change either b1 or b2 , otherwise the term a⋆1 hb⋆1 , b1 i2 hb⋆1 , b2 i2 is still
1, and all other terms (a⋆i hb⋆i , bj i2 hb⋆i , bk i2 ) are non-negative.
p
p
Assume we have a local perturbation B ′ , where b′1 = 1 − ε21 e1 + ε1 u1 , b′2 = 1 − ε22 e1 + ε2 u2 .
Here u1 , u2 are unit vectors that are orthogonal to e1 . Also, since this is a local perturbation, we
make sure ε1 , ε2 ≤ ε, and b3 (2) ≥ 1 − ε, [b4 (3)]2 , [b4 (4)]2 ≥ 0.5 − ε. We will show that when ε is
small enough, the objective function P ′ (B ′ ) ≥ 1.
To see this, notice that the term a⋆1 hb⋆1 , b1 i2 hb⋆1 , b2 i2 is now equal to (1 − ε21 )(1 − ε22 ). On the
other hand, for b1 , we have
4
X
i=2
a⋆i
4
X
k=3
hb⋆i , b1 i2 hb⋆i , bk i2 = ε21 (2 + δ)
= ε21 (2 +
≥ ε21 (2 +
4 X
4
X
hb⋆i , u1 i2 hb⋆i , bk i2
i=2 k=3
4
4
X
X
hb⋆i , u1 i2 ( hb⋆i , bk i2 )
δ)
i=2
k=3
4
4
X
X
4
hb⋆i , u1 i2 · min{ hb⋆i , bk i2 }
δ)
i=2
i=2
k=3
≥ ε21 (2 + δ)(0.5 − ε).
Similarly we have the same equation for b2 . Note that all the terms we analyzed are disjoint,
therefore
P ′ (B ′ ) ≥ (1 − ε21 )(1 − ε22 ) + ε21 (2 + δ)(0.5 − ε) + ε22 (2 + δ)(0.5 − ε).
44
By removing higher order terms of ε, it is easy to see that P ′ (B ′ ) ≥ 1 when ε is small enough.
Therefore B is a local minima of P ′ .
45
| 8cs.DS
|
Submitted to the Annals of Statistics
EFFICIENT ESTIMATION FOR DIMENSION REDUCTION WITH
CENSORED DATA
arXiv:1710.05377v1 [math.ST] 15 Oct 2017
By Ge Zhao, Yanyuan Ma and Wenbin Lu∗
We propose a general index model for survival data, which generalizes many commonly used semiparametric survival models and
belongs to the framework of dimension reduction. Using a combination of geometric approach in semiparametrics and martingale treatment in survival data analysis, we devise estimation procedures that
are feasible and do not require covariate-independent censoring as
assumed in many dimension reduction methods for censored survival
data. We establish the root-n consistency and asymptotic normality
of the proposed estimators and derive the most efficient estimator
in this class for the general index model. Numerical experiments are
carried out to demonstrate the empirical performance of the proposed
estimators and an application to an AIDS data further illustrates the
usefulness of the work.
1. Introduction. Cox proportional hazards model (Cox, 1972) is probably the most widely
used semiparametric model for analyzing survival data. In the Cox model, covariate effect is described by a single linear combination of covariates in an exponential function and is multiplicative
in modeling the hazard function. Although this special way of modeling the hazard function permits a convenient estimation procedure, such as the maximum partial likelihood estimation (Cox,
1975), it has its limitations. As widely studied in the literature, there are many situations where
the Cox model may not be proper. Due to the limitations of the Cox model, many other semiparametric survival models have been proposed in the literature, such as the accelerated failure
time model (Buckley and James, 1979), proportional odds model (McCullagh, 1980) and linear
transformation model (Dabrowska and Doksum, 1988), etc. Despite of all these efforts, the link
between the summarized covariate effect, typically in the form of a linear combination of covariates, and the possibly transformed event time remains to have a predetermined form and hence
can be restrictive sometimes.
The single index feature of the above mentioned semiparametric survival models is appealing
since the covariates effect has a nice interpretation. It also naturally achieves dimension reduction
when there is a large number of covariates. However, the specific model form to link the covariate
index to the event time may be restrictive, and it is often difficult to check the goodness-of-fit of the
specific link function form. To achieve a model that is flexible yet is feasible in practice, we borrow
and extend the idea of linear summary of the covariate effects, while free up the specific functional
relation between the event time and the linear summaries. Thus, we propose the following general
index model
(1)
prpT ď t | Xq “ prpT ď t | β T
0 Xq,
tą0
where T is the survival time of interest, X is the p-dimensional baseline covariates, and β 0 P Rpˆd
is the regression coefficient matrix, with p ą d. Several properties of model (1) is worth mentioning.
Ge Zhao is graduate student (E-mail: [email protected]) and Yanyuan Ma is professor (E-mail: [email protected]),
Department of Statistics, Penn State University, University Park, PA, 16802. Wenbin Lu is professor (E-mail:
[email protected]), Department of Statistics, North Carolina State University, Raleigh, NC 27695. The work is supported by NSF grant DMS-1608540.
Keywords and phrases: Dimension reduction, General index model, Kernel estimation, Semiparametric theory,
Survival analysis
∗
1
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
2
1) First of all, instead of a single linear summary, we allow d linear summaries described by the d
columns of β 0 . This increases the flexibility of how the covariate effects are combined. We can view
this as a generalization from single index to multi index covariate summary. Imagine an extreme
case when d “ p, this model degenerates to the restriction free case where the dependence of T on
X is arbitrary. Of course, in practice, when d is large, the estimation will encounter difficulties and
it is not feasible to carry out the analysis. However conceptually this provides a way of appreciating
the flexibility of the model. In addition, we will see that in practice, when d is often smaller than
p, this model framework allows us to find and incorporate the suitable number of indices d. 2)
Second, we do not specify any functional form of the conditional probability. Thus, the conditional
probability in (1) is simply a function of both t and β T
0 X. This relaxes both the exponential form
of the covariate relation and the multiplicative form of the hazard function in the Cox model and is
also much more flexible than other popular semiparametric survival models, such as the accelerated
failure time and linear transformation models. Despite of the flexibility of the model in (1), we
show that through properly incorporating semiparametric treatment and martingale techniques,
estimation and inference is still possible. 3) In addition, the analysis can be carried out under
the usual conditional independent censoring assumption, where the censoring time is allowed to
depend on the covariates.
The proposed general index model and associated semiparametric estimation method naturally
provide a dimension reduction tool for survival data. It has a few advantages over existing dimension
reduction methods for survival data. 1) First, many existing dimension reduction methods for survival data require a stronger assumption on the censoring time, such as the covariate-independent
censoring assumption (e.g. Li, Wang and Chen, 1999; Lu and Li, 2011), or requires nonparametric estimation of the conditional survival function of censored survival times (Xia, Zhang and Xu,
2010) or censoring times (Li, Wang and Chen, 1999) given all the covariates, which may suffer
from the curse of dimensionality. All these drawbacks are avoided here. 2) Second, most of existing
methods (Xia, Zhang and Xu, 2010; Li, Wang and Chen, 1999) are constructed based on general
inverse probability weighted estimation techniques in one way or another, and are thus not efficient. In contrast, our proposed method is built on the semiparametric theory (Tsiatis, 2006) and
achieves the optimal semiparametric efficient estimator.
The rest of the paper is organized as the following. In Section 2, we develop the estimation
procedures for both the index parameters in β and functional relation between event time and
the multiple indices. In Section 3, we establish the large sample properties to enable inference. We
perform extensive numerical experiments in Section 4, where both simulation and analysis of an
AIDS data are included. We conclude the paper with a discussion in Section 5, while relegate all
the technical details in an Appendix.
2. Methodology Development. Define Z “ minpT, Cq and ∆ “ IpT ď Cq, where C is the
censoring time. Assume C T | X and the relation between T and X follows the model in (1). The
observed data consist of pXi , Zi , ∆i q, i “ 1, . . . , n, which are independent copies of pX, Z, ∆q. Note
that even without censoring, β 0 in (1) is not identifiable because for any d ˆ d full rank matrix A,
β 0 and β 0 A suite the model (1) equally well. Thus, we fix a parameterization of β 0 by assuming
the upper d ˆ d block of β 0 to be the identity matrix Id . This ensures the unique identification of
β 0 . Here we consider a fixed d, and our focus will be in estimating the lower block of β 0 , which
has dimension pp ´ dq ˆ d. We then proceed to estimate the conditional distribution function in
d
p´d . Note that under the
T T
(1). For convenience, write X “ pXT
u , Xl q , where Xu P R and Xl P R
assumption of C T | X and (1), we have
T
Etf1 pCqf2 pT q | β T
0 Xu “ ErEtf1 pCqf2 pT q | Xu | β 0 Xs
T
“ ErEtf1 pCq | XuEtf2 pT q | β T
0 Xu | β 0 Xs
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
3
T
“ Etf1 pCq | β T
0 XuEtf2 pT q | β 0 Xu
for any functions f1 , f2 , hence C T | β T
0 X.
To facilitate further development, in describing the censoring process, write Sc pz, xq “ prpC ě z |
X “ xq, Λc pz, xq “ ´logSc pz, xq, λc pz, xq “ BΛc pz, xq{Bz and fc pz, xq “ ´BSc pz, xq{Bz. Similarly,
to describe the event process, for any parameter vector β, define Spz, β T xq “ prpT ě z | β T X “
β T xq, f pz, β T xq “ ´BSpz, β T xq{Bz, Λpz, β T xq “ ´logSpz, β T xq and λpz, β T xq “ BΛpz, β T xq{Bz.
Note that the functions S, Λ, λ and f are in fact different when the parameter β changes, so the
more precise notations are Spz, β T x, βq, Λpz, β T x, βq, λpz, β T x, βq and f pz, β T x, βq. Here, without
causing confusion, we omit the last parameter for notational simplicity. Using these notation, the
pdf of the model described in (1) is
(2)
fX,Z,∆ px, z, δ, β, λ, λc , fX q “ fX pxqλpz, β T xqδ e´
şz
0
λps,βT xqds
λc pz, xq1´δ e´
şz
0
λc ps,xqds
,
where fX pxq is the pdf of X. We assume the true data generation process is based on fX,Z,∆ px, z, δ, β 0 , λ0 , λc0 , fX0 q.
T
Note that following our notation, λ0 pz, β T
0 xq “ λpz, β 0 xq.
We view (2) as a semiparametric model, where β is a finite dimensional parameter of interest and
all the remaining unknown components of the model are treated as infinite dimensional nuisance
parameters. In survival analysis, the most popular approaches to estimation are martingale based
estimators (Fleming and Harrington, 1991) and nonparametric maximum likelihood estimators
(NPMLE) (Zeng and Lin, 2007). Here we find that NPMLE does not suit well without adaption
due to the inseparable relation between the hazard function and the covariates. Martingale approach may enable us to obtain one specific estimator for β, while we aim at obtaining a more
comprehensive understanding of the estimation of β. Thus we use a less conventional approach
by adopting the geometrical treatment in semiparametrics. Similar practice has been performed
in Tsiatis (2006) to rediscover the partial likelihood estimator for Cox proportional hazard model.
To this end, we first characterize the nuisance tangent space as described in Proposition 1. The
proof utilizes properties of şmartingale integration and the details are given
şt in Appendix A.1. Det
T
T
fine M pt, β 0 Xq ” N ptq ´ 0 Y psqλ0 ps, β 0 Xqds and Mc pt, Xq ” Nc ptq ´ 0 Y psqλc ps, Xqds, where
N ptq “ ∆IpZ ď tq, Nc ptq “ p1 ´ ∆qIpZ ď tq and Y ptq “ IpZ ě tq. Then M pt, β T
0 Xq and Mc pt, Xq
are mean-zero martingale processes.
Proposition 1.
Λ1
Λ2
Λ3
The nuisance tangent space Λ “ Λ1 ‘ Λ2 ‘ Λ3 , where
ı
”
“ apXq : EtapXqu “ 0, apXq P Rpp´dqd ,
"ż 8
*
T
T
T
pp´dqd
hps, β 0 XqdM ps, β 0 Xq : @hpZ, β 0 Xq P R
“
,
"ż08
*
“
hps, XqdMc ps, Xq : @hpZ, Xq P Rpp´dqd .
0
Having found the nuisance tangent space, we can now proceed to identify the efficient score function through projecting the score function onto Λ and calculating the residual. The score function is
defined as Sβ p∆, Z, Xq ” BlogfX,Z,∆px, z, δ, β, λ, λc , fX q{Bβ. Let λ1 ps, β T Xq “ Bλps, β T Xq{Bpβ T Xq
T
T
and λ10 ps, β T
0 Xq “ Bλ0 ps, β 0 Xq{Bpβ 0 Xq. Straightforward calculation yields
Sβ p∆, Z, Xq “
ż8
0
λ1 ps, β T Xq
b Xl dM ps, β T Xq.
λps, β T Xq
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
4
We can verify that Sβ 0 p∆, Z, Xq K Λ1 and Sβ 0 p∆, Z, Xq K Λ3 due to the martingale properties.
Thus to look for the efficient score, we only need to project Sβ p∆, Z, Xq onto Λ2 and calculate its
residual. To this end, we search for h˚ ps, β T
0 Xq so that
ż8
T
h˚ ps, β T
Seff p∆, Z, Xq “ Sβ0 p∆, Z, Xq ´
0 XqdM ps, β 0 Xq
0
*
ż8"
λ10 ps, β T
0 Xq
˚
T
b
X
´
h
ps,
β
Xq
dM ps, β T
“
l
0
0 Xq
T
λ0 ps, β 0 Xq
0
is orthogonal to Λ2 . This entails that for any hps, β T
0 Xq,
„ż 8
ż8"
*
λ10 ps, β T
0 Xq
˚
T
T
h
0 “ E
b Xl ´ h ps, β 0 Xq dM ps, β 0 Xq
λ0 ps, β T
0
0
0 Xq
"
„ż 8
*
λ10 ps, β T
0 Xq
T
T
˚
T
T
h ps, β 0 Xq
“ E
b Xl ´ h ps, β 0 Xq Y psqλ0 ps, β 0 Xqds .
λ0 ps, β T
0
0 Xq
T
T
ps, β T
0 XqdM ps, β 0 Xq
T
T
By letting hps, β T
0 Xq “ Ips “ tqapβ 0 Xq for any apβ 0 Xq, we obtain that
*
λ10 pt, β T
0 Xq
˚
T
T
T
b Xl ´ h pt, β 0 Xq Y ptqλ0 pt, β 0 Xq | β 0 X
0 “ E
λ0 pt, β T
0 Xq
„"
*
λ10 pt, β T
0 Xq
˚
T
T
“ E
b
X
´
h
pt,
β
Xq
Y
ptq
|
β
X
.
l
0
0
λ0 pt, β T
0 Xq
„"
Note that
(
(
E Xl Sc pt, Xq | β T
E Xl Y ptq | β T
0X
0X
( “
( .
E Y ptq | β T
E Sc pt, Xq | β T
0X
0X
(3)
This leads to
h
˚
λ10 pt, β T
0 Xq
pt, β T
Xq
0
T
λ0 pt, β 0 Xq
(
E Xl Sc pt, Xq | β T
0X
( .
b
E Sc pt, Xq | β T
0X
Thus the efficient score is
«
(ff
ż8
T
E
X
S
ps,
Xq
|
β
X
λ10 ps, β T
Xq
c
l
0
0
( dM ps, β T
(4)
b Xl ´
Seff p∆, Z, Xq “
0 Xq.
T
λ0 ps, β T
Xq
E
S
ps,
Xq
|
β
X
0
c
0
0
Several properties of the efficient score is worth pointing out. First of all, EtSeff p∆, Z, Xq | Xu “
0 is ensured by EtdM pt, β T
0 Xq | Xu “ 0, hence to preserve the mean zero property, we can replace
T
T
λ10 ps, β T
Xq{λ
ps,
β
Xq
by
any function of s and β T
0
0
0
0 X, say gps, β 0 Xq, and still obtain
«
(ff
ż8
E Xl Sc ps, Xq | β T
0X
T
( dM ps, β T
gps, β 0 Xq b Xl ´
E
0 Xq “ 0.
T
E Sc ps, Xq | β 0 X
0
This implies that if we are only aiming at a consistent estimator, we can use an arbitrary function
T
T
gps, β T
0 Xq to replace λ10 ps, β 0 Xq{λ0 ps, β 0 Xq in the efficient score to get a more general martingale
integration.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
5
Second, using (3), by first taking expectation conditional on β T X, we can verify that
«
(ff
ż8
E Xl Sc ps, Xq | β T
0X
T
( Y psqλ0 ps, β T
gps, β 0 Xq b Xl ´
E
0 Xqds
T
E Sc ps, Xq | β 0 X
0
ff
«
(
ż8
E Xl Sc ps, Xq | β T
0X
T
T
T
( EtY psq | β 0 Xu
gps, β 0 Xq b EtXl Y psq | β 0 Xu ´
“ E
E Sc ps, Xq | β T
0
0X
ˆλ0 ps, β T
0 Xqds
“ 0.
As a consequence,
E
ż8
0
«
E Xl Sc ps, Xq | β T
0X
(
gps, β T
Xq
b
X
´
l
0
T
E Sc ps, Xq | β 0 X
(ff
dN psq “ 0.
This implies that we can construct estimating equations of the form
«
(ff
n
p Xli Yi pZi q | β T Xi
ÿ
E
∆i gpZi , β T Xi q b Xli ´
(5)
“0
(
p Yi pZi q | β T Xi
E
i“1
for any g, where
(6)
(7)
p Yi pZi q | β Xi
E
T
p Xli Yi pZi q | β T Xi
E
(
(
řn
T
T
j“1 Kh pβ Xj ´ β Xi qIpZj ě Zi q
,
řn
T
T
j“1 Kh pβ Xj ´ β Xi q
řn
T
T
j“1 Kh pβ Xj ´ β Xi qXlj IpZj ě Zi q
.
řn
T
T
j“1 Kh pβ Xj ´ β Xi q
“
“
Here EtYi pZi q | β T Xi u ” EtYi ptq | β T Xi u|t“Zi and similarly for other terms, Kp¨q is a kernel
function and Kh p¨q “ Kp¨{hq{h.
Third, we can further relax the estimating equation form to
«
(ff
n
p apXli qYi pZi q | β T Xi
ÿ
E
T
(8)
“0
∆i gpZi , β Xi q b apXli q ´
(
p Yi pZi q | β T Xi
E
i“1
by taking advantage of the fact that
«
(ff
T
E
apX
qY
pZq
|
β
X
l
0
(
E∆gpZ, β T
“0
0 Xq b apXl q ´
E Y pZq | β T
X
0
for any apXl q.
T
Fourth, when we choose to estimate λ10 ps, β T
0 Xq{λ0 ps, β 0 Xq, using, for example, the smoothed
Kaplan-Meier estimator, we can then obtain the efficient estimator from solving
«
(ff
n
p Xli Yi pZi q | β T Xi
ÿ
p 1 pZi , β T Xi q
E
λ
∆i
(9)
“ 0.
b Xli ´
(
p i , β T Xi q
p Yi pZi q | β T Xi
E
λpZ
i“1
Here we can use a local Nelson-Aalen estimator to estimator the cumulative hazard function
p
ΛpZ,
β T Xq “
ÿ
Zi ďZ
∆i Kh pβ T Xi ´ β T Xq
.
T
T
j“1 IpZj ě Zi qKh pβ Xj ´ β Xq
řn
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
6
The local Nelson-Aalen estimator of hazard function can be obtained from
ż8
T
p
p
Kb pt ´ ZqdΛpt|β
Xq
λpZ,
β T Xq “
0
(10)
“
and we estimate the derivative
n
ÿ
i“1
∆i Kh pβ T Xi ´ β T Xq
,
T
T
j“1 IpZj ě Zi qKh pβ Xj ´ β Xq
Kb pZi ´ Zq řn
p 1 pZ, β T Xq “ BλpZ,
p
λ
β T Xq{Bpβ T Xq
n
ÿ
Kb pZi ´ Zq řn
“ ´
(11)
∆i K1h pβ T Xi ´ β T Xq
T
T
j“1 IpZj ě Zi qKh pβ Xj ´ β Xq
i“1
ř
n
T
T
n
1
ÿ
j“1 IpZj ě Zi qKh pβ Xj ´ β Xq
T
T
Kb pZi ´ Zq∆i Kh pβ Xi ´ β Xq řn
`
.
t j“1 IpZj ě Zi qKh pβ T Xj ´ β T Xqu2
i“1
Here K1h pvq “ BKh pvq{Bv is the first derivative of Kh with respect to its variables, which is a
vector, and b is a bandwidth. For any vector or matrix a, let ab2 “ aaT .
Among the different constructions of consistent estimators, the estimator obtained from (9) will
be shown to achieve the smallest possible variability, hence this estimator is efficient and is what
we recommend. The efficient estimator will be the focus of our study. We provide the detailed
algorithm of the efficient estimation procedure below.
1. Obtain an initial estimator of β through, for example, hmave (Xia, Zhang and Xu, 2010).
r
Denote the result β.
2. Replacing EtY pZq | β T Xu, EtXl Y pZq | β T Xu, λpZ, β T Xq and λ1 pZ, β T Xq with their
nonparametric estimated versions given in (6), (7), (10) and (11) respectively. Write the
p
p 1 pZ, β T Xq.
p l Y pZq | β T Xu, EtY
p pZq | β T Xu, λpZ,
resulting estimators as EtX
β T Xq and λ
T
T
T
T
p
p
p
p
3. Plug EtXl Y pZq | β Xu, EtY pZq | β Xu, λpZ, β Xq and λ1 pZ, β Xq into (9) and solve
p using β
r as starting value.
the estimating equation to obtain the efficient estimator β,
In performing the nonparametric estimations , bandwidths need to be selected. Because the final
estimator is insensitive to the bandwidths, as indicated by condition C2, Lemma 1, Theorems 1
and 2, where a range of different bandwidths all lead to the same asymptotic property, we suggest
that one should select the corresponding bandwidths by taking the sample size n to its suitable
power to satisfy C2, and then multiply a constant to scale it. For example, when d “ 1, we let
r T Xi , let b be n´1{4`1{8 multiplying the
h be n´1{4´1{32 multiplying the standard deviation of β
standard deviation of Zi .
3. Asymptotics. We will show that the efficient estimator described in Section 2 is rootn consistent, asymptotically normally distributed and achieves the optimal efficiency. Let the
parameter space of β be B. We first list some regularity conditions.
C1 (The kernel function.) The univariate kernel function Kpxq is symmetric, monotonically
ş j
decreasing when x ą 0şand differentiable,ş with bounded derivative.
In addition,
x Kpxqdx “
ş
ş
ν Kpxqdx ă 8, K 2 pxqdx ă 8, x2 K 2 pxqdx ă 8, K 12 pxqdx ă 8,
0,
for
1
ď
j
ă
ν,
0
ă
x
ş 2 12
ş
ş
x K pxqdx ă 8, K 22 pxqdx ă 8, x2 K 22 pxqdx ă 8. The d-dimension kernel function is
ś
a product of d univariate kernel functions, that is Kpuq “ dj“1 Kpuj q for u “ pu1 , ..., ud qT .
For simplicity, we use the same K for both univariate and multivariate kernel functions.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
7
C2 (The bandwidths.) The bandwidths satisfy h Ñ 0, b Ñ 0, nhd`2 b Ñ 8 and nh2ν Ñ 0, where
2ν ą d ` 1.
C3 (The boundedness.) The parameter space B is bounded.
C4 (The density functions of covariates.) For all β P B, the probability density function of
β T X, fβT X pβ T xq, has a compact support and has four derivatives. The function fβT X pβ T xq
is bounded away from zero and infinity on the support, and its first four derivatives are
bounded uniformly on the support.
C5 (The smoothness.) For all β P B, the absolute value of EtXj IpZj ě Zq | β T xu, EtIpZj ě Zq |
β T xu, and their first four derivatives are bounded uniformly component wise. The absolute
T
value of EtXj XT
j IpZj ě Zq | β xu and its first two derivatives are bounded uniformly
component wise.
C6 (The survival function.) For all β P B, the survival function of the event process Spt, β T xq, the
conditional expectation of the survival function of the censoring processes EtSc pt, Xq | β T xu
and the probability density function of the survival process f pt, β T xq satisfy Bi`j Spt, β T xq{Bti Bpβ T xqj ,
Bi`j EtSc pt, Xq | β T xu{Bti Bpβ T xqj , Bi`j f pt, β T xq{Bti Bpβ T xqj exist and are bounded and
bounded away from zero, for all i ě 0, j ě 0, i ` j ď 4.
C7 (The uniqueness.) The equation
«
˜
( ff¸
E Xl Y pZq | β T X
λ1 pZ, β T Xq
(
“0
b Xl ´
E ∆
λpZ, β T Xq
E Y pZq | β T X
has a unique solution on B. Because the true parameter β 0 satisfies the equation, hence the
unique solution is β 0 .
These conditions are quite commonly imposed in nonparametrics, survival analysis and estimating equations and are generally mild. Conditions C1 and C2 contain some basic requirements on
the kernel function and the bandwidths, which are common in kernel related works. The boundedness of the parameter space B in C3 is satisfied in most practical problems. Condition C4-C6
impose certain boundedness condition of event time, censoring time, covariates, their expectations
and corresponding derivatives. The unique solution requirement in C7 is needed to ensure the
convergence of the estimator, which could be further relaxed to local uniqueness if needed.
Before presenting the main results, we summarize several preliminary results first. These results
highlight the theoretical properties of the kernel based estimators of several conditional expectations, as well as the estimation properties of the hazard function and its derivative, hence are of
their own interest. These properties also play an important role in the proof of the Theorems 1
and 2. The proofs of Lemma 1 and both theorems are respectively in the Appendice A.2, A.3 and
A.4.
Lemma 1.
(12)
(13)
(14)
(15)
(16)
(17)
Under the regularity conditions C1-C7 listed above,
(
p Y pZq | β T X “ EtY pZq | β T Xu ` Op tpnhq´1{2 ` h2 u,
E
(
p XY pZq | β T X “ EtXY pZq | β T Xu ` Op tpnhq´1{2 ` h2 u,
E
(
B
B p
E Y pZq | β T X “
EtY pZq | β T Xu ` Op tpnh3 q´1{2 ` h2 u
T
T
Bβ X
Bβ X
(
B p
B
E XY pZq | β T X “
EtXY pZq | β T Xu ` Op tpnh3 q´1{2 ` h2 u
T
Bβ X
Bβ T X
p β T xq “ λpz, β T xq ` Op tpnhbq´1{2 ` h2 ` b2 u
λpz,
p 1 pz, β T xq “ λ1 pz, β T xq ` Op tpnbh3 q´1{2 ` h2 ` b2 u
λ
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
8
uniformly for all Z, β T X.
p ´ β Ñ 0 in probaTheorem 1. The estimator obtained from solving (9) is consistent, i.e. β
0
bility when n Ñ 8.
Theorem 2.
The estimator obtained from solving (9) satisfies
?
p ´ β q Ñ N p0, rEtSb2 p∆, Z, Xqus´1 q
npβ
0
eff
in distribution when n Ñ 8. Here Seff p∆, Z, Xq is given in (4). Thus, the estimator is efficient.
Note that because Seff is a martingale, we have
EtSb2
eff p∆, Z, Xqu
«
ż8˜
λ10 ps, β T
0 Xq
“ E
b Xl ´
λ0 ps, β T
0
0 Xq
«
ż8˜
λ10 ps, β T
Xq
0
b Xl ´
“ E
T
λ
ps,
β
0
0
0 Xq
E Xl Sc ps, Xq | β T
0X
(
T
E Sc ps, Xq | β 0 X
E Xl Sc ps, Xq | β T
0X
(
T
E Sc ps, Xq | β 0 X
( ff¸b2
( ff¸b2
λps, β T
0 XqY psqds
dN psq.
Therefore, a natural estimator of the estimation variance is the inverse of
) fi˛b2
!
¨
»
T
p T xi
p Xl Sc pzi , Xq | β
n
E
ÿ
p
p
1
λ1 pzi , β xi q –
) fl‚ .
!
δi ˝
b xil ´
n i“1
p i, β
p T xi q
p T xi
p Sc pzi , Xq | β
λpz
E
4. Numerical Experiments.
4.1. Simulation. To evaluate the finite sample performance of our method, we perform four
simulation studies. In the first study, we generate event times from
“
`
˘
(
‰
T “ Φ 5ǫ exp 3β T X ` 1 ´ 2
where Φ is the cumulative distribution function (cdf) of the standard normal distribution, ǫ has
an exponential distribution with parameter 1, and X follows a standard normal distribution independent with ǫ. We consider d “ 1, p “ 7 and the true parameter values are taken to be
β “ p1, 0, ´1, 0, 1, 0, ´1qT . We further generate the covariate dependent censoring times using
C “ Φp2X2 ` 2X3 q ` U where U denotes a random variable uniformly distributed on p0, c1 q, where
c1 is a constant controlling the proportion of censoring.
In the second study, we generate the event times from
T “ exppβ T X ` ǫq
where ǫ follows an exponential distribution with parameter 1 and each component in X follows
independent uniform distribution on p0, 1q. We consider d “ 1, p “ 7 and set the true parameter
value to be β “ p1, 1.3, ´1.3, 1, ´0.5, 0.5, ´0.5qT . We generate the censoring time from a uniform
distribution on p0, c2 q, where different values of c2 are used to achieve various censoring rate.
In the third study, we generate the event times from
(
T “ exp 5 ´ 10p1 ´ β T Xq2 ` ǫ
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
9
where ǫ „ Normalp0, 1q, and each component of X is independently distributed with uniform
distribution on p0, 1q. We consider d “ 1, p “ 10 and set the true parameter value to be β “
p1, ´0.6, 0, ´0.3, ´0.1, 0, 0.1, 0.3, 0, 0.6qT . The censoring time is generated from C “ U β T
c X where
T
β c “ p0, 0, 0, 1, 1, 0, 0, 0, 0, 0q and U is uniformly distributed on p0, c3 q, and c3 is a constant
controlling the censoring proportion.
In the last simulation study, we increase d to 2 to further evaluate the performance of the
proposed method. We set the event times
+
#
2
ÿ
2
p1 ´ β T
T “ exp 5 ´ 10
j Xq ` ǫ
j“1
where ǫ „ Normalp0, 1q and each component of X is independently distributed with uniform
distribution on p0, 1q, β j , j “ 1, 2, denotes the jth column of β with p “ 6. The censoring time is
generated from a uniform distribution on p0, c4 q, where c4 controls the censoring rate.
These studies are designed to resemble and extend the simulation studies considered in Xia, Zhang and Xu
(2010), which proposed hmave, the best method so far in the literature for dimension reduction
under censored data. In fact, we compared our results with those from hmave. In all the studies,
we generated 1000 data sets. In the first study, sample size n “ 100 was considered. We set the
sample sizes to n “ 500 for the second study and to n “ 200 for all the remaining studies.
The results of the first simulation study are given in Table 1 and Figure 1, where we considered
three different censoring rates, 0%, 20% and 40% separately. From these results, we can see that
the semiparametric method we proposed generally performs better, and often is much better than
hmave, in that it has much smaller absolute biases as well as smaller sample standard errors. The
semiparametric method also yields smaller difference between the estimated projection matrix
p β
p T βq
p ´1 β
p T and the true projection matrix P ” βpβ T βq´1 β T , in that both the mean and
p ” βp
P
p ´ P are much smaller based on the semiparametric
variance of the largest singular value of P
method than based on hmave. The same results are also presented in Figure 1 to provide a quick
visual inspection.
Table 1
p ´ β of
Results of study 1, based on 1000 simulations with sample size 100. “bias” is the absolute value of meanpβq
each component in β, “sd” is the sample standard errors of the corresponding estimation. The last column is the
p ´ P.
mean and standard errors of the largest singular value of P
hmave
semi
hmave
semi
hmave
semi
true
β2
0
β3
´1
bias
sd
bias
sd
0.0010
0.1700
0.0025
0.1298
0.0256
0.2344
0.0129
0.1333
bias
sd
bias
sd
0.0747
0.1688
0.0003
0.1301
0.0994
0.2236
0.0143
0.1339
bias
sd
bias
sd
0.0056
0.1812
0.0012
0.1345
0.0261
0.2502
0.0130
0.1352
β4
β5
0
1
No censoring
0.0099 0.0058
0.1710 0.2320
0.0071 0.0059
0.1314 0.1277
20% censoring
0.0095 0.0042
0.1663 0.2281
0.0064 0.0079
0.1300 0.1268
40% censoring
0.0078 0.0079
0.1784 0.2462
0.0064 0.0090
0.1353 0.1305
β6
0
β7
´1
λmax
0.0052
0.1643
0.0071
0.1337
0.0182
0.2264
0.0033
0.1335
0.2208
0.1578
0.0903
0.0622
0.0099
0.1612
0.0055
0.1320
0.0228
0.2217
0.0054
0.1294
0.2256
0.1560
0.0928
0.0574
0.0169
0.1707
0.0103
0.1351
0.0189
0.2416
0.0056
0.1354
0.2314
0.1604
0.0948
0.0694
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
10
β =0
β3=−1
0% censoring
2
β6=0
5
β7=−1
2
2
2
2
2
1.5
1.5
1.5
1.5
1.5
1.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
0
−0.5
−0.5
−0.5
−0.5
−0.5
−0.5
−1
−1
−1
−1
−1
−1
−1.5
−1.5
−1.5
−1.5
−1.5
−1.5
−2
hmave
−2
semi
β2=0
20% censoring
β =1
β4=0
2
hmave
−2
semi
hmave
β3=−1
−2
semi
hmave
−2
semi
hmave
β5=1
β4=0
−2
semi
2
2
2
2
2
1.5
1.5
1.5
1.5
1.5
1.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
0
−0.5
−0.5
−0.5
−0.5
−0.5
−0.5
−1
−1
−1
−1
−1
−1
−1.5
−1.5
−1.5
−1.5
−1.5
−1.5
hmave
semi
−2
hmave
−2
hmave
β3=−1
β2=0
40% censoring
semi
semi
−2
hmave
β4=0
semi
−2
hmave
β5=1
semi
−2
hmave
β6=0
2
2
2
2
2
1.5
1.5
1.5
1.5
1.5
1.5
1
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
0
−0.5
−0.5
−0.5
−0.5
−0.5
−0.5
−1
−1
−1
−1
−1
−1
−1.5
−1.5
−1.5
−1.5
−1.5
−1.5
hmave
semi
−2
hmave
semi
−2
hmave
semi
−2
hmave
semi
−2
hmave
semi
β7=−1
2
−2
semi
β7=−1
β6=0
2
−2
hmave
semi
−2
hmave
Fig 1. Boxplot of hmave and the semiparametric methods of study 1. First row: no censoring; Second row: 20%
censoring rate; Third row: 40% censoring rate. Dashed line: True β.
The results of the second study study are presented in Tables 2, and Figures 2, where the same
conclusion can be drawn as in the first study. The superiority of the semiparametric method to
hmave is even more prominent in the third study, as reflected in Table 3 and Figure 3. Here, the
semiparametric method is substantially more accurate in estimating each component in β, yielding
smaller biases and variances. The largest singular value of the difference between the estimated and
true projection matrices is also much smaller for the semiparametric method in comparison with
hmave. When we increased d to 2 in the last simulation, the semiparametric method continues to
generate satisfactory results, see Table 4 and Figure 4. In this case, the performance of hmave is
rather concerning, possibly caused by the difficulties associated with multiple indices.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
semi
11
Table 2
p ´ β of
Results of study 2, based on 1000 simulations with sample size 200. “bias” is the absolute value of meanpβq
each component in β, “sd” is the sample standard errors of the corresponding estimation. The last column is the
p ´ P.
mean and standard errors of the largest singular value of P
hmave
semi
hmave
semi
hmave
semi
true
β2
1.3
β3
´1.3
bias
sd
bias
sd
0.1958
7.0919
0.2483
5.3549
0.3428
7.6671
0.0617
4.3952
bias
sd
bias
sd
0.5650
4.2947
0.0864
1.9872
0.5097
3.4980
0.0289
0.5161
bias
sd
bias
sd
0.0998
0.4603
0.0158
0.3976
0.1024
0.4678
0.0107
0.4678
β4
β5
1
´0.5
No censoring
0.1644 0.0137
9.8179 6.2261
0.0929 0.2354
2.5283 5.9618
20% censoring
0.3841 0.1947
2.4377 1.5136
0.0167 0.0109
1.2467 0.8233
40% censoring
0.0566 0.0381
0.3877 0.3085
0.0053 0.0128
0.3492 0.5517
β6
0.5
β7
´0.5
λmax
0.3699
10.976
0.1273
3.2538
0.1672
5.6033
0.0646
1.9437
0.4247
0.1442
0.2915
0.1133
0.2155
2.2722
0.0537
0.9729
0.2656
2.4448
0.1300
3.8902
0.3212
0.1356
0.1448
0.0909
0.0386
0.2939
0.0063
0.2718
0.0371
0.3085
0.0062
0.2873
0.1991
0.1303
0.0781
0.0720
Table 3
p ´ β of
Results of study 3, based on 1000 simulations with sample size 200. “bias” is the absolute value of meanpβq
each component in β, “sd” is the sample standard errors of the corresponding estimation. The last column is the
p ´ P.
mean and standard errors of the largest singular value of P
hmave
semi
hmave
semi
hmave
semi
true
β2
´0.6
β3
0
β4
-0.3
bias
sd
bias
sd
0.3711
17.267
0.0124
0.1639
0.3643
10.472
0.0044
0.1538
1.3056
35.110
0.0175
0.1523
bias
sd
bias
sd
1.4974
41.451
0.0035
0.1600
2.3355
44.735
0.0003
0.1584
1.6021
66.302
0.0239
0.1722
bias
sd
bias
sd
0.5909
20.946
0.0209
0.1500
5.5304
146.58
0.0004
0.1514
3.2835
68.000
0.0198
0.1518
β5
-0.1
β6
β7
0
0.1
No censoring
0.5394 0.0913 0.1954
16.246
9.449
16.931
0.0035 0.0026 0.0111
0.1563 0.1585 0.1535
20% censoring
0.4699 2.3620 1.6553
49.280 40.673 47.488
0.0063 0.0023 0.0072
0.1716 0.1555 0.1615
40% censoring
0.8370
1.775
5.6482
25.442 38.877 145.82
0.0062 0.0041 0.0138
0.1506 0.1492 0.1560
β8
0.3
β9
0
β10
0.6
λmax
0.4499
26.510
0.0323
0.1590
0.2092
12.342
0.0013
0.1543
0.3660
17.304
0.0180
0.1631
0.8706
0.2953
0.2337
0.0637
0.5311
58.485
0.0184
0.1531
1.6596
57.025
0.0035
0.1633
2.9426
72.228
0.0177
0.1595
0.8822
0.2952
0.2148
0.0691
4.8272
90.530
0.0244
0.1453
1.2442
58.178
0.0021
0.1503
0.6951
23.032
0.0166
0.1579
0.8382
0.2900
0.2656
0.0556
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
12
Table 4
p ´ β of
Results of study 4, based on 1000 simulations with sample size 200. “bias” is the absolute value of meanpβq
each component in β, “sd” is the sample standard errors of the corresponding estimation. The last column is the
p ´ P.
mean and standard errors of the largest singular value of P
hmave
semi
hmave
semi
hmave
semi
true
β3,1
2.75
β4,1
-0.75
β5,1
-1
bias
sd
bias
sd
5.8883
130.6
0.1609
0.3393
3.4252
109.3
0.0914
0.2143
0.4704
24.214
0.0609
0.2545
bias
sd
bias
sd
4.2011
64.499
0.0846
0.3433
2.8472
39.786
0.0358
0.2068
1.7685
56.389
0.0311
0.2018
bias
sd
bias
sd
3.2328
14.712
0.0986
0.3604
0.5529
19.789
0.0555
0.2173
2.0826
17.661
0.0246
0.2168
β6,1
β3,2
2.0
-3.125
No censoring
2.9062 43.257
101.81 1091.9
0.1388 0.1257
0.2837 0.3674
20% censoring
3.4267 3.5963
41.570 20.496
0.0596 0.1133
0.2646 0.4051
40% censoring
1.0118 4.2734
19.938 26.892
0.0808 0.1420
0.2991 0.4864
β4,2
-1.125
β5,2
1.0
β6,2
-2.0
λmax
38.079
939.06
0.1002
0.2619
8.0696
174.9
0.0372
0.2559
33.551
855.8
0.0887
0.3121
0.8351
0.1116
0.1791
0.0788
0.4313
24.564
0.0726
0.2690
0.6443
23.323
0.0305
0.2374
0.9892
22.339
0.0724
0.2952
0.9273
0.1038
0.0971
0.0777
3.6817
73.550
0.0868
0.2627
2.0900
29.838
0.0451
0.2645
2.9417
31.115
0.0950
0.3099
0.9363
0.1071
0.0915
0.0898
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
13
β =1.3
β =−1.3
0% censoring
2
β =0.5
5
β =−0.5
6
7
3
3
3
3
2
2
2
2
2
2
1
1
1
1
1
1
0
0
0
0
0
0
−1
−1
−1
−1
−1
−1
−2
−2
−2
−2
−2
−2
−3
hmave
semi
−3
hmave
β =1.3
semi
−3
hmave
β =−1.3
2
20% censoring
β =−0.5
4
3
−3
semi
β =1
3
−3
hmave
−3
semi
hmave
β =−0.5
4
semi
hmave
β =0.5
5
7
3
3
3
3
3
2
2
2
2
2
2
1
1
1
1
1
1
0
0
0
0
0
0
−1
−1
−1
−1
−1
−1
−2
−2
−2
−2
−2
−2
−3
hmave
semi
−3
hmave
β =1.3
semi
β =−1.3
semi
−3
hmave
β =1
3
2
−3
hmave
semi
−3
hmave
β =−0.5
4
semi
hmave
β =0.5
5
7
3
3
3
3
3
2
2
2
2
2
2
1
1
1
1
1
1
0
0
0
0
0
0
−1
−1
−1
−1
−1
−1
−2
−2
−2
−2
−2
−2
−3
hmave
semi
−3
hmave
semi
−3
hmave
semi
−3
hmave
semi
semi
β =−0.5
6
3
−3
semi
β =−0.5
6
3
−3
40% censoring
β =1
3
3
−3
hmave
semi
hmave
Fig 2. Boxplot of hmave and the semiparametric methods of study 2. First row: no censoring; Second row: 20%
censoring rate; Third row: 40% censoring rate. Dashed line: True β.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
semi
14
β =−0.6
β =0
0% censoring
2
β =0
5
β =0.1
6
β =0.3
7
β =0
8
β =0.6
9
10
4
4
4
4
4
4
4
3
3
3
3
3
3
3
3
3
2
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
−1
−1
−1
−1
−1
−1
−1
−1
−1
−2
−2
−2
−2
−2
−2
−2
−2
−2
−3
−3
−3
−3
−3
−3
−3
−3
−3
−4
−4
−4
−4
−4
−4
−4
−4
hmave
semi
hmave
semi
β3=0
2
20% censoring
β =−0.1
4
4
β =−0.6
hmave
semi
β4=−0.3
hmave
semi
β5=−0.1
hmave
semi
β6=0
hmave
semi
β7=0.1
hmave
semi
β8=0.3
hmave
semi
−4
β9=0
4
4
4
4
4
4
4
4
3
3
3
3
3
3
3
3
3
2
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
−1
−1
−1
−1
−1
−1
−1
−1
−1
−2
−2
−2
−2
−2
−2
−2
−2
−2
−3
−3
−3
−3
−3
−3
−3
−3
−3
−4
−4
−4
−4
−4
−4
−4
−4
hmave
semi
hmave
semi
β =0
2
hmave
semi
β =−0.3
3
hmave
semi
β =−0.1
4
hmave
semi
β =0
5
hmave
semi
β =0.1
6
hmave
semi
β =0.3
7
hmave
semi
−4
β =0
8
4
4
4
4
4
4
4
3
3
3
3
3
3
3
3
3
2
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
0
−1
−1
−1
−1
−1
−1
−1
−1
−1
−2
−2
−2
−2
−2
−2
−2
−2
−2
−3
−3
−3
−3
−3
−3
−3
−3
−3
−4
−4
−4
−4
−4
−4
−4
−4
hmave
semi
hmave
semi
hmave
semi
hmave
semi
hmave
semi
hmave
semi
hmave
semi
semi
10
4
semi
hmave
β =0.6
9
4
hmave
hmave semi
β10=0.6
4
β =−0.6
40% censoring
β =−0.3
3
4
−4
hmave semi
Fig 3. Boxplot of hmave and the semiparametric methods of study 3. First row: no censoring; Second row: 20%
censoring rate; Third row: 40% censoring rate. Dashed line: True β.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
15
β =2.75
β =−0.75
0% censoring
2
β =2
4
β =−3.125
5
β =−1.125
6
β =1
7
β =−2
8
9
8
8
8
8
8
8
8
6
6
6
6
6
6
6
6
4
4
4
4
4
4
4
4
2
2
2
2
2
2
2
2
0
0
0
0
0
0
0
0
−2
−2
−2
−2
−2
−2
−2
−2
−4
−4
−4
−4
−4
−4
−4
−4
−6
−6
−6
−6
−6
−6
−6
−6
−8
−8
−8
−8
−8
−8
−8
hmave
semi
hmave
β =2.75
semi
hmave
β3=−0.75
2
20% censoring
β =−1
3
8
semi
hmave
β4=−1
semi
hmave
β5=2
semi
hmave
β6=−3.125
semi
−8
hmave
β7=−1.125
semi
hmave
β8=1
β9=−2
5
5
5
5
5
5
5
5
4
4
4
4
4
4
4
4
3
3
3
3
3
3
3
3
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
−1
−1
−1
−1
−1
−1
−1
−1
−2
−2
−2
−2
−2
−2
−2
−2
−3
−3
−3
−3
−3
−3
−3
−3
−4
−4
−4
−4
−4
−4
−4
−5
hmave
semi
−5
β =2.75
semi
−5
β =−0.75
2
40% censoring
hmave
hmave
semi
−5
β =−1
3
hmave
semi
−5
β =2
4
hmave
semi
−5
β =−3.125
5
hmave
semi
−5
β =−1.125
6
−4
hmave
semi
−5
β =1
7
5
5
5
5
5
5
4
4
4
4
4
4
4
3
3
3
3
3
3
3
3
2
2
2
2
2
2
2
2
1
1
1
1
1
1
1
1
0
0
0
0
0
0
0
0
−1
−1
−1
−1
−1
−1
−1
−1
−2
−2
−2
−2
−2
−2
−2
−2
−3
−3
−3
−3
−3
−3
−3
−3
−4
−4
−4
−4
−4
−4
−4
semi
−5
hmave
semi
−5
hmave
semi
−5
hmave
semi
−5
hmave
semi
−5
hmave
semi
−5
semi
9
5
4
hmave
hmave
β =−2
8
5
−5
semi
−4
hmave
semi
−5
hmave
Fig 4. Boxplot of hmave and the semiparametric methods of study 4. First row: no censoring; Second row: 20%
censoring rate; Third row: 40% censoring rate. Dashed line: True β.
We also performed an additional experiment to further assess the finite sample performance
of the asymptotic results established in Section 3. To this end, we generate covariates X from a
standard normal distribution and event times T from a distribution with hazard function
+
#
2
ÿ
` T ˘
λpt|Xq “ λ0 ptq
exp β j X ,
j“1
where the baseline hazard λ0 ptq “ t and the dimension of β is d “ 2, p “ 6. We use the parameter values β “ tp2.75, ´0.75, ´1, 2qT ; p´3.125, ´1.125, 1, ´2qT uT , and adopt the same censoring
process as in the second study to yield 40% censoring rate. We carry out 1000 simulations and
consider sample sizes n “ 100, 500 and 1000. The estimation results, together with sample standard errors, average of the estimated standard deviations and coverage probabilities of the 95%
confidence intervals are given in Table 5. These results indicate that the large sample properties
of the estimator requires more sample size than 1000. However, the general trend is that when
sample size increases, the results are approaching what we expect based on the asymptotic results,
in that the sample standard errors and their estimated versions are becoming closer to each other,
and the 95% coverage probabilities are getting closer to the nominal level. The phenomenon that
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
semi
16
asymptotic result requires very large sample size to illustrate itself is quite common in survival
data analysis and is not unique to our semiparametric method. Due to the limited sample size in
practice, we recommend to use bootstrap to assess estimation variability.
Table 5
Results of study 5, based on 1000 simulations with sample size 100, 500, 1000 respectively. “bias” is the absolute
p ´ β of each component in β, “sd” is the sample standard errors of the corresponding estimation,
value of meanpβq
p component, “95%” is the sample coverage of the 95%
“p
σ ” is the mean of the estimated standard errors of β
confidence intervals.
β31
2.75
β41
-0.75
β51
-1
bias
sd
σ
p
95%
0.3995
0.5760
0.3868
0.7100
0.5031
0.4236
0.3188
0.6577
0.2066
0.4673
0.3312
0.8051
0.1790
0.2741
0.1585
0.6663
0.1258
0.1714
0.1371
0.8022
0.0714
0.2177
0.1644
0.8298
bias
sd
σ
p
95%
0.0611
0.1951
0.1062
0.8060
0.0492
0.1451
0.1113
0.8830
0.0188
0.1555
0.1134
0.8783
bias
sd
σ
p
95%
β61
β32
2.0
-3.125
n “ 100
0.3799 0.5515
0.5608 0.6163
0.3427 0.3956
0.7034 0.6416
n “ 500
0.1338 0.2100
0.2380 0.2979
0.1659 0.2683
0.7566 0.8127
n “ 1000
0.0423 0.0695
0.1538 0.1867
0.1190 0.1823
0.8621 0.9268
β42
-1.125
β52
1.0
β62
-2.0
0.5349
0.4376
0.3131
0.6292
0.1757
0.4772
0.3331
0.8089
0.3395
0.5377
0.3602
0.7414
0.1489
0.1897
0.2179
0.8773
0.07386
0.2202
0.2538
0.9125
0.1340
0.2244
0.2558
0.8764
0.0467
0.1433
0.1712
0.9705
0.0209
0.1650
0.1749
0.9515
0.0448
0.1711
0.1740
0.9287
4.2. AIDS Application. We apply the proposed method to analyze the HIV data from AIDS
Clinical Trials Group Protocol 175 (ACTG175) (Hammer et al. (1996)). In this study, 2137 HIVinfected subjects were randomized to receive one of four treatments: zidovudine (ZDV) monotherapy, ZDV plus didanosine, ZDV plus zalcitabine and ddI monotherapy. As in Geng, Lu and Zhang
(2015) and Jiang et al. (2017), the survival time of interest was chosen as the time to having a
larger than 50% decline in the CD4 count, or progressing to AIDS or death, whichever comes first.
Besides the treatments, there were 12 covariates included in our study, specifically, patient age
in years at baseline (X1 ), patient weight in kilogram at baseline (X2 ), hemophilia indicator (X3 ),
homosexual activity (X4 ), history of IV drug use (X5 ), Karnofsky score on a scale of 0-100 (X6 ),
race (X7 ), gender (X8 ), antiretroviral history (X9 ), symptomatic indicator (X10 ), number of CD4
at baseline (X11 ), number of CD8 at baseline (X12 ), treatment indicator (X13 ), where we coded
X13 “ 0 for treatment ZDV+ddl and X13 “ 1 for treatment ZDV+Zal. As in Jiang et al. (2017),
we only analyzed data from the two composite treatments: ZDV plus didanosine and ZDV plus
zalcitabine, which had been shown to have significantly better survival than the other two treatments (Geng, Lu and Zhang, 2015). This subset of data contains 1046 subjects with the censoring
rate around 75%. In addition, each covariate is standardized respectively with no obvious outliers
and no missing values.
To determine the proper reduced space dimension d, we employ the Validated Information
Criterion (VIC) (Ma et al., 2015), where the d corresponding to the smallest VIC value is selected.
In the example, the VIC value at d “ 1 is 90.38. Further, when d ě 2, the VIC values are all greater
than 180.7 which result from the penalty term alone. Hence we choose d “ 1 as the final model.
p under the selected model, with the corresponding
Table 6 contains the estimated coefficient β’s
estimation standard errors and p-values. Here, we implemented the semiparametric estimator to
obtain these results due to its superior theoretical and numerical performance.
p , all covariates are
The results in Table 6 indicate that in forming the index described by β
¨,1
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
17
significant except hemophilia indicator (X3 ), gender (X8 ) and number of CD4 at baseline (X11 ).
The estimated cumulative hazard functions are also reported in Figure 5, where it is plotted as
a function of time (upper left panel), a function of the covariate index β T x (upper right panel)
and as a function of both (bottom panel). Specifically, in plotting the cumulative hazard as a
function of time t, we fixed the covariate index at three different sets of covariate values, respectively X1:12 “ p40, 60, 1, 0, 0, 80, 0, 0, 0, 1, 200, 800qT , X1:12 “ p20, 70, 1, 0, 0, 80, 0, 1, 0, 1, 200, 800qT ,
X1:12 “ p60, 70, 1, 0, 0, 20, 0, 0, 0, 1, 200, 200qT , in combination with the treatment indicator of both
X13 “ 0 and X13 “ 1. Based on the plots, the estimated cumulative hazard of treatment ZDV+ddl
is slightly larger than that of treatment ZDV+Zal in all scenarios. In plotting the estimated cup T x, we fixed the time at t “ 100, 500 and 1000.
p as a function of the index β
mulative hazard Λ
Finally, we also plotted the cumulative hazard as a function of two variables t and β T x using the
contour plot, where the hazard values are explicitly written out on each contour.
Table 6
The estimated coefficients, standard errors and p-value in AIDS data.
est
std
p-value
p
β
2,1
0.115
0.039
0.003
p
β
3,1
-0.002
0.039
0.965
p
β
4,1
0.093
0.039
0.017
p
β
5,1
0.088
0.037
0.017
p
β
6,1
-0.090
0.043
0.036
p
β
7,1
0.231
0.046
0.001
p
β
8,1
-0.003
0.036
0.928
p
β
9,1
-0.178
0.046
0.001
p
β
10,1
0.058
0.035
0.100
p
β
11,1
-0.031
0.042
0.457
p
β
12,1
0.201
0.033
0.001
p
β
13,1
0.156
0.038
0.001
5. Discussion. We have considered a very general model for analyzing time to event data
subject to censoring. The model allows the event times to link to the covariate indices in an
unspecified fashion. Because both the number of indices and the functional form of the linkage to
the indices are data determined, conceptually the model is maximally flexible. In practice, relatively
low number of indices are expected to avoid curse of dimensionality. The work is conducted without
requiring covariate independent censoring. Instead, it only requires event independent censoring
conditional on covariates, which is the minimum requirement for identification. We derived a class
of estimators which are consistent and asymptotically normal. We also proposed a procedure to
construct the semiparametric efficient estimator that achieves the optimal estimation variability
among all possible consistent estimators.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
18
25
8
Estimated cumulative hazard function
7
6
Estimated cumulative hazard function
Scenario_1, ZDV + ddI
Scenario_2, ZDV + ddI
Scenario_3, ZDV + ddI
Scenario_1, ZDV + Zal
Scenario_2, ZDV + Zal
Scenario_3, ZDV + Zal
5
4
3
2
T=100
T=500
T=1000
20
15
10
5
1
0
200
400
600
Event time
(a)
800
1000
0
−6
1200
−4
−2
0
5
2
12
2
2
4
10
4
0
2
T
β X
1
6
6
4
−2
8
6
8
−3
6
8
10
4
8
12
10
200
12
14
14
8
10
10
62 4
−5
6
14
2
3
−4
4
16
4
−1
2
T
βX
(b)
16
400
600
Event time
(c)
Estimated cumulative hazard function
0
800
16
1000
2
p in AIDS data. (a). Comparisons of Λ
p as a function of t between
Fig 5. Estimated cumulative hazard function Λ
p as a function of β
p T X at
treatments ZDV+ddl and ZDV+Zal when other covariates are fixed at three indices. (b). Λ
p T X.
p as a function of T and β
T “ 100, 500, 1000. (c). Contour plot of Λ
Appendix.
A.1. Proof of Proposition 1. The result of Λ1 is obvious. To obtain Λ2 , let hpt, β T
0 X, γq “
T
T
Xq.
Hence
X,
γq
is
a
submodel
of
λpt,
β
X,
γq{Bγ,
where
λpt,
β
Blogλpt, β T
0
0
0
Blogf pX, Z, ∆q
Bγ
żZ
Bλps, β T
BlogλpZ, β T
0 X, γq
0 X, γq
“ ∆
´
ds
γ
Bγ
0
żZ
T
T
hps, β T
“ ∆hpZ, β 0 X, γq ´
0 X, γqλps, β 0 Xqds
0
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
19
“
ż8
0
T
hps, β T
0 X, γqdM ps, β 0 Xq.
T
Because λ0 pt, β T
0 Xq can be any positive function, hps, β 0 X, γq can be any function. We denote it
hps, β T
0 Xq. This leads to the form of Λ2 .
Similar derivation leads to Λ3 .
It is easy to verify that Λ1 K Λ2 and Λ1 K Λ3 . Because C T | X, the martingale integrations
associated with M pt, β T
0 Xq and MC pz, Xq are also independent conditional on X, hence Λ2 K Λ3 .
This completes the proof.
A.2. Proof of Lemma 1. For notational convenience, we prove the results for d “ 1 and assume
the first component of β is 1. The first four results are obtained from the uniform convergence
property of the kernel estimation (Mack and Silverman, 1982) under conditions C1-C2. Specifically,
to derive the first four results, we first establish the following preliminary conclusion for any Z and
β T X,
(A.1)
n
1 ÿ
Kh pβ T Xj ´ β T xq “ fβT X pβ T xq ` Op pn´1{2 h´1{2 ` h2 q.
n j“1
To see this, we compute the absolute bias of the left hand side of (A.1) as
ˇ #
ˇ
+
n
ˇ
ˇ
1 ÿ
ˇ
T
T
T ˇ
E
Kh pβ Xj ´ β xq ´ fβT X pβ xqˇ
ˇ
ˇ
ˇ
n j“1
ˇ
ˇ
ˇ
ˇ
“ ˇEKh pβ T Xj ´ β T xq ´ fβT X pβ T xqˇ
ˇ
ˇż
¸
˜
ˇ
ˇ 1
β T xj ´ β T x
ˇ
ˇ
fβT X pβ T xj qdβ T xj ´ fβT X pβ T xqˇ
K
“ ˇ
ˇ
ˇ h
h
ˇż
ˇ
ˇ
ˇ
“ ˇˇ KpuqfβT X pβ T x ` huqdu ´ fβT X pβ T xqˇˇ
ˇż
ˇ
*
"
ˇ
ˇ
1 2
T ˚ 2 2
T ˇ
T
1
T
ˇ
“ ˇ Kpuq fβT X pβ xq ` fβT X pβ xqhu ` fβT X pβ x qh u du ´ fβT X pβ xqˇ
2
ż
2
h
sup |f 2T pβ T xq| u2 Kpuqdu,
ď
2 βT x β X
where throughout the text, β T x˚ is on the line connecting β T x and β T x ` hu, and the variance
to be
#
+
n
1 ÿ
T
T
var
Kh pβ Xj ´ β xq
n j“1
“
“
“
“
1
varKh pβ T Xj ´ β T xq
n
(2 ı
1”
EKh2 pβ T Xj ´ β T xq ´ EKh pβ T Xj ´ β T xq
n „ż
1
1 2
T
T
T
T
2
T
2
K tpβ xj ´ β xq{hufβ T x pβ xj qdβ xj ´ fβT x pβ xq ` Oph q
n
h2
ż
1
1
K 2 puqfβT x pβ T x ` huqdu ´ fβ2 T x pβ T xq ` Oph2 {nq
nh
n
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
20
ď
1
f T pβ T xq
nh β x
ż
K 2 puqdu `
h
sup |f 2T pβ T xq|
2n βT x β x
ż
u2 K 2 puqdu `
1 2
|f T pβ T xq| ` Oph2 {nq.
n β x
Therefore, we have
n
1 ÿ
Kh pβ T Xj ´ β T xq “ fβT X pβ T xq ` Op pn´1{2 h´1{2 ` h2 q
n j“1
uniformly under conditions C1, C2 and C4.
Following similar derivations, we have
#
+
n
1 ÿ
E ´
Xj IpZj ě ZqKh1 pβ T Xj ´ β T xq
n j“1
(
“ E ´Xj IpZj ě ZqKh1 pβ T Xj ´ β T xq
ż
(
1
xj Ipc ě ZqIpt ě ZqK 1 pβ T xj ´ β T xq{h
“ ´
2
h
ˆfc pc, xj qf pt, β T xj qfX|βT X pxj | β T xj qfβT X pβ T xj qdcdtdxj dβ T xj
ż
(
1
“ ´
xj K 1 pβ T xj ´ β T xq{h Sc pZ, xjl , β T xj qSpZ, β T xj qfXl |βT X pxjl | β T xj q
2
h
ˆfβT X pβ T xj qdxjl dβ T xj
ż
1
T 1
T
T
T
pβ T x ` hu, xT
“ ´
jl q K puqSc pZ, xjl , β x ` huqSpZ, β x ` huqfXl |βT X pxjl | β x ` huq
h
ˆfβT X pβ T x ` huqdxjl du
ż
ż
1
T
T
T
T
T
pβ T x, xT
q
S
pZ,
x
,
β
xqSpZ,
β
xqf
px
|
β
xqf
pβ
xqdx
K 1 puqdu
“ ´
T
T
c
jl
jl
jl
Xl |β X jl
β X
h
ż
)
B ! T
1
T T
T
T
T
T
pβ
x,
x
q
S
pZ,
x
,
β
xqSpZ,
β
xqf
px
|
β
xqf
pβ
xq
dxjl
´
c
jl
jl
Xl |βT X jl
βT X
h Bβ T x
ˆhuK 1 puqdu
ż
!
)
1
B2
T
T T
T
T
T
T
pβ
x,
x
q
S
pZ,
x
,
β
xqSpZ,
β
xqf
px
|
β
xqf
pβ
xq
dxjl
´
T
T
c
jl
jl
Xl |β X jl
β X
2h Bpβ T xq2
ż
ˆ h2 u2 K 1 puqdu
ij
)
!
1
B3
T ˚
T ˚
T ˚
T ˚
T
T T
´
q
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
px
|
β
x
qf
pβ
x
q
dxjl
pβ
x,
x
T
T
c
jl
jl
Xl |β X jl
β X
6h
Bpβ T xq3
ˆh3 u3 K 1 puqdu
ż
)
B ! T
T T
T
T
T
T
pβ
x,
x
q
S
pZ,
x
,
β
xqSpZ,
β
xqf
px
|
β
xqf
pβ
xq
dxjl
“
T
T
c
jl
jl
jl
Xl |β X
β X
Bβ T x
ż ż
)
!
h2
B3
T ˚
T ˚
T ˚
T ˚
T
T T
q
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
pβ
x,
x
´
T pxjl | β x qf T pβ x q dxjl
c
jl
jl
X
|β
X
β
X
l
6
Bpβ T xq3
ˆu3 K 1 puqdu
B
“
SpZ, β T xqfβT X pβ T xqEtXj Sc pZ, Xj q | β T xu
Bβ T x
ż
(
B3
h2
SpZ, β T x˚ qfβT X pβ T x˚ qE Xj Sc pZ, Xj q | β T x˚ u3 K 1 puqdu
´
T
6
Bpβ xq3
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
21
“
B
fβT X pβ T xqEtXj IpZj ě Zq | β T xu
T
Bβ x
ż
(
h2
B3
´
fβT X pβ T x˚ qE Xj IpZj ě Zq | β T x˚ u3 K 1 puqdu.
T
6
Bpβ xq3
Hence the absolute bias is
ˇ #
ˇ
+
n
ˇ
ˇ
1 ÿ
B
ˇ
T
T ˇ
f
Xj IpZj ě ZqKh1 pβ T Xj ´ β T xq ´
T pβ xqEtXj pZj ě Zq | β xuˇ
ˇE ´
ˇ
ˇ
n j“1
Bβ T x β X
ˇ
ˇ 2ż
ˇ
ˇ h
( 3 1
B3
T ˚
T ˚
ˇ
ˇ
f
u
K
puqdu
“ ˇ´
T pβ x qE Xj IpZj ě Zq | β x
β
X
ˇ
6
Bpβ T xq3
ˇ
ˇ
"
*
ż
ˇ B3
(ˇ
h2
T
T
2
ˇ
ˇ
sup
ď
f T pβ xqE Xj IpZj ě Zq | β x ˇ 3 u Kpuqdu
6 βT x ˇ Bpβ T xq3 β X
and
+
n
1 ÿ
Xj IpZj ě ZqKh1 pβ T Xj ´ β T xq
var ´
n j“1
(T
1”
E Xj IpZj ě ZqKh1 pβ T Xj ´ β T xq ut Xj IpZj ě ZqKh1 pβ T Xj ´ β T xq
“
n
(T ı
(
´ EXj IpZj ě ZqKh1 pβ T Xj ´ β T xq EXj IpZj ě ZqKh1 pβ T Xj ´ β T xq
ˆż
(
1
1
12
“
xj xT
pβ T xj ´ β T xq{h
j IpZj ě ZqK
4
n
h
#
ˆfc pc, xj qf pt, β T xj qfXl |βT X pxjl | β T xj qfβT X pβ T xj qdcdtdxjl dβ T xj
¸
„
T
„
B
B
f T pβ T xqEtXj IpZj ě Zq | β T xu
f T pβ T xqEtXj IpZj ě Zq | β T xu ` Oph2 q
´
Bβ T x β X
Bβ T x β X
ż
1
1 T
T
T
T
T
“
pβ x ` hu, xT
jl q pβ x ` hu, xjl qSc pZ, xjl , β x ` huq
n h3
ˆSpZ, β T x ` huqfXl |βT X pxjl | β T x ` huqfβ T X pβ T x ` huqdxj K 12 puqdu
`Op1{nq
ż
1
T
T
T
T
T
T
T
“
pβ T x, xT
jl q pβ x, xjl qSc pZ, xjl , β xqSpZ, β xqfβT X pβ xqfXl |βT X pxjl | β xqdxjl
nh3
ż
ˆ K 12 puqdu
ij
1
B2
T ˚
T ˚
T
T ˚ T
`
pβ T x˚ , xT
jl q pβ x , xjl qSc pZ, xjl , β x qSpZ, β x q
2nh3 Bpβ T xq2
ˆfβT X pβ T x˚ qfXl |βT X pxjl | β T x˚ qdxjl h2 u2 K 12 puqdu ` Op1{nq
ˇ
ˇż
1
ˇ
ˇ
T
T
T
sup ˇfβT X pβ xqEtXj Xj IpZj ě Zq | β xuˇ K 12 puqdu
ď
3
nh βT x
ˇ
ˇż
ˇ B2
ˇ
1
T
˚
T
T
˚
`
sup ˇˇ
fβT X pβ x qEtXj Xj IpZj ě Zq | β x uˇˇ u2 K 12 puqdu ` Op1{nq.
T
2nh βT x˚ Bpβ xq2
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
22
So
n
1 ÿ
Xj IpZj ě ZqKh1 pβ T Xj ´ β T xq
n j“1
“ ´
(A.2)
B
f T pβ T xqEtXj IpZj ě Zq | β T xu ` Op pn´1{2 h´3{2 ` h2 q
Bβ T x β X
uniformly under conditions C1-C2.
We next show
´
(A.3)
n
1 ÿ 1 T
K pβ Xj ´ β T xq “ fβ1 T X pβ T xq ` Op pn´1{2 h´3{2 ` h2 q.
n j“1 h
To see this, we compute the absolute bias of the left hand side as
ˇ
ˇ #
+
n
ˇ
ˇ
1 ÿ 1 T
ˇ
ˇ
Kh pβ Xj ´ β T xq ´ fβT X pβ T xqˇ
ˇE ´
ˇ
ˇ
n j“1
ˇ
ˇ
(
ˇ
ˇ
“ ˇE ´Kh1 pβ T Xj ´ β T xq ´ fβT X pβ T xqˇ
ˇ
ˇż
˜
¸
ˇ
ˇ
1 1 β T xj ´ β T x
ˇ
T
T
T ˇ
fβT X pβ xj qdβ xj ´ fβT X pβ xqˇ
“ ˇ ´ 2K
ˇ
ˇ
h
h
ˇż
ˇ
ˇ
ˇ
1
“ ˇˇ ´ K 1 puqfβ T X pβ T x ` huqdu ´ fβT X pβ T xqˇˇ
h
ˇż
"
*
ˇ
1
1
1
p3q
T
˚
3
3
1
T
1
2
T
T
2
2
“ ˇˇ ´ K puq fβT X pβ xq ` fβT X pβ xqhu ` fβT X pβ xqh u ` fβT X pβ x qh u du
h
2
6
ˇ
ˇ
´fβT X pβ T xqˇ
ˇ
ˇ
ż
ˇ
ˇ 1
h2
p3q
T ˚ 3 1
T ˇ
T
ˇ
fβT X pβ x qu K puqdu ´ fβT X pβ xqˇ
“ ˇfβT X pβ xq ´
6
ż
2
h
p3q
T
ď
pβ xq| u2 Kpuqdu,
sup |f
2 β T x βT X
and the variance to be
#
+
n
1 ÿ 1 T
var
K pβ Xj ´ β T xq
n j“1 h
“
“
“
“
ď
1
varKh1 pβ T Xj ´ β T xq
n
(2 ı
1”
T
T
T
T
1
12
EK pβ Xj ´ β xq ´ EKh pβ Xj ´ β xq
n „ż h
1 12
1
T
T
T
T
2
T
2
K tpβ xj ´ β xq{hufβ T x pβ xj qdβ xj ´ fβT x pβ xq ` Oph q
n
h4
ż
1
1
K 12 puqfβ T x pβ T x ` huqdu ´ fβ2 T x pβ T xq ` Oph2 {nq
3
nh
n
ż
ż
1
1
2
T
12
T
sup |f T pβ xq| u2 K 12 puqdu ` Op1{nq.
f T pβ xq K puqdu `
nh3 β x
2nh βT x β x
Therefore, (A.3) holds uniformly under conditions C1, C2 and C4.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
23
We compute the expectation
+
#
n
1 ÿ
IpZj ě ZqKh1 pβ T Xj ´ β T xq
E ´
n j“1
(
“ E ´IpZj ě ZqKh1 pβ T Xj ´ β T xq
ż
(
1
“ ´
Ipc ě ZqIpt ě ZqK 1 pβ T xj ´ β T xq{h
2
h
ˆfc pc, xj qf pt, β T xj qfX|βT X pxj | β T xj qfβT X pβ T xj qdcdtdxj dβ T xj
ż
(
1 1
T
T
K
pβ
x
´
β
xq{h
Sc pZ, xjl , β T xj qSpZ, β T xj qfXl |βT X pxjl | β T xj q
“ ´
j
h2
ˆfβT X pβ T xj qdxjl dβ T xj
ż
1
“ ´
K 1 puqSc pZ, xjl , β T x ` huqSpZ, β T x ` huqfXl |βT X pxjl | β T x ` huq
h
ˆfβT X pβ T x ` huqdxjl du
ż
ż
1
“ ´
Sc pZ, xjl , β T xqSpZ, β T xqfXl |βT X pxjl | β T xqfβT X pβ T xqdxjl K 1 puqdu
h
ż
)
B !
1
T
T
T
T
S
pZ,
x
,
β
xqSpZ,
β
xqf
px
|
β
xqf
pβ
xq
dxjl
´
T
T
c
jl
Xl |β X jl
β X
h Bβ T x
ˆhuK 1 puqdu
ż
!
)
1
B2
T
T
T
T
´
S
pZ,
x
,
β
xqSpZ,
β
xqf
T pxjl | β xqf T pβ xq dxjl
c
jl
Xl |β X
β X
2h Bpβ T xq2
ż
ˆ h2 u2 K 1 puqdu
ij
!
)
1
B3
T ˚
T ˚
T ˚
T ˚
´
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
T pxjl | β x qf T pβ x q dxjl
c
jl
Xl |β X
β X
6h
Bpβ T xq3
ˆh3 u3 K 1 puqdu
ż
B
Sc pZ, xjl , β T xqSpZ, β T xqfXl |βT X pxjl | β T xqfβT X pβ T xqdxjl
“
Bβ T x
ż ż
h2
B3
´
S pZ, xjl , β T x˚ qSpZ, β T x˚ qfXl |βT X pxjl | β T x˚ qfβ T X pβ T x˚ qdxjl
T 3 c
6
Bpβ xq
3 1
ˆu K puqdu
B
“
SpZ, β T xqfβT X pβ T xqEtSc pZ, Xj q | β T xu
Bβ T x
ż
(
B3
h2
SpZ, β T x˚ qfβT X pβ T x˚ qE Sc pZ, Xj q | β T x˚ u3 K 1 puqdu
´
T
3
6
Bpβ xq
B
“
fβT X pβ T xqEtIpZj ě Zq | β T xu
T
Bβ x
ż
(
h2
B3
´
fβT X pβ T x˚ qE IpZj ě Zq | β T x˚ u3 K 1 puqdu.
T
6
Bpβ xq3
Hence the absolute bias is
ˇ
ˇ #
+
n
ˇ
ˇ
ÿ
B
1
ˇ
T
T ˇ
IpZj ě ZqKh1 pβ T Xj ´ β T xq ´
f
T pβ xqEtIpZj ě Zq | β xuˇ
ˇE ´
ˇ
ˇ
n j“1
Bβ T x β X
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
24
and
ˇ
ˇ 2ż
ˇ
ˇ h
( 3 1
B3
T ˚
T ˚
ˇ
ˇ
f
u
K
puqdu
“ ˇ´
T pβ x qE IpZj ě Zq | β x
ˇ
6
Bpβ T xq3 β X
ˇ
ˇ
"
*
ż
ˇ B3
(ˇ
h2
T
T
2
ˇ
ˇ
sup
f T pβ xqE IpZj ě Zq | β x ˇ
u Kpuqdu
ď
2 βT x ˇ Bpβ T xq3 β X
#
+
n
1 ÿ
var ´
IpZj ě ZqKh1 pβ T Xj ´ β T xq
n j“1
(T
1”
E IpZj ě ZqKh1 pβ T Xj ´ β T xq ut IpZj ě ZqKh1 pβ T Xj ´ β T xq
“
n
(
(T ı
´ EIpZj ě ZqKh1 pβ T Xj ´ β T xq EIpZj ě ZqKh1 pβ T Xj ´ β T xq
ˆż
(
1
1
“
IpZj ě ZqK 12 pβ T xj ´ β T xq{h
4
n
h
ˆfc pc, xj qf pt, β T xj qfXl |βT X pxjl | β T xj qfβT X pβ T xj qdcdtdxjl dβ T xj
¸
„
T
„
B
B
2
T
T
T
T
f T pβ xqEtIpZj ě Zq | β xu
f T pβ xqEtIpZj ě Zq | β xu ` Oph q
´
Bβ T x β X
Bβ T x β X
ż
1
1
“
Sc pZ, xjl , β T x ` huq
n h3
ˆSpZ, β T x ` huqfXl |βT X pxjl | β T x ` huqfβ T X pβ T x ` huqdxj K 12 puqdu ` Op1{nq
ż
ż
1
T
T
T
T
S
pZ,
x
,
β
xqSpZ,
β
xqf
pβ
xqf
px
|
β
xqdx
K 12 puqdu
“
T
T
c
jl
jl
β X
Xl |β X jl
nh3
ij
B2
1
Sc pZ, xjl , β T x˚ qSpZ, β T x˚ q
`
2nh3 Bpβ T xq2
ˆfβT X pβ T x˚ qfXl |βT X pxjl | β T x˚ qdxjl h2 u2 K 12 puqdu ` Op1{nq
ˇ
ˇż
1
ˇ
T
T ˇ
ď
K 12 puqdu
sup
T pβ xqEtIpZj ě Zq | β xuˇ
ˇf
nh3 βT x β X
ˇ
ˇż
ˇ B2
ˇ
1
T ˚
T ˚ ˇ
2 12
ˇ
`
sup ˇ
f
T pβ x qEtIpZj ě Zq | β x u
ˇ u K puqdu ` Op1{nq.
2nh βT x˚ Bpβ T xq2 β X
So
´
n
B
1 ÿ
IpZj ě ZqKh1 pβ T Xj ´ β T xq “
fβT X pβ T xqEtIpZj ě Zq | β T xu ` Op pn´1{2 h´3{2 ` h2 q
T
n j“1
Bβ x
(A.4)
uniformly under conditions C1-C2.
Following similar derivations, we have
#
+
n
1 ÿ
E
Xj IpZj ě ZqKh pβ T Xj ´ β T xq
n j“1
(
“ E Xj IpZj ě ZqKh pβ T Xj ´ β T xq
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
25
ż
(
1
xj Ipc ě ZqIpt ě ZqK pβ T xj ´ β T xq{h
h
ˆfc pc, xj qf pt, β T xj qfX|βT X pxj | β T xj qfβT X pβ T xj qdcdtdxj dβ T xj
ż
(
1
“
xj K pβ T xj ´ β T xq{h Sc pZ, xjl , β T xj qSpZ, β T xj qfXl |βT X pxjl | β T xj q
h
ˆfβT X pβ T xj qdxjl dβ T xj
ż
T
T
T
T
“
pβ T x ` hu, xT
jl q KpuqSc pZ, xjl , β x ` huqSpZ, β x ` huqfXl |βT X pxjl | β x ` huq
“
ˆfβT X pβ T x ` huqdxjl du
ż
ż
T
T T
T
T
T
T
“
pβ x, xjl q Sc pZ, xjl , β xqSpZ, β xqfXl |βT X pxjl | β xqfβT X pβ xqdxjl Kpuqdu
ż
)
B ! T
T
T
T
T
T T
q
S
pZ,
x
,
β
xqSpZ,
β
xqf
px
|
β
xqf
pβ
xq
dxjl
pβ
x,
x
`
T
T
c
jl
jl
jl
Xl |β X
β X
Bβ T x
ż
ˆ huKpuqdu
ij
)
!
1
B2
T ˚
T ˚
T ˚
T ˚
T ˚ T T
`
q
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
pβ
x
,
x
T pxjl | β x qf T pβ x q dxjl
c
jl
jl
Xl |β X
β X
2
Bpβ T xq2
ˆh2 u2 Kpuqdu
ż
T
T
T
T
T
“
pβ T x, xT
jl q Sc pZ, xjl , β xqSpZ, β xqfXl |βT X pxjl | β xqfβT X pβ xqdxjl
ij
!
)
B2
h2
T ˚
T T
T ˚
T ˚
T ˚
T ˚
pβ
x
,
x
q
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
px
|
β
x
qf
pβ
x
q
dxjl
`
T
T
c
jl
jl
jl
Xl |β X
β X
2
Bpβ T xq2
ˆu2 Kpuqdu
“ SpZ, β T xqfβT X pβ T xqEtXj Sc pZ, Xj q | β T xu
ż
( 2
h2
B2
T ˚
T ˚
T ˚
`
SpZ,
β
x
qf
u Kpuqdu
T pβ x qE Xj Sc pZ, Xj q | β x
β
X
2
Bpβ T xq2
“ fβT X pβ T xqEtXj IpZj ě Zq | β T xu
ż
(
B2
h2
fβT X pβ T x˚ qE Xj IpZj ě Zq | β T x˚ u2 Kpuqdu.
`
T
2
Bpβ xq2
Hence the absolute bias is
ˇ
ˇ #
+
n
ˇ
ˇ
1 ÿ
ˇ
ˇ
Xj IpZj ě ZqKh pβ T Xj ´ β T xq ´ fβT X pβ T xqEtXj pZj ě Zq | β T xuˇ
ˇE
ˇ
ˇ
n j“1
ˇ 2ż
ˇ
ˇh
ˇ
( 2
B2
T ˚
T ˚
ˇ
“ ˇˇ
f
u
Kpuqdu
T pβ x qE Xj IpZj ě Zq | β x
ˇ
2
Bpβ T xq2 β X
ˇ
ˇ
"
*
ż
ˇ B2
(ˇ
h2
T
T
2
ˇ
ˇ
sup
f T pβ xqE Xj IpZj ě Zq | β x ˇ
u Kpuqdu
ď
2 βT x ˇ Bpβ T xq2 β X
and
+
n
1 ÿ
Xj IpZj ě ZqKh pβ T Xj ´ β T xq
var ´
n j“1
#
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
26
“
“
(T
1”
E Xj IpZj ě ZqKh pβ T Xj ´ β T xq ut Xj IpZj ě ZqKh pβ T Xj ´ β T xq
n
(
(T ı
´ EXj IpZj ě ZqKh pβ T Xj ´ β T xq EXj IpZj ě ZqKh pβ T Xj ´ β T xq
ˆż
(
1
1
2
xj xT
pβ T xj ´ β T xq{h
j IpZj ě ZqK
2
n
h
ˆfc pc, xj qf pt, β T xj qfXl |βT X pxjl | β T xj qfβT X pβ T xj qdcdtdxjl dβ T xj
”
ı”
ıT
´ fβT X pβ T xqEtXj IpZj ě Zq | β T xu fβT X pβ T xqEtXj IpZj ě Zq | β T xu
˘
`Oph2 q
ż
1 1 T
T
T
T
T
pβ x ` hu, xT
“
jl q pβ x ` hu, xjl qSc pZ, xjl , β x ` huq
n h
ˆSpZ, β T x ` huqfXl |βT X pxjl | β T x ` huqfβ T X pβ T x ` huqdxj K 2 puqdu
`Op1{nq
ż
1
T
T
T
T
T
T
T
“
pβ T x, xT
jl q pβ x, xjl qSc pZ, xjl , β xqSpZ, β xqfβ T X pβ xqfXl |βT X pxjl | β xqdxjl
nh
ż
ˆ K 2 puqdu
ij
B2
1
T
T ˚
T
T ˚
T ˚
pβ T x˚ , xT
`
jl q pβ x , xjl qSc pZ, xjl , β x qSpZ, β x q
2nh Bpβ T xq2
So
ˆfβT X pβ T x˚ qfXl |βT X pxjl | β T x˚ qdxjl h2 u2 K 2 puqdu ` Op1{nq
ˇ
ˇż
1
ˇ
T ˇ
sup ˇfβT X pβ T xqEtXj XT
IpZ
ě
Zq
|
β
xu
ď
ˇ K 2 puqdu
j
j
nh β T x
ˇ
ˇż
ˇ B2
ˇ
h
T ˚
T
T ˚ ˇ
ˇ
`
sup ˇ
fβT X pβ x qEtXj Xj IpZj ě Zq | β x uˇ u2 K 2 puqdu ` Op1{nq.
T
2
2n βT x˚ Bpβ xq
n
1 ÿ
Xj IpZj ě ZqKh pβ T Xj ´ β T xq “ fβT X pβ T xqEtXj IpZj ě Zq | β T xu ` Op pn´1{2 h´1{2 ` h2 q
n j“1
(A.5)
uniformly under conditions C1-C2.
+
n
1 ÿ
IpZj ě ZqKh pβ T Xj ´ β T xq
E
n j“1
(
“ E IpZj ě ZqKh pβ T Xj ´ β T xq
ż
(
1
“
Ipc ě ZqIpt ě ZqK pβ T xj ´ β T xq{h
h
ˆfc pc, xj qf pt, β T xj qfX|βT X pxj | β T xj qfβT X pβ T xj qdcdtdxj dβ T xj
ż
(
1
“
K pβ T xj ´ β T xq{h Sc pZ, xjl , β T xj qSpZ, β T xj qfXl |βT X pxjl | β T xj q
h
ˆfβT X pβ T xj qdxjl dβ T xj
#
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
27
“
ż
KpuqSc pZ, xjl , β T x ` huqSpZ, β T x ` huqfXl |βT X pxjl | β T x ` huq
ˆfβT X pβ T x ` huqdxjl du
ż
ż
T
T
T
T
“
Sc pZ, xjl , β xqSpZ, β xqfXl |βT X pxjl | β xqfβT X pβ xqdxjl Kpuqdu
ż
)
B !
T
T
T
T
S
pZ,
x
,
β
xqSpZ,
β
xqf
px
|
β
xqf
pβ
xq
dxjl
`
T
T
c
jl
Xl |β X jl
β X
Bβ T x
ż
ˆ huKpuqdu
ij
!
)
B2
1
T ˚
T ˚
T ˚
T ˚
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
px
|
β
x
qf
pβ
x
q
dxjl
`
T
T
c
jl
Xl |β X jl
β X
2
Bpβ T xq2
ˆh2 u2 Kpuqdu
ż
“
Sc pZ, xjl , β T xqSpZ, β T xqfXl |βT X pxjl | β T xqfβT X pβ T xqdxjl
ij
!
)
h2
B2
T ˚
T ˚
T ˚
T ˚
`
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
px
|
β
x
qf
pβ
x
q
dxjl
T
T
c
jl
Xl |β X jl
β X
2
Bpβ T xq2
ˆu2 Kpuqdu
“ SpZ, β T xqfβT X pβ T xqEtSc pZ, Xj q | β T xu
ż
(
h2
B2
`
SpZ, β T x˚ qfβT X pβ T x˚ qE Sc pZ, Xj q | β T x˚ u2 Kpuqdu
T
2
2
Bpβ xq
“ fβT X pβ T xqEtIpZj ě Zq | β T xu
ż
( 2
B2
h2
T ˚
T ˚
f
u Kpuqdu.
`
T pβ x qE IpZj ě Zq | β x
β
X
2
Bpβ T xq2
Hence the absolute bias is
ˇ
ˇ #
+
n
ˇ
ˇ
1 ÿ
ˇ
ˇ
IpZj ě ZqKh pβ T Xj ´ β T xq ´ fβT X pβ T xqEtIpZj ě Zq | β T xuˇ
ˇE
ˇ
ˇ
n j“1
ˇ 2ż
ˇ
ˇh
ˇ
( 2
B2
T ˚
T ˚
ˇ
ˇ
“ ˇ
f
u
Kpuqdu
T pβ x qE IpZj ě Zq | β x
β
X
ˇ
2
Bpβ T xq2
ˇ
ˇ
"
*
ż
ˇ B2
(ˇ
h2
T
T
2
ˇ
ˇ
sup
f T pβ xqE IpZj ě Zq | β x ˇ
u Kpuqdu
ď
2 βT x ˇ Bpβ T xq2 β X
and
var
“
“
#
+
n
1 ÿ
T
T
IpZj ě ZqKh pβ Xj ´ β xq
n j“1
(T
1”
E IpZj ě ZqKh pβ T Xj ´ β T xq ut Xj IpZj ě ZqKh pβ T Xj ´ β T xq
n
(
(T ı
´ EIpZj ě ZqKh pβ T Xj ´ β T xq EIpZj ě ZqKh pβ T Xj ´ β T xq
ˆż
(
1
1
2
T
T
IpZ
ě
ZqK
pβ
x
´
β
xq{h
j
j
n
h2
ˆfc pc, xj qf pt, β T xj qfXl |βT X pxjl | β T xj qfβT X pβ T xj qdcdtdxjl dβ T xj
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
28
˙
”
ı”
ıT
T
T
T
T
2
´ fβT X pβ xqEtIpZj ě Zq | β xu fβT X pβ xqEtIpZj ě Zq | β xu ` Oph q
ż
1 1
Sc pZ, xjl , β T x ` huq
“
n h
ˆSpZ, β T x ` huqfXl |βT X pxjl | β T x ` huqfβ T X pβ T x ` huqdxj K 2 puqdu ` Op1{nq
ż
ż
1
T
T
T
T
“
Sc pZ, xjl , β xqSpZ, β xqfβ T X pβ xqfXl |βT X pxjl | β xqdxjl K 2 puqdu
nh
ij
B2
1
Sc pZ, xjl , β T x˚ qSpZ, β T x˚ q
`
2nh Bpβ T xq2
So
n
1 ÿ
n j“1
ˆfβT X pβ T x˚ qfXl |βT X pxjl | β T x˚ qdxjl h2 u2 K 2 puqdu ` Op1{nq
ˇż
ˇ
1
ˇ
T
T ˇ
sup ˇf T pβ xqEtIpZj ě Zq | β xuˇ K 2 puqdu
ď
nh β T x β X
ˇ
ˇż
ˇ B2
ˇ
h
T ˚
T ˚ ˇ
ˇ
sup ˇ
`
fβT X pβ x qEtIpZj ě Zq | β x uˇ u2 K 2 puqdu ` Op1{nq.
T
2n βT x˚ Bpβ xq2
IpZj ě ZqKh pβ T Xj ´ β T xq “ fβT X pβ T xqEtIpZj ě Zq | β T xu ` Op pn´1{2 h´1{2 ` h2 q
(A.6)
uniformly under conditions C1-C2.
+
#
n
1 ÿ
T
T
2
IpZj ě ZqKh pβ Xj ´ β xq
E
n j“1
(
“ E IpZj ě ZqKh2 pβ T Xj ´ β T xq
ż
(
1
“
Ipc ě ZqIpt ě ZqK 2 pβ T xj ´ β T xq{h
3
h
ˆfc pc, xj qf pt, β T xj qfX|βT X pxj | β T xj qfβT X pβ T xj qdcdtdxj dβ T xj
ż
(
1 2
“
K pβ T xj ´ β T xq{h Sc pZ, xjl , β T xj qSpZ, β T xj qfXl |βT X pxjl | β T xj q
3
h
ˆfβT X pβ T xj qdxjl dβ T xj
ż
1 2
K puqSc pZ, xjl , β T x ` huqSpZ, β T x ` huqfXl |βT X pxjl | β T x ` huq
“
h2
ˆfβT X pβ T x ` huqdxjl du
ż
!
)
1
B2
T
T
T
T
“
S
pZ,
x
,
β
xqSpZ,
β
xqf
px
|
β
xqf
pβ
xq
dxjl
T
T
c
jl
Xl |β X jl
β X
2h2 Bpβ T xq2
ż
2
ˆh
u2 K 2 puqdu
ij
!
)
h2
B4
T ˚
T ˚
T ˚
T ˚
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
px
|
β
x
qf
pβ
x
q
dxjl
`
T
T
c
jl
Xl |β X jl
β X
24
Bpβ T xq4
ˆu4 K 2 puqdu
ż
!
)
B2
T
T
T
T
S
pZ,
x
,
β
xqSpZ,
β
xqf
px
|
β
xqf
pβ
xq
dxjl
“
T
T
c
jl
Xl |β X jl
β X
Bpβ T xq2
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
29
h2
`
24
ij
!
)
B4
T ˚
T ˚
T ˚
T ˚
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
T pxjl | β x qf T pβ x q dxjl
c
jl
Xl |β X
β X
Bpβ T xq4
ˆu4 K 2 puqdu
B2
SpZ, β T xqfβT X pβ T xqEtSc pZ, Xj q | β T xu
“
Bpβ T xq2
ij
!
)
h2
B4
T ˚
T ˚
T ˚
T ˚
`
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
T pxjl | β x qf T pβ x q dxjl
c
jl
Xl |β X
β X
24
Bpβ T xq4
ˆu4 K 2 puqdu
B2
“
f T pβ T xqEtIpZj ě Zq | β T xu
Bpβ T xq2 β X
ij
!
)
h2
B4
T ˚
T ˚
T ˚
T ˚
`
S
pZ,
x
,
β
x
qSpZ,
β
x
qf
T pxjl | β x qf T pβ x q dxjl
c
jl
Xl |β X
β X
24
Bpβ T xq4
ˆu4 K 2 puqdu.
Hence the absolute bias is
ˇ
ˇ #
+
n
ˇ
ˇ
B2
1 ÿ
ˇ
T
T
2
T
T ˇ
pβ
X
´
β
xq
´
IpZ
ě
ZqK
f
pβ
xqEtIpZ
ě
Zq
|
β
xu
E
T
ˇ
ˇ
j
j
j
h
ˇ
ˇ
n j“1
Bpβ T xq2 β X
ˇ 2ij
ˇ
!
)
4
ˇh
ˇ
B
T
˚
T
˚
T
˚
T
˚
4
2
“ ˇˇ
Sc pZ, xjl , β x qSpZ, β x qfXl |βT X pxjl | β x qfβ T X pβ x q dxjl u K puqduˇˇ
T
24
Bpβ xq4
ˇ "ż
ˇ
*
ˇ B4
(ˇ
h2
T
T
2
ď
sup ˇ
f T pβ xqE IpZj ě Zq | β x ˇˇ
u Kpuqdu .
2 β T x˚ ˇ Bpβ T xq4 β X
To analyze the variance,
+
#
n
1 ÿ
2
T
T
IpZj ě ZqKh pβ Xj ´ β xq
var
n j“1
(T
1”
“
E IpZj ě ZqKh2 pβ T Xj ´ β T xq ut Xj IpZj ě ZqKh2 pβ T Xj ´ β T xq
n
(
(T ı
´ EIpZj ě ZqKh2 pβ T Xj ´ β T xq EIpZj ě ZqKh2 pβ T Xj ´ β T xq
ˆż
(
1
1
22
T
T
“
IpZ
ě
ZqK
pβ
x
´
β
xq{h
j
j
n
h6
ˆfc pc, xj qf pt, β T xj qfXl |βT X pxjl | β T xj qfβT X pβ T xj qdcdtdxjl dβ T xj
¸
„
T
„
2
B
B2
f T pβ T xqEtIpZj ě Zq | β T xu
f T pβ T xqEtIpZj ě Zq | β T xu ` Oph2 q
´
Bpβ T xq2 β X
pBβ T xq2 β X
ż
1
1
“
Sc pZ, xjl , β T x ` huq
n h5
ˆSpZ, β T x ` huqfXl |βT X pxjl | β T x ` huqfβ T X pβ T x ` huqdxj K 22 puqdu ` Op1{nq
ż
ż
1
T
T
T
T
“
Sc pZ, xjl , β xqSpZ, β xqfβT X pβ xqfXl |βT X pxjl | β xqdxjl K 22 puqdu
nh5
ij
B2
1
Sc pZ, xjl , β T x˚ qSpZ, β T x˚ q
`
2nh5 Bpβ T xq2
ˆfβT X pβ T x˚ qfXl |βT X pxjl | β T x˚ qdxjl h2 u2 K 2 puqdu ` Op1{nq
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
30
ď
So
ˇ
ˇż
1
ˇ
T
T ˇ
K 22 puqdu
sup
T pβ xqEtIpZj ě Zq | β xuˇ
ˇf
nh5 βT x β X
ˇ
ˇż
ˇ B2
ˇ
1
T ˚
T ˚ ˇ
ˇ
`
sup ˇ
fβT X pβ x qEtIpZj ě Zq | β x uˇ u2 K 22 puqdu ` Op1{nq.
T
3
2
2nh βT x˚ Bpβ xq
n
1 ÿ
B2
f
pβ T xqEtIpZj ě Zq | β T xu ` Op pn´1{2 h´5{2 ` h2 q
IpZj ě ZqKh pβ T Xj ´ β T xq “
T 2 βT X
n j“1
Bpβ xq
(A.7)
uniformly under conditions C1-C2.
Combining the results of (A.1) and (A.6), we have
(
p Y pZq | β X “
E
T
“
“
řn
T
T
j“1 IpZj ě ZqKh pβ Xj ´ β Xq
řn
T
T
j“1 Kh pβ Xj ´ β Xq
fβT X pβ T XqEtIpZj ě Zq | β T Xu ` Op pn´1{2 h´1{2
fβT X pβ T Xq ` Op pn´1{2 h´1{2 ` h2 q
fβT X pβ T XqEtY pZq | β T Xu
fβT X pβ T Xq
` h2 q
` Op pn´1{2 h´1{2 ` h2 q
“ EtY pZq | β T Xu ` Op tpnhq´1{2 ` h2 u,
uniformly under conditions C1-C2. Combining the results of (A.1) and (A.5), we have
řn
T
T
(
j“1 Xj IpZj ě ZqKh pβ Xj ´ β Xq
T
p
E XY pZq | β X “
řn
T
T
j“1 Kh pβ Xj ´ β Xq
“
“
fβT X pβ T XqEtXj IpZj ě Zq | β T Xu ` Op pn´1{2 h´1{2 ` h2 q
fβT X pβ T Xq ` Op pn´1{2 h´1{2 ` h2 q
fβT X pβ T XqEtXY pZq | β T Xu
T
fβT X pβ Xq
` Op pn´1{2 h´1{2 ` h2 q
“ EtXY pZq | β T Xu ` Op tpnhq´1{2 ` h2 u
uniformly under conditions C1-C2. Combining the results of (A.1), (A.3), (A.6) and (A.4), we have
(
B p
E Y pZq | β T X
T
Bβ X
řn
T
T
1
j“1 IpZj ě ZqKh pβ Xj ´ β Xq
“ ´
řn
T
T
j“1 Kh pβ Xj ´ β Xq
řn
ř
t j“1 IpZj ě ZqKh pβ T Xj ´ β T Xqut nj“1 Kh1 pβ T Xj ´ β T Xqu
`
ř
t nj“1 Kh pβ T Xj ´ β T Xqu2
“
Bfβ T X pβ T XqEtIpZj ě Zq | β T Xu{Bβ T X ` Op pn´1{2 h´3{2 ` h2 q
fβT X pβ T Xq ` Op pn´1{2 h´1{2 ` h2 q
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
31
`
“
fβ1 T X pβ T XqEtIpZj ě Zq | β T Xu
fβT X pβ T Xq
´
“
”
ı”
ı
fβT X pβ T XqEtIpZj ě Zq | β T Xu ` Op pn´1{2 h´1{2 ` h2 q ´fβ1 T X pβ T Xq ` Op pn´1{2 h´3{2 ` h2 q
fβ2 T X pβ T Xq ` Opn´1{2 h´1{2 ` h2 q
`
fβ1 T X pβ T XqEtIpZj ě Zq | β T Xu
fβ2 T X pβ T Xq
BEtIpZj ě Zq | β T Xu
Bβ T X
` Op pn´1{2 h´3{2 ` h2 q
B
EtY pZq | β T Xu ` Op tpnh3 q´1{2 ` h2 u
Bβ T X
uniformly under conditions C1-C2. Finally, combining the results of (A.1), (A.3), (A.5) and (A.2),
we have
(
B p
E XY pZq | β T X
T
Bβ X
řn
T
T
1
j“1 Xj IpZj ě ZqKh pβ Xj ´ β Xq
“ ´
řn
T
T
j“1 Kh pβ Xj ´ β Xq
ř
ř
t nj“1 Xj IpZj ě ZqKh pβ T Xj ´ β T Xqut nj“1 Kh1 pβ T Xj ´ β T Xqu
`
ř
t nj“1 Kh pβ T Xj ´ β T Xqu2
“
Bfβ T X pβ T XqEtXj IpZj ě Zq | β T Xu{Bβ T X ` Op pn´1{2 h´3{2 ` h2 q
fβT X pβ T Xq ` Op pn´1{2 h´1{2 ` h2 q
`
“
fβ1 T X pβ T XqEtXY pZq | β T Xu
fβT X pβ T Xq
´
“
”
ı”
ı
fβT X pβ T XqEtXj IpZj ě Zq | β T Xu ` Op pn´1{2 h´1{2 ` h2 q ´fβ1 T X pβ T Xq ` Op pn´1{2 h´3{2 ` h2 q
fβ2 T X pβ T Xq ` Opn´1{2 h´1{2 ` h2 q
`
fβ1 T X pβ T XqEtXY pZq | β T Xu
fβT X pβ T Xq
fβT X pβ T XqBEtXY pZq | β T Xu{Bβ T X
fβT X pβ T Xq
` Op pn´1{2 h´3{2 ` h2 q
B
EtXY pZq | β T Xu ` Op tpnh3 q´1{2 ` h2 u
T
Bβ X
uniformly under conditions C1-C2.
Now we inspect the consistency
of the Kaplan Meier estimator on the hazard function and its
ř
derivatives. Let A “ n´1 nj“1 IpZj ě Zi qKh pβ T Xj ´ β T xq ´ fβT X pβ T xqEtIpZ ě Zi q | β T xu.
ż8
T
T
p
p
Kb pt ´ zqdΛpt|β
xq
λpz, β xq “
0
“
“
“
n
ÿ
i“1
∆i Kh pβ T Xi ´ β T xq
T
T
j“1 IpZj ě Zi qKh pβ Xj ´ β xq
Kb pZi ´ zq řn
n
1 ÿ
∆i Kh pβ T Xi ´ β T xq
Kb pZi ´ zq
n i“1
fβ T X pβ T xqEtIpZ ě Zi q | β T xu ` A
n
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
t1 ` Op pAqu,
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
32
We first inspect
First,
n
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
.
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
«
ff
n
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
E
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
«
ff
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
“ E
fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
ij
Kb pzi ´ zqKh pβ T xi ´ β T xq
“
f pzi , β T xi qSc pzi , xi qfX|βT X pxi , β T xi q
fβT X pβ T xqSpzi , β T xqEtSc pzi , Xi q | β T xu
ˆfβT X pβ T xi qdzi dxi dβ T xi
ij
Kb pzi ´ zqKh pβ T xi ´ β T xqEtSc pzi , Xi q | β T xi u
f pzi , β T xi q
“
T
T
T
fβT X pβ xqSpzi , β xqEtSc pzi , Xi q | β xu
ˆfβT X pβ T xi qdzi dβ T xi
ij
KpvqKpuqEtSc pz ` bv, Xi q | β T x ` huu
“
f pz ` bv, β T x ` huq
fβT X pβ T xqSpz ` bv, β T xqEtSc pz ` bv, Xi q | β T xu
ˆfβT X pβ T x ` huqdvdu
ij
ij
KpvqKpuq
b2 B 2
T
f pz ˚ , β T xqv 2 dvdu
“
KpvqKpuqλpz, β xqdvdu `
2
2Bz
Spz ˚ , β T xq
ij
KpvqKpuqEtSc pz, Xi q | β T x˚ uf pz, β T x˚ qfβT X pβ T x˚ q 2
h2 B2
`
u dvdu
2Bpβ T xq2
fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu
ij
KpvqKpuq
b2 B2
T
f pz ˚ , β T xqv 2 dvdu
“ λpz, β xq `
2
2Bz
Spz ˚ , β T xq
ij
KpvqKpuqEtSc pz, Xi q | β T x˚ uf pz, β T x˚ qfβT X pβ T x˚ q 2
h2 B2
`
u dvdu,
2Bpβ T xq2
fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu
where throughout the text, z ˚ is on the line connecting z and z ` bv. Thus, the absolute bias is
ˇ #
ˇ
+
n
ˇ
ˇ
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
ˇ
T ˇ
´
λpz,
β
xq
ˇE
ˇ
ˇ
ˇ
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
ˇ 2
ˇ
ż
ˇ B f pz ˚ , β T xq ˇ
ˇ v 2 Kpvqdv
ď b2 sup ˇˇ
ˇ
2
T
˚
z ˚ ,βT x 2Bz Spz , β xq
ˇ
ˇ
T ˚
T ˚
T ˚ ˇż
ˇ
2
EtS
pz,
X
q
|
β
x
uf
pz,
β
x
qf
T pβ x q ˇ
c
i
B
ˇ
β
X
`h2
sup
ˇ u2 Kpuqdu
ˇ
T
T
T
T
ˇ
T
T ˚ ˇ 2Bpβ xq2
f
pβ
xqSpz,
β
xqEtS
pz,
X
q
|
β
xu
T
c
i
z,β x,β x
β X
“ Oph2 ` b2 q
under conditions C1-C6. Following the same procedure, noting that A “ Op tpnhq´1{2 ` h2 u uniformly, we can show that
n
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
Op pAq “ Op th2 ` pnhq´1{2 u,
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
33
p β T xq is of order Op tpnhq´1{2 ` h2 ` b2 u uniformly. On the other hand, the
hence the bias of λpz,
p β T xq is
variance of λpz,
!
)
p β T xq
var λpz,
ff
«
n
∆i Kh pβ T Xi ´ β T xq
1ÿ
Kb pZi ´ zq
“ var
n i“1
fβ T X pβ T xqEtIpZ ě Zi q | β T xu ` A
«
ff
n
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
“ var
t1 ` Op pAqu
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
«
ff
«
ff
n
n
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
ď 2var
` 2var
Op pAq .
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
We inspect the first term first.
ff
«
n
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
2var
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
ff
«
2
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
“
var
n
fβT X pβ T xqEtIpZ ě Zi q | β T xu
˛
¨ «
ff2
T
T
2˝
Kb pZi ´ zq∆i Kh pβ Xi ´ β xq
“
´ λ2 pz, β T xq ` Opn´1{2 h´1{2 ` h2 ` b2 q‚
E
T
T
n
fβT X pβ xqEtIpZ ě Zi q | β xu
¨ «
ff2 ˛
T
T
Kb pZi ´ zq∆i Kh pβ Xi ´ β xq ‚
2˝
` Op1{nq
E
“
n
fβT X pβ T xqEtIpZ ě Zi q | β T xu
ij
2
K 2 tpZi ´ zq{buK 2 tpβ T xi ´ β T xq{hu
“ 2 2
b h n
fβ2T X pβ T xqS 2 pzi , β T xqE 2 tSc pzi , Xi q | β T xu
ˆf pzi , β T xi qSc pzi , xi qfX|β T X pxi , β T xi qfβT X pβ T xi qdzi dxi dβ T xi ` Op1{nq
ij
2
K 2 pvqK 2 puq
“
bhn
fβ2 T X pβ T xqS 2 pz ` bv, β T xqE 2 tSc pz ` bv, Xi q | β T xu
ˆf pz ` bv, β T x ` huqEtSc pz ` bv, Xi q | β T x ` huufβ T X pβ T x ` huqdvdu ` Op1{nq
ij
2
f pz, β T xqK 2 pvqK 2 puq
“
dvdu
bhn
fβT X pβ T xqS 2 pz, β T xqEtSc pz, Xi q | β T xu
ij
f pz ˚ , β T xqK 2 pvqK 2 puq
bB2
v 2 dvdu
`
nhBz 2
fβT X pβ T xqS 2 pz ˚ , β T xqEtSc pz ˚ , Xi q | β T xu
ij
fβT X pβ T x˚ qf pz, β T x˚ qEtSc pz, Xi q | β T x˚ uK 2 pvqK 2 puq 2
hB2
`
u dvdu ` Op1{nq
nbBpβ T xq2
fβ2 T X pβ T xqS 2 pz, β T xqEtSc pz, Xi q2 | β T xu
"ż
*2
f pz, β T xq
2
2
ď
K puqdu
bhn fβT X pβ T xqS 2 pz, β T xqEtSc pz, Xi q | β T xu
ˇ "ż
ˇ
* "ż
*
ˇ
ˇ B2
b
f pz ˚ , β T xq
ˇ
ˇ
2
2
2
`
sup ˇ
K puqu du
K puqdu
ˇ
nh z ˚ ,βT x ˇ Bz 2 fβT X pβ T xqS 2 pz ˚ , β T xqEtSc pz ˚ , Xi q | β T xu ˇ
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
34
ˇ
ˇ
ˇ B2
fβT X pβ T x˚ qf pz, β T x˚ qEtSc pz, Xi q | β T x˚ u ˇˇ
h
ˇ
sup
`
ˇ
ˇ
nb z,βT x,βT x˚ ˇ Bpβ T xq2 f 2 T pβ T xqS 2 pz, β T xqEtSc pz, Xi q2 | β T xu ˇ
β X
"ż
* "ż
*
2
2
2
ˆ
K puqu du
K puqdu ` Op1{nq
“ Ot1{pnhbq ` h{pnbq ` b{pnhq ` 1{nu
“ Ot1{pnhbqu
uniformly under conditions C1-C6 and β T x˚ is on the line connecting β T x and β T x ` hu. For the
second term
«
ff
n
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
2var
Op pAq
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
ff2
«
n
1 ÿ Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
Op pAq
ď 2E
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
˛
¨
ff2
«
n
T
T
ÿ
Kb pZi ´ zq∆i Kh pβ Xi ´ β xq
1
suptOp2 pAqu‚
ď 2E ˝
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
˛
¨
ff2
«
n
T
T
ÿ
“
‰
1
Kb pZi ´ zq∆i Kh pβ Xi ´ β xq
Op tpnhq´1 ` h4 u ‚
“ 2E ˝
n i“1 fβT X pβ T xqEtIpZ ě Zi q | β T xu
ff2
«
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
“ 2E
Otpnhq´1 ` h4 u
fβT X pβ T xqEtIpZ ě Zi q | β T xu
«
ff
"ż
*2
2
f pz, β T xq
´1 ´1
´1 ´1
2
“
K puqdu ` Opn b h ` n h bq
bhn fβT X pβ T xqS 2 pz, β T xqEtSc pz, Xi q | β T xu
ˆOtpnhq´1 ` h4 u
“ Otpnhq´2 b´1 ` n´1 h3 b´1 u
p β T xq is of
under conditions C1-C6 uniformly. Summarizing the above results, the variance of λpz,
T
p β xq, specifically
order 1{pnhbq uniformly. Hence we have the consistency of estimator λpz,
p β T xq “ λpz, β T xq ` Op tpnhbq´1{2 ` h2 ` b2 u
λpz,
uniformly under condition C1-C6. Next we inspect the estimator for the first derivative of hazard
function λpz, β T xq. Let
p 12
λ
n
ÿ
∆i Kh1 pβ T Xi ´ β T xq
T
T
j IpZj ě zqKh pβ Xj ´ β xq
i“1
řn
T
T
n
1
ÿ
j“1 IpZj ě zqKh pβ Xj ´ β xq
T
T
Kb pZi ´ zq∆i Kh pβ Xi ´ β xq řn
“
.
t j“1 IpZj ě zqKh pβ T Xj ´ β T xqu2
i“1
p 11 “ ´
λ
Kb pZi ´ zq ř
p 1 px, β T xq “ λ
p 11 ` λ
p 12 . To analyze λ
p 11 ,
Then λ
#
n
ÿ
p
Kb pZi ´ zq ř
E λ11 “ E ´
i“1
∆i Kh1 pβ T Xi ´ β T xq
n
T
T
j“1 IpZj ě zqKh pβ Xj ´ β xq
+
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
35
ff
n
∆i Kh1 pβ T Xi ´ β T xq
1 ÿ
´Kb pZi ´ zq
“ E
n i“1
fβ T X pβ T xqEtIpZ ě Zi q | β T xu ` A
«
ff
n
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
1 ÿ
“ E
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
«
ff
n
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
1ÿ
`E
Op pAq
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
«
ff
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
“ E
fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
«
ff
n
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
1ÿ
`E
Op pAq .
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
«
We inspect the first term to obtain
ff
«
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
E
fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
ij
Kb pzi ´ zqKh1 pβ T xi ´ β T xq
“ ´
f pzi , β T xi qSc pzi , xi qfX|β T X pxi , β T xi q
fβT X pβ T xqSpzi , β T xqEtSc pzi , Xi q | β T xu
ˆfβT X pβ T xi qdzi dxi dβ T xi
ij
1 KpvqK 1 puqf pz ` bv, β T x ` huqEtSc pz ` bv, Xi q | β T x ` huu
fβT X pβ T x ` huqdvdu
“ ´
h
fβT X pβ T xqSpz ` bv, β T xqEtSc pz ` bv, Xi q | β T xu
”
ı
ij B f pz, β T xqEtSc pz, Xi q | β T xuf T pβ T xq {Bβ T x
β X
“ ´
KpvquK 1 puqdvdu
T
T
fβT X pβ xqSpz, β xqEtSc pz, Xi q | β T xu
ij
fβT X pβ T x˚ qf pz ˚ , β T x˚ qEtSc pz ˚ , Xi q | β T x˚ u 2
b2 B3
´
v KpvquK 1 puqdvdu
2Bz 2 Bβ T x
fβT X pβ T xqSpz ˚ , β T xqEtSc pz ˚ , Xi q | β T xu
ij
fβT X pβ T x˚ qf pz, β T x˚ qEtSc pz, Xi q | β T x˚ u
h2 B3
´
Kpvqu3 K 1 puqdvdu
T 3
T
T
T
6Bpβ xq
fβT X pβ xqSpz, β xqEtSc pz, Xi q | β xu
”
ı
B f pz, β T xqEtSc pz, Xi q | β T xufβT X pβ T xq {Bβ T x
“
fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu
ij
fβT X pβ T x˚ qf pz ˚ , β T x˚ qEtSc pz ˚ , Xi q | β T x˚ u 2
b2 B3
v KpvquK 1 puqdvdu
´
2Bz 2 Bβ T x
fβT X pβ T xqSpz ˚ , β T xqEtSc pz ˚ , Xi q | β T xu
ij
fβT X pβ T x˚ qf pz, β T x˚ qEtSc pz, Xi q | β T x˚ u
h2 B3
´
Kpvqu3 K 1 puqdvdu.
6Bpβ T xq3
fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu
Hence the absolute bias is given by
ˇ «
ff
ˇ
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
ˇ
ˇE
ˇ
fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
”
ı
ˇ
B f pz, β T xqEtSc pz, Xi q | β T xufβT X pβ T xq {Bβ T x ˇˇ
ˇ
´
ˇ
fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu
ˇ
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
36
ˇż
ˇ
ˇ
fβT X pβ T x˚ qf pz ˚ , β T x˚ qEtSc pz ˚ , Xi q | β T x˚ u ˇˇ
B3
ˇ
ď b
sup
ˇ v 2 Kpvqdv
ˇ
T
T
T
T
2
˚
˚
ˇ
T
T
z ˚ ,β x,β x˚ 2Bz Bβ x fβT X pβ xqSpz , β xqEtSc pz , Xi q | β xu ˇ
ˇ
ˇż
ˇ
fβT X pβ T x˚ qf pz, β T x˚ qEtSc pz, Xi q | β T x˚ u ˇˇ
B3
ˇ
2
`h
sup
ˇ
ˇ u2 Kpuqdu
T
T
T
T
3
ˇ
fβT X pβ xqSpz, β xqEtSc pz, Xi q | β xu ˇ
z,βT x,βT x˚ 2Bpβ xq
2
“ Opb2 ` h2 q
uniformly under condition C1-C6.
Following the same procedure, we conclude that
n
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
1 ÿ
Op pAq “ Op th2 ` pnhq´1{2 u
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
uniformly under conditions C1-C6 due to A “ Op th2 ` pnhq´1{2 u. Therefore, we have
”
ı
B f pz, β T xqEtSc pz, Xi q | β T xufβ T X pβ T xq {Bβ T x
p 11 “
` Otpnhq´1{2 ` b2 ` h2 u
Eλ
T
T
T
fβT X pβ xqSpz, β xqEtSc pz, Xi q | β xu
p 12 , let B “ ´1{n
For λ
then
řn
j“1 IpZj
ě zqKh1 pβ T Xj ´ βT xq ´ BfβT X pβ T xqEtIpZj ě zq | β T xu{Bβ T x,
řn
T
T
1
j“1 IpZj ě zqKh pβ Xj ´ β xq
T
T
p
λ12 “ Kb pZi ´ zq∆i Kh pβ Xi ´ β xq řn
t j“1 IpZj ě zqKh pβ T Xj ´ β T xqu2
”
ı
T
T
T
n
B
f
T pβ xqEtIpZ ě Zi q | β xu {Bβ x ` B
ÿ
X
β
1
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
“ ´
n i“1
rfβT X pβ T Xi qEtIpZ ě Zi q | β T xu ` As2
”
ı
T
T
T
n
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
ÿ
β X
1
T
T
“ ´
Kb pZi ´ zq∆i Kh pβ Xi ´ β xq
n i“1
fβ2T X pβ T xqE 2 tIpZ ě Zi q | β T xu
ˆ t1 ` Op pBq ` Op pAqu .
We have
»
”
ı
fi
T
T
B
f
pβ
zqEtIpZ
ě
Z
q
|
β
xu
{Bβ T x
T
i
β
X
1
fl
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
E –´
n i“1
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
”
ı
fi
»
B fβT X pβ T zqEtIpZ ě Zi q | β T xu {Bβ T x
fl
“ ´E –Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
”
ı
ij
B fβT X pβ T xqSpzi , β T xqEtSc pzi , Xi q | β T xu {Bβ T x
“ ´
Kb pzi ´ zqKh pβ T xi ´ β T xq
fβ2 T X pβ T xqS 2 pzi , β T xqE 2 tSc pzi , Xi q | β T xu
n
ÿ
ˆf pzi , β T xi qSc pzi , xi qfX|β T X pxi , β T xi qfβT X pβ T xi qdzi dxi dβ T xi
”
ı
ij
B fβT X pβ T xqSpzi , β T xqEtSc pzi , Xi q | β T xu {Bβ T x
“ ´
Kb pzi ´ zqKh pβ T xi ´ β T xq
fβ2 T X pβ T xqS 2 pzi , β T xqE 2 tSc pzi , Xi q | β T xu
ˆf pzi , β T xi qEtSc pzi , Xi q | β T xi ufβT X pβ T xi qdzi dβ T xi
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
37
“ ´
ij
KpvqKpuq
”
ı
B fβT X pβ T xqSpz ` bv, β T xqEtSc pz ` bv, Xi q | β T xu {Bβ T x
fβ2T X pβ T xqS 2 pz ` bv, β T xqE 2 tSc pz ` bv, Xi q | β T xu
ˆf pz ` bv, β T x ` huqEtSc pz ` bv, Xi q | β T x ` huufβ T X pβ T x ` huqdvdu
”
ı
ij
B fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu {Bβ T x
“ ´
KpvqKpuq
f pz, β T xqdvdu
fβT X pβ T xqS 2 pz, β T xqEtSc pz, Xi q | β T xu
”
ı
T
T
T
T
˚
˚
ij
B
f
2
2
T pβ xqSpz , β xqEtSc pz , Xi q | β xu {Bβ x
β X
b B
f pz ˚ , β T xqv 2 dvdu
KpvqKpuq
´
T
T
T
2
˚
˚
2Bz 2
fβT X pβ xqS pz , β xqEtSc pz , Xi q | β xu
”
ı
T
T
T
T
ij
2
2
B
f
T pβ xqSpz, β xqEtSc pz, Xi q | β xu {Bβ x
β X
h B
´
KpvqKpuq
2Bpβ T xq2
fβ2 T X pβ T xqS 2 pz, β T xqE 2 tSc pz, Xi q | β T xu
ˆf pz, β T x˚ qEtSc pz, Xi q | β T x˚ ufβT X pβ T x˚ qu2 dvdu
”
ı
B fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu {Bβ T x
f pz, β T xq
“ ´
T
T
T
2
fβT X pβ xqS pz, β xqEtSc pz, Xi q | β xu
”
ı
ij B f T pβ T xqSpz ˚ , β T xqEtSc pz ˚ , Xi q | β T xu {Bβ T x
2
2
β X
b B
´
f pz ˚ , β T xqv 2 KpvqKpuqdvdu
2
2Bz
fβT X pβ T xqS 2 pz ˚ , β T xqEtSc pz ˚ , Xi q | β T xu
”
ı
ij B f T pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu {Bβ T x
2
2
β X
h B
´
T 2
2Bpβ xq
fβ2 T X pβ T xqS 2 pz, β T xqE 2 tSc pz, Xi q | β T xu
ˆf pz, β T x˚ qEtSc pz, Xi q | β T x˚ ufβT X pβ T x˚ qu2 KpvqKpuqdvdu.
Therefore
”
ı
ˇ »
fi
T
T
ˇ
n
B
f
pβ
zqEtIpZ
ě
Z
q
|
β
xu
{Bβ T x
T
ÿ
i
ˇ
β
X
1
ˇE –´
fl
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
ˇ
n
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
ˇ
i“1
”
ı
ˇ
ˇ
B fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu {Bβ T x
ˇ
T ˇ
`
f pz, β xqˇ
T
T
T
fβT X pβ xqS 2 pz, β xqEtSc pz, Xi q | β xu
ˇ
”
ı
ˇ
ˇ
ˇ
ˇ 2 B f T pβ T xqSpz ˚ , β T xqEtS pz ˚ , X q | β T xu {Bβ T x
c
i
ˇ
ˇ
β X
B
˚
T ˇ
2
ˇ
f
pz
,
β
xq
ď b sup ˇ
ˇ
2
fβT X pβ T xqS 2 pz ˚ , β T xqEtSc pz ˚ , Xi q | β T xu
ˇ
z ˚ ,β T x ˇ 2Bz
"ż
*
ˆ
v 2 Kpvqdv
”
ı
ˇ
ˇ
T
T
T
2
B fβT X pβ xqSpz, β xqEtSc pz, Xi q | β xu {Bβ T x
ˇ
B
2
ˇ
`h
sup
ˇ
T 2
fβ2 T X pβ T xqS 2 pz, β T xqE 2 tSc pz, Xi q | β T xu
z,βT x,βT x˚ ˇ 2Bpβ xq
ˇ "ż
*
ˇ
2
T ˚
T ˚
T ˚ ˇ
u Kpuqdu
ˆf pz, β x qEtSc p, Xi q | β x ufβT X pβ x qˇ
ˇ
“ Opb2 ` h2 q
uniformly under conditions C1-C6. Noting that B “ Op pn´1{2 h´3{2 ` h2 q, based on similar proceimsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
38
dure, we have
´
n
ÿ
1
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
n i“1
”
ı
B fβT X pβ T zqEtIpZ ě Zi q | β T xu {Bβ T x
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
Op pBq
“ Op pn´1{2 h´3{2 ` h2 q,
”
ı
T
T
T
n
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
ÿ
β X
1
T
T
´
Op pAq
Kb pZi ´ zq∆i Kh pβ Xi ´ β xq
n i“1
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
“ Op pn´1{2 h´1{2 ` h2 q
uniformly under condition C1-C6. Therefore, we can conclude that
”
ı
B fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu {Bβ T x
p 12 “ ´
Eλ
f pz, β T xq ` Opn´1{2 h´3{2 ` b2 ` h2 q
T
T
T
2
fβT X pβ xqS pz, β xqEtSc pz, Xi q | β xu
In addition, we have
”
ı
B f pz, β T xqEtSc pz, Xi q | β T xufβ T X pβ T xq {Bβ T x
´
“
fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu
”
ı
B fβT X pβ T xqSpz, β T xqEtSc pz, Xi q | β T xu {Bβ T x
fβT X pβ
T
xqS 2 pz, β T xqEtS
c pz, Xi q
T
| β xu
f pz, β T xq
T
T
1
Bf pz, β T xq{Bβ T x f pz, β T xqBEtSc pz, Xi q | β T xu{Bβ T x fβT X pβ xqf pz, β xq
`
`
Spz, β T xq
Spz, β T xqEtSc pz, Xi q | β T xu
fβT X pβ T xqSpz, β T xq
´
fβ1 T X pβ T xqf pz, β T xq
fβT X pβ T xqSpz, β T xq
´
f pz, β T xqBSpz, β T xq{Bβ T x f pz, β T xqBEtSc pz, Xi q | β T xu{Bβ T x
´
S 2 pz, β T xq
Spz, β T xqEtSc pz, Xi q | β T xu
Bf pz, β T xq{Bβ T x f pz, β T xqBSpz, β T xq{Bβ T x
´
Spz, β T xq
S 2 pz, β T xq
B
“
λpz, β T xq
Bβ T x
“ λ1 pz, β T xq.
“
p 11 and E λ
p 12 , we readily obtain
Combining E λ
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
ˇ
p 11 ` E λ
p 12 ´ B λ1 pz, β T xqˇ
p 1 pz, β T xq ´ B λ1 pz, β T xqˇ “ ˇE λ
ˇE λ
ˇ
ˇ
ˇ
ˇ
Bβ T x
Bβ T x
“ Opn´1{2 h´3{2 ` b2 ` h2 q
uniformly under conditions C1-C6.
p 1 pz, β T xq is given by
The variance of λ
!
)
!
)
p 1 pz, β T xq “ var λ
p 11 q ` 2varpλ
p 12 q.
p 11 ` λ
p 12 ď 2varpλ
var λ
We exam each term separately.
p 11 q
varpλ
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
39
+
´∆i Kh1 pβ T Xi ´ β T xq
“ var
Kb pZi ´ zq ř
T
T
j IpZj ě zqKh pβ Xj ´ β xq
i“1
+
#
n
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
1ÿ
“ var
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu ` A
ff
«
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
2
ď
var
n
fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
«
ff
n
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
1ÿ
`2var
Op pAq
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
#
n
ÿ
The first part is given by
ff
«
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
2
var
n
fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
ff2
«
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
2
“
E
n
fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
ff¸2
˜ «
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
2
E
´
n
fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
ij
Kb2 pzi ´ zqKh12 pβ T xi ´ β T xq
2
“
n
fβ2 T X pβ T xqS 2 pzi , β T xqE 2 tSc pzi , Xi q | β T xu
ˆf pzi , β T xi qSc pzi , xi qfX|β T X pxi , β T xi qfβT X pβ T xi qdzi dxi dβ T xi ` Op1{nq
ij
K 2 pvqK 12 puqEtSc pz ` bv, Xi q | β T x ` huu
2
“
nbh3
fβ2 T X pβ T xqS 2 pz ` bv, β T xqE 2 tSc pz ` bv, Xi q | β T xu
ˆf pz ` bv, β T x ` huqfβ T X pβ T x ` huqdvdu ` Op1{nq
ij
2
K 2 pvqK 12 puqf pz, β T xq
“
dvdu
nbh3
fβT X pβ T xqS 2 pz, β T xqEtSc pz, Xq | β T xu
ij
b B2
f pz ˚ , β T xq
` 3 2
v 2 K 2 pvqK 12 puqdvdu
T
T
T
2
˚
˚
nh Bz
fβT X pβ xqS pz , β xqEtSc pz , Xi q | β xu
ij
f pz, β T x˚ qfβT X pβ T x˚ qEtSc pz, Xi q | β T x˚ u 2
1
B2
`
K pvqu2 K 12 puqdvdu ` Op1{nq
T
T
T
2
2
2
nbh Bpβ T xq2
fβT X pβ xqS pz, β xqE tSc pz, Xi q | β xu
"ż
* "ż
*
f pz, β T xq
2
2
12
ď
K pvqdv
K puqdu
nbh3 fβT X pβ T xqS 2 pz, β T xqEtSc pz, Xq | β T xu
«
ff "ż
* "ż
*
B2
b
f pz ˚ , β T xq
2 2
12
` 3 sup
v K pvqdv
K puqdu
nh z ˚ ,βT x Bz 2 fβT X pβ T xqS 2 pz ˚ , β T xqEtSc pz ˚ , Xi q | β T xu
ff
«
f pz, β T x˚ qfβ T X pβ T x˚ qEtSc pz, Xi q | β T x˚ u
B2
1
sup
`
nbh z,βT x,βT x˚ Bpβ T xq2 f 2 T pβ T xqS 2 pz, β T xqE 2 tSc pz, Xi q | β T xu
β X
"ż
* "ż
*
2
2 12
ˆ
K pvqdv
u K puqdu ` Op1{nq
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
40
“ Ot1{pnbh3 q ` b{pnh3 q ` 1{pnbhq ` 1{nu
“ Ot1{pnbh3 qu
uniformly under condition C1-C6. Noting that A “ Op pn´1{2 h´1{2 ` h2 q, the second part is
«
ff
n
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
1 ÿ
2var
Op pAq
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
ff2
«
n
´Kb pZi ´ zq∆i Kh1 pβ T Xi ´ β T xq
1 ÿ
Op pAq
ď 2E
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
¨
˛
«
ff2
n
T
T
1
ÿ
(
´Kb pZi ´ zq∆i Kh pβ Xi ´ β xq
1
ď 2E ˝
sup Op2 pAq ‚
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
¨
«
ff2 ˛
n
1 pβ T X ´ β T xq
ÿ
´K
pZ
´
zq∆
K
1
i
i
i h
b
‚Ot1{pnhq ` h4 u
“ 2E ˝
n i“1 fβT X pβ T xqSpZi , β T xqEtSc pZi , Xi q | β T xu
«
"ż
˙ff
* "ż
*
ˆ
f pz, β T xq
1
b
2
2
12
`
K pvqdv
K puqdu ` O
“
nbh3 fβT X pβ T xqS 2 pz, β T xqEtSc pz, Xq | β T xu
nh3 nbh
ˆOt1{pnhq ` h4 u
“ Otn´2 b´1 h´4 ` pnbq´1 hu
uniformly under conditions C1-C6. Therefore
p 11 q “ Ot1{pnbh3 qu
varpλ
uniformly under conditions C1-C6.
p 12 ,
For λ
p q
varpλ
ff
« 12
řn
n
1 pβ T X ´ β T xq
ÿ
IpZ
ě
zqK
j
j
j“1
h
“ var
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq řn
t j“1 IpZj ě zqKh pβ T Xj ´ β T xqu2
i“1
”
ı
»
fi
T
T
T
n
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x ` B
ÿ
β
X
1
fl
“ var –
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
n i“1
tfβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu ` Au2
”
ı
»
fi
T
T
T
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
β
X
2
fl
ď
var –Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
n
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
”
ı
fi
»
T
T
T
n
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
ÿ
β
X
1
Op pBqfl
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
`4var –
n i“1
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
”
ı
fi
»
T
T
T
n
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
ÿ
β X
1
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
Op pAqfl .
`4var –
n i“1
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
The first part is given by
”
ı
»
fi
T
T
T
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
β X
2
fl
var –Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
n
f 2 T pβ T xqE 2 tIpZ ě Zi q | β T xu
β X
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
41
”
ı
fi2
T
T
T
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
β
X
2 –
fl
E Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
“
n
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
”
ı
fi˛2
¨ »
T
T
T
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
β X
2
fl‚
` ˝E –Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
n
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
´ ”
ı
¯2
T
T
T
ij
B
f
T pβ xqEtIpZ ě zi q | β xu {Bβ x
β X
2
“
Kb2 pzi ´ zqKh2 pβ T xi ´ β T xq
n
fβ4 T X pβ T xqE 4 tIpZ ě zi q | β T xu
»
ˆf pzi , β T xi qSc pzi , xi qfX|β T X pxi , β T xi qfβT X pβ T xi qdzi dxi dβ T xi ` Op1{nq
´ ”
ı
¯2
T
T
T
T
ij
B
f
T pβ xqSpzi , β xqEtSc pzi , Xi q | β xu {Bβ x
β X
2
“
Kb2 pzi ´ zqKh2 pβ T xi ´ β T xq
n
fβ4 T X pβ T xqS 4 pzi , β T xqE 4 tSc pzi , Xi q | β T xu
ˆf pzi , β T xi qEtSc pzi , Xi q | β T xi ufβT X pβ T xi qdzi dxi dβ T xi ` Op1{nq
´ ”
ı
¯2
T
T
T
T
ij
B
f
T pβ xqSpz ` bv, β xqEtSc pz ` bv, Xi q | β xu {Bβ x
β X
2
K 2 pvqK 2 puq
“
nbh
fβ4 T X pβ T xqS 4 pz ` bv, β T xqE 4 tSc pz ` bv, Xi q | β T xu
ˆf pz ` bv, β T x ` huqEtSc pz ` bv, Xi q | β T x ` huufβT X pβ T x ` huqdvdu ` Op1{nq
´ ”
ı
¯2
T
T
T
T
ij
B
f
T pβ xqSpz, β xqEtSc pz, Xi q | β xu {Bβ x
β X
2
“
f pz, β T xqdvdu
K 2 pvqK 2 puq
nbh
fβ3 T X pβ T xqS 4 pz, β T xqE 3 tSc pz, Xi q | β T xu
´ ”
ı
¯2
T
T
T
T
˚
˚
ij
2
B
f
T pβ xqSpz , β xqEtSc pz , Xi q | β xu {Bβ x
β X
b B
f pz ˚ , β T xqv 2 dvdu
K 2 pvqK 2 puq
`
2
nh Bz
fβ3 T X pβ T xqS 4 pz ˚ , β T xqE 3 tSc pz ˚ , Xi q | β T xu
´ ”
ı
¯2
T
T
T
T
ij
2
B
f
pβ
xqSpz,
β
xqEtS
pz,
X
q
|
β
xu
{Bβ
x
c
i
βT X
h
B
`
K 2 pvqK 2 puq
T
T
T
nb Bpβ xq2
fβ4 T X pβ xqS 4 pz, β xqE 4 tSc pz, Xi q | β T xu
ˆf pz, β T xqEtSc pz, Xi q | β T x˚ ufβT X pβ T x˚ qu2 dvdu ` Op1{nq
´ ”
ı
¯2
T
T
T
T
"ż
* "ż
*
B
f
pβ
xqSpz,
β
xqEtS
pz,
X
q
|
β
xu
{Bβ
x
c
i
βT X
2
2
2
T
ď
K pvqdv
K puqdu
f pz, β xq
nbh
fβ3 T X pβ T xqS 4 pz, β T xqE 3 tSc pz, Xi q | β T xu
ˇ
´ ”
ı
¯2 ˇˇ
ˇ
T
T
T
T
˚
˚
ˇ
B fβT X pβ xqSpz , β xqEtSc pz , Xi q | β xu {Bβ x ˇˇ
b
ˇ B2
˚
T
`
sup ˇ
f pz , β xq
ˇ
ˇ
nh z ˚ ,βT x ˇ Bz 2
fβ3 T X pβ T xqS 4 pz ˚ , β T xqE 3 tSc pz ˚ , Xi q | β T xu
ˇ
ˇ
"ż
* "ż
*
ˆ
v 2 K 2 pvqdv
K 2 puqdu
ˇ´ ”
ı
¯2
ˇ
T
T
T
T
ˇ
B
f
T pβ xqSpz, β xqEtSc pz, Xi q | β xu {Bβ x
β X
h
ˇ
sup
`
ˇ
ˇ
nb z,βT x,βT x˚
fβ4 T X pβ T xqS 4 pz, β T xqE 4 tSc pz, Xi q | β T xu
ˇ
ˇ "ż
* "ż
*
ˇ
T
T ˚
T ˚ ˇ
K 2 pvqdv
u2 K 2 puqdu ` Op1{nq
ˆf pz, β xqEtSc pz, Xi q | β x ufβT X pβ x qˇ
ˇ
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
42
“ Ot1{pnbhq ` b{pnhq ` h{pnbq ` 1{nu
uniformly under conditions C1-C6. Noting that A “ Op pn´1{2 h´1{2 `h2 q and B “ Op pn´1{2 h´3{2 `
h2 q, the second part is
”
ı
fi
»
T
T
T
n
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
ÿ
β
X
1
Op pBqfl
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
4var –
n i“1
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
¨
”
ı
»
fi2 ˛
T
T
T
n
B
f
pβ
zqEtIpZ
ě
Z
q
|
β
xu
{Bβ
x
T
i
β X
˚1 ÿ –
fl ‹
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
ď 4E ˝
‚
T
2
n i“1
fβT X pβ xqE 2 tIpZ ě Zi q | β T xu
ˆOt1{pnhq ` h4 u
´ ”
ı
¯2
T
T
T
T
B
f
T pβ xqSpz, β xqEtSc pz, Xi q | β xu {Bβ x
β X
˚ 4
f pz, β T xq
“ ˝
nbh
fβ3 T X pβ T xqS 4 pz, β T xqE 3 tSc pz, Xi q | β T xu
"ż
* "ż
*
˙
2
2
ˆ
K pvqdv
K puqdu ` Otb{pnhq ` h{pnbqu Opn´1 h´3 ` h4 q
¨
“ Otn´2 b´1 h´4 ` pnbq´1 h3 u
under conditions C1-C6 uniformly. The last part is
”
ı
fi
»
T
T
n
B
f
pβ
zqEtIpZ
ě
Z
q
|
β
xu
{Bβ T x
T
ÿ
i
β
X
1
Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
Op pAqfl
4var –
n i“1
fβ2 T X pβ T xqE 2 tIpZ ě Zi q | β T xu
¨
”
ı
»
fi2 ˛
T
T
T
n
B
f
T pβ zqEtIpZ ě Zi q | β xu {Bβ x
ÿ
β X
˚1
–Kb pZi ´ zq∆i Kh pβ T Xi ´ β T xq
fl ‹
ď 4E ˝
‚
T
T
2
2
n i“1
f T pβ xqE tIpZ ě Zi q | β xu
β X
ˆOtpnhq´1 ` h4 u
´ ”
ı
¯2
T
T
T
T
B
f
T pβ xqSpz, β xqEtSc pz, Xi q | β xu {Bβ x
β X
˚ 4
“ ˝
f pz, β T xq
nbh
fβ3 T X pβ T xqS 4 pz, β T xqE 3 tSc pz, Xi q | β T xu
"ż
* "ż
*
˙
ˆ
K 2 pvqdv
K 2 puqdu ` Otb{pnhq ` h{pnbqu Opn´1 h´1 ` h4 q
¨
“ Otpnhq´2 b´1 ` h3 pnbq´1 u
under conditions C1-C6 uniformly. Therefore,
p 12 q “ Ot1{pnbhqu
varpλ
uniformly under conditions C1-C6.
p 1 pz, β T xqu “ Ot1{pnbh3 qu uniformly. Hence the estimator
Summarizing the results above, vartλ
T
p 1 pz, β xq satisfies
λ
p 1 px, β T xq “ λ1 px, β T xq ` Op tpnbh3 q´1{2 ` h2 ` b2 u
λ
uniformly under conditions C1-C6.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
43
p satisfy
A.3. Proof of Theorem 1. For each n, let β
n
)fi
!
»
T
p T Xi
p Xli Yi pZi q | β
n
E
ÿ
p
p
n
1
λ1 pZi , β n Xi q –
) fl “ 0.
!
b Xli ´
∆i
T
T
n i“1
p
p
p
p
λpZi , β Xi q
E Yi pZi q | β Xi
n
n
p , n “ 1, 2, . . . , that converges. For notational
Under condition C3, there exists a subsequence of β
n
p , n “ 1, 2, . . . , as the subsequence that converges and let the limit be
simplicity, we still write β
n
β˚.
From the uniform convergence in (12), (13), (16), (17) given in Lemma 1, we obtain
)fi
!
»
T
p T Xi
p Xli Yi pZi q | β
n
E
ÿ
p
p
n
λ1 pZi , β n Xi q –
1
) fl
!
∆i
b Xli ´
n i“1
p i, β
p T Xi q
p T Xi
p Yi pZi q | β
λpZ
E
n
n
n
p T Xi q ` Op tpnbh3 q´1{2 ` h2 ` b2 u
1 ÿ λ1 pZi , β
n
∆i
T
n i“1
p Xi q ` Op tpnbhq´1{2 ` h2 ` b2 u
λpZi , β
n
)
!
fi
»
p T Xi ` Op tpnhq´1{2 ` h2 u
E Xli Yi pZi q | β
n
fl
)
!
b –Xli ´
p T Xi ` Op tpnhq´1{2 ` h2 u
E Yi pZi q | β
n
«
ff
T
n
p Xi q
1 ÿ
λ1 pZi , β
n
“
∆i
` Op tpnbh3 q´1{2 ` h2 ` b2 u
T
n i“1
p
λpZi , β n Xi q
)
!
fi
»
p T Xi
E Xli Yi pZi q | β
n
) ` Op tpnhq´1{2 ` h2 ufl
!
b –Xli ´
T
p
E Yi pZi q | β n Xi
)fi
!
»
T
p T Xi
n
E
X
Y
pZ
q
|
β
ÿ
p
i
i
li
n
1
λ1 pZi , β n Xi q –
) fl ` op p1q.
!
b Xli ´
“
∆i
T
T
n i“1
p
p
λpZi , β Xi q
E Yi pZi q | β Xi
“
n
n
Thus, for sufficiently large n, we have
)fi
!
T
p
X
E
X
Y
pZ
q
|
β
i
li i
n i
1
) fl
!
b –Xli ´
T
n
p Xi
E Yi pZi q | β
n
)fi
!
»
˚T
n
˚T
X
E
X
Y
pZ
q
|
β
ÿ
i
i
i
li
λ1 pZi , β Xi q –
1
p ´ β˚ q
) fl ` Op pβ
!
∆i
b Xli ´
n
˚T
n i“1
λpZi , β ˚ T Xi q
E Yi pZi q | β Xi
)fi
!
»
˚T
n
˚T
X
E
X
Y
pZ
q
|
β
ÿ
i
i
i
li
λ1 pZi , β Xi q –
1
) fl ` op p1q,
!
∆i
b Xli ´
n i“1
λpZi , β ˚ T Xi q
E Yi pZi q | β ˚ T Xi
p T Xi q
λ1 pZi , β
n
∆i
T
p
λpZi , β n Xi q
i“1
n
ÿ
“
“
»
where the first equality is because the first derivative of the summation with respect to β is bounded
p converges to
uniformly under conditions C1-C2 by Lemma 1, and the last equality is because β
n
˚
β . In addition,
)fi
!
»
˚T
n
˚T
X
E
X
Y
pZ
q
|
β
ÿ
i
i
i
li
λ1 pZi , β Xi q –
1
) fl
!
b Xli ´
∆i
n i“1
λpZi , β ˚ T Xi q
E Yi pZi q | β ˚ T Xi
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
44
) fi˛
!
˚T
X
E
X
Y
pZq
|
β
l
λ1 pZ, β Xq –
) fl‚` op p1q
!
b Xl ´
“ E ˝∆
λpZ, β ˚ T Xq
E Y pZq | β ˚ T Xi
¨
˚T
»
under conditions C1-C2. Thus, for sufficient large n we have
)fi
!
»
T
p T Xi
p Xli Yi pZi q | β
n
E
p
p
ÿ
n
1
λ1 pZi , β n Xi q –
) fl
!
0 “
b Xli ´
∆i
T
T
n i“1
p
p
p
p
λpZi , β n Xi q
E Yi pZi q | β n Xi
)fi
!
»
T
p T Xi
n
E
X
Y
pZ
q
|
β
ÿ
p
i
i
li
n
1
λ1 pZi , β n Xi q –
) fl ` op p1q
!
“
∆i
b Xli ´
T
T
n i“1
p
p
λpZi , β n Xi q
E Yi pZi q | β n Xi
)fi
!
»
˚T
n
T
˚
E
X
Y
pZ
q
|
β
X
i
i
li i
1 ÿ λ1 pZi , β Xi q –
) fl ` op p1q
!
∆i
b
X
´
“
li
T
n i“1
λpZi , β ˚ Xi q
E Yi pZi q | β ˚ T Xi
!
) fi˛
»
¨
˚T
˚T
X
Y
pZq
|
β
E
X
l
λ1 pZ, β Xq –
) fl‚` op p1q
!
b Xl ´
“ E ˝∆
˚T
λpZ, β ˚ T Xq
E Y pZq | β Xi
under conditions C1-C2 and C3. Note that
!
) fi˛
¨
»
˚T
˚T
E
X
Y
pZq
|
β
X
l
λ1 pZ, β Xq –
) fl‚
!
E ˝∆
b
X
´
l
T
λpZ, β ˚ Xq
E Y pZq | β ˚ T Xi
is a nonrandom quantity that does not depend on n, hence it is zero. Thus the uniqueness requirement in Condition C7 ensures that β ˚ “ β 0 .
We now show that the subsequence that converges includes all but a finite number of n’s. Assume
p n ’s that do not converge to β ˚ . As
this is not the case, then we can obtain an infinite sequence of β
an infinite sequence in a compact set B, we can thus obtain another subsequence that converges,
say to β ˚˚ ‰ β ˚ . Identical derivation as before then leads to β ˚˚ “ β 0 , which is a contradiction to
p ´ β Ñ 0 in probability when n Ñ 8 under condition C1-C7.
β ˚˚ ‰ β ˚ . Thus we conclude β
0
A.4. Proof of Theorem 2. We first expand (9) as
)fi
!
»
T
T
p
p
n
E
X
Y
pZ
q
|
β
X
ÿ λ
p 1 pZi , β
p Xi q
i
i
li i
) fl
!
b –Xli ´
∆i
0 “ n´1{2
T
T
p i, β
p Xi q
p Xi
p Yi pZi q | β
λpZ
E
i“1
«
(ff
n
p Xli Yi pZi q | β T Xi
ÿ
p 1 pZi , β T Xi q
E
λ
0
0
(A.8) “ n´1{2
∆i
b Xli ´
(
p i , β T Xi q
p Yi pZi q | β T Xi
E
λpZ
0
0
i“1
˜
#
+ˇ
«
( ff¸
n
ˇ
p Xli Yi pZi q | β T Xi
p 1 pZi , β T Xi q
E
λ
B
1ÿ
T ˇ
∆
`
b
X
b
X
´
(
i
li
li ˇ
p i , β T Xi q
p Yi pZi q | β T Xi
ˇ
n
BpXT
E
λpZ
i βq
i“1
(A.9)
?
p ´ β q,
ˆ npβ
0
r is on the line connecting β and β.
p
where β
0
r
β“β
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
45
We first consider (A.9). Because of Theorem 1 and Lemma 1, we have
#
+ˇ
˜
«
( ff¸
n
ˇ
p Xli Yi pZi q | β T Xi
p 1 pZi , β T Xi q
E
B
1 ÿ
λ
T ˇ
b
X
b
X
´
∆
(
i
li
li ˇ
p i , β T Xi q
p Yi pZi q | β T Xi
ˇ r
n i“1 BpXT
E
λpZ
i βq
β“β
+
ff¸
˜
«
#
(
n
T
p Xli Yi pZi q | β Xi
p 1 pZi , β T Xi q
E
λ
1 ÿ
B
0
0
b XT
∆
b
X
´
“
(
i
li
li
T
T
p
p
n i“1 BpXT
E Yi pZi q | β 0 Xi
λpZi , β 0 Xi q
i β0 q
`op p1q
˜
¸
«
(ff
n
p Xli Yi pZi q | β T
p b2 pZi , β T Xi q
E
X
1ÿ
λ
i
0
0
“ ´
∆i 1
(A.10)
b XT
b Xli ´
(
li
p2 pZi , β T Xi q
p Yi pZi q | β T
n i“1
E
X
λ
0 i
0
«
˜
( ff¸
n
p Xli Yi pZi q | β T
E
X
∆i
1ÿ
B
i
0
p 1 pZi , β T Xi q b Xli ´
λ
`
b XT
(
0
li
T
p i , β T Xi q BpXT
p
n i“1 λpZ
β
q
E
Y
pZ
q
|
β
X
i
i
i 0
0 i
0
(A.11)
`op p1q.
Because of Lemma 1, (A.10) converges uniformly in probability to
˜ż
¸
«
(ff
T
8 b2
E
X
Y
psq
|
β
X
λ1 ps, β T
Xq
l
0
0
(
´E
b XT
b Xl ´
l dN psq
T
2 ps, β T Xq
λ
E
Y
psq
|
β
X
0
0
0
¸
«
˜ż
(ff
8 b2
E Xl Y psq | β T
λ1 ps, β T
0X
0 Xq
T
T
(
b Xl Y psqλps, β 0 Xqds
b Xl ´
“ ´E
λ2 ps, β T
E Y psq | β T
0
0 Xq
0X
¨
˛
«
( ffT
(ff «
ż 8 b2
T
T
T
E
X
Y
psq
|
β
X
E
X
Y
psq
|
β
X
λ
ps,
β
Xq
l
l
0
0
0
1
(
(
“ ´E ˝
b Xl ´
b Xl ´
Y psqds‚
T
T
T
λps,
β
E
Y
psq
|
β
E
Y
psq
|
β
Xq
X
0
0
0
0X
«
¸
˜ż
(T
(ff
8 b2
E Xl Y psq | β T
E Xl Y psq | β T
λ1 ps, β T
0X
0X
0 Xq
(
( Y psqds
b
b Xl ´
´E
λps, β T
E Y psq | β T
E Y psq | β T
0
0 Xq
0X
0X
“ ´EtSeff p∆, Z, Xqb2 u,
where the last equality is because the second term above is zero by first taking expectation conditional on β T
0 X.
Similarly, from Lemma 1, the term in (A.12) converges uniformly in probability to the limit of
#
+
«
˜
( ff¸
T
E
X
Y
pZ
q
|
β
X
∆i
B
i
i
i
li
0
p 1 pZi , β T Xi q b Xli ´
(
E
b XT
λ
li .
0
Tβ q
T
BpX
λpZi , β T
X
q
E
Y
pZ
q
|
β
X
i
i
i
i
0
i
0
0
p 1,´i pZ, β T Xq be the leave-one-out version of λ
p 1 pZ, β T Xq, i.e. it is constructed the same
Now let λ
0
0
T
p
as λ1 pZ, β 0 Xq except that the ith observation is not used. Obviously,
«
˜
( ff¸
T
E
X
Y
pZ
q
|
β
X
B
∆i
i
i
i
li
0
p 1 pZi , β T Xi q b Xli ´
(
b XT
λ
0
li
Tβ q
T
BpX
λpZi , β T
X
q
E
Y
pZ
q
|
β
X
i
i
i
i 0
0
0 i
«
˜
( ff¸
E Xli Yi pZi q | β T
∆i
B
0 Xi
T
p
(
´
b XT
λ1,´i pZi , β 0 Xi q b Xli ´
li
Tβ q
T
BpX
λpZi , β T
X
q
E
Y
pZ
q
|
β
X
i
i
i
i 0
0
0 i
“ op p1q.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
46
Now let Ei mean taking expectation with respect to the ith observation conditional on all other
observations, then
+
˜
«
#
( ff¸
T
E
X
Y
pZ
q
|
β
X
B
∆i
i
i
i
li
0
T
T
p 1,´i pZi , β Xi q b Xli ´
(
b Xli
λ
Ei
0
T
λpZi , β T
E Yi pZi q | β T
0 Xi q BpXi β 0 q
0 Xi
˜
#ż
«
+
( ff¸
T
E
X
Y
psq
|
β
X
1
B
i
i
li
0
p 1,´i ps, β T Xi q b Xli ´
(
“ Ei
λ
b XT
0
li dNi psq
Tβ q
T
BpX
λps, β T
E
Y
psq
|
β
X
q
X
i
i
i
0
i
0
0
˜
#ż
«
+
( ff¸
T
E
X
Y
psq
|
β
X
B
i
i
li
0
p 1,´i ps, β T Xi q b Xli ´
(
“ Ei
λ
b XT
0
li Yi psqds
T
BpXT
β
q
E
Y
psq
|
β
X
i
i
i 0
0
#
+
«
(ff
ż
T
E
X
Y
psq
|
β
X
B
i
li i
0
p 1,´i ps, β T Xi q b Xli ´
(
“ Ei
EtYi psq | Xi uds
λ
0
Bβ 0
E Yi psq | β T
X
0 i
+
#ż
«
(ff
T
E
X
Y
psq
|
β
X
B
i
i
li
0
T
p 1,´i ps, β Xi q b XlI ´
(
EtYi psq | Xuds
Ei
λ
“
0
Bβ 0
E Yi psq | β T
0 Xi
+
#ż
«
(ff
T
E
X
Y
psq
|
β
X
B
i
i
li
0
T
p 1,´i ps, β Xi q b Xli ´
(
Yi psqds
“
Ei
λ
0
Bβ 0
E Yi psq | β T
0 Xi
“ 0.
Here, the last equality is because the integrand has expectation zero conditional on β T
0 Xi and all
other observations, and the third last equality is because the expectation is with respect to Xi and
does not involve β 0 . Therefore, the term in (A.12) converges in probability uniformly to
+
«
˜
#
( ff¸
T
E
X
Y
pZ
q
|
β
X
B
∆i
i
i
i
li
0
p 1,´i pZi , β T Xi q b Xli ´
(
“0
b XT
λ
E
li
0
Tβ q
T
BpX
λpZi , β T
X
q
E
Y
pZ
q
|
β
X
i
i
i
i 0
0
0 i
Combining the results concerning (A.10) and (A.12), we thus have obtained that the expression in
(A.9) is ´EtSeff p∆, Z, Xqb2 u ` op p1q.
Next we decompose (A.8) into
«
(ff
n
p Xli Yi pZi q | β T Xi
ÿ
p 1 pZi , β T Xi q
E
λ
0
0
∆i
(A.12)n´1{2
“ T1 ` T2 ` T3 ` T4 ,
b Xli ´
(
p i , β T Xi q
p Yi pZi q | β T Xi
E
λpZ
0
i“1
where
T1 “
T2 “
T3 “
T4 “
0
«
(ff
T
T
E
X
Y
pZ
q
|
β
X
λ
pZ
,
β
X
q
i
i
i
li
1
i
i
0
0
(
,
n´1{2
∆i
b Xli ´
T
T
λpZ
,
β
X
q
E
Y
pZ
q
|
β
X
i
i
i
i
0
0 i
i“1
#
+ «
(ff
n
T
ÿ
p 1 pZi , β T Xi q λ1 pZi , β T Xi q
E
X
Y
pZ
q
|
β
X
λ
i
i
i
li
0
0
0
(
∆i
n´1{2
,
´
b Xli ´
T
T
p i , β T Xi q
λpZ
,
β
X
q
E
Y
pZ
q
|
β
X
λpZ
i
i
i
i
0 i
0
0
i“1
«
(ff
(
n
p Xli Yi pZi q | β T
ÿ
E
E Xli Yi pZi q | β T
λ1 pZi , β T
0 Xi
0 Xi
0 Xi q
´1{2
( ´
∆i
n
,
b
(
p Yi pZi q | β T
λpZi , β T
E Yi pZi q | β T
E
0 Xi q
0 Xi
0 Xi
i“1
+
#
n
ÿ
p 1 pZi , β T Xi q λ1 pZi , β T Xi q
λ
0
0
´1{2
∆i
n
´
p i , β T Xi q
λpZi , β T
λpZ
0 Xi q
0
i“1
«
(
(ff
T
p Xli Yi pZi q | β T
E Xli Yi pZi q | β 0 Xi
E
0 Xi
( ´
b
.
(
p Yi pZi q | β T
E Yi pZi q | β T
E
0 Xi
0 Xi
n
ÿ
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
47
First, note that
T2
+ «
(ff
T
p 1 ps, β T Xi q λ1 ps, β T Xi q
E
X
Y
psq
|
β
X
λ
i
i
li
0
0
0
(
dNi psq
b Xli ´
´
“ n´1{2
T
p β T Xi q
λps,
β
E Yi psq | β T
Xi q
λps,
0
0 Xi
0
i“1
«
˜
¸
(ff
n ż
T
ÿ
E
X
Y
psq
|
β
X
i
i
li
0
(
Xli ´
“ op n´1{2
Yi psqλps, β T
0 Xli qds
T
E
Y
psq
|
β
X
i
i
0
i“1
n ż
ÿ
#
“ op p1q,
where the last equality above is because the quantity inside the parenthesis is a mean zero normal
random quantity of order Op p1q. Further,
˜
(
n
p Xli Yi pZi q | β T
ÿ
E
λ1 pZi , β T
0 Xi
0 Xi q
´1{2
(
∆i
T3 “ n
b ´
T
T
λpZ
,
β
X
q
E
Y
pZ
q
|
β
X
i
i
i
i
i
0
0
i“1
(¸
(
T
p
E Yi pZi q | β T
0 Xi E Xli Yi pZi q | β 0 Xi
(
`
` op p1q
2
rE Yi pZi q | β T
0 Xi s
˜
řn
T
T
n
´1
T
ÿ
n
λ
pZ
,
β
X
q
j“1 Kh pβ 0 Xj ´ β 0 Xi qXlj IpZj ě Zi q
1
i
i
0
´1{2
(
∆i
“ n
b ´
T
λpZi , β T
fβT X pβ T
0 Xi q
0 Xi qE Yi pZi q | β 0 Xi
i“1
0
( ´1 řn
T
T
pβ T
E Xli Yi pZi q | β T
0 Xi qu
0 Xi tn
j“1 Kh pβ 0 Xj ´ β 0 Xi q ´ fβT
0X
(
`
T
T
fβT X pβ 0 Xi qE Yi pZi q | β 0 Xi
0
( « ´1 řn
T
T
T
n
E Xli Yi pZi q | β 0 Xi
j“1 Kh pβ 0 Xj ´ β 0 Xi qIpZj ě Zi q
(
`
2
rE Yi pZi q | β T
fβT X pβ T
0 Xi s
0 Xi q
0
ff¸
(
ř
n
T
T
T
Xi qu
E Yi pZi q | β 0 Xi tn´1 j“1 Kh pβ 0 Xj ´ β 0 Xi q ´ fβT X pβ T
0
0
` op p1q
´
fβT X pβ T
X
q
i
0
0
«
n
n ÿ
T
T
ÿ
Kh pβ T
λ1 pZi , β 0 Xi q
0 Xj ´ β 0 Xi qXlj IpZj ě Zi q
´3{2
(
∆i
“ n
b
´
T
T
T
λpZ
,
β
X
q
f
T pβ 0 Xi qE Yi pZi q | β 0 Xi
i
i
0
β
X
i“1 j“1
0
ff
(
T
T
T
E Xli Yi pZi q | β 0 Xi Kh pβ 0 Xj ´ β 0 Xi qIpZj ě Zi q
(
`
` op p1q
T
2
fβT X pβ T
0 Xi qrE Yi pZi q | β 0 Xi s
0
“ T31 ` T32 ` T33 ` op p1q,
where
T31
T32
«
T
Kh pβ T
λ1 pZi , β T
0 Xj ´ β 0 Xi qXlj IpZj ě Zi q
0 Xi q
(
b
E
´
∆i
“ n
T
T
T
λpZ
,
β
X
q
f
T pβ 0 Xi qE Yi pZi q | β 0 Xi
i
i
0
β0 X
i“1
ff
(
T
T
E Xli Yi pZi q | β 0 Xi Kh pβ 0 Xj ´ β T
0 Xi qIpZj ě Zi q
(
`
| ∆ i , Z i , Xi
T
2
fβT X pβ T
0 Xi qrE Yi pZi q | β 0 Xi s
0
«
˜
n
T
T
ÿ
Kh pβ T
λ
pZ
,
β
X
q
1
i
i
0 Xj ´ β 0 Xi qXlj IpZj ě Zi q
0
(
b
´
E ∆i
“ n´1{2
λpZi , β T
fβT X pβ T
Xi qE Yi pZi q | β T
Xi
0 Xi q
0
0
j“1
0
¸
ff
(
T
T
T
E Xli Yi pZi q | β 0 Xi Kh pβ 0 Xj ´ β 0 Xi qIpZj ě Zi q
(
`
| ∆j , Zj , Xj
T
2
fβT X pβ T
0 Xi qrE Yi pZi q | β 0 Xi s
´1{2
n
ÿ
0
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
48
T33
«
T
Kh pβ T
λ1 pZi , β T
0 Xj ´ β 0 Xi qXlj IpZj ě Zi q
0 Xi q
(
b
E
´
“ ´n E ∆i
λpZi , β T
fβT X pβ T
Xi qE Yi pZi q | β T
Xi
0 Xi q
0
0
0
ff¸
(
T
T
E Xli Yi pZi q | β 0 Xi Kh pβ 0 Xj ´ β T
0 Xi qIpZj ě Zi q
(
`
.
T
2
fβT X pβ T
0 Xi qrE Yi pZi q | β 0 Xi s
1{2
˜
0
Here we used U-statistic property in the last equality above. Now when nh4 Ñ 0,
«
n
T
ÿ
EtXli Yi pZi q | β T
λ
pZ
,
β
X
q
1
i
i
0 Xi u
0
(
b
´
∆i
T31 “ n´1{2
T
T
λpZi , β 0 Xi q
E Yi pZi q | β 0 Xi
i“1
(ff
(
T
E
Y
pZ
q
|
β
X
E Xli Yi pZi q | β T
X
i
i
i
i
0
0
(
`
` Opn1{2 h2 q
2
s
rE Yi pZi q | β T
X
i
0
“ op p1q.
Thus, T33 “ op p1q as well. To analyze T32 ,
˜
«
n
T
T
ÿ
Kh pβ T
X
q
λ
pZ
,
β
i
1
i
0 Xj ´ β 0 Xi qXlj IpZj ě Zi q
0
(
T32 “ n´1{2
E ∆i
b
´
T
T
λpZi , β T
fβT X pβ T
0 Xi q
0 Xi qE IpZ ě Zi q | β 0 X “ β 0 Xi , Zi
j“1
0
¸
ff
(
T
T
T
E Xl IpZ ě Zi q | β T
0 X “ β 0 Xi , Zi Kh pβ 0 Xj ´ β 0 Xi qIpZj ě Zi q
(
| ∆j , Zj , Xj
`
fβT X pβ T
Xi qrE IpZ ě Zi q | β T
X “ βT
Xi , Zi s2
0
0
0
˜0
«
n
T
T
ÿ
Kh pβ T
λ
pZ
,
β
X
q
1
i
i
0 xj ´ β 0 Xi qxlj Ipzj ě Zi q
0
(
“ n´1{2
E ∆i
b
´
T
T
λpZi , β T
fβT X pβ T
0 Xi q
0 Xi qE IpZ ě Zi q | β 0 X “ β 0 Xi , Zi
j“1
0
ff¸
(
T
T
T
E Xl IpZ ě Zi q | β T
0 X “ β 0 Xi , Zi Kh pβ 0 xj ´ β 0 Xi qIpzj ě Zi q
(
`
fβT X pβ T
Xi qrE IpZ ě Zi q | β T
X “ βT
Xi , Zi s2
0
0
0
#0 ˜
«
n
T
ÿ
xlj Ipzj ě Zi q
λ
pZ
,
β
X
q
1
i
i
0
(
E E ∆i
“ n´1{2
b ´
T
T
T
λpZi , β 0 Xi q
fβT X pβ 0 Xi qE IpZ ě Zi q | β T
0 X “ β 0 Xi , Zi
j“1
0
¸
+
ff
(
T
E Xl IpZ ě Zi q | β T
0 X “ β 0 Xi , Zi Ipzj ě Zi q
T
T
T
(
| β 0 Xi Kh pβ 0 xj ´ β 0 Xi q
`
fβT X pβ T
Xi qrE IpZ ě Zi q | β T
X “ βT
Xi , Zi s2
0
0
0
0
˜
«
n
ÿ
xlj Ipzj ě Zi q
λ1 pZi , β T
0 xj q
´1{2
(
E ∆i
“ n
b ´
T
T
T
λpZ
,
β
E
IpZ
ě
Z
q
|
β
x
q
X
“
β
x
,
Z
i
i
j
j
i
0
0
0
j“1
¸
ff
(
T
Ipz
ě
Z
q
E Xl IpZ ě Zi q | β T
X
“
β
x
,
Z
j
i
i
0
0 j
1{2 2
T
(
h q
| βT
`
0 Xi “ β 0 xj ` Op pn
T
2
rE IpZ ě Zi q | β T
s
X
“
β
x
,
Z
i
0
0 j
˜
n
ÿ
∆i Ipzj ě Zi qλ1 pZi , β T
0 xj q
´1{2
(
“ n
E
T
T
T
SpZi , β 0 xj qλpZi , β 0 xj qE Sc pZi , Xq | β T
0 X “ β 0 xj , Zi
j“1
¸
ff
«
(
T
E Xl Sc pZi , Xq | β T
0 X “ β 0 xj , Zi
T
( ´ xlj | β T
` Op pn1{2 h2 q
b
0 Xi “ β 0 xj
T
E Sc pZi , Xq | β T
X
“
β
x
,
Z
j
i
0
0
˜ż
n
zj
ÿ
λ1 ps, β T
0 xj q
(
“ n´1{2
E
T
T
0 E Sc ps, Xq | β 0 X “ β 0 xj
j“1
¸
ff
«
(
T
E Xl Sc ps, Xq | β T
0 X “ β 0 xj
T
T
( ´ xlj Sc ps, Xi qds | β 0 Xi “ β 0 xj ` Op pn1{2 h2 q
b
T
T
E Sc ps, Xq | β 0 X “ β 0 xj
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
49
ff
(
E Xlj Sc ps, Xj q | β T
0 xj
( ´ xlj ds ` Op pn1{2 h2 q
b
“ n
T
E Sc ps, Xj q | β 0 xj
j“1 0
ff
«
(
ż
n
T
T
ÿ
E
X
Y
psq
|
β
x
λ
ps,
β
x
q
j
j
lj
1
0
0 j
( ´ xlj ds ` Op pn1{2 h2 q.
b
Yj psqλps, β T
“ n´1{2
0 xj q
T
T
λps,
β
x
q
E
Y
psq
|
β
x
j
0 j
0 j
j“1
´1{2
n ż zj
ÿ
λ1 ps, β T
0 xj q
«
When nh4 Ñ 0, plugging the results of T1 and T32 to (A.12), we obtain that the expression in
(A.8) is
«
(ff
n
p Xli Yi pZi q | β T
ÿ
p 1 pZi , β T Xi q
E
λ
0 Xi
0
´1{2
b Xli ´
∆i
n
(
p i , β T Xi q
p Yi pZi q | β T Xi
E
λpZ
0
0
i“1
«
(ff
n ż
ÿ
E Xli Yi ptq | β T
λ1 pt, β T
0 Xi
0 Xi q
´1{2
(
“ n
dMi ptq ` op p1q
b Xli ´
T
T
λpt,
β
E
Y
ptq
|
β
X
q
X
i
i
i
0
0
i“1
n
ÿ
“ n´1{2
Seff p∆i , Zi , Xi q ` op p1q.
i“1
Finally, note that
T4
+
p 1 pZi , β T Xi q λ1 pZi , β T Xi q
λ
0
0
´
∆i
“ n
p i , β T Xi q
λpZi , β T
λpZ
0 Xi q
0
i“1
«
(
(ff
T
p Xli Yi pZi q | β T
E Xli Yi pZi q | β 0 Xi
E
0 Xi
( ´
ˆ
(
p Yi pZi q | β T Xi
E Yi pZi q | β T
E
0 Xi
0
«
˜
( ff¸
(
n
T
p Xli Yi pZi q | β T
ÿ
E
E
X
Y
pZ
q
|
β
X
X
i
i
i
i
li
0
0
( ´
∆i
“ op n´1{2
(
p Yi pZi q | β T Xi
E Yi pZi q | β T
E
0 Xi
0
i“1
ff ¸
«
˜
(
n ż
T
ÿ
E
X
Y
psq
|
β
x
i
i
li
0
( ´ xli ds ` op pn1{2 h2 q
Yi psqλps, β T
“ op n´1{2
0 xi q
T
E
Y
psq
|
β
x
i
0 i
i“1
´1{2
n
ÿ
#
“ op p1q,
where the last equality is because the integrands have mean zero conditional on β T
0 X, and the
second last equality is obtained following the same derivation of T3 . Using these results in (A.8),
combined with the results on (A.9), it is now clear that the theorem holds.
References.
Buckley, J. and James, I. (1979). Linear regression with censored ta. Biometrika 66 429–436.
Cox, D. R. (1972). Regression models and life tables (with Discussion). Journal of the Royal Statistical Society,
Series B. 34 187–220.
Cox, D. R. (1975). Partial likelihood. Biometrika 62 269–276.
Dabrowska, D. M. and Doksum, K. A. (1988). Partial likelihood in transformation models with censored data.
Journal of the Royal Statistical Society, Series B. 18 1–23.
Fleming, T. R. and Harrington, D. P. (1991). Counting processes and survival analysis. Wiley, New York.
Geng, Y., Lu, W. and Zhang, H. H. (2015). On optimal treatment regimes selection for mean survival time.
Statistics in Medicine 34 1169–1184.
Hammer, S. M., Katzenstein, D. A., Hughes, M. D., Gundacker, H., Schooley, R. T., Haubrich, R. H.,
Henry, W. K., Lederman, M. M., Phair, J. P., Niu, M. et al. (1996). A trial comparing nucleoside monotherapy
with combination therapy in HIV-infected adults with CD4 cell counts from 200 to 500 per cubic millimeter. New
England Journal of Medicine 335 1081–1090.
Jiang, R., Lu, W., Song, R. and Davidian, M. (2017). On estimation of optimal treatment regimes for maximizing
t-year survival probabilities. Journal of the Royal Statistical Society: Series B 79 1165–1185.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
50
Li, K. C., Wang, J. L. and Chen, C. H. (1999). Dimension reduction for censored regression data. The Annals of
Statistics 27 1–23.
Lu, W. and Li, L. (2011). Sufficient dimension reduction for censored regressions. Biometrics 67 513–523.
Ma, Y., Zhang, X. et al. (2015). A validated information criterion to determine the structural dimension in dimension reduction models. Biometrika 102 409–420.
Mack, Y. P. and Silverman, B. W. (1982). Weak and strong uniform consistency of kernel regression estimates.
Probability Theory and Related Fields 61 405–415.
McCullagh, P. (1980). Regression models for ordinal data. Journal of the Royal Statistical Society, Series B. 42
109–142.
Tsiatis, A. A. (2006). Semiparametric Theory and Missing Data. Springer, New York.
Xia, Y., Zhang, D. and Xu, J. (2010). Dimension reduction and semiparametric estimation of survival models.
Journal of the American Statistical Association 105 278–290.
Zeng, D. L. and Lin, D. Y. (2007). Maximum likelihood estimation in semiparametric regression models with
censored data (with discussion). Journal of the Royal Statistical Society, Series B. 69 507–564.
imsart-aos ver. 2014/10/16 file: SDR_survival.tex date: October 17, 2017
β4=0
β5=1
β6=0
β7=0
β8=0
1.5
1.5
1.5
1.5
1.5
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
−0.5
hmave semi
−0.5
hmave semi
−0.5
hmave semi
−0.5
hmave semi
−0.5
hmave semi
β4=0
β5=1
β6=0
β7=0
β8=0
1.5
1.5
1.5
1.5
1.5
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
−0.5
hmave semi
−0.5
hmave semi
−0.5
hmave semi
−0.5
hmave semi
−0.5
hmave sem
β4=0
β5=1
β6=0
β7=0
β8=0
1.5
1.5
1.5
1.5
1.5
1
1
1
1
1
0.5
0.5
0.5
0.5
0.5
0
0
0
0
0
−0.5
hmave semi
−0.5
hmave semi
−0.5
hmave semi
−0.5
hmave semi
−0.5
hmave semi
| 10math.ST
|
DATA AUGMENTATION BY PAIRING S AMPLES
FOR I MAGES C LASSIFICATION
Hiroshi Inoue
IBM Research - Tokyo
[email protected]
arXiv:1801.02929v1 [cs.LG] 9 Jan 2018
A BSTRACT
Data augmentation is a widely used technique in many machine learning tasks,
such as image classification, to virtually enlarge the training dataset size and
avoid overfitting. Traditional data augmentation techniques for image classification tasks create new samples from the original training data by, for example,
flipping, distorting, adding a small amount of noise to, or cropping a patch from
an original image. In this paper, we introduce a simple but surprisingly effective
data augmentation technique for image classification tasks. With our technique,
named SamplePairing, we synthesize a new sample from one image by overlaying another image randomly chosen from the training data (i.e., taking an average
of two images for each pixel). By using two images randomly selected from the
training set, we can generate N 2 new samples from N training samples. This simple data augmentation technique significantly improved classification accuracy for
all the tested datasets; for example, the top-1 error rate was reduced from 33.5% to
29.0% for the ILSVRC 2012 dataset with GoogLeNet and from 8.22% to 6.93%
in the CIFAR-10 dataset. We also show that our SamplePairing technique largely
improved accuracy when the number of samples in the training set was very small.
Therefore, our technique is more valuable for tasks with a limited amount of training data, such as medical imaging tasks.
1
I NTRODUCTION
For machine learning tasks, such as image classification, machine translation, and text-to-speech
synthesis, the amount of the samples available for training is critically important to achieve high
performance by better generalization. Data augmentation, applying a small mutation in the original
training data and synthetically creating new samples, is widely used to virtually increase the amount
of training data (e.g. Krizhevsky et al. (2012), Szegedy et al. (2015), Fadaee et al. (2017)). There are
a wide variety of approaches to synthesizing new samples from original training data; for example,
traditional data augmentation techniques include flipping or distorting the input image, adding a
small amount of noise, or cropping a patch from a random position. Use of data augmentation is
mostly the norm for winning image classification contests.
In this paper, we introduce a simple but surprisingly effective data augmentation technique for image
classification tasks. With our technique, named SamplePairing, we create a new sample from one
image by overlaying another image randomly picked from the training data (i.e., simply taking an
average of two images for each pixel). By using two images randomly selected from the training
set, we can generate N 2 new samples from N training samples. Even if we overlay another image,
we use the label for the first image as the correct label for the overlaid (mixed) sample. This simple
data augmentation technique gives significant improvements in classification accuracy for all tested
datasets: ILSVRC 2012, CIFAR-10, CIFAR-100, and SVHN. For example, the top-1 error rate was
reduced from 33.5% to 29.0% for the ILSVRC 2012 dataset with GoogLeNet and from 8.22% to
6.93% in the CIFAR-10 dataset with a simple 6-layer convolutional network. We also conducted
data augmentation by overlaying an image picked from outside the training set; this approach also
gives some improvements, but our technique, which picks an image to overlay from the training
set, gives a far more significant gain. To show that our SamplePairing technique often gives larger
benefits when the number of samples in the training data is smaller, we conducted the evaluations
by reducing the number of samples used for training with the CIFAR-10. The results showed that
1
our technique yields larger improvements in accuracy when the number of samples are smaller than
the full CIFAR-10 dataset. When we used only 100 samples per label (thus, 1,000 samples in total),
the classification error rate was reduced from 43.1% to 31.0% with our technique. Based on these
results, we believe that our technique is more valuable for tasks with a limited number of training
datasets available, such as medical image classification tasks.
2
R ELATED WORK
When a model is trained with a small training set, the trained model tends to overly fit to the samples
in the training set and results in poor generalization. Data augmentation has been widely used to
avoid this overfitting (poor generalization) problem by enlarging the size of the training set; thus,
it allows the use of a larger network without significant overfitting. For example, Krizhevsky et al.
employ a couple of data augmentation techniques in their Alexnet paper (Krizhevsky et al. (2012))
using random numbers: 1) randomly cropping patches of 224 × 224 pixels from the original image
of 256 × 256 pixels, 2) horizontally flipping the extracted patches, and 3) changing the intensity
of RGB channels. Many submissions in the ImageNet Large Scale Visual Recognition Challenge
(ILCVRC) employed the first two techniques, which can give a 2048x increase in the dataset size.
Our SamplePairing technique is another way to enlarge the dataset size. Since it creates pairs of
images, it gives N 2 samples from the original N samples. Also, our technique is orthogonal to
other data augmentation techniques; as detailed later, our implementation employs the above two
basic data augmentation techniques used in Alexnet in each epoch. In this case, we can synthesize
(2048 × N )2 samples in total.
There are many ways to avoid overfitting other than the data augmentation. Dropout (Hinton et al.
(2012)) and its variants (e.g. Wan et al. (2013)) are famous techniques used to avoid overfitting
by randomly disabling connections during training. Batch normalization (Ioffe & Szegedy (2015)),
which intends to solve the problem of internal covariate shift, is also known to be (potentially)
effective to prevent overfitting. Our SamplePairing data augmentation is again orthogonal to these
techniques and we can employ both techniques to achieve better generalization performance.
Our technique randomly picks an image to overlay from the training set and creates pairs to synthesize a new sample. Our technique is not the first to create a new sample from two randomly
picked original samples; existing work such as Chawla et al. (2002) and Wang & Perez (2017) also
create a new sample from two samples. SMOTE, Synthetic Minority Over-sampling Technique, by
Chawla et al. is a well-known technique for an imbalanced dataset; i.e., the dataset consists of a lot
of ”normal” samples and a small number of ”abnormal” samples. SMOTE does over-sampling of
the minority class by synthesizing a new sample from two randomly picked samples in the feature
space. Comparing our technique against SMOTE, the basic idea of pairing two samples to synthesize a new sample is common between the two techniques; however, we apply the pairing for the
entire dataset instead of for the minority class. Wang and Perez conducted data augmentation by
pairing images using an augmentation network, a separate convolutional network, to create a new
sample. Compared to their work, we simply average the intensity of each channel of pixels from
two images to create a new sample and we do not need to create an augmentation network. Also,
our technique yields more significant improvements compared to their reported results.
A paper on a technique called mixup (Anonymous (2018), Zhang et al. (2017)) is submitted to the
same conference (ICLR 2018). Although developed totally independently, mixup also mixes two
randomly selected samples to create a new training data as we do in this paper. The differences
between our SamplePairing and mixup include 1) mixup overlaying sample labels as well while we
use the label from only one sample, 2) they use weighted average (with randomly selected weight)
to overlay two samples and 3) they tested tasks other than image classification including speech
recognition and Generative Adversarial Network (Goodfellow et al. (2014)). About the label to
be used in training, Huszár (2017) pointed out that using the label from only one sample without
overlaying the labels from two examples will resulted in the same learning result. This way is
simpler than blending two labels and also is suitable for a semi-supervised setting. We also tested
overlaying labels from two samples in our SamplePairing and we did not see the differences in the
results larger than those due to a random number generator (see appendix).
2
label A
image A
(256 x 256)
basic data
augmentation
label A
label A
patch A
mixed
patch
(224 x 224)
(224 x 224)
training
network
mixing
unused
label B
image B
(256 x 256)
basic data
augmentation
randomly picked
from training set
label B
averaging intensity of two
patches for each pixel
(RGB channels)
patch B
(224 x 224)
1) extracting a 224x224 patch from a random
position, 2) horizontal flipping randomly
Figure 1: Overview of our SamplePairing data augmentation for ILSVRC dataset. For other datasets,
the size of the original images is 32 × 32 and the size of the extracted patches is 28× 28.
3
M ETHOD – DATA AUGMENTATION BY SamplePairing
This section describes our SamplePairing data augmentation technique. The basic idea of our technique is simple: We synthesize a new image from two images randomly picked from the training
set as done in some existing techniques, such as Chawla et al. (2002) and Wang & Perez (2017). To
synthesize a new image, we take the most naive approach, just averaging the intensity of two images
for each pixel.
Figure 1 depicts the overview of our sample pairing technique. For each training epoch, all samples
are fed into the network for training in randomized order. For our technique, we randomly picked
another image (image B in Figure 1) from the training set, took an average of the two images,
and fed the mixed image associated with the label for the first image into the network for training.
Hence, essentially, our SamplePairing randomly creates pairs of samples from the training set and
synthesizes a new sample from them. The label for the second image was not used in the training.
Since two images are equally weighted in the mixed image, a classifier cannot correctly predict the
label of the first image (label A in Figure 1) from the mixed image unless label A and label B are
the same label. So, the training loss cannot become zero even using a huge network and the training
accuracy cannot surpass about 50% on average; for N -class classification, the maximum training
accuracy is 0.5 + 1/(N × 2) if all classes have the same number of samples in the training set.
Even though the training accuracy will not be high with SamplePairing, both the training accuracy
and validation accuracy quickly improve when we stop the SamplePairing as the final fine-tuning
phase of training. After the fine-tuning, the network trained with our SamplePairing can achieve
much higher accuracy than the network trained without our technique, as we empirically show in
this paper. The network trained with SamplePairing shows higher training error rate and training loss
than the network trained without SamplePairing even after the fine tuning since the SamplePairing
is a strong regularizer.
When other types of data augmentation are employed in addition to our technique, we can apply
them for each image before mixing them into the final image for training. The data augmentation
incurs additional time to prepare the input image, but this can be done on the CPU while the GPU
is executing the training through back propagation. Therefore, the additional execution time of the
CPU does not visibly increase the total training time per image.
As for the entire training process, we do as follows:
1) We start training without our SamplePairing data augmentation (but with basic data augmentations, including random flipping and random extraction).
3
Table 1: Training and validation sets error rates with and without our SamplePairing data augmentation.
Dataset
training error
without
SamplePairing
CIFAR-10
0.55%
CIFAR-100
SVHN
ILSVRC with
100 classes
ILSVRC with
1000 classes
top-1 error
top-5 error
top-1 error
top-5 error
with
SamplePairing
1.25%
validation error
without
SamplePairing
8.22%
6.93%
-15.68%
10.56%
0.95%
3.21%
26.21%
21.02%
-19.82%
17.58%
33.51%
29.01%
-13.46%
-
1.52%
-
2.08%
-
-
4.28%
8.58%
13.15%
27.9%
reduction in
error rate
5.78%
0.84%
30.5%
with
SamplePairing
4.15%
6.11%
11.36%
-8.58%
-3.05%
-28.74%
-13.55%
2) After completing one epoch (for the ILSVRC) or 100 epochs (for other datasets) without SamplePairing, we enable SamplePairing data augmentation.
3) In this phase, we intermittently disable SamplePairing. For the ILSVRC, we enable SamplePairing for 300,000 images and then disable it for the next 100,000 images. For other datasets, we enable
SamplePairing for 8 epochs and disable it for the next 2 epochs.
4) To complete the training, after the training loss and the accuracy become mostly stable during the
progress of the training, we disable the SamplePairing as the fine-tuning.
We evaluated various training processes; for example, intermittently disabling SamplePairing with
the granularity of a batch or without intermittently disabling it. Although SamplePairing yielded
improvements even with the other training processes, we empirically observed that the above process
gave the highest accuracy in the trained model.
4
4.1
E XPERIMENTS
I MPLEMENTATION
In this section, we investigate the effects of SamplePairing data augmentation using various image classification tasks: ILSVRC 2012, CIFAR-10, CIFAR-100, and Street View House Numbers
(SVHN) datasets.
For the ILSVRC dataset, we used GoogLeNet (Szegedy et al. (2015)) as the network architecture and
trained the network using stochastic gradient descent with momentum as the optimization method
with the batch size of 32 samples. For other datasets, we used a network that has six convolutional
layers with batch normalization (Ioffe & Szegedy (2015)) followed by two fully connected layers
with dropout (Hinton et al. (2012)). We used the same network architecture except for the number
of neurons in the output layer. We trained the network using Adam (Kingma & Ba (2014)) as the
optimizer with the batch size of 100. During the training, we used data augmentations by extracting
a patch from a random position of the input image and using random horizontal flipping as the
baseline regardless of whether or not it was with our SamplePairing as shown in Figure 1. In this
paper, we do not ensemble predictions in the results. For the validation set, we extracted the patch
from the center position and fed it into the classifier without horizontal flipping. We implemented
our algorithm using Chainer (Tokui et al. (2015)) as the framework.
4.2
R ESULTS
We show the improvements in the accuracy by our SamplePairing data augmentation in Table 1. For
the ILSVRC dataset, as well as the full dataset with 1,000 classes, we tested a shrinked dataset with
only the first 100 classes. For all the datasets we evaluated, SamplePairing reduced the classification
error rates for validation sets from 3.1% (for SVHN) up to 28.8% (for the top-5 error rate of the
ILSVRC with 100 classes). For training sets, our SamplePairing increased training errors for all
datasets by avoiding overfitting. When comparing the ILSVRC with 100 classes and with 1,000
4
0.2
Validation error for CIFAR-10
lower is better
0.18
0.14
0.12
fine tuning
0.1
0.08
0.06
0.04
SamplePairing enabled intermittently
(enabled for 8 epochs and then disabled for 2 epochs)
0.02
0
1
0.9
0
50
100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950
epoch
without SamplePairing
with SamplePairing
Validation error for ILSVRC
lower is better
validation set error rate
0.16
validation set top-1 error rate
0.8
0.7
0.6
fine tuning
0.5
0.4
0.3
0.2
0.1
SamplePairing enabled intermittently
(enabled for 300k images and then disabled for 100k images)
0
# trained images (x100k)
without SamplePairing
with SamplePairing
Figure 2: Changes in validation error rates for CIFAR-10 and ILSVRC datasets with and without
our SamplePairing data augmentation.
classes (or the CIFAR-10 and CIFAR-100), the case with 100 classes (CIFAR-10) has a much lower
training error rate and potentially suffers from the overfitting more severely. Therefore, the benefits
of our SamplePairing data augmentation are more significant for the case with 100 classes than the
case with 1,000 classes (for the CIFAR-10 than CIFAR-100). These results show that SamplePairing
yields better generalization performance and achieves higher accuracy.
Figure 2 and Figure 3 illustrate the validation error rates and the training error rates for the CIFAR10 and ILSVRC datasets. From Figure 2, we can see much better final accuracy in trade for longer
training time. Because we disabled SamplePairing data augmentation intermittently (as described in
the previous section), both the training error rates and validation error rates fluctuated significantly.
When the network was under training with SamplePairing enabled, both the validation error rate
and training error rate were quite poor. In particular, the training error rate was about 50% as we
expected; the theoretical best error rate for a 10-class classification task was 45%. Even after the
fine tuning, the networks trained with SamplePairing show more than 2x higher training error rate
for CIFAR-10 and more than 10x higher for ILSVRC. For the validation set, however, when we
disabled the SamplePairing, the validation error rates became much lower than the baseline, which
did not employ SamplePairing at all, after the fine tuning as already shown in Table 1.
We also show the changes in training and valiadation losses for ILSVRC dataset in Figure 4. The
losses match with the training and validation error rates. Without SamplePairing, we achieved the
5
1
SamplePairing enabled intermittently
(enabled for 8 epochs and then disabled for 2 epochs)
0.7
0.6
0.5
0.4
0.3
0.2
fine tuning
0.1
0
0
50
100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950
epoch
without SamplePairing
1
0.9
training set top-1 error rate
0.8
with SamplePairing
Training error for ILSVRC
SamplePairing enabled intermittently
(enabled for 300k images and then disabled for 100k images)
0.7
0.6
lower is better
training set error rate
0.8
lower is better
Training error for CIFAR-10
0.9
0.5
0.4
fine tuning
0.3
0.2
0.1
0
# trained images (x100k)
without SamplePairing
with SamplePairing
Figure 3: Changes in training error rates for CIFAR-10 and ILSVRC datasets with and without our
SamplePairing data augmentation.
minimum validation loss in an early stage of the training and the validation loss gradually increased
after that point. With SamplePairing, we did not see such increase in the validation loss.
Figure 5 shows how our SamplePairing data augmentation causes improvements when the number
of samples available for training is limited; we change the number of training samples in the CIFAR10. The CAFAR-10 dataset provides 5,000 samples for each of 10 classes in the training set. We
gradually reduced the number of images used for training from 5,000 samples to only 10 samples
per class. As shown in Figure 5, the largest gain of SamplePairing was 28% when we used 100
samples per class; the error rate was reduced from 43.1% to 31.0%. When we used only 50 samples
per class, the second largest reduction of 27% was achieved. The error rate using 50 samples with
SamplePairing was actually better than the error rate using 100 samples without SamplePairing.
When we didn’t use all the images in the training set, we also evaluated the effect of overlaying an
image randomly chosen from an image not in the training set. For example, if we used only 100
images for training, we made a pool of 100 images selected from the other 4,900 unused images
and then randomly picked the image to overlay (i.e., image B in Figure 1) from this pool. The
results are shown in Figure 5 as ”pairing with non-training sample.” This method also reduced error
rates (except for the case of 10 images per class), but our SamplePairing, which picks an image
from the training set, yielded more significant improvements regardless of the training set size. Our
SamplePairing, which does not require images other than the training set, is easier to implement and
also gives larger gains compared to naively overlaying another image.
6
9
8
validation loss
7
Validation loss for ILSVRC
SamplePairing enabled intermittently
(enabled for 300k images and then disabled for 100k images)
6
lower is better
10
5
4
fine tuning
3
2
1
0
# trained images (x100k)
12
10
with SamplePairing
Training loss for ILSVRC
SamplePairing enabled intermittently
(enabled for 300k images and then disabled for 100k images)
training loss
8
lower is better
without SamplePairing
6
4
fine tuning
2
0
# trained images (x100k)
without SamplePairing
with SamplePairing
Figure 4: Changes in validation and training losses for ILSVRC datasets with and without our
SamplePairing data augmentation.
In this paper, we disabled SamplePairing intermittently. Figure 6 shows how the ratio with and without SamplePairing enabled affects the final classification error on CIFAR-10. In the configuration
we used in this paper (disabling SamplePairing in 2 epochs per 10 epochs, i.e. enabled in 80% of
the training epochs), we got slightly better final results compared to the case without disabling SamplePairing. However, the differences were not significant if we enabled SamplePairing in more than
a half of epochs; hence the tuning of the rate to disable SamplePairing is not that sensitive.
5
C ONCLUSION
This paper presents our new data augmentation technique named SamplePairing. SamplePairing is
quite easy to implement; simply mix two randomly picked images before they are fed into a classifier
for training. Surprisingly, this simple technique gives significant improvements in classification
accuracy by avoiding overfitting, especially when the number of samples available for training is
limited. Therefore, our technique is valuable for tasks with a limited number of samples, such as
medical image classification tasks.
In this paper, we only discussed empirical evidences without theoretical explanations or proof. In
future work, we like to provide a theoretical background on why SamplePairing helps generalization
so much and how we can maximize the benefits of data augmentation. Another important future
work is applying SamplePairing to other machine learning tasks especially Generative Adversarial
Networks (Goodfellow et al. (2014)).
7
lower is better
0.8
0.7
0.6
classification error rate
errors reduced
by 28%
0.5
0.4
0.3
0.2
0.1
0
5000
2500
without SamplePairing
500
100
# images per class
with SamplePairing
50
20
10
pairing with non-training sample
Figure 5: Validation error rates for CIFAR-10 with reduced number of samples in training set.
lower is better
8.5%
validation set error rate
8.0%
7.5%
7.0%
6.5%
6.0%
(non-zero
origin)
without
enabled in
SamplePairing
20% of
epochs
enabled in
40% of
epochs
enabled in
50% of
epochs
enabled in
60% of
epochs
enabled in
70% of
epochs
enabled in
80% of
epochs
(default in this
paper)
enabled in
90% of
epochs
enabled in all
epochs
with SamplePairing
Figure 6: Validation error rates for CIFAR-10 with different ratios for intermittently disabling SamplePairing.
R EFERENCES
Anonymous. mixup: Beyond empirical risk minimization. International Conference on Learning
Representations, 2018. URL https://openreview.net/forum?id=r1Ddp1-Rb.
Nitesh V. Chawla, Kevin W. Bowyer, Lawrence O. Hall, and W. Philip Kegelmeyer. Smote: Synthetic minority over-sampling technique. Journal of Artificial Intelligence Research, 16(1):321–
357, 2002.
Marzieh Fadaee, Arianna Bisazza, and Christof Monz. Data augmentation for low-resource neural
machine translation. arXiv:1705.00440, 2017.
Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial networks. arXiv:1406.2661, 2014.
8
Geoffrey Hinton, Nitish Srivastava, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.
Improving neural networks by preventing co-adaptation of feature detectors. arXiv:1207.0580,
2012.
Ferenc Huszár. mixup: Data-dependent data augmentation, 2017. URL http://www.
inference.vc/mixup-data-dependent-data-augmentation/.
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. arXiv:1502.03167, 2015.
Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv:1412.6980,
2014.
Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton. Imagenet classification with deep convolutional neural networks. In Annual Conference on Neural Information Processing Systems (NIPS),
pp. 1106–1114, 2012.
Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In
Conference on Computer Vision and Pattern Recognition (CVPR), 2015.
Seiya Tokui, Kenta Oono, Shohei Hido, and Justin Clayton. Chainer: a next-generation open source
framework for deep learning. In Workshop on Machine Learning Systems (LearningSys), 2015.
Li Wan, Matthew Zeiler, Sixin Zhang, Yann LeCun, and Rob Fergus. Regularization of neural
networks using dropconnect. In International Conference on Machine Learning (ICML), pp. III–
1058–III–1066, 2013.
Jason Wang and Luis Perez. The effectiveness of data augmentation in image classification using
deep learning. Stanford University research report, 2017.
Hongyi Zhang, Moustapha Cissé, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv:1710.09412, 2017. URL http://arxiv.org/abs/1710.
09412.
9
6
A PPENDIX
In our SamplePairing, we take the label from only one sample (label A in Figure 1) and discard
the label from another (label B). Figure 7 shows the validation error with CIFAR-10 if we use both
labels for each synthesized sample, i.e. with 0.5 assigned to each in the softmax target of two
samples. From Figure 7, we can see that difference due to the use of one label or two labels was not
significant. This result is consistent with what Huszár (2017) pointed out on mixup (Zhang et al.
(2017)).
0.2
Validation error for CIFAR-10
lower is better
validation set error rate
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0
50
100 150 200 250 300 350 400 450 500 550 600 650 700 750 800 850 900 950
epoch
without SamplePairing
with SamplePairing with label from only one sample
with SamplePairing with label from two samples
Figure 7: Validation error rates for CIFAR-10 when using lebels from both samples.
10
| 1cs.CV
|
On kernel design for regularized LTI system identification ⋆
Tianshi Chen
arXiv:1612.03542v3 [cs.SY] 31 Jul 2017
School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, 518172, China, (e-mail:
[email protected]).
Abstract
There are two key issues for the kernel-based regularization method: one is how to design a suitable kernel to embed in the
kernel the prior knowledge of the LTI system to be identified, and the other one is how to tune the kernel such that the
resulting regularized impulse response estimator can achieve a good bias-variance tradeoff. In this paper, we focus on the issue
of kernel design. Depending on the type of the prior knowledge, we propose two methods to design kernels: one is from a
machine learning perspective and the other one is from a system theory perspective. We also provide analysis results for both
methods, which not only enhances our understanding for the existing kernels but also directs the design of new kernels.
Key words: System identification, regularization methods, kernel methods, kernel design, prior knowledge.
1
Introduction
Among diverse system identification problems, the
linear time-invariant (LTI) system identification is a
classical and fundamental problem. As well-known, an
LTI system is uniquely characterized by its impulse
response, and thus the LTI system identification is
equivalent to its impulse response estimation that could
be ill-conditioned in practice. So to tackle the LTI system identification, one needs to first find a way to make
the ill-conditioned problem well-conditioned. Then one
faces the core issue of system identification, i.e., how to
construct a model estimator able to achieve a good biasvariance tradeoff, or equivalently, to suitably balance
the adherence to the data and the model complexity.
There are different routes to handle these two issues. The
⋆ Abridged and preliminary results of this paper was presented in the 19th IFAC World Congress, Cape Town, South
Africa, 2014, the European Research Network on System
Identification (ERNSI) workshop, Ostend, Belgium, 2014,
and the 17th IFAC Symposium on System Identification,
Beijing, China, 2015.
This work is supported by the Thousand Youth Talents Plan
funded by the central government of China, the Shenzhen
Projects Ji-20170189 and Ji-20160207 funded by Shenzhen
Science and Technology Innovation Council, the President’s
grant under contract No. PF. 01.000249 and the Start-up
grant under contract No. 2014.0003.23 funded by the Chinese
University of Hong Kong, Shenzhen, as well as by a research
grant for junior researchers under contract No. 2014-5894,
funded by Swedish Research Council.
Preprint submitted to Automatica
most widely used route is to adopt instead a parametric
model structure with a finite number of parameters. The
so-called maximum likelihood/prediction error method
(ML/PEM) is one method along this route. It has optimal asymptotic properties and is the current standard
method to LTI system identification. It first postulates
a parametric model structure, e.g., a rational transfer
function, then forms the prediction error criterion, and
finally obtains the model estimator by minimizing the
prediction error criterion; see e.g., [1,2]. The model complexity of the parametric model structure is governed by
the number of parameters and tuned in a discrete manner. The bias-variance tradeoff issue is handled by cross
validation or model structure selection criteria such as
Akaike’s information criterion (AIC), Bayesian information criterion (BIC), and etc, which correspond to combinatorial optimization problems; see e.g., [1, Chapter
16]. However, these techniques are not as reliable as expected when the data is short and/or has low signal-tonoise ratio; see e.g., [3,4].
Another route is to adopt regularization by adding an extra regularization term in the estimation criterion. The
kernel-based regularization method (KRM) proposed in
the seminal paper [5] and further developed in [6,3,7]
is one method along this route. It first proposes a kernel parameterized by some hyper-parameters for the impulse response, then estimates these hyper-parameters
with certain methods, and finally obtains the estimate
of the impulse response by minimizing a regularized
least squares criterion with the regularization term in
quadratic form and the regularization matrix defined
1 August 2017
tails. This finding leads to the machine learning method
to design general kernels: design the rank-1 kernel and
the stationary kernel to account for the decay and the
varying rate of the impulse response, respectively.
through the kernel; see [4] for a recent survey. While the
impulse response model is used, the underlying model
structure is determined by the kernel. The model complexity is governed by the hyper-parameters of the kernel, and tuned in a continuous manner by e.g., the empirical Bayes method, Stein’s unbiased risk estimator
[4,8,9] and etc., which correspond to non-convex but
non-combinatorial optimization problems.
To develop the system theory method, we recall the general guideline to design a kernel in [3] that is to let the
kernel mimic the behavior of the optimal kernel [3, Thm.
1] and [4, Prop. 19]. Moreover, the prior knowledge for
the impulse response, or equivalently, the LTI system to
be identified, shall be made use of, since the optimal kernel depends on the true impulse response. Following this
guideline and employing the multiplicative uncertainty
configuration in robust control, see e.g. [20], we design
the simulation induced (SI) kernel. In particular, the
prior knowledge is embedded in the nominal model, the
uncertainty is assumed to be stable and finally, the multiplicative uncertainty configuration is used to take into
account both the nominal model and the uncertainty
and is simulated with an impulsive input to get the SI
kernel. Then we make analysis for SI kernels in terms of
stability, maximum entropy property, and Markov property, with the SS and DC kernels as examples. In particular, we give conditions under which the SI kernel is
stable, solves a suitably defined entropy maximization
problem, and induces a Gaussian Markov process. The
maximum entropy interpretation enhances our understanding about a kernel, and the Markov property of a
kernel ensures that the inverse of the kernel matrix is
banded and this special structure can be exploited to
derive more stable and efficient implementation for the
KRM, see e.g., [21,18].
The route to adopt regularization is by no means new,
see e.g., [10], [11], [1, p. 504-505] and also [12] for a historic review, but no important progress along this route
has been reported until [5]. The major obstacle is that
it was unclear whether or not it is possible to design
the regularization to embed the prior knowledge of the
LTI system to be identified. The intriguing finding disclosed by the KRM is that when considering impulse
response estimation problem, it is possible to design a
regularization in quadratic form or equivalently a kernel
to embed the prior knowledge of the impulse response
of the LTI system to be identified. So far several kernels
have been invented to embed various prior knowledge,
e.g., [5,6,3,7,13,14,15,16,17,18,19] and some analysis results of the corresponding kernels have been derived,
e.g., [5,3,7,13,14,15,16,17,18]. Still, there lack systematic
ways to design kernels to embed in the kernel the prior
knowledge of the impulse response of the LTI system to
be identified.
In this paper, we try to develop systematic ways to design kernels. Clearly, this issue relates to the type of the
prior knowledge. Interestingly, even for the same impulse
response estimation problem, users with different background may come up with different types of prior knowledge because they may treat the impulse response in different ways. For instance, users from machine learning
may treat the impulse response as a function, whose amplitude varies with a certain rate and decays exponentially to zero in the end; users from signals and systems
may associate the impulse response with an LTI system
that is stable and may be overdamped, underdamped,
has multiple distinct time constants and resonant frequencies, and etc. Accordingly, we provide two methods to design kernels to embed the corresponding type
of prior knowledge from a machine learning perspective
and from a system theory perspective, respectively.
The remaining part of this paper is organized as follows.
In Section 2, the KRM is recapped. The machine learning and system theory methods to design kernels are introduced in Section 3 and 4, respectively. Then a case
study is provided in Section 5 to demonstrate how to
design kernels from the proposed two methods to model
impulse responses with damped oscillation. Finally, this
paper is concluded in Section 6 with a take-home message. The proofs of all propositions and theorems are
given in the Appendix.
2
To develop the machine learning method, we first point
out a common feature of the two most widely used single
kernels, i.e., the stable spline (SS) kernel in [5] and the
diagonal-correlated (DC) kernel in [3], that they belong
to the class of the so-called amplitude modulated locally
stationary (AMLS) kernel. The AMLS kernel is a multiplication of two kernels: a rank-1 kernel and a stationary
kernel. It has a neat implication: the zero mean Gaussian
process with the AMLS kernel as the covariance function
can be seen as an amplitude modulated (by the rank-1
kernel) zero mean Gaussian processes with the stationary kernel as covariance function; see Section 3 for de-
LTI System Identification with Kernel-based
Regularization Method
2.1
Impulse Response Estimation
Consider a single-input-single-output, causal and LTI
system described by
y(t) = (g ∗ u)(t) + v(t),
t = Ts , 2Ts , · · · , N Ts ,
(1)
where t is the time index, y(t), u(t), v(t) ∈ R and g(t) ∈
R are the measured output, input, disturbance, and impulse response of the LTI system at time t, respectively,
2
inite kernel k(t, s; θ) 1 with t, s ∈ X, where for the DT
case, X = N and for the CT case, X = {t|t ≥ 0}, and
θ ∈ Rm is a hyper-parameter vector that contains the
parameters used to parameterize the kernel, and then
solves the following regularized least squares problem:
(g∗u)(t) is the convolution between the impulse response
g(·) and the input u(·) evaluated at time t, and Ts is the
sampling period and for convenience chosen to be 1.
The system (1) represents both discrete-time (DT) system with t = 1, 2, · · · , and continuous-time system (CT)
with t ≥ 0. In particular, we have
(g ∗ u)(t) =
( P∞
τ =1 g(τ )u(t
R∞
0
− τ ), DT,
g(τ )u(t − τ )dτ, CT,
ĝ(t) = arg min
g∈Hk
(2)
Kernel-based Regularization Method
a(t, s; θ) = (k(t, ·; θ) ∗ u)(s), At,s (θ) = (a(·, s; θ) ∗ u)(t)
As well-known [10], [1, p. 504-505], straightforward impulse response estimation, i.e.,
g
N
X
t=1
(3)
We will discuss the issue of kernels in the next subsections. For the time being, we assume that a suitable kernel k(t, s; θ) has been designed, but its hyper-parameter
θ is left to be determined. The current most widely
used method to determine θ is the so-called empirical
Bayes method [23]. It first embeds the regularization in a
Bayesian framework, and then estimates θ and possibly
also the noise variance σ 2 by maximizing the marginal
likelihood p(Y |η), where Y ∈ RN with y(t) being its tth
element, and η could be θ or the concatenation of θ and
σ 2 . Specifically, we define A(θ) ∈ RN ×N with its (t, s)th
element At,s (θ) defined by
The goal of LTI system identification is to estimate the
impulse response g(t) as well as possible given the data
y(t), u(t) with t = 1, 2, · · · , N for the DT case and y(t)
with t = 1, 2, · · · , N and u(t) with t ≥ 0 for the CT case.
minimize
t=1
(y(t) − (g ∗ u)(t))2 + γkgk2Hk ,
where kgkHk is the norm of g in Hk , and γ > 0 is a regularization parameter and controls the tradeoff between
PN
the data fit t=1 (y(t) − (g ∗ u)(t))2 and the regularization term kgk2Hk .
where the unmeasured portions of u(t) are set to zero.
Moreover, the system (1) is assumed to be stable, i.e., g ∈
ℓ1 for the DT case and g ∈ L1 for the CT case, where ℓ1
denotes the space of absolutely convergent sequences and
L1 denotes the space of absolutely Lebesgue integrable
functions on t ≥ 0, and the disturbance v(t) is assumed
to be a white Gaussian noise with zero mean and variance
σ 2 and independent of u(t).
2.2
N
X
and moreover, we let Σ(η) = A(θ) + σ 2 IN , where IN is
the N -dimensional identity matrix. Then we get
(y(t) − (g ∗ u)(t))2
minimize Y T Σ(η)−1 Y + log det Σ(η),
η∈Γ
where Γ is a constraint set where we search for η. When
an estimate of η is obtained, the solution to (3) is given
by the representer theorem [4, Theorem 3]:
could be an ill-conditioned problem in practice and one
way to overcome this problem is to adopt regularization.
Moreover, the recent progresses for KRM [5,6,3,7] show
that if the regularization is well designed and tuned, the
resulting regularized impulse response estimator can also
achieve a good bias-variance tradeoff.
ĝ(t) =
N
X
ĉs a(t, s; θ),
(4)
s=1
To introduce the KRM, we first recall the definition
of the positive semidefinite kernel and its associated
reproducing kernel Hilbert space (RKHS). Let (X, d)
be a metric space with d being its metric. A function
k : X × X → R is called a positive
Pmsemidefinite kernel,
if it is symmetric and satisfies i,j=1 ai aj k(xi , xj ) ≥
0 for any finite set of points {x1 , · · · , xm } ⊂ X and
{a1 , ..., am } ⊂ R. As well-known from e.g., [22], to every positive semidefinite kernel k there corresponds to
one and only one class of functions with a unique determined inner product in it, leading to the so-called reproducing kernel Hilbert space (RKHS) Hk with k as the
reproducing kernel.
where ĉs is the sth element of ĉ = (A(θ) + σ 2 IN )−1 Y .
2.3
Existing Single Kernels
So far the two mostly widely used single kernels are the
stable spline (SS) kernel [5] and the diagonal/correlated
(DC) kernel [3]. The SS kernel is defined as
1
k SS (t, s; θ) = c exp(−β(t + s) − β max{t, s})
2
1
− c exp(−3β max{t, s}), θ = [c β]T , c, β ≥ 0, (5)
6
1
The KRM first introduces a suitable positive semidef-
3
Sometimes the dependence of k(t, s; θ) on θ is ignored.
in [13, Lemma 8]. More generally, the designed kernel
should guarantee that its associated RKHS Hk is a subspace of ℓ1 or L1 and a kernel that has such property is
said to be stable [13,4]. Sufficient and necessary conditions for a kernel to be stable exist and are given below.
and the DC kernel is defined as
k DC (t, s; θ) = cλ(t+s)/2 ρ|t−s| , θ = [c λ]T
(6)
k TC (t, s; θ) = c min(λt , λs ), θ = [c λ ρ]T
(
|ρ| ≤ 1, DT
c ≥ 0, 0 ≤ λ < 1,
0 ≤ ρ ≤ 1, CT
(7)
Lemma 2.2 ([24,13]) A positive semidefinite kernel k
is stable if and only if
where the TC
√ kernel is a special case of the DC kernel
with ρ = λ and is also called the first order stable
spline kernel.
DT:
CT:
P∞
P∞
s=1 |
R∞ R∞
0
0
t=1
k(t, s)l(t)| < ∞, ∀ l(t) ∈ ℓ∞
k(t, s)l(t)dt ds < ∞, ∀ l(t) ∈ L∞
(10)
Remark 2.1 It is worth to note that for ρ < 0 and
|t − s| 6∈ N, ρ|t−s| is complex and thus k DC (t, s; θ) with
ρ < 0 is not well defined for the CT case.
where L∞ and ℓ∞ denote the space of bounded functions
on R≥0 and bounded sequences, respectively.
2.4
Corollary 2.1 ([13,4]) A positive semidefinite kernel
k is stable if
Optimal Kernel and Stable Kernel
For the KRM, the optimal kernel in the sense of minimizing the mean square error (MSE) exists [3,4] and motivates a general guideline to design a kernel. To state this
result, we assume that the data has been generated by
(1) for a true impulse response g 0 (t) and we let ḡ 0 and
ĝ¯ to represent any finite dimensional vector obtained by
sampling g 0 (t) and its estimate ĝ(t) at the same arbitrary time instants. Then the following result holds.
DT:
CT:
2.5
Lemma 2.1 (Optimal kernel, [3,4]) Letting γ = σ .
Then for the KRM, the MSE matrix
(8)
k opt (t, s) = g 0 (t)g 0 (s),
(9)
where E is the mathematical expectation, is minimized by
the kernel
t, s ∈ X
(11)
Kernel Design
Hereafter, we focus on the problem of kernel design. The
problem is stated as follows: for the given prior knowledge of the impulse response to be identified, our goal is
to design a kernel such that the prior knowledge is embedded in the kernel. The answer should depend on the
type of the prior knowledge. We will consider two types
of prior knowledge and discuss how to design kernels
accordingly from two perspectives: a machine learning
perspective and a system theory perspective.
2
MSE(k) , E (ĝ¯ − ḡ 0 )(ĝ¯ − ḡ 0 )T ,
P∞
t=1 k(t, s)| < ∞,
s=1 |
R∞ R∞
0 k(t, s)dt ds < ∞.
0
P∞
It should be noted that the LTI system (1) is assumed to
be stable, meaning that g ∈ ℓ1 or L1 becomes our first
prior knowledge. In this regard, Theorem 2.2 becomes a
rule that the designed kernel should obey.
in the sense that MSE(k)−MSE(k opt ) is positive semidefinite for any positive semidefinite kernel k.
The optimal kernel k opt cannot be applied in practice
as it depends on the true impulse response g 0 (t). However, it motivates a general guideline to design a kernel:
let the kernel mimic the behavior of the optimal kernel,
and moreover, the prior knowledge of the true impulse
response should be used in the design of the kernel.
Remark 2.2 Other than stability, the cases where the
impulse response or the corresponding LTI system is
known to have relative degree, to be monotonic, smooth
and to have delay, have been discussed in [13] and in particular, sufficient and/or necessary conditions for a kernel to have such properties are given. For instance, if a
CT kernel k is m-times continuously differentiable, then
every function h ∈ Hk is m-times continuously differentiable [25, Corollary 4.36, p. 131], [13, Lemma 5].
For instance, if the LTI system is known to be stable,
i.e., g ∈ ℓ1 or L1 , then the designed kernel k should reflect this, and a necessary condition to satisfy is that
k(t, t) should tend to 0 as t goes to infinity. This observation basically rules out the possibility to model the
impulse response of stable LTI systems with stationary
kernels 2 and a rigorous proof for this has been given
3
A Machine Learning Perspective
We first show that both the SS kernel and the DC kernel
belong to the class of the so-called amplitude modulated
locally stationary (AMLS) kernel. Accordingly, we propose to treat the impulse response as a function, whose
2
Recall that a kernel k(t, s) with t, s ∈ X is said to be
stationary if k(t, s) is a function of t − s for any t, s ∈ X.
4
• The stationary kernel k c (t − s) in (13) accounts for
the varying rate of the function.
Note that part (b) shows that k c (t − s) is the correlation coefficient for g(t), which implies that
amplitude varies with a certain rate and decays exponentially to zero in the end, and moreover, design AMLS
kernels to model the impulse response.
3.1
Amplitude Modulated Locally Stationary Kernels
E(g(t) − g(s))2
= (b(t))2 + (b(s))2 − 2b(t)b(s)k c (t − s).
Recall that a kernel k(t, s) is said to be a locally stationary (LS) kernel in the sense of Silverman [26] if
t+s
d
k(t, s) = k
k c (t − s), t, s ∈ X
(12)
2
The above equation shows that for any fixed t, s ∈ X,
as k c (t − s) varies from 1 to −1, (g(t) − g(s))2 tends to
become larger, that is, g(s) tends to vary more quickly
away from g(t). This observation also holds for h(t) as
k c (t − s) is also the correlation coefficient for h(t).
• The rank-1 kernel k d (t, s) in (13) account for the
change of the amplitude of the function.
Note that part (c) shows that the amplitude of g(t)
is modulated by the factor b(t). If h(t) does not converge to 0 as t goes to ∞, then a suitable b(t) can be
chosen such that g(t) does.
where k d ≥ 0 and k c is a stationary kernel 3 . Motivated
by the LS kernel, we introduce a kernel suitable for modeling impulse response, which is called the amplitude
modulated locally stationary (AMLS) kernel.
Definition 3.1 A kernel k(t, s) is said to be an amplitude modulated locally stationary kernel if
k(t, s) = k d (t, s)k c (t − s), k d (t, s) = b(t)b(s),
k c (0) = 1, t, s ∈ X,
(13)
Interestingly, both the SS and DC kernels can be put in
the form of (13). Setting λ = exp(−β/2) and using the
equality max{t, s} = (t + s + |t − s|)/2 yields that the
SS kernel (5) is rewritten as
where b(t) > 0 is bounded and k c is a stationary kernel.
It can be proved that the AMLS kernels have the following properties (proof given in the Appendix).
k SS (t, s; c, λ) = cλ3(t+s)
Proposition 3.1 Consider the AMLS kernel (13).
Then the following results hold:
1 |t−s| 1 3|t−s|
.
λ
− λ
2
6
(15)
Then we identify, for the SS kernel (15),
c 3(t+s) c
3
1
λ
, k (t − s) = λ|t−s| − λ3|t−s| ,
3
2
2
(16)
and for the DC kernel (6),
(a) k d (t, s) is a rank-1 kernel 4 and moreover satisfy
q
k d (t, s) = k d (t, t)k d (s, s).
(14)
k d (t, s) =
(b) Let g(t) be a stochastic process with zero mean and
the AMLS kernel as the covariance function. Then
k c (t − s) is the correlation coefficient between g(t)
and g(s).
(c) Let h(t) be a stochastic process with zero mean
and k c (t − s) as the covariance function. Then the
stochastic process g(t) , b(t)h(t) has zero mean
and the AMLS kernel as its covariance function.
1
k d (t, s) = cλ 2 (t+s) , k c (t − s) = ρ|t−s| .
(17)
Moreover, one can prove the following result (proof given
in the Appendix).
Proposition 3.2 The SS kernel (15) and the DC kernel
(6) are AMLS kernels.
Now we demonstrate for the SS and DC kernels the role
of the rank-1 kernel k d and the stationary kernel k c .
Remark 3.1 Since k c (t−s) is the correlation coefficient
and k c (0) = 1, then k c (t − s) is bounded and moreover,
|k c (t − s)| ≤ 1 for any t, s ∈ X.
Example 3.1 As seen from the left panel of Fig. 1,
1
1
as λ changes from 0.9 2 to 0.5 2 , the realizations of the
DT zero mean Gaussian process with k c (t − s; [c, λ]T )
in (16) as the covariance function varies more quickly,
1
1
because k c (1; [1, 0.9 2 ]T ) > k c (1; [1, 0.5 2 ]T ). Similarly,
as ρ changes from 0.99, to 0 and to −0.99, the realizations of the DT zero mean Gaussian process with k(t −
s; [1, 0.9, ρ]T ) in (17) varies more quickly, and especially
for the case with ρ = −0.99, the realization tends to
change its sign at the next time instant. The above observations carry over to the right panel of Fig. 1. Finally,
For a function estimation problem, if a zero mean Gaussian processes with the AMLS kernel as the covariance
function is chosen to model the function, then Proposition 3.1 has the following implications:
3
Clearly, if kd is a positive constant, then the LS kernel
reduces to a stationary kernel.
4
A kernel k(t, s) with t, s ∈ X is said to be a rank-1 kernel if
for any ti , si ∈ X, i = 1, . . . , n and for any n ∈ N, the kernel
matrix K, defined by Ki,j = k(ti , tj ), is a rank-1 matrix.
5
3.2.1
0.2
0.04
0
0.02
-0.2
0
0
20
40
60
80
100
0
0.5
20
0
10
-0.5
20
40
60
80
100
0
20
40
60
80
100
0
20
40
60
80
100
0
20
40
60
80
100
×10
An important class of stationary kernels is the isotropic
stationary (IS) kernel. Recall that a stationary kernel
k c (t − s) is said to be an IS kernel if it depends on |t − s|.
IS kernels have been studied extensively in the literature
in statistics and machine learning, see e.g. [27, Section
4.2.1]. There are many choices of IS kernels that could
be used instead of the ones in (16) and (17).
-3
0
0
20
40
60
80
100
0.2
0.1
0.1
0.05
0
Most of the IS kernels introduced in [27, pages 83 -88]
decay monotonically w.r.t. |t−s| and are always positive.
For example, the squared exponential (SE) kernel and
the Matèrn class of kernels:
0
0
20
40
60
80
100
0.5
0.2
0
0
-0.5
-0.2
0
20
40
60
80
100
0.2
0.2
0
0
-0.2
2
k c (r) = e−βr , β > 0, “SE”,
√
ν
21−ν √
k c (r) =
β 2νr K ν β 2νr ,
Γ(ν)
β > 0, ν > 0“Matèrn”,
-0.2
0
20
40
60
80
100
0
20
40
60
80
100
Fig. 1. Realizations of the DT zero mean Gaussian processes
with the SS kernel (15) and the DC kernel (6) as covariance
function. For each row, the left panel shows the realizations
of the DT zero mean Gaussian process with the IS kernel
kc (t − s) as covariance function and the right panel shows
the realization of the DT zero mean Gaussian process with
the SS or DC kernel as covariance function. The corresponding realizations for the SS kernel (15) are shown in the top
1
1
two rows, which correspond to c = 1 and λ = 0.9 2 , 0.5 2 , respectively. The corresponding realizations for the DC kernel
(6) are shown in the bottom three rows, which correspond
to c = 1, λ = 0.9 and ρ = 0.99, 0, −0.99, respectively.
where r = |t − s|, Γ(·) is the Gamma function and K ν is
a modified Bessel function with order ν.
There are IS kernels that do not decay monotonically,
can take negative values, and can have the form of
damped oscillation w.r.t. |t − s|. For example, the kernel
in [27, p. 89], [28]:
k c (r) = c(αr)−ν Jν (αr),
α > 0, ν ≥ −1/2,
(18a)
where r = |t − s|, c is a scalar such that k c (0) = 1, and
Jν (·) is the Bessel function of the first kind with order ν.
The kernel (18a) is defined for any r ≥ 0 and can thus be
used for both CT and DT impulse response estimation.
In particular, for ν = −1/2 and ν = 1/2, the kernel
(18a) takes the following form
the realizations on the left panel of Fig. 1 do not go to
zero for large t but the realizations on the right panel do,
because of k d (t, s).
Remark 3.2 Recall from Remark 2.1 that the DC kernel
(6) with ρ < 0 is only defined for the DT case but not
for the CT case. Now we see that the DC kernel is not
good to model quickly varying impulse responses for the
CT case, which is also true for the SS kernel (15).
k c (r) = cos(αr), α > 0,
sin(αr)
k c (r) =
, α > 0.
αr
3.2.2
3.2
Stationary Kernel k c (t − s)
Construct AMLS Kernels for Regularized Impulse
Response Estimation
(18b)
(18c)
Rank-1 Kernel k d (t, s)
The design of k d (t, s) is equivalent to that of the strictly
positive function b(t) in (13). The bottom line is that
b(t) should ensure that the designed AMLS kernel (13)
is stable, i.e, Hk is a subspace of ℓ1 or L1 . Our following result gives a characterization of the stability of the
AMLS kernel (proof given in the Appendix).
Proposition 3.2 and its implications on the role of the
rank-1 kernel k d and the stationary kernel k c suggest
a machine learning method to design kernels for regularized impulse response estimation. If the impulse response is treated as a function and the prior knowledge
is about its decay and varying rate, then AMLS kernels
can be designed by choosing suitable rank-1 kernel k d
and stationary kernel k c to account for the decay and
varying rate of the impulse response, respectively.
Proposition 3.3 Consider the AMLS kernel (13).
Then the following results hold.
(a) If b(t) ∈ ℓ1 for the DT case, and b(t) ∈ L1 for the CT
case, then the AMLS kernel (13) is stable.
6
(b) Assume that there exists a sequence of positive numbers
λi and linearly independent functions φi (t) defined on
X such that
c
k (t − s) =
∞
X
G0 (q)
G4 (q)
λi φi (t)φi (s),
G(q)
i=1
t, s ∈ X, λi > 0, i = 1, · · · , ∞,
Fig. 2. The block diagram for the multiplicative uncertainty
paradigm in robust control. The single-input-single-output
system G(q) = G0 (q)(1 + G△ (q)) consists of two parts: the
nominal part G0 (q) and the uncertainty part G△ (q), where
q is the forward shift operator and the differential operator
for the DT and CT case, respectively. The real-valued signals
δ(t), ū(t) and g(t) are the impulsive input, the input of G0 (q)
and the output of G(q), respectively.
(19)
where the convergence is absolute and uniform on Y1 ×
Y2 with Y1 , Y2 being any compact subsets of X. If the
AMLS kernel (13) is stable, then there exists no ǫ > 0
such that b(t) ≥ ǫ for all t ∈ X.
Since the b(t) in (16) and (17) are exponential decay
functions and clearly satisfy b(t) ∈ ℓ1 or L1 , the SS and
DC kernels are stable by Proposition 3.3.
accordingly the ECLS kernel [15] instead of the AMLS
kernel. The price to pay is that the role of k d and k c
would become obscure and in particular, k d would not describe the decay rate and k c would not be the correlation
coefficient of the underlying Gaussian process.
Remark 3.3 It is reasonable to check wether the condition (19) is too strong to be satisfied. In fact, the series expansion (19) can be obtained by Mercer’s Theorem
[29,30,31]. For instance, a sufficient condition for (19)
to hold is given in [31]. To state this result, we define the
kernel section of k c at a fixed s ∈ X as ksc , k c (t − s) and
we let L2 (X,
R µ) denote the space of functions f : X → R
such that X |f (t)|2 dµ(t) < ∞, where µ is a nondegenerate Boreal measure on X. Then (19) holds if k c (t − s)
is continuous, ksc ∈ L2 (X, µ) and
Z Z
(k c (t − s))2 dµ(t)dµ(s) < ∞.
(20)
X
g(t)
ū(t)
δ(t)
4
A System Theory Perspective
Instead of simply treating the impulse response as a
function, we now associate the impulse response with an
LTI system which is stable and may be overdamped, underdamped, have multiple distinct time constants and
resonant frequencies, and etc., and our goal is to design
kernels to embed this kind of prior knowledge.
X
4.1
It is easy to check that many stationary kernels satisfy
the above sufficient condition, e.g., the SE kernel and the
k c in (16) and (17).
Sketch of The idea
Suppose that the prior knowledge of an LTI system is
embedded in a stochastic process g(t) that is used to
model the corresponding impulse response. Then following the guideline to design kernels in Section 2.4 that
is to mimic the behavior of the optimal kernel (9), we
should design the kernel as
Remark 3.4 It follows from (A.6) that if there exists an
ǫ > 0 such that |h(t)| ≥ ǫ for all t ≥ 0, then b(t) ∈ L1 ,
implying that b(t) ∈ L1 is also necessary for the stability
of the AMLS kernel (13). The above observations show
that under the assumption that the AMLS kernel (13) is
stable, the result we can draw on the properties of b(t) is
determined by the property of Hkc .
k(t, s) = Cov(g(t), g(s)),
(21)
where Cov(g(t), g(s)) is the covariance between g(t) and
g(s). Now the problem of kernel design becomes “how
to embed the prior knowledge of an LTI system in a
stochastic process g(t) that is used to model the corresponding impulse response”?
Remark 3.5 The reason why we force b(t) > 0
for t ∈ X is because for a kernel in the form of
k(t, s) = k d (t, s)k c (t − s), we expect the two kernels
k d and k c have somewhat independent role. As shown
above, this idea eases the kernel design and the corresponding analysis. If this idea is not taken, then we can
design more general b(t) and even more general k d . For
instance, we could allow b(t) to be arbitrary real-valued
function and we could also allow k d to be the more
general exponentially convex (EC) kernel 5 and design
A natural way to tackle the above question is by using simulation. To this goal, it is useful to employ the
multiplicative uncertainty configuration in robust control, see e.g., [20], as shown in Fig. 2. Here, the nominal
model G0 (q) is used to embed the prior knowledge on
the LTI system to be identified, the uncertainty G△ (q) is
assumed to be stable and finally, the multiplicative uncertainty configuration is used to take into account both
the nominal model and the uncertainty, and is simulated
with an impulsive input to get a zero-mean Gaussian
5
For the LS kernel (12), if kd is also a kernel, then kd is
called an exponentially convex (EC) kernel [32] and in this
case k is called an exponentially convex locally stationary
(ECLS) kernel [26].
7
δ(t)+ l(t), we obtain the simulation induced (SI) kernel:
δ(t)
w(t)
l(t)
b(t)
z(t + 1) = Az(t) + Bδ(t) + Bl(t)
z(t + 1) = Az(t) + Bδ(t) + Bb(t)w(t),
ż(t) = Az(t) + Bδ(t) + Bb(t)w(t),
g(t) = Cz(t) + Dδ(t) + Db(t)w(t),
z(0) ∼ N (0, Q)
g(t)
z(t)
_
= Az(t) + Bδ(t) + Bl(t)
g(t) = Cz(t) + Dδ(t) + Dl(t)
z(0) ∼ N (0; Q)
Fig. 3. The block diagram for the SI kernel (24).
k SI (t, s) = Cov(g(t), g(s)),
process to model the impulse response g(t). This idea
leads to the system theory method to design kernels and
we call such kernels the simulation induced (SI) kernels.
k SI (t, s) = CAt Q(As )T C T + D2 b(t)b(s)δ(t − s)
s−1
X
δ(t − k)b(k)B T (As−1−k )T C T
+ Db(t)
+ Db(s)
G (q) :
t ∈ X,
b(k)2 CAt−1−k BB T (As−1−k )T C T
(25a)
For the CT case with D = 0, it is
k SI (t, s) = CeAt Q(eAs )T C T +
Z min{t,s}
C
b2 (τ )eA(t−τ ) BB T (eA(s−τ ) )T dτ C T
(25b)
0
Remark 4.2 It should be noted that both DT and CT
cases are considered in (22) and (24). For the CT case,
(24b) could be more rigorously written as an Itô stochastic differential equation (SDE) as in [14, eq. (9)]. However, we decide to take the same point of view as in [40] to
use (24b) instead in order to save the space for the introduction of SDE stuff, which is only used to derive (25b).
(22a)
(22b)
(22c)
(22d)
Remark 4.3 The SI kernels (25) may or may not have
closed form expressions. The related computational difficulty and cost depend on whether or not the SI kernels
(25) have closed form expressions. If they do, then the
computation would be easier and similar to the SS and
DC kernels. If they do not, the computation of the hyperparameter estimate and the regularized impulse response
estimate would become more demanding. In this regard,
the particle filtering based technique for nonlinear statespace model identification in [41] could be adopted. However, the technique in [41] cannot be applied trivially as
the measurement output (1) for the state space model (24)
does not depend on the current state z(t) solely but also
the past state due to the presence of convolution. More
details will be reported in an independent paper.
Since the uncertainty G△ (q) is stable, the simplest way
to model the impulse response of G△ (q) with a stochastic
process l(t), is to have
l(t) = b(t)w(t),
δ(s − k)b(k)B T (At−1−k )T C T
k=0
where A, B, C, D, and Q have compatible dimensions,
ū(t) and g(t) are the input and output of G0 (q), respectively. In what follows, we will use the quintuple
(A, B, C, D, Q) to represent the state-space model (22).
G△ (q) :
X
+
More specifically, the prior knowledge is embedded in
G0 (q) or equivalently, in its state-space model
z(t + 1) = Az(t) + B ū(t)
ż(t) = Az(t) + B ū(t)
z(0) ∼ N (0, Q)
g(t) = Cz(t) + Dū(t)
k=0
t−1
X
k=0
min{t,s}−1
Simulation Induced Kernel
0
(24e)
whose block diagram is shown in Fig. 3. From linear
system theory, e.g., [39], the formal expression of the SI
kernel (24) is available. For the DT case, it is
Remark 4.1 It is worth to note that the problem of marrying system identification and robust control is not new
and has been studied over two decades in the system identification community, see e.g., [33,34]. Both the additive uncertainty configuration and the multiplicative uncertainty configuration have been used and many results
have been obtained, see e.g., [35,33,34,36,37,38]. In particular, [35,33,36,37,38] used the nominal model G0 (q)
and the uncertainty G△ (q) to represent a model estimate
and the corresponding model error, respectively. Here the
nominal model G0 (q) and the uncertainty G△ (q) are used
to embed the prior knowledge of the LTI system and describe the corresponding uncertainty, respectively.
4.2
(24a)
(24b)
(24c)
(24d)
(23)
where w(t) is a white Gaussian noise with zero mean and
unit variance (independent of z(0)), and b(t) > 0 and
b(t) ∈ ℓ1 or L1 for the DT and CT case, respectively.
Clearly, l(t) is a zero mean Gaussian process with the
AMLS kernel b(t)b(s)δ(t − s) as the covariance function,
where δ(·) is the Dirac Delta function and the AMLS
kernel b(t)b(s)δ(t − s) is stable by Proposition 3.3.
Note that g(t) in (24) is a Gaussian process with the
impulse response of the LTI system (A, B, C, D, Q) as
its mean and the SI kernel k SI (t, s) as its covariance
Finally, we simulate the system G(q) in Fig. 2 with the
impulsive input δ(t) and noting (22), (23) and ū(t) =
8
The DC kernel is in the form of (24) with
function. If δ(t) is set to 0 in (24), g(t) in (24) is a zero
mean Gaussian process with the SI kernel k SI (t, s) as its
covariance function. In what follows, when considering
kernel design, we will set δ(t) to zero for convenience.
1
1
c
A = −α − β, B = 1, C = (2α) 2 , Q =
,
4
2α
β
1
D = 0, b(t) = c 2 e− 4 t .
(30)
Interestingly, the AMLS kernel (13) is closely related
with the SI kernel (24), and in fact many AMLS kernels
such as the SS and DC kernels can be put in the form of
(24) and are thus SI kernels. To state this result, recall
that the power spectral density of k c (t − s), denoted by
Ψ(ω), is defined as
Ψ(ω) =
( P+∞
c
−iωτ
,
τ =−∞ k (τ )e
c
−iωτ
k (τ )e
dτ,
−∞
R +∞
DT,
CT.
Proposition 4.2 enhances our understanding about the
SS and DC kernels and in particular, the prior knowledge embedded in the two kernels from a system theory
perspective is as follows:
1) For the SS kernel, the corresponding nominal model
G0 (q) in Fig. 2 is a second order system. For the DC
kernel, the corresponding nominal model G0 (q) is a
first order system. For both the SS and DC kernels
and for the CT case, G0 (q) has negative real poles
corresponding to overdamped impulse responses. For
the DT case, the two poles for the SS kernel are real
positive and correspond to impulse response without
oscillation, and the pole for the DC kernel can be positive or negative (depending on ρ) and corresponds to
impulse response without or with oscillation. The pole
for the TC kernel is positive (equal to λ) and corresponds to impulse response without oscillation.
2) For both the SS and DC kernels, the decay rate of the
impulse response of the uncertainty G△ (q) in Fig. 2
are all described by exponential decay functions.
(26)
Proposition 4.1 Consider the AMLS kernel (13). Assume that k d (t, s) = cλt+s with c ≥ 0 and 0 ≤ λ < 1and
Ψ(ω) is a proper rational function of eiω or cos(ω) for
the DT case, and ω for the CT case, respectively. Then
the AMLS kernel (13) can be put in the form of (24) and
thus is SI kernel.
Proposition 4.2 Consider the SS kernel (15) and the
DC kernel (6). Then the following results hold:
• For the DT case, let
q
p
1 + λ2 − 1 + λ2 + λ4 ,
q
p
b̄ = 1 + λ2 + 1 + λ2 + λ4 .
ā =
Remark 4.4 It is worth to note that to study the link
between a kernel and its state-space model realization has
a long history, see e.g., [42,40]. This link is important
because it opens the body of a kernel from a system theory perspective and accordingly helps to understand the
underlying behavior of a kernel.
Then the SS kernel is in the form of (24) with
A = λ3
"
λ 0
0 λ3
#
,B =
"
λ3
λ3
#
c
,Q =
3
"
1
1
1−λ2 1−λ4
1
1
1−λ4 1−λ6
#
4.3
r
(1 − λ2 )3 h ā+b̄λ −λ2 (ā+b̄λ3 ) i
C=
1−λ2
1−λ2
r 2
2
3
(1 − λ )
c 12 3t
λ .
, b(t) =
D = b̄
2
3
For the SI kernel (25), we provide the following sufficient
condition to guarantee its stability.
(27)
Proposition 4.3 Consider the SI kernel (25).
The DC kernel is in the form of (24) with
• For the DT case, assume that A has distinct eigenvalues and moreover, A is stable, i.e., A has all eigenvalues strictly inside the unit circle. Assume further that
c
A = λ ρ, B = λ , C = ρ(1 − ρ2 ) , Q =
,
1 − ρ2
1
1
t
D = (1 − ρ2 ) 2 , b(t) = c 2 λ 2 .
(28)
1
2
1
2
1
2
t
b(t) ≤ c̄|λ̄| 2
• For the CT case, let λ = exp(−β/2) and ρ = exp(−α).
Then the SS kernel is in the form of (24) with
− 23 β
1
#
(31a)
for some c̄ > 0, where λ̄ is the eigenvalue of A with the
largest modulus. Then the SI kernel (25a) is stable.
• For the CT case, assume that A has distinct eigenvalues and moreover A is stable, i.e., A has all eigenvalues
with strictly negative real parts. Assume further that
"
#
c 3β1 3 0
,B =
,Q =
A=
1
3
− 34 β 2 − 72 β
0 4β
1
i
c 12
h 1
3
3
(29)
C = 3 2 β 2 0 , D = 0, b(t) =
e− 2 βt .
3
"
Stability of SI kernels
" #
0
1
b(t) ≤ c̄e− 2 |Re(λ̄)|t
9
(31b)
where Cov(z̄(0)) is the covariance matrix of z(0) and
V(f (t)) is the variance of f (t).
• For the case D = 0, assume that CB 6= 0 and define
for some c̄ > 0, where λ̄ is the eigenvalue of A with the
largest real part and Re(λ̄) is the real part of λ̄. Then
the SI kernel (25b) is stable.
Pt−2
ḡ(t) − CAt z̄(0) − k=0 CAt−1−k Bb(k)f (k)
f (t − 1) =
CBb(t − 1)
ḡ(t) ∈ R, z̄(0) ∈ Rn , t = 0, · · · , s − 1
(34)
Remark 4.5 The conditions (31) are sufficient but not
necessary. For instance, consider the DT case. The SS
kernel (27) satisfies (31a) but the DC kernel (28) does
not. It is possible to derive a slower upper bound for b(t).
Another intension to have Proposition 4.3 is to give some
guidelines about the relations between b(t), A and the
stability of SI kernel: in order to guarantee stability of the
SI kernel, the impulse response of the uncertainty G△ (q)
(23), i.e., b(t) should decay a bit faster than the slowest
mode of the nominal model G0 (q).
4.4
Then for any s ∈ N, the Gaussian process g(t) in (24)
is the solution to the MaxEnt problem
z̄(0),ḡ(t)
Since the prior knowledge is never complete, it is worth
to note Jaynes’s maximum entropy (MaxEnt) rationale
[43] to derive from incomplete prior knowledge the optimal statistical prior distribution. Jaynes’s idea is to formulate a MaxEnt problem with respect to the prior, and
then solve the problem to obtain the optimal prior. The
constraints of the problem describes the prior knowledge (in the MaxEnt sense) of the underlying stochastic
process (the system) to be identified. Interestingly, the
SI kernel lends itself easily to a MaxEnt interpretation,
leading to a new facet to understand the underlying behavior of the SI kernel.
Corollary 4.1 For any s ∈ N, the zero mean Gaussian
process g(t) with the DC kernel (6) defined on N × N as
its covariance is the solution to the MaxEnt problem
ḡ(t)
subject to
V(ḡ(t) − λ
maximize
ḡ(t)
CAt−1−k Bb(k)f (k)
Db(t)
ḡ(t) ∈ R, z̄(0) ∈ Rn , t = 0, 1, · · · , s
(32)
k=0
Remark 4.7 By using Corollary 4.1 and the trick in
[17, Theorem 2], it is possible to derive a more concise
proof for [18, Proposition IV.1] which shows that, the fact
that the DC kernel has tridiagonal inverse can be given
a MaxEnt covariance completion interpretation.
Then for any s ∈ N, the Gaussian process g(t) in (24)
is the solution to the MaxEnt problem
subject to
E(z̄(0)) = 0, E(ḡ(t)) = 0,
Cov(z̄(0)) = Q, V(f (t)) = 1, t = 0, · · · , s
H(ḡ(0), ḡ(1), · · · , ḡ(s))
subject to E(ḡ(t)) = 0, t = 0, · · · , s, V(ḡ(0)) = c,
V(ḡ(t) − λḡ(t − 1)) = c(1 − λ)λt , t = 1, · · · , s
(37)
which is different from the MaxEnt interpretation in [17].
It shall be noted that both interpretation are correct but
derived in different ways: the MaxEnt problems are different but have the same optimal solution.
Pt−1
H(z̄(0), ḡ(0), ḡ(1), · · · , ḡ(s))
ρḡ(t − 1)) = c(1 − ρ2 )λt , t = 1, · · · , s
(36)
Remark 4.6 When ρ = λ1/2 , the DC kernel (6) becomes the TC kernel (7). For the TC kernel defined on
N × N, (36) becomes
• For the case D 6= 0, define
z̄(0),ḡ(t)
E(ḡ(t)) = 0, t = 0, · · · , s, V(ḡ(0)) = c,
1/2
Proposition 4.4 Consider the SI kernel (24) with
(25a).
maximize
H(ḡ(0), ḡ(1), · · · , ḡ(s))
maximize
Recall that the differential entropy H(X) of a real-valued
continuous
random variable X is defined as H(X) =
R
− S p(x) log p(x)dx, where p(x) is the probability density function of X and S is the support set of X. Then
we prove the next MaxEnt interpretation of SI kernels 6 .
ḡ(t) − CAt z̄(0) −
(35a)
subject to
E(z̄(0)) = 0, E(ḡ(t)) = 0,
Cov(z̄(0)) = Q, V(f (t)) = 1, t = 0, · · · , s − 1
(35b)
Maximum Entropy Property of SI Kernels
f (t) =
H(z̄(0), ḡ(1), · · · , ḡ(s))
maximize
(33a)
Remark 4.8 The CT case is not discussed here because
according to our best knowledge the entropy is not well
defined for CT stochastic processes.
(33b)
4.5
Markov Property of SI Kernels
6
Let X1 , X2 be two jointly distributed random variables.
Then the joint differential entropy of H([X1 , X2 ]T ) is simply
written as H(X1 , X2 ) below.
As shown in [44,19,18,17], the kernel matrix of DC kernel
(6) has tridiagonal inverse. Here, we further show that
10
the Gaussian process with the DC kernel as its covariance function is also a Markov process with order 1 and
moreover, we are able to design SI kernels which correspond to more general Gaussian Markov processes and
have banded 7 inverses of their kernel matrices.
2
100
4
6
50
First, recall from e.g., [27, Appendix B] that a Gaussian
Markov process is a stochastic process that is both a
Gaussian process and a Markov process. A well-known
instance is the DT autoregressive process of order p:
x(t + 1) =
p−1
X
10
p−1
X
k=0
a(t, k)x(t − k) + b(t)w(t)
1
1
t+1
2
w(t + 1)
G0 (q) =
10
b̄q
,
(q + a1 )(q + a2 )
(41)
Remark 4.10 The Markov property of a kernel and its
associated special structure (the banded inverse of the
kernel matrix) can be used to develop numerically more
stable and efficient implementations for this kernel based
regularization method, see e.g., [18, Section 5].
(39)
Remark 4.11 The CT case is not discussed here because according to [27, p. 217], regular sampling of a
CT Gaussian Markov process entropy in general would
not lead to a DT Gaussian Markov process. That is to
say, even if a CT SI kernel with Markov property is constructed, its corresponding kernel matrix evaluated on the
sampling instants may not have banded matrix inverse.
Remark 4.9 Interestingly, (39) can also be derived from
(36), i.e., from the MaxEnt property of the DC kernel.
It is possible to construct more general SI kernels with
Markov property.
Proposition 4.6 Consider the DT SI kernel (24) with
(A, B, C, D) being a realization of G0 (q) which is an nth
order DT system
q n−1 b̄
,
q n + a1 q n−1 + · · · + an
8
where b̄ ∈ R and |ai | < 1, i = 1, 2. We consider the
inverse of the kernel matrix K defined by Ki,j = k SI (i, j)
with i, j = 1, . . . , 10. By Proposition 4.6, K −1 should be
a 2-band matrix, which is confirmed by Fig. 4.
and thus a Markov process with order 1, and moreover,
its kernel matrix has a 1-band matrix inverse.
G0 (q) =
6
Example 4.1 Consider the DT SI kernel (24) with
(A, B, C, D) being a realization of G0 (q), which is a 2nd
order DT system having two stable real poles, i.e.,
(38b)
Proposition 4.5 Consider a zero mean DT Gaussian
process g(t) with the DC kernel (6) as its covariance.
Then g(t) with t ∈ N can be put in the form of
1
4
Fig. 4. Scaled image of K −1 in Example 4.1. When generating
K, we choose b̄ = 1, a1 = 0.5, a2 = 0.9, b(t) = 0.8t , and
Q = I2 . The image is drawn by using imagesc in MATLAB,
where the colder the color the smaller the element of K −1 .
where t ∈ N, x(t), a(t, k), b(t) ∈ R, x(0) is Gaussian
distributed and assumed to be independent of w(t), a
zero mean white Gaussian noise with unit variance. The
stochastic process (38) is a Markov process with order p
since x(t + p + 1) only depends on x(t + p), · · · , x(t + 1)
given the history x(s) with s ≤ t + p.
g(t + 1) = λ 2 ρg(t) + (1 − ρ2 ) 2 c 2 λ
0
2
a(t, k)x(t − k) + b(t + 1)w(t + 1) (38a)
k=0
or, x(t + 1) =
8
5
A Case Study: Impulse Response with
Damped Oscillation
In this section, we consider the estimation of impulse
responses with damped oscillation and demonstrate how
to design kernels from the proposed two perspectives.
(40)
where b̄, a1 , . . . , an ∈ R. Then the Gaussian process g(t)
in (24) with any b(t) > 0 and positive semidefinite Q is
also a Markov process with order n and thus the SI kernel
has an n-band matrix inverse.
5.1
The Machine Learning Perspective
As shown in Section 3, the machine learning perspective
treats the impulse response as a function, and designs
AMLS kernels with the rank-1 kernel and the stationary
kernel to account for the decay and varying rate of the
impulse response, respectively. Now we show that by
further exploiting the rank-1 kernel or the stationary
kernel, we can design AMLS kernels capable to embed
For illustration, we consider an example.
7
A real symmetric matrix A with dimension n > m + 1 is
called an m−band matrix if Ai,j = 0 for |i − j| > m.
11
5.2
5
The System Theory Perspective
2
0
0
-5
-2
0
20
40
60
80
100
5
0
20
40
60
80
100
0
20
40
60
80
100
0
20
40
60
80
100
As shown in Section 4, the system theory perspective
associates the impulse response with an LTI system, and
designs SI kernels with the nominal model to embed the
prior knowledge on the LTI system. Now we show that
by choosing the nominal model to be a second order
LTI system with a pair of complex conjugate poles, we
design a SI kernel capable to model impulse responses
with damped oscillation.
1
0
0
-5
-1
0
20
40
60
80
100
5
2
0
0
-5
-2
0
20
40
60
80
100
2
2
0
0
-2
More specifically, we choose the transfer function of the
nominal model G0 (q) in Fig. 2 to be
-2
0
20
40
60
80
100
0
2
2
0
0
-2
20
40
60
80
100
G0 (s) =
-2
0
20
40
60
80
100
0
20
40
60
80
A=
"
(42)
k
t+s
(t, s) = cλ
5.3.1
(43a)
1
h
i
, C = 1 0 , D = 0.
Numerical Simulation
Test Data-bank
The way in [45] is used to generate the test systems
and data bank. In particular, we first generate 200 test
systems with strong oscillation:
[cos(ωt) + 1 + ǫ]
× [cos(ωs) + 1 + ǫ]ρ|t−s| ,
,B =
" #
0
To illustrate that the AMLS kernls (43) and the SI kernel (45) are capable to model LTI systems with strong
oscillation, we consider the following numerical example.
where c > 0, 0 ≤ λ < 1, ω ≥ 0, and ǫ > 0 is a tiny
number to ensure that b(t) > 0 for t ∈ X. The above
idea leads to the following two kernels, respectively:
AMLS-2Od
−α2 − β 2 −2α
#
e2(α−γ) min{t,s} cos(2β min{t, s} − φ − β(t + s))],
2(α − γ)
.
(45)
φ = arccos
4β 2 + 4(α − γ)2
5.3
k AMLS-2Os (t, s) = cλt+s cos(α|t − s|),
1
k SI-2Od (t, s) = e−α(t+s) [cos(βt) cos(βs)
α
α2 + 1
+ sin(β(t + s)) +
sin(βt) sin(βs)]
β
β2
e−α(t+s) cos(β(t − s))(e2(α−γ) min{t,s} − 1)
+
4β 2 (α − γ)
−α(t+s)
e
p
× [cos(φ + β(t + s))−
+
2
2β 4β 2 + 4(α − γ)2
To have oscillation behavior in the AMLS kernel, we can
choose to have it either in the stationary kernel or in the
rank-1 kernel, i.e., b(t). For example, we can choose b(t)
as an exponential decay function, i.e., b(t) = cλt and
then choose (18b) as the stationary kernel, because (18b)
has oscillation behavior. Or we can choose the stationary
kernel in (17) and b(t) as an exponential decay function
with oscillation, i.e.,
b(t) = c λ [cos(ωt) + 1 + ǫ],
0
Finally, setting Q = I2 and b(t) = e−γt with γ > 0 yields
that the SI kernel (25b) takes the form:
the extra prior knowledge that the impulse response has
damped oscillation.
t
(44)
where ωp
0 > 0 and 0 ≤ ξ < 1. By setting α = w0 ξ and
β = w0 1 − ξ 2 , a state space model for (44) is described
by (A, B, C, D) with
100
Fig. 5. Realizations of DT zero mean Gaussian process with
AMLS kernels (43). For each row, the left panel shows the
realizations of the DT zero mean Gaussian process with the
IS kernel kc (t − s) as covariance function and the right panel
shows the realization of the DT zero mean Gaussian process with the AMLS kernel as covariance function. The top
three rows correspond to the AMLS kernel (43b) with c = 1,
1
λ = 0.9 2 , ω = 0.2π and ρ = 0.99, 0, −0.99, respectively. The
bottom two rows correspond to the AMLS kernel (43a) with
1
c = 1, λ = 0.9 2 and α = π, 0.2π, respectively.
1
2
1
s2 + 2w0 ξs + w02
(43b)
G(q) =
Fig. 5 illustrates that the AMLS kernels (43) are capable
to describe functions with damped oscillation behavior.
12
Nr +1
q + 0.99 X
Gi (q)
q
i=1
(46)
where
100
90
q + 0.9
Gi (q) = Ki
, i = 1, . . . , Nr ,
(q − pi )(q − p∗i )
80
70
60
and GNr +1 (q) is a 4th order system randomly generated
by the function drmodel in MATLAB with its poles inside the disk of radius 0.95. The parameters Nr , pi , Ki
are randomly generated as follows: Nr ∼ U[3, 8], Ki ∼
π
U[2, 10], pi = ρi ej[φ0 + 2Nr (i−1)] with φ0 ∼ U[0, π2 ] and
ρi ∼ U[0.8, 0.99].
50
40
30
20
10
0
1
For each of the 200 DT systems (46), we generated the
test data as follows. The function idinput in MATLAB
is used to generate a random Gaussian input u(t) with
normalized band [0, 0.95] and length 210. The chosen
DT system is simulated with u(t) to get the noise-free
output G(q)u(t), and G(q)u(t) is then perturbed by a
white Gaussian noise, whose standard deviation is equal
to U[0.5, 1] times the standard deviation of G(q)u(t).
The following model fit is used to measure how good the
estimated impulse response is:
fit = 100 1 −
"P
100
Pk=1
100
k=1
# 12
100
X
, ḡ 0 = 1
gk0
100
|gk0 − ḡ 0 |2
|gk0 − ĝk |2
The AMLS kernels (43) and the SI kernel (45) are compared with the TC kernel (7), the DC kernel (6) and the
SS kernel (15) enriched with a second order parametric
part proposed in [6] and denoted by SSp below.
6
The simulation result is summarized below. The average
model fits for the tested kernels are shown below:
TC
DC
SSp
AMLS-2Os
AMLS-2Od
SI-2Od
47.5
50.0
50.9
48.4
49.7
53.3
6
Conclusion
Kernel methods or regularization methods to estimate
the impulse response of a linear time invariant systems is a most useful relatively recent addition to the
techniques of system identification. Earlier results have
demonstrated that important improvements in estimation quality can be achieved with kernels devised in a
more or less ad hoc way. In this contribution, we have
focused on systematic mechanisms and design concepts
for how to construct kernels that are capable of adjusting
its hyperparameters to capture the unknown system’s
properties in useful ways. The two main avenues to
such thinking have been a machine learning perspective
focusing on function properties of the impulse response
and a system theory perspective focusing on the LTI system that produces the impulse response. This has lead
The distribution of the model fits are shown in Fig. 6.
5.3.3
5
Remark 5.1 To embed the prior knowledge such as multiple distinct time constants and multiple resonant frequencies, we proposed to use the multiple kernel introduced in [7] where the multiple kernel is a conic combination of some fixed kernels. The fixed kernels can be instances of the SS, TC, DC and SI kernels (24) evaluated
on a grid in their hyper-parameter space. The interested
readers are referred to [7] for more details.
k=1
where gk0 and ĝk are the true impulse response and its
estimate at the kth time instant, respectively.
Avg. Fit
4
better than the TC kernel in both the average accuracy
and robustness and it is just a bit worse than the DC
kernel in the average accuracy. The AMLS kernel (43a)
works not very well, though it has better average accuracy than the TC kernel. One possible explanation why
the SI kernel (45) is best to model impulse response with
strong oscillation is that the prior knowledge embedded
in the SI kernel (45) is closest to the truth. This may be
true because it makes use of the prior knowledge from a
system theory perspective directly by choosing the nominal model to be a second order system with a pair of
complex conjugate poles.
Simulation Results
3
Fig. 6. Boxplot of the model fits for the TC, DC, SSp,
AMLS-2Os, AMLS-2Od and SI-2Od kernels [from left to
right]. The six kernels have 5, 3, 3, 3, 4, and 3 fits smaller
than zero, respectively, which are not shown in the boxplot
for better display.
In this way, we generate 200 test DT systems and for
each system a data set with 210 data points.
5.3.2
2
Findings
For the test systems and data bank, we have the following findings. The SI kernel (45) is best among all
the tested kernels in both the average accuracy and robustness and thus is best to model systems with strong
oscillation. The AMLS kernel (43b) works well as it is
13
References
to understanding and insights into general properties
of the earlier suggested methods, e.g. that the so called
DC and SS kernels (derived from more ad hoc thinking)
belong to the class of amplitude modulated locally stationary kernels (Proposition 3.2) and that they are simulation induced from certain LTI systems (Proposition
4.2). They are also related to maximum entropy optimal
choices (Proposition 4.4) which is a valuable feature.
[1] L. Ljung. System Identification - Theory for the User.
Prentice-Hall, Upper Saddle River, N.J., 2nd edition, 1999.
[2] T. Söderström and P. Stoica. System Identification. PrenticeHall Int., London, 1989.
[3] T. Chen, H. Ohlsson, and L. Ljung. On the estimation of
transfer functions, regularizations and Gaussian processes Revisited. Automatica, 48:1525–1535, 2012.
[4] G. Pillonetto, F. Dinuzzo, T. Chen, G. De Nicolao, and
L. Ljung. Kernel methods in system identification, machine
learning and function estimation: A survey. Automatica,
50(3):657–682, 2014.
The take-home message of this contribution is as follows.
The issue of kernel design should relate to the type of the
prior knowledge and different types of prior knowledge
should lead to different ways to design the corresponding kernels. Here, a machine learning perspective and a
system theory perspective are introduced accordingly:
[5] G. Pillonetto and G. De Nicolao. A new kernel-based
approach for linear system identification.
Automatica,
46(1):81–93, 2010.
[6] G. Pillonetto, A. Chiuso, and G. De Nicolao. Prediction error
identification of linear systems: a nonparametric Gaussian
regression approach. Automatica, 47(2):291–305, 2011.
• Machine Learning perspective: If the impulse response
is treated as a function and the prior knowledge is on
its decay and varying rate, then we can design the amplitude modulated locally stationary (AMLS) kernel.
In particular, we design a rank-1 kernel and a stationary kernel to account for the decay and varying rate of
the impulse response, respectively. Moreover, by further exploiting the rank-1 kernel or the stationary kernel, it is possible to design AMLS kernels capable to
embed more general prior knowledge.
• System Theory perspective: If the impulse is associated with an LTI system and the prior knowledge is
that the LTI system is stable and may be overdamped,
underdamped, have multiple distinct time constants
and resonant frequencies and etc., then we can design
the simulation induced (SI) kernel. In particular, the
nominal model is used to embed the prior knowledge,
the uncertainty is assumed to be stable and finally,
the multiplicative uncertainty configuration is used to
take into account both the nominal model and the uncertainty, and is simulated with an impulsive input to
get the SI kernel or equivalently the zero-mean Gaussian process to model the impulse response.
[7] T. Chen, M. S. Andersen, L. Ljung, A. Chiuso,
and G. Pillonetto.
System identification via sparse
multiple kernel-based regularization using sequential convex
optimization techniques. IEEE Transactions on Automatic
Control, (11):2933–2945, 2014.
[8] Gianluigi Pillonetto and Alessandro Chiuso.
Tuning
complexity in regularized kernel-based regression and linear
system identification: The robustness of the marginal
likelihood estimator. Automatica, 51:106 – 117, 2015.
[9] B. Mu, T. Chen, and L. Ljung. Tuning of hyperparameters
for fir models – an asymptotic theory. In Proceedings of the
IFAC 2017 World Congress, page under review, Toulouse,
France., 2017.
[10] A. N. Tikhonov and V. Y. Arsenin. Solutions of Ill-posed
Problems. Winston/Wiley, Washington, D.C., 1977.
[11] J Sjöberg, T McKelvey, and L Ljung. On the use of
regularization in system identification. In Proceedings of
the IFAC 2012 World Congress, pages 381–386, Sydney,
Australia., 1993.
[12] A. Chiuso.
Regularization and Bayesian learning in
dynamical systems: Past, present and future. Annual Reviews
in Control, 41:24 – 38, 2016.
[13] F. Dinuzzo.
Kernels for linear time invariant system
identification. SIAM Journal on Control and Optimization,
53(5):3299–3317, 2015.
Finally, finding a suitable kernel structure is only one
leg of the kernel-based regularization method. Tuning
its hyperparameters regardless of structure is the other
main topic. This has been discussed in some detail e.g.
in [4] and [8]. Some related new asymptotic results are
recently presented in [9].
[14] Tianshi Chen and Lennart Ljung. Constructive state-space
model induced kernels for regularized system identification.
In 19th IFAC World Congress, pages 1047–1052, Cape town,
South Africa, 2014.
[15] T. Chen and L. Ljung. On kernel structure for regularized
system identification (i): a machine learning perspective. In
Proceedings of the IFAC Symposium on System Identification,
pages 1035–1040, Beijing, China., 2015.
[16] T. Chen and L. Ljung. On kernel structure for regularized
system identification (ii): a system theory perspective. In
Proceedings of the IFAC Symposium on System Identification,
pages 1041–1046, Beijing, China., 2015.
Acknowledgements
[17] Tianshi Chen, Tohid Ardeshiri, Francesca P. Carli,
Alessandro Chiuso, Lennart Ljung, and Gianluigi Pillonetto.
Maximum entropy properties of discrete-time first-order
stable spline kernel. Automatica, 66:34 – 38, 2016.
The authors would like to thank Prof. Lennart Ljung for
his suggestions and comments at the early stage of this
project. The authors also would like to thank Prof. Gianluigi Pillonetto, Prof. Alessandro Chiuso, Prof. Håkan
Hjalmarsson, Dr. John Latarie and Dr. Giulio Bottegal
for their helpful comments.
[18] F. P. Carli, T. Chen, and L. Ljung. Maximum entropy kernels
for system identification. IEEE Transactions on Automatic
Control, 2017.
14
[19] A. Marconato, M. Schoukens, and J. Schoukens. Filter-based
regularisation for impulse response modelling. IET Control
Theory & Applications, to appear 2016.
[41] T. B. Schön, A. Wills, and B. Ninness. System identification
of nonlinear state-space models. Automatica, 47(1):39–49,
January 2011.
[20] K. Zhou, J. C. Doyle, and K. Glover. Robust and Optimal
Control. Prentice-Hall, Englewood Cliffs, NJ, 1996.
[42] K. J. Åström. Introduction to stochastic control theory.
Academic Press, New York and London, 1970.
[21] T. Chen and L. Ljung. Implementation of algorithms for
tuning parameters in regularized least squares problems in
system identification. Automatica, 49:2213–2220, 2013.
[43] E. T Jaynes. On the rationale of maximum-entropy methods.
Proceedings of the IEEE, 70(9):939–952, 1982.
[44] F. P. Carli. On the maximum entropy property of the firstorder stable spline kernel and its implications. In IEEE
Multi-Conference on Systems and Control, pages 409–414,
Nice/Antibes, France, 2014.
[22] Nachman Aronszajn.
Theory of reproducing kernels.
Transactions of the American mathematical society, pages
337–404, 1950.
[23] Gianluigi Pillonetto and Alessandro Chiuso.
Tuning
complexity in regularized kernel-based regression and linear
system identification: The robustness of the marginal
likelihood estimator. Automatica, 58:106–117, 2015.
[45] A. Chiuso, T. Chen, L. Ljung, and G. Pillonetto. On the
design of multiple kernels for nonparametric linear system
identification. In Proceedings of the IEEE Conference on
Decision and Control, pages 3346–3351, Los Angeles, CA.,
2014.
[24] Claudio Carmeli, Ernesto De Vito, and Alessandro Toigo.
Vector valued reproducing kernel hilbert spaces of integrable
functions and mercer theorem. Analysis and Applications,
4(04):377–408, 2006.
[46] Marc G Genton. Classes of kernels for machine learning:
a statistics perspective. The Journal of Machine Learning
Research, 2:299–312, 2002.
[25] Ingo Steinwart and Andreas Christmann. Support vector
machines. Springer Science & Business Media, New York,
2008.
[47] Theodoros Evgeniou, Massimiliano Pontil, and Tomaso
Poggio.
Regularization networks and support vector
machines. Advances in computational mathematics, 13(1):1–
50, 2000.
[26] R Silverman.
Locally stationary random processes.
Information Theory, IRE Transactions on, 3(3):182–187,
1957.
[48] F. Cucker and S. Smale. On the mathematical foundations of
learning. American Mathematical Society, 39(1):1–49, 2002.
[27] C. E. Rasmussen and C. K. I. Williams. Gaussian Processes
for Machine Learning. MIT Press, Cambridge, MA, 2006.
[49] T. M. Cover and J. A. Thomas. Elements of information
theory. John Wiley & Sons, 2012.
[28] Akira Moiseevich Yaglom. Correlation theory of stationary
and related random functions. Springer, 1987.
[50] A. P. Dempster. Covariance selection. Biometrics, 28(1):157–
175, 1972.
[29] James Mercer. Functions of positive and negative type,
and their connection with the theory of integral equations.
Philosophical transactions of the royal society of London.
Series A, containing papers of a mathematical or physical
character, pages 415–446, 1909.
Proof of Proposition 3.1
[30] Harry Hochstadt. Integral equations. John Wiley & Sons,
1973.
Part (a). For any n ∈ N and for any ti ∈ X, i = 1, . . . , n,
let b̄ be the column vector containing b(t1 ), . . . , b(tn ) in
d
=
order. Then the kernel matrix K d defined by Ki,j
d
d
T
d
k (ti , tj ) is rewritten as K = b̄b̄ . Clearly, K is positive
semidefinite, and moreover rank(K d ) = 1. So k d (t, s) is
a rank-1 kernel by definition. The identity (14), Part (b)
and Part (c) can be verified in a straightforward way.
[31] Hongwei Sun. Mercer theorem for RKHS on noncompact
sets. Journal of Complexity, 21(3):337–349, 2005.
[32] Michel Loève. Probability theory, vol. II. Graduate texts in
mathematics, 1978.
[33] Brett Ninness and Graham C. Goodwin. Estimation of model
quality. Automatica, 31(12):1771 – 1797, 1995.
[34] L. Ljung. Model validation and model error modeling.
In B. Wittenmark and A. Rantzer, editors, The strm
Symposiium on Control, pages 15 –42, Lund, Sweden, Aug
1999. Studentlitteratur.
Proof of Proposition 3.2
[35] G. C. Goodwin, M. Gevers, and B. Ninness. Quantifying
the error in estimated transfer functions with application
to model order selection. IEEE Trans. Automatic Control,
37(7):913–929, 1992.
We give the proof for the CT case with X = {t|t ≥ 0}.
The results for the DT case with X = N hold by noting
that the DT kernel is the CT kernel restricted to N.
[36] G. C. Goodwin, J. H. Braslavsky, and M. M. Seron.
Non-stationary stochastic embedding for transfer function
estimation. Automatica, 38:47–62, 2002.
From (16) and (17), we see the proof will be done if it can
be shown that the kernels k c (t − s) in (16) and (17) are
stationary kernels. To show this, note from e.g. [27, page
85] that e−β|t−s| with β > 0 is called the exponential
kernel and belongs to the class of Matérn covariance
functions with order 1/2. So it remains to show k c (t − s)
in (16) is also a kernel. Note that the exponential kernel
is an isotropic stationary (IS) kernel (see Section 3.2.1
for its definition). Below we show that k c (t − s) in (16) is
actually also an IS kernel. As shown in [46], the spectral
[37] Wolfgang Reinelt, Andrea Garulli, and Lennart Ljung.
Comparing different approaches to model error modeling in
robust identification. Automatica, 38(5):787 – 803, 2002.
[38] Model Error Modeling and Stochastic Embedding, volume 48,
Beijing, China, 2015.
[39] C. Chen. Linear system theory and design. Oxford University
Press, New York, 3 edition, 1999.
[40] Torkel Glad and Lennart Ljung.
Control theory:
Multivariable and nonlinear methods. Taylor & Francis, 2000.
15
representation of an IS kernel k c (t, s) takes the form of
Z ∞
c
cos(ω|t − s|)F (dω)
(A.1)
k (t, s) =
Noting (19) and by Lemma A.1, we have
Hkc =
0
where F is any nondecreasing bounded function. For the
exponential kernel e−β|t−s| with β > 0, (A.1) is satisfied
with the spectral density
k(t, s) =
3
2
3β
=
4π
ω2)
−
1
2
Hk =
!
≥ 0,
Z
We give the proof for the CT case with X = {t|t ≥ 0}.
The proof for the DT case with X = N can be derived
in a similar way.
R∞
Part (a). Assume that b(t) ∈ L1 , i.e., 0 b(t)dt < ∞.
Since |k c (t − s)| ≤ 1 for any t, s ≥ 0. Then from
Z ∞ Z ∞
Z ∞
Z ∞
k(t, s)dt ds ≤
b(s)ds
b(t)dt < ∞,
Z
0
(A.3)
(
g|g =
∞
X
µi ρ i ,
∞
X
µ2
i
i=1
i=1
λi
)
<∞ .
(A.4)
(A.5)
∞
|g(t)|dt =
Z
0
∞
b(t)|h(t)|dt < ∞,
(A.6)
∞
|g(t)|dt =
Z
∞
0
b(t)|h(t)|dt ≥ ǫ
Z
∞
0
|h(t)|dt = ∞,
Proof of Proposition 4.1
Lemma A.1 [47, p. 16], [48, p. 37] Assume that k̄(t, s)
with t, s ∈ X is a positive semidefinite kernel and moreover, there exists a sequence of positive numbers νi and
linearly independent functions ψi (t) such that k̄(t, s) =
P
∞
i=1 νi ψi (t)ψi (s), where the convergence is absolute and
uniform on Y1 × Y2 with Y1 , Y2 being any compact subsets
of X. Then
)
(
∞
∞
X
X
µ2i
<∞
µi ψi ,
Hk̄ = g|g =
ν
i=1 i
i=1
To prove the result, we need a lemma.
Lemma A.2 [42,40] Consider a zero mean stationary
Gaussian process h(t) with covariance function k c (t − s).
Assume that k c (t − s) has rational power spectral density
Ψ(ω) 8 . Then the following results hold:
• For the DT case, there exists a rational function G
which has all poles inside the unit circle and all zeros
inside or on the unit circle such that
P∞
and for any f1 , f2 ∈ Hk̄ with f1 = i=1 ci ψi and f2 =
P
∞
i=1 di ψi , the inner product over Hk̄ is defined as
i=1
λi ρi (t)ρi (s), ρi (t) = b(t)φi (t),
which contradicts with (A.6). This completes the proof.
Part (b). Assume that the AMLS kernel (13) is stable,
i.e., Hk ⊂ L1 . We need a lemma to prove the result.
νi
∞
X
Since k c is not stable, Hkc * L1 and there exists h ∈ Hkc
but h ∈
/ L1 . For such h, if there exits an ǫ > 0 such that
b(t) ≥ ǫ for all t ≥ 0, then
0
∞
X
ci di
(A.2)
where the equality follows from (A.5) and h ∈ Hkc is the
function associated with g ∈ Hk .
and by Corollary 2.1, the AMLS kernel (13) is stable.
hf, giHk̄ =
<∞ .
Since Hk ⊂ L1 , for any g ∈ Hk , g ∈ L1 , i.e.,
0
0
λi
Hk = {g|g(t) = b(t)h(t), t ≥ 0, h ∈ Hkc }.
Proof of Proposition 3.3
0
i=1
i=1
)
By (A.4) and (A.2), we have
∀ω
Therefore, k c (t−s) in (16) is an IS kernel. This completes
the proof.
0
µi φi ,
i
and by Lemma A.1, we have
ω2)
1
1
− 3β
β 2
2
2
(2) + ω
( 2 ) + ω2
h|h =
∞
X
µ2
i=1
So k c (t − s) in (16) with λ = exp(−β/2) satisfies (A.1)
with a well defined spectral density described by
3
2β
3β 2
π(( 2 ) +
∞
X
Then from (19), the AMLS kernel (13) is rewritten as
β
f (ω) =
.
π(β 2 + ω 2 )
1
2β
β 2
π(( 2 ) +
(
Ψ(ω) = G(eiω )G(e−iω )
8
(A.7)
Check the statement of Theorem 4.1 for the definition of
rational power spectral density.
.
16
where | · | denotes the modulus of a complex number. Moreover, let the quintuple (Ā, B̄, C̄, D̄, Q̄) be
a state space realization of G(eiω ), i.e., G(eiω ) =
C̄(eiω Idim(Ā) − Ā)−1 B̄ + D̄, where dim(Ā) is the dimension of Ā and Idim(Ā) is the identity matrix with
dimension dim(Ā). Then the stationary Gaussian process h(t) has the following state space representation
Proof of Proposition 4.2
We only sketch the proof for the SS kernel and the proof
for the DC kernel can be derived in a similar way.
For the DT case, the IS kernel k c (t − s) in (16) has the
power spectral density
x(t + 1) = Āx(t) + B̄w(t), x(0) ∼ N (0, Q̄) (A.8a)
h(t) = C̄x(t) + D̄w(t)
(A.8b)
Ψ(ω) =
τ =−∞
(A.8c)
• For the CT case, there exists a rational function G
which has all poles in the left half plane and all zeros in
the left half plane or on the imaginary axis such that
Ψ(ω) = G(iω)G(−iω).
By spectral factorization technique, we have Ψ(ω) =
G(eiω )G(eiω ) with
(A.9)
G(e ) =
(āe−iω + b̄)
(1 − λ2 )3
.
2
(1 − λe−iω )(1 − λ3 e−iω )
For the CT case, the IS kernel k c (t − s) has the power
spectral density
(A.10a)
(A.10b)
Ψ(ω) =
Z
+∞
−∞
where w(t) is the zero mean white Gaussian noise with
unit power spectral density and Q̄ is the solution of the
Lyapunov equation
ĀQ̄ + Q̄ĀT + B̄ B̄ T = 0
r
iω
Moreover, let the quintuple (Ā, B̄, C̄, D̄, Q̄) be a state
space realization of G(iω), i.e., G(iω) = C̄(iωIdim Ā −
Ā)−1 B̄+ D̄. Then the stationary Gaussian process h(t)
has the following state space representation
ẋ(t) = Āx(t) + B̄w(t), x(0) ∼ N (0, Q̄)
h(t) = C̄x(t) + D̄w(t)
k c (τ )e−iωτ
3
1 − λ2
1 − λ6
1
=
−
−iω
iω
3
−iω
2 (1 − λe )(1 − λe ) 2 (1 − λ e )(1 − λ3 eiω )
(1 − λ2 )3
2 + 2λ2 + λ(e−iω + e−iω )
=
−iω
2
(1 − λe )(1 − λeiω )(1 − λ3 e−iω )(1 − λ3 eiω )
(āe−iω + b̄)(āeiω + b̄)
(1 − λ2 )3
=
−iω
2
(1 − λe )(1 − λeiω )(1 − λ3 e−iω )(1 − λ3 eiω )
where w(t) is the zero mean white Gaussian noise with
unit variance and Q̄ is the solution of the Lyapunov
equation
Q̄ = ĀQ̄ĀT + B̄ B̄ T
+∞
X
3 − 1 β|τ | 1 − 3 β|τ | −iwτ
2
2
e
dτ
− e
e
2
2
1
3
32 β 2
=
1
(iω + 2 β)(iω + 23 β)
(A.10c)
2
, |G(iω)|2 .
From the realization theory of linear systems, see e.g.,
[39], we can derive the corresponding state-space model
representation of the IS kernel k c (t − s), based on which
and by using the argument in the end of the proof of
Proposition 4.1, we derive (27) and (29), respectively.
Proof: For the DT case, the first part is a result of
the Spectral Factorization Theorem [42, Theorem 3.1 in
Chapter 4] and the second part is a result of the Representation Theorem [42, Theorem 3.2 in Chapter 4] and
[40, Theorem 5.3 and Equation (5.92)]. For the CT case,
the first part is a result of the Spectral Factorization Theorem [42, Theorem 5.1 in Chapter 4] and the second part
is a result of the Representation Theorem [42, Theorem
5.2 in Chapter 4] and [40, Theorem 5.3].
♦.
Proof of Proposition 4.3
For the DT case, note that there exists l > 0 such that
|CAt Q(As )T C T | ≤ l|λ̄|t+s ,
1
2
t+s
For the DT case, let z(t) = c λt x(t). Then simple calculation shows that the AMLS kernel (13) with k d (t, s) =
cλt+s is in the form of (24) with A = λĀ, B = λB̄,
1
C = C̄, D = D̄, Q = cQ̄ and b(t) = c 2 λt . For the
1
CT case, let z(t) = c 2 e−βt x(t) with λ = e−β . Then
simple calculation shows that the AMLS kernel (13)
with k d (t, s) = cλt+s is in the form of (24) with A =
Ā − βIdim Ā , B = B̄, C = C̄, D = D̄, Q = cQ̄ and
1
b(t) = c 2 e−βt .
|D2 b(t)b(s)δ(t − s)| ≤ l|λ̄| 2 ,
s−1
X
t+s
|Db(t)
δ(t − k)b(k)B T (As−1−k )T C T | ≤ l|λ̄| 2 ,
k=0
min{t,s}−1
X
b(k)2 CAt−1−k BB T (As−1−k )T C T
k=0
≤ l(|λ̄|t+s + |λ̄|t+s−min{t,s} ).
17
P∞ P∞
Then it is easy to show t=1 s=1 k SI (t, s) < ∞, and
thus the SI kernel (25a) is stable by Corollary 2.1.
(33) is equivalently rewritten as
maximize
For the CT case, note that there exists l > 0 such that
z̄(0),f (t)
H(z̄(0), f (0), f (1), · · · , f (s))
(A.12a)
subject to
E(z̄(0)) = 0, E(f (t)) = 0,
Cov(z̄(0)) = Q, V(f (t)) = 1, t = 0, · · · , s. (A.12b)
|CeAt Q(eAs )T C T | ≤ le−|Re(λ̄)|(t+s) ,
Z min{t,s}
C
b2 (τ )eA(t−τ ) BB T (eA(s−τ ) )T dτ C T
0
Note that the constraints in (A.12) are separable
with respect to z̄(0), f (t), t = 0, · · · , s, and that
H(z̄(0), f (0), f (1), · · · , f (s)) is maximized if z̄(0), f (0),
f (1), · · · , f (s) are independent with each other. Therefore, (A.12) is equivalent to
≤ l(e−|Re(λ̄)|(t+s) + e−|Re(λ̄)|(t+s−min{t,s}) ).
R∞R∞
Then it is easy to show 0 0 k SI (t, s) dtds < ∞, and
thus the SI kernel (25b) is stable by Corollary 2.1.
maximize
Proof of Proposition 4.4
H(z̄(0))
z̄(0)
subject to E(z̄(0)) = 0, Cov(z̄(0)) = Q
We only give the proof for the case D 6= 0 and the proof
for the case (b) is similar and thus omitted.
and for t = 0, · · · , s,
By chain rule, the differential entropy in (33) becomes
maximize H(f (t))
H(z̄(0), ḡ(0), ḡ(1), · · · , ḡ(s)) = H(z̄(0)) + H(ḡ(0)|z̄(0))
s
X
+
H(ḡ(t)|z̄(0), ḡ(0), · · · , ḡ(t − 1))
(A.11)
f (t)
subject to E(f (t)) = 0, V(f (t)) = 1
t=1
It is well known that multivariate normal distribution
maximizes the differential entropy over all distributions with the same covariance, see e.g., [49, Theorem 8.6.5]. Then we have the optimal solution to
(A.12) is that z̄(0) ∼ N (0, Q) and f (t) ∼ N (0, 1),
t = 0, · · · , s, and moreover, z̄(0), f (0), f (1), · · · , f (s)
are independent with each other. Finally, (32) implies
P
t−1−k
ḡ(t) = CAt z̄(0)+ t−1
Bb(k)f (k)+Db(t)f (t),
k=0 CA
same as g(t) in (25a). This completes the proof.
Note from (32) that
H(ḡ(0)|z̄(0)) = H(f (0)Db(0) + C z̄(0)|z̄(0))
= H(f (0)Db(0)|z̄(0))
= H(f (0)|z̄(0)) + log |Db(0)|
where the first equation follows because translation does
not change the differential entropy and the second equation follows because of the scaling property of differential entropy. Analogously, we have
Proof of Corollary 4.1
H(ḡ(t)|z̄(0), ḡ(0), · · · , ḡ(t − 1))
= H(ḡ(t)|z̄(0), f (0), · · · , f (t − 1))
= H(f (t)Db(t)|z̄(0), f (0), · · · , f (t − 1))
= H(f (t)|z̄(0), f (0), · · · , f (t − 1)) + log |Db(t)|
From Proposition 4.2, the DC kernel is a SI kernel with
A, B, C, D, Q, b(t) given in (28), which implies that Theorem 4.4 holds for the DC kernel. Comparing (32) with
(36) shows that the proof is completed if we can show
Therefore, (A.11) is rewritten as
H(z̄(0), ḡ(0), ḡ(1), · · · , ḡ(s)) = H(z̄(0)) + H(f (0)|z̄(0))
s
s
X
X
+
H(f (t)|z̄(0), f (0), · · · , f (t − 1)) +
log |Db(t)|
t=1
= H(z̄(0), f (0), f (1), · · · , f (s)) +
Cov(z̄(0)) = Q
t=0
V(ḡ(0)) = c,
V(f (t)) = 1 =⇒ V(ḡ(t) − λ1/2 ρḡ(t − 1))
= c(1 − ρ2 )λt , t = 1, · · · , s (A.13)
t=0
s
X
=⇒
log |Db(t)|
The first line of (A.13) is straightforward. In what follows, we consider the second line of (A.13). Clearly, the
results holds for the case either λ = 0 or ρ = 0. So we
further consider below the case where λ 6= 0 and ρ 6= 0.
Since the second term in the above equation is independent of z̄(0) and f (t), t = 0, · · · , s, the MaxEnt problem
18
It is easy to see that for t = 0, . . . , s − 1,
ḡ(t + 1) = ACAt z̄(0) + A
t−1
X
with order 1. Finally, the result that the kernel matrix of
the DC kernel has a 1-band matrix inverse follows from
Lemma A.3.
CAt−1−k Bb(k)f (k)
k=0
Proof of Proposition 4.6
+ ACA−1 Bb(t)f (t) + Db(t + 1)f (t + 1)
t−1
X
= ACAt z̄(0) + A
CAt−1−k Bb(k)f (k)
By using the realization of G0 (q) in controllable canonical form, the DT SI kernel (24) is written as follows:
k=0
+ ADb(t)f (t) + Db(t + 1)f (t + 1)
= Aḡ(t) + Db(t + 1)f (t + 1)
z(t + 1) =
where the second equality holds because CA−1 B = D.
Therefore (A.13) holds, which completes the proof.
g(t + 1) = −b̄
(a) For any t ∈ N and k > p, x(t) and x(t + k) are
conditionally independent given x(s) with s 6= t,
s 6= t + k and s ∈ N.
(b) Let i1 < i2 < · · · < in with n > p + 1 be any
strictly increasing subsequence taken from N. Then
the covariance matrix of [x(i1 ) · · · x(in )]T is a pband matrix.
1
1
t
1
1
1
t
g(t) = ρ(1 − ρ2 ) 2 z(t) + (1 − ρ2 ) 2 c 2 λ 2 w(t).
n
X
ai zi (t) + b̄b(t)w(t)
(A.15)
i=1
Therefore, (A.15) becomes
g(t + 1) = −
n
X
i=1
ai g(t − i + 1) + b̄b(t)w(t)
which is in the form of (38b) with order n. Thus the
Gaussian process g(t) in (24) is a Markov process with
order n and the fact that the kernel has an n-band matrix
inverse follows from Lemma A.3.
(A.14)
Note from Proposition 4.2 that g(t) can be written in
the following form
1
#
b̄zi (t) = g(t − i + 1), i = 1, · · · , n
where x, y ∈ R, z ∈ Rn and K ∈ R(n+2)×(n+2) . Then for
a given z, x, y are conditionally independent if and only
if (K −1 )1,2 = (K −1 )2,1 = 0, where (K −1 )i,j denote the
(i, j)th element of K −1 . Therefore, Part (b) follows from
the above observation and part (a).
♦
z(t + 1) = λ 2 ρz(t) + λ 2 c 2 λ 2 w(t), z(0) ∼ N (0,
1
z(t)
where zi (t) is the ith element of z(t). Note that
Proof: Part (a) follows from the pth order Markov property of x(t) and the fact that w(t) is white Gaussian
noise. To prove Part (b), recall from e.g., [50] that, for
Gaussian random variables, the zeros in the inverse of
the covariance matrix indicate conditional independence
of the two corresponding elements conditioned on the remaining ones. To be specific, consider a Gaussian random variable with mean m and covariance matrix K:
∼ N (m, K),
0n×1
From the above equation, we have
Lemma A.3 Consider the pth order Gaussian Markov
process (38). Then the following results hold:
iT
In
"
#
b(t)w(t)
0n×1
i
h
g(t) = b̄ 0 0 · · · 0 z(t), z(0) ∼ N (0, Q).
We need a lemma to prove this result.
xy z
−a1 · · · − an−1 −an
+
Proof of Proposition 4.5
h
"
c
)
1 − ρ2
which implies that (39) holds. Clearly, (39) is in the form
of (38a) with p = 1 and thus g(t) is a Markov process
19
| 3cs.SY
|
Meta Inverse Reinforcement Learning via Maximum Reward Sharing
for Human Motion Analysis
arXiv:1710.03592v2 [cs.AI] 12 Oct 2017
Kun Li1 , Joel W. Burdick1
Abstract— This work handles the inverse reinforcement
learning (IRL) problem where only a small number of demonstrations are available from a demonstrator for each highdimensional task, insufficient to estimate an accurate reward
function. Observing that each demonstrator has an inherent
reward for each state and the task-specific behaviors mainly
depend on a small number of key states, we propose a meta IRL
algorithm that first models the reward function for each task as
a distribution conditioned on a baseline reward function shared
by all tasks and dependent only on the demonstrator, and then
finds the most likely reward function in the distribution that
explains the task-specific behaviors. We test the method in a
simulated environment on path planning tasks with limited
demonstrations, and show that the accuracy of the learned
reward function is significantly improved. We also apply the
method to analyze the motion of a patient under rehabilitation.
(a)
(b)
(c)
(d)
I. I NTRODUCTION
Inverse reinforcement learning (IRL) [1] algorithms estimate a reward function that explains the motions demonstrated by an operator or other agents on a task described by
a Markov Decision Process (MDP) [2]. The recovered reward
function can be used by a robot to replicate the demonstrated
task [3], or by an algorithm to analyze the demonstrator’s
preference [4]. Therefore, IRL algorithms can make multitask robot control simpler by alleviating the need to explicitly
set a cost function for each task, and make robot friendlier
by personalizing services based on the recovered condition
and preference of the operator.
The accuracy of the recovered function depends heavily on
the ratio of visited states in the demonstrations to the whole
state space, because the demonstrator’s motion policy can be
estimated more accurately if every state is repeatedly visited.
However, the ratio is low for many useful applications, since
they usually have huge or high-dimensional state spaces,
while the demonstrations are relatively rare for each task. For
example, in a path planning task on a mild 100 ∗ 100 grid,
the demonstrator chooses paths based on the destination, but
may not move to the same destination hundreds of times
in practice. For robot manipulation tasks based on ordinary
640 ∗ 480 ∗ 3 RGB images, east task specifies a final result,
but it is expensive to repeat each task millions of times. For
human motion analysis, it is physically improbable to follow
an instruction thousands of times in the huge state space of
human poses. Therefore, it is difficult to estimate an accurate
reward function for a single task with limited data.
*This work was supported by the National Institutes of Health, NIBIB.
Li and Joel W. Burdick are with Department of Mechanical and
Civil Engineering, California Institute of Technology, Pasadena, CA 91125,
USA [email protected]
1 Kun
Fig. 1: Different behaviors under different goal states and
goal rewards: Figure 1a and Figure 1b share the same goal
state, but the goal reward of Figure 1b is larger than Figure
1a.Figure 1c and Figure 1d share the same goal state, but
the goal reward of Figure 1d is larger than Figure 1c. Five
trajectories are plotted in each figure, where red dots denote
the starting point and black dots denote the ending point.
In practice, usually multiple tasks can be observed from
the same demonstrator, and the problem of rare demonstrations can be handled by combining data from all tasks,
hence the meta-learning problem. Existing solutions mainly
classification problems, like using the data from all tasks
to learn an optimizer for each task, using the data from all
tasks to learn a metric space where a single task can be more
accurate with limited data, using the data from all tasks to
learn a good initialization or a good initial parameter for each
task, etc. Some of these methods are applicable to inverse
reinforcement learning problems, but they mainly consider
transfer of motion policy.
In many IRL applications, we observe that a demonstrator
usually has an inherent reward for each state, materialized
as the innate state preferences of a human, the hardwaredependent cost function of a robot, the default structure of
an environment, etc. For a given task, the demonstrators are
usually reluctant to drastically change the inherent reward
function to complete the task; instead, they alter the innate
reward function minimally to generate a task-specific reward
function and plan the motion. For example, in path planning,
the C-space of a mobile robot at home rarely changes, and the
robot’s motion depends on the goal state; in human motion
analysis, the costs of different poses are mostly invariant,
while the actual motion depends on the desired directions.
Based on this observation, we propose a meta inverse
reinforcement learning algorithm by maximizing the shared
rewards among all tasks. We model the reward function
for each task as a probabilistic distribution conditioned on
an inherent baseline function, and estimate the most likely
reward function in the distribution that explains the observed
task-specific demonstrations.
We review existing IRL and meta-learning algorithms
in Section II, and then introduce the proposed method in
Section III. We show a simulated experiments for evaluation
and a real-world experiment for application in Section IV,
with conclusions in Section V.
II. R ELATED W ORKS
The idea of inverse optimal control is proposed by Kalman
[5], white the inverse reinforcement learning problem is
firstly formulated in [1], where the agent observes the states
resulting from an assumingly optimal policy, and tries to
learn a reward function that makes the policy better than
all alternatives. Since the goal can be achieved by multiple
reward functions, this paper tries to find one that maximizes
the difference between the observed policy and the second
best policy. This idea is extended by [6], in the name of
max-margin learning for inverse optimal control. Another
extension is proposed in [3], where the purpose is not
to recover the real reward function, but to find a reward
function that leads to a policy equivalent to the observed one,
measured by the amount of rewards collected by following
that policy.
Since a motion policy may be difficult to estimate from
observations, a behavior-based method is proposed in [7],
which models the distribution of behaviors as a maximumentropy model on the amount of reward collected from each
behavior. This model has many applications and extensions.
For example, [8] considers a sequence of changing reward
functions instead of a single reward function. [9] and [10]
consider complex reward functions, instead of linear one, and
use Gaussian process and neural networks, respectively, to
model the reward function. [11] considers complex environments, instead of a well-observed Markov Decision Process,
and combines partially observed Markov Decision Process
with reward learning. [12] models the behaviors based on the
local optimality of a behavior, instead of the summation of
rewards. [13] uses a multi-layer neural network to represent
nonlinear reward functions.
Another method is proposed in [14], which models the
probability of a behavior as the product of each state-action’s
probability, and learns the reward function via maximum a
posteriori estimation. However, due to the complex relation
between the reward function and the behavior distribution, the author uses computationally expensive Monte-Carlo
methods to sample the distribution. This work is extended
by [15], which uses sub-gradient methods to simplify the
problem. Another extensions is shown in [16], which tries to
find a reward function that matches the observed behavior.
For motions involving multiple tasks and varying reward
functions, methods are developed in [17] and [18], which
try to learn multiple reward functions.
Most of these methods need to solve a reinforcement
learning problem in each step of reward learning, thus practical large-scale application is computationally infeasible.
Several methods are applicable to large-scale applications.
The method in [1] uses a linear approximation of the value
function, but it requires a set of manually defined basis
functions. The methods in [10], [19] update the reward function parameter by minimizing the relative entropy between
the observed trajectories and a set of sampled trajectories
based on the reward function, but they require a set of
manually segmented trajectories of human motion, where
the choice of trajectory length will affect the result. Besides,
these methods solve large-scale problems by approximating
the Bellman Optimality Equation, thus the learned reward
function and Q function are only approximately optimal. In
our previous work, we proposed an approximation method
that guarantees the optimality of the learned functions as well
as the scalability to large state space problems [20].
To learn a model from limited data, meta learning algorithms are developed. A survey of different work is given
in [21], viewing meta-learner as a way to improve biases
for base-learners. The method in [22] uses neural memory
machine to do the meta learning. The method in [23]
minimizes the representations. The method in [24] learns by
gradient descent. The method [25] learns optimizers.
Meta learning algorithms are also applied to reinforcement
learning problems. The method in [26] tunes meta parameters
for reinforcement learning, learning rate for TD learning,
action selection trade-off, and discount factor. The method in
[27] uses one network to play multiple games. The method in
[28] trains reinforcement learning with slower rl. The method
in [29] learns a good initial parameter that reaches optimal
parameters with limited gradient descent. Meta learning in
inverse reinforcement learning focuses on imitation learning,
like one-shot imitation learning[30].
III. M ETA I NVERSE R EINFORCEMENT L EARNING
A. Meta Inverse Reinforcement Learning
We assume that an agent needs to handle multiple tasks in
an environment, denoted by {Ti |i = 1, NT }, where Ti denotes
the ith task and NT denotes the number of tasks.
We describe a task Ti as a Markov Decision Process,
consisting of the following variables:
• Si = {s}, a set of states
• Ai = {a}, a set of actions
a
• Pi,ss′ , a state transition function that defines the probability that state s becomes s′ after action a.
Ri = {ri (s)}, a reward function that defines the immediate reward of state s.
• γi , a discount factor that ensures the convergence of the
MDP over an infinite horizon.
For a task Ti , the agent performs a set of demonstrations
ζi = {ζi, j | j = 1, · · · , Nζi }, represented by Nζi sequences of
state-action pairs:
•
ζi, j = {(sti, j , ati, j )|t = 0, · · · , Nζi, j },
where Nζi, j denotes the length of the jth sequence ζi, j .
Given the observed sequences ζ = {ζi |i = 1, · · · , NT } for the
NT tasks, inverse reinforcement learning algorithms try to
recover a reward function ri (s) for each task.
Our key observation in multi-task IRL is that the demonstrator has an inherent reward function rb (s), generating a
baseline reward for each state in all tasks. To complete the
ith task, the agent generates a reward function ri (s) from a
distribution P(ri |rb ) conditioned on rb (s) to plan the motion.
Therefore, the motion ζi is generated as:
P(ζi |ri )P(ri |rb )
For the ith task, we want to find the most likely ri (s)
sampled from P(ri |rb ) that explains the demonstration ζi .
Assuming all the tasks are independent from each other, the
following joint distribution is formulated:
where θi denotes the parameter of VR function. The function
value of a state is named as VR value.
Substituting Equation (4) into Bellman Optimality Equation, the optimal Q function is given as:
Q∗i (s, a) =
Vi∗ (s) = max
The reward functions can be found via maximumlikelihood estimation:
NT
, ri (s)) + L(ri (s), rb (s))
∑ |Li (ζi{z
} |
{z
}
rb (s),r1 (s),··· ,rT (s)∈F
min
i=1
IRL loss
(1)
reward sharing loss
where F denotes a function space, Li (ζi , ri (s)) is the negative
loglikelihood of P(ζi |ri ), and L(ri (s), rb (s)) is the negative
loglikelihood P(ri |rb ).
B. Loss for Inverse Reinforcement Learning
While many solutions exist for the inverse reinforcement
learning problem, we adopt the solution based on function
approximation developed in our earlier work [20] to handle
the practical high-dimensional state spaces.
The core idea of the method is to approximate the Bellman Optimality Equation [2] with a function approximation
framework. The Bellman Optimality Equation is given as:
Vi∗ (s) = max
a∈A
Q∗i (s, a)
=
∗ ′
′
a
Pi,ss
′ [ri (s ) + γ ∗ Vi (s )],
∑
′
s |s,a
∗ ′ ′
a
′
Pi,ss
′ [ri (s ) + γ ∗ max Qi (s , a )].
′ ∈A
a
s′ |s,a
∑
(2)
(3)
′
a
′ f i (s , θi ),
∑ Pi,ss
a∈A ′
s |s,a
(6)
and the reward function can be computed as:
ri (s) = fi (s, θi ) − γ ∗ max
′
a
′ f i (s , θi ).
∑ Pi,ss
a∈A ′
s |s,a
(7)
This framework avoids solving the Bellman Optimality
Equation. Besides, this formulation can be generalized to
other extensions of Bellman Optimality Equation by replacing the max operator with other types of Bellman backup operators. For example, V ∗ (s) = log ∑a∈A exp Q∗ (s, a) is used in
the maximum-entropy method[7]; V ∗ (s) = 1k log ∑a∈A exp k ∗
Q∗ (s, a) is used in Bellman Gradient Iteration [31].
To apply this framework to IRL problems, this work
chooses a motion model p(a|s) based on the optimal Q
function Q∗i (s, a) [14]:
P(a|s) =
i=1
(5)
the optimal value function is given as:
NT
∏ P(ζi |ri )P(ri |rb )
a
′
′ f i (s , θi ),
∑ Pi,ss
s′ |s,a
expb ∗ Q∗i (s, a)
,
∑ã∈A exp b ∗ Q∗i (s, ã)
(8)
where b is a parameter controlling the degree of confidence in
the agent’s ability to choose actions based on Q values. Other
models can also be used, like p(a|s) = exp(Q(s, a) − V (s))
in [7].
Assuming the approximation function is a neural network,
the parameter θi = {w, b}-weights and biases, the negative
log-likelihood of P(ζi |θi ) is given by:
Li (θi ) = −
∑
(b ∗ Q∗i (s, a) − log ∑ exp b ∗ Q∗i (s, â)), (9)
(s,a)∈ζi
â∈A
where the optimal Q function is given by Equation (5). After
estimating the parameter θi , the value function and reward
function can be computed with Equation (4), (6), and (7).
C. Loss for Reward Sharing
Since the demonstrator makes minimal changes to adapt
the inherent reward function rb (s) into task-specific one ri (s),
we model the distribution as:
P(ri (s)|rb (s)) ∝ exp(D(ri (s), rb (s))))
where D(ri (s), rb (s))) measures the difference between ri (s)
and rb (s). Thus the loss function for reward sharing is given
as:
L(ri (s), rb (s)) = log Z − D(ri (s), rb (s)))
It is computationally prohibitive to solve in high-dimensional
state spaces.
But with a parameterized VR function, we describe the
summation of the reward function and the discounted optimal
value function as:
where log Z is the partition function and remains the same
for all ri (s).
We test several functions as D(ri (s), rb (s))). The first
choice is L2 loss, where
fi (s, θi ) = ri (s) + γ ∗ Vi∗ (s),
D(ri (s), rb (s)) = ||{ri (s) − rb (s)}||
(4)
Algorithm 1 Meta IRL
1:
2:
3:
4:
5:
6:
7:
8:
9:
Data: ζ , S, A, P, γ , b, α
Result: reward value R[S]
create variable θb = {W, b} for a neural network
create variable θi = {W, b}, i = 1, · · · , T for each task
initialize θb , θi = {W, b}, i = 1, · · · , T
while Not converging do
update θb , θi = {W, b}, i = 1, · · · , T based on optimization (1)
end while
compute and return all Ri [S]
where {ri (s)− rb (s)} denotes the set of differences, evaluated
on the full state space or only the visited states.
The second choice is Huber loss with δ = 1, a differentiable approximation of the L1 loss popular in sparse models:
(
1 2
a
for |a| ≤ δ ,
,
Lδ (a) = 2
δ (|a| − 21 δ ), otherwise,
each task, the agent demonstrates ten trajectories, where the
length of a trajectory depends on how many steps to reach
the goal state.
We evaluate the proposed method with different reward
sharing loss functions under different number of tasks and
different number of trajectories. The evaluated loss functions
include no reward sharing, reward sharing with standard
deviation, information entropy, L2 loss, and huber loss. The
number of tasks ranges from 1 to 16, and for each task, the
number of trajectories ranges from 1 to 10. The learning rate
is 0.01, with Adam optimizer. The accuracy of a reward is
computed as the correlation coefficient between the learned
reward function and the ground truth one. The results are
shown in Figure 2.
The result shows that the meta learning step can significantly improve the accuracy of reward learning, among
which the huber loss function leads to the best performance
in average. L2 loss and standard deviation have similar performance, not surprisingly. However, the information entropy
has a really bad performance.
B. Motion Analysis
and
D(ri (s), rb (s)) = ∑ L1 (ri (s) − rb (s)).
s
The third choice is standard deviation:
D(ri (s), rb (s)) = stdev({ri (s) − rb (s)})
The fourth choice is information entropy, after converting
{ri (s) − rb (s)} into a probabilistic distribution with sofmax
function:
D(ri (s), rb (s)) = entropy(so f tmax({ri (s) − rb (s)}))
With the loss function for IRL and reward sharing, the
reward functions can be learned via gradient method. The
algorithm is shown in Algorithm 1.
IV. E XPERIMENTS
A. Path Planning
We consider a path planning problem on an uneven terrain,
where an agent can observe the whole terrain to find the
optimal paths from random starting points to arbitrary goal
points, but a mobile robot can only observe the agent’s
demonstrations to learn how to plan paths. Given a starting
point and a goal point, an optimal path depends solely on
the costs to move across the terrain. To learn the costs, we
formulate a Markov Decision Process for each goal point,
where a state denotes a small region of the terrain and
an action denotes a possible movement. The reward of a
state equals to the negative of the cost to move across the
corresponding region, while the goal state has an additional
reward to attract movements.
In this work, we create a discretized terrain with several
hills, where each hill is defined as a peak of cost distribution
and the costs around each hill decay exponentially, and the
true cost of a region is the summation of the costs from all
hills. Ten worlds are randomly generated, and in each world,
ten tasks are generated, each with a different goal state. For
During rehabilitation, a patient with spinal cord injuries
sits on a box, with a flat plate force sensor mounted on box
to capture the center-of-pressure (COP) of the patient during
movement. Each experiment is composed of two sessions,
one without transcutaneous stimulation and one with stimulation. The electrodes configuration and stimulation signal
pattern are manually selected by the clinician [32].
In each session, the physician gives eight (or four) directions for the patient to follow, including left, forward
left, forward, forward right, right, right backward, backward,
backward left, and the patient moves continuously to follow
the instruction. The physician observes the patient’s behaviors and decides the moment to change the instruction.
Six experiments are done, each with two sessions. The
COP trajectories in Figure 3 denote the case with four
directional instructions; Figure 4, 5, 6, 7, and 8 denote the
sessions with eight directional instructions.
The COP sensory data from each session is discretized on
a 100×100 grid, which is fine enough to capture the patient’s
small movements. The problem is formulated into a MDP,
where each state captures the patient’s discretized location
and velocity, and the set of actions changes the velocity into
eight possible directions. The velocity is represented with
a two-dimensional vector showing eight possible velocity
directions. Thus the problem has 80000 states and 8 actions,
and each action is assumed to lead to a deterministic state.
To learn the reward function from the observed trajectories
based on the formulated MDP, we use the coordinate and
velocity direction of each grid as the feature, and learn
the reward function parameter from each set of data after
segmentation based on peak detection on distances from the
origin. The function approximator is a neural network with
three hidden layers and [100, 50, 25] nodes. The huber loss
function is used in reward sharing, and the result is show in
Table I.
Experiments with 1 tasks
0.55
Experiments with 3 tasks
Experiments with 2 tasks
0.50
Experiments with 4 tasks
0.5
0.20
2
3
4
5
6
7
Number of Trajectories
8
9
0.10
10
1
2
3
4
(a) 1 task
0.3
no sharing
stdev
entropy
huber
l2
0.2
0.1
3
4
5
6
7
Number of Trajectories
8
9
Mean Correlation with the Ground Truth
0.3
no sharing
stdev
entropy
huber
l2
0.2
0.1
1
2
3
4
5
6
7
Number of Trajectories
8
9
0.2
0.1
3
4
5
6
7
Number of Trajectories
8
9
0.4
0.3
no sharing
stdev
entropy
huber
l2
0.1
1
2
3
no sharing
stdev
entropy
huber
l2
0.2
0.1
2
3
4
5
6
7
Number of Trajectories
8
9
(m) 13 tasks
10
Mean Correlation with the Ground Truth
0.3
1
4
2
3
no sharing
stdev
entropy
huber
l2
0.2
2
3
4
5
6
7
Number of Trajectories
4
5
6
7
Number of Trajectories
8
9
10
Experiments with 8 tasks
0.6
8
9
0.5
0.4
0.3
no sharing
stdev
entropy
huber
l2
0.2
0.1
10
1
2
3
4
5
6
7
Number of Trajectories
8
9
10
(h) 8 tasks
Experiments with 11 tasks
Experiments with 12 tasks
0.6
5
6
7
Number of Trajectories
8
9
0.5
0.4
0.3
no sharing
stdev
entropy
huber
l2
0.2
0.1
10
1
2
3
0.4
0.3
no sharing
stdev
entropy
huber
l2
0.2
0.1
2
3
4
5
6
7
Number of Trajectories
5
6
7
Number of Trajectories
8
9
0.4
0.3
no sharing
stdev
entropy
huber
l2
0.2
0.1
10
1
2
3
8
9
10
0.4
0.3
no sharing
stdev
entropy
huber
l2
0.2
0.1
2
(n) 14 tasks
3
4
5
6
7
Number of Trajectories
5
6
7
Number of Trajectories
8
9
10
Experiments with 16 tasks
0.6
0.5
1
4
(l) 12 tasks
Experiments with 15 tasks
0.6
0.5
1
4
0.5
(k) 11 tasks
Experiments with 14 tasks
0.6
0.4
1
(d) 4 tasks
0.3
(j) 10 tasks
0.5
0.1
10
0.4
0.7
0.2
Experiments with 13 tasks
no sharing
stdev
entropy
huber
l2
0.2
0.6
(i) 9 tasks
0.6
9
0.3
(g) 7 tasks
0.5
10
8
Experiments with 7 tasks
1
Mean Correlation with the Ground Truth
no sharing
stdev
entropy
huber
l2
Mean Correlation with the Ground Truth
0.3
5
6
7
Number of Trajectories
0.1
0.6
0.4
4
0.5
10
Experiments with 10 tasks
0.7
0.5
2
3
(f) 6 tasks
Experiments with 9 tasks
1
2
0.6
0.4
10
0.6
Mean Correlation with the Ground Truth
1
0.4
(c) 3 tasks
0.5
(e) 5 tasks
0.7
0.1
10
Mean Correlation with the Ground Truth
Mean Correlation with the Ground Truth
0.4
2
9
Experiments with 6 tasks
0.6
0.5
1
8
no sharing
stdev
entropy
huber
l2
0.2
(b) 2 tasks
Experiments with 5 tasks
0.6
5
6
7
Number of Trajectories
Mean Correlation with the Ground Truth
1
no sharing
stdev
entropy
huber
l2
0.20
0.15
Mean Correlation with the Ground Truth
no sharing
stdev
entropy
huber
l2
0.25
0.3
Mean Correlation with the Ground Truth
0.30
0.30
0.25
0.4
Mean Correlation with the Ground Truth
0.35
0.40
0.35
8
9
10
Mean Correlation with the Ground Truth
0.40
Mean Correlation with the Ground Truth
0.45
Mean Correlation with the Ground Truth
Mean Correlation with the Ground Truth
0.50
Mean Correlation with the Ground Truth
0.5
0.45
0.5
0.4
0.3
no sharing
stdev
entropy
huber
l2
0.2
0.1
1
(o) 15 tasks
2
3
4
5
6
7
Number of Trajectories
8
9
10
(p) 16 tasks
Fig. 2: The result with five reward sharing loss functions on 16 tasks with at most 10 demonstrations for each task in 10
environment.
TABLE I: Evaluation of the learned rewards: ”forward” etc. denote the instructed direction; item name”1u” denotes the
patient id ”1”, with ”u” denoting unstimulated session and ”s” denoting stimulated sessions. The table shows the correlation
coefficient between the ideal reward and the recovered reward.
1u
1s
2u
2s
3u
3s
4u
4s
5u
5s
6u
6s
forward
0.411741
0.200355
0.161595
-0.0310265
0.362801
-0.265834
0.301472
-0.104719
0.360293
-0.212823
-0.0416432
-0.157148
backward
0.257564
0.486723
-0.17814
-0.0803484
-0.2995
0.146516
-0.281474
0.0930068
-0.311692
0.0414435
0.0570847
0.178879
left
0.0691989
0.129839
0.153376
0.0474505
0.245916
0.379665
0.377787
0.327783
-0.253715
0.0908994
0.210028
0.0880393
right
-0.210216
0.436533
-0.16767
-0.00146299
-0.178778
-0.272437
-0.320403
-0.229091
0.260426
-0.124174
-0.104113
-0.0718817
top left
top right
bottom left
bottom right
0.162906
0.0442916
0.386421
0.138805
0.410212
0.175432
0.0863029
0.00414109
0.0363181
-0.102579
0.105993
0.0874981
0.0148849
-0.2683
-0.119599
-0.161819
0.495134
-0.107462
-0.0672399
-0.298918
-0.0211192
0.00668849
-0.00335653
0.212331
0.136309
0.323862
-0.38137
0.122018
0.0704143
0.307328
-0.220457
0.0742221
-0.385605
0.00301386
-0.306677
-0.0521654
-0.140836
0.0453461
-0.00392284
0.171319
origin
0.49016
0.207188
0.156034
0.0726437
0.0507719
-0.182916
0.171433
-0.202197
-0.160687
0.145686
0.190253
0.359168
Fig. 3: Patient 1 under four directional instructions: ”unstimulated motion” means that the patient moves without transcutaneous stimulations, while ”stimulated motion” represents
the motion under stimulations.
Fig. 6: Patient 4 under eight directional instructions: ”unstimulated motion” means that the patient moves without
transcutaneous stimulations, while ”stimulated motion” represents the motion under stimulations.
Fig. 4: Patient 2 under eight directional instructions: ”unstimulated motion” means that the patient moves without
transcutaneous stimulations, while ”stimulated motion” represents the motion under stimulations.
Fig. 7: Patient 5 under eight directional instructions: ”unstimulated motion” means that the patient moves without
transcutaneous stimulations, while ”stimulated motion” represents the motion under stimulations.
Fig. 5: Patient 3 under eight directional instructions: ”unstimulated motion” means that the patient moves without
transcutaneous stimulations, while ”stimulated motion” represents the motion under stimulations.
Fig. 8: Patient 6 under eight directional instructions: ”unstimulated motion” means that the patient moves without
transcutaneous stimulations, while ”stimulated motion” represents the motion under stimulations.
It shows that the patient’s ability to following instructions
vary among different directions, and the values will assist
physicians to design the stimulating signals.
V. C ONCLUSIONS
This work proposes a solution to learn an accurate reward
function for each task with limited demonstrations but from
the same demonstrator, by maximizing the shared rewards
among different tasks. We proposed several loss functions to
maximize the shared reward, and compared their accuracies
in a simulated environment. It shows that huber loss has the
best performance.
In future work, we will apply the proposed method to
imitation learning.
R EFERENCES
[1] A. Y. Ng and S. Russell, “Algorithms for inverse reinforcement
learning,” in in Proc. 17th International Conf. on Machine Learning,
2000.
[2] R. S. Sutton and A. G. Barto, Reinforcement learning: An introduction.
MIT press Cambridge, 1998, vol. 1, no. 1.
[3] P. Abbeel and A. Y. Ng, “Apprenticeship learning via inverse reinforcement learning,” in Proceedings of the twenty-first international
conference on Machine learning. ACM, 2004, p. 1.
[4] B. Najafi, K. Aminian, A. Paraschiv-Ionescu, F. Loew, C. J. Bula, and
P. Robert, “Ambulatory system for human motion analysis using a
kinematic sensor: monitoring of daily physical activity in the elderly,”
IEEE Transactions on biomedical Engineering, vol. 50, no. 6, pp.
711–723, 2003.
[5] R. Kalman and M. M. C. B. D. R. I. for Advanced Studies. Center for
Control Theory, When is a Linear Control System Optimal?., ser. RIAS
technical report. Martin Marietta Corporation, Research Institute for
Advanced Studies, Center for Control Theory, 1963.
[6] N. D. Ratliff, J. A. Bagnell, and M. A. Zinkevich, “Maximum margin
planning,” in Proceedings of the 23rd international conference on
Machine learning. ACM, 2006, pp. 729–736.
[7] B. D. Ziebart, A. Maas, J. A. Bagnell, and A. K. Dey, “Maximum
entropy inverse reinforcement learning,” in Proc. AAAI, 2008, pp.
1433–1438.
[8] Q. P. Nguyen, B. K. H. Low, and P. Jaillet, “Inverse reinforcement
learning with locally consistent reward functions,” in Advances in
Neural Information Processing Systems, 2015, pp. 1747–1755.
[9] S. Levine, Z. Popovic, and V. Koltun, “Nonlinear inverse reinforcement
learning with gaussian processes,” in Advances in Neural Information
Processing Systems 24, J. Shawe-Taylor, R. S. Zemel, P. L. Bartlett,
F. Pereira, and K. Q. Weinberger, Eds. Curran Associates, Inc., 2011,
pp. 19–27.
[10] C. Finn, S. Levine, and P. Abbeel, “Guided cost learning: Deep
inverse optimal control via policy optimization,” arXiv preprint
arXiv:1603.00448, 2016.
[11] J. Choi and K.-E. Kim, “Inverse reinforcement learning in partially
observable environments,” Journal of Machine Learning Research,
vol. 12, no. Mar, pp. 691–730, 2011.
[12] S. Levine and V. Koltun, “Continuous inverse optimal control with
locally optimal examples,” arXiv preprint arXiv:1206.4617, 2012.
[13] M. Wulfmeier, P. Ondruska, and I. Posner, “Deep inverse reinforcement learning,” arXiv preprint arXiv:1507.04888, 2015.
[14] D. Ramachandran and E. Amir, “Bayesian inverse reinforcement
learning,” in Proceedings of the 20th International Joint Conference
on Artifical Intelligence, ser. IJCAI’07. San Francisco, CA, USA:
Morgan Kaufmann Publishers Inc., 2007, pp. 2586–2591.
[15] G. Neu and C. Szepesvári, “Apprenticeship learning using inverse reinforcement learning and gradient methods,” arXiv preprint
arXiv:1206.5264, 2012.
[16] K. Mombaur, A. Truong, and J.-P. Laumond, “From human to humanoid locomotionan inverse optimal control approach,” Autonomous
robots, vol. 28, no. 3, pp. 369–383, 2010.
[17] C. Dimitrakakis and C. A. Rothkopf, “Bayesian multitask inverse
reinforcement learning,” in European Workshop on Reinforcement
Learning. Springer, 2011, pp. 273–284.
[18] J. Choi and K.-E. Kim, “Nonparametric bayesian inverse reinforcement learning for multiple reward functions,” in Advances in Neural
Information Processing Systems, 2012, pp. 305–313.
[19] A. Boularias, J. Kober, and J. R. Peters, “Relative entropy inverse
reinforcement learning,” in International Conference on Artificial
Intelligence and Statistics, 2011, pp. 182–189.
[20] K. Li and J. W. Burdick, “Large-scale inverse reinforcement learning via function approximation for clinical motion analysis,” arXiv
preprint arXiv:1707.09394, 2017.
[21] R. Vilalta and Y. Drissi, “A perspective view and survey of metalearning,” Artificial Intelligence Review, vol. 18, no. 2, pp. 77–95,
2002.
[22] A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap,
“Meta-learning with memory-augmented neural networks,” in International conference on machine learning, 2016, pp. 1842–1850.
[23] B. Hariharan and R. Girshick, “Low-shot visual recognition by shrinking and hallucinating features,” arXiv preprint arXiv:1606.02819,
2016.
[24] M. Andrychowicz, M. Denil, S. Gomez, M. W. Hoffman, D. Pfau,
T. Schaul, and N. de Freitas, “Learning to learn by gradient descent
by gradient descent,” in Advances in Neural Information Processing
Systems, 2016, pp. 3981–3989.
[25] D. Ha, A. Dai, and Q. V. Le, “Hypernetworks,” arXiv preprint
arXiv:1609.09106, 2016.
[26] N. Schweighofer and K. Doya, “Meta-learning in reinforcement learning,” Neural Networks, vol. 16, no. 1, pp. 5–9, 2003.
[27] E. Parisotto, J. L. Ba, and R. Salakhutdinov, “Actor-mimic:
Deep multitask and transfer reinforcement learning,” arXiv preprint
arXiv:1511.06342, 2015.
[28] Y. Duan, J. Schulman, X. Chen, P. L. Bartlett, I. Sutskever, and
P. Abbeel, “Rl2: Fast reinforcement learning via slow reinforcement
learning,” arXiv preprint arXiv:1611.02779, 2016.
[29] C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for
fast adaptation of deep networks,” arXiv preprint arXiv:1703.03400,
2017.
[30] Y. Duan, M. Andrychowicz, B. Stadie, J. Ho, J. Schneider, I. Sutskever,
P. Abbeel, and W. Zaremba, “One-shot imitation learning,” arXiv
preprint arXiv:1703.07326, 2017.
[31] K. Li and J. W. Burdick, “Bellman Gradient Iteration for Inverse
Reinforcement Learning,” ArXiv e-prints, Jul. 2017.
[32] S. Harkema, Y. Gerasimenko, J. Hodes, J. Burdick, C. Angeli, Y. Chen,
C. Ferreira, A. Willhite, E. Rejc, R. G. Grossman et al., “Effect of
epidural stimulation of the lumbosacral spinal cord on voluntary movement, standing, and assisted stepping after motor complete paraplegia:
a case study,” The Lancet, vol. 377, no. 9781, pp. 1938–1947, 2011.
| 2cs.AI
|
An Adaptive Amoeba Algorithm for Shortest Path Tree
Computation in Dynamic Graphs
Xiaoge Zhang
a,f ,
Qi Liub,c,g , Yong Hud , Felix T. S. Chane , Sankaran Mahadevanf , Zili Zhanga
,g , Yong Deng a,f*
arXiv:1311.0460v1 [cs.NE] 3 Nov 2013
a School
of Computer and Information Science, Southwest University, Chongqing 400715,
China; b Department of Biomedical Informatics, Medical Center, Vanderbilt University,
Nashiville, 37235, USA; c School of Life Sciences and Biotechnology, Shanghai Jiao Tong
University, Shanghai, 200030, China; d Institute of Business Intelligence and Knowledge
Discovery, Guangdong University of Foreign Studies, Guangzhou 510006, China; e Department
of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hum,
Kowloon, Hong Kong f School of Engineering, Vanderbilt University, Nashiville, 37235, USA g
These authors contribute equally
* Correspondence and requests for materials should be addressed to Y. D.
([email protected])
This paper presents an adaptive amoeba algorithm to address the shortest path tree (SPT) problem in
dynamic graphs. In dynamic graphs, the edge weight updates consists of three categories: edge weight increases,
edge weight decreases, the mixture of them. Existing work on this problem solve this issue through analyzing
the nodes influenced by the edge weight updates and recompute these affected vertices. However, when the
network becomes big, the process will become complex. The proposed method can overcome the disadvantages
of the existing approaches. The most important feature of this algorithm is its adaptivity. When the edge
weight changes, the proposed algorithm can recognize the affected vertices and reconstruct them spontaneously.
To evaluate the proposed adaptive amoeba algorithm, we compare it with the Label Setting algorithm and
Bellman-Ford algorithm. The comparison results demonstrate the effectiveness of the proposed method.
The shortest path tree problem (SPT) is one of the basic network optimization problems and it is a variance
of the shortest path problem (SPP), which has been widely used in many fields, such as multicast routing,1
Route Information Protocol (RIP),2 wireless network,3 network design,4, 5 IS-IS,6 and complex networks.7–9 Its
objective is to find the set of edges which connect all the nodes in the network so that the sum of the edge
lengths from the source to the other nodes is minimized. As the SPT problem is frequently as subproblem when
solving many combinatorial and network optimization problems, many researchers payed their attention to this
problem.10, 11
In practical environment, the network is changing with time. It lead to the occurence of the variance of
SPTs named Dynamic Shortest Path (DSP) problem. Assume
G (V,
E, ω) be a simple network and the edge
′
′
weights in this network are nonnegative numbers. Let G V, E, ω be another network obtained from G in
′
′
which some edge weights change. Suppose Ts and Ts are the SPTs rooted at s in G and G . The DSP problem is
′
to compute Ts from Ts . Many methods have been proposed to deal with this problem. For example, the classical
Dijkstra algorithm12 and Bellman algorithm solve this problem through recalculating the SPTs whenever there is
a change to edge weights. Due to the high computational time, it cannot meet the requirement of the emergent
accidents. After the idea of using an SPTs update program which only reconstruct on the affected vertices
proposed by Frigioni13 appeared, many dynamic SPTs approaches have implemented this concept into realworld applications10, 11 to reduce the computational time. They have divided the edge weight updates into three
categories: edge weight increase, edge weight decreases, the mixture of them. For the edge updates belonging
to different categories, various operations are processed. The main feature of these approaches is recognize the
affected vertices, such as the intelligent semidynamic DSP algorithm named BallString in,14 the fully dynamic
algorithm called DynamicSWSF-FP proposed in.15
However, the above algorithms have obvious disadvantages. When the scale of the network becomes very big
or the weights of multiple edges decrease while that of other multiple edges increase, the procedure analyzing the
affected vertices will become very complex. Secondly, from the practical viewpoint, this will cost lots of time.
Especially in recent years, the networks with big scale become more and more. This may still cause long latency
and unnecessary overheads. As a consequence, it is meaningful to explore new methods to handle the SPTs
problem. Recently, a large amoeboid organism, the plasmodium of Physarum polycephalum, has been shown to
be capable of solving many graph theoretical problems,16–19 including finding the shortest path,20–23 network
design,24–26 population migration27 and others.28–32 Moreover, this organism has been shown to be able to form
networks with features comparable to or better than the Tokyo rail network.18 In addition, Baumgarten has
proved the mass of mold will eventually convergence to the shortest path of the network that the mold lies on.17
Inspired by this intelligent organism, a path finding mathematical model has been established.33 To the best of
knowledge, the amoeba model is not used to deal with SPTs by now.
In what follows, based on the amoeba model, an adaptive amoeba appraoch to SPTs in dynamic graphs is
presented. The main characteristic of the propose method is its adaptivity. More specially, the algorithm can
recognize the affected vertices and reconstruct them spontaneously. Those unaffected nodes will not be computed
again. We will introduce how to implement amoeba model to deal with SPTs in the following sections.
Results
System Environment and Data Sets. In order to evaluate the performance of the proposed method, we introduce the experimental environment and the problem instance generator is presented. Besides, the performance
of our algorithm is compared with the Label Setting algorithm34 and Bellman-Ford algorithm.35
The proposed adaptive amoeba algorithm for shortest path tree in dynamic graphs is tested on networks
with random and varying topologies through computer simulations using Matlab on an Intel Pentium Dual-Core
E5700 processor (3.00 GHz) with 2 GB of RAM under Windows Seven. The random directed graphs can be
generated using the erdos.renyi.game function of the igraph package in R language (for details, please refer
to http://igraph.sourceforge.net/doc/R/erdos.renyi.game.html). The weight for an edge is randomly generated
ranging from 1 to 1000. The data for random graph is shown in Table 1.
Performance Indicators. In each testing graph, the first node in the random generated graph is denoted
as the source node. Then, a set λ of edges is randomly selected to decrease or increase their edge weights. If both
the increased case and the decreased case associated with the edge weights appear in the network, we denote
λ as mixed. In this paper, we pay attention to the CPU runtime for each algorithm. In order to examine the
efficiency of the presented method, the following parameters are taken into consideration.
• Graph size (graphsize). It denotes the size of the network. We will focus on how the CPU runtime changes
with the change of network size.
• Ratio of updated edges (rue). This parameter represents the percentage of updated edges occupied in the
whole network. For instance, in a network with 1000 edges, when 100 edges get their weight updated, then
we will say rue is 0.1.
• Ratio of changed weight (rcw ). This variable reflects the degree of the changed weight which is decreased
or increased from its original value. As for this parameter, in the increased case, the rcw for the edge
weight ranges from 1 to a large number 10. For example, assume the original weight value associated with
the edge is 100. If the parameter rcw is 1, the updated weight for this edge will be 200. In the decreased
case, the rcw for the edge weight ranges from 0 to 0.9.
In order to evaluate the performance of the proposed method, we compare it with modified Dijkstra algorithm.
We will show how the parameter rue affects the investigated algorithms. For the following computational results,
the program is run for 10 times for each instance.
Edge Weight Increases. For the increase case, Fig. 1 shows the influence of rue on the networks with
different sizes when the parameter rue changes from 0.1 to 0.6 (here, the parameter rcw is set a constant value
0.1). As can be seen in Fig. 1, in the increase cases, regardless of the change of the parameter rue, the CPU time
of all the three algorithms remain constant. On the other hand, considering CPU runtime, the Bellman-Ford
algorithm has less computational time than the Label Setting algorithm in the four networks. The proposed
adaptive amoeba algorithm outperforms when compared with the Label Setting algorithm and Bellman-Ford
algorithm. The reason for this phenomenon is due to the change of the edge weights. When the weights
associated with the edge change, the proposed adaptive amoeba algorithm can recognize the affected vertices
and reconstruct them spontaneously. However, for the Label Setting Algorithm and Bellman-Ford Algorithm,
it must reconstruct the whole network, which consumes more time. As a result, there is a big gap between the
CPU runtime of the proposed method and that of Label Setting algorithm and Bellman-Ford algorithm.
Label Setting Algorithm
4 Bellman−Ford Algorithm
Adaptive Amoeba Algorithm
graphsize=500
graphsize=1000
30
3.5
CPU Time/SPT (s)
CPU Time/SPT (s)
25
3
2.5
2
1.5
20
15
10
1
0.5
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
5
0.1
0.6
0.15
0.2
0.25
300
100
250
80
60
40
20
0
0.1
0.35
rue
0.4
0.45
0.5
0.55
0.6
0.45
0.5
0.55
0.6
graphsize=2000
120
CPU Time/SPT (s)
CPU Time/SPT (s)
graphsize=1500
0.3
200
150
100
50
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0.6
0
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
Figure 1. Comparison in edge weight increases when the parameter rue changes from 0.1 to 0.6 (here, the parameter rcw
is set a constant value 0.1)
In what follows, we focus on how the change of the parameter rcw affects the CPU runtime of the different
algorithms in the increase case. As shown in Fig. 2, it shows the influence of the parameter rcw on the three
algorithms’ computational time when the parameter rcw changes from 0.1 to 0.6 (here, the parameter rue is set
a constant value 0.2). As we can see, the parameter rcw has different influences on the three algorithms. For the
Label Setting algorithm and Bellman-Ford algorithm, the computational time remains relatively constant, which
is shown as straight lines in Fig. 2. On the contrary, for the proposed adaptive amoeba algorithm, it is influenced
strongly by the parameter rcw, especially when the graph size is greater than 1500. From Fig. 2, it can be seen
that the CPU runtime of the presented method becomes more and more with the increase of parameter rcw.
The reason lies that when the parameter rcw becomes bigger and bigger, the edge length is increased a lot from
its original value. As a result, most of the paths in the SPT need to be recomputed. Although the CPU time of
the proposed adaptive amoeba algorithm increases a lot, it is still less than that of the Label Setting algorithm
and Bellman-Ford algorithm.
In summary, in the increase case, the change of the parameter rue has little influence on these algorithms. The
adaptive amoeba algorithm proposed in this paper has the least CPU runtime when dealing with the networks
5
graphsize=500
Label Setting Algorithm
Bellman−Ford Algorithm
Adaptive Amoeba Algorithm
graphsize=1000
35
30
CPU Time/SPT (s)
CPU Time/SPT (s)
4
3
2
25
20
15
10
1
5
0
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0
0.1
0.6
0.15
0.2
0.25
300
100
250
80
60
40
20
0
0.1
0.35
rue
0.4
0.45
0.5
0.55
0.6
0.45
0.5
0.55
0.6
graphsize=2000
120
CPU Time/SPT (s)
CPU Time/SPT (s)
graphsize=1500
0.3
200
150
100
50
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0.6
0
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
Figure 2. Comparison in edge weight increases when the parameter rcw changes from 0.1 to 0.6 (here, the parameter rue
is set a constant value 0.2)
with different sizes. On the contrary, the change of the parameter rcw has great influence on the CPU runtime of
the proposed method. The CPU runtime of the presented method becomes bigger with the increase of parameter
rcw.
Edge Weight Decreases. Similarly, in the decrease case, we have observed the influence of the two
parameters on the CPU runtime of the above three algorithms. Figs. 3 and 4 show the experimental results,
when the parameters rue, rcw change respectively.
As can be seen in Fig. 3, the parameters are set the same as that of the edge weight increases. From Fig. 3, it
can be concluded that, in the decrease case, when the parameter rue changes from 0.1 to 0.6, the CPU runtime
of the three algorithms fluctuates slightly in the networks with different sizes. In other words, the parameter
rcw has very little effect on the computational efficiency of these algorithms. The proposed method has obvious
advantage over the Label Setting algorithm and Bellman-Ford algorithm when dealing with the SPT in dynamic
graphs.
As shown in Fig. 4, when the parameter rcw changes from 0.1 to 0.6, the CPU runtime of the proposed
method fluctuates strongly. When rcw increases from 0.1 to 0.6, the CPU runtime for the adaptive amoeba
algorithm deteriorates rapidly. In the network with graph size equal to 500, 1500, 2000, when rcw reaches 0.6,
it can be observed that the performance of the Label Setting algorithm and Bellman-Ford algorithm outperform
compared to the proposed method. It can be observed that when rcw is less than a certain threshold value, the
proposed method has better performance. The threshold value varies with the size of a graph. When rcw is
bigger than 0.4, it is more appropriate to adopt the Label Setting algorithm and Bellman-Ford algorithm. The
reason why the performance of the proposed algorithm gets worse when rcw is bigger than the threshold value is
that more edges are influenced, and more time is spent to reconstruct the SPT and reallocate the flux associated
with each edge in the amoeba algorithm.
In summary, in the decreased case, the situation is similar to that of the increased case. The parameter
rcw has different effect with that of parameter rue. Besides, the performance of the adaptive amoeba algorithm
5
Label Setting Algorithm
Bellman−Ford Algorithm
Adaptive Amoeba Algorithm
graphsize=500
graphsize=1000
35
30
CPU Time/SPT (s)
CPU Time/SPT (s)
4
3
2
25
20
15
10
1
5
0
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0
0.1
0.6
0.15
0.2
0.25
300
100
250
80
60
40
0.4
0.45
0.5
0.55
0.6
0.45
0.5
0.55
0.6
200
150
100
20
0
0.1
0.35
rue
graphsize=2000
120
CPU Time/SPT (s)
CPU Time/SPT (s)
graphsize=1500
0.3
50
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0
0.1
0.6
0.15
0.2
0.25
0.3
0.35
rue
0.4
Figure 3. Comparison in edge weight decreases when the parameter rue changes from 0.1 to 0.6 (here, the parameter rcw
is set a constant value 0.1)
5
graphsize=500
Label Setting Algorithm
Bellman−Ford Algorithm
Adaptive Amoeba Algorithm
graphsize=1000
35
30
CPU Time/SPT (s)
CPU Time/SPT (s)
4
3
2
25
20
15
10
1
5
0
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0
0.1
0.6
0.15
0.2
0.25
graphsize=1500
0.4
0.45
0.5
0.55
0.6
0.45
0.5
0.55
0.6
1200
1000
CPU Time/SPT (s)
200
CPU Time/SPT (s)
0.35
rue
graphsize=2000
250
150
100
50
0
0.1
0.3
800
600
400
200
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0.6
0
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
Figure 4. Comparison in edge weight increases when the parameter rcw changes from 0.1 to 0.6 (here, the parameter rue
is set a constant value 0.2)
gets worse when rcw is more than a threshold value. As a consequence, it is appropriate to carry out different
algorithms according to the specific value of this parameter.
Mixed Edge Weight Changes. In the mixed case, both the increased case and the decreased case share
the same percentage in the total ratio of updated edges. Simply speaking, suppose there are 200 edges in the
network whose weights change, then 100 of them increase their edge weights while the left 100 edges’ weights
decrease. As for the parameter rcw, it is implemented for both the increased case and the decreased case. It is
to say suppose rcw is 0.1, then half of the randomly selected edges will increase their weights by 10 percentages
while the left edges will decrease their weights by 10 percentages.
graphsize=500
5
graphsize=1000
Label Setting Algorithm
Bellman−Ford Algorithm
Adaptive Amoeba Algorithm
30
25
CPU Time/SPT (s)
CPU Time/SPT (s)
4
3
2
1
0
0.1
20
15
10
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
5
0.1
0.6
0.15
0.2
0.25
graphsize=1500
0.3
0.35
rue
0.4
0.45
0.5
0.55
0.6
0.45
0.5
0.55
0.6
graphsize=2000
300
100
250
80
CPU Time/SPT (s)
CPU Time/SPT (s)
90
70
60
50
200
150
100
40
50
30
20
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0.6
0
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
Figure 5. Comparison in mixed change of the edge weight when the parameter rue changes from 0.1 to 0.6 (here, the
parameter rcw is set a constant value 0.1)
Fig. 5 shows the CPU runtime of the three algorithms when the parameter rue changes from 0.1 to 0.6.
Similarly, it has slight influence like in the increased and decreased cases. Obviously, the adaptive amoeba
algorithm outperforms when compared with the other algorithms.
As for the change of parameter rcw, Fig. 6 displays the computational results of the three algorithms. It can
be concluded that the Label Setting algorithm and Bellman-Ford algorithm remain constant, regardless of the
changed weights. For the adaptive amoeba algorithm, it fluctuates greatly when rcw changes. As can be seen in
Fig. 6, when the rcw is less than 0.5, the proposed method outperforms certainly than the other two algorithms.
When rcw reaches 0.6 and the graph size reaches 2000, the CPU runtime of the presented algorithm has spent
more time than the other two algorithms. In the four sub figures, the results are not consistent. It is due to
the difference of the network topology. Every time, the affected edges are randomly selected, which may lead to
different influences on the network. In one word, similar to the increased and decreased case, the parameter rue
affect the CPU runtime a little while rcw playing an important role regarding the CPU runtime.
Discussion
In this paper, we investigated the previous algorithms used to deal with SPT problem in dynamic graphs.
These algorithms solving SPT problem with edge updates by identifying the affected nodes and reconstruct the
5
Label Setting Algorithm
Bellman−Ford Algorithm
Adaptive Amoeba Algorithm
graphsize=500
graphsize=1000
35
30
CPU Time/SPT (s)
CPU Time/SPT (s)
4
3
2
25
20
15
10
1
5
0
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0
0.1
0.6
0.15
0.2
0.25
graphsize=1500
0.3
0.35
rue
0.4
0.45
0.5
0.55
0.6
0.45
0.5
0.55
0.6
graphsize=2000
120
220
200
100
CPU Time/SPT (s)
CPU Time/SPT (s)
180
80
60
40
160
140
120
100
80
20
60
0
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
0.45
0.5
0.55
0.6
40
0.1
0.15
0.2
0.25
0.3
0.35
rue
0.4
Figure 6. Comparison in mixed change of the edge weight when the parameter rcw changes from 0.1 to 0.6 (here, the
parameter rue is set a constant value 0.2)
shortest path among these nodes. However, when these algorithms are faced with the network with big scale,
on the one hand, this procedure becomes very complicated. On the other hand, it cost much time. In order
to address the above problems, we proposed a fully adaptive amoeba algorithm for solving SPT in dynamic
graphs. The efficiency of the presented algorithm is demonstrated by implementing it in all the edge updates
including the increase, decrease, mixed change of the edge weight. In order to evaluate the performance of the
proposed method, we conducted experiments on randomly generated graphs, in terms of the CPU execution time.
Moreover, we compared our algorithm with the existing algorithms, such as the Label Setting algorithm, BellmanFord algorithm. The purpose of the experiment is to obverse how these algorithms behave for different graph sizes
and various mixes of changed edges. We randomly generate four graphs with different sizes: 500, 1000, 1500, 2000.
We evaluate the performance of the three algorithms according to three factors: graphsize, ratio of updated
edges (rue), ratio of changed weight (rcw ). For the increased case, the change of the parameter rue has little
influence on these algorithms while the change of the parameter rcw has great influence on the CPU runtime of
the proposed method. For the decreased case, the influence of parameter rue is similar to that of the increased
case. On the other hand, the performance of the adaptive amoeba algorithm gets worse when rcw is more than
a threshold value. As for the other left algorithms, they are slightly influenced by parameter rcw. For the mixed
change of the edge weights, the influences of parameters rue and rcw on the computational time are consistent
with that of the increased and decreased cases. We conclude the following for the above three algorithms. For
the increased case, decreased case, and mixed change case, in spite of the ratio of updated edges, the proposed
method has the best overall performance. On the contrary, when the parameter rcw changes, it is appropriate
to carry out different algorithms according to the specific value of rcw.
Methods
Physarum Polycephalum Inspired Shortest Path Finding Model. Physarum Polycephalum is a large,
single-celled amoeboid organism forming a dynamic tubular network connecting the discovered food sources
during foraging. The mechanism of tube formation can be described as: tubes thicken in a given direction when
shuttle streaming of the protoplasm persists in that direction for a certain time. It implies positive feedback
between flux and tube thickness, as the conductance of the sol is greater in a thicker channel. With this
mechanism, a mathematical model illustrating the shortest path finding has been constructed.33
Suppose the shape of the network formed by the Physarum is represented by a graph, in which a plasmodial
tube refers to an edge of the graph and a junction between tubes refers to a node. Two special nodes labeled
as N1 , N2 act as the starting node and ending node respectively. The other nodes are labeled as N3 , N4 , N5 , N6
etc. The edge between node Ni and Nj is expressed as Mij . The parameter Qij denotes the flux through tube
Mij from node Ni to Nj . Regard the flow along the tube as an approximately poiseuille flow, the flux Qij can
be expressed as:
Qij =
Dij
(pi − pj )
Lij
(1)
where pi is the pressure at the node Ni , Dij is the conductivity of the tube Mij , Lij is its length.
By considering that the inflow and outflow must be balanced, we have:
X
Qij = 0(j 6= 1, 2)
For the source node N1 and the sink node N2 the following two equations hold
X
Qi1 + I0 = 0
(2)
(3)
i
X
Qi2 − I0 = 0
(4)
i
where I0 is the flux flowing from the source node and I0 is a constant value here.
In order to describe such an adaptation of tubular thickness we assume that the conductivity Dij changes
over time according to the flux Qij . The following equation for the evolution of Dij (t) can be used
d
Dij = f (|Qij |) − rDij
dt
(5)
where r is a decay rate of the tube. It can be obtained that the equation implies that the conductivity ends
to vanish if there is no flux along the edge, while it is enhanced by the flux. The f is monotonically increasing
continuous function satisfying f (0) = 0.
Then the network poisson equation for the pressure can be obtained from the Eq. (8-4) as follows:
f or
j = 1,
+1
X Dij
−1
f or
j = 2,
(pi − pj ) =
Lij
i
0 otherwise
(6)
By setting p2 =0 as a basic pressure level, all pi can be determined by solving Eq. (6) and Qij can also be
obtained.
In this paper, f (Q) = |Q| is used. With the flux calculated, the conductivity can be derived, where Eq. (7)
is used instead of Eq. (5), adopting the functional form f (Q) = |Q|.
n+1
n
Dij
− Dij
n+1
= |Q| − Dij
δt
(7)
The amoeba model is modified to solve SPTs in dynamic graphs. The algorithm is composed of three parts.
First of all, the amoeba model is extended to solve the shortest path in the directed networks. Then a modified
model is presented to handle SPTs in static directed networks. Finally, we analyze the changing trend of the
amoeba algorithm is applied to deal with SPTs in dynamic graphs.
Shortest Path Problem in Directed Network. It is observed that the original amoeba model can only
handle the shortest path problem (SPP) in the undirected network according to.16, 20, 33 As a result, we need to
extend its application field to directed networks by modifying the original model.
Let G = (N, E, L) be a directed network, where N denotes a set of n nodes, E denotes an edge set with m
directed edges, and L denotes a weight set for E. Given a source node s and a sink node t, the directed shortest
path problem can be defined as how to find a path from s to t, which only consists of directed edges of E, with
the minimum sum of weights on the edges.
The original amoeba model is designed to solve the shortest path problem in the undirected graphs. For the
directed graphs, the following equations are constructed according to the Kirchhoff’s laws:
+1 f or i = s
X Dij
Dji
−1 f or i = t
(pi − pj ) =
(8)
+
Lij
Lji
j∈N
0
otherwise
where s denotes the starting node, t denotes the ending node, Lij is the length of the edge Mij , Dij is the
conductivity of the edge Mij . In the directed graphs, Mij is different from that of the undirected networks.
Here, Mij denotes the tube starting from node i to node j. As a result, the way that Dij is initialized in the
directed graphs is different from that in the undirected networks. If Mij ∈ E, then Dij = 1. Otherwise, Dij = 0.
In what follows, the flux Qij of every edge can be obtained. The direction of the edge is related with the
pressure of each node. Naturally, the flux starts at the node with high pressure and ends at the node with low
pressure. In the directed graphs, in order to keep the direction of each edge, the following check procedure is
inevitable. Assume there is an edge Mij starting from node i and ending at node j. If the pressure pj is larger
than pi , it means the flux is flowing from node j to node i, which is opposite with the direction of the edge in the
directed graph. Once the phenomenon is found, then the flux needs to be cut off. As a consequence, we change
the flux of this edge Qij to be 0. Next, the following Eq. (9) for the evolution of Dij (t) can be summarized. As
can be seen, the changing process of D is continuous. In the process of the algorithm, a discrete procedure is
applied as shown in Eq. (10). The general flow of the proposed method is detailed in Table 2.
dDij (t)
=
dt
Dij (n + 1) =
Qij (t) − Dij (t)
pi (t) ≥ pj (t)
−Dij (t)
pi (t) < pj (t)
(Qij (n) − Dij (n)) ∗ ∆t + Dij (n) pi (n) ≥ pj (n)
(0 − Dij (n)) ∗ ∆t + Dij (n)
pi (n) < pj (n)
(9)
(10)
where ∆t is the time interval, 0 < ∆t < 1.
There are several possible solutions to decide when to stop execution of Algorithm 1, such as the maximum
number of iterations is arrived, conductivity of each tube converges to 0 or 1, flux through each tube remains
unchanged, etc. In this paper, when the conductivity matrix changes between the current iteration and the
N P
N
P
n−1
n
≤ δ, where δ is a threshold), then the program
Dij
− Dij
previous iteration very little (in other words,
i=1 j=1
will stop.
Proofs of Physarum Polycephalum Model in Directed Networks. In undirected graphs, the model
is able to converge to an equilibrium state and Baumgarten17 has proved its stability . In directed graphs, a
lot of our experiments have shown that, it will also convergence to an equilibrium state in directed networks
and the process of the convergence is similar to that of the undirected graphs. In this section, we prove that all
the flux flowing from s to t consists of the shortest path in the directed networks when the network reaches the
equilibrium state.
First of all, the following parameters are defined:
F (t): F (t) is one set of edges whose Qij is non-negative number at time t. Let F = lim F (t).
t→∞
B (t): B (t) is another set of edges whose Qij is is negative number at time t. Let B = lim B (t).
t→∞
uij (t): uij (t) is the pressure difference between node i and node j at time t. uij (t) = pi (t) − pj (t). Let
uij = lim uij (t).
t→∞
lemma 3.1. For any directed edge Mij , lim (Qij (t) − Dij (t)) = 0.
t→∞
Proof When t goes to a infinite number, the network reaches stable. As a result, for any Mij , lim
t→∞
According to Eq. (9), it can be seen that:
(
lim (Qij (t) − Dij (t)) = 0
t→∞
lim Dij (t) = 0
t→∞
When j ∈ B, lim Qij (t) = lim
t→∞
t→∞
Dij (t)uij (t)
Lij
dDij (t)
dt
= 0.
Mij ∈ F
Mij ∈ B
= 0.
Lemma 3.1 is established.
lemma 3.2. For any Mij , if lim Qij (t) 6= 0, then Mij ∈ F and ue = Le .
t→∞
Proof According to lemma 3.1, if lim Qij (t) 6= 0, then lim Dij (t) = lim Qij (t) 6= 0.
t→∞
t→∞
t→∞
Based on Eq. (8), it can be seen that uij = Lij > 0 and Mij ∈ F .
Lemma 3.2 is established.
lemma 3.3. If Mij ∈ F , then uij ≤ Lij .
Proof According to lemma 3.2, when lim Qij (t) 6= 0, lemma 3.3 is established.
t→∞
When lim Qij (t) = 0 and Mij ∈ F , then lim Dij (t) = 0, ∃T, ∀t > T,
t→∞
t→∞
dDij (t)
dt
< 0. Thus, Mij ∈ F (t).
According to Eq. (9), we can obtain Qij (t) < Dij (t).
According to Eq. (8), it can be seen uij (t) < Lij .
In summary, uij ≤ Lij . Therefore, lemma 3.3 is established.
lemma 3.4. When the model reaches the equilibrium state:
(i) All the flow converges to some paths from s to t.
(ii) These directed paths have the same path length and their length is equal to ust .
(iii) These directed paths have the shortest path length.
Proof
(i) According to lemma 3.1 and lemma 3.2, in the equilibrium state, the directions of all the edges are in
accordance with the flow directions. Consequently, the flow converges to the paths existing in the directed
graph G.
(ii) Assume v is one directed
path
Pwhich the flow converges to, Lv is the length of v. According to lemma 3.2,
P
Lij = Lv .
uij =
there are ust =
Mij ∈v
Mij ∈v
P
(iii) Assume v is one path from s to t, it is known to us that ust =
uij . For any edge Mij , if uij ≥ 0,
Mij ∈v
then Mij ∈ F . According
to lemma
P 3.3, uij ≤ Lij . If uij < 0, it is obviously seen that uij ≤ Lij .
P
Lij = Lv . It means that all the other paths’ length is not less than
uij ≤
Consequently, ust =
Mij ∈v
Mij ∈v
ust . Incorporating with (ii), these paths are the shortest ones.
Based on the above proofs, lemma 3.4 is established.
Shortest Path Tree Problem in Directed Networks. The above algorithm can only compute the
shortest path between two nodes at a time. Assume there are N nodes in the network, if we want to construct
the shortest path tree, the algorithm must be run for N − 1 times. However, this will consume lots of time. In
this section, after we modify the above model further, the shortest path tree can be constructed by running the
algorithm one time. In what follows, the modified model used to find the shortest path tree is introduced.
In the original amoeba model, there are only one starting node s and one ending node t in it. If there are one
starting node s and all the other nodes are the ending nodes, the Kirchhoffs laws can be transformed as below:
X Dij
Lij
i
Dji
+
Lji
(pi − pj ) =
+1
−1
N −1
f or j = s
f or j 6= s
(11)
The general flow of the modified model used to construct SPTs rooted at node s are shown in Table 3. By
this way, SPT can be constructed through running the program once.
K
G
E
F
H
I
S
R
E
Q
L
F
H
I
O
N
J
(a) An SPT Ts rooted at node s. In this figure, the vertex
are represented by letters and the numbers along the arcs
denote the path length. The arrow from the tail to the
head means the direction of the edge.
M
V
N
J
O
G
P
K
V
P
S
M
R
Q
L
(b) The SPTs rooted at node s constructed by the adaptive
amoeba algorithm for shortest path tree
Figure 7. The amoeba algorithm for SPT
Example 0.1. In order to illustrate the algorithm shown in Table 3, an example is shown in Fig. 7(a). There
are 16 nodes in this network. In this example, the amoeba model is applied to construct SPT rooted at node
s. According to Algorithm 2, first of all, initialize the parameters, such as the length matrix L, the initial
conductivity matrix D, the initial pressure matrix Q etc. Then, according to Eq. (11), the pressure of each node
can be obtained during the first iteration. In turn, the flux of each node can be computed according to Eq. (8).
Next, the conductivity matrix during the following iteration can be constructed based on Eq. (7). This procedure
will continue until the flux of each arc do not change any more. Fig. 8 shows the flux variation of each edge
during different iterations in the graph shown in Fig. 7(a) using the adaptive amoeba model. As we can see, the
flux of some edges converges to 0 during the iterations. Those edges whose flux are not equal to zero constitutes
the following graph shown in Fig. 7(b).
1
0.9
0.8
0.7
Flux
Edge(c,g)
Edge(S,c)
0.6
Edge(S,b)
0.5
Edge(S,d)
Edge(d,h)
Edge(g,j)
Edge(g,k) Edge(b,f)
0.4
0.3
0.2
0.1
0
0
10
20
30
40
50
60
Iteration
Figure 8. Flux variation during different iterations in the graph shown in Fig. 7(a) using the adaptive amoeba model.
Fig. 7(b) shows the SPTs rooted at vertex s. The trees represented by the edges in blue color are consistent
with the result in.10 The edges with red color appear in the adaptive amoeba algorithm, but not in.10 The reason
lies that there are more than one shortest paths between the rooted node s and nodes e, k. For example, the length
of the path S → b → e is equal to that of the path S → b → f → e. Similarly, the length of path S → d → h → k
is equal to that of S → c → g → k. This is the first advantage of our presented method.
The edge weight change can be divided into three categories: increase only, decrease only, the mixture of
them. In this section, the adaptive amoeba algorithm for SPTs in dynamic graphs will be analyzed from the
three perspectives.
Edge Weight Increase. As shown in Fig. 9(a), the weights of edges (c, g) and (g, j) are increased. For
traditional approaches to SPT in dynamic graphs, such as DynDijkInc, MBallString 10 etc, firstly, these methods
locates all locally affected vertices and reconstruct the shortest path between these affected nodes. In this
example, the vertices g, k, o, p, i, n are affected, which are shown in blue circles in Fig. 9(a). For the above
methods, when the scale of the network become very big, on the one hand, the process to locate the affected
nodes become very complex. On the other hand, it costs much time. For the adaptive amoeba algorithm proposed
in this paper, the following Fig. 10 displays the changing trend of the flux existing in each edge.
P
K
G
V
E
F
H
I
R
E
Q
F
H
I
O
N
J
L
(a) An SPT rooted at node S. In this example, the length
of edge (c, g) is increased by 2 and the length of edge (g,
j) is increased by 3. The changed edges are denoted by
red dotted arrows.
S
M
V
N
J
O
G
P
K
S
M
R
Q
L
(b) The SPT after the increase of the edge weights.
Figure 9. The SPT with edge weight increase
As can be seen in Fig. 10, the red doted line divides the whole process into two parts. The first part is the
changing trend of the flux associated with each edge before the edge weights change. The second part illustrates
how the flux associated with every edge change after the increase of edge weights. During the iteration process,
1
0.9
0.8
0.7
Edge(S,c)
Edge(S,b)
Flux
0.6
Edge(d,h)
Edge(c,g)
Edge(g,k)
Edge(g,j)
0.5
Edge(f,j)
Edge(b,f)
Edge(S,d)
0.4
0.3
0.2
0.1
0
0
50
100
150
Iteration
Part 1
200
250
Part 2
Figure 10. The changing trend of the flux when the length of edge (c, g) is increased by 2 and the length of edge (g, j) is
increased by 3
the flux of the edge (g, j) decreases to 0 while that of the edge (f, j) become one of the edges constructing
SPT. This example shows the adaptivity of the presented algorithm. It can recognize the affected vertices and
reconstruct them spontaneously after the increase of the edge weights. Fig. 9(b) displays the SPT after the
increase of the edge weights and the result is consistent with that of.10
Edge Weight Decrease. Given a graph shown in Fig. 11(a), there are a source vertex S, an SPT rooted
at node S. The weights of edge (c, g) and edge (g, j) are decreased by 3, 1 respectively. Different from the
increase case, the locally affected nodes can be predicted. In the decrease case, the traditional algorithms such
as DynDijkstra, MBallString recognize the all affected heads and then recompute the shortest path among the
affected heads. When the network becomes very big, the traditional algorithms are faced with the same problems
mentioned in the increase case. For the adaptive amoeba algorithm, the following Fig. 12 illustrates the changing
trend of the flux of each edge when the edge weights decrease.
P
K
G
V
E
F
H
I
R
E
Q
S
M
V
N
J
O
G
P
K
F
H
I
L
(a) An SPT rooted at node S. In this example, the length
of edge (c, g) is decreased by 3 and the length of edge (g,
j) is decreased by 1. The changed edges are denoted by
red dotted arrows.
N
O
J
S
M
R
Q
L
(b) The SPT after the decrease of the edge weights.
Figure 11. The SPT with edge weight decrease
As can be seen in Fig. 12, the whole process can be separated into two procedures by the red dotted line: the
part before edge weight decrease, the part after edge weight decrease. It can be seen that the flux associated with
every edge changes when the edge weights decrease. All the edges can be divided into 2 categories: the edges
which is not influenced by the edge weights decrease, such as edges (k, p) and (b, e) shown in Fig. 12 belongs to
the first category; the second category includes the edges that are influence by the edge weights decrease, such
as edges (h, k), (S, d), (S, c) etc. From Fig. 12, it can be concluded that the adaptive amoeba algorithm have the
1
0.9
0.8
Edge(S,b)
0.7
Flux
Edge(c,g)
Edge(b,f)
Edge(h,k)
0.6
Edge(S,c)
Edge(g,k)
Edge(k,p) Edge(b,e)
Edge(d,h)
Edge(g,j)
0.5
Edge(S,d)
0.4
0.3
0.2
0.1
0
0
20
40
60
80
Iteration
Part 1
100
120
140
Part 2
Figure 12. The changing trend of the flux when the length of edge (c, g) is decreased by 3 and the length of edge (g, j) is
decreased by 1
advantage of recognizing all the affected vertices spontaneously over traditional algorithms such as DynDijkstra,
MBallString. Fig. 11(b) displays SPT after the edge weights decrease and the result is the same as that of.10
Mixed Edge Weight Changes. In Fig. 13(a), the weights of the following edges updates: edge (c, g) is
decreased by 1, the weight of edge (g, j) is increased by 3, and the weight of edge (f, i) is decreased by 8. Based
on this example, we will pay attention to the changing trend of the flux associated with each edge in the adaptive
amoeba algorithm when both the decrease and increase of the edges appear in the graph.
P
K
G
V
E
F
H
I
R
E
Q
L
(a) An SPT rooted at node S. In this example, the weight
of edge (c, g) is decreased by 1, the weight of edge (g, j)
is increased by 3, and the weight of edge (f, i) is
decreased by 8. The changed edges are denoted by red
dotted arrows.
S
M
V
N
J
O
G
P
K
F
N
H
I
O
J
S
M
R
Q
L
(b) The SPT after the update of the edge weights.
Figure 13. The SPT with mixed edge weight changes
As can be seen in Fig. 14, the changing trend of the flux associated with every edge is shown. Several edges
such as edge (h, k) disappear from the SPT while edges (f, i) and (i, n) become the elements constructing SPT
after the update of edge weight. The other edges such as edges (k, p) and (b, e) are not affected by the update
of the edge weight. Fig. 13(b) gives the final SPT after the update of the edge weights. From Fig. 13(b), it can
be concluded that the amoeba algorithm can distinguish all the nodes adaptively.
In summary, all the edge updates including the increase, decrease, mixed change of the edge weight can be
handled efficiently by the adaptive amoeba algorithm.
1
0.9
0.8
0.7 Edge(S,b)
Flux
0.6
0.5
Edge(S,c)
Edge(g,j)
Edge(b,f)
Edge(k,p)
Edge(g,k)
Edge(d,h)
Edge(h,k)
Edge(c,g)
Edge(S,d)
Edge(j,n)
Edge(b,e)
Edge(i,n)
0.4
0.3
0.2
0.1
0
0
20
40
Part 1
60
80
100
120
Iteration
140
160
180
Part 2
Figure 14. The changing trend of the flux associated with each edge when the length of edge (c, g) is decreased by 1, the
length of edge (g, j) is increased by 3, and the weight of edge (f, i) is decreased by 810
Acknowledgment
All the authors of the cited papers for providing their network data. The work is partially supported by
Chongqing Natural Science Foundation, Grant No. CSCT, 2010BA2003, National Natural Science Foundation
of China, Grant No. 61174022, 61364030, National High Technology Research and Development Program of
China (863 Program) (No.2013AA013801), the Fundamental Research Funds for the Central Universities, Grant
No. XDJK2012D009.
1. Bauer, F., Varma, A. Distributed algorithms for multicast path setup in data networks. IEEE/ACM Trans.
Networking 4, 181–191 (1996).
2. Rescigno, A. Optimally balanced spanning tree of the star network. IEEE Trans. Comput. 50, 88–91
(2001).
3. Zhao, M., Yang, Y. Bounded relay hop mobile data gathering in wireless sensor networks. IEEE Trans.
Comput. 61, 265–277 (2012).
4. Wang, P., Hunter, T., Bayen A.M., Schechtner, K., González M.C. Understanding road usage patterns in
urban areas. Sci. Rep. 2 (2012).
5. Barthelemy, M., Bordin, P., Berestycki, H., Gribaudi, M. Self-organization versus top-down planning in
the evolution of a city. Sci. Rep. 3 (2013).
6. Perlman, R. A comparison between two routing protocols: OSPF and IS-IS. IEEE Network 5, 18–24 (1991).
7. Estrada, E., Vargas-Estrada, E. How peer pressure shapes consensus, leadership, and innovations in social
groups. Sci. Rep. 3 (2013).
8. Wei, D.J. et al. Box-covering algorithm for fractal dimension of weighted networks. Sci. Rep. 3 (2013).
9. Szell, M., Sinatra, R., Petri, G., Thurner, S., Latora, V. Understanding mobility in a social petri dish. Sci.
Rep. 2 (2012).
10. Chan, E., Yang, Y. Shortest path tree computation in dynamic graphs. IEEE Trans. Comput. 58, 541–557
(2009)
11. Chen, H., Tseng, P. A low complexity shortest path tree restoration scheme for IP networks. IEEE
Commun. Lett. 14, 566–568 (2010).
12. Dijkstra, E. A note on two problems in connection with graphs. Numerische Mathematik 1, 269C271
(1959).
13. Frigioni, D., Marchetti-Spaccamela, A., Nanni, U. Incremental algorithms for the single-source shortest
path problem. Foundation of Software Technology and Theoretical Computer Science 14, 113–124 (1994).
14. Narváez, P., Siu, K., Tzeng, H. New dynamic spt algorithm based on a ball-and-string model. IEEE/ACM
Trans. Networking 9, 706–718 (2001).
15. Nguyen, S., Pallottino, S., Scutella, M.G. A new dual algorithm for shortest path reoptimization. Transportation and Network Analysis: Current Trends: Miscellanea in honor of Michael Florian 63, 221–221
(2002).
16. Tero, A., Kobayashi, R., Nakagaki, T. Physarum solver: A biologically inspired method of road-network
navigation. Physica A 363, 115–119 (2006).
17. Baumgarten, W., Ueda, T., Hauser, M. Plasmodial vein networks of the slime mold physarum polycephalum
form regular graphs. Phys. Rev. E 82, 046113 (2010).
18. Tero, A. et al. Rules for biologically inspired adaptive network design. Science Signalling 327, 439 (2010).
19. Watanabe, S., Tero, A., Takamatsu, A., Nakagaki, T. Traffic optimization in railroad networks using an
algorithm mimicking an amoeba-like organism, physarum plasmodium. BioSystems 105, 225–232 (2011).
20. Nakagaki, T. et al. Minimum-risk path finding by an adaptive amoebal network. Phys. Rev. Lett. 99,
068104 (2007).
21. Adamatzky, A., Prokopenko, M. Slime mould evaluation of australian motorways. Int. J. Parallel Emergent
Distrib. Syst. 27, 275–295 (2012).
22. Zhang, X. et al. Solving 0-1 knapsack problems based on amoeboid organism algorithm. Appl. Math.
Comput. 219, 9959–9970 (2013)
23. Zhang, X., Zhang, Z., Zhang, Y., Wei, D., Deng, Y. Route selection for emergency logistics management:
A bio-inspired algorithm. Saf. Sci. 54, 87–91.
24. Gunji, Y.P., Shirakawa, T., Niizato, T., Haruna, T. Minimal model of a cell connecting amoebic motion
and adaptive transport networks. J. Theor. Biol. 253, 659–667 (2008).
25. Adamatzky, A., Martı́nez, G.J., Chapa-Vergara, S.V., Asomoza-Palacio, R., Stephens, C.R. Approximating
mexican highways with slime mould. Natural Computing 10, 1195–1214 (2011).
26. Adamatzky, A., de Oliveira, P.P. Brazilian highways from slime molds point of view. Kybernetes 40,
1373–1394 (2011).
27. Adamatzky, A. The worlds colonization and trade routes formation as imitated by slime mould. Int. J.
Bifurcation Chaos 22 (2012).
28. Adamatzky, A. Growing spanning trees in plasmodium machines. Kybernetes 37, 258–264 (2008).
29. Aono, M., Hara, M., Aihara, K., Munakata, T. Amoeba-based emergent computing: combinatorial optimization and autonomous meta-problem solving. International Journal of Unconventional Computing 6,
89–108 (2010).
30. Jones, J. Characteristics of pattern formation and evolution in approximations of physarum transport
networks. Artificial Life 16, 127–153 (2010).
31. Aono, M., Zhu, L., Hara, M. Amoeba-based neurocomputing for 8-city traveling salesman problem. International Journal of Unconventional Computing 7, 463–480 (2011).
32. Shirakawa, T., Yokoyama, K., Yamachiyo, M., Gunji, Y.P., Miyake, Y. Multi-scaled adaptability in motility
and pattern formation of the physarum plasmodium. International Journal of Bio-Inspired Computation
4, 131–138 (2012).
33. Tero, A., Kobayashi, R., Nakagaki, T. A mathematical model for adaptive transport network in path
finding by true slime mold. J. Theor. Biol. 244, 553–564 (2007).
34. Meyer, U. Average-case complexity of single-source shortest-paths algorithms: lower and upper bounds.
Journal of Algorithms 48, 91–134 (2003).
35. Nguyen, U.T., Xu, J. Multicast routing in wireless mesh networks: Minimum cost trees or shortest path
trees? IEEE Commun. Mag. 45, 72–77 (2007).
Table 1. The related parameters of the erdos.renyi.game function. In this function, the parameter p denotes the probability
for drawing an edge between two arbitrary vertices.
Dataset
dataset
dataset
dataset
dataset
1
2
3
4
Number of Nodes
p
Number of Edges
500
1000
1500
2000
0.0200
0.0100
0.0060
0.0050
5000
9921
13475
19903
Table 2. Adaptive Amoeba Algorithm(L, V, E) for Directed Networks
// L is an N × N matrix, Lij denotes the length between node i and
node j
// V denote the set of nodes, E denotes the set of edges
// s is the starting node, t is the ending node
Dij ← (0, 1] (∀i, j = 1, 2, . . . , N ∧ Lij 6= 0)
Qij ← 0 (∀i, j = 1, 2, . . . , N )
pi ← 0 (∀i = 1, 2, . . . , N )
count ← 1
repeat
pt ← 0 // the pressure at the ending node t
Calculate the pressure of every node using Eq. (8)
+1 f or j = s
X Dij
Dji
−1 f or j = t
(pi − pj ) =
+
Lij
Lji
i
0
otherwise
Qij ← Dij × (pi − pj )/Lij // Using Eq. (8)
if Qij < 0 then
Qij = 0
end if
Dij ← Qij + Dij // Using Eq. (7)
count ← count + 1
until a termination criterion is met
Author contributions
X. Z., Q. L., Z. Z. designed and performed research. X. Z. wrote the paper. Y. H. performed the computation. S.
M., F. T.S. C. and Y. D. analyzed the data. All authors discussed the results and commented on the manuscript.
Additional information
Competing financial interests: The authors declare no competing financial interests.
Table 3. Adaptive Amoeba Algorithm(L, V, E) for Shortest Path Tree
// L is an N × N matrix, Lij denotes the length between node i and
node j
// V denote the set of nodes, E denotes the set of edges
// s is the root node
Dij ← (0, 1] (∀i, j = 1, 2, . . . , N ∧ Lij 6= 0)
Qij ← 0 (∀i, j = 1, 2, . . . , N )
pi ← 0 (∀i = 1, 2, . . . , N )
count ← 1
repeat
Calculate the pressure of every node using Eq. (11)
X Dij
+1 f or j = s
Dji
(pi − pj ) =
+
−1
f or j 6= s
Lij
Lji
N−1
i
Qij ← Dij × (pi − pj )/Lij // Using Eq. (8)
if Qij < 0 then
Qij = 0
end if
Dij ← Qij + Dij // Using Eq. (7)
count ← count + 1
until a termination criterion is met
| 9cs.NE
|
arXiv:1602.01952v2 [math.GR] 26 Feb 2017
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
DANIEL J. WOODHOUSE
Abstract. We consider a finitely generated virtually abelian group G acting
properly and without inversions on a CAT(0) cube complex X. We prove that G
stabilizes a finite dimensional CAT(0) subcomplex Y ⊆ X that is isometrically
embedded in the combinatorial metric. Moreover, we show that Y is a product of
finitely many quasilines. The result represents a higher dimensional generalization
of Haglund’s axis theorem.
1. Introduction
A CAT(0) cube complex X is a cell complex that satisfies two properties: it is
a geodesic metric space satisfying the CAT(0) comparison triangle condition, and
each n-cell is isometric to [0, 1]n . We will call this metric the CAT(0) metric dX and
refer to [2] for a comprehensive account. A hyperplane Λ ⊆ X is the subset of points
equidistant between two adjacent vertices. Despite the brevity of this definition,
hyperplanes are better understood via their combinatorial definition, and the reader
is urged to consult the literature; see [10] [7] [12] for the required background. There
also exists an alternative metric on the 0-cubes of X, that we will refer to as the
combinatorial metric dcX , sometimes referred to as the ℓ1 -metric. The combinatorial
distance between two 0-cubes is the length of the shortest combinatorial path in X
joining the 0-cubes. Equivalently, the combinatorial distance between two 0-cubes
is the number of hyperplanes in X separating them. We will always assume that
a group G acting on a CAT(0) cube complex preserves its cell structure and maps
cubes isometrically to cubes. A group G acts without inversions if the stabilizer
of a hyperplane also stabilizes each complementary component. The requirement
that the action be without inversions is not a serious restriction as G acts without
inversions on the cubical subdivision.
A connected CAT(0) cube complex X is a quasiline if it is quasiisometric to R.
The rank of a virtually abelian group commensurable to Zn is n. The goal of this
paper will be the following theorem:
Theorem 4.3. Let G be virtually Zn . Suppose G acts properly and without inversions on a CAT(0) cube complex X. Then G stabilizes a finite dimensional
subcomplex Y ⊆ X that is isometrically embedded in the combinatorial metric, and
1
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
2
Qm
Y ∼
= i=1 Ci , where each Ci is a cubical quasiline and m ≥ n. Moreover, StabG (Λ)
is a codimension-1 subgroup for each hyperplane Λ in Y .
Note that Y will not in general be a convex subcomplex.
Corollary 1.1. Let A be a finitely generated virtually abelian group acting properly
on a CAT(0) cube complex X. Then A acts metrically properly on X.
Corollary 1.2. Let G be a finitely generated group acting properly on a CAT(0)
cube complex X. Then virtually Zn subgroups are undistorted in G.
Let g be an isometry of X, and let x ∈ X. The displacement of g at x, denoted τx (g), is the distance dX (x, gx). The translation length of g, denoted τ (g), is
inf{τx (g) | x ∈ X}. Similarly, if x is a 0-cube of X, we can define the combinatorial
displacement of g at x, denoted τxc (g), as dcX (x, gx) and the combinatorial translation length, denoted τ c (g), is inf{τxc (g) | x ∈ X}. Note that τ , and τ c are conjugacy
invariant. An isometry g of a CAT(0) space is semisimple if τx (g) = τ (g) for some
x ∈ X, and G acts semisimply on a CAT(0) space X if each g ∈ G is semisimple.
If a virtually Zn group G acts metrically properly by semisimple isometries on a
CAT(0) space X, then the Flat Torus Theorem [2] provides a G-invariant, convex,
flat En ⊆ X. A group acting on a CAT(0) cube complex does not, in general, have to
do so semisimply. See [1] for examples of non-semisimple isometries in Thompson’s
group F acting on an infinite dimensional CAT(0) cube complex. Alternatively,
in [6] a free-by-cyclic group G is shown not to permit a semisimple action on a
CAT(0) space. Yet in [13] it is shown that G does act freely on a CAT(0) cube
complex. Thus Theorem 4.3 can be applied to such actions, whereas the classical
Flat Torus Theorem cannot.
A virtually abelian subgroup is highest if it is not virtually contained in a higher
rank abelian subgroup. If G is a highest virtually abelian subgroup of a group
acting properly and cocompactly on a CAT(0) cube complex X, then G cocompactly
stabilizes a convex subcomplex Y which is a product of quasilines, as above [14].
However, this theorem fails without the highest hypothesis. Moreover, most actions
do not arise in the above fashion.
Despite the fact that the Flat Torus Theorem will not hold under the hypotheses
of Theorem 4.3, we can deduce the following:
Corollary 4.4. Let G be virtually Zn . Suppose G acts properly and without inversions on a CAT(0) cube complex X. Then G cocompactly stabilizes a subspace
F ⊆ X homeomorphic to Rn such that for each hyperplane Λ ⊆ X, the intersection
Λ ∩ F is either empty or homeomorphic to Rn−1 .
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
3
The initial motivation for Theorem 4.3 and Corollary 4.4 was to resolve the following question posed by Wise. Although we have not found a combinatorial flat,
Corollary 4.4 is perhaps better suited to applications (see [15]).
Problem 1.3. Let Z2 act freely on a CAT(0) cube complex Y . Does there exists
a Z2 -equivariant map F → Y where F is a square 2-complex homeomorphic to R2 ,
and such that no two hyperplanes of F map to the same hyperplane in Y ?
A combinatorial geodesic axis for g is a g-invariant, isometrically embedded, subcomplex γ ⊆ X with γ ∼
= R. Note that γ realizes the minimal combinatorial translation length of g. Theorem 4.3 is a high dimensional generalization of Haglund’s
combinatorial geodesic axis theorem. Haglund’s proof involved an argument by
contradiction, exploiting the geometry of hyperplanes. We reprove the result in
Section 5 by using the dual cube complex construction of Sageev. The results are
further support for Haglund’s slogan “in CAT(0) cube complexes the combinatorial
geometry is as nice as the CAT(0) geometry”.
The following is an application of Theorem 4.3, and the argument is inspired by
the solvable subgroup theorem [2]. Note that since we do not require that the action
of G on a CAT(0) cube complex be semisimple the following is not covered by the
solvable subgroup theorem.
Corollary 1.4. Let H be virtually Zn , and let φ : H → H be an injection with
φ 6= φi for all i > 1. Then G = hH, t | t−1 ht = φ(h) : h ∈ Hi cannot act properly
on a CAT(0) cube complex.
Proof. Suppose that G acts properly on a CAT(0) cube complex X. After subdividing X we can assume that G acts without inversions. As H is finitely generated,
there exists an a in the finite generating set such that φi (a) 6= a for all i ∈ N,
otherwise φi = φ for some i, contradicting our hypothesis. Thus, |{φi (a)}| = ∞. By
Theorem 4.3 there is an H-equivariant isometrically embedded subcomplex Y ⊆ X
Qm
such that Y ∼
= i=1 Ci where each Ci is a cubical quasiline.
As Y is isometrically embedded in X in the combinatorial metric, the combinatorial translation length τ c (φi (a)) is the same in Y as it is in X. The set {τ c (φi (a))}i∈N
must be unbounded since the action of H on Y is proper and Y is locally finite. However, since τ c is conjugacy invariant in G, we conclude that τ c (φi (a)) = τ c (φj (a))
for all i, j ∈ N. Thus, we arrive at the contradiction that {τ c (φi (a))}i∈N is both
bounded and unbounded.
However, we have the following example of a solvable group which does act freely
on a CAT(0) cube complex.
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
4
Example 1.5. Let H = ha1 , a2 , . . . | [ai , aj ] : i 6= ji. Note that H is the fundamental
group of the nonpositively curved cube complex Y obtained from a 0-cube v, and
1-cubes e1 , e2 , e3 . . . with n-cubes inserted for every cardinality n collection of 1cubes to create an n-torus. One should think of Y as an infinite cubical torus. The
oriented loop ei represents the element ai .
Let φ : H → H be the monomorphism such that φ(ai ) = ai+1 . Let G = H∗φ =
ht, a1 , a2 , . . . | [ai , aj ] : i 6= j , t−1 ai t = ai+1 i be the associated ascending HNN
extension. Note that G is generated by a1 and t. There is a graph of spaces X
obtained by letting Y be the vertex space and Y × [0, 1] be the edge space and
identifying (v, 1) and (v, 0) with v, and the 1-cube ei × {1} with ei and ei × {0} with
ei+1 . Note that X is nonpositively curved, and therefore G = π1 X acts freely on
e the universal cover of X.
the CAT(0) cube complex X,
Acknowledgements: I would like to thank Daniel T. Wise, Mark F Hagen,
Jack Button, Piotr Przytycki, and Dan Guralnik.
2. Dual Cube Complexes
←
− −
→
Let S be a set. A wall Λ = { Λ , Λ } in S is a partition of S into two disjoint,
←
− −
→
nonempty subsets. The subsets Λ , Λ are the halfspaces of Λ. A wall Λ separates
x, y ∈ S if they belong to distinct halfspaces of Λ. Let K ⊆ S. A wall Λ intersects
←
−
→
−
K if K nontrivially intersects both Λ and Λ . Let W be a set of walls in S, then
(S, W) is a wallspace if for all x, y ∈ S, the number of walls separating x and y is
finite. If Λ intersects K, then the restriction of Λ to K, is the wall in K determined
←
−
−
→
by Λ K = { Λ ∩ K, Λ ∩ K}.
In this paper duplicate walls are not permitted in W. Let H be the set of halfspaces of corresponding to W.
Example 2.1. Let X be a CAT(0) cube complex, and let Λ ⊆ X be a hyperplane
in X. The complement X − Λ has two components, therefore defining a wall in X
←
−
→
−
such that Λ is an open halfspace not containing Λ, and Λ is a closed halfspace
←
− −
→
containing Λ. Note that Λ ⊔ Λ = X. Let L(Λ) and R(Λ) denote the maximal
←
−
→
−
subcomplexes contained in Λ and Λ respectively. Note that L(Λ) and R(Λ) are
convex subcomplexes. Let W be the set of walls determined by the hyperplanes in
X. Then (X, W) is the wallspace associated to X. Note that we are using Λ to
denote both the hyperplane and the wall corresponding to the hyperplane.
←
− −
→
A function c : W → H is a 0-cube if c[Λ] ∈ { Λ , Λ } and the following two
conditions are satisfied:
(1) For all Λ1 , Λ2 ∈ W the intersection c[Λ1 ] ∩ c[Λ2 ] is nonempty.
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
5
(2) For all x ∈ S, the set {Λ ∈ W | x ∈
/ c[Λ]} is finite.
The dual cube complex C(S, W) is the connected CAT(0) cube complex obtained
by letting the union of all 0-cubes be the 0-skeleton. Two 0-cubes c1 6= c2 are
endpoints of a 1-cube if c1 [Λ] = c2 [Λ] for all but precisely one Λ ∈ W. An n-cube
is then inserted wherever there is the 1-skeleton of an n-cube. The hyperplanes
in C(S, W) are identified naturally with the walls in W. A proof of the fact that
C(S, W) is in fact a CAT(0) cube complex can be found in [9].
A point x ∈ S determines a 0-cube cx defined such that x ∈ cx [Λ] for all Λ ∈ W.
Condition (1) holds immediately since x ∈ cx [Λ] for all Λ ∈ W. Condition (2)
holds for cx , since if y ∈ S a wall Λ does not separate x and y, we can deduce that
y ∈ cx [Λ], hence all but finitely many Λ satisfy y ∈ cx [Λ]. Such 0-cubes are called
the canonical 0-cubes.
Lemma 2.2. Let X be a CAT(0) cube complex. Let W be a set of walls obtained
from the hyperplanes in X. Let Z be a connected subcomplex of X, and let WZ ⊆
W be the subset of walls intersecting Z. Let V be walls in WZ restricted to Z.
Then (Z, V) is a wallspace and C(Z, V) embeds in C(X, W) isometrically in the
combinatorial metric.
Proof. We first claim that the map WZ → V is an injection. Suppose that Λ1 , Λ2 ∈
WZ are distinct walls. As Λ1 , Λ2 intersects Z, and since Z is connected, there
are 1-cubes e1 , e2 in Z that are dual to the hyperplanes corresponding to Λ1 , Λ2 .
Therefore, both 0-cubes in e1 belong in a single halfspace of Λ2 Z , so Λ1 Z 6= Λ2 Z .
We construct a map φ : C(Z, V) → C(X, W) on the 0-skeleton first. Let c be a
0-cube in C(Z, V). We let φ(c) ∈ C(X, W) be the uniquely defined 0-cube such that
φ(c)[Λ] ⊇ c[Λ Z ] for Λ Z ∈ V, and φ(c)[Λ] ⊇ Z for Λ ∈ W − WZ . To verify that φ(c)
is a 0-cube, first observe that φ(c)[Λ1 ]∩φ(c)[Λ2 ] is nonempty since Λ1 Z ∩Λ2 Z ⊆ X.
Secondly, if x ∈ X we need to show that x ∈ φ(c)[Λ] for all but finitely many Λ ∈ W.
Choose z ∈ Z, then z ∈ c[Λ Z ] for all Λ Z ∈ V − {Λ1 Z , . . . , Λk Z }, hence z ∈ φ(c)[Λ]
for all Λ ∈ WZ − {Λ1 , . . . , Λk }. Let {Λk+1 , . . . , Λk+ℓ } be the set of walls in W
separating x and z. Then x ∈ φ(c)[Λ] for all Λ ∈ W − {Λ1 , . . . Λk+ℓ }.
The 0-cubes are embedded since if c1 6= c2 , there exists Λ Z ∈ V such that
c1 [Λ Z ] 6= c2 [Λ Z ], hence φ(c1 )[Λ] 6= φ(c2 )[Λ]. If c1 , c2 are adjacent 0-cubes in
C(Z, V), then c1 [Λ Z ] = c2 [Λ Z ] for all Λ Z ∈ V, with the exception of precisely one
wall Λ̂ . Therefore, we can deduce that φ(c1 )[Λ] = φ(c2 )[Λ] for all walls in W, with
Z
the precise exception of Λ̂. Therefore, the 1-skeleton of C(Z, V) embeds in C(X, W),
which is sufficient for φ to extend to an embedding of the entire cube complex.
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
6
Consider C(Z, V) as a subcomplex of C(X, W). The set of hyperplanes in C(Z, V)
embeds into the set of hyperplanes in C(X, W). To see that C(Z, V) is an isometrically embedded subcomplex, let z1 , z2 be 0-cubes in Z and γ be a geodesic
combinatorial path in C(Z, V) joining them. Each hyperplane dual to γ in C(Z, V)
intersects γ precisely once, and since the hyperplanes in C(Z, V) inject to hyperplanes in C(X, W), it is geodesic there as well.
Given a wall Λ associated to a hyperplane in X we let N (Λ) denote the carrier
of Λ, by which we mean the union of all cubes intersected by Λ.
The following Lemma decribes what is called the restriction quotient in [3].
Lemma 2.3. Let S be a set and let W be a set of walls of S. Let G be a group
acting on (S, W). Let V ⊆ W be a G-invariant subset. Then there is a G-equivariant
function φ : C(S, W)0 → C(S, V)0 . Moreover, φ−1 (z) is nonempty for all 0-cubes z
in C(S, V).
Proof. Let c be a 0-cube in C(S, W). Let φ(c)[Λ] = c[Λ] for Λ ∈ V. It is immediate
that φ is G-equivariant.
To verify φ(c)[Λ] is a 0-cube in C(S, V) first note that φ(c1 )[Λ1 ] ∩ φ(c2 )[Λ2 ] 6= ∅
for all Λ1 , Λ2 ∈ V, since c1 [Λ1 ] ∩ c2 [Λ2 ] 6= ∅ for all Λ1 , Λ2 ∈ W. Secondly, for all
x ∈ S observe that x ∈ φ(c)[Λ] for all but finitely many Λ ∈ V. Indeed, this is true
for all but finitely many Λ ∈ W.
To see that φ−1 (z) is non-empty for all 0-cubes z in C(S, V) we determine a 0cube x in C(S, W) such that φ(x) = z. Fix s ∈ S. Let x[Λ] = z[Λ] for Λ ∈ V.
−
→
→
−
Suppose that Λ ∈ W − V. If Λ ⊇ z[Λ′ ] for some Λ′ ∈ V let x[Λ] = Λ . Similarly if
←
−
Λ ⊇ z[Λ′ ]. Otherwise, if Λ intersects z[Λ′ ] for all Λ′ ∈ V then let s ∈ x[Λ].
To verify that x is a 0-cube, consider the following cases to show x[Λ1 ] ∩ x[Λ2 ] 6= ∅
for Λ1 , Λ2 ∈ W. If Λ1 , Λ2 ∈ V then x[Λ1 ] ∩ x[Λ2 ] = z[Λ1 ] ∩ z[Λ2 ] 6= ∅. Suppose
that Λ1 ∈ W − V and x[Λ1 ] ⊆ z[Λ′1 ] for some Λ′1 ∈ V. If Λ2 ∈ V, then x[Λ1 ] ∩
x[Λ2 ] ⊇ z[Λ′1 ] ∩ z[Λ2 ] 6= ∅. If Λ2 ∈ W − V and x[Λ2 ] ⊆ z[Λ′2 ] for some Λ′2 ∈ V
then x[Λ1 ] ∩ x[Λ2 ] ⊆ z[Λ′1 ] ∩ z[Λ′2 ] 6= ∅. If Λ2 intersects z[Λ] for all Λ ∈ V, then
x[Λ1 ] ∩ x[Λ2 ] ⊇ z[Λ′1 ] ∩ x[Λ2 ] 6= ∅. Finally if both s ∈ x[Λ1 ] and x[Λ2 ], then their
intersection will contain at least s.
Finally, we verify that for s′ ∈ S there are only finitely many Λ ∈ W such that
s′ ∈
/ x[Λ]. Suppose, by way of contradiction, that there is an infinite subset of walls
{Λ1 , Λ2 , . . .} ⊆ W such that s′ ∈
/ x[Λi ] for all i ∈ N. We can assume, by excluding
at most finitely many walls, that each Λi ∈ W − V. Similarly, by excluding finitely
many walls, we can assume that Λi does not separate s and s′ . Therefore, s ∈
/ x[Λi ]
′
′
for i ∈ N. Therefore, by construction of x, there exist Λi ∈ V such that z[Λi ] ⊆ x[Λi ],
which implies that s′ ∈
/ z[Λ′i ]. There are infinitely many distinct Λ′i , as otherwise
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
7
there is a Λ′ ∈ V such that z[Λ′ ] ⊆ x[Λi ] for infinitely many i, which would imply
that infinitely many Λi separate s′ from an element in the complement of z[Λ′ ].
Therefore, infinitely many distinct walls Λ′i ∈ V have s′ ∈
/ z[Λ′i ], contradicting that
z is a 0-cube in C(S, V).
3. Minimal Zn -invariant convex subcomplexes
The following is Theorem 2 from [5]. As this paper is written in Russian, we give
a proof in Appendix A based on the work in [8] as well as stating the definition of
codimension-1.
Theorem 3.1 (Gerasimov [5]). Let G be a finitely generated group that acts on
a CAT(0) cube complex X without a fixed point or inversions. Then there is a
hyperplane in X that is stabilized by a codimension-1 subgroup of G.
The goal of this section is to prove the following:
Lemma 3.2. Let G be a finitely generated group acting without fixed point or inversions on a CAT(0) cube complex X. There exists a minimal, G-invariant, convex
subcomplex Xo ⊆ X such that Xo contains only finitely many hyperplane orbits, and
every Xo hyperplane stabilizer is a codimension-1 subgroup of G.
Proof. Since G is finitely generated, by taking the convex hull of a G-orbit we
obtain a G-invariant convex subcomplex Xo ⊆ X containing finitely many G-orbits
of hyperplanes. Assume that Xo is a minimal such subcomplex in terms of the
number of hyperplane orbits.
Let (Xo , W) be the wallspace obtained from the hyperplanes in Xo . Suppose that
StabG (Λ) is not a codimension-1 subgroup of G for some Λ ∈ W. Let GΛ ⊆ W be
the G-orbit of Λ. By Lemma 2.3 there is an G-invariant map φ : Xo0 → C(Xo , GΛ)0 .
Since StabG (Λ) is not commensurable to a codimension-1 subgroup, Theorem 3.1
implies that there is a fixed 0-cube x in C(Xo , GΛ). Lemma 2.3 then implies that
T
←
−
φ−1 (x) is non-empty. Assuming that φ−1 (x) ⊆ Λ , then the intersection g∈G gL(Λ)
contains a proper, convex, G-invariant subcomplex of Xo , with one less hyperplane
orbit. This contradicts the minimality of Xo .
The following Corollary follows since all codimension-1 subgroups of a rank n
virtually abelian group are of rank (n − 1).
Corollary 3.3. Let G be a rank n, virtually abelian group acting without fixed
point or inversions on a CAT(0) cube complex X. Then there exists a minimal,
G-invariant, convex subcomplex Xo ⊆ X such that Xo contains only finitely many
hyperplane orbits, and every hyperplane stabilizer is a rank (n − 1) subgroup of G.
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
8
4. Proof of Main Theorem
Definition 4.1. Regard R as a CAT(0) cube complex whose 0-skeleton is Z. Let g
be an isometry of X. A geodesic combinatorial axis for g is a g-invariant subcomplex
homeomorphic to R that embeds isometrically in X.
Definition 4.2. Let (M, d) be a metric space. The subspaces N1 , N2 ⊆ M are
coarsely equivalent if each lies in an r-neighbourhood of the other for some r > 0.
Theorem 4.3. Let G be virtually Zn . Suppose G acts properly and without inversions on a CAT(0) cube complex X. Then G stabilizes a finite dimensional
subcomplex Y ⊆ X that is isometrically embedded in the combinatorial metric, and
Qm
Y ∼
= i=1 Ci , where each Ci is a cubical quasiline and m ≥ n. Moreover, StabG (Λ)
is a codimension-1 subgroup for each hyperplane Λ in Y .
Proof. By Corollary 3.3 there is a minimal, non-empty, convex subcomplex Xo ⊆ X
stabilized by G, containing finitely many hyperplane orbits, and StabilizerG (Λ) is a
rank (n − 1) subgroup of G, for each hyperplane Λ ⊆ Xo .
Let S = {g1 , . . . , gr } be a generating set for G. Let x ∈ Xo be a 0-cube. Let Υ
be the Cayley graph of G with respect to S. Let φ : Υ → Xo be a G-equivariant
map that sends vertices to vertices, and edges to combinatorial paths or vertices in
Xo . Let Q = φ(Υ). As G acts properly on X, and cocompactly on Υ, the graph
Q is quasiisometric to G. Let WQ be the set of hyperplanes intersecting Q, and
let (Q, WQ ) be the associated wallspace. By Lemma 2.2 we know that C(Q, WQ )
is an isometrically embedded subcomplex of Xo . Fix a proper action of G on Rn ,
and let q : Q → Rn be a G-equivariant quasiisometry. Note that StabilizerG (Λ) is
a quasiisometrically embedded subgroup of G, for all Λ ∈ WQ . Thus q(Λ ∩ Q) is
←
−
coarsely equivalent to a codimension-1 affine subspace H ⊆ Rn . Moreover, q( Λ ∩ Q)
→
−
and q( Λ ∩ Q) are coarsely equivalent to the halfspaces of H.
Let n > 0. Since there are finitely many orbits of hyperplanes in Xo , there
are only finitely many commensurability classes of stabilizers. Therefore, we may
F
partition WQ as the disjoint union m
i=1 Wi where each Wi contains all walls with
commensurable stabilizers. For each Λi ∈ Wi let q(Λi ∩ Q) be coarsely equivalent to
a codimension-1 affine subspace Hi ⊆ Rn , stabilized by StabilizerG (Λi ). If i 6= j then
Hi and Hj are nonparallel affine subspaces, and therefore Λi and Λj will intersect
in Q. Therefore, every wall in Wi intersects every wall in Wj if i 6= j, and thus
Qm
C(Q, WQ ) ∼
= i=1 C(Q, Wi ).
Finally, we show that C(Q, Wi ) is a quasiline for each 1 ≤ i ≤ m. As G permutes
Q
′
the factors in m
i=1 C(Q, Wi ), there is a finite index subgroup G 6 G that preserves
each factor. For each i, the stabilizers StabG (Λ) are commensurable for all Λ ∈
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
9
Wi . Therefore, there is a cyclic subgroup Zi that is not virtually contained in
any StabG (Λ) and thus acts freely on C(Q, Wi ). As the stabilizers of Λ ∈ Wi
are commensurable, all q(Λ ∩ Q) will be quasi-equivalent to parallel codimension-1
affine subspaces of Rn , which implies that only finitely many Zi -translates of Λ can
pairwise intersect. As there are finitely many Zi -orbits of Λ in Wi , there is an upper
bound on the number of pairwise intersecting hyperplanes in Wi . Thus, there are
finitely many Zi -orbits of maximal cubes in C(Q, Wi ), which implies that C(Q, Wi )
is CAT(0) cube complex quasiisometric to R.
We can now prove Corollary 4.4.
Corollary 4.4. Let G be virtually Zn . Suppose G acts properly and without inversions on a CAT(0) cube complex X. Then G cocompactly stabilizes a subspace
F ⊆ X homeomorphic to Rn such that for each hyperplane Λ ⊆ X, the intersection
Λ ∩ F is either empty or homeomorphic to Rn−1 .
Proof. By Theorem 4.3 there is a G-equivariant, isometrically embedded, subcomQ
plex Y ⊆ X, such that Y = m
i=1 Ci , where each Ci is a quasiline, and StabG (Λ)
is a codimension-1 subgroup. Considering Y with the CAT(0) metric, note that
Y is a complete CAT(0) metric space in its own right, and G acts semisimply on
Y . By the Flat Torus Theorem [2] there is an isometrically embedded flat F ⊆ Y .
Note that F ⊆ X is not isometrically embedded. As StabG (Λ) is a codimension-1
subgroup of G for each hyperplane Λ in X, the intersection Λ ∩ F = (Λ ∩ Y ) ∩ F is
either empty or, as F ⊆ Y is isometrically embedded, the hyperplane intersection
is an isometrically embedded copy of Rn−1 .
5. Haglund’s Axis
The goal of this section is to reprove the following result of Haglund as a consequence of Corollary 4.4.
Theorem 5.1 (Haglund [7]). Let G be a group acting on a CAT(0) cube complex
without inversions. Every element g ∈ G either fixes a 0-cube of G, or stabilizes a
combinatorial geodesic axis.
Proof. As finite groups don’t contain codimension-1 subgroups, Theorem 3.1 implies
that if g is finite order then it fixes a 0-cube. Suppose that G does not fix a 0-cube,
then hgi must act properly on X. By Corollary 4.4, there is a line L ⊂ X stabilized
by G, that intersects each hyperplane at most once at a single point in L. Let WL
be the set of hyperplanes intersecting L. Note that the intersection points of the
walls in WL with L is locally finite subset.
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
10
Fix a basepoint p ∈ L that doesn’t belong to a hyperplane intersecting L, and let x
be the canonical 0-cube corresponding to p. Let Λ1 , . . . , Λk be the set of hyperplanes
←
−
separating p and gp, and assume that p ∈ Λ i . Reindex the hyperplanes such that
←
−
←
−
←
−
Λ 1 ∩ L ⊆ Λ 2 ∩ L ⊆ · · · ⊆ Λ k ∩ L. The ordering of the hyperplanes separating
p and gp determines a combinatorial geodesic joining x and gx of length k, where
the i-th edge is a 1-cube dual to Λi . This can be extended hgi-equivariantly, to
obtain a combinatorial geodesic axis Lc , since each hyperplanes intersects Lc at
most once.
Appendix A. Codimension-1 Subgroups
Definition A.1. Let G be a finitely generated group. Let Υ denote the Cayley
graph of G with respect to some finite generating set. A subgroup H 6 G is
codimension-1 if K/Υ has more than one end.
Let ⊕ denote the operation of symmetric difference. A subset A ⊆ G is H-finite
if A ⊆ HF where F is some finite subset of G. We will use the following equivalent
formulation (see [11]) of codimension-1: A subgroup H ≤ G is a codimension-1
subgroup if there exists some A ⊆ G such that
(1) A = HA,
(2) A is H-almost invariant, that is to say that A⊕ Ag is H-finite for any g ∈ G.
(3) A is H-proper, that is to say that neither A nor G − A is H-finite.
We will reprove the following theorem from [5] using techniques from [8].
Theorem A.2. Let G be a finitely generated group acting on a CAT(0) cube complex
X without edge inversions or fixing a 0-cube. Then the stabilizer of some hyperplane
in X is a codimension-1 subgroup of G.
Proof. Suppose that no hyperplane stabilizer is a codimension-1 subgroup of G. We
will find a 0-cube fixed by G.
Let H denote the set of hyperplanes in X. We can assume that X has finitely
many G-orbits of hyperplanes after possibly passing to the convex hull of a single
0-cube orbit in X. If x, y are 0-cubes in X, then let ∆(x, y) ⊆ H denote the
hyperplanes separating x and y. Note that
dcX (x, y) = |∆(x, y)|.
Let Λ1 , . . . , Λn be a minimal set of representatives of those orbits. Let
←
−
Hi = StabG (Λi ) and Ai = {g ∈ G | gx0 ∈ Λ i }.
We can verify that Ai satisfies the first two criteria in Definition A.1.
(1): It is immediate that Ai = Hi Ai , as G doesn’t invert the hyperplanes in X.
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
11
(2): Let xor denote the exclusive or. For f ∈ G we can deduce that Ai ⊕ Ai f is
Hi -finite:
←
−
←
−
g ∈ Ai ⊕ Ai f ⇐⇒ gx0 ∈ Λ i xor gf −1 x0 ∈ Λ i
←
−
←
−
⇐⇒ x0 ∈ g−1 Λ i xor f −1 x0 ∈ g −1 Λ i
⇐⇒ g ∈ G such that g −1 Λi separates x0 and f −1 x0 .
As (X, H) is a wallspace, there are only finitely many g ∈ G such that g−1 Λi
separates x0 and f −1 x0 . If g1 Λi , . . . , gk Λi are the translates then
Ai + Ai f = {g1 , . . . , gk }Hi
which implies almost Hi -invariant.
Therefore, Ai cannot be Hi -proper, for any i, as we have assumed that none of the
Hi are codimension-1. This means that Ai is Hi -finite so Ai ⊆ Hi Fi where Fi ⊆ G
is finite.
Claim. dX (x0 , f x0 ) < 2 maxi (|Fi |) for all f ∈ G.
Proof.
gΛi ∈ ∆(x0 , f x0 ) ⇐⇒ x0 [gΛi ] 6= f x0 [gΛi ]
←
−
←
−
⇐⇒ x0 ∈ g Λ i xor f x0 ∈ g Λ i
←
−
←
−
⇐⇒ g −1 x0 ∈ g Λ i xor g −1 f x0 ∈ g Λ i
⇐⇒ g −1 ∈ Ai xor g−1 ∈ Af −1
⇐⇒ g −1 ∈ Ai + Ai f −1
As the final set is covered by 2|Fi | translates of Hi , we can deduce that there are at
most 2|Fi | hyperplanes in ∆(x0 , f x0 ).
Thus, we can conclude that the G-orbit of x0 is a bounded set. If G has a
finite orbit in X, then the convex hull of the orbit is a compact, finite dimensional,
complete CAT(0) cube complex, and we can apply Corollary II.2.8 (1) from [2] to
find a fixed point p. As p is in the interior of some n-cube that is fixed by G, and
since G doesn’t invert hyperplanes we can deduce that G fixes a 0-cube in that cube.
If the G-orbits in X are infinite, then their convex hull may not be complete, so the
above argument will not hold.
Let C(H) denote the connected cube, a graph with vertices given by functions
c : H → {0, 1} with finite support, and edges that join a pair of distinct vertices if
and only if they differ on precisely one hyperplane.
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
12
Fix a 0-cube x0 . Then there is an embedding
φ : X 1 ֒→ C(H)
that maps the 0-cube x to cx where
1 if x[Λ] 6= x [Λ]
0
cx (Λ) =
0 if x[Λ] = x0 [Λ]
A hyperplan Λ ∈ H separates two vertices c1 , c2 in C(H) if c1 (Λ) 6= c2 (Λ). Note that
Λ separates 0-cubes x, y in X if and only if it separates φ(x) and φ(y). Therefore,
we can define ∆(c1 , c2 ) for vertices in C(H) and conclude that if x, y are 0-cubes in
X then ∆(x, y) = ∆(φ(x), φ(y)). This implies that φ is an isometric embedding in
the combinatorial metric.
We will show that a bounded orbit in X implies there is a fixed 0-cube in C(H)
and then argue that we can go one step further and find a fixed 0-cube in X.
Let ℓ2 (H) be the Hilbert Space of square summable functions s : H → R. There
is an embedding ρ : C(H) → ℓ2 (H) given by
ρ(c)(Λ) = c[Λ]
It is straight forward to verify that kρ(c1 ) − ρ(c2 )k2 = dC(H) (c1 , c2 ). There is a
G-action on ℓ2 (H) such that if s ∈ ℓ2 (H), Λ ∈ H, g ∈ G then
s(g−1 Λ)
if cx0 (g−1 Λ) = cx0 (Λ)
gs(Λ) =
1 − s(g−1 Λ) if cx (g−1 Λ) 6= cx (Λ)
0
0
It is again straight forward to verify that this action is by isometries, and that ρ is
G-equivariant.
As Gx0 is bounded, so is G(ρ ◦ φ(x0 )). It then follows that G has a fixed point in
2
ℓ (H) ([8] gives a proof and also cites Lemma 3.8 in [4]). Let s : H → R be the fixed
point. For all g ∈ G we can deduce that s(gΛ) is either s(Λ) or 1 − s(Λ). Therefore
s can only take two values on the hyperplanes in a single G-orbit. As s has to be
square summable the two values have to be 0 and 1, and s can only take the value
1 on finitely many hyperplanes. Thus, s is the image of a point c in C(S).
Let c ∈ C(S) be a G-invariant vetex which minimizes the distance to the image of
1
X in C(S). Let Z be a G-orbit of 0-cubes in X such that φ(Z) realize the minimal
distance from c.
Let V be the set of hyperplanes that intersect {c} ∪ V. Every hyperplane in V
must intersect Z otherwise if F ⊆ V is the finite, G-invariant subset of hyperplanes
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
13
separating c from Z we can define a 0-cube c′ such that
c(Λ)
if Λ ∈
/F
c′ (Λ) =
1 − c(Λ′ ) if Λ ∈ F
and deduce that c′ is G-invariant and is |F| closer to Z than c.
Let z0 , z1 , z2 , . . . an enumeration of 0-cubes in Z. Each hyperplane separating
z0 and z1 must lie in either ∆(z0 , c) or ∆(z1 , c). As z0 is minimal distance in X
from c, the edges in X incident to z0 must be dual to hyperplanes not in ∆(z0 , c),
and instead belongs to ∆(z1 , c). Therefore, the hyperplane Λ0 ∈ V dual to the first
edge in a combinatorial geodesic joining z0 to z1 must lie in ∆(z1 , c). Similarly,
there exists a hyperplane Λ1 dual to the first edge of the combinatorial geodesic in
X joining z1 to z2 that belongs to ∆(z2 , c) but not ∆(z1 , c). Note that Λ1 cannot
intersect Λ0 in X, otherwise Λ0 would be dual to an edge incident to z1 , which would
imply that there exists a 0-cube in X adjacent to z1 that is closer to c. Therefore
Λ0 , Λ1 separates z0 from z2 in X. Iterating this argument produces a sequence of
disjoint hyperplanes Λ0 , Λ2 , Λ3 , . . . such that Λ0 , . . . , Λk separates z0 from zk+1 in
X. This contradicts the hypothesis that Z is a bounded set in X.
References
[1] Yael Algom-Kfir, Bronislaw Wajnryb, and Pawel Witowicz. A parabolic action on a proper,
CAT(0) cube complex. J. Group Theory, 16(6):965–984, 2013.
[2] Martin R. Bridson and André Haefliger. Metric spaces of non-positive curvature. SpringerVerlag, Berlin, 1999.
[3] Pierre-Emmanuel Caprace and Michah Sageev. Rank rigidity for cat(0) cube complexes. Geometric And Functional Analysis, 21:851–891, 2011. 10.1007/s00039-011-0126-7.
[4] Pierre de la Harpe and Alain Valette. La propriété (T ) de Kazhdan pour les groupes localement
compacts. Number 175 in Astérisque. 1989. With an appendix by M. Burger.
[5] V. N. Gerasimov. Semi-splittings of groups and actions on cubings. In Algebra, geometry,
analysis and mathematical physics (Russian) (Novosibirsk, 1996), pages 91–109, 190. Izdat.
Ross. Akad. Nauk Sib. Otd. Inst. Mat., Novosibirsk, 1997.
[6] S. M. Gersten. The automorphism group of a free group is not a CAT(0) group. Proc. Amer.
Math. Soc., 121(4):999–1002, 1994.
[7] Frédéric Haglund. Isometries of CAT(0) cube complexes are semi-simple. pages 1–17, 2007.
[8] Graham A. Niblo and Martin A. Roller. Groups acting on cubes and Kazhdan’s property (T).
Proc. Amer. Math. Soc., 126(3):693–699, 1998.
[9] Michah Sageev. Ends of group pairs and non-positively curved cube complexes. Proc. London
Math. Soc. (3), 71(3):585–617, 1995.
[10] Michah Sageev. Codimension-1 subgroups and splittings of groups. J. Algebra, 189(2):377–389,
1997.
[11] Peter Scott. Ends of pairs of groups. J. Pure Appl. Algebra, 11(1-3):179–198, 1977/78.
A GENERALIZED AXIS THEOREM FOR CUBE COMPLEXES
14
[12] Daniel T. Wise. From riches to raags: 3-manifolds, right-angled Artin groups, and cubical
geometry, volume 117 of CBMS Regional Conference Series in Mathematics. Published for the
Conference Board of the Mathematical Sciences, Washington, DC, 2012.
[13] Daniel T. Wise. Cubular tubular groups. Trans. Amer. Math. Soc., 366(10):5503–5521, 2014.
[14] Daniel T. Wise and Daniel J. Woodhouse. A cubical flat torus theorem and the bounded
packing property. Isreal J. Math., pages 1–12. Accepted.
[15] Daniel J. Woodhouse. Classifying virtually special tubular groups. 2015. In Preparation.
E-mail address: [email protected]
| 4math.GR
|
arXiv:1712.01815v1 [cs.AI] 5 Dec 2017
Mastering Chess and Shogi by Self-Play with a
General Reinforcement Learning Algorithm
David Silver,1∗ Thomas Hubert,1∗ Julian Schrittwieser,1∗
Ioannis Antonoglou,1 Matthew Lai,1 Arthur Guez,1 Marc Lanctot,1
Laurent Sifre,1 Dharshan Kumaran,1 Thore Graepel,1
Timothy Lillicrap,1 Karen Simonyan,1 Demis Hassabis1
1
DeepMind, 6 Pancras Square, London N1C 4AG.
∗
These authors contributed equally to this work.
Abstract
The game of chess is the most widely-studied domain in the history of artificial intelligence. The strongest programs are based on a combination of sophisticated search techniques, domain-specific adaptations, and handcrafted evaluation functions that have been
refined by human experts over several decades. In contrast, the AlphaGo Zero program
recently achieved superhuman performance in the game of Go, by tabula rasa reinforcement learning from games of self-play. In this paper, we generalise this approach into
a single AlphaZero algorithm that can achieve, tabula rasa, superhuman performance in
many challenging domains. Starting from random play, and given no domain knowledge
except the game rules, AlphaZero achieved within 24 hours a superhuman level of play in
the games of chess and shogi (Japanese chess) as well as Go, and convincingly defeated a
world-champion program in each case.
The study of computer chess is as old as computer science itself. Babbage, Turing, Shannon, and von Neumann devised hardware, algorithms and theory to analyse and play the game
of chess. Chess subsequently became the grand challenge task for a generation of artificial intelligence researchers, culminating in high-performance computer chess programs that perform at
superhuman level (9, 13). However, these systems are highly tuned to their domain, and cannot
be generalised to other problems without significant human effort.
A long-standing ambition of artificial intelligence has been to create programs that can instead learn for themselves from first principles (26). Recently, the AlphaGo Zero algorithm
achieved superhuman performance in the game of Go, by representing Go knowledge using
deep convolutional neural networks (22, 28), trained solely by reinforcement learning from
games of self-play (29). In this paper, we apply a similar but fully generic algorithm, which we
1
call AlphaZero, to the games of chess and shogi as well as Go, without any additional domain
knowledge except the rules of the game, demonstrating that a general-purpose reinforcement
learning algorithm can achieve, tabula rasa, superhuman performance across many challenging
domains.
A landmark for artificial intelligence was achieved in 1997 when Deep Blue defeated the human world champion (9). Computer chess programs continued to progress steadily beyond human level in the following two decades. These programs evaluate positions using features handcrafted by human grandmasters and carefully tuned weights, combined with a high-performance
alpha-beta search that expands a vast search tree using a large number of clever heuristics and
domain-specific adaptations. In the Methods we describe these augmentations, focusing on the
2016 Top Chess Engine Championship (TCEC) world-champion Stockfish (25); other strong
chess programs, including Deep Blue, use very similar architectures (9, 21).
Shogi is a significantly harder game, in terms of computational complexity, than chess (2,
14): it is played on a larger board, and any captured opponent piece changes sides and may subsequently be dropped anywhere on the board. The strongest shogi programs, such as Computer
Shogi Association (CSA) world-champion Elmo, have only recently defeated human champions (5). These programs use a similar algorithm to computer chess programs, again based on a
highly optimised alpha-beta search engine with many domain-specific adaptations.
Go is well suited to the neural network architecture used in AlphaGo because the rules of
the game are translationally invariant (matching the weight sharing structure of convolutional
networks), are defined in terms of liberties corresponding to the adjacencies between points
on the board (matching the local structure of convolutional networks), and are rotationally and
reflectionally symmetric (allowing for data augmentation and ensembling). Furthermore, the
action space is simple (a stone may be placed at each possible location), and the game outcomes
are restricted to binary wins or losses, both of which may help neural network training.
Chess and shogi are, arguably, less innately suited to AlphaGo’s neural network architectures. The rules are position-dependent (e.g. pawns may move two steps forward from the
second rank and promote on the eighth rank) and asymmetric (e.g. pawns only move forward,
and castling is different on kingside and queenside). The rules include long-range interactions
(e.g. the queen may traverse the board in one move, or checkmate the king from the far side
of the board). The action space for chess includes all legal destinations for all of the players’
pieces on the board; shogi also allows captured pieces to be placed back on the board. Both
chess and shogi may result in draws in addition to wins and losses; indeed it is believed that the
optimal solution to chess is a draw (17, 20, 30).
The AlphaZero algorithm is a more generic version of the AlphaGo Zero algorithm that was
first introduced in the context of Go (29). It replaces the handcrafted knowledge and domainspecific augmentations used in traditional game-playing programs with deep neural networks
and a tabula rasa reinforcement learning algorithm.
Instead of a handcrafted evaluation function and move ordering heuristics, AlphaZero utilises
a deep neural network (p, v) = fθ (s) with parameters θ. This neural network takes the board position s as an input and outputs a vector of move probabilities p with components pa = P r(a|s)
2
for each action a, and a scalar value v estimating the expected outcome z from position s,
v ≈ E[z|s]. AlphaZero learns these move probabilities and value estimates entirely from selfplay; these are then used to guide its search.
Instead of an alpha-beta search with domain-specific enhancements, AlphaZero uses a generalpurpose Monte-Carlo tree search (MCTS) algorithm. Each search consists of a series of simulated games of self-play that traverse a tree from root sroot to leaf. Each simulation proceeds by
selecting in each state s a move a with low visit count, high move probability and high value
(averaged over the leaf states of simulations that selected a from s) according to the current
neural network fθ . The search returns a vector π representing a probability distribution over
moves, either proportionally or greedily with respect to the visit counts at the root state.
The parameters θ of the deep neural network in AlphaZero are trained by self-play reinforcement learning, starting from randomly initialised parameters θ. Games are played by selecting
moves for both players by MCTS, at ∼ π t . At the end of the game, the terminal position sT is
scored according to the rules of the game to compute the game outcome z: −1 for a loss, 0 for
a draw, and +1 for a win. The neural network parameters θ are updated so as to minimise the
error between the predicted outcome vt and the game outcome z, and to maximise the similarity
of the policy vector pt to the search probabilities π t . Specifically, the parameters θ are adjusted
by gradient descent on a loss function l that sums over mean-squared error and cross-entropy
losses respectively,
(p, v) = fθ (s),
l = (z − v)2 − π > log p + c||θ||2
(1)
where c is a parameter controlling the level of L2 weight regularisation. The updated parameters
are used in subsequent games of self-play.
The AlphaZero algorithm described in this paper differs from the original AlphaGo Zero
algorithm in several respects.
AlphaGo Zero estimates and optimises the probability of winning, assuming binary win/loss
outcomes. AlphaZero instead estimates and optimises the expected outcome, taking account of
draws or potentially other outcomes.
The rules of Go are invariant to rotation and reflection. This fact was exploited in AlphaGo
and AlphaGo Zero in two ways. First, training data was augmented by generating 8 symmetries
for each position. Second, during MCTS, board positions were transformed using a randomly
selected rotation or reflection before being evaluated by the neural network, so that the MonteCarlo evaluation is averaged over different biases. The rules of chess and shogi are asymmetric,
and in general symmetries cannot be assumed. AlphaZero does not augment the training data
and does not transform the board position during MCTS.
In AlphaGo Zero, self-play games were generated by the best player from all previous iterations. After each iteration of training, the performance of the new player was measured against
the best player; if it won by a margin of 55% then it replaced the best player and self-play games
were subsequently generated by this new player. In contrast, AlphaZero simply maintains a single neural network that is updated continually, rather than waiting for an iteration to complete.
3
Figure 1: Training AlphaZero for 700,000 steps. Elo ratings were computed from evaluation
games between different players when given one second per move. a Performance of AlphaZero
in chess, compared to 2016 TCEC world-champion program Stockfish. b Performance of AlphaZero in shogi, compared to 2017 CSA world-champion program Elmo. c Performance of
AlphaZero in Go, compared to AlphaGo Lee and AlphaGo Zero (20 block / 3 day) (29).
Self-play games are generated by using the latest parameters for this neural network, omitting
the evaluation step and the selection of best player.
AlphaGo Zero tuned the hyper-parameter of its search by Bayesian optimisation. In AlphaZero we reuse the same hyper-parameters for all games without game-specific tuning. The
sole exception is the noise that is added to the prior policy to ensure exploration (29); this is
scaled in proportion to the typical number of legal moves for that game type.
Like AlphaGo Zero, the board state is encoded by spatial planes based only on the basic
rules for each game. The actions are encoded by either spatial planes or a flat vector, again
based only on the basic rules for each game (see Methods).
We applied the AlphaZero algorithm to chess, shogi, and also Go. Unless otherwise specified, the same algorithm settings, network architecture, and hyper-parameters were used for all
three games. We trained a separate instance of AlphaZero for each game. Training proceeded
for 700,000 steps (mini-batches of size 4,096) starting from randomly initialised parameters,
using 5,000 first-generation TPUs (15) to generate self-play games and 64 second-generation
TPUs to train the neural networks.1 Further details of the training procedure are provided in the
Methods.
Figure 1 shows the performance of AlphaZero during self-play reinforcement learning, as
a function of training steps, on an Elo scale (10). In chess, AlphaZero outperformed Stockfish
after just 4 hours (300k steps); in shogi, AlphaZero outperformed Elmo after less than 2 hours
(110k steps); and in Go, AlphaZero outperformed AlphaGo Lee (29) after 8 hours (165k steps).2
We evaluated the fully trained instances of AlphaZero against Stockfish, Elmo and the previous version of AlphaGo Zero (trained for 3 days) in chess, shogi and Go respectively, playing
100 game matches at tournament time controls of one minute per move. AlphaZero and the
previous AlphaGo Zero used a single machine with 4 TPUs. Stockfish and Elmo played at their
1
The original AlphaGo Zero paper used GPUs to train the neural networks.
AlphaGo Master and AlphaGo Zero were ultimately trained for 100 times this length of time; we do not
reproduce that effort here.
2
4
Game
White
Black
Win
Draw
Loss
Chess
AlphaZero
Stockfish
Stockfish
AlphaZero
25
3
25
47
0
0
Shogi
AlphaZero
Elmo
Elmo
AlphaZero
43
47
2
0
5
3
Go
AlphaZero
AG0 3-day
AG0 3-day
AlphaZero
31
29
–
–
19
21
Table 1: Tournament evaluation of AlphaZero in chess, shogi, and Go, as games won, drawn
or lost from AlphaZero’s perspective, in 100 game matches against Stockfish, Elmo, and the
previously published AlphaGo Zero after 3 days of training. Each program was given 1 minute
of thinking time per move.
strongest skill level using 64 threads and a hash size of 1GB. AlphaZero convincingly defeated
all opponents, losing zero games to Stockfish and eight games to Elmo (see Supplementary Material for several example games), as well as defeating the previous version of AlphaGo Zero
(see Table 1).
We also analysed the relative performance of AlphaZero’s MCTS search compared to the
state-of-the-art alpha-beta search engines used by Stockfish and Elmo. AlphaZero searches just
80 thousand positions per second in chess and 40 thousand in shogi, compared to 70 million
for Stockfish and 35 million for Elmo. AlphaZero compensates for the lower number of evaluations by using its deep neural network to focus much more selectively on the most promising
variations – arguably a more “human-like” approach to search, as originally proposed by Shannon (27). Figure 2 shows the scalability of each player with respect to thinking time, measured
on an Elo scale, relative to Stockfish or Elmo with 40ms thinking time. AlphaZero’s MCTS
scaled more effectively with thinking time than either Stockfish or Elmo, calling into question
the widely held belief (4, 11) that alpha-beta search is inherently superior in these domains.3
Finally, we analysed the chess knowledge discovered by AlphaZero. Table 2 analyses the
most common human openings (those played more than 100,000 times in an online database
of human chess games (1)). Each of these openings is independently discovered and played
frequently by AlphaZero during self-play training. When starting from each human opening,
AlphaZero convincingly defeated Stockfish, suggesting that it has indeed mastered a wide spectrum of chess play.
The game of chess represented the pinnacle of AI research over several decades. State-ofthe-art programs are based on powerful engines that search many millions of positions, leveraging handcrafted domain expertise and sophisticated domain adaptations. AlphaZero is a generic
reinforcement learning algorithm – originally devised for the game of Go – that achieved superior results within a few hours, searching a thousand times fewer positions, given no domain
3
The prevalence of draws in high-level chess tends to compress the Elo scale, compared to shogi or Go.
5
A10: English Opening
D06: Queens Gambit
rmblkans
7
opopopop
6
0Z0Z0Z0Z
5
Z0Z0Z0Z0
4
0ZPZ0Z0Z
3
Z0Z0Z0Z0
2
PO0OPOPO
1
SNAQJBMR
rmblkans
7
opo0opop
6
0Z0Z0Z0Z
5
Z0ZpZ0Z0
4
0ZPO0Z0Z
3
Z0Z0Z0Z0
2
PO0ZPOPO
1
SNAQJBMR
8
a
b
c
d
e
f
g
8
h
w 20/30/0, b 8/40/2
a
1...e5 g3 d5 cxd5 Nf6 Bg2 Nxd5 Nf3
b
c
d
e
f
g
h
w 16/34/0, b 1/47/2
A46: Queens Pawn Game
rmblka0s
opopopop
6
0Z0Z0m0Z
5
Z0Z0Z0Z0
4
0Z0O0Z0Z
3
Z0Z0ZNZ0
2
POPZPOPO
1
SNAQJBZR
rmblka0s
opopZpop
0Z0Zpm0Z
5
Z0Z0Z0Z0
4
0ZPO0Z0Z
3
Z0Z0Z0Z0
2
PO0ZPOPO
1
SNAQJBMR
8
8
7
7
a
b
c
d
e
f
g
6
h
w 24/26/0, b 3/47/0
a
2...d5 c4 e6 Nc3 Be7 Bf4 O-O e3
b
c
d
e
f
g
h
w 17/33/0, b 5/44/1
E61: Kings Indian Defence
rmblka0s
opopopZp
6
0Z0Z0mpZ
5
Z0Z0Z0Z0
4
0ZPO0Z0Z
3
Z0M0Z0Z0
2
PO0ZPOPO
1
S0AQJBMR
rmblkans
opo0Zpop
0Z0ZpZ0Z
5
Z0ZpZ0Z0
4
0Z0OPZ0Z
3
Z0Z0Z0Z0
2
POPZ0OPO
1
SNAQJBMR
8
7
7
a
b
c
d
e
f
g
6
h
a
3...d5 cxd5 Nxd5 e4 Nxc3 bxc3 Bg7 Be3
b
c
d
e
f
g
h
w 39/11/0, b 4/46/0
B50: Sicilian Defence
rmblkans
opZ0opop
0Z0o0Z0Z
5
Z0o0Z0Z0
4
0Z0ZPZ0Z
3
Z0Z0ZNZ0
2
POPO0OPO
1
SNAQJBZR
rZblkans
opZpopop
0ZnZ0Z0Z
5
Z0o0Z0Z0
4
0Z0ZPZ0Z
3
Z0Z0ZNZ0
2
POPO0OPO
1
SNAQJBZR
8
7
7
6
6
a
b
c
d
e
f
g
h
a
3.d4 cxd4 Nxd4 Nf6 Nc3 a6 f3 e5
b
c
d
e
f
B40: Sicilian Defence
rmblkans
opZpZpop
0Z0ZpZ0Z
5
Z0o0Z0Z0
4
0Z0ZPZ0Z
3
Z0Z0ZNZ0
2
POPO0OPO
1
SNAQJBZR
7
6
6
c
d
e
f
g
h
a
3.d4 cxd4 Nxd4 Nc6 Nc3 Qc7 Be3 a6
b
c
d
e
f
g
h
w 27/22/1, b 6/44/0
B10: Caro-Kann Defence
rmblkans
opZpopop
0ZpZ0Z0Z
5
Z0Z0Z0Z0
4
0Z0ZPZ0Z
3
Z0Z0Z0Z0
2
POPO0OPO
1
SNAQJBMR
rmblka0s
opopopop
0Z0Z0m0Z
5
Z0Z0Z0Z0
4
0Z0Z0Z0Z
3
Z0Z0ZNZ0
2
POPOPOPO
1
SNAQJBZR
8
7
7
6
6
a
b
c
d
e
f
g
4.Ba4 Be7 O-O Nf6 Re1 b5 Bb3 O-O
A05: Reti Opening
8
h
w 25/25/0, b 4/45/1
3.Bb5 e6 O-O Ne7 Re1 a6 Bf1 d5
rZblkans
ZpopZpop
pZnZ0Z0Z
5
ZBZ0o0Z0
4
0Z0ZPZ0Z
3
Z0Z0ZNZ0
2
POPO0OPO
1
SNAQJ0ZR
7
b
h
C60: Ruy Lopez (Spanish Opening)
8
a
g
w 11/39/0, b 3/46/1
8
w 17/31/2, b 3/40/7
3.Nc3 Nf6 e5 Nd7 f4 c5 Nf3 Be7
B30: Sicilian Defence
8
w 17/32/1, b 4/43/3
3.Nf3 d5 Nc3 Bb4 Bg5 h6 Qa4 Nc6
C00: French Defence
8
w 16/34/0, b 0/48/2
2...c6 Nc3 Nf6 Nf3 a6 g3 c4 a4
E00: Queens Pawn Game
a
2.d4 d5 e5 Bf5 Nf3 e6 Be2 a6
b
c
d
e
f
g
h
w 13/36/1, b 7/43/0
Total games: w 242/353/5, b 48/533/19
2.c4 e6 d4 d5 Nc3 Be7 Bf4 O-O
Overall percentage: w 40.3/58.8/0.8, b 8.0/88.8/3.2
Table 2: Analysis of the 12 most popular human openings (played more than 100,000 times
in an online database (1)). Each opening is labelled by its ECO code and common name. The
plot shows the proportion of self-play training games in which AlphaZero played each opening,
against training time. We also report the win/draw/loss results of 100 game AlphaZero vs.
Stockfish matches starting from each opening, as either white (w) or black (b), from AlphaZero’s
perspective. Finally, the principal variation (PV) of AlphaZero is provided from each opening.
6
Figure 2: Scalability of AlphaZero with thinking time, measured on an Elo scale. a Performance of AlphaZero and Stockfish in chess, plotted against thinking time per move. b Performance of AlphaZero and Elmo in shogi, plotted against thinking time per move.
knowledge except the rules of chess. Furthermore, the same algorithm was applied without
modification to the more challenging game of shogi, again outperforming the state of the art
within a few hours.
References
1. Online chess games database, 365chess, 2017. URL: https://www.365chess.com/.
2. Victor Allis. Searching for Solutions in Games and Artificial Intelligence. PhD thesis,
University of Limburg, Netherlands, 1994.
3. Thomas Anthony, Zheng Tian, and David Barber. Thinking fast and slow with deep learning and tree search. In Advances in Neural Information Processing Systems 30: Annual
Conference on Neural Information Processing Systems 2017, 4-9 December 2017, Long
Beach, CA, USA, pages 5366–5376, 2017.
4. Oleg Arenz. Monte Carlo chess. Master’s thesis, Technische Universitat Darmstadt, 2012.
5. Computer Shogi Association. Results of the 27th world computer shogi championship. http://www2.computer-shogi.org/wcsc27/index_e.html. Retrieved November 29th, 2017.
6. J. Baxter, A. Tridgell, and L. Weaver. Learning to play chess using temporal differences.
Machine Learning, 40(3):243–263, 2000.
7. Donald F. Beal and Martin C. Smith. Temporal difference learning for heuristic search and
game playing. Inf. Sci., 122(1):3–21, 2000.
7
8. Donald F. Beal and Martin C. Smith. Temporal difference learning applied to game playing
and the results of application to shogi. Theoretical Computer Science, 252(1–2):105–119,
2001.
9. M. Campbell, A. J. Hoane, and F. Hsu. Deep Blue. Artificial Intelligence, 134:57–83, 2002.
10. R. Coulom. Whole-history rating: A Bayesian rating system for players of time-varying
strength. In International Conference on Computers and Games, pages 113–124, 2008.
11. Omid E David, Nathan S Netanyahu, and Lior Wolf. Deepchess: End-to-end deep neural
network for automatic learning in chess. In International Conference on Artificial Neural
Networks, pages 88–96. Springer, 2016.
12. Kunihito Hoki and Tomoyuki Kaneko. Large-scale optimization for evaluation functions
with minimax search. Journal of Artificial Intelligence Research (JAIR), 49:527–568, 2014.
13. Feng-hsiung Hsu. Behind Deep Blue: Building the Computer that Defeated the World
Chess Champion. Princeton University Press, 2002.
14. Hiroyuki Iida, Makoto Sakuta, and Jeff Rollason. Computer shogi. Artificial Intelligence,
134:121–144, 2002.
15. Norman P. Jouppi, Cliff Young, Nishant Patil, et al. In-datacenter performance analysis of
a tensor processing unit. In Proceedings of the 44th Annual International Symposium on
Computer Architecture, ISCA ’17, pages 1–12. ACM, 2017.
16. Tomoyuki Kaneko and Kunihito Hoki. Analysis of evaluation-function learning by comparison of sibling nodes. In Advances in Computer Games - 13th International Conference,
ACG 2011, Tilburg, The Netherlands, November 20-22, 2011, Revised Selected Papers,
pages 158–169, 2011.
17. John Knudsen. Essential Chess Quotations. iUniverse, 2000.
18. D. E. Knuth and R. W Moore. An analysis of alphabeta pruning. Artificial Intelligence,
6(4):293–326, 1975.
19. Matthew Lai. Giraffe: Using deep reinforcement learning to play chess. Master’s thesis,
Imperial College London, 2015.
20. Emanuel Lasker. Common Sense in Chess. Dover Publications, 1965.
21. David N. L. Levy and Monty Newborn. How Computers Play Chess. Ishi Press, 2009.
22. Chris J. Maddison, Aja Huang, Ilya Sutskever, and David Silver. Move evaluation in Go
using deep convolutional neural networks. In International Conference on Learning Representations, 2015.
8
23. Tony Marsland. Computer chess methods. In S. Shapiro, editor, Encyclopedia of Artificial
Intelligence. John Wiley & sons, New York, 1987.
24. Raghuram Ramanujan, Ashish Sabharwal, and Bart Selman. Understanding sampling style
adversarial search methods. In Proceedings of the 26th Conference on Uncertainty in Artificial Intelligence (UAI), 2010.
25. Tord Romstad, Marco Costalba, Joona Kiiski, et al. Stockfish: A strong open source chess
engine. https://stockfishchess.org/. Retrieved November 29th, 2017.
26. A. L. Samuel. Some studies in machine learning using the game of checkers II - recent
progress. IBM Journal of Research and Development, 11(6):601–617, 1967.
27. Claude E Shannon. Xxii. programming a computer for playing chess. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 41(314):256–275,
1950.
28. David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den
Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot,
Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of Go with deep neural networks and tree search. Nature,
529(7587):484–489, January 2016.
29. David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang,
Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, Yutian Chen,
Timothy Lillicrap, Fan Hui, Laurent Sifre, George van den Driessche, Thore Graepel, and
Demis Hassabis. Mastering the game of go without human knowledge. Nature, 550:354–
359, 2017.
30. Wilhelm Steinitz. The Modern Chess Instructor. Edition Olms AG, 1990.
31. Sebastian Thrun. Learning to play the game of chess. In Advances in neural information
processing systems, pages 1069–1076, 1995.
32. J. Veness, D. Silver, A. Blair, and W. Uther. Bootstrapping from game tree search. In
Advances in Neural Information Processing Systems, pages 1937–1945, 2009.
9
Methods
Anatomy of a Computer Chess Program
In this section we describe the components of a typical computer chess program, focusing
specifically on Stockfish (25), an open source program that won the 2016 TCEC computer chess
championship. For an overview of standard methods, see (23).
Each position s is described by a sparse vector of handcrafted features φ(s), including
midgame/endgame-specific material point values, material imbalance tables, piece-square tables, mobility and trapped pieces, pawn structure, king safety, outposts, bishop pair, and other
miscellaneous evaluation patterns. Each feature φi is assigned, by a combination of manual and
automatic tuning, a corresponding weight wi and the position is evaluated by a linear combination v(s, w) = φ(s)> w. However, this raw evaluation is only considered accurate for positions that are “quiet”, with no unresolved captures or checks. A domain-specialised quiescence
search is used to resolve ongoing tactical situations before the evaluation function is applied.
The final evaluation of a position s is computed by a minimax search that evaluates each leaf
using a quiescence search. Alpha-beta pruning is used to safely cut any branch that is provably
dominated by another variation. Additional cuts are achieved using aspiration windows and
principal variation search. Other pruning strategies include null move pruning (which assumes
a pass move should be worse than any variation, in positions that are unlikely to be in zugzwang,
as determined by simple heuristics), futility pruning (which assumes knowledge of the maximum possible change in evaluation), and other domain-dependent pruning rules (which assume
knowledge of the value of captured pieces).
The search is focused on promising variations both by extending the search depth of promising variations, and by reducing the search depth of unpromising variations based on heuristics
like history, static-exchange evaluation (SEE), and moving piece type. Extensions are based on
domain-independent rules that identify singular moves with no sensible alternative, and domaindependent rules, such as extending check moves. Reductions, such as late move reductions, are
based heavily on domain knowledge.
The efficiency of alpha-beta search depends critically upon the order in which moves are
considered. Moves are therefore ordered by iterative deepening (using a shallower search to
order moves for a deeper search). In addition, a combination of domain-independent move
ordering heuristics, such as killer heuristic, history heuristic, counter-move heuristic, and also
domain-dependent knowledge based on captures (SEE) and potential captures (MVV/LVA).
A transposition table facilitates the reuse of values and move orders when the same position
is reached by multiple paths. A carefully tuned opening book is used to select moves at the
start of the game. An endgame tablebase, precalculated by exhaustive retrograde analysis of
endgame positions, provides the optimal move in all positions with six and sometimes seven
pieces or less.
Other strong chess programs, and also earlier programs such as Deep Blue, have used very
similar architectures (9,23) including the majority of the components described above, although
10
important details vary considerably.
None of the techniques described in this section are used by AlphaZero. It is likely that
some of these techniques could further improve the performance of AlphaZero; however, we
have focused on a pure self-play reinforcement learning approach and leave these extensions
for future research.
Prior Work on Computer Chess and Shogi
In this section we discuss some notable prior work on reinforcement learning in computer chess.
NeuroChess (31) evaluated positions by a neural network that used 175 handcrafted input
features. It was trained by temporal-difference learning to predict the final game outcome, and
also the expected features after two moves. NeuroChess won 13% of games against GnuChess
using a fixed depth 2 search.
Beal and Smith applied temporal-difference learning to estimate the piece values in chess (7)
and shogi (8), starting from random values and learning solely by self-play.
KnightCap (6) evaluated positions by a neural network that used an attack-table based on
knowledge of which squares are attacked or defended by which pieces. It was trained by a
variant of temporal-difference learning, known as TD(leaf), that updates the leaf value of the
principal variation of an alpha-beta search. KnightCap achieved human master level after training against a strong computer opponent with hand-initialised piece-value weights.
Meep (32) evaluated positions by a linear evaluation function based on handcrafted features.
It was trained by another variant of temporal-difference learning, known as TreeStrap, that
updated all nodes of an alpha-beta search. Meep defeated human international master players
in 13 out of 15 games, after training by self-play with randomly initialised weights.
Kaneko and Hoki (16) trained the weights of a shogi evaluation function comprising a million features, by learning to select expert human moves during alpha-beta serach. They also performed a large-scale optimization based on minimax search regulated by expert game logs (12);
this formed part of the Bonanza engine that won the 2013 World Computer Shogi Championship.
Giraffe (19) evaluated positions by a neural network that included mobility maps and attack
and defend maps describing the lowest valued attacker and defender of each square. It was
trained by self-play using TD(leaf), also reaching a standard of play comparable to international
masters.
DeepChess (11) trained a neural network to performed pair-wise evaluations of positions. It
was trained by supervised learning from a database of human expert games that was pre-filtered
to avoid capture moves and drawn games. DeepChess reached a strong grandmaster level of
play.
All of these programs combined their learned evaluation functions with an alpha-beta search
enhanced by a variety of extensions.
An approach based on training dual policy and value networks using AlphaZero-like policy
iteration was successfully applied to improve on the state-of-the-art in Hex (3).
11
MCTS and Alpha-Beta Search
For at least four decades the strongest computer chess programs have used alpha-beta search
(18, 23). AlphaZero uses a markedly different approach that averages over the position evaluations within a subtree, rather than computing the minimax evaluation of that subtree. However, chess programs using traditional MCTS were much weaker than alpha-beta search programs, (4, 24); while alpha-beta programs based on neural networks have previously been unable to compete with faster, handcrafted evaluation functions.
AlphaZero evaluates positions using non-linear function approximation based on a deep
neural network, rather than the linear function approximation used in typical chess programs.
This provides a much more powerful representation, but may also introduce spurious approximation errors. MCTS averages over these approximation errors, which therefore tend to cancel
out when evaluating a large subtree. In contrast, alpha-beta search computes an explicit minimax, which propagates the biggest approximation errors to the root of the subtree. Using MCTS
may allow AlphaZero to effectively combine its neural network representations with a powerful,
domain-independent search.
Domain Knowledge
1. The input features describing the position, and the output features describing the move,
are structured as a set of planes; i.e. the neural network architecture is matched to the
grid-structure of the board.
2. AlphaZero is provided with perfect knowledge of the game rules. These are used during
MCTS, to simulate the positions resulting from a sequence of moves, to determine game
termination, and to score any simulations that reach a terminal state.
3. Knowledge of the rules is also used to encode the input planes (i.e. castling, repetition,
no-progress) and output planes (how pieces move, promotions, and piece drops in shogi).
4. The typical number of legal moves is used to scale the exploration noise (see below).
5. Chess and shogi games exceeding a maximum number of steps (determined by typical
game length) were terminated and assigned a drawn outcome; Go games were terminated
and scored with Tromp-Taylor rules, similarly to previous work (29).
AlphaZero did not use any form of domain knowledge beyond the points listed above.
Representation
In this section we describe the representation of the board inputs, and the representation of the
action outputs, used by the neural network in AlphaZero. Other representations could have been
used; in our experiments the training algorithm worked robustly for many reasonable choices.
12
Go
Feature
Chess
Planes
Feature
Shogi
Planes
Feature
Planes
P1 stone
P2 stone
1
1
P1 piece
P2 piece
Repetitions
6
6
2
P1 piece
P2 piece
Repetitions
P1 prisoner count
P2 prisoner count
14
14
3
7
7
Colour
1
Colour
Total move count
P1 castling
P2 castling
No-progress count
1
1
2
2
1
Colour
Total move count
1
1
Total
17
Total
119
Total
362
Table S1: Input features used by AlphaZero in Go, Chess and Shogi respectively. The first set
of features are repeated for each position in a T = 8-step history. Counts are represented by
a single real-valued input; other input features are represented by a one-hot encoding using the
specified number of binary input planes. The current player is denoted by P1 and the opponent
by P2.
The input to the neural network is an N × N × (M T + L) image stack that represents state
using a concatenation of T sets of M planes of size N × N . Each set of planes represents the
board position at a time-step t − T + 1, ..., t, and is set to zero for time-steps less than 1. The
board is oriented to the perspective of the current player. The M feature planes are composed
of binary feature planes indicating the presence of the player’s pieces, with one plane for each
piece type, and a second set of planes indicating the presence of the opponent’s pieces. For shogi
there are additional planes indicating the number of captured prisoners of each type. There are
an additional L constant-valued input planes denoting the player’s colour, the total move count,
and the state of special rules: the legality of castling in chess (kingside or queenside); the
repetition count for that position (3 repetitions is an automatic draw in chess; 4 in shogi); and
the number of moves without progress in chess (50 moves without progress is an automatic
draw). Input features are summarised in Table S1.
A move in chess may be described in two parts: selecting the piece to move, and then
selecting among the legal moves for that piece. We represent the policy π(a|s) by a 8 × 8 × 73
stack of planes encoding a probability distribution over 4,672 possible moves. Each of the 8 × 8
positions identifies the square from which to “pick up” a piece. The first 56 planes encode
possible ‘queen moves’ for any piece: a number of squares [1..7] in which the piece will be
moved, along one of eight relative compass directions {N, N E, E, SE, S, SW, W, N W }. The
next 8 planes encode possible knight moves for that piece. The final 9 planes encode possible
13
Chess
Feature
Shogi
Feature
Planes
Queen moves
Knight moves
Underpromotions
56
8
9
Queen moves
Knight moves
Promoting queen moves
Promoting knight moves
Drop
Total
73
Total
Planes
64
2
64
2
7
139
Table S2: Action representation used by AlphaZero in Chess and Shogi respectively. The policy
is represented by a stack of planes encoding a probability distribution over legal moves; planes
correspond to the entries in the table.
underpromotions for pawn moves or captures in two possible diagonals, to knight, bishop or
rook respectively. Other pawn moves or captures from the seventh rank are promoted to a
queen.
The policy in shogi is represented by a 9 × 9 × 139 stack of planes similarly encoding a
probability distribution over 11,259 possible moves. The first 64 planes encode ‘queen moves’
and the next 2 moves encode knight moves. An additional 64 + 2 planes encode promoting
queen moves and promoting knight moves respectively. The last 7 planes encode a captured
piece dropped back into the board at that location.
The policy in Go is represented identically to AlphaGo Zero (29), using a flat distribution
over 19 × 19 + 1 moves representing possible stone placements and the pass move. We also
tried using a flat distribution over moves for chess and shogi; the final result was almost identical
although training was slightly slower.
The action representations are summarised in Table S2. Illegal moves are masked out by
setting their probabilities to zero, and re-normalising the probabilities for remaining moves.
Configuration
During training, each MCTS used 800 simulations. The number of games, positions, and thinking time varied per game due largely to different board sizes and game lengths, and are shown
in Table S3. The learning rate was set to 0.2 for each game, and was dropped three times (to
0.02, 0.002 and 0.0002 respectively) during the course of training. Moves are selected in proportion to the root visit count. Dirichlet noise Dir(α) was added to the prior probabilities in the
root node; this was scaled in inverse proportion to the approximate number of legal moves in a
typical position, to a value of α = {0.3, 0.15, 0.03} for chess, shogi and Go respectively. Unless
otherwise specified, the training and search algorithm and parameters are identical to AlphaGo
Zero (29).
14
Mini-batches
Training Time
Training Games
Thinking Time
Chess
Shogi
Go
700k
9h
44 million
800 sims
40 ms
700k
12h
24 million
800 sims
80 ms
700k
34h
21 million
800 sims
200 ms
Table S3: Selected statistics of AlphaZero training in Chess, Shogi and Go.
During evaluation, AlphaZero selects moves greedily with respect to the root visit count.
Each MCTS was executed on a single machine with 4 TPUs.
Evaluation
To evaluate performance in chess, we used Stockfish version 8 (official Linux release) as a
baseline program, using 64 CPU threads and a hash size of 1GB.
To evaluate performance in shogi, we used Elmo version WCSC27 in combination with
YaneuraOu 2017 Early KPPT 4.73 64AVX2 with 64 CPU threads and a hash size of 1GB with
the usi option of EnteringKingRule set to NoEnteringKing.
We evaluated the relative strength of AlphaZero (Figure 1) by measuring the Elo rating of
each player. We estimate the probability that player a will defeat player b by a logistic function
p(a defeats b) = 1+exp (celo1(e(b)−e(a)) , and estimate the ratings e(·) by Bayesian logistic regression, computed by the BayesElo program (10) using the standard constant celo = 1/400. Elo
ratings were computed from the results of a 1 second per move tournament between iterations
of AlphaZero during training, and also a baseline player: either Stockfish, Elmo or AlphaGo
Lee respectively. The Elo rating of the baseline players was anchored to publicly available
values (29).
We also measured the head-to-head performance of AlphaZero against each baseline player.
Settings were chosen to correspond with computer chess tournament conditions: each player
was allowed 1 minute per move, resignation was enabled for all players (-900 centipawns for 10
consecutive moves for Stockfish and Elmo, 5% winrate for AlphaZero). Pondering was disabled
for all players.
Example games
In this section we include 10 example games played by AlphaZero against Stockfish during the
100 game match using 1 minute per move.
15
White: Stockfish Black: AlphaZero
1. e4 e5 2. Nf3 Nc6 3. Bb5 Nf6 4. d3 Bc5 5. Bxc6 dxc6 6. 0-0 Nd7 7. Nbd2 0-0 8. Qe1 f6 9. Nc4 Rf7
10. a4 Bf8 11. Kh1 Nc5 12. a5 Ne6 13. Ncxe5 fxe5 14. Nxe5 Rf6 15. Ng4 Rf7 16. Ne5 Re7 17. a6 c5
18. f4 Qe8 19. axb7 Bxb7 20. Qa5 Nd4 21. Qc3 Re6 22. Be3 Rb6 23. Nc4 Rb4 24. b3 a5 25. Rxa5
Rxa5 26. Nxa5 Ba6 27. Bxd4 Rxd4 28. Nc4 Rd8 29. g3 h6 30. Qa5 Bc8 31. Qxc7 Bh3 32. Rg1 Rd7
33. Qe5 Qxe5 34. Nxe5 Ra7 35. Nc4 g5 36. Rc1 Bg7 37. Ne5 Ra8 38. Nf3 Bb2 39. Rb1 Bc3 40. Ng1
Bd7 41. Ne2 Bd2 42. Rd1 Be3 43. Kg2 Bg4 44. Re1 Bd2 45. Rf1 Ra2 46. h3 Bxe2 47. Rf2 Bxf4 48.
Rxe2 Be5 49. Rf2 Kg7 50. g4 Bd4 51. Re2 Kf6 52. e5+ Bxe5 53. Kf3 Ra1 54. Rf2 Re1 55. Kg2+ Bf4
56. c3 Rc1 57. d4 Rxc3 58. dxc5 Rxc5 59. b4 Rc3 60. h4 Ke5 61. hxg5 hxg5 62. Re2+ Kf6 63. Kf2
Be5 64. Ra2 Rc4 65. Ra6+ Ke7 66. Ra5 Ke6 67. Ra6+ Bd6 0-1
White: Stockfish Black: AlphaZero
1. e4 e5 2. Nf3 Nc6 3. Bb5 Nf6 4. d3 Bc5 5. Bxc6 dxc6 6. 0-0 Nd7 7. c3 0-0 8. d4 Bd6 9. Bg5 Qe8 10.
Re1 f6 11. Bh4 Qf7 12. Nbd2 a5 13. Bg3 Re8 14. Qc2 Nf8 15. c4 c5 16. d5 b6 17. Nh4 g6 18. Nhf3
Bd7 19. Rad1 Re7 20. h3 Qg7 21. Qc3 Rae8 22. a3 h6 23. Bh4 Rf7 24. Bg3 Rfe7 25. Bh4 Rf7 26. Bg3
a4 27. Kh1 Rfe7 28. Bh4 Rf7 29. Bg3 Rfe7 30. Bh4 g5 31. Bg3 Ng6 32. Nf1 Rf7 33. Ne3 Ne7 34. Qd3
h5 35. h4 Nc8 36. Re2 g4 37. Nd2 Qh7 38. Kg1 Bf8 39. Nb1 Nd6 40. Nc3 Bh6 41. Rf1 Ra8 42. Kh2
Kf8 43. Kg1 Qg6 44. f4 gxf3 45. Rxf3 Bxe3+ 46. Rfxe3 Ke7 47. Be1 Qh7 48. Rg3 Rg7 49. Rxg7+
Qxg7 50. Re3 Rg8 51. Rg3 Qh8 52. Nb1 Rxg3 53. Bxg3 Qh6 54. Nd2 Bg4 55. Kh2 Kd7 56. b3 axb3
57. Nxb3 Qg6 58. Nd2 Bd1 59. Nf3 Ba4 60. Nd2 Ke7 61. Bf2 Qg4 62. Qf3 Bd1 63. Qxg4 Bxg4 64. a4
Nb7 65. Nb1 Na5 66. Be3 Nxc4 67. Bc1 Bd7 68. Nc3 c6 69. Kg1 cxd5 70. exd5 Bf5 71. Kf2 Nd6 72.
Be3 Ne4+ 73. Nxe4 Bxe4 74. a5 bxa5 75. Bxc5+ Kd7 76. d6 Bf5 77. Ba3 Kc6 78. Ke1 Kd5 79. Kd2
Ke4 80. Bb2 Kf4 81. Bc1 Kg3 82. Ke2 a4 83. Kf1 Kxh4 84. Kf2 Kg4 85. Ba3 Bd7 86. Bc1 Kf5 87.
Ke3 Ke6 0-1
White: AlphaZero Black: Stockfish
1. Nf3 Nf6 2. c4 b6 3. d4 e6 4. g3 Ba6 5. Qc2 c5 6. d5 exd5 7. cxd5 Bb7 8. Bg2 Nxd5 9. 0-0 Nc6 10.
Rd1 Be7 11. Qf5 Nf6 12. e4 g6 13. Qf4 0-0 14. e5 Nh5 15. Qg4 Re8 16. Nc3 Qb8 17. Nd5 Bf8 18.
Bf4 Qc8 19. h3 Ne7 20. Ne3 Bc6 21. Rd6 Ng7 22. Rf6 Qb7 23. Bh6 Nd5 24. Nxd5 Bxd5 25. Rd1 Ne6
26. Bxf8 Rxf8 27. Qh4 Bc6 28. Qh6 Rae8 29. Rd6 Bxf3 30. Bxf3 Qa6 31. h4 Qa5 32. Rd1 c4 33. Rd5
Qe1+ 34. Kg2 c3 35. bxc3 Qxc3 36. h5 Re7 37. Bd1 Qe1 38. Bb3 Rd8 39. Rf3 Qe4 40. Qd2 Qg4 41.
Bd1 Qe4 42. h6 Nc7 43. Rd6 Ne6 44. Bb3 Qxe5 45. Rd5 Qh8 46. Qb4 Nc5 47. Rxc5 bxc5 48. Qh4
Rde8 49. Rf6 Rf8 50. Qf4 a5 51. g4 d5 52. Bxd5 Rd7 53. Bc4 a4 54. g5 a3 55. Qf3 Rc7 56. Qxa3 Qxf6
57. gxf6 Rfc8 58. Qd3 Rf8 59. Qd6 Rfc8 60. a4 1-0
16
White: AlphaZero Black: Stockfish
1. d4 e6 2. Nc3 Nf6 3. e4 d5 4. e5 Nfd7 5. f4 c5 6. Nf3 Nc6 7. Be3 Be7 8. Qd2 a6 9. Bd3 c4 10. Be2 b5
11. a3 Rb8 12. 0-0 0-0 13. f5 a5 14. fxe6 fxe6 15. Bd1 b4 16. axb4 axb4 17. Ne2 c3 18. bxc3 Nb6 19.
Qe1 Nc4 20. Bc1 bxc3 21. Qxc3 Qb6 22. Kh1 Nb2 23. Nf4 Nxd1 24. Rxd1 Bd7 25. h4 Ra8 26. Bd2
Rfb8 27. h5 Rxa1 28. Rxa1 Qb2 29. Qxb2 Rxb2 30. c3 Rb3 31. Ra8+ Rb8 32. Ra2 Rb3 33. g4 Ra3 34.
Rb2 Kf7 35. Kg2 Bc8 36. Rb6 Ra6 37. Rb1 Ke8 38. Kg3 h6 39. Ng6 Ra3 40. Rb6 Bd7 41. g5 hxg5 42.
Kg4 Bd8 43. Rb2 Bc8 44. Nxg5 Ra1 45. Nf3 Ra3 46. Be1 Ba5 47. Rf2 Ra1 48. Bd2 Bd8 49. Rh2 Ne7
50. Bg5 Nf5 51. Bxd8 Kxd8 52. Rb2 Rc1 53. Ngh4 Nxh4 54. Nxh4 Bd7 55. Rb8+ Bc8 56. Ng2 Rxc3
57. Nf4 Rc1 58. Ra8 Kd7 59. Kf3 Rc3+ 60. Kf2 Ke7 61. Kg2 Kf7 62. Ng6 Ke8 63. Ra1 Rc7 64. Kh3
Rf7 65. Kg4 Kd8 66. Nf4 Bd7 67. Ra7 Kc8 68. Kg3 Re7 69. Nd3 Kb8 70. Ra6 Bc8 71. Rb6+ Kc7 72.
Rd6 Kb8 73. Nc5 g6 74. h6 Rh7 75. Nxe6 Rxh6 76. Nf4 Rh1 77. Nxd5 Rh3+ 78. Kf4 Rh4+ 79. Ke3
Rh3+ 80. Kd2 Bf5 81. Ne7 Rh2+ 82. Ke3 Bh3 83. Nxg6 Rh1 84. Nf4 Bg4 85. Rf6 Kc7 86. Nd3 Bd7
87. d5 Bb5 88. Nf4 Ba4 89. Kd4 Be8 90. Rf8 Rd1+ 91. Kc5 Rc1+ 92. Kb4 Rb1+ 93. Kc3 Bb5 94. Kd4
Ba6 95. Rf7+ 1-0
White: AlphaZero Black: Stockfish
1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb7 5. Bg2 Be7 6. 0-0 0-0 7. d5 exd5 8. Nh4 c6 9. cxd5 Nxd5 10.
Nf5 Nc7 11. e4 Bf6 12. Nd6 Ba6 13. Re1 Ne8 14. e5 Nxd6 15. exf6 Qxf6 16. Nc3 Nb7 17. Ne4 Qg6
18. h4 h6 19. h5 Qh7 20. Qg4 Kh8 21. Bg5 f5 22. Qf4 Nc5 23. Be7 Nd3 24. Qd6 Nxe1 25. Rxe1 fxe4
26. Bxe4 Rf5 27. Bh4 Bc4 28. g4 Rd5 29. Bxd5 Bxd5 30. Re8+ Bg8 31. Bg3 c5 32. Qd5 d6 33. Qxa8
Nd7 34. Qe4 Nf6 35. Qxh7+ Kxh7 36. Re7 Nxg4 37. Rxa7 Nf6 38. Bxd6 Be6 39. Be5 Nd7 40. Bc3 g6
41. Bd2 gxh5 42. a3 Kg6 43. Bf4 Kf5 44. Bc7 h4 45. Ra8 h5 46. Rh8 Kg6 47. Rd8 Kf7 48. f3 Bf5 49.
Bh2 h3 50. Rh8 Kg6 51. Re8 Kf7 52. Re1 Be6 53. Bc7 b5 54. Kh2 Kf6 55. Re3 Ke7 56. Re4 Kf7 57.
Bd6 Kf6 58. Kg3 Kf7 59. Kf2 Bf5 60. Re1 Kg6 61. Kg1 c4 62. Kh2 h4 63. Be7 Nb6 64. Bxh4 Na4 65.
Re2 Nc5 66. Re5 Nb3 67. Rd5 Be6 68. Rd6 Kf5 69. Be1 Ke5 70. Rb6 Bd7 71. Kg3 Nc1 72. Rh6 Kd5
73. Bc3 Bf5 74. Rh5 Ke6 75. Kf2 Nd3+ 76. Kg1 Nf4 77. Rh6+ Ke7 78. Kh2 Nd5 79. Kg3 Be6 80. Rh5
Ke8 81. Re5 Kf7 82. Bd2 Ne7 83. Bb4 Nd5 84. Bc3 Ke7 85. Bd2 Kf6 86. f4 Ne7 87. Rxb5 Nf5+ 88.
Kh2 Ke7 89. Ra5 Nh4 90. Bb4+ Kf7 91. Rh5 Nf3+ 92. Kg3 Kg6 93. Rh8 Nd4 94. Bc3 Nf5+ 95. Kxh3
Bd7 96. Kh2 Kf7 97. Rb8 Ke6 98. Kg1 Bc6 99. Rb6 Kd5 100. Kf2 Bd7 101. Ke1 Ke4 102. Bd2 Kd5
103. Rf6 Nd6 104. Rh6 Nf5 105. Rh8 Ke4 106. Rh7 Bc8 107. Rc7 Ba6 108. Rc6 Bb5 109. Rc5 Bd7
110. Rxc4+ Kd5 111. Rc7 Kd6 112. Rc3 Ke6 113. Rc5 Nd4 114. Be3 Nf5 115. Bf2 Nd6 116. Rc3 Ne4
117. Rd3 1-0
White: AlphaZero Black: Stockfish
1. d4 Nf6 2. Nf3 e6 3. c4 b6 4. g3 Be7 5. Bg2 Bb7 6. 0-0 0-0 7. d5 exd5 8. Nh4 c6 9. cxd5 Nxd5 10.
Nf5 Nc7 11. e4 Bf6 12. Nd6 Ba6 13. Re1 Ne8 14. e5 Nxd6 15. exf6 Qxf6 16. Nc3 Bc4 17. h4 h6 18.
b3 Qxc3 19. Bf4 Nb7 20. bxc4 Qf6 21. Be4 Na6 22. Be5 Qe6 23. Bd3 f6 24. Bd4 Qf7 25. Qg4 Rfd8
26. Re3 Nac5 27. Bg6 Qf8 28. Rd1 Rab8 29. Kg2 Ne6 30. Bc3 Nbc5 31. Rde1 Na4 32. Bd2 Kh8 33.
f4 Qd6 34. Bc1 Nd4 35. Re7 f5 36. Bxf5 Nxf5 37. Qxf5 Rf8 38. Rxd7 Rxf5 39. Rxd6 Rf7 40. g4 Kg8
41. g5 hxg5 42. hxg5 Nc5 43. Kf3 Nb7 44. Rdd1 Na5 45. Re4 c5 46. Bb2 Nc6 47. g6 Rc7 48. Kg4
Nd4 49. Rd2 Rf8 50. Bxd4 cxd4 51. Rdxd4 Rfc8 52. Kg5 Rf8 53. Rd2 Rc6 54. Rd5 Rc7 55. f5 Rb7 56.
a3 Rc7 57. a4 a6 58. Red4 Rcc8 59. Re5 Rc7 60. a5 Rc5 61. Rxc5 bxc5 62. Rd6 Ra8 63. Re6 Kf8 64.
Rc6 Ke7 65. Kf4 Kd7 66. Rxc5 Rh8 67. Rd5+ Ke7 68. Re5+ Kd7 69. Re6 Rh4+ 70. Kg5 1-0
17
White: AlphaZero Black: Stockfish
1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb7 5. Bg2 Bb4+ 6. Bd2 Bxd2+ 7. Qxd2 d5 8. 0-0 0-0 9. cxd5 exd5
10. Nc3 Nbd7 11. b4 c6 12. Qb2 a5 13. b5 c5 14. Rac1 Qe7 15. Na4 Rab8 16. Rfd1 c4 17. Ne5 Qe6
18. f4 Rfd8 19. Qd2 Nf8 20. Nc3 Ng6 21. Rf1 Qd6 22. a4 Rbc8 23. e3 Ne7 24. g4 Ne8 25. f5 f6 26.
Nf3 Qd7 27. Qf2 Nd6 28. Nd2 Rf8 29. Qg3 Rcd8 30. Rf4 Nf7 31. Rf2 Rfe8 32. h3 Qd6 33. Nf1 Qa3
34. Rcc2 h5 35. Qc7 Qd6 36. Qxd6 Rxd6 37. Ng3 h4 38. Nh5 Ng5 39. Rf1 Kh7 40. Nf4 Rdd8 41. Kh2
Rd7 42. Bh1 Rd6 43. Ng2 g6 44. Nxh4 gxf5 45. gxf5 Rh8 46. Nf3 Kg7 47. Nxg5 fxg5 48. Rg2 Kf6 49.
Rg3 Re8 50. Bf3 Rdd8 51. Be2 Rf8 52. Bg4 Nc8 53. Bf3 Rfe8 54. h4 Rh8 55. h5 Rhe8 56. Bg2 Ne7
57. h6 Rh8 58. Rh3 Rh7 59. Kg1 Ba8 60. Nd1 g4 61. Rh5 g3 62. Nc3 Ng8 63. Ne2 Rxh6 64. Nxg3
Rxh5 65. Nxh5+ Kf7 66. Kf2 Nf6 67. Nxf6 Kxf6 68. Rh1 c3 69. Rc1 Rh8 70. Rxc3 Kxf5 71. Rc7 Kf6
72. Bf3 Rg8 73. Rh7 Rg6 74. Bd1 Rg8 75. Rh6+ Ke7 76. Rxb6 Kd7 77. Rf6 Ke7 78. Rh6 Rg7 79. Rh8
Bb7 80. Rh5 Kd6 81. Rh3 Rf7+ 82. Ke1 Bc8 83. Rh6+ Kc7 84. Rc6+ Kb8 85. Rd6 Bb7 86. b6 Ba6 87.
Rxd5 Rf6 88. Rxa5 Rxb6 89. Kd2 Bb7 90. Rb5 Rf6 91. Bb3 Kc7 92. Re5 Ba6 93. Kc3 Rf1 94. Bc2
Rh1 95. a5 Kd6 96. e4 Bf1 97. Rf5 Bg2 98. Rf4 Rc1 99. Kb2 Rh1 100. a6 1-0
White: AlphaZero Black: Stockfish
1. d4 Nf6 2. c4 e6 3. Nf3 b6 4. g3 Bb7 5. Bg2 Bb4+ 6. Bd2 Be7 7. Nc3 c6 8. e4 d5 9. e5 Ne4 10. 0-0
Ba6 11. b3 Nxc3 12. Bxc3 dxc4 13. b4 b5 14. Nd2 0-0 15. Ne4 Bb7 16. Qg4 Nd7 17. Nc5 Nxc5 18.
dxc5 a5 19. a3 axb4 20. axb4 Rxa1 21. Rxa1 Qd3 22. Rc1 Ra8 23. h4 Qd8 24. Be4 Qc8 25. Kg2 Qc7
26. Qh5 g6 27. Qg4 Bf8 28. h5 Rd8 29. Qh4 Qe7 30. Qf6 Qe8 31. Rh1 Rd7 32. hxg6 fxg6 33. Qh4
Qe7 34. Qg4 Rd8 35. Bb2 Qf7 36. Bc1 c3 37. Be3 Be7 38. Qe2 Bf8 39. Qc2 Bg7 40. Qxc3 Qd7 41.
Rc1 Qc7 42. Bg5 Rf8 43. f4 h6 44. Bf6 Bxf6 45. exf6 Qf7 46. Ra1 Qxf6 47. Qxf6 Rxf6 48. Ra7 Rf7
49. Bxg6 Rd7 50. Kf2 Kf8 51. g4 Bc8 52. Ra8 Rc7 53. Ke3 h5 54. gxh5 Kg7 55. Ra2 Re7 56. Be4 e5
57. Bxc6 exf4+ 58. Kxf4 Rf7+ 59. Ke5 Rf5+ 60. Kd6 Rxh5 61. Rg2+ Kf6 62. Kc7 Bf5 63. Kb6 Rh4
64. Ka5 Bg4 65. Bxb5 Ke7 66. Rg3 Bc8 67. Re3+ Kf7 68. Be2 1-0
White: AlphaZero, Black: Stockfish
1. d4 e6 2. e4 d5 3. Nc3 Nf6 4. e5 Nfd7 5. f4 c5 6. Nf3 cxd4 7. Nb5 Bb4+ 8. Bd2 Bc5 9. b4 Be7 10.
Nbxd4 Nc6 11. c3 a5 12. b5 Nxd4 13. cxd4 Nb6 14. a4 Nc4 15. Bd3 Nxd2 16. Kxd2 Bd7 17. Ke3 b6
18. g4 h5 19. Qg1 hxg4 20. Qxg4 Bf8 21. h4 Qe7 22. Rhc1 g6 23. Rc2 Kd8 24. Rac1 Qe8 25. Rc7 Rc8
26. Rxc8+ Bxc8 27. Rc6 Bb7 28. Rc2 Kd7 29. Ng5 Be7 30. Bxg6 Bxg5 31. Qxg5 fxg6 32. f5 Rg8 33.
Qh6 Qf7 34. f6 Kd8 35. Kd2 Kd7 36. Rc1 Kd8 37. Qe3 Qf8 38. Qc3 Qb4 39. Qxb4 axb4 40. Rg1 b3
41. Kc3 Bc8 42. Kxb3 Bd7 43. Kb4 Be8 44. Ra1 Kc7 45. a5 Bd7 46. axb6+ Kxb6 47. Ra6+ Kb7 48.
Kc5 Rd8 49. Ra2 Rc8+ 50. Kd6 Be8 51. Ke7 g5 52. hxg5 1-0
White: AlphaZero, Black: Stockfish
1. Nf3 Nf6 2. d4 e6 3. c4 b6 4. g3 Bb7 5. Bg2 Be7 6. 0-0 0-0 7. d5 exd5 8. Nh4 c6 9. cxd5 Nxd5 10.
Nf5 Nc7 11. e4 d5 12. exd5 Nxd5 13. Nc3 Nxc3 14. Qg4 g6 15. Nh6+ Kg7 16. bxc3 Bc8 17. Qf4 Qd6
18. Qa4 g5 19. Re1 Kxh6 20. h4 f6 21. Be3 Bf5 22. Rad1 Qa3 23. Qc4 b5 24. hxg5+ fxg5 25. Qh4+
Kg6 26. Qh1 Kg7 27. Be4 Bg6 28. Bxg6 hxg6 29. Qh3 Bf6 30. Kg2 Qxa2 31. Rh1 Qg8 32. c4 Re8 33.
Bd4 Bxd4 34. Rxd4 Rd8 35. Rxd8 Qxd8 36. Qe6 Nd7 37. Rd1 Nc5 38. Rxd8 Nxe6 39. Rxa8 Kf6 40.
cxb5 cxb5 41. Kf3 Nd4+ 42. Ke4 Nc6 43. Rc8 Ne7 44. Rb8 Nf5 45. g4 Nh6 46. f3 Nf7 47. Ra8 Nd6+
48. Kd5 Nc4 49. Rxa7 Ne3+ 50. Ke4 Nc4 51. Ra6+ Kg7 52. Rc6 Kf7 53. Rc5 Ke6 54. Rxg5 Kf6 55.
Rc5 g5 56. Kd4 1-0
18
Program
Chess
Shogi
Go
AlphaZero
Stockfish
Elmo
80k
70,000k
40k
16k
35,000k
Table S4: Evaluation speed (positions/second) of AlphaZero, Stockfish, and Elmo in chess,
shogi and Go.
19
| 2cs.AI
|
Toward Deeper Understanding of Neural Networks:
The Power of Initialization and a Dual View on Expressivity
arXiv:1602.05897v2 [cs.LG] 19 May 2017
Amit Daniely∗
Roy Frostig†
Yoram Singer‡
May 23, 2017
Abstract
We develop a general duality between neural networks and compositional kernels,
striving towards a better understanding of deep learning. We show that initial representations generated by common random initializations are sufficiently rich to express
all functions in the dual kernel space. Hence, though the training objective is hard
to optimize in the worst case, the initial weights form a good starting point for optimization. Our dual view also reveals a pragmatic and aesthetic perspective of neural
networks and underscores their expressive power.
∗
Email: [email protected]
Email: [email protected]. Work performed at Google.
‡
Email: [email protected]
†
Contents
1 Introduction
1
2 Related work
2
3 Setting
3
4 Computation skeletons
4.1 From computation skeletons to neural networks . . . . . . . .
4.2 From computation skeletons to reproducing kernels . . . . . .
4
6
7
5 Main results
5.1 Incorporating bias terms . . . . . . . . . . . . . . . . . . . . .
9
11
6 Mathematical background
11
7 Compositional kernel spaces
13
8 The dual activation function
15
9 Proofs
20
9.1 Well-behaved activations . . . . . . . . . . . . . . . . . . . . . 20
9.2 Proofs of Thms. 2 and 3 . . . . . . . . . . . . . . . . . . . . . 23
9.3 Proofs of Thms. 4 and 5 . . . . . . . . . . . . . . . . . . . . . 26
10 Discussion
29
1
Introduction
Neural network (NN) learning has underpinned state of the art empirical results in numerous applied machine learning tasks (see for instance [31, 33]). Nonetheless, neural network
learning remains rather poorly understood in several regards. Notably, it remains unclear
why training algorithms find good weights, how learning is impacted by the network architecture and activations, what is the role of random weight initialization, and how to choose
a concrete optimization procedure for a given architecture.
We start by analyzing the expressive power of NNs subsequent to the random weight initialization. The motivation is the empirical success of training algorithms despite inherent
computational intractability, and the fact that they optimize highly non-convex objectives
with potentially many local minima. Our key result shows that random initialization already positions learning algorithms at a good starting point. We define an object termed a
computation skeleton that describes a distilled structure of feed-forward networks. A skeleton induces a family of network architectures along with a hypothesis class H of functions
obtained by certain non-linear compositions according to the skeleton’s structure. We show
that the representation generated by random initialization is sufficiently rich to approximately express the functions in H. Concretely, all functions in H can be approximated by
tuning the weights of the last layer, which is a convex optimization task.
In addition to explaining in part the success in finding good weights, our study provides
an appealing perspective on neural network learning. We establish a tight connection between network architectures and their dual kernel spaces. This connection generalizes several
previous constructions (see Sec 2). As we demonstrate, our dual view gives rise to design
principles for NNs, supporting current practice and suggesting new ideas. We outline below
a few points.
• Duals of convolutional networks appear a more suitable fit for vision and acoustic tasks
than those of fully connected networks.
• Our framework surfaces a principled initialization scheme. It is very similar to common
practice, but incorporates a small correction.
• By modifying the activation functions, two consecutive fully connected layers can be
replaced with one while preserving the network’s dual kernel.
• The ReLU activation, i.e. x 7→ max(x, 0), possesses favorable properties. Its dual
kernel is expressive, and it can be well approximated by random initialization, even
when the initialization’s scale is moderately changed.
• As the number of layers in a fully connected network becomes very large, its dual kernel
converges to a degenerate form for any non-linear activation.
• Our result suggests that optimizing the weights of the last layer can serve as a convex proxy for choosing among different architectures prior to training. This idea was
advocated and tested empirically in [49].
1
2
Related work
Current theoretical understanding of NN learning. Understanding neural network
learning, particularly its recent successes, commonly decomposes into the following research
questions.
(i) What functions can be efficiently expressed by neural networks?
(ii) When does a low empirical loss result in a low population loss?
(iii) Why and when do efficient algorithms, such as stochastic gradient, find good weights?
Though still far from being complete, previous work provides some understanding of questions (i) and (ii). Standard results from complexity theory [28] imply that essentially all
functions of interest (that is, any efficiently computable function) can be expressed by a
network of moderate size. Biological phenomena show that many relevant functions can be
expressed by even simpler networks, similar to convolutional neural networks [32] that are
dominant in ML tasks today. Barron’s theorem [7] states that even two-layer networks can
express a very rich set of functions. As for question (ii), both classical [10, 9, 3] and more
recent [40, 22] results from statistical learning theory show that as the number of examples
grows in comparison to the size of the network the empirical loss must be close to the population loss. In contrast to the first two, question (iii) is rather poorly understood. While
learning algorithms succeed in practice, theoretical analysis is overly pessimistic. Direct
interpretation of theoretical results suggests that when going slightly deeper beyond single
layer networks, e.g. to depth two networks with very few hidden units, it is hard to predict
even marginally better than random [29, 30, 17, 18, 16]. Finally, we note that the recent
empirical successes of NNs have prompted a surge of theoretical work around NN learning
[47, 1, 4, 12, 39, 35, 19, 52, 14].
Compositional kernels and connections to networks. The idea of composing kernels
has repeatedly appeared throughout the machine learning literature, for instance in early
work by Schölkopf et al. [51], Grauman and Darrell [21]. Inspired by deep networks’ success,
researchers considered deep composition of kernels [36, 13, 11]. For fully connected two-layer
networks, the correspondence between kernels and neural networks with random weights has
been examined in [46, 45, 38, 56]. Notably, Rahimi and Recht [46] proved a formal connection (similar to ours) for the RBF kernel. Their work was extended to include polynomial
kernels [27, 42] as well as other kernels [6, 5]. Several authors have further explored ways to
extend this line of research to deeper, either fully-connected networks [13] or convolutional
networks [24, 2, 36]. Our work sets a common foundation for and expands on these ideas.
We extend the analysis from fully-connected and convolutional networks to a rather broad
family of architectures. In addition, we prove approximation guarantees between a network
and its corresponding kernel in our more general setting. We thus extend previous analyses
that only applies to fully connected two-layer networks. Finally, we use the connection as
an analytical tool to reason about architectural design choices.
2
3
Setting
Notation. We denote vectors by bold-face letters (e.g. x), and matrices by upper case
Greek letters (e.g. Σ). The 2-norm of x ∈ Rd is denoted by kxk. For functions σ : R → R
we let
q
p
R∞
x2
kσk := EX∼N (0,1) σ 2 (X) = √12π −∞ σ 2 (x)e− 2 dx .
Let G = (V, E) be a directed acyclic graph. The set of neighbors incoming to a vertex v is
denoted in(v) := {u ∈ V | uv ∈ E}. The d − 1 dimensional sphere is denoted Sd−1 = {x ∈
Rd | kxk = 1}. We provide a brief overview of reproducing kernel Hilbert spaces in the sequel
and merely introduce notation here. In a Hilbert space H, we use a slightly non-standard
notation HB for the ball of radius B, {x ∈ H | kxkH ≤ B}. We use [x]+ to denote max(x, 0)
and 1[b] to denote the indicator function of a binary variable b.
Input space. Throughout the paper we assume that each example is a sequence of n
elements, each of which is represented
as a unit vector. Namely, we fix n and take the input
n
space to be X = Xn,d = Sd−1 . Each input example is denoted,
x = (x1 , . . . , xn ), where xi ∈ Sd−1 .
(1)
We refer to each vector xi as the input’s ith coordinate, and use xij to denote it jth scalar
entry. Though this notation is slightly non-standard, it unifies input types seen in various
domains. For example, binary features can be encoded by taking d = 1, in which case
X = {±1}n . Meanwhile, images and audio signals are often represented as bounded and
continuous numerical values—we can assume in full generality that these values lie in [−1, 1].
1
To match
setup
the πx
above, we embed [−1, 1] into the circle S , e.g. via the map x 7→
πx
sin 2 , cos 2 . When each coordinate is categorical—taking one of d values—we can
represent category j ∈ [d] by the unit vector ej ∈ Sd−1 . When d may be very large or the
basic units exhibits some structure, such as when the input is a sequence of words, a more
0
concise encoding may be useful, e.g. as unit vectors in a low dimension space Sd where
d0 d (see for instance Mikolov et al. [37], Levy and Goldberg [34]).
Supervised learning. The goal in supervised learning is to devise a mapping from the
input space X to an output space Y based on a sample S = {(x1 , y1 ), . . . , (xm , ym )}, where
(xi , yi ) ∈ X × Y, drawn i.i.d. from a distribution D over X × Y. A supervised learning
problem is further specified by an output length k and a loss function ` : Rk × Y → [0, ∞),
and the goal is to find a predictor h : XP→ Rk whose loss, LD (h) := E(x,y)∼D `(h(x), y),
is small. The empirical loss LS (h) := m1 m
i=1 `(h(xi ), yi ) is commonly used as a proxy for
the loss LD . Regression problems correspond to Y = R and, for instance, the squared loss
`(ŷ, y) = (ŷ − y)2 . Binary classification is captured by Y = {±1} and, say, the zero-one loss
`(ŷ, y) = 1[ŷy ≤ 0] or the hinge loss `(ŷ, y) = [1 − ŷy]+ , with standard extensions to the
multiclass case. A loss ` is L-Lipschitz if |`(y1 , y) − `(y2 , y)| ≤ L|y1 − y2 | for all y1 , y2 ∈ Rk ,
y ∈ Y, and it is convex if `(·, y) is convex for every y ∈ Y.
3
Neural network learning. We define a neural network N to be a vertices weighted
directed acyclic graph (DAG) whose nodes are denoted V (N ) and edges E(N ). The weight
function will be denoted by δ : V (N ) → [0, ∞), and its sole role would be to dictate the
distribution of the initial weights (see definition 3). Each of its internal units, i.e. nodes with
both incoming and outgoing edges, is associated with an activation function σv : R → R.
In this paper’s context, an activation can be any function that is square integrable with
respect to the Gaussian measure on R. We say that σ is normalized if kσk = 1. The set
of nodes having only incoming edges are called the output nodes. To match the setup of a
supervised learning problem, a network N has nd input nodes and k output nodes, denoted
o1 , . . . , ok . A network N together with a weight vector w = {wuv | uv ∈ E} defines a
predictor hN ,w : X → Rk whose prediction is given by “propagating” x forward through
the network. Formally, we define hv,w (·) to be the output of the subgraph of the node v
as follows: for an input node v, hv,w outputs the corresponding coordinate in x, and for all
other nodes, we define hv,w recursively as
P
w
h
(x)
.
hv,w (x) = σv
uv u,w
u∈in(v)
Finally, we let hN ,w (x) = (ho1 ,w (x), . . . , hok ,w (x)). We also refer to internal nodes as hidden
units. The output layer of N is the sub-network consisting of all output neurons of N
along with their incoming edges. The representation induced by a network N is the network
rep(N ) obtained from N by removing the output layer. The representation function induced
by the weights w is RN ,w := hrep(N ),w . Given a sample
Pm S, a learning algorithm searches for
1
weights w having small empirical loss LS (w) = m i=1 `(hN ,w (xi ), yi ). A popular approach
is to randomly initialize the weights and then use a variant of the stochastic gradient method
to improve these weights in the direction of lower empirical loss.
Kernel learning. A function κ : X × X → R is a reproducing kernel, or simply a kernel,
if for every x1 , . . . , xr ∈ X , the r × r matrix Γi,j = {κ(xi , xj )} is positive semi-definite.
Each kernel induces a Hilbert space Hκ of functions from X to R with a corresponding norm
k · kHκ . A kernel and its corresponding space are normalized if ∀x ∈ X , κ(x, x) = 1. Given
a convex loss function `, a sample S, and a kernel κ, a kernel learning
algorithm finds a
P
1
k
function f = (f1 , . . . , fk ) ∈PHκ whose empirical loss, LS (f ) = m i `(f (xi ), yi ), is minimal
among all functions with i kfi k2κ ≤ R2 for some R > 0. Alternatively, kernel algorithms
minimize the regularized loss,
LR
S (f )
m
k
1 X
1 X
=
`(f (xi ), yi ) + 2
kfi k2κ ,
m i=1
R i=1
a convex objective that often can be efficiently minimized.
4
Computation skeletons
In this section we define a simple structure which we term a computation skeleton. The
purpose of a computational skeleton is to compactly describe a feed-forward computation
4
S1
S2
S3
S4
Figure 1: Examples of computation skeletons.
from an input to an output. A single skeleton encompasses a family of neural networks that
share the same skeletal structure. Likewise, it defines a corresponding kernel space.
Definition 1. A computation skeleton S is a DAG whose non-input nodes are labeled by
activations.
Though the formal definition of neural networks and skeletons appear identical, we make
a conceptual distinction between them as their role in our analysis is rather different. Accompanied by a set of weights, a neural network describes a concrete function, whereas
the skeleton stands for a topology common to several networks as well as for a kernel. To
further underscore the differences we note that skeletons are naturally more compact than
networks. In particular, all examples of skeletons in this paper are irreducible, meaning that
for each two nodes v, u ∈ V (S), in(v) 6= in(u). We further restrict our attention to skeletons
with a single output node, showing later that single-output skeletons can capture supervised
problems with outputs in Rk . We denote by |S| the number of non-input nodes of S.
Figure 1 shows four example skeletons, omitting the designation of the activation functions. The skeleton S1 is rather basic as it aggregates all the inputs in a single step. Such
topology can be useful in the absence of any prior knowledge of how the output label may be
computed from an input example, and it is commonly used in natural language processing
where the input is represented as a bag-of-words [23]. The only structure in S1 is a single
fully connected layer:
5
Terminology (Fully connected layer of a skeleton). An induced subgraph of a skeleton with
r + 1 nodes, u1 , . . . , ur , v, is called a fully connected layer if its edges are u1 v, . . . , ur v.
The skeleton S2 is slightly more involved: it first processes consecutive (overlapping) parts
of the input, and the next layer aggregates the partial results. Altogether, it corresponds
to networks with a single one-dimensional convolutional layer, followed by a fully connected
layer. The two-dimensional (and deeper) counterparts of such skeletons correspond to networks that are common in visual object recognition.
Terminology (Convolution layer of a skeleton). Let s, w, q be positive integers and denote
n = s(q − 1) + w. A subgraph of a skeleton is a one dimensional convolution layer of width
w and stride s if it has n + q nodes, u1 , . . . , un , v1 , . . . , vq , and qw edges, us(i−1)+j vi , for
1 ≤ i ≤ q, 1 ≤ j ≤ w.
The skeleton S3 is a somewhat more sophisticated version of S2 : the local computations are first aggregated, then reconsidered with the aggregate, and finally aggregated
again. The last skeleton, S4 , corresponds to the networks that arise in learning sequenceto-sequence mappings as used in translation, speech recognition, and OCR tasks (see for
example Sutskever et al. [55]).
4.1
From computation skeletons to neural networks
The following definition shows how a skeleton, accompanied with a replication parameter
r ≥ 1 and a number of output nodes k, induces a neural network architecture. Recall that
inputs are ordered sets of vectors in Sd−1 .
Definition 2 (Realization of a skeleton). Let S be a computation skeleton and consider
input coordinates in Sd−1 as in (1). For r, k ≥ 1 we define the following neural network
N = N (S, r, k). For each input node in S, N has d corresponding input neurons with
weight 1/d. For each internal node v ∈ S labeled by an activation σ, N has r neurons
v 1 , . . . , v r , each with an activation σ and weight 1/r. In addition, N has k output neurons
o1 , . . . , ok with the identity activation σ(x) = x and weight 1. There is an edge v i uj ∈ E(N )
whenever uv ∈ E(S). For every output node v in S, each neuron v j is connected to all
output neurons o1 , . . . , ok . We term N the (r, k)-fold realization of S. We also define the
r-fold realization of S as1 N (S, r) = rep (N (S, r, 1)).
Note that the notion of the replication parameter r corresponds, in the terminology of
convolutional networks, to the number of channels taken in a convolutional layer and to the
number of hidden units taken in a fully-connected layer.
Figure 2 illustrates a (5, 4)- and 5-realizations of a skeleton with coordinate dimension
d = 2. The (5, 4)-realization is a network with a single (one dimensional) convolutional layer
having 5 channels, stride of 2, and width of 4, followed by three fully-connected layers. The
global replication parameter r in a realization is used for brevity; it is straightforward to
extend results when the different nodes in S are each replicated to a different extent.
1
Note that for every k, rep (N (S, r, 1)) = rep (N (S, r, k)).
6
N (S, 5, 4)
S
N (S, 5)
Figure 2: A (5, 4)-fold and 5-fold realizations of the computation skeleton S with d = 2.
We next define a scheme for random initialization of the weights of a neural network,
that is similar to what is often done in practice. We employ the definition throughout the
paper whenever we refer to random weights.
Definition 3 (Random weights). A random initialization of a neural network N is a multivariate Gaussian w = (wuv )uv∈E(N ) such that each weight wuv is sampled independently from
a normal distribution with mean 0 and variance2 dδ(u)/δ(in(v)) if u is an input neuron and
δ(u)/(kσu k2 δ(in(v))) otherwise.
Architectures such as convolutional nets have weights that are shared across different edges.
Again, it is straightforward to extend our results to these cases and for simplicity we assume
no explicit weight sharing.
4.2
From computation skeletons to reproducing kernels
In addition to networks’ architectures, a computation skeleton S also defines a normalized
kernel κS : X × X → [−1, 1] and a corresponding norm k · kS on functions f : X → R. This
norm has the property that kf kS is small if and only if f can be obtained by certain simple
compositions of functions according to the structure of S. To define the kernel, we introduce
a dual activation and dual kernel. For ρ ∈ [−1, 1], we denote by Nρ the multivariate Gaussian
distribution on R2 with mean 0 and covariance matrix ρ1 ρ1 .
Definition 4 (Dual activation and kernel). The dual activation of an activation σ is the
function σ̂ : [−1, 1] → R defined as
σ̂(ρ) =
σ(X)σ(Y ) .
E
(X,Y )∼Nρ
The dual kernel w.r.t. to a Hilbert space H is the kernel κσ : H1 × H1 → R defined as
κσ (x, y) = σ̂(hx, yiH ) .
2
For U ⊂ V (N ) we denote δ(U ) =
P
u∈U
δ(u).
7
Activation
Identity
x
2nd Hermite
2 −1
x√
2
ReLU
Step
Exponential
√
√
e
Kernel
ρ
linear
2
ρ
2 [x]+
2 1[x ≥ 0]
x−2
Dual Activation
1
π
1
2
1
e
√
+
+
+
ρ
2
ρ
π
ρ
e
+
+
+
Ref
poly
1−ρ2 +(π−cos−1 (ρ))ρ
ρ2
ρ4
+ 24π
+ ... =
2π
π
π−cos−1 (ρ)
ρ3
3ρ5
+ 40π + . . . =
6π
π
ρ2
ρ3
ρ−1
+
+
.
.
.
=
e
2e
6e
arccos1
[13]
arccos0
[13]
RBF
[36]
Table 1: Activation functions and their duals.
Section 7 shows that κσ is indeed a kernel for every activation σ that adheres with the
square-integrability requirement. In fact, any continuous µ : [−1, 1] → R, such that (x, y) 7→
µ(hx, yiH ) is a kernel for all H, is the dual of some activation. Note that κσ is normalized iff
σ is normalized. We show in Section 8 that dual activations are closely related to Hermite
polynomial expansions, and that these can be used to calculate the duals of activation
functions analytically. Table 1 lists a few examples of normalized activations and their
corresponding dual (corresponding derivations are in Section 8). The following definition
gives the kernel corresponding to a skeleton having normalized activations.3
Definition 5 (Compositional kernels). Let S be a computation skeleton with normalized
activations and (single) output node o. For every node v, inductively define a kernel κv :
X × X → R as follows. For an input node v corresponding to the ith coordinate, define
κv (x, y) = hxi , yi i. For a non-input node v, define
!
P
κ
(x,
y)
u
u∈in(v)
.
κv (x, y) = σ̂v
|in(v)|
The final kernel κS is κo , the kernel associated with the output node o. The resulting Hilbert
space and norm are denoted HS and k · kS respectively, and Hv and k · kv denote the space
and norm when formed at node v.
As we show later, κS is indeed a (normalized) kernel for every skeleton S. To understand the
kernel in the context of learning, we need to examine which functions can be expressed as
moderate norm functions in HS . As we show in section 7, these are the functions obtained
by certain simple compositions according to the feed-forward structure of S. For intuition,
the following example contrasts two commonly used skeletons.
Example 1 (Convolutional vs. fully connected skeletons). Consider a network whose activations are all ReLU, σ(z) = [z]+ , and an input space Xn,1 = {±1}n . Say that S1 is a skeleton
comprising a single fully connected layer, and that S2 is one comprising a convolutional layer
3
For a skeleton S with unnormalized activations, the corresponding kernel is the kernel of the skeleton
S obtained by normalizing the activations of S.
0
8
of stride 1 and width q = log0.999 (n), followed by a single fully-connected layer. (The skeleton
S2 from Figure 1 is a concrete example of the convolutional skeleton with q = 2 and n = 4.)
The kernel κS1 takes the form κS1 (x, y) = σ̂ (hx, yi/n). It is a symmetric kernel and therefore
functions with small norm in HS1 are essentially low-degree polynomials. For instance, fix a
bound R = n1.001 on the norm of the functions. In this case, the space HSR1 contains multiplication of one or two input coordinates. However, multiplication of 3 or more coordinates are
no-longer in HSR1 . Moreover, this property holds true regardless of the choice of activation
function. On the other hand, HSR2 contains functions whose dependence on adjacent input
coordinates is far more complex. It includes, for instance, any function f : X → {±1} that
is symmetric (i.e. f (x) = f (−x)) and that depends on q adjacent coordinates xi , . . . , xi+q .
Furthermore, any sum of n such functions is also in HSR2 .
5
Main results
We review our main results. Let us fix a compositional kernel S. There are a few upshots to
underscore upfront. First, our analysis implies that a representation generated by a random
initialization of N = N (S, r, k) approximates the kernel κS . The sense in which the result
holds is twofold. First, with the proper rescaling we show that hRN ,w (x), RN ,w (x0 )i ≈
κS (x, x0 ). Then, we also show that the functions obtained by composing bounded linear
functions with RN ,w are approximately the bounded-norm functions in HS . In other words,
the functions expressed by N under varying the weights of the last layer are approximately
bounded-norm functions in HS . For simplicity, we restrict the analysis to the case k = 1.
We also confine the analysis to either bounded activations, with bounded first and second
derivatives, or the ReLU activation. Extending the results to a broader family of activations
is left for future work. Through this and remaining sections we use & to hide universal
constants.
Definition 6. An activation σ : R → R is C-bounded if it is twice continuously differentiable
and kσk∞ , kσ 0 k∞ , kσ 00 k∞ ≤ kσkC.
Note that many activations are C-bounded for some constant C > √
0. In particular, most
of the popular sigmoid-like functions such as 1/(1 + e−x ), erf(x), x/ 1 + x2 , tanh(x), and
tan−1 (x) satisfy the boundedness requirements. We next introduce terminology that parallels
the representation layer of N with a kernel space. Concretely, let N be a network whose
representation part has q output neurons. Given weights w, the normalized representation
√
Ψw is obtained from the representation RN ,w by dividing each output neuron v by kσv k q.
The empirical kernel corresponding to w is defined as κw (x, x0 ) = hΨw (x), Ψw (x0 )i. We also
define the empirical kernel space corresponding to w as Hw = Hκw . Concretely,
Hw = {hv (x) = hv, Ψw (x)i | v ∈ Rq } ,
and the norm of Hw is defined as khkw = inf{kvk | h = hv }. Our first result shows that the
empirical kernel approximates the kernel kS .
9
Theorem 2. Let S be a skeleton with C-bounded activations. Let w be a random initialization of N = N (S, r) with
r≥
(4C 4 )depth(S)+1 log (8|S|/δ)
.
2
Then, for all x, x0 , with probability of at least 1 − δ,
|kw (x, x0 ) − kS (x, x0 )| ≤ .
We note that if we fix the activation and assume that the depth of S is logarithmic, then
the required bound on r is polynomial. For the ReLU activation we get a stronger bound
with only quadratic dependence on the depth. However, it requires that ≤ 1/depth(S).
Theorem 3. Let S be a skeleton with ReLU activations. Let w be a random initialization
of N (S, r) with
depth2 (S) log (|S|/δ)
.
r&
2
Then, for all x, x0 and . 1/depth(S), with probability of at least 1 − δ,
|κw (x, x0 ) − κS (x, x0 )| ≤ .
For the remaining theorems, we fix a L-Lipschitz loss ` : R × Y → [0, ∞). For a distribution
D on X × Y we denote by kDk0 the cardinality of the support of the distribution. We note
that log (kDk0 ) is bounded by, for instance, the number of bits used to represent an element
in X × Y. We use the following notion of approximation.
Definition 7. Let D be a distribution on X ×Y. A space H1 ⊂ RX -approximates the space
H2 ⊂ RX w.r.t. D if for every h2 ∈ H2 there is h1 ∈ H1 such that LD (h1 ) ≤ LD (h2 ) + .
Theorem 4. Let S be a skeleton with C-bounded activations. Let w be a random initialization of N (S, r) with
LRC|S|
4 4
4 depth(S)+1
L R (4C )
log
δ
r&
.
4
√
Then, with√probability of at least 1−δ over the choices of w we have that Hw2R -approximates
R
HSR and HS 2R -approximates Hw
.
Theorem 5. Let S be a skeleton with ReLU activations and . 1/depth(C). Let w be a
random initialization of N (S, r) with
L4 R4 depth2 (S) log kDkδ0 |S|
r&
.
4
√
Then, with√probability of at least 1−δ over the choices of w we have that Hw2R -approximates
R
HSR and HS 2R -approximates Hw
.
10
As in Theorems 2 and 3, for a fixed C-bounded activation and logarithmically deep S, the
required bounds on r are polynomial. Analogously, for the ReLU activation the bound is
polynomial even without restricting the depth. However, the polynomial growth in Theorems 4 and 5 is rather large. Improving the bounds, or proving their optimality, is left to
future work.
5.1
Incorporating bias terms
Our results can be extended to incorporate bias terms. Namely, in addition to the weights
we can add a bias vector b = {bv | v ∈ V } and let each neuron compute the function
P
hv,w,b (x) = σv
w
h
(x)
+
b
.
uv
u,w
v
u∈in(v)
To do so, we extend the definition of random initialization and compositional kernel as
follows:
Definition 8 (Random weights with bias terms). Let 0 ≤ β ≤ 1. A β-biased random initialization of a neural network N is a multivariate Gaussian (b, w) = ((wuv )uv∈E(N ) , (bv )v∈V (N ) )
such that each weight wuv is sampled independently from a normal distribution with mean 0
and variance (1 − β)dδ(u)/δ(in(v)) if u is an input neuron and (1 − β)δ(u)/(kσu k2 δ(in(v)))
otherwise. Finally, each bias term bv is sampled independently from a normal distribution
with mean 0 and variance β.
Definition 9 (Compositional kernels with bias terms). Let S be a computation skeleton with
normalized activations and (a single) output node o, and let 0 ≤ β ≤ 1. For every node v,
inductively define a kernel κβv : X × X → R as follows. For an input node v corresponding
to the ith coordinate, define κβv (x, y) = hxi , yi i. For a non-input node v, define
!
P
β
(x,
y)
κ
u∈in(v) u
+β .
κβv (x, y) = σ̂v (1 − β)
|in(v)|
The final kernel κβS is κβo , the kernel associated with the output node o.
Note that our original definitions correspond to β = 0. With the above definitions,
Theorems 2, 3, 4 and 5 extend to the case when there exist bias terms. To simplify the
notation, we focus on the case when there are no biases.
6
Mathematical background
Reproducing kernel Hilbert spaces (RKHS). The proofs of all the theorems we quote
here are well-known and can be found in Chapter 2 of [48] and similar textbooks. Let H
be a Hilbert space of functions from X to R. We say that H is a reproducing kernel Hilbert
space, abbreviated RKHS or kernel space, if for every x ∈ X the linear functional f 7→ f (x)
is bounded. The following theorem provides a one-to-one correspondence between kernels
and kernel spaces.
11
Theorem 6. (i) For every kernel κ there exists a unique kernel space Hκ such that for every
x ∈ X , κ(·, x) ∈ Hκ and for all f ∈ Hκ , f (x) = hf (·), κ(·, x)iHκ . (ii) A Hilbert space
H ⊆ RX is a kernel space if and only if there exists a kernel κ : X × X → R such that
H = Hκ .
The following theorem describes a tight connection between embeddings of X into a
Hilbert space and kernel spaces.
Theorem 7. A function κ : X × X → R is a kernel if and only if there exists a mapping
Φ : X → H to some Hilbert space for which κ(x, x0 ) = hΦ(x), Φ(x0 )iH . In addition, the
following two properties hold,
• Hκ = {fv : v ∈ H}, where fv (x) = hv, Φ(x)iH .
• For every f ∈ Hκ , kf kHκ = inf{kvkH | f = fv }.
Positive definite functions. A function µ : [−1, 1] → R is positive definite (PSD) if there
are non-negative numbers b0 , b1 , . . . such that
∞
X
bi < ∞ and ∀x ∈ [−1, 1], µ(x) =
∞
X
i=0
The norm of µ is defined as kµk :=
bi x i .
i=0
pP
i bi
=
p
µ(1). We say that µ is normalized if kµk = 1
Theorem 8 (Schoenberg, [50]). A continuous function µ : [−1, 1] → R is PSD if and only
if for all d = 1, 2, . . . , ∞, the function κ : Sd−1 × Sd−1 → R defined by κ(x, x0 ) = µ(hx, x0 i)
is a kernel.
The restriction to the unit sphere of many of the kernels used in machine learning applications
corresponds to positive definite functions. An example is the Gaussian kernel,
kx − x0 k2
0
.
κ(x, x ) = exp −
2σ 2
Indeed, note that for unit vectors x, x0 we have
kxk2 + kx0 k2 − 2hx, x0 i
1 − hx, x0 i
0
κ(x, x ) = exp −
= exp −
.
2σ 2
σ2
Another example is the Polynomial kernel κ(x, x0 ) = hx, x0 id .
12
Hermite polynomials. The normalized Hermite polynomials is the sequence h0 , h1 , . . . of
orthonormal polynomials obtained by applying the Gram-Schmidt process to the sequence
R∞
x2
1, x, x2 , . . . w.r.t. the inner-product hf, gi = √12π −∞ f (x)g(x)e− 2 dx. Recall that we define
activations as square integrable functions w.r.t. the Gaussian measure. Thus, Hermite polynomials form an orthonormal basis to the space of activations. In particular, each activation
σ can be uniquely described in the basis of Hermite polynomials,
σ(x) = a0 h0 (x) + a1 h1 (x) + a2 h2 (x) + . . . ,
(2)
where the convergence holds in `2 w.r.t. the Gaussian measure. This decomposition is called
the Hermite expansion. Finally, we use the following facts (see Chapter 11 in [41] and the
relevant entry in Wikipedia):
r
n
x
hn−1 (x) ,
(3)
hn (x) −
∀n ≥ 1, hn+1 (x) = √
n+1
n+1
√
∀n ≥ 1, h0n (x) =
nh (x)
(4)
( n−1
n
ρ n=m
E hm (X)hn (Y ) =
where n, m ≥ 0, ρ ∈ [−1, 1] ,
(5)
(X,Y )∼Nρ
0 n 6= m
(
0,
if n is odd
n
hn (0) =
,
(6)
√1 (−1) 2 (n − 1)!! if n is even
n!
where
7
n≤0
1
n!! = n · (n − 2) · · · 5 · 3 · 1 n > 0 odd .
n · (n − 2) · · · 6 · 4 · 2 n > 0 even
Compositional kernel spaces
We now describe the details of compositional kernel spaces. Let S be a skeleton with normalized activations and n input nodes associated with the input’s coordinates. Throughout
the rest of the section we study the functions in HS and their norm. In particular, we show
that κS is indeed a normalized kernel. Recall that κS is defined inductively by the equation,
!
P
0
u∈in(v) κu (x, x )
0
κv (x, x ) = σ̂v
.
(7)
|in(v)|
The recursion (7) describes a means for generating a kernel form another kernel. Since kernels
correspond to kernel spaces, it also prescribes an operator that produces a kernel space from
other kernel spaces. If Hv is the space corresponding to v, we denote this operator by
⊕u∈in(v) Hu
Hv = σ̂v
.
(8)
|in(v)|
13
The reason for using the above notation becomes clear in the sequel. The space HS is obtained by starting with the spaces Hv corresponding to the input nodes and propagating them
according to the structure of S, where at each node v the operation (8) is applied. Hence,
to understand HS we need to understand this operation as well as the spaces corresponding
to input nodes. The latter spaces are rather simple: for an input node v corresponding to
the variable xi , we have that Hv = {fw | ∀x, fw (x) = hw, xi i} and kfw kHv = kwk. To
understand (8), it is convenient to decompose it into two operations.P The first operation,
κu (x,x0 )
, and the
termed the direct average, is defined through the equation κ̃v (x, x0 ) = u∈in(v)
|in(v)|
⊕
H
u
resulting kernel space is denoted Hṽ = u∈in(v)
. The second operation, called the extension
|in(v)|
0
according to σ̂v , is defined through κv (x, x ) = σ̂v (κ̃v (x, x0 )). The resulting kernel space is
denoted Hv = σ̂v (Hṽ ). We next analyze these two operations.
The direct average of kernel spaces. Let H1 , . . . , Hn be kernel spaces with kernels
n
κ1 , . . . , κn : X × X → R. Their direct average,
denoted H = H1 ⊕···⊕H
, is the kernel space
n
P
n
1
0
0
corresponding to the kernel κ(x, x ) = n i=1 κi (x, x ).
Lemma 9. The function κ is indeed a kernel. Furthermore, the following properties hold.
1. If H1 , . . . , Hn are normalized then so is H.
n
2. H = f1 +...+f
| fi ∈ Hi
n
o
n kf k2 +...+kf k2
n Hn
1 H
f1 +...+fn
1
s.t.
f
=
,
f
∈
H
3. kf k2H = inf
i
i
n
n
Proof. (outline) The fact that κ is a kernel follows directly from the definition of a kernel
and the fact that an average of PSD matrices is PSD. Also, it is straight forward to verify item
1. We now proceed to items 2 and 3. By Theorem 7 there are Hilbert spaces G1 , . . . , Gn and
mappings Φi : X → Gi such that κi (x, x0 ) = hΦi (x), Φi (x0 )iGi . Consider now the mapping
Φ1 (x)
Φn (x)
√ ,..., √
.
Ψ(x) =
n
n
It holds that κ(x, x0 ) = hΨ(x), Ψ(x0 )i. Properties 2 and 3 now follow directly form Thm. 7
applied to Ψ.
The extension
P of a kernel space. Let H be a normalized kernel space with a kernel κ.
Let µ(x) = i bi xi be a PSD function. As we will see shortly, a function is PSD if and only
if it is a dual of an activation function. The extension of H w.r.t. µ, denoted µ (H), is the
kernel space corresponding to the kernel κ0 (x, x0 ) = µ(κ(x, x0 )).
Lemma 10. The function κ0 is indeed a kernel. Furthermore, the following properties hold.
1. µ(H) is normalized if and only if µ is.
14
(
2. µ(H) = span
)
Y
g | A ⊂ H, b|A| > 0
where span(A) is the closure of the span of A.
g∈A
3. kf kµ(H)
(
)
Q
X g∈A kgkH
XY
p
≤ inf
s.t. f =
g, A ⊂ H
b|A|
A
A g∈A
Proof. (outline) Let Φ : X → G be a mapping from X to the unit ball of a Hilbert space G
such that κ(x, x0 ) = hΦ(x), Φ(x0 )i. Define
p p
p
p
Ψ(x) =
b0 , b1 Φ(x), b2 Φ(x) ⊗ Φ(x), b3 Φ(x) ⊗ Φ(x) ⊗ Φ(x), . . .
It is not difficult to verify that hΨ(x), Ψ(x0 )i = µ(κ(x, x0 )). Hence, by Thm. 7, κ0 is indeed a
kernel. Verifying property 1 is a straightforward task. Properties 2 and 3 follow by applying
Thm. 7 on the mapping Ψ.
8
The dual activation function
The following lemma describes a few basic properties of the dual activation. These properties
follow easily from the definition of the dual activation and equations (2), (4), and (5).
Lemma 11. The following properties of the mapping σ 7→ σ̂ hold:
P
P
(a) If σ = i ai hi is the Hermite expansion of σ, then σ̂(ρ) = i a2i ρi .
(b) For every σ, σ̂ is positive definite.
(c) Every positive definite function is a dual of some activation.
(d) The mapping σ 7→ σ̂ preserves norms.
(e) The mapping σ 7→ σ̂ commutes with differentiation.
(f ) For a ∈ R, ac
σ = a2 σ̂.
(g) For every σ, σ̂ is continuous in [−1, 1] and smooth in (−1, 1).
(h) For every σ, σ̂ is non-decreasing and convex in [0, 1].
(i) For every σ, the range of σ̂ is [−kσk2 , kσk2 ].
2
(j) For every σ, σ̂(0) = EX∼N (0,1) σ(X) and σ̂(1) = kσk2 .
We next discuss a few examples for activations and calculate their dual activation and kernel.
Note that the dual of the exponential activation was calculated in [36] and the duals of the
step and the ReLU activations were calculated in [13]. Here, our derivations are different
and may prove useful for future calculations of duals for other activations.
15
The exponential activation. Consider the activation function σ(x) = Ceax where C > 0
2
is a normalization constant such that kσk = 1. The actual value of C is e−2a but it will not
be needed for the derivation below. From properties (e) and (f) of Lemma 11 we have that,
(σ̂)0 = σb0 = ac
σ = a2 σ̂ .
The the solution of ordinary differential equation (σ̂)0 = a2 σ̂ is of the form σ̂(ρ) = b exp (a2 ρ).
2
Since σ̂(1) = 1 we have b = e−a . We therefore obtain that the dual activation function is
2 ρ−a2
σ̂(ρ) = ea
2 (ρ−1)
= ea
.
Note that the kernel induced by σ is the RBF kernel, restricted to the d-dimensional sphere,
2 (hx,x0 i−1)
κσ (x, x0 ) = ea
= e−
a2 kx−x0 k2
2
.
The Sine activation and the Sinh kernel. Consider the activation σ(x) = sin(ax). We
iax
−iax
can write sin(ax) = e −e
. We have
2i
σ̂(ρ) =
E
(X,Y )∼Nρ
eiaX − e−iaX
2i
eiaY − e−iaY
2i
1
E
eiaX − e−iaX eiaY − e−iaY
4 (X,Y )∼Nρ
ia(X+Y )
1
E
e
− eia(X−Y ) − eia(−X+Y ) + eia(−X−Y ) .
= −
4 (X,Y )∼Nρ
= −
1 2
Recall that the characteristic function, E[eitX ], when X is distributed N (0, 1) is e− 2 t . Since
X + Y and −X − Y are normal variables with expectation 0 and variance of 2 + 2ρ, it follows
that,
E
eia(X+Y ) =
(X,Y )∼Nρ
e−ia(X+Y ) = e−
E
a2 (2+2ρ)
2
.
(X,Y )∼Nρ
Similarly, since the variance of X − Y and Y − X is 2 − 2ρ, we get
E
eia(X−Y ) =
(X,Y )∼Nρ
eia(−X+Y ) = e−
E
a2 (2−2ρ)
2
.
(X,Y )∼Nρ
We therefore obtain that
2 (1−ρ)
σ̂(ρ) =
e−a
2 (1+ρ)
− e−a
2
2
= e−a sinh(a2 ρ) .
Hermite activations and polynomial kernels. From Lemma 11 it follows that the dual
activation of the Hermite polynomial hn is ĥn (ρ) = ρn . Hence, the corresponding kernel is
the polynomial kernel.
16
The normalized step activation. Consider the activation
(√
2 x>0
σ(x) =
.
0
x≤0
To calculate σ̂ we compute the Hermite expansion of σ. For n ≥ 0 we let
Z ∞
Z ∞
2
x2
1
1
− x2
σ(x)hn (x)e dx = √
hn (x)e− 2 dx .
an = √
π 0
2π −∞
Since h0 (x) = 1, h1 (x) = x, and h2 (x) =
a2
we get the corresponding coefficients,
1
[σ(X)] = √
X∼N(0,1)
2
1
1
E [|X|] = √
=
E [σ(X)X] = √
X∼N(0,1)
π
2 X∼N(0,1)
1
1
= √
E [σ(X)(X 2 − 1)] =
E [X 2 − 1] = 0 .
2 X∼N(0,1)
2 X∼N(0,1)
a0 =
a1
2 −1
x√
,
2
E
x2
For n ≥ 3 we write gn (x) = hn (x)e− 2 and note that
x2
gn0 (x) = [h0n (x) − xhn (x)] e− 2
√
x2
=
nhn−1 (x) − xhn (x) e− 2
√
x2
= − n + 1 hn+1 (x)e− 2
√
= − n + 1 gn+1 (x) .
Here, the second equality follows from (4) and the third form (3). We therefore get
Z ∞
1
an = √
gn (x)dx
π 0
Z ∞
1
g 0 (x)dx
= −√
nπ 0 n−1
=0
z
}|
{
1
= √ gn−1 (0) − gn−1 (∞)
nπ
1
= √ hn−1 (0)
nπ
n−1
2 (n−2)!!
(−1)
=
√ √
nπ (n−1)!
=
0
(−1)
n−1
2 (n−2)!!
√
πn!
if n is odd
if n is even
17
.
The second equality follows from
and the last equality follows from (6). Finally, from
P(3)
∞
Lemma 11 we have that σ̂(ρ) = n=0 bn ρn where
((n−2)!!)2
if n is odd
πn!
1
bn = 2
.
if n = 0
0
if n is even ≥ 2
1
3
In particular, (b0 , b1 , b2 , b3 , b4 , b5 , b6 ) = 21 , π1 , 0, 6π
, 0, 40π
, 0 . Note that from the Taylor ex−1
pansion of cos−1 it follows that σ̂(ρ) = 1 − cos π (ρ) .
The normalizedP
ReLU activation. Consider the activation σ(x) =
now write σ̂(ρ) = i bi ρi . The first coefficient is
b0 =
√
2 max(0, x). We
2
2
1
1
E σ(X) =
E |X| = .
X∼N(0,1)
2 X∼N(0,1)
π
To calculate the remaining coefficients we simply note that the derivative of the ReLU
activation is the step activation and the mapping σ 7→ σ̂ commutes with differentiation.
Hence, from the calculation of the step activation we get,
((n−3)!!)2
if n is even
πn!
bn = 12
.
if n = 1
0
if n is odd ≥ 3
1
1
1
In particular, (b0 , b1 , b2 , b3 , b4 , b5 , b6 ) = π1 , 12 , 2π
, 0, 24π
, 0, 80π
. We see that the coefficients
corresponding to the degrees 0, 1, and 2 sum to 0.9774. The sums up to degrees 4 or 6 are
0.9907 and 0.9947 respectively. That is, we get an excellent approximation of less than 1%
error with a dual activation of degree 4.
The collapsing tower of fully connected layers. To conclude this section we discuss
the case of very deep networks. The setting is taken for illustrative purposes but it might
surface when building networks with numerous fully connected layers. Indeed, most deep
architectures that we are aware of do not employ more than five consecutive fully connected
layers.
Consider a skeleton Sm consisting of m fully connected layers, each layer associated with
the same (normalized) activation σ. We would like to examine the form of the compositional
kernel as the number of layers becomes very large. Due to the repeated structure and
activation we have
κSm (x, y) = αm
hx, yi
n
m times
z }| {
where αm = σ̂ = σ̂ ◦ . . . ◦ σ̂ .
m
18
Hence, the limiting properties of κSm can be understood from the limit of αm . In the case
that σ(x) = x or σ(x) = −x, σ̂ is the identity function. Therefore αm (ρ) = σ̂(ρ) = ρ for
all m and κSm is simply the linear kernel. Assume now that σ is neither the identity nor
its negation. The following claim shows that αm has a point-wise limit corresponding to a
degenerate kernel.
Claim 1. There exists a constant 0 ≤ ασ ≤ 1 such that for all −1 < ρ < 1,
lim αm (ρ) = ασ
m→∞
Before proving the claim, we note that for ρ = 1, αm (1) = 1 for all m, and therefore
limm→∞ αm (1) = 1. For ρ = −1, if σ is anti-symmetric then αm (−1) = −1 for all m,
and in particular limm→∞ αm (−1) = −1. In any other case, our argument can show that
limm→∞ αm (−1) = ασ .
P
i
Proof. Recall that σ̂(ρ) = ∞
i=0 bi ρ where the bi ’s are non-negative numbers that sum to 1.
By the assumption that σ is not the identity or its negation, b1 < 1. We first claim that
there is a unique ασ ∈ [0, 1] such that
∀x ∈ (−1, ασ ) , σ̂(ρ) > ρ and ∀x ∈ (ασ , 1) , ασ < σ̂(ρ) < ρ
(9)
To prove (9) it suffices to prove the following properties.
(a) σ̂(ρ) > ρ for ρ ∈ (−1, 0)
(b) σ̂ is non-decreasing and convex in [0, 1]
(c) σ̂(1) = 1
(d) the graph of σ̂ has at most a single intersection in [0, 1) with the graph of f (ρ) = ρ
If the above properties hold we can take ασ to be the intersection point or 1 if such a point
does not exist. We first show (a). For ρ ∈ (−1, 0) we have that
σ̂(ρ) = b0 +
∞
X
i=1
i
bi ρ ≥ b0 −
∞
X
i
bi |ρ| > −
i=1
∞
X
bi |ρ| ≥ −|ρ| = ρ .
i=1
Here, the third inequality follows form the fact that b0 ≥ 0 and for all i, −bi |ρ|i ≥ −bi |ρ|.
Moreover since b1 < 1, one of these inequalities must be strict. Properties (b) and (c) follows
from
Finally, to show (d), we note that the second derivative of σ̂(ρ) − ρ is
P Lemma 11.
i−2
which is non-negative in [0, 1). Hence, σ̂(ρ) − ρ is convex in [0, 1] and in
i≥2 i(i − 1)bi ρ
particular intersects with the x-axis at either 0, 1, 2 or infinitely many times in [0, 1]. As we
assume that σ̂ is not the identity, we can rule out the option of infinitely many intersections.
Also, since σ̂(1) = 1, we know that there is at least one intersection in [0, 1]. Hence, there
are 1 or 2 intersections in [0, 1] and because one of them is in ρ = 1, we conclude that there
is at most one intersection in [0, 1).
19
Lastly, we derive the conclusion of the claim from equation (9). Fix ρ ∈ (−1, 1). Assume
first that ρ ≥ ασ . By (9), αm (ρ) is a monotonically non-increasing sequence that is lower
bounded by ασ . Hence, it has a limit ασ ≤ τ ≤ ρ < 1. Now, by the continuity of σ̂ we have
σ̂(τ ) = σ̂ lim αm (ρ) = lim σ̂(αm (ρ)) = lim αm+1 (ρ) = τ .
m→∞
m→∞
m→∞
Since the only solution to σ̂(ρ) = ρ in (−1, 1) is ασ , we must have τ = ασ . We next
deal with the case that −1 < ρ < ασ . If for some m, αm (ρ) ∈ [ασ , 1), the argument for
ασ ≤ ρ shows that ασ = limm→∞ αm (ρ). If this is not the case, we have that for all m,
αm (ρ) ≤ αm+1 (ρ) ≤ ασ . As in the case of ρ ≥ ασ , this can be used to show that αm (ρ)
converges to ασ .
9
Proofs
9.1
Well-behaved activations
The proof of our main results applies to activations that are decent, i.e. well-behaved, in a
sense defined in the sequel. We then show that C-bounded activations as well as the ReLU
activation are decent. We first need to extend the definition of the dual activation and kernel
to apply to vectors in Rd , rather than just Sd . We denote by M+ the collection of 2 × 2
positive semi-define matrices and by M++ the collection of positive definite matrices.
Definition 10. Let σ be an activation. Define the following,
σ̄ :
M2+
kxk2 hx, yi
→ R , σ̄(Σ) =
E
σ(X)σ(Y ) , kσ (x, y) = σ̄
.
hx, yi kyk2
(X,Y )∼N(0,Σ)
We underscore the following properties of the extension of a dual activation.
(a) The following equality holds,
1 ρ
σ̂(ρ) = σ̄
ρ 1
(b) The restriction of the extended kσ to the sphere agrees with the restricted definition.
(c) The extended dual activation and kernel are defined for every activation σ such that
for all a ≥ 0, x 7→ σ(ax) is square integrable with respect to the Gaussian measure.
(d) For x, y ∈ Rd , if w ∈ Rd is a multivariate normal distribution with zero mean vector
and identity covariance matrix, then
kσ (x, y) = E σ(hw, xi)σ(hw, yi) .
w
Denote
Mγ+
:=
Σ11 Σ12
Σ12 Σ22
∈ M+ | 1 − γ ≤ Σ11 , Σ22 ≤ 1 + γ
20
.
Definition 11. A normalized activation σ is (α, β, γ)-decent for α, β, γ ≥ 0 if the following
conditions hold.
(i) The dual activation σ̄ is β-Lipschitz in Mγ+ with respect to the ∞-norm.
(ii) If (X1 , Y1 ), . . . , (Xr , Yr ) are independent samples from N (0, Σ) for Σ ∈ Mγ+ then
Pr
2
σ(X
)σ(Y
)
r
i
i
i=1
− σ̄(Σ) ≥ ≤ 2 exp − 2 .
Pr
r
2α
Lemma 12 (Bounded activations are decent). Let σ : R → R be a C-bounded normalized
activation. Then, σ is (C 2 , 2C 2 , γ)-decent for all γ ≥ 0.
Proof. It is enough to show that the following properties hold.
1. The (extended) dual activation σ̄ is 2C 2 -Lipschitz in M++ w.r.t. the ∞-norm.
2. If (X1 , Y1 ), . . . , (Xr , Yr ) are independent samples from N (0, Σ) then
Pr
r2
i=1 σ(Xi )σ(Yi )
− σ̄(Σ) ≥ ≤ 2 exp − 4
Pr
r
2C
From the boundedness of σ it holds that |σ(X)σ(Y )| ≤ C 2 . Hence, the second property
follows directly from Hoeffding’s bound. We next prove the first part. Let z = (x, y) and
φ(z) = σ(x)σ(y). Note that for Σ ∈ M++ we have
Z
z> Σ−1 z
1
φ(z)e− 2 dz .
σ̄(Σ) = p
2π det(Σ) R2
Thus we get that,
" p
#
p
1
1
−1
−1
> −1
det(Σ)Σ
−
det(Σ)(Σ
zz
Σ
)
z> Σ−1 z
2
φ(z) 2
e− 2 dz
det(Σ)
R2
Z
z> Σ−1 z
1
1
p
=
φ(z) Σ−1 − Σ−1 zz> Σ−1 e− 2 dz
2
2π det(Σ) R2
∂ σ̄
1
=
∂Σ
2π
Let g(z) = e−
z> Σ−1 z
2
Z
. Then, the first and second order partial derivatives of g are
z> Σ−1 z
∂g
= −Σ−1 ze− 2
∂z
−1
− z> Σ−1 z
∂ 2g
−1
> −1
2
=
−Σ
+
Σ
zz
Σ
e
.
∂ 2z
We therefore obtain that,
∂ σ̄
1
=− p
∂Σ
4π det(Σ)
21
Z
φ
R2
∂ 2g
dz .
∂ 2z
By the product rule we have
∂ σ̄
1
1
=− p
∂Σ
2π det(Σ) 2
Z
R2
2
1
∂ φ
∂ 2φ
gdz = −
E
(X, Y )
∂ 2z
2 (X,Y )∼N(0,Σ) ∂ 2 z
We conclude that σ̄ is differentiable in M++ with partial derivatives that are point-wise
2
bounded by C2 . Thus, σ̄ is 2C 2 -Lipschitz in M+ w.r.t. the ∞-norm.
We next show that the ReLU activation is decent.
Lemma 13 (ReLU is decent). There exists
√ a constant αReLU ≥ 1 such that for 0 ≤ γ ≤ 1,
the normalized ReLU activation σ(x) = 2 max(0, x) is (αReLU , 1 + o(γ), γ)-decent.
Proof. The measure concentration property follows from standard concentration bounds for
sub-exponential random variables (e.g. [53]). It remains to show that σ̄ is (1+o(γ))-Lipschitz
in Mγ+ . We first calculate an exact expression for σ̄. The expression was already calculated
in [13], yet we give here a derivation for completeness.
Claim 2. The following equality holds for all Σ ∈ M2+ ,
p
Σ12
σ̄(Σ) = Σ11 Σ22 σ̂ √
.
Σ11 Σ22
Proof. Let us denote
Σ̃ =
1
√ Σ12
Σ11 Σ12
√ Σ12
Σ11 Σ12
1
!
.
By the positive homogeneity of the ReLU activation we have
σ̄ (Σ) =
E
σ(X)σ(Y )
(X,Y )∼N(0,Σ)
X
Y
E
σ √
σ √
(X,Y )∼N(0,Σ)
Σ11
Σ22
p
=
Σ11 Σ22
E
σ X̃ σ Ỹ
(X̃,Ỹ )∼N(0,Σ̃)
p
Σ12
=
.
Σ11 Σ22 σ̂ √
Σ11 Σ22
p
=
Σ11 Σ22
which concludes the proof.
For brevity, we henceforth drop the argument from σ̄(Σ) and use the abbreviation σ̄. In
order to show that σ̄ is (1 + o(γ))-Lipschitz w.r.t. the ∞-norm it is enough to show that for
every Σ ∈ Mγ+ we have,
k∇σ̄k1 =
∂ σ̄
∂ σ̄
∂ σ̄
+
+
≤ 1 + o(γ) .
∂Σ12
∂Σ11
∂Σ22
22
(10)
First, Note that ∂ σ̄/∂Σ11 and ∂ σ̄/∂Σ22 have the same sign, hence,
k∇σ̄k1 =
∂ σ̄
∂ σ̄
∂ σ̄
+
+
.
∂Σ12
∂Σ11 ∂Σ22
Next we get that,
∂ σ̄
∂Σ22
∂ σ̄
∂Σ12
r
Σ12
Σ22
σ̂ √
−
Σ11
Σ11 Σ22
r
1 Σ11
Σ12
=
σ̂ √
−
2 Σ22
Σ11 Σ22
Σ12
0
= σ̂ √
.
Σ11 Σ22
1
∂ σ̄
=
∂Σ11
2
r
Σ12
Σ22 Σ12
0
√
σ̂ √
Σ11 Σ11 Σ22
Σ11 Σ22
r
1 Σ11 Σ12
Σ12
0
√
σ̂ √
2 Σ22 Σ11 Σ22
Σ11 Σ22
1
2
We therefore get that the 1-norm of ∇σ̄ is,
Σ12
Σ12
Σ12
Σ12
1 Σ11 + Σ22
0
0
σ̂ √
σ̂ √
+ σ̂ √
−√
.
k∇σ̄k1 = √
2 Σ11 Σ22
Σ11 Σ22
Σ11 Σ22
Σ11 Σ22
Σ11 Σ22
Σ11 +Σ22
The gradient of 12 √
at (Σ11 , Σ22 ) = (1, 1) is (0, 0). Therefore, from the mean value
Σ11 Σ22
Σ11 +Σ22
1√
theorem we get, 2 Σ11 Σ22 = 1 + o(γ). Furthermore, σ̂, σ̂ 0 and √ΣΣ1112Σ22 are bounded by 1 in
absolute value. Hence, we can write,
Σ12
Σ12
Σ12
Σ12
0
0
σ̂ √
+ σ̂ √
−√
+ o(γ) .
k∇σ̄k1 = σ̂ √
Σ11 Σ22
Σ11 Σ22
Σ11 Σ22
Σ11 Σ22
Finally, if we let t =
√ Σ12
,
Σ11 Σ22
we can further simply the expression for ∇σ̄,
k∇σ̄(Σ)k1 = |σ̂(t) − tσ̂ 0 (t)| + |σ̂ 0 (t)| + o(γ)
√
cos−1 (t)
1 − t2
+1−
+ o(γ) .
=
π
π
√
−1
Finally, the proof is obtained from the fact that the function f (t) = 1−t
+ 1 − cos π (t)
π
satisfies 0 ≤ f (t) ≤ 1 for every t ∈ [−1, 1]. Indeed, it is simple to verify that f (−1) = 0 and
f (1) = 1. Hence, it suffices to show that f 0 is non-negative in [−1, 1] which is indeed the
case since,
r
1 1−t
1 1−t
0
f (t) = √
=
≥ 0.
π 1 − t2
π 1+t
9.2
2
Proofs of Thms. 2 and 3
We start by an additional theorem which serves as a simple stepping stone for proving the
aforementioned main theorems.
23
Theorem
Pd−1 i 14. Let S be a skeleton with (α, β, γ)-decent activations, 0 < ≤ γ, and Bd =
i=0 β . Let w be a random initialization of the network N = N (S, r) with
2 2
2α Bdepth(S) log 8|S|
δ
.
r≥
2
Then, for every x, y with probability of at least 1 − δ, it holds that
|κw (x, y) − κS (x, y)| ≤ .
Before proving the theorem we show that together with Lemmas 12 and 13, Theorems 2 and 3
follow from Theorem 14. We restate them as corollaries, prove them, and then proceed to
the proof of Theorem 14.
Corollary 15. Let S be a skeleton with C-bounded activations. Let w be a random initialization of N = N (S, r) with
4 depth(S)+1
(4C )
log 8|S|
δ
r≥
.
2
Then, for every x, y, w.p. ≥ 1 − δ,
|κw (x, y) − κS (x, y)| ≤ .
Proof. From Lemma 12, for all γ > 0, each activation is (C 2 , 2C 2 , γ)-decent. By Theorem
14, it suffices to show that
2
depth(S)−1
X
2
2 C2
(2C 2 )i ≤ (4C 4 )depth(S)+1 .
i=0
The sum of can be bounded above by,
depth(S)−1
X
i=0
(2C 2 )i =
(2C 2 )depth(S) − 1
(2C 2 )depth(S)
≤
.
2C 2 − 1
C2
Therefore, we get that,
2
depth(S)−1
X
2
2C 4 (4C 4 )depth(S)
2 C2
(2C 2 )i ≤
≤ (4C 4 )depth(S)+1 ,
4
C
i=0
which concludes the proof.
24
Corollary 16. Let S be a skeleton with ReLU activations, and w a random initialization of
8|S|
depth2 (S) log( δ )
1
N (S, r) with r ≥ c1
. For all x, y and ≤ min(c2 , depth(S)
), w.p. ≥ 1 − δ,
2
|κw (x, y) − κS (x, y)| ≤
Here, c1 , c2 > 0 are universal constants.
Proof. From Lemma 13, each activation is (αReLU , 1 + o(), )-decent. By Theorem 14, it is
enough to show that
depth(S)−1
X
(1 + o())i = O(depth(S)) .
i=0
This claim follows from the fact that (1 + o())i ≤ eo()depth(S) as long as i ≤ depth(S). Since
we assume that ≤ 1/depth(S), the expression is bounded by e for sufficiently small .
We next prove Theorem 14.
Proof. (Theorem 14) For a node u ∈ S we denote by Ψu,w : X → Rr the normalized
representation of S’s sub-skeleton rooted at u. Analogously, κu,w denotes the empirical
kernel of that network. When u is the output node of S we still use Ψw and κw for Ψu,w
and κu,w . Given two fixed x, y ∈ X and a node u ∈ S, we denote
κu,w (x, x) κu,w (x, y)
κu (x, x) κu (x, y)
u
u
Kw =
, K =
κu,w (x, y) κu,w (y, y)
κu (x, y) κu (y, y)
P
P
v
v
v∈in(u) K
v∈in(u) Kw
←u
←u
, K =
.
Kw =
|in(u)|
|in(u)|
For a matrix K ∈ M+ and a function f : M+ → R, we denote
K11 K11
f (K)
f K11 K11
p
f (K) =
K22 K22
f (K)
f
K22 K22
Note that Ku = σ̄up (K←u ). We say that a node u ∈ S, is well-initialized if
u
kKw
− K u k∞ ≤
Bdepth(u)
.
Bdepth(S)
(11)
Here, we use the convention that B0 = 0. It is enough to show that with probability of at
least ≥ 1 − δ all nodes are well-initialized. We first note that input nodes are well-initialized
u
by construction since Kw
= Ku . Next, we show that given that all incoming nodes for a
certain node are well-initialized, then w.h.p. the node is well-initialized as well.
Claim 3. Assume that all the nodes in in(u) are well-initialized. Then, the node u is wellδ
initialized with probability of at least 1 − |S|
.
25
u
Proof. It is easy to verify that Kw
is the empirical covariance matrix of r independent
←u
variables distributed according to (σ(X), σ(Y )) where (X, Y ) ∼ N (0, Kw
). Given the
assumption that all nodes incoming to u are well-initialized, we have,
P
P
v
v
K
w
v∈in(v) K
v∈in(v)
←u
←u
−
kKw − K k∞ =
|in(v)|
|in(v)|
∞
X
1
v
v
kKw − K k∞
(12)
≤
|in(v)|
v∈in(v)
≤
Bdepth(u)−1
.
Bdepth(S)
←u
Further, since ≤ γ then Kw
∈ Mγ+ . Using the fact that σu is (α, β, γ)-decent and that
8|S|
2
2α2 Bdepth(S)
log( δ )
δ
r≥
,
, we get that w.p. of at least 1 − |S|
2
u
←u
kKw
− σ̄up (Kw
)k∞ ≤
Bdepth(S)
.
(13)
Finally, using (12) and (13) along with the fact that σ̄ is β-Lipschitz, we have
u
u
kKw
− Ku k∞ = kKw
− σ̄up (K←u )k∞
u
←u
←u
≤ kKw
− σ̄up (Kw
)k∞ + kσ̄up (Kw
) − σ̄up (K←u )k∞
←u
+ β kKw
− K←u k∞
≤
Bdepth(S)
Bdepth(u)−1
Bdepth(u)
+ β
=
.
≤
Bdepth(S)
Bdepth(S)
Bdepth(S)
We are now ready to conclude the proof. Let u1 , . . . , u|S| be an ordered list of the nodes
in S in accordance to their depth, starting with the shallowest nodes, and ending with the
output node. Denote by Aq the event that u1 , . . . , uq are well-initialized. We need to show
qδ
.
that Pr(A|S| ) ≥ 1 − δ. We do so using an induction on q for the inequality Pr(Aq ) ≥ 1 − |S|
Indeed, for q = 1, . . . , n, uq is an input node and Pr(Aq ) = 1. Thus, the base of the induction
δ
hypothesis holds. Assume that q > n. By Claim (3) we have that Pr(Aq |Aq−1 ) ≥ 1 − |S|
.
Finally, from the induction hypothesis we have,
δ
(q − 1)δ
qδ
Pr(Aq ) ≥ Pr(Aq |Aq−1 ) Pr(Aq−1 ) ≥ 1 −
1−
≥1−
.
|S|
|S|
|S|
9.3
Proofs of Thms. 4 and 5
Theorems 4 and 5 follow from using the following lemma combined with Theorems 2 and 3.
When we apply the lemma, we always focus on the special case where one of the kernels is
constant w.p. 1.
26
Lemma 17. Let D be a distribution on X × Y, ` : R × Y → R be an L-Lipschitz loss,
δ > 0, and κ1 , κ2 : X × X → R be two independent random kernels sample from arbitrary
distributions. Assume that the following properties hold.
• For some C > 0, ∀x ∈ X , κ1 (x, x), κ2 (x, x) ≤ C.
2
δ
• ∀x, y ∈ X , Prκ1 ,κ2 (|κ1 (x, y) − κ2 (x, y)| ≥ ) ≤ δ̃ for δ̃ < c2 C 2 log
2
( 1δ )
where c2 > 0 is a
universal constant.
√
Then, w.p. ≥ 1 − δ over√the choices of κ1 , κ2 , for every f1 ∈ HκM1 there is f2 ∈ Hκ22M such
that LD (f2 ) ≤ LD (f1 ) + 4LM .
To prove the above lemma, we state another lemma below followed by a basic measure
concentration result.
d
∗
d
Lemma 18. Let
Pmx1 , . . . , xm ∈ R , w ∈ R and > 0. There are weights α1 , . . . , αm such
that for w := i=1 αi xi we have,
P
∗
• L(w) := m1 m
i=1 |hw, xi i − hw , xi i| ≤
P
kw∗ k2
•
i |αi | ≤
• kwk ≤ kw∗ k
Proof. Denote M = kw∗ k, C = maxi kxi k, and yi = hw∗ , xi i. Suppose that we run stochastic
gradient decent on the sample {(x1 , y1 ), . . . , (xm , ym )} w.r.t. the loss L(w), with learning rate
η = C2 , and with projections onto the ball of radius M . Namely, we start with w0 = 0 and
at each iteration t ≥ 1, we choose at random it ∈ [m] and perform the update,
(
wt−1 − ηxit hwt−1 , xit i ≥ yit
w̃t =
wt−1 + ηxit hwt−1 , xit i < yit
(
w̃t
kw̃t k ≤ M
wt = M w̃t
kw̃t k > M
kw̃t k
2
2
After T = M2C iterations the loss in expectation would be at most (see for instance
2 2
Chapter 14 in [53]). In particular, there exists a sequence of at most M2C gradient steps
that attains a solution w with L(w) ≤ . Each update adds or subtracts C2 xi from the
current solution. Hence w can be written as a weighted sum of xi ’s where the sum of each
2
coefficient is at most T C2 = M .
Theorem 19 (Bartlett and Mendelson [8]). Let D be a distribution over X ×Y, ` : R×Y → R
a 1-Lipschitz loss, κ : X × X → R a kernel, and , δ > 0. Let S = {(x1 , y1 ), . . . , (xm , ym )}
M 2 maxx∈X κ(x,x)+log( 1δ )
be i.i.d. samples from D such that m ≥ c
where c is a constant. Then,
2
with probability of at least 1 − δ we have,
∀f ∈ HκM , |LD (f ) − LS (f )| ≤ .
27
√
Proof. (of Lemma 17) By rescaling `, we can assume w.l.o.g that L = 1. Let 1 = M and
S = {(x1 , y1 ), . . . , (xm , ym )} ∼ D be i.i.d. samples which are independent of the choice of
CM 2 log( 1δ )
C log( 1δ )
κ1 , κ2 . By Theorem 19, for a large enough constant c, if m = c
=
c
, then
2
w.p. ≥ 1 −
δ
2
1
over the choice of the samples we have,
√
∀f ∈ HκM1 ∪ Hκ22M , |LD (f ) − LS (f )| ≤ 1
Now, if we choose c2 = 2c12 then w.p. ≥ 1 − m2 δ̃ ≥ 1 −
and the kernel), we have that
δ
2
(14)
(over the choice of the examples
∀i, j ∈ [m], |κ1 (xi , xj ) − κ2 (xi , xj )| < .
(15)
In particular, w.p. ≥ 1 − δ (14) and (15) hold and therefore it suffices to prove the conclusion
of the theorem under these conditions. Indeed, let Ψ1 , Ψ2 : X → H be two mapping from X
to a Hilbert space H so that κi (x, y) = hΨ
Ψi (y)i. Let f1 ∈ HκM1 . By lemma 18 there
Pi (x),
m
are α1 , . . . , αm so that for the vector w = i=1 α1 Ψ1 (xi ) we have
m
1 X
|hw, Ψ1 (xi )i − f1 (xi )| ≤ 1 , kwk ≤ M ,
m i=1
and
m
X
(16)
M2
|αi | ≤
.
(17)
1
i=1
P
Consider the function f2 ∈ H2 defined by f2 (x) = m
i=1 α1 hΨ2 (xi ), Ψ2 (x)i. We note that
kf2 k2Hk
2
≤
m
X
2
αi Ψ2 (xi )
i=1
=
≤
m
X
i,j=1
m
X
αi αj κ2 (xi , xj )
αi αj κ1 (xi , xj ) +
i,j=1
m
X
|αi αj |
i,j=1
m
X
2
= kwk +
!2
|αi |
i=1
≤ M2 +
M4
= 2M 2 .
21
Denote by f˜1 (x) = hw, Ψ1 (x)i and note that for every i ∈ [m] we have,
|f˜1 (xi ) − f2 (xi )| =
m
X
≤
αj (κ1 (xi , xj ) − κ2 (xi , xj ))
j=1
m
X
|αi | ≤
i=1
28
M2
= 1 .
1
Finally, we get that,
LD (f2 ) ≤ LS (f2 ) + 1
m
1 X
=
` (f2 (xi ), yi ) + 1
m i=1
m
1 X ˜
≤
` f1 (xi ), yi + 1 + 1
m i=1
m
1 X
≤
` (f1 (xi ), yi ) + |f˜1 (xi ) − f1 (xi )| + 21
m i=1
m
≤
1 X
` (f1 (xi ), yi ) + 31
m i=1
≤ LS (f1 ) + 31 ≤ LD (f1 ) + 41 ,
which concludes the proof.
10
Discussion
Role of initialization and training. Our results surface the question of the extent to
which random initialization accounts for the success of neural networks. While we mostly
leave this question for future research, we would like to point to empirical evidence supporting the important role of initialization. First, numerous researchers and practitioners
demonstrated that random initialization, similar to the scheme we analyze, is crucial to the
success of neural network learning (see for instance [20]). This suggests that starting from
arbitrary weights is unlikely to lead to a good solution. Second, several studies show that the
contribution of optimizing the representation layers is relatively small [49, 26, 44, 43, 15]. For
example, competitive accuracy on CIFAR-10, STL-10, MNIST and MONO datasets can be
achieved by optimizing merely the last layer [36, 49]. Furthermore, Saxe et al. [49] show that
the performance of training the last layer is quite correlated with training the entire network.
The effectiveness of optimizing solely the last layer is also manifested by the popularity of
the random features paradigm [46]. Finally, other studies show that the metrics induced by
the initial and fully trained representations are not substantially different. Indeed, Giryes et
al. [19] demonstrated that for the MNIST and CIFAR-10 datasets the distances’ histogram of
different examples barely changes when moving from the initial to the trained representation.
For the ImageNet dataset the difference is more pronounced yet still moderate.
The role of architecture. By using skeletons and compositional kernel spaces, we can
reason about functions that the network can actually learn rather than merely express. This
may explain in retrospect past architectural choices and potentially guide future choices. Let
us consider for example the task of object recognition. It appears intuitive, and is supported
by visual processing mechanisms in mammals, that in order to perform object recognition,
29
the first processing stages are confined to local receptive fields. Then, the result of the local
computations are applied to detect more complex shapes which are further combined towards
a prediction. This processing scheme is naturally expressed by convolutional skeletons. A
two dimensional version of Example 1 demonstrates the usefulness of convolutional networks
for vision and speech applications.
The rationale we described above was pioneered by LeCun and colleagues [32]. Alas,
the mere fact that a network can express desired functions does not guarantee that it can
actually learn them. Using for example Barron’s theorem [7], one may claim that visionrelated functions are expressed by fully connected two layer networks, but such networks are
inferior to convolutional networks in machine vision applications. Our result mitigates this
gap. First, it enables use of the original intuition behind convolutional networks in order to
design function spaces that are provably learnable. Second, as detailed in Example 1, it also
explains why convolutional networks perform better than fully connected networks.
The role of other architectural choices. In addition to the general topology of the
network, our theory can be useful for understanding and guiding other architectural choices.
We give two examples. First, suppose that a skeleton S has a fully connected layer with the
dual activation σ̂1 , followed by an additional fully connected layer with dual activation σ̂2 .
It is straightforward to verify that if these two layers are replaced by a single layer with dual
activation σ̂2 ◦ σ̂1 , the corresponding compositional kernel space remains the same. This
simple observation can be useful in potentially saving a whole layer in the corresponding
networks.
The second example is concerned with the ReLU activation, which is one of the most
common activations used in practice. Our theory suggests a somewhat surprising explanation for its usefulness. First, the dual kernel of the ReLU activation enables expression of
non-linear functions. However, this property holds true for many activations. Second, Theorem 3 shows that even for quite deep networks with ReLU activations, random initialization
approximates the corresponding kernel. While we lack a proof at the time of writing, we
conjecture that this property holds true for many other activations. What is then so special
about the ReLU? Well, an additional property of the ReLU is being positive homogeneous,
i.e. satisfying σ(ax) = aσ(x) for all a ≥ 0. This fact makes the ReLU activation robust
to small perturbations in the distribution used for initialization. Concretely, if we multiply
the variance of the random weights by a constant, the distribution of the generated representation and the space Hw remain the same up to a scaling. Note moreover that training
algorithms are sensitive to the initialization. Our initialization is very similar to approaches
used in practice, but encompasses a small “correction”, in the form of a multiplication by a
small constant which depends on the activation. For most activations, ignoring this correction, especially in deep networks, results in a large change in the generated representation.
The ReLU activation is more robust to such changes. We note that similar reasoning applies
to the max-pooling operation.
30
Future work. Though our formalism is fairly general, we mostly analyzed fully connected
and convolutional layers. Intriguing questions remain, such as the analysis of max-pooling
and recursive neural network components from the dual perspective. On the algorithmic
side, it is yet to be seen whether our framework can help in understanding procedures such
as dropout [54] and batch-normalization [25]. Beside studying existing elements of neural
network learning, it would be interesting to devise new architectural components inspired
by duality. More concrete questions are concerned with quantitative improvements of the
main results. In particular, it remains open whether the dependence on 2O(depth(S)) can be
made polynomial and the quartic dependence on 1/, R, and L can be improved. In addition
to being interesting in their own right, improving the bounds may further underscore the
effectiveness of random initialization as a way of generating low dimensional embeddings of
compositional kernel spaces. Randomly generating such embeddings can be also considered
on its own, and we are currently working on design and analysis of random features a la
Rahimi and Recht [45].
Acknowledgments
We would like to thank Yossi Arjevani, Elad Eban, Moritz Hardt, Elad Hazan, Percy Liang,
Nati Linial, Ben Recht, and Shai Shalev-Shwartz for fruitful discussions, comments, and
suggestions.
References
[1] A. Andoni, R. Panigrahy, G. Valiant, and L. Zhang. Learning polynomials with neural
networks. In Proceedings of the 31st International Conference on Machine Learning,
pages 1908–1916, 2014.
[2] F. Anselmi, L. Rosasco, C. Tan, and T. Poggio. Deep convolutional networks are
hierarchical kernel machines. arXiv:1508.01084, 2015.
[3] M. Anthony and P. Bartlet. Neural Network Learning: Theoretical Foundations. Cambridge University Press, 1999.
[4] S. Arora, A. Bhaskara, R. Ge, and T. Ma. Provable bounds for learning some deep representations. In Proceedings of The 31st International Conference on Machine Learning,
pages 584–592, 2014.
[5] F. Bach.
Breaking the curse of dimensionality with convex neural networks.
arXiv:1412.8690, 2014.
[6] F. Bach. On the equivalence between kernel quadrature rules and random feature
expansions. 2015.
[7] A.R. Barron. Universal approximation bounds for superposition of a sigmoidal function.
IEEE Transactions on Information Theory, 39(3):930–945, 1993.
31
[8] P. L. Bartlett and S. Mendelson. Rademacher and Gaussian complexities: Risk bounds
and structural results. Journal of Machine Learning Research, 3:463–482, 2002.
[9] P.L. Bartlett. The sample complexity of pattern classification with neural networks: the
size of the weights is more important than the size of the network. IEEE Transactions
on Information Theory, 44(2):525–536, March 1998.
[10] E.B. Baum and D. Haussler. What size net gives valid generalization? Neural Computation, 1(1):151–160, 1989.
[11] L. Bo, K. Lai, X. Ren, and D. Fox. Object recognition with hierarchical kernel descriptors. In Computer Vision and Pattern Recognition (CVPR), 2011 IEEE Conference on,
pages 1729–1736. IEEE, 2011.
[12] J. Bruna and S. Mallat. Invariant scattering convolution networks. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 35(8):1872–1886, 2013.
[13] Y. Cho and L.K. Saul. Kernel methods for deep learning. In Advances in neural
information processing systems, pages 342–350, 2009.
[14] A. Choromanska, M. Henaff, M. Mathieu, G. Ben Arous, and Y. LeCun. The loss
surfaces of multilayer networks. In AISTATS, pages 192–204, 2015.
[15] D. Cox and N. Pinto. Beyond simple features: A large-scale feature search approach to
unconstrained face recognition. In Automatic Face & Gesture Recognition and Workshops (FG 2011), 2011 IEEE International Conference on, pages 8–15. IEEE, 2011.
[16] A. Daniely. Complexity theoretic limitations on learning halfspaces. In STOC, 2016.
[17] A. Daniely and S. Shalev-Shwartz. Complexity theoretic limitations on learning DNFs.
In COLT, 2016.
[18] A. Daniely, N. Linial, and S. Shalev-Shwartz. From average case complexity to improper
learning complexity. In STOC, 2014.
[19] R. Giryes, G. Sapiro, and A.M. Bronstein. Deep neural networks with random gaussian
weights: A universal classification strategy? arXiv preprint arXiv:1504.08291, 2015.
[20] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward
neural networks. In International conference on artificial intelligence and statistics,
pages 249–256, 2010.
[21] K. Grauman and T. Darrell. The pyramid match kernel: Discriminative classification
with sets of image features. In Tenth IEEE International Conference on Computer
Vision, volume 2, pages 1458–1465, 2005.
[22] M. Hardt, B. Recht, and Y. Singer. Train faster, generalize better: Stability of stochastic
gradient descent. arXiv:1509.01240, 2015.
32
[23] Z.S. Harris. Distributional structure. Word, 1954.
[24] T. Hazan and T. Jaakkola. Steps toward deep kernel methods from infinite neural
networks. arXiv:1508.05133, 2015.
[25] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. arXiv:1502.03167, 2015.
[26] K. Jarrett, K. Kavukcuoglu, M. Ranzato, and Y. LeCun. What is the best multi-stage
architecture for object recognition? In Computer Vision, 2009 IEEE 12th International
Conference on, pages 2146–2153. IEEE, 2009.
[27] P. Kar and H. Karnick. Random feature maps for dot product kernels. arXiv:1201.6530,
2012.
[28] R.M. Karp and R.J. Lipton. Some connections between nonuniform and uniform complexity classes. In Proceedings of the twelfth annual ACM symposium on Theory of
computing, pages 302–309. ACM, 1980.
[29] M. Kearns and L.G. Valiant. Cryptographic limitations on learning Boolean formulae
and finite automata. In STOC, pages 433–444, May 1989.
[30] A.R. Klivans and A.A. Sherstov. Cryptographic hardness for learning intersections of
halfspaces. In FOCS, 2006.
[31] A. Krizhevsky, I. Sutskever, and G.E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages
1097–1105, 2012.
[32] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to
document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998.
[33] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521(7553):436–444, 2015.
[34] O. Levy and Y. Goldberg. Neural word embedding as implicit matrix factorization. In
Advances in Neural Information Processing Systems, pages 2177–2185, 2014.
[35] R. Livni, S. Shalev-Shwartz, and O. Shamir. On the computational efficiency of training
neural networks. In Advances in Neural Information Processing Systems, pages 855–863,
2014.
[36] J. Mairal, P. Koniusz, Z. Harchaoui, and Cordelia Schmid. Convolutional kernel networks. In Advances in Neural Information Processing Systems, pages 2627–2635, 2014.
[37] T. Mikolov, I. Sutskever, K. Chen, G.S. Corrado, and J. Dean. Distributed representations of words and phrases and their compositionality. In NIPS, pages 3111–3119,
2013.
33
[38] R.M. Neal. Bayesian learning for neural networks, volume 118. Springer Science &
Business Media, 2012.
[39] B. Neyshabur, R. R Salakhutdinov, and N. Srebro. Path-SGD: Path-normalized optimization in deep neural networks. In Advances in Neural Information Processing
Systems, pages 2413–2421, 2015.
[40] B. Neyshabur, N. Srebro, and R. Tomioka. Norm-based capacity control in neural
networks. In COLT, 2015.
[41] R. O’Donnell. Analysis of boolean functions. Cambridge University Press, 2014.
[42] J. Pennington, F. Yu, and S. Kumar. Spherical random features for polynomial kernels.
In Advances in Neural Information Processing Systems, pages 1837–1845, 2015.
[43] N. Pinto and D. Cox. An evaluation of the invariance properties of a biologicallyinspired system for unconstrained face recognition. In Bio-Inspired Models of Network,
Information, and Computing Systems, pages 505–518. Springer, 2012.
[44] N. Pinto, D. Doukhan, J.J. DiCarlo, and D.D. Cox. A high-throughput screening
approach to discovering good forms of biologically inspired visual representation. PLoS
Computational Biology, 5(11):e1000579, 2009.
[45] A. Rahimi and B. Recht. Random features for large-scale kernel machines. In NIPS,
pages 1177–1184, 2007.
[46] A. Rahimi and B. Recht. Weighted sums of random kitchen sinks: Replacing minimization with randomization in learning. In Advances in neural information processing
systems, pages 1313–1320, 2009.
[47] I. Safran and O. Shamir. On the quality of the initial basin in overspecified neural
networks. arxiv:1511.04210, 2015.
[48] S. Saitoh. Theory of reproducing kernels and its applications. Longman Scientific &
Technical England, 1988.
[49] A. Saxe, P.W. Koh, Z. Chen, M. Bhand, B. Suresh, and A.Y. Ng. On random weights
and unsupervised feature learning. In Proceedings of the 28th International Conference
on Machine Learning (ICML-11), pages 1089–1096, 2011.
[50] I.J. Schoenberg et al. Positive definite functions on spheres. Duke Mathematical Journal,
9(1):96–108, 1942.
[51] B. Schölkopf, P. Simard, A. Smola, and V. Vapnik. Prior knowledge in support vector
kernels. In Advances in Neural Information Processing Systems 10, pages 640–646. MIT
Press, 1998.
34
[52] H. Sedghi and A. Anandkumar. Provable methods for training neural networks with
sparse connectivity. arXiv:1412.2693, 2014.
[53] S. Shalev-Shwartz and S. Ben-David. Understanding Machine Learning: From Theory
to Algorithms. Cambridge University Press, 2014.
[54] N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout:
A simple way to prevent neural networks from overfitting. The Journal of Machine
Learning Research, 15(1):1929–1958, 2014.
[55] I. Sutskever, O. Vinyals, and Q.V. Le. Sequence to sequence learning with neural
networks. In Advances in neural information processing systems, pages 3104–3112,
2014.
[56] C.K.I. Williams. Computation with infinite neural networks. pages 295–301, 1997.
35
| 2cs.AI
|
1
Centralized versus Decentralized
Infrastructure Networks
arXiv:1510.08792v1 [physics.soc-ph] 29 Oct 2015
Paul D. H. Hines, Senior Member, IEEE, Seth Blumsack, Member, IEEE, and Markus Schläpfer
Abstract—While many large infrastructure networks, such as
power, water, and natural gas systems, have similar physical
properties governing flows, these systems tend to have distinctly
different sizes and topological structures. This paper seeks
to understand how these different size-scales and topological
features can emerge from relatively simple design principles.
Specifically, we seek to describe the conditions under which it is
optimal to build decentralized network infrastructures, such as
a microgrid, rather than centralized ones, such as a large highvoltage power system. While our method is simple it is useful
in explaining why sometimes, but not always, it is economical
to build large, interconnected networks and in other cases it
is preferable to use smaller, distributed systems. The results
indicate that there is not a single set of infrastructure cost
conditions under which optimally-designed networks will have
highly centralized architectures. Instead, as costs increase we
find that average network sizes increase gradually according to
a power-law. When we consider the reliability costs, however,
we do observe a transition point at which optimally designed
networks become more centralized with larger geographic scope.
As the losses associated with node and edge failures become more
costly, this transition becomes more sudden.
I. I NTRODUCTION
One of the key goals of the United States and its allies
during the United States/NATO conflict in Afghanistan was
the improvement of infrastructure, particularly electricity infrastructure, in order to build good will among the Afghan
people. One of these major infrastructure projects was the
upgrade of the Kajaki Hydroelectric plant from 33 to 51 MW, a
project designed to bring additional power to Kandahar, about
80 km to the southeast. However, moving the necessary heavy
equipment from Kandahar to Kajaki through hostile territory
proved to be one of the most difficult and costly operations
of the Afghanistan conflict [1]. And in the end (at least as of
late 2015) the project was never completed because it proved
impossible to move the concrete necessary to complete the
project. At around the same time, the U.S. Agency for International Development (USAID) embarked on a less ambitious
project to install several smaller (10 MW) diesel power plants
in the outskirts of Kandahar. These plants are more expensive
P.H. was supported by NSF awards ECCS-1254549 and DGE-1144388,
and DTRA award HDTRA110-1-0088. S.B. acknowledges support from NSF
award CNS-1331761. M.S. was supported by the Army Research Office
Minerva Programme (grant no. W911NF-12-1-0097).
P. Hines is with the University of Vermont, School of Engineering and
Complex Systems Center, Burlington, USA, [email protected]
S. Blumsack is with Penn State University, University Park PA, USA,
[email protected]
M. Schläpfer is with the Santa Fe Institute, Santa Fe, USA,
[email protected]
to operate given that they require diesel fuel, but because they
were built close to the city, they were not nearly as complicated
to install and continue to provide relatively reliable power
for residents in that portion of the city [1]. While there are
many factors that contributed to the demise of the Kajaki
Hydroelectric plant and the relative success of the more
distributed diesel plants, it seems reasonable to ask whether
a solution that is less reliant on long-distance transmission is,
under these particular conditions, fundamentally better. More
generally, one might surmise that there exist general conditions
under which decentralized solutions are fundamentally better.
But what are those conditions?
Consider a second example: water distribution networks. In
the United States alone there are more than 150,000 public
drinking water networks that serve at least 25 people [2].
In contrast, there are only 3 power networks: the Eastern,
Western, and Texas interconnections. Why is it that in the
case of drinking water, smaller, more decentralized networks
seem to be optimal, whereas for electric power, larger systems
that span continents seem to be preferable? Both systems have
similar physical properties that govern flows. Both systems
transport largely interchangeable goods: one electron is as
good as another, just as one water molecule is as good as
another (given appropriate standards for cleanliness). However,
these two systems have fundamentally different size scales.
A drinking water system serves, on average, 2000 people. A
power network serves, on average, 100,000,000 people.
Motivated by these, and many other, examples, this paper
seeks to identify conditions under which it may be preferable
to build and maintain large, centralized, interconnected infrastructure networks, versus constructing smaller, decentralized
networks that effectively operate independently from one another.
Large infrastructure systems are designed to deliver services
in a way that balances a variety of potentially conflicting
objectives including economic cost, environmental impact and
reliability. These fundamental trade-offs become particularly
challenging when societies face the potential for rapid infrastructure transitions. This paper is motivated by two such
distinct global infrastructure transitions.
The first transition is the growing push toward decentralized
electric energy systems in more developed countries. The
energy infrastructures in most industrialized countries have
evolved into complex network structures [3]. However the
growing movement toward the use of microgrids is a push back
toward small, relatively independent systems [4]. The historical
case for large interconnected systems (economies of scale
in generation and transmission) and increased redundancy
through interconnection is being challenged by falling costs
2
for distributed power generation [5], increased interest in smart
micro-grids, and the insight that distributed systems may offer
improved local reliability in some cases [6].
The second transition is the rapid growth of infrastructure,
including electric power, in less developed countries [7]. Prior
work has argued that for political, geographic and economic
reasons, the greenfield build-out of highly interconnected electric power infrastructure may not be desirable in developingnation contexts, particularly in locations that are subject to
elevated levels of stress [8].
While the contexts for infrastructure decisions in moredeveloped and less-developed nations differ, the basic question remains the same—given the need to build out power
generation and delivery systems either incrementally (as in
more-developed countries) or as a greenfield project (closer to
the case for many less-developed nations), what mix of smallscale versus large-scale system architectures will best balance
cost and reliability goals? Under what assumptions about cost,
reliability and other factors would it be more advantageous to
make either incremental or greenfield investments in decentralized system architecture? This paper seeks to understand
the conditions under which the transformation of large-scale
systems into multiple smaller-scale systems, or the greenfield
construction of multiple smaller-scale systems to serve a large
geographic area, would yield improvements in cost and other
measures of performance.
Our work is fundamentally concerned with the optimal
planning of networks that deliver services or otherwise provide connectivity over physical space. As such, we model a
single planner making optimal resource decisions, as distinct
from game-theoretic approaches of network generation or the
literature in random generation of synthetic networks [9], [10].
While this topic has been of interest to geographers since the
1960s [11], [12], spatial network design has emerged only
more recently as an area of scientific research [13]. Many
applications of spatial network analysis focus on traffic or
transportation networks [14], [13], [15], [16], [17] or physical
infrastructures that deliver information, such as the internet or
mobile telephony [13], [18]. Others find that the constraints
of geographic space can dramatically change the implications
found in abstract network models [19], [20].
Research on the spatial aspects of network design or performance, as distinct from data-driven empirical investigations of
spatial network structure, has largely focused in two areas. The
first is how the cost of adding edges or otherwise connecting
nodes in space influences network structure and design choices
[13], [21]. The second strand utilizes known or theoretical
spatial properties of networks to understand their performance
in the case of attacks, failures or other contingency events [6],
[17], [22].
We build on this extensive body of work, and add to
its relevance for electrical networks, in two ways. The first
is to embed some most salient properties of electric power
networks (namely Kirchhoff’s Current Law) into the type of
cost-driven spatial network design problem discussed in [13]
and [21]. These properties are important for electric power
networks specifically because while expansion costs may be
straightforward to parametrize in terms of spatial distance,
actual network flows are not so simply represented. Further,
production costs in real electrical networks are heterogeneous
by technology and in space, because of regional variations
in resource endowments and technology choices. The second is to consider a design objective that incorporates the
costs of network operation, network expansion, and network
(un)reliability. While joint planning and operations models
have been devised for incremental expansion decisions [23],
[24], and for optimal topology control applications (e.g., [25],
[26]), our approach is different in its consideration of a
flexible greenfield infrastructure build problem that does not,
for example, restrict infrastructure expansion options to predefined paths or represent branches as binary integer variables.
The utility in our approach is for the discovery of more general
principles describing the optimal geographic scope of network
design.
II.
O PTIMAL I NFRASTRUCTURE N ETWORK D ESIGN
Consider a system planner who has the task of designing
(or modifying) an infrastructure system to provide a particular
infrastructure product (water, natural gas, electricity, etc.) for
a set of a set of n locations (towns, buildings, etc.). Each
location i has some known demand, di , for the product and
also has the ability to produce this product locally with an
incremental production cost, ci , that varies with geography. In
order to satisfy the demand at node i one can either produce
locally at a cost of ci per unit, or build an interconnection to
some nearby node, with the intention of satisfying di at a cost
that is less than ci .
Given a model of this sort we can ask a number of
important questions. Under what conditions is it optimal to
produce locally, rather than building interconnections? If the
goal of the planner is to minimize overall cost, what type of
network would one want to build? Should one build many
small networks or one large one? Should the planner build a
meshed network that allows redundancy, or a radial network
that provides only one path between sources and sinks?
In this section we introduce two relatively simple optimization models that allow one to address questions of this sort.
Both models use a “greenfield” approach, in which we seek to
find the optimal network configuration that satisfies the total
demand for a particular infrastructure product, given a set of
objectives and constraints.
A. Basic model
To start, we assume that each of the n locations is a vertex
(v ∈ V ) that has coordinates xv , yv in some 2d space, demand
dv , a per unit production cost of cv and a maximum potential
production capacity gv . In addition we assume that there is
a maximum feasible set of undirected edges e ∈ E that one
might choose to build. For example one might allow into E
all possible vertex pairs, thus allowing at most n(n − 1) edges.
The cost of building any one particular edge ei↔j depends on
two factors: the length of the edge
q
lei↔j = (xi − xj )2 + (yi − yj )2
(1)
3
and the cost of one unit length and one unit capacity of edge
construction, w.
Given these input data, the following formulation allows one
to compute an “optimal” infrastructure network design.
X
√ X
`e fe
(2a)
min
cv gv + w n
f ,f ,g
v∈V
e∈E
s.t. 0 ≤ gv ≤ gv , ∀v
fe ≥ 0, ∀e
− fe ≤ fe ≤ fe , ∀e
g − d = E| f
(2b)
(2c)
(2d)
(2e)
where fe and fe are the undirected flow capacity and actual
directed flow on
√ edge e, gv is the actual amount of production
at vertex v, w n is an interconnection cost parameter (cost
per-unit length·capacity), and E is an m × n edge matrix with
1 and -1 on the rows indicating the two endpoints for each of
m edges in the maximum feasible network.
Our objective (2a) is to minimize the combined cost of
production c| g and interconnection `| f , while satisfying constraints (2b)-(2e). Constraint (2b) defines locational production
limits; (2c) ensures that we do not build negative quantities of
interconnection capacity; (2d) constrains flows to be less than
the chosen flow limits; and (2e) ensures that the net flow into
and out of each vertex must be zero.
It is important to note that the cost function (2a) is designed
so that both the production and the edge construction cost
terms grow linearly with n. In order to implement this, we
first observe that (at least for the case of uniformly distributed
node locations, see Sec. III-A) edge lengths `e fall with n
according to: `e ∼ n−1/2 . As a result, ensuring linear
√ growth
n, thus
of the edge cost term requires
that
we
multiply
by
√
producing the term w0 = w n.
Implied in this formulation are a number of important
assumptions. First, we assume that interconnections can be
built at any size scale and that construction costs scale linearly
with the capacity of the edge. It is certainly possible to think
of particular examples, such as transmission line construction, where costs are “lumpy,” such that building a 1 MW
transmission line is more than 1/100 of the cost of building
a 100 MW transmission line. However, if we consider that
the edge might be either a large transmission line or a small
distribution line, this assumption is not as obviously incorrect.
And modeling lumpiness of this sort would require knowledge
about the details of a particular infrastructure system at a
particular place and time. In this paper we are more interested
to identify general trends that appear in optimal infrastructure
designs. Second, formulation (2) models a single snapshot of
demand, whereas all real infrastructure systems have demand
that varies in time. As a result the production cost term cv
and the interconnection cost term w meld together the capital
and operating costs associated with supplying the demand dv .
Finally, we make the assumption that nodes do not include
any storage, leading us to include constraint (2e), which is
equivalent to Kirchhoff’s Current Law. Since our model aims
to represent the long-term average operating pattern of a
network, rather than short-term time-domain details, we argue
that this assumption provides at least some insight, even for
networks that do include some storage in them, such as water
networks. Clearly, if one were wanting to design a particular
infrastructure at a particular place and time, one would want to
relax these assumptions and model additional details. However,
this paper focuses on general trends that appear in optimal
network designs.
B. Considering reliability
An obvious limitation in the formulation above is the
complete disregard of reliability. In reality, reliability has an
enormous role in the design of infrastructure systems. In a
power system, for example, electric utilities frequently argue
(in their rate-case filings) that the construction of a new
transmission line is justified purely on reliability grounds. It is
thus useful to understand what impact reliability considerations
have on the topological structure of infrastructure networks.
In order to model the impact of reliability we add a third
term to our objective function (2a) to capture the cost of
(un)reliability. In addition, we separate the production cost
term to include separate terms for capital and operating costs,
since one will sometimes want to build surplus production
capacity to prepare for plausible component failures. With
these additions to the objective and the associated constraints
we get the following formulation.
X
√ X
min
(cv gv + kv gv ) + w n
f e le
f,f ,g,g
v∈V
e∈E
r X |
+
1 ∆d(p)
n
(3a)
p∈P
s.t.
0 ≤ gv ≤ gv , ∀v
0 ≤ |fe | ≤ fe , ∀e
g − d = E| f
∆gv (pv ) = −gv , ∀v ∈ {1 . . . n}
∆fe (pe ) = −fe , ∀e ∈ {1 . . . m}
|fe + ∆fe (p)| ≤ fe , ∀p ∈ P, ∀e ∈ E
0 ≤ gv + ∆gv (p) ≤ gv , ∀p ∈ P, ∀v ∈ V
dv ≤ ∆dv (p) ≤ 0, ∀p ∈ P, ∀v ∈ V
(g + ∆g(p)) − (d + ∆d(p))
= E| (f + ∆f (p))
(3b)
(3c)
(3d)
(3e)
(3f)
(3g)
(3h)
(3i)
(3j)
In this formulation, (3a) is the modified objective, which now
includes the reliability term. In this term, r is the reliability
cost parameter that allows us to adjust the relative importance
of reliability and ∆d(p) is the change (loss) of demand
that results from perturbation p, which is one of the set of
all perturbations, P . Eqns. (3b) and (3d) are equivalent to
(2b)-(2e) in (2). Eqs. (3e) and (3f) cause specific node, pv ,
and edge, pe , failures that together make up the set of all
perturbations, P , by forcing the production or flow to be
zero for the appropriate edge/perturbation combination. While
this approach could be used to model many different types
of failures, here we consider P to be the set of all single
component (either production unit or edge) outages. Eq. (3g)
ensures that all flows are below edge capacities, after all
4
A
1
B
160
0.9
140
0.8
0.7
120
0.6
100
0.5
80
0.4
60
0.3
40
0.2
0.1
0
20
0
Fig. 1. Illustrative results for random placement of “towns” on a 2d plane. Panel A shows an initial, maximum feasible graph for a n = 200 network, with
node colors indicating the production cost, cv , at each “town.” Panel B shows the optimal network configuration, after solving (2a)-(2e) for w = 0.001, with
node colors indicating the amount of production at each node, gv , and edge thicknesses indicating the flow capacity. This optimal network has two connected
components and thus an average component size of < ns >= n/2 = 100.
perturbations. As a result there are m(n + m) constraints
of this type within the formulation. Similarly, (3h) constrains
production at every node after each perturbation (a total of
n(m + n) constraints), to be below the chosen production
capacities for each node. Eq. (3i) ensures that demand can
only decrease, and only up to the total demand at node v, as a
result of each p. Finally, (3j) enforces a nodal supply/demand
balance after each perturbation. This forces the formulation to
compute production, demand loss, and flow patterns that obey
Kirchhoff’s Current Law for each disturbance in P .
As a whole this formulation allows us to observe how
network size and structure changes as we increase the relative
importance of reliability. If r = 0, demand losses are effectively deemed irrelevant, and the problem will produce results
that essentially identical to those obtained from 2. On the other
hand, as r increases we hypothesize that networks are likely to
become more meshed (rather than tree-like) and more likely to
include surplus production capacity. It is not obvious, ex ante,
how r will impact optimal network sizes. On the one hand,
small, local networks will be more robust to edge failures and
thus may be more optimal when reliability is very important.
On the other hand, large interconnected systems provide a high
level of redundancy, which also has tremendous value. In the
sections that follow we explore this tradeoff.
III. R ESULTS
Here we explore the structural properties of the networks
that emerge from formulations (2) and (3), under a variety
of different cost and reliability assumptions. Section III-A
explores the case of nodes distributed uniformly within a 2d
square, ignoring reliability cost. Section III-B extends this to
the reliability case and Sec. III-C applies our approach to data
from the country of Senegal.
A. Uniform distribution of load, ignoring reliability
Consider the case of n nodes randomly located within a 1×1
2d square, such that each node location xv and yv is a uniform
random variable in [0, 1]. Each of these nodes has a production
cost cv that is also a uniform random variable in [0, 1] and a
demand dv = 1. The set of feasible edges that we might decide
to build (the feasible graph E) comes from initially setting E
to be a modified form of the random geometric graph [27].
In this case, rather than connecting each node to nodes that
lie within a fixed radius, for each node i an edge is added to
connect from i to i’s k nearest (Euclidian) neighbors, while
avoiding the addition of duplicate edges. Because it is possible
that j is one of i’s k nearest neighbors, but i is not one of j’s k
nearest neighbors, the resulting E has an average degree that
is slightly larger than k. If one were to set k ≥ n − 1 the
result would be the full graph of n(n − 1)/2 possible edges.
However, optimizing over the full graph makes the resulting
problem computationally impractical for all but the smallest
problems. Instead the results in this section come from the
(somewhat arbitrary) choice of k = 5.
Figure 1 illustrates the application of this approach to a
system with n = 200 nodes and w = 10−3 . From this figure a
few observation can be made. First, we see that the algorithm
tends to produce tree-like graphs in which the number of
edges in each connected component, ms , is one less than
the number of nodes in that component ns . The reason for
this is fairly straightforward: creating a loop means that there
are redundant paths between node pairs. Given a network
with a loop,
P can always reduce the edge-construction
√ one
cost term w n e∈E `e fe by removing one edge in the loop,
without loss of functionality. As a result the graphs that result
from (2) are always treelike, with precisely ms = ns − 1
edges in each component. Secondly, we see that there are
5
two connected components in the optimal network and thus
an average of < ns >= n/2 = 100 nodes per component.
In this illustration the two least-expensive production nodes
had costs of cv1 = 0.0003 and cv2 = 0.0046, with the less
expensive node supplying the larger sub-component. While it
would have been feasible to connect the two components with
a fairly short additional edge, supplying the whole network
from the less expensive node v1 would have required building
additional capacity along the spidery path from v1 to v2 . Doing
so would have cost more than the additional cost of supplying
the second component from the more expensive unit, a cost of
ns2 (cv2 − cv1 ) = 0.149. For comparison purposes, the cost of
building a length `e = 1 edge that could supply
√ the whole of
the 35 node second component would be 35w n = 0.495.
Given that this approach can determine “optimal” network
sizes, it is natural to ask how those network sizes change as the
cost of building network infrastructure changes. For example,
as w increases one might expect to see a relatively sudden
phase transition from optimal networks that span the entire
network to optimal networks with many small, decentralized
sub-systems. In order to investigate this further and understand
the cost conditions under which centralized, or decentralized,
networks are optimal, we performed the following experiment.
For several values of n, we computed optimal infrastructure
networks using (2) over a range of w from 10−4 to 1. For each
value of w we re-initialized the random node locations, the
feasible network E and production costs cv , and computed the
optimal network configuration using (2) 200 times. Then we
recorded the mean size of the connected components < ns >
over the 200 optimal networks.
Mean component size, < ns >
1000
n
n
n
n
=
=
=
=
100
200
500
1000
100
10
1
10−4
10−3
0.01
0.1
1
Interconnection cost, w
Fig. 2. Mean connected component sizes as a function of the interconnection
cost parameter w. For each of the four different network sizes considered, we
find that component sizes scale approximately as w−2/3 .
Figure 2 shows the resulting relationship between the edge
construction cost, w and optimal component sizes. As one
would expect, as network construction costs increase the size
of the optimal network decreases. However, what is somewhat
surprising is that the change from large networks to small
networks does not occur suddenly as does the first-order
phase transition from a solid ice to liquid water. Instead, this
transition occurs gradually over several orders of magnitude
in w. In fact, fitting the data in Figure 2 to a power-law
distribution indicates that mean component sizes fall as
< ns >∼ w−0.648 ∼ w−2/3 .
(4)
B. Uniform distribution of load, with reliability
Next we turn our attention to understanding how the results
described above change after modeling reliability costs as
in (3). As in the simple model, we consider nodes scattered
uniformly on a 2d plane. Also as before, we assume that
each location has an overall production cost that is a uniform
random variable in [0, 1], however unlike in the simple model
we assume that this cost is split evenly between marginal and
capital costs, cv and kv . Here we restrict our attention to the
case of networks with n = 100 nodes, since solving (3) for
larger systems leads to prohibitively large solution times.
First we show a few illustrative results for a n = 50 node
network (see Fig. 3) that clearly show the importance of reliability to network structure. For small values of r the solutions
are nearly identical to what we get from the simple model: treelike network that satisfy demand with no redundancy. However
as r increases, we find a (rather sudden) transition to meshed
networks that include substantial supply redundancy. For the
example in Fig. 3, the system
P builds a network with total
generation capacity equal to
v∈V gv = 82.2, much more
than what is needed to supply the 50 nodes in the system.
Next we computed optimal networks for several different
values of the reliability parameter r and interconnection costs
w for 100 nodes networks For each value of w and r the
random variables (xv , yv and cv ) were re-initialized 100 times
in order to minimize variance.1 Figure 4 shows the resulting
mean component sizes for various values of w and r. For
w ≤ 0.001 and w ≥ 0.1 the network sizes do not change substantially with r. However for the intermediate case w = 0.01
we see a sudden jump in optimal network size as r passes
0.01. For small r the optimal size is around 20 nodes, whereas
as reliability becomes more important the optimal network
size increases toward the size of the network. We also see
a more sensitive relationship between component sizes and w
as r increases. For example, Fig. 4B shows that for r = 1
optimal network sizes suddenly decrease from full networks
< ns >= n to relatively small ones < ns >∼
= 4, as w
increases above 0.03.
Not only do the optimal network sizes change, but the level
of redundancy also changes with r and w. One way to measure
the level of redundancy is by the number of edges constructed
in the optimal network. In the tree-like networks that result
from the simple model there are always fewer than n edges.
But, as shown in Fig. 4, as r increases the number of edges in
the optimal networks also increase, frequently quite suddenly.
1 A few of these cases failed to solve, which means that a few of the results
are averaged over fewer than 100 trials
6
A
r=0.1
1
0.9
B
r=1
25
C
40
35
0.8
20
0.7
30
0.6
15
0.5
25
20
0.4
10
15
0.3
10
0.2
5
5
0.1
0
0
0
Fig. 3. Illustrating the impact of adding reliability to the optimal network construction formulation. Panel A shows the full feasible network E, with colors
indicating production costs cv . Panel B shows the optimal network for w = 0.01 and r = 0.1, which is identical to the tree-like network that results from
the simple model.
Panel C shows the optimal network for w = 0.01 and r = 1, which shows the emergence of a meshed topology and substantial supply
w = 0.0001
redundancy.w Colors
= 0.001 in panels B and C indicate the amount of production capacity at each node.
Number of edges built
Mean component size, < ns >
w = 0.01
w = 0.1
w=1
10 3
A
B
C
D
r
r
r
r
r
10 2
= 0.0001
= 0.001
= 0.01
= 0.1
=1
10 1
10 0
10 3
w
w
w
w
w
= 0.0001
= 0.001
= 0.01
= 0.1
=1
r
r
r
r
r
= 0.0001
= 0.001
= 0.01
= 0.1
=1
10 2
10 1 -4
10
10 -3
10 -2
10 -1
Reliability cost, r
10 0 10 -4
10 -3
10 -2
10 -1
10 0
Interconnection cost, w
Fig. 4. Statistical results from the reliability model. Panels A and B show
mean component sizes as a function of the reliability cost and interconnection
cost parameters, r and w. Panels C and D show the average number of edges
in the optimal network, also as a function of r and w.
C. Senegal application
As a real-world case study, we applied the infrastructure
network design model to the geographic distribution of cities
and rural towns in Senegal. About half of the countries’
population still has no access to electricity, and the electrification rate in rural areas is as low as 28% [28]. In contrast
to Senegal’s electric power grid, the mobile communication
infrastructure is highly developed, with 1666 mobile phone
towers distributed across the country and a mobile phone
penetration rate of almost 100%. This allows for the use of data
from the mobile communication system as a robust prediction
for the geographic distribution of electricity needs, see [28]
for technical details. Data on the mobile phone infrastructure
has been made available by ORANGE / SONATEL within
the framework of the D4D Challenge [29]. Figure 5A depicts
both the existing electricity infrastructure and the location of
the mobile phone towers.
In order to use data from the communication system to
model demand for electricity, we first partitioned the country
into a rectangular grid with cell size 5km × 5km. Following the
approach in [28] we then used the number of cell phone towers
that are located in each grid cell as a proxy for the relative
electricity demand within that cell. Note that for the purpose
of our analysis we are not interested in estimating absolute
demand, but rather the relative amount of electricity that might
be consumed in a particular location. The center points of the
grid cells were used as locations xv , yv for the load nodes. As
with the uniformly distributed vertices, we randomly assigned
production costs to each node in the network, using the load
locations described above, using uniform random variables
over [0, 1].
Figure 5B shows the result of applying the basic optimization model (2) for the case of w = 0.01. Interestingly, without
reliability constraints, our model produces tree-like networks
that i) are similar in structure to what we found with randomly
distributed vertices in Sec III-A and ii) closely resemble the
tree-like topology of the existing electricity grid in Senegal.
Moreover, we find that the scaling of optimal component sizes
clearly follows the behavior previously (4) for synthetically
generated load points (see Fig. 5C). This suggests that the
power law decay in optimal network sizes is largely robust to
changes in the geographic distribution of load locations.
IV. C ONCLUSIONS
This paper presents results from a model of optimal infrastructure network design for electric power production and
delivery, with which we aimed to better understand the conditions under which electricity production and delivery were
best managed through a highly connected network with large
geographic reach, versus more localized networks with less
global connectivity. Several interesting observations emerged.
First we find, unsurprisingly, that as network construction costs
7
B
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
C
Mean component size, < ns >
A
103
10
2
10
1
0.2
0.1
0
100 -4
10
10-3
10-2
10-1
100
101
Interconnection cost, w
Fig. 5. Application of the optimal network design model to the geographic distribution of electricity demand centers in Senegal. Panel A shows the existing
high and medium voltage network together with the geographic distribution of the mobile phone towers (adopted from [28]). Panel B shows the optimal network
for w = 0.01, given random production costs and ignoring reliability. Panel C shows the mean component size versus interconnection cost, w.
increase the optimal size of infrastructure networks decreases
and the local provision of electrical services becomes preferable. In the case where cost is the primary network design
objective and reliability is not important, we unsurprisingly
find that optimal topologies always have tree structure. More
surprising, however, is the decrease in optimal network size
occurs gradually, over several orders of magnitude in our
network cost parameter, w. More specifically, we find that
optimal network sizes decrease with the power-law ∼ w−2/3 .
This same scaling property appears both in the random graphs
that we generated for simulation purposes and when we apply
our infrastructure design model to a spatial distribution of
demand centers taken from data from the country of Senegal.
This suggests that when cost is the most important design
criterion there is no single optimal size for infrastructure
networks, but rather that different sizes are likely to be optimal
for different locations. The distinction between which type of
network architecture (local or global) is “better” is not clear.
We do find that this gradual scaling becomes a more sudden
transition once reliability is added to the network design
objectives. When the failure to supply demand after vertex or
edge outages is deemed costly (large r), the optimal network
is a single interconnected system that spans the entire network
for a wide range of values for infrastructure costs w, and
then a small increase in w causes the optimal network size
to be small. Also, as the importance of reliability increases,
the optimal network topology transitions from being a tree, in
which there are no duplicate paths, to a meshed system with
substantial redundancy.
While the model that we used to reach these conclusions is
simple, the results have important implications that may yield
insight into difficult global problems such as the expansion
of infrastructure in less developed countries and the potential
transition from large power networks to smaller microgrids.
R EFERENCES
1
G. Zorpette, “Re-engineering Afghanistan,” IEEE Spectrum, vol. 48,
pp. 30–43, 30 September 2011.
[2] “Factoids: Drinking water and ground water statistics for 2007,” tech.
rep., US Environmental Protection Agency, March 2008.
[3]
E. Cotilla-Sanchez, P. D. H. Hines, C. Barrows, and S. Blumsack,
“Comparing the topological and electrical structure of the North American electric power infrastructure,” IEEE Systems Journal, vol. 6, no. 4,
2012.
[4]
R. Lasseter, “Smart distribution: Coupled microgrids,” Proceedings of
the IEEE, vol. 99, pp. 1074 – 1082, June 2011.
[5]
“Distributed generation energy technology capital costs,” tech. rep.,
National Energy Technology Laboratory, 2013.
[6]
H. Zerriffi, H. Dowlatabadi, and A. Farrell, “Incorporating stress in
electric power systems reliability models,” Energy Policy, vol. 35, no. 1,
pp. 61–75, 2007.
[7]
A. Zvoleff, A. S. Kocaman, W. T. Huh, and V. Modi, “The impact
of geography on energy infrastructure costs,” Energy Policy, vol. 37,
no. 10, pp. 4066–4078, 2009.
[8]
H. Zerriffi, Rural electrification: Strategies for distributed generation.
Springer Science & Business Media, 2010.
[9]
A. Fabrikant, A. Luthra, E. Maneva, C. Papadimitriou, and S. Shenker,
“On a network creation game,” in Proceedings of the Twenty-second
Annual Symposium on Principles of Distributed Computing, PODC ’03,
pp. 347–351, 2003.
[10]
Z. Wang, A. Scaglione, and R. Thomas, “Generating statistically correct
random topologies for testing smart grid communication and control
networks,” IEEE Trans. Smart Grid, vol. 1, pp. 28–39, June 2010.
[11]
K. Kanskey, Structure of transportation networks: relationships between
network geometry and regional characteristics. University of Chicago
Press, 1960.
[12]
P. Haggett and R. Chorley, Network analysis in geography. St. Martin’s
Press, 1969.
[13]
M. T. Gastner and M. E. Newman, “The spatial structure of networks,”
The European Physical Journal B - Condensed Matter and Complex
Systems, vol. 49, pp. 247–252, Jan 2006.
[14]
W. Garrison, “Connectivity in the interstate highway system,” Papers
in Regional Science, vol. 6, no. 1, pp. 121–137, 1960.
[15]
M. Kurant and P. Thiran, “Extraction and analysis of traffic and
topologies of transportation networks,” Phys. Rev. E, vol. 74, p. 036114,
Sep 2006.
[16]
S. Chan, R. Donner, and S. Lämmer, “Urban road networks – spatial
networks with universal geometric features?,” The European Physical
Journal B, vol. 84, no. 4, pp. 563–577, 2011.
[17]
S. Wilkinson, S. Dunn, and M. Shu, “The vulnerability of the European
air traffic network to spatial hazards,” Natural Hazards, vol. 60, no. 3,
p. 036114, 2012.
[18]
R. Lambiotte, V. Blondel, C. de Kerchove, E. Huens, C. Prieur,
Z. Smoreda, and P. V. Dooren, “Geographical dispersal of mobile
[1]
8
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
communication networks,” Physica A: Statistical Mechanics and its
Applications, vol. 387, pp. 5317 – 5325, 2008.
M. Barthelemy, “Spatial networks,” Physics Reports, vol. 499, no. 1–3,
pp. 1 – 101, 2011.
A. Bashan, Y. Berezin, S. V. Buldyrev, and S. Havlin, “The extreme
vulnerability of interdependent spatially embedded networks,” Nature
Physics, vol. 9, pp. 667–672, 10 2013.
M. Barthelemy, “Spatial networks,” Physics Reports, vol. 499, pp. 1–
101, Apr 2011.
T. C. McAndre, C. M. Danforth, and J. P. Bagrow, “Robustness of
spatial micronetworks,” Phys. Rev. E, vol. 91, p. 042813, Apr 2015.
C. Unsihuay-Vila, J. Marangon-Lima, A. de Souza, J. Perez-Arriaga,
and P. Balestrassi, “A model to long-yerm, multiarea, multistage, and
integrated expansion planning of electricity and natural gas systems,”
IEEE Trans. Power Systems, vol. 25, no. 2, pp. 1154–1168, 2010.
C. B. Sanchez, R. Bent, S. Backhaus, S. Blumsack, H. Hijazi, and
P. van Hentenryck, “Convex optimization for joint expansion planning
of natural gas and power systems,” in 49th Hawai’i Int’l. Conf. Sys.
Sci., pp. 1–10, Jan 2016.
E. Fisher and a. M. F. R. O’Neill, “Optimal transmission switching,”
IEEE Trans. Power Systems, vol. 23, pp. 1346–1355, Aug 2008.
K. Hedman, M. Ferris, R. O’Neill, E. Fisher, and S. Oren, “Cooptimization of generation unit commitment and transmission switching
with n-1 reliability,” IEEE Trans. Power Systems, vol. 25, pp. 1052–
1063, May 2010.
J. Dall and M. Christensen, “Random geometric graphs,” Physical
Review E, vol. 66, no. 1, p. 016121, 2002.
E. Martinez-Cesena, P. Mancarella, M. Ndiaye, and M. Schläpfer,
“Using mobile phone data for electricity infrastructure planning,” arXiv
preprint arXiv:1504.03899, 2015.
“Data For Development (D4D) Senegal Challenge.” http://www.d4d.
orange.com. Accessed: 2015-08-30.
ACKNOWLEDGEMENT
The authors gratefully acknowledge the hospitality of the
Santa Fe Institute in Santa Fe, NM, USA, where Hines and
Blumsack were sabbatical visitors in 2014-2015 and where
much of this work was completed. The authors would also
like to acknowledge Christa Brelsford, Luis Bettencourt, and
participants at the Santa Fe Institute workshop, “Reinventing
the Grid” for helpful suggestions and discussions.
AUTHOR BIOGRAPHIES
Paul D. H. Hines (S‘96,M‘07,SM‘14) received the Ph.D. in Engineering and
Public Policy from Carnegie Mellon University in 2007 and M.S. (2001)
and B.S. (1997) degrees in Electrical Engineering from the University of
Washington and Seattle Pacific University, respectively.
He is currently an Associate Professor in the School of Engineering, and
the L. Richard Fisher professor of electrical engineering, at the University
of Vermont. Formerly he worked at the U.S. National Energy Technology
Laboratory, the U.S. Federal Energy Regulatory Commission, Alstom ESCA,
and for Black and Veatch. He currently serves as the chair of the Green
Mountain Section of the IEEE, as the vice-chair of the IEEE PES Working
Group on Cascading Failure, and as an Associate Editor for the IEEE
Transactions on Smart Grid. He is a National Science Foundation CAREER
award winner.
Seth Blumsack (M ‘06) received the Ph.D. in Engineering and Public Policy
from Carnegie Mellon University in 2006, the M.S. degree in Economics from
Carnegie Mellon in 2003 and the B.A. degree in Mathematics and Economics
from Reed College in 1998. He is currently an Associate Professor in the John
and Willie Leone Family Department of Energy and Mineral Engineering at
The Pennsylvania State University, Chair of the Energy Business and Finance
program and is the John T. Ryan, Jr. Fellow in the College of Earth and
Mineral Sciences.
Markus Schläpfer received the Ph.D. in Mechanical and Process Engineering (2010) and the M.S. in Environmental Engineering (2003), both from ETH
Zurich. He is currently a Postdoctoral Fellow at the Santa Fe Institute and a
Research Affiliate at MIT’s Senseable City Lab.
| 3cs.SY
|
Submitted to the Annals of Statistics
arXiv:1803.07793v1 [math.ST] 21 Mar 2018
HIGH-DIMENSIONAL COVARIANCE MATRICES IN
ELLIPTICAL DISTRIBUTIONS WITH APPLICATION TO
SPHERICAL TEST
By Jiang Hu∗,¶ ,
Weiming Li†,k ,
‡,∗∗
Zhi Liu
and Wang Zhou§,††
Key Laboratory for Applied Statistics of MOE & Northeast Normal
University¶ , Shanghai University of Finance and Economicsk , University
of Macau & UM Zhuhai Research Institute∗∗ and National University of
Singapore††
This paper discusses fluctuations of linear spectral statistics of
high-dimensional sample covariance matrices when the underlying
population follows an elliptical distribution. Such population often
possesses high order correlations among their coordinates, which have
great impact on the asymptotic behaviors of linear spectral statistics.
Taking such kind of dependency into consideration, we establish a
new central limit theorem for the linear spectral statistics in this
paper for a class of elliptical populations. This general theoretical
result has wide applications and, as an example, it is then applied to
test the sphericity of elliptical populations.
1. Introduction. Large-scale statistical inference develops rapidly in
the last two decades. This type of inference often relies on spectral statistics
of certain random matrices in high-dimensional frameworks, where both the
dimension p of the observations and the sample size n tend to infinity. Recall
that a linear spectral statistic (LSS) (Bai and Silverstein, 2010) of a p × p
∗
Jiang Hu’s research is partially supported by NSFC (No. 11771073) and the Fundamental Research Funds for the Central Universities.
†
Weiming Li’s research is partially supported by NSFC (No. 11401037), MOE Project
of Humanities and Social Sciences, No. 17YJC790057, and Program of IRTSHUFE.
‡
Zhi Liu’s research is supported by FDCT127/2016/A3 of Macau and NSFC
(No.11401607).
§
Wang Zhou’s research is partially supported by the MOE Tier 2 grant MOE2015-T22-039 (R-155-000-171-112) at the National University of Singapore.
MSC 2010 subject classifications: Primary 62H15; secondary 62H10
Keywords and phrases: Covariance matrix, High-dimensional data, Elliptical distribution, Sphericity test.
1
2
J. HU ET AL.
Hermitian random matrix Rn is of the form
Z
p
1X
f (λi ) = f (x)dF Rn (x),
(1.1)
p
i=1
where λ1 , . . . , λp P
are the p eigenvalues of Rn , f is a function defined on R,
and F Rn = (1/p) pi=1 δλi is called the spectral distribution (SD) of Rn . Here
δa denotes the Dirac measure at the point a. In Ledoit and Wolf (2002) and
Schott (2007), most test statistics are actually LSSs of sample covariance
matrices. Bai et al. (2009) made systematic corrections to several classical
likelihood ratio tests to overcome the effect of high-dimension using LSSs of
sample covariance matrices and F-matrices. Later, Bai et al. (2015) derived
the CLT for LSSs of a high-dimensional Beta matrix, which can be broadly
used in multivariate statistical analysis, such as testing the equality of several covariance matrices, multivariate analysis of variance, and canonical
correlation analysis, see Anderson (2003). Most recently, based on an LSS of
regularized canonical correlation matrices, Yang and Pan (2015) proposed
a test for the independence between two large random vectors. Gao et al.
(2017) applied LSSs of sample correlation matrices to the complete independence test for p random variables and the equivalence test for p factor
loadings in a factor model. Clearly, it is of great interests to investigate the
behaviors of LSSs under various circumstances.
Specifically, let x1 , . . . , xn be n observations of x ∈ Rp , whose mean is
zero and covariance matrix is Σ. The sample covariance matrix is
n
Bn =
1X
xj x′j .
n
j=1
Our attention in this paper is focused on the asymptotic properties of LSSs
of Bn . The earliest study on this problem dates back to Jonsson (1982), who
obtained the central limit theorem (CLT) for LSSs of Bn by assuming the
population to be standard multivariate normal. A remarkable breakthrough
was done in Bai and Silverstein (2004), where the population is allowed to
be a linear transform of a vector of independent and identically distributed
(i.i.d.) random variables, i.e.,
(1.2)
x = Az.
Here A ∈ Rp×p is a non-random transformation matrix with rank(A) = p,
and z = (z1 , . . . , zp )′ with i.i.d. zi ’s satisfying
(1.3)
E(z1 ) = 0,
E(z12 ) = 1 and
E(z14 ) = 3.
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
3
The fourth moment condition was later extended by Pan and Zhou (2008)
to E(z14 ) < ∞. Though these assumptions are fairly weak, their requirement
of linearly dependent structure in (1.2) still excludes a lot of important
distributions. In particular, it excludes almost all distributions from the
elliptical family (Fang and Zhang, 1990).
Elliptical distributions were originally introduced by Kelker (1970) to generalize the multivariate normal distributions. A random vector x with zero
mean follows an elliptical distribution if and only if it has a stochastic representation (Fang and Zhang, 1990):
(1.4)
x = ξAu,
where the matrix A ∈ Rp×p is non-random with rank(A) = p, ξ ≥ 0 is a
scalar variable representing the radius of x, and u ∈ Rp is the random direction, which is independent of ξ and uniformly distributed on the unit sphere
S p−1 in Rp , denoted by u ∼ U (S p−1 ) in the sequel. This family of distributions has been widely applied in many areas, such as statistics, economics
and finance, which can describe fat (or light) tails and tail dependence among
components of a population, see Fang and Zhang (1990) and Gupta et al.
(2013). Evidently such distributions with high order correlations can not be
modeled by the linear transform model in (1.2).
A question raised immediately is that how the nonlinear dependency affects the asymptotic behaviors of LSSs in high-dimensional frameworks?
Indeed, Bai and Zhou (2008) proved that the SD F Bn of Bn converges to a
common generalized Marčenko-Pastur law almost surely if, for any sequence
of symmetric matrices {Cp } with bounded spectral norm,
(1.5)
V ar(x′ Cp x) = o(p2 )
as p, n → ∞. This condition is also sharp for the convergence, see Li and Yao
(2017) for an example. What is more, this condition holds for a list of
well known elliptical distributions, such as multivariate normal distributions, multivariate Pearson type II distributions, power exponential distributions, and a more general family of multivariate Kotz-type distributions
(Kotz, 1975), see Section 2 for more details. However, the limit of SD is
not enough for many procedures of statistical inference, such as confidence
interval and hypothesis testing. Therefore, in this paper, we will explore the
fluctuations of LSSs of Bn , when the population belongs to elliptical distributions that satisfy the condition (1.5). Compared with the pioneer work of
Bai and Silverstein (2004), the main difficulty of the current study lies in the
fact that both the radius ξ and direction u introduce nonlinear dependence
to the coordinates of the population x, which can not be handled through
4
J. HU ET AL.
the same way as they did for the linearly dependent structure. Technically,
we are facing the following three challenges. First, for paying the cost of
dropping linearly dependent structure, we have to add more moment condi√
tions on ξ, because the finite fourth moment of ξ/ p is no longer sufficient
for the nonlinear dependence case (see (2.5)). This is totally different from
the linearly dependent structure case. Second, we need to figure out how
the dependence of the entries of ξu influences the fluctuations of LSSs of Bn
(see Remark 2.3). Third, we have to extend many fundamental conclusions
in the independent case (Bai and Silverstein (2004)) to accommodate our
non-linearly dependent structure; see Lemma A.1-A.4 for example.
The structure of this paper is as follows. Firstly in Section 2, we set up a
new CLT for LSSs of Bn under elliptical distributions satisfying (1.5). Then
in Section 3, based on the derived results, we theoretically investigate the
problem of sphericity test for covariance matrices. This is done by discussing
a John’s-type test from Tian et al. (2015) for general alternative models
and a likelihood ratio test from Onatski et al. (2013) for spiked covariances
under arbitrary elliptical distributions. For illustration, the John-type test is
applied to analyze a dataset of weekly stock returns in Section 4. Technical
proofs of the main results are gathered in Section 5. Some supporting lemmas
are postponed to Appendix. The paper has also an on-line supplementary
file which includes the following materials: (i) CLT for general moment LSSs;
(ii) simulations regarding the John-type test; (iii) Proofs of some lemmas.
2. High-dimensional theory for eigenvalues of Bn . This section
investigates asymptotic behaviors of the eigenvalues of Bn , referred as sample
eigenvalues. We begin with proposing an equivalent condition of (1.5) under
the settings of the elliptical model in (1.4).
Lemma 2.1. Suppose that a p-dimensional random vector x has a stochastic form x = ξAu as defined in (1.4) with the radius ξ normalized as
E(ξ 2 ) = p. If the spectral norm of Σ = AA′ is uniformly bounded in p,
then the following two conditions are equivalent:
a)
V ar(x′ Cp x) = o(p2 ),
b)
E(ξ 4 ) = p2 + o(p2 ),
as p → ∞, where {Cp } is any sequence of symmetric matrices with bounded
spectral norm.
Remark 2.1. The fourth moment condition b) together with the normalization E(ξ 2 ) = p characterize the class of elliptical distributions discussed
in this paper. For the normal case, the squared radius ξ 2 ∼ χ2p and thus
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
5
E(ξ 2 ) = p and E(ξ 4 ) = p2 + 2p. In general, the typical order of E(ξ 4 ) is
p2 + τ p + o(p) with τ ≥ 0 a constant. Hence a specific elliptical distribution
can be recognized by evaluating the ratio
(2.1)
E(ξ 4 )/E 2 (ξ 2 ) = 1 + τ /p + o(p−1 ),
as p → ∞. We note that the parameter τ has a non-negligible effect on the
limiting distributions of LSSs of Bn , see Theorem 2.2. The proof of Lemma
2.1 is given in the supplementary material (Hu et al., 2017).
In the following we provide three examples of elliptical family satisfying
the condition (2.1). Some commonly seen elliptical distributions are also
checked and the results are summarized in Table 1.
Example 2.1. A p-dimensional centered multivariate Pearson type II
distribution has a density function
(2.2)
β −1
1
2
,
f (x) = cp |Σp |− 2 1 − x′ Σ−1
p x
where cp = π −p/2 Γ[(β + p)/2]/Γ(β/2) and β > 0. The stochastic representa1/2
tion of such a distribution is x = ξΣp u, where ξ 2 follows the distribution
Beta(p/2, β/2), see Fang and Zhang (1990). Therefore, we have
E(ξ 4 )/E 2 (ξ 2 ) = 1 + 2β/(p2 + βp + 2p),
which verifies the condition in (2.1) with τ = 0.
Example 2.2. The family of Kotz-type distributions introduced by Kotz
(1975) is an important class of elliptical distributions, which includes normal distributions, exponential power distributions, and double exponential
distribution as special cases. The density function of a centered Kotz-type
random variable x is
k−1
s
1
exp −β x′ Σ−1
,
(2.3)
f (x) = cp |Σp |− 2 x′ Σ−1
p x
p x
where cp = sβ α π −p/2 Γ(p/2)Γ(α) with α = (k−1+p/2)/s > 0 and (β, s) > 0.
1/2
s
Write x = ξΣp u. The 2s power of the radius is ξ 2s = [x′ Σ−1
p x] which has
the characteristic function
Z
Z
s
it[x′ Σ−1
itξ 2s
p x]
f (x)dx ∝ eitx xα−1 e−βx dx,
(2.4)
E(e
) = cp e
6
J. HU ET AL.
where the seconded integral is derived by polar coordinates transformation.
This characteristic function implies that ξ 2s follows the Gamma distribution
Gamma(α, β). Simple calculations reveal that
Γ(α + 2/s)Γ(α)
1
E(ξ 4 )
=
= 1 + 2 + o(α−1 ),
2
2
2
E (ξ )
Γ (α + 1/s)
s α
which verifies the condition in (2.1) with τ = 2/s. For the mentioned three
special cases, their details are presented in the 2-4th rows of Table 1.
P
Example 2.3. Let x = ξAu with ξ 2 = pj=1 yj2 independent of u, where
(yj ) is a sequence of i.i.d. random variables with
E(y12 ) = 1
and
E(y14 ) = µ4 < ∞.
Then it is simple to check that E(ξ 2 ) = p and E(ξ 4 ) = p2 + (µ4 − 1)p which
verifies the condition in (2.1) with τ = µ4 − 1.
We should note that the condition (2.1) excludes some elliptical distributions, such as multivariate student-t distributions and normal scale mixtures,
as shown in the 5-6th rows of Table 1. Indeed, sample eigenvalues from these
distributions do not obey the generalized Marčenko-Pastur law (El Karoui,
2009; Li and Yao, 2017), which are then out of the scope of this paper.
Table 1
Some elliptical distributions and their verification of the condition (2.1). The notation
“⊥
⊥” in the last row denotes independence.
x = ξAu ∈ Rp
Normal
Double exponential
Exponential power
Student-t
Distribution of ξ
ξ 2 ∼ χ2p
ξ ∼ Gamma(p, 1)
p 1
ξ 2s ∼ Gamma 2s
,2
ξ 2 /p ∼ F (p, v), v > 4
E(ξ 4 )/E 2 (ξ 2 )
1 + p2
4p+6
1 + p(p+1)
2
1 + sp + o(p−1 )
2
+ 2(v−2)
1 + v−4
p(v−4)
Normal scale mixture ξ 2 = w · v, w⊥
⊥v, v ∼ χ2p 1 +
V ar(w)
E 2 (w)
+
2
2 E(w )
p E 2 (w)
Condition (2.1)
Holds (τ = 2).
Holds (τ = 4).
Holds (τ = 2s ).
Not hold.
Not hold.
Now we are ready to investigate the asymptotic properties of sample eigenvalues in high-dimensional frameworks, under the following assumptions.
Assumption (a). Both the sample size n and dimension p tend to infinity
in such a way that cn := p/n → c ∈ (0, ∞).
Assumption (b). There are two independent arrays of i.i.d. random variables (uj )j≥1 , u1 ∼ U (S p−1 ), and (ξj )j≥1 satisfying for some τ ≥ 0 and
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
7
ε > 0,
(2.5)
E(ξ12 ) = p,
E(ξ14 ) = p2 + τ p + o(p) and E
ξ12 − p
√
p
2+ε
< ∞,
such that for each p and n the observation vectors can be represented as
xj = ξj Auj , where A is a p × p matrix.
Assumption (c). The spectral distribution Hp of the matrix Σ := AA′
weakly converges to a probability distribution H, as p → ∞, referred as
Population Spectral Distribution (PSD). Moreover, the spectral norm of the
sequence (Σ) is uniformly bounded in p.
In the sequel, for any function G of bounded variation on the real line,
its Stieltjes transform is defined by
Z
1
(2.6)
dG(λ), z ∈ C \ SG ,
m(z) =
λ−z
where SG stands for the support of G. Then we have the following theorems.
Theorem 2.1. Suppose that Assumptions (a)-(c) hold. Then, almost
surely, the empirical spectral distribution F Bn converges weakly to a probability distribution F c,H , whose Stieltjes transform m = m(z) is the only
solution to the equation
Z
1
(2.7)
dH(t) , z ∈ C+ ,
m=
t(1 − c − czm) − z
in the set {m ∈ C : −(1 − c)/z + cm ∈ C+ } where C+ ≡ {z ∈ C : ℑ(z) > 0}.
Remark 2.2. Theorem 2.1 follows from Lemma 2.1 and Theorem 1.1
in Bai and Zhou (2008), and thus we omit its proof here. Let m = m(z) be
the Stieltjes transform of F c,H = cF c,H + (1 − c)δ0 . Then Equation (2.7) can
be re-expressed as
Z
t
1
(2.8)
z = − +c
dH(t) , z ∈ C+ ,
m
1 + tm
which is the so-called Silverstein equation (Silverstein, 1995).
Let F cn ,Hp be the distribution defined by (2.7) with the parameters (c, H)
replaced by (cn , Hp ) and denote Gn = F Bn − F cn ,Hp . We next study the
fluctuation of centralized LSSs with form
Z
Z
f (x)dGn (x) = f (x)d[F Bn (x) − F cn ,Hp (x)],
where f is a function on the real line.
8
J. HU ET AL.
Theorem 2.2. Suppose that Assumptions (a)-(c) hold. Let f1 , . . . , fk be
functions analytic on an open interval containing
√ 2
√ 2
Σ
Σ
(2.9)
lim inf λmin δ(0,1) (c)(1 − c) , lim sup λmax (1 + c) .
p→∞
p→∞
Then the random vector
Z
Z
f1 (x)dGn (x), . . . , fk (x)dGn (x)
p
converges weakly to a Gaussian vector (Xf1 , . . . , Xfk ), with mean function
Z
I
1
c(m′ (z)t)2
EXf = −
dH(t)dz
f (z)
2πi C1
m(z)(1 + tm(z))3
Z
I
τ −2
(zm(z) + 1)m′ (z)t
−
dH(t)dz
f (z)
2πi C1
(1 + tm(z))2
and covariance function
I
I
f (z1 )g(z2 )m′ (z1 )m′ (z2 )
dz1 dz2
(m(z1 ) − m(z2 ))2
C1 C2
Z
Z
+ c(τ − 2) xf ′ (x)dF (x) xg ′ (x)dF (x),
1
Cov (Xf , Xg ) = − 2
2π
(f, g ∈ {f1 , · · · , fk }), where the contours C1 and C2 are non-overlapping,
closed, counter-clockwise orientated in the complex plane, and each enclosing
the support of the limiting spectral distribution F c,H .
Remark 2.3. When the population is normal, or rather τ = 2, this theorem coincides with the main result in Bai and Silverstein (2004). It implies
that the high order correlation among the components of the population affects both the limiting mean vectors and the covariance matrices of LSSs by
additive quantities proportional to τ − 2. This factor can be further decomposed into two parts, τ and −2, which correspond respectively to the effect
from the radius ξ and that from the direction u (considering the case ξ 2 ≡ p).
It’s interesting to see that these two kinds of dependency have opposite effects
and they may cancel each other for normal population.
R
As an application of Theorem 2.2, we consider β̂nj = xj dF Bn (x), j =
1, 2, the first two moments of sample eigenvalues. Theorem 2.2 implies
β̂ − βn1 D
v1
ψ11 ψ12
p n1
−
→N
,
(2.10)
,
v2
ψ12 ψ22
β̂n2 − βn2
9
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
where the parameters possess explicit expressions as
2
βn2 = γn2 + cn γn1
,
βn1 = γn1 ,
ψ11 = 2cγ2 + c(τ −
2)γ12 ,
v1 = 0,
2
v2 = cγ2 + c(τ − 2)γ1 ,
ψ12 = 4cγ3 + 4c γ1 γ2 + 2c(τ − 2)γ1 (cγ12 + γ2 ),
ψ22 = 8cγ4 + 4c2 γ22 + 16c2 γ1 γ3 + 8c3 γ12 γ2 + 4c(τ − 2)(cγ12 + γ2 )2 ,
R
R
where γnj = tj dHp (t) and γj = tj dH(t) for j ≥ 1. For LSSs of higher
order moments, explicit formulas of their limiting means and covairances are
discussed in the supplementary material (Hu et al., 2017).
0
2
4
−2
0
2
4
4
2
−4
−4
−4
(p,n)=(400,200)
0
Sample Quantiles
0
Sample Quantiles
2
4
(p,n)=(400,200)
−2
0
Sample Quantiles
−4
−2
0
−4
−2
0
2
4
−4
−2
0
2
4
Theoretical Quantiles
Theoretical Quantiles
Theoretical Quantiles
^
QQ Plot for normalized βn1
^
QQ Plot for normalized βn2
^
QQ Plot for normalized βn1
^
QQ Plot for normalized βn2
0
2
4
−4
−2
0
2
Theoretical Quantiles
4
4
0
Sample Quantiles
2
(p,n)=(400,200)
−4
−2
0
Sample Quantiles
−4
−4
−2
Theoretical Quantiles
−2
0
Sample Quantiles
0
−2
−4
−4
(p,n)=(400,200)
2
4
(p,n)=(200,400)
2
4
(p,n)=(200,400)
−2
Sample Quantiles
−2
−4
−2
^
QQ Plot for normalized βn2
Theoretical Quantiles
2
4
−4
Sample Quantiles
(p,n)=(200,400)
2
4
(p,n)=(200,400)
^
QQ Plot for normalized βn1
−2
^
QQ Plot for normalized βn2
2
4
^
QQ Plot for normalized βn1
−4
−2
0
2
Theoretical Quantiles
4
−4
−2
0
2
4
Theoretical Quantiles
Fig 1. Normal QQ-plots for normalized β̂n1 and√β̂n2 from 10,000 independent replications.
Upper panels: ξ ∼ k1 Gamma(p, 1) with k1 = 1/ p + 1. Lower panels: ξ 2 ∼ k2 Beta(p/2, 2)
with k2 = p + 4. The dimensional settings are (p, n, c) = (200, 400, 0.5), (400, 200, 2).
We conduct a small simulation experiment to examine the fluctuations of
β̂n1 and β̂n2 . In the experiment, the PSD Hp is fixed at Hp = 0.5δ1 + 0.5δ2 .
The distribution of ξ is selected as (1) ξ ∼ k1 Gamma(p, 1) with k1 =
√
1/ p + 1 and (2) ξ 2 ∼ k2 Beta(p/2, 2) with k2 = p + 4, which correspond
the CLT with τ = 4 and τ = 0, respectively. The factors k1 and k2 are
10
J. HU ET AL.
selected to satisfy E(ξ 2 ) = p. The dimensional settings are (p, n, c) =
(200, 400, 0.5), (400, 200, 2) and the number of independent replications
√ is
ψ11
10, 000. Normal QQ-plots
for
normalized
statistics,
i.e.
p(
β̂
−
β
)/
n1
n1
√
and [p(β̂n1 − βn1 ) − v2 ]/ ψ22 , are displayed in Figure 1. Their asymptotic
standard normality is well confirmed in all studied cases.
3. Testing for high-dimensional spherical distributions.
3.1. John’s test and its extension. In this section, we revisit the sphericity test for covariance matrices in high-dimensional frameworks. For this
particular test probelm, the underlying population can follow arbitrary elliptical distribution, which may violate the condition in (1.5).
The sphericity test on the covariance matrix Σ is
H 0 : Σ = σ 2 Ip
(3.1)
v.s. H1 : Σ 6= σ 2 Ip ,
where σ 2 is an unknown scalar parameter. When the dimension p is fixed, for
normal populations, John (1972) proposed a locally most powerful invariant
test statistic to deal with the sphericity hypothesis based on the spectrum of
sample covariance matrices. Due to its concise form and broad applicability,
this kind of test is quite favorable for high dimensional situations and has
been extensively studied in recent years. See, for example, Ledoit and Wolf
(2002), Wang and Yao (2013), Tian et al. (2015) for the linear transform
model in (1.2), while Zou et al. (2014) and Paindaveine and Verdebout (2016)
for the elliptical model in (1.4). In particular, the test statistic in Tian et al.
(2015) synthesizes the first four moments of sample eigenvalues, by which
it gains extra powers for spike-like alternative covariance matrices. However this statistic is not valid for general elliptical populations (Li and Yao,
2017). Hence, we next develop an analogue test procedure with the help of
the theoretical results in Section 2, and then compare it numerically with
that from Paindaveine and Verdebout (2016).
Since the hypotheses in (3.1) are only concerned with the shape component of Σ, by convention, we transform the original samples into the so-called
spatial-sign samples, that is,
x̌j :=
√
pxj /||xj || =
√
pAuj /||Auj ||, j = 1, . . . , n.
Therefore, testing the sphericity of Σ can now be converted to testing the
identity of Σ̌ = E(x̌1 x̌′1 ). P
This inference can be realized by constructing
spectral statistics of B̌n = nj=1 x̌j x̌′j /n. Specifically, let
αnj = p−1 tr(Σ̌j ) and
β̌nj = p−1 tr(B̌nj ),
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
11
j = 0, 1, 2, . . . . By verifying the condition in (1.5) for x̌1 , one may conclude
that Theorem 2.1 also holds for (Σ̌, B̌n ) with all conditions on ξ removed.
Then, similar to Tian et al. (2015), from the fact that β̌n1 ≡ 1, one may
obtain estimators of αn2 and αn4 as
α̌n2 = β̌n2 − cn ,
α̌n4 = β̌n4 − 4cn β̌n3 − 2cn (β̌n2 )2 + 10c2n β̌n2 − 5c3n ,
respectively, and two simple statistics for the sphericity test as
T1 = α̌n2 − 1
and T2 = α̌n4 − 1.
Moreover, their joint null distribution is directly from (2.10) with τ = 0.
Theorem 3.1. Suppose that Assumptions (a)–(c) [removing the moment
conditions in (2.5)] hold. Under the null hypothesis,
D
n (T1 , T2 ) −
→ N2 (µ, Ω),
where µ = (−1, −6 + c) and the covariance matrix Ω = (ωij ) with ω11 =
4, ω12 = 24, and ω22 = 8(18 + 12c + c2 ).
The two statistics T1 and T2 , together with their null distributions, provide two test procedures for the identity of Σ̌ (thus the sphericity of Σ).
The test statistic T1 agrees with that in Paindaveine and Verdebout (2016),
where its null asymptotic distribution is proved to be universal whenever
min{n, p} → ∞. For the case where the population mean is unknown, see
Zou et al. (2014). The test statistic T2 is new. Compared with T1 , it is more
sensitive to extreme eigenvalues of Σ̌ and thus can serve as a complement
of T1 . Parallel to Tian et al. (2015), a joint statistic of T1 and T2 can be
constructed as
(
)
nT1 + 1
nT2 + 6 − cn
Tm = max
,
,p
2
8(18 + 12cn + c2n )
where the two original statistics are both standardized according to their
asymptotic null distributions.
Theorem 3.2. Suppose that Assumptions (a)–(c) [removing the moment
conditions in (2.5)] hold and let δp = ptr(Σ2 )/tr2 (Σ) − 1.
(i) Under the null hypothesis, for any x ∈ R,
2
Z x Z x
u − 2ρuv + v 2
1
p
exp −
P (Tm ≤ x) →
dudv,
2(1 − ρ2 )
−∞ −∞ 2π 1 − ρ2
p
where ρ = 6/ 2(18 + 12c + c2 ).
12
J. HU ET AL.
(ii) Under the alternative hypothesis, if nδp → ∞ then the power of the
test Tm goes to 1 as (n, p) → ∞.
The asymptotic null distribution of Tm is an immediate consequence of
Theorem 3.1. The consistency of Tm can be proved by showing either the
consistency of T1 or T2 . As the consistency of T1 has been given in Zou et al.
(2014), we omit its proof.
We have run a simulation experiment for the tests T1 , T2 , and Tm to check
their finite-sample properties under similar model settings as in Tian et al.
(2015). The results show that all the three tests have satisfactory empirical sizes and powers. In addition, compared with T1 and T2 , the test Tm
exhibits its robustness against different types of alternative models, see the
supplementary material Hu et al. (2017).
3.2. Sphericity test under spiked alternative model. The sphericity test
Tm applies to general alternative models. However, its consistency requires
nδp → ∞ which excludes the well-known spiked covariance model (Johnstone,
2001). For the simplest spiked model, the covariance matrix can be expressed
as Σ = σ 2 (Ip + hvv ′ ) where σ 2 and Ip are as before, h is a constant, and
v is a unit vector in Rp . Both h and v are unknown parameters. Thus the
sphericity hypotheses in (3.1) reduce to
(3.2)
H0 : h = 0 v.s. H1 : h > 0.
It’s obvious that Tm will asymptotically fail to reject such alternatives
since nδp → 0. What’s more, this testing problem will become more dif√
ficult but attractive when the signal h falls below the threshold c, see
Berthet and Rigollet (2013); Onatski et al. (2013, 2014); Donoho and Jin
(2015), and references therein. Hence, applying the CLT for LSSs under
elliptical distributions, we discuss a test procedure for (3.2) proposed by
Onatski et al. (2013), which was built under normal populations.
In Onatski et al. (2013), the authors discussed a likelihood ratio test based
on the joint distribution of sample eigenvalues from normal populations. This
√
test was especially designed for the local alternative H1 : h ∈ (0, c) and
the employed statistic was approximated by a special LSS. In our settings,
this LSS can be formulated as
Z
Z
B̌n
(3.3)
TLR (s) = ln(z(s) − x)dF (z) − ln(z(s) − x)dF cn ,δ1 (x),
where s ∈ (0, s̄) is a testing parameter and z(s) = (1 + s)(cn + s)/s. The
√
√
upper bound s̄ of s is chosen as s̄ = cn for h ∈ [0, cn ] and s̄ = h−1 cn for
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
13
√
h ∈ ( cn , ∞) such that z(s) is larger than the limit of λmax (B̌n ), the largest
sample eigenvalues. Applying Theorem 2.2, one may get the asymptotic
distribution of TLR (s) under general elliptical distributions.
Theorem 3.3. Suppose that Assumptions (a)–(c) [removing the moment
conditions in (2.5)] hold. Under the null hypothesis, for any fixed s ∈ (0, s̄),
D
pTLR (s) −
→ N (µs , σs2 ),
(3.4)
where the respective mean and variance functions are
µs =
1
ln(1 − c−1 s2 ) + c−1 s2 and σs2 = −2 ln(1 − c−1 s2 ) − 2c−1 s2 .
2
The proof of Theorem 3.3 is given in the supplementary material (Hu et al.,
2017). Given a value of s ∈ (0, s̄) and a significance level α ∈ (0, 1), the
test TLW (s) rejects H0 if pTLW (s) < σs Φ−1 (α) + µs , where Φ(x) denotes
the standard normal distribution function. Unlike Onatski et al. (2013), the
theoretical power of this test is not available at present since pTLR (s) is not
a likelihood ratio statistic in elliptical distributions. Another reason is that
Theorem 2.2 is inapplicable to this situation since the spatial-sign sample is
not anymore elliptically distributed under H1 .
Let’s take a step back and only consider the testing problem in elliptical
distributions satisfying (2.1). For simplicity, we assume σ is known and set
σ = 1, so that the test TLR (s) is still valid by simply substituting the sample
covairance Bn into B̌n , i.e.,
Z
Z
Bn
e
TLR (s) = ln(z(s) − x)dF (z) − ln(z(s) − x)dF cn ,δ1 (x),
whose asymptotic distribution under both the null and alternative hypotheses is described in the following theorem.
Theorem 3.4. Suppose that Assumptions (a)–(c) hold. Let h0 be the
true value of h and σ = 1, then for any fixed s ∈ (0, s̄),
D
pTeLR (s) −
→ N (µ̃s (h0 ), σ̃s2 ),
(3.5)
where the respective mean and variance functions are
1
ln(1 − c−1 s2 ) + (1 − τ /2)c−1 s2 + ln(1 − c−1 sh),
2
σ̃s2 = −2 ln(1 − c−1 s2 ) − (2 − τ )c−1 s2 .
µ̃s (h) =
14
J. HU ET AL.
This theorem is a direct conclusion of Theorem 2.2. It proof is similar to
that of Theorem 3.3 and we thus omit it here. From Theorem 3.4, the power
function of TeLW (s) is
µ̃s (h0 ) − µ̃s (0)
−1
e
PH1 (TLW (s) reject H0 ) = Φ Φ (α) −
, h0 > 0.
σ̃s
For normal populations (τ = 2), this power function reaches its maximum
at s = h0 , which agrees with (5.1) in Proposition 9 of Onatski et al. (2013).
In general, the maximizer may not locate at h0 . An interesting case is τ = 0,
for which the power function tends to 1 as s → 0+ . This is from the fact
that
µ̃s (h0 ) − µ̃s (0) = −c−1 sh0 + o(s) and σs2 = 2c−2 s4 + o(s4 ).
At this time, TeLW (s) will successfully detect any positive h0 as long as s is
close to zero.
4. An empirical study. For illustration, we apply the test procedure
based on Tm to analyze weekly returns of the stocks from S&P 500. The
tests TLR and T̃LR are not included in this analysis since there is a lack of
evidence to fit the data using the simplest spiked model. According to The
North American Industry Classification System (NAICS), which is used by
business and government to classify business establishments, the 500 stocks
can be divided into 20 sectors. Nine of them are removed from our analysis
since their numbers of stocks are all less than 10. The remaining 11 sectors
as well as their numbers of stocks are listed in Table 2. Usually the stocks
Table 2
Number of stocks in each NAICS Sectors.
Sector
Number of stocks
1
30
2
32
3
189
4
17
5
36
6
14
7
37
8
65
9
14
10
17
11
11
in the same sector are correlated, and the stocks in different sectors are
uncorrelated. So it is expected that the weekly returns of stocks in the
same sector are not spherically distributed, and it is interesting to see if
the weekly returns of stocks in different sectors are spherically distributed.
In the following, we apply Tm to stocks in the same sector and stocks in
different sectors respectively.
The original data are the closing prices or the bid/ask average of these
stocks for the trading days in the first half of 2013, i.e., from 1 January 2013
to 30 June 2013, with total 124 trading days. This dataset is downloadable
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
15
from the Center for Research in Security Prices Daily Stock in Wharton
Research Data Services. The testing model is established as follows. Denote
pl as the number of stocks in the lth sector, uij (l) as the price of the ith
stock in the lth sector on Wednesday of the jth week. The reason that we
choose Wednesday’s price here is to avoid the weekend effect in stock market.
Thus we get 22 returns for each stock. In order to meet the condition of the
proposed test, the original data uij (l) should be transformed by logarithmic
difference, which is a very commonly used procedure in finance. There are a
number of theoretical and practical advantages of using logarithmic returns.
One of them is that the sequence of logarithmic returns are independent
of each other for big time scales (e.g. ≥ 1 day, see Rama (2001)). Denote
xij (l) = ln(ui(j+1) (l)/uij (l)), j = 1, . . . , 21 and X(l) = (xij (l))pl ×n , where
n = 21 is the sample size.
Now applying Tm to the dataset X(l), l = 1, . . . , 11, respectively, we
obtain 11 p-values, which are all below 10−9 . Therefore, we have strong
evidence to believe that stocks in the same sector are not spherically distributed. This is consistent with our intuition. Next, we consider stocks in
different sectors. Specifically, we choose one stock from each sector to make
up a group of 11 cross-sectoral stocks and then test whether these stocks are
spherically distributed. Because there are about 9.79 × 1015 different groups,
we just randomly draw 1,000,000 groups from them to analyze. It turns out
that the largest p-value is 0.3889, 231 p-values are bigger than 0.05, and 69
p-values are bigger than 0.1. These results again demonstrate that, when
the number of stocks is not very small, it is hard to say weekly logarithmic
returns for the stocks are spherically distributed. It is also very interesting to
analyze these spherically distributed stocks in different sectors, which have
almost the same variances.
5. Proof of Theorem 2.2. The proof of Theorem 2.2 relies on analyzing the resolvent of the sample covariance matrix Bn and the general strategy
follows the approach in Bai and Silverstein (2004). Also see Bai et al. (2015)
and Gao et al. (2017) for recent developments. However, as we are dealing
with the new model equipped with nonlinear dependency, all technical steps
of implementing this strategy have to be updated, or at least revalidated.
They are presented in this section.
5.1. Sketch of the proof of Theorem 2.2. Let v0 > 0 be arbitrary, xr
any number greater than the right end point of interval (2.9), and xl any
negative number if the left end point of (2.9) is zero, otherwise choose xl ∈
√ 2
(0, lim inf p→∞ λΣ
min (1 − c) ). Let Cu = {x ± iv0 : x ∈ [xl , xr ]} and define a
16
J. HU ET AL.
contour C
(5.1)
C = {x + iv : x ∈ {xr , xl }, v ∈ [−v0 , v0 ]} ∪ Cu .
By definition, this contour encloses a rectangular region in the complex plane
containing the support of the LSD F c,H . Denote by mn (z) and mF cn ,Hp (z)
the Stieltjes transforms of the ESD F Bn and the LSD F cn ,Hp , respectively.
Their companion Stieltjes transforms are given by
mn (z) = −
1 − cn
+ cn mn (z)
z
and mF cn ,Hp (z) = −
1 − cn
+ cn mF cn ,Hp (z).
z
With these notation, we define an empirical process on C as
Mn (z) = p mn (z) − mF cn,Hp (z) = n mn (z) − mF cn ,Hp (z) ,
z ∈ C.
Since fℓ , ℓ = 1, . . . , k, in Theorem 2.2 are analytic on an open region containing the interval (2.9) (thus analytic on the region enclosed by C), by
Cauchy’s integral formula, we have for any k complex numbers a1 , . . . , ak ,
k
X
paℓ
ℓ=1
Z
I
k
X
aℓ
fℓ (x)dGn (x) = −
fℓ (z)Mn (z)dz,
2πi C
ℓ=1
when all sample eigenvalues fall in the interval (xl , xr ), which is correct with
overwhelming probability. In order to remove the small probability event
that some sample eigenvalues fall outside the interval, we need a truncated
cn (z). Specifically, let {εn } be a sequence
version of Mn (z), denoted by M
decreasing to zero satisfying εn > n−a for some a ∈ (3/4, 1). The truncated
cn (z) for z = x + iv ∈ C is given by
process M
z ∈ Cn ,
Mn (z)
−1
c
(5.2) Mn (z) = Mn (x + in εn ) x ∈ {xl , xr }, v ∈ [0, n−1 εn ],
Mn (x − in−1 εn ) x ∈ {xl , xr }, v ∈ [−n−1 εn , 0],
where
Cn = Cu ∪ {x ± iv : x ∈ {xl , xr }, v ∈ [n−1 εn , v0 ]},
cn (z) agrees with Mn (z), is a regularized set of C excluding a
on which M
small segment near the real line. Then we have
Lemma 5.1.
any ℓ > 0,
Under the same assumptions in Theorem 2.2, we have for
I
fℓ Mn (z)dz =
C
Z
Cn
cn (z)dz + op (1).
fℓ (z)M
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
17
The proof of this lemma will be put in the supplementary material,
Hu et al. (2017). Hence, Theorem 2.2 follows by similar arguments on Pages
562-563 in Bai and Silverstein (2004) and the following Lemma.
cn (·) conLemma 5.2. Under Assumptions (a)-(c), the random process M
verges weakly to a two-dimensional Gaussian process M (·) with the mean
function
Z
Z
c(m′ (z)t)2 dH(t)
(zm(z) + 1)m′ (z)tdH(t)
(5.3) EM (z) =
+
(τ
−
2)
m(z)(1 + tm(z))3
(1 + tm(z))2
and covariance function
Cov (M (z1 ), M (z2 )) =
(5.4)
where z, z1 , z2 ∈ C.
2
2m′ (z1 )m′ (z2 )
−
2
(m(z1 ) − m(z2 ))
(z1 − z2 )2
τ −2
+
m(z1 ) + z1 m′ (z1 ) m(z2 ) + z2 m′ (z2 ) ,
c
The proof of this lemma is the main task of this section and can be
achieved by four steps as described below. Notice that in the proof we will use
several inequalities frequently, which are presented as lemmas in Appendix.
We will show how and where to use these lemmas in the following. Write for
z ∈ Cn ,
cn (z) = p[mn (z) − Emn (z)] + p[Emn (z) − m cn,Hp (z)]
M
F
:= Mn(1) (z) + Mn(2) (z).
Step 1: Truncation and
of ξ. This step regularizes the variPn rescaling
2 Au u′ A′ /n such that {ξ } have proper
ables {ξj } in Bn =
ξ
j
j
j
j=1 j
bound for finite (n, p) while maintaining the limiting distribution of
the LSSs. Compared with the proof in Bai and Silverstein (2004), this
result is entirely new since their model doesn’t include a radius variable at all. Moreover, our key inequalities (Lemmas A.2-A.4) are all
built on this regularization and thus their proofs have to be updated
to accommodate the elliptical model.
(1)
Step 2: Finite dimensional convergence of Mn (z) in distribution.
This step finds the joint limiting distribution of an r-dimensional vec
(1)
tor Mn (zℓ ) 1≤ℓ≤r by the martingale CLT. Lemmas A.2 and A.3 are
used to simplify the expression of the martingale difference and verify
Lindeberg’s condition, respectively. The (limiting) covariance function
18
J. HU ET AL.
is calculated based on Lemma A.1 with the help of Lemma A.3. A new
finding here is that the nonlinear dependency comes up with an extra
term in the covariance function (Lemma A.1), which results in a novel
procedure of proving the convergence of this term.
(1)
Step 3: Tightness of Mn (z) on Cn . This step presents the basic idea
for establishing the tightness. A key element is the uniform boundedness of E||(Bn − zI)−q || for q > 0 which is derived by Lemma A.4.
By virtue of this and Lemmas A.2-A.4, the tightness can be verified
following similar arguments in Bai and Silverstein (2004).
(2)
Step 4: Convergence of Mn (z). This final step mainly calculates
(2)
the limit of Mn (z), the limiting mean function of the LSSs. Akin to
deriving the covariance function in Step 2, the nonlinear effect again
brings an additional quantity to the mean function. Hence, most work
in this part is to handle this new quantity. Note that Lemma A.4 is
useful in this step to obtain several convergence results and uniform
boundedness on Cn .
These detailed four steps are presented in the next subsection. We note that
(1)
(2)
when simplifying Mn (z) and Mn (z), one more straightforward method is
used (see (5.11) and (5.32) respectively), which are different from Bai and Silverstein
(2004).
5.2. Truncation and rescaling of the ξ-variable. From the moment con√
dition E|(ξ12 − p)/ p|2+ε < ∞ for some ε > 0 in Assumption (b), we can
choose a sequence of δn > 0 such that
δn → 0, δn n1/2 → ∞, δn−2 p−1 E[(ξ12 − p)2 I{|ξ12 −p|≥δnp} ] → 0.
Pn
′
ˆ
Let B̂n =
j=1 x̂j x̂j /n where x̂j = ξ̂j Auj with ξj = ξj I{|ξj2 −p|<δn p} . We
then have
(5.5)
(5.6)
P (B̂n 6= Bn ) ≤ nP (|ξ12 − p| ≥ δn p)
≤ δn−2 np−2 E[(ξ12 − p)2 I{|ξ12 −p|≥δn p} ] → 0.
P
Define B̃n = nj=1 x̃j x̃′j /n where x̃j = ξ̃j Auj with ξ˜j = ξ̂j /σn and σn2 =
E(ξ̂12 )/p. By the assumptions in (5.5),
|1 − σn2 | = E(ξ12 /p − 1)I{|ξ12 −p|≥δn p} + EI{|ξ12 −p|≥δn p}
(5.7)
≤ δn−1 (1 + δn−1 )p−2 E(ξ12 − p)2 I{|ξ12 −p|≥δn p} = o(p−1 ).
Therefore we have
E(ξ̃12 ) = p
and E(ξ̃14 ) =
1
4
4
= p2 + τ p + o(p).
E(ξ
)
−
Eξ
I
2 −p|≥δ p}
1
1
{|ξ
n
1
σn4
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
19
On the other hand, write uj = zj /||zj || where, and in the following zj ∼
N (0, Ip ) and k · k denotes the spectral norm for a matrix, or L2 norm for a
vector. By the strong law of large numbers, for any fixed 0 < η < 1, we have
max{||zj ||2 /p : j = 1, . . . , n} ≥ 1 − η holds almost surely for large p. Hence
we have for large p,
||B̃n || =
n
n
(1 + δn )
1X
1 X ξ̃j2 /p
′ ′
Azj zj A ≤
zj z′j
||Σ||
n
||zj ||2 /p
(1 − η)σn2
n
j=1
j=1
n
almost surely. Thus, from Yin et al. (1988), we know that lim supn λB̃
max (and
√
B̂
n
similarly lim supn λmax ) are almost surely bounded by lim supp ||Σ||(1+ c)2 .
Let G̃n (x) and Ĝn (x) be the analogues of Gn (x) with the matrix Bn replaced by B̃n and B̂n , respectively. From the arguments in Bai and Silverstein
(2004) and (5.7), we can get for f (x, z) = 1/(x − z) (z ∈ C), almost surely,
p2
Z
f (x)dĜn (x) −
≤ 4K 2
p q
X
j=1
Z
n
λB̂
j −
2
f (x)dG̃n (x)
q
≤
p
X
j=1
2
B̃n
n
K|λB̂
j − λj |
2 X
p
B̂n
B̃n
n
λB̃
λ
+
λ
j
j
j
j=1
n
1X
B̃k
2
B̂n
(x̂j − x̃j )′ (x̂j − x̃j )
≤ 4K p(λmax + λmax )
n
j=1
B̃k
n
= 4K 2 p(λB̂
max + λmax )
(1 − σn2 )2
tr(B̂n ) → 0,
σn2 (1 + σn2 )
where K is an upper bound of |fx′ (x, z)|. As a consequence of this and (5.6),
Z
Z
Mn (z) = p f (x)dGn (x) = p f (x)dG̃n (x) + op (1).
R
Therefore, we only need to find the limiting distribution of f (x)dG̃n (x).
For simplicity, we still use Bn , xj , ξj instead of B̃n , x̃j , ξ˜j , respectively, and
assume that
(5.8)
∀j, |ξj2 − p| < δn p, E ξ12 = p, E ξ14 = p2 + τ p + o(p),
in the sequel.
20
J. HU ET AL.
(1)
5.3. Finite dimensional convergence of Mn (z) in distribution. We will
show in this part that for any positive integer r and any complex numbers
z1 , . . . , zr ∈ Cn , the random vector
[Mn(1) (z1 ), . . . , Mn(1) (zr )]
converges to a 2r-dimensional Gaussian vector. Because of Assumption (c),
without loss of generality, we may assume kΣk ≤ 1 for all p. We will denote
by K any constant appearing in inequalities and it may take different values
at different places.
We first define some quantities which are frequently used in the sequel:
√
rj = (1/ n)xj , D(z) = Bn − zI, Dj (z) = D(z) − rj rj′ ,
1
Dij (z) = D(z) − ri ri′ − rj rj′ , εj (z) = rj′ Dj−1 (z)rj − trΣDj−1 (z),
n
1
1
,
ζj (z) = rj′ Dj−2 (z)rj − trΣDj−2 (z), βj (z) =
′
n
1 + rj Dj−1 (z)rj
1
1
, bn (z) =
.
β̄j (z) =
−1
−1
−1
1 + n trΣDj (z)
1 + n EtrΣDj−1 (z)
Note that, for any z = u + iv ∈ C+ , the last three quantities are bounded
in absolute value by |z|/v. Moreover, D −1 (z) and Dj−1 (z) satisfy
D −1 (z) − Dj−1 (z) = −Dj−1 (z)rj rj′ Dj−1 (z)βj (z).
(5.9)
From Lemma 2.6 in Silverstein and Bai (1995), for any p × p matrix B,
|tr(D −1 (z) − Dj−1 (z))B| ≤
(5.10)
||B||
.
v
Let E0 (·) denote expectation, and Ej (·) the conditional given the σ-field
generated by r1 , . . . , rj . Using the martingale decomposition, we can express
(1)
Mn (z) as
n
n
h
i
X
X
(Ej − Ej−1 )tr D −1 (z) − Dj−1 (z)
(Ej − Ej−1 )trD −1 (z) =
j=1
j=1
=−
n
X
j=1
(Ej −
Ej−1 )βj (z)rj′ Dj−2 rj
n
X
d log(βj (z)/β̄j (z))
(Ej − Ej−1 )
=
,
dz
j=1
where the second equality uses the identity (5.9). By the fact that
βj (z) = β̄j (z) − β̄j (z)βj (z)εj (z) = β̄j (z) − β̄j2 εj (z) + β̄j2 (z)βj (z)ε2j (z),
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
21
we have
(5.11)
Mn(1) (z)
n
d X
(Ej − Ej−1 ) log[1 − β̄j (z)εj (z) + β̄j (z)βj (z)ε2j (z)].
=
dz
j=1
Notice that for all j > 0 and any z ∈ Cn , β̄j (z)εj (z) and β̄j (z)βj (z)ε2j (z) are
almost surely away from 1 when n is large enough. In addition, by Lemma
A.2 and Burkholder’s inequality (Lemma 2.1 in Bai and Silverstein (2004)),
we have
E
n
X
j=1
2
(Ej − Ej−1 )β̄j (z)βj (z)ε2j (z)
= O(δn2 ) → 0.
Therefore, applying Taylor expansion,
Mn(1) (z) = −
=−
n
d X
(Ej − Ej−1 )β̄j (z)εj (z) + op (1)
dz
d
dz
j=1
n
X
Ej (β̄j (z)εj (z)) + op (1).
j=1
For any ǫ > 0,
n
2
d
1 X
d
E Ej εj (z)β̄j (z)
E Ej εj (z)β̄j (z) I(|Ej d εj (z)β̄j (z)|≥ǫ) ≤ 2
dz
dz
ǫ
dz
j=1
j=1
n
|z|8 p4 E|εj (z)|4
K X |z|4 E|ζj (z)|4
≤
+
ε2
v4
v 16 n4
n
X
j=1
which tends to zero according to Lemma A.3, and thus Lindeberg’s condition
is verified. Therefore, from the martingale CLT (Theorem 35.12 Billingsley
(1)
(1995)), the random vector (Mn (zj )) tends to a 2r-dimensional zero-mean
Gaussian vector (M (zj )) with covariance function Cov(M (z1 ), M (z2 )) being
(5.12)
lim
n→∞
n
X
j=1
∂2
Ej−1 Ej εj (z1 )β̄j (z1 ) · Ej εj (z2 )β̄j (z2 ) ,
∂z1 ∂z2
provided that this limit exits in probability. The same argument in page 571
of Bai and Silverstein (2004) implies that it suffices to show
(5.13)
n
X
j=1
Ej−1
2
Y
k=1
Ej β̄j (zk )εj (zk )
4
22
J. HU ET AL.
converges in probability. In addition, by the martingale decomposition,
2
2 −2
E|β̄j (z) − bn (z)| =|bn (z)| n
n
X
−1
E β̄1 (z)
(Ek − Ek−1 )tr(D1−1 (z) − D1k
(z))
k=2
4 −6 −1
(5.14)
≤K|z| v
n
,
where the inequality is from (5.10).Thus it is sufficient to study the convergence of
(5.15)
bn (z1 )bn (z2 )
n
X
Ej−1 (Ej εj (z1 )Ej εj (z2 )) ,
j=1
whose second mixed partial derivative yields the limit of (5.12). Applying
Lemma A.1, we know that
Eξ 4
2Eξ 4
(5.16)
(5.15) = n
− 1 T1 +
T2 ,
p(p + 2)
p(p + 2)
where
T1 = bn (z1 )bn (z2 )
n
1 X
tr[ΣEj Dj−1 (z1 )]tr[ΣEj Dj−1 (z2 )],
n3
j=1
n
1 X
tr[ΣEj Dj−1 (z1 )ΣEj Dj−1 (z2 )].
T2 = bn (z1 )bn (z2 ) 2
n
j=1
We note that the statistic T2 has the same form as Equation (2.8) in
Bai and Silverstein (2004), which has been handled under their model. Following their calculations and using Lemmas A.2-A.3 instead, one may get
i.p.
T2 −−→
(5.17)
Z
0
a(z1 ,z2 )
1
dz,
1−z
where
a(z1 , z2 ) =
Z
m(z1 )m(z2 )(z1 − z2 )
cm(z1 )m(z2 )t2 dH(t)
=1+
,
(1 + tm(z1 ))(1 + tm(z2 ))
m(z2 ) − m(z1 )
and
(5.18)
m′ (z1 )m′ (z2 )
∂ 2 T2 i.p.
1
−−→
−
.
∂z1 ∂z2
(m(z1 ) − m(z2 ))2 (z1 − z2 )2
2
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
23
Now we derive the limit of T1 and its second mixed partial derivative,
which is new compared with the linear transform model. Denote
−1
βij (z) = (1 + ri′ Dij
(z)ri )−1 ,
−1
b1 (z) = (1 + n−1 EtrΣD12
(z))−1 .
By similar proofs of (5.14) and Equation (4.3) of Bai and Silverstein (1998),
one may get |b1 (z) − bn (z)| ≤ Kn−1 and |bn (z) − Eβ1 (z)| ≤ Kn−1/2 , respectively. Also, by Equation (2.2) of Silverstein (1995) and discussions in
Section 5 of Bai and Silverstein (1998), we obtain
Eβ1 (z) = −zEmn (z)
and |Emn (z) − mF cn,Hp (z)| ≤ Kn−1 ,
respectively. Therefore, we get
|b1 (z) + zmF cn ,Hp (z)| ≤ Kn−1/2 .
(5.19)
With the quantity b1 (z), we define a nonrandom matrix L(z) for the purpose
of replacing Dj (z) in T1 ,
L(z) = −zI +
n−1
b1 (z)Σ,
n
which satisfies
(5.20)
||L(z)||−1 ≤
1 + p/(nv)
|b−1
|b−1
1 (z)|
1 (z)|
≤
.
≤
−1
ℑ(z)
v
ℑ(zb1 (z))
−1
By the identity ri′ Dj−1 (z) = βij (z)ri′ Dij
(z), we get their difference
(5.21)
Dj−1 (z) − L−1 (z) = b1 (z)R1 (z) + R2 (z) + R3 (z),
where
R1 (z) = −
R2 (z) = −
X
i6=j
X
i6=j
−1
L−1 (z)(ri ri′ − n−1 Σ)Dij
(z),
−1
(βij (z) − b1 (z))L−1 (z)ri ri′ Dij
(z),
R3 (z) = − n−1 b1 (z)L−1 (z)Σ
X
i6=j
−1
Dij
(z) − Dj−1 (z) .
For any p × p (nonrandom) matrix M , from (5.10), (5.20), and Lemma A.3,
we get
−1
−1
E|trR1 (z)M | ≤nE 1/2 |r1′ D12
(z)M L−1 (z)r1 − n−1 trΣD12
(z)M L−1 (z)|2
(1 + p/(nv))
(5.22)
,
≤n1/2 K||M ||
v2
24
J. HU ET AL.
2
−1
M L−1 (z)r1
E|trR2 (z)M | ≤nE 1/2 (|β12 (z) − b1 (z)|2 )E 1/2 r1′ D12
|z|2 (1 + p/(nv))
,
v5
|z|(1 + p/(nv))
|trR3 (z)M | ≤ ||M ||
.
v3
≤n1/2 K||M ||
(5.23)
(5.24)
Hence, plugging (5.21) into T1 and applying the inequalities (5.19), (5.22)(5.24), we have
2
Y
trEj Dj−1 (zk )Σ =
k=1
2
Y
trL−1 (zk )Σ + Q1 (z1 , z2 )
k=1
Z
2
Y
tdHp (t)
1
+ Q2 (z1 , z2 ),
=p
zk
1 + tmF cn ,Hp (zk )
2
k=1
where E|Qk (z1 , z2 )| ≤ Kn3/2 , k = 1, 2. We thus get
T1 =
2
Y
mF cn ,Hp (zk )
k=1
Z
2
cn tdHp (t)
i.p. Y
+ op (1) −−→
(1 + zk m(zk ))
1 + tmF cn ,Hp (zk )
k=1
whose second mixed partial derivative is
(5.25)
i.p.
∂ 2 T1 /(∂z1 ∂z2 ) −−→ (m(z1 ) + z1 m′ (z1 ))(m(z2 ) + z2 m′ (z2 )).
The result in (5.25) can be obtained by Vitali’s convergence theorem (Lemma
2.3 in Bai and Silverstein (2004)).
Collecting results in (5.16), (5.18) and (5.25), we finally get
Cov(M (z1 ), M (z2 )) =(m(z1 ) + z1 m′ (z1 ))(m(z2 ) + z2 m′ (z2 ))
+ 2m′ (z1 )m′ (z2 )/(m(z1 ) − m(z2 ))2 − 2/(z1 − z2 )2 ,
which completes the proof of Step 1.
(1)
(1)
5.4. Tightness of Mn (z). The tightness of Mn (z) can be established
by verifying the moment condition (12.51) of Billingsley (1968), i.e.,
(5.26)
sup
n,z1 ,z2 ∈Cn
E|Mn(1) (z1 ) − Mn(1) (z2 )|2 /|z1 − z2 |2 < ∞.
By the martingale decomposition and the equality
mn (z1 ) − mn (z2 ) = (z1 − z2 )p−1 tr(D −1 (z1 )D −1 (z2 )),
25
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
to show (5.26), it is sufficient to prove the absolute second moment of
n
X
(Ej − Ej−1 )tr[D −1 (z1 )D −1 (z2 )]
j=1
is bounded uniformly. We first show the uniformly boundedness of E||D −q (z)||
on C for any fixed q > 0. Note that D −1 (z) is bounded on z ∈ Cu . While for
z ∈ Cl ∪ Cr , applying Lemma A.4 with suitable large s, we have uniformly
E||D −1 (z)||q ≤ K +
1
n
P (||Bn || > ηr or λB
min < ηl ) ≤ K + o(1),
vq
√ 2
√
where lim supp ||Σ||(1 + c)2 < ηr < xr and xl < ηl < lim inf p λΣ
min (1 − c) .
Analogously E||Dj−1 (z)||q has the same order, and we thus get
−1
(z)||q } ≤ Kq .
max{E||D −1 (z)||q , E||Dj−1 (z)||q , E||Dij
(5.27)
Then, (5.26) can be obtained by the same procedure in Section 3 of Bai and Silverstein
(2004), applying Lemmas A.2-A.4 together with (5.27). We omit the details.
(2)
(2)
5.5. Convergence of Mn (z). Next we will show that for z ∈ Cn , {Mn (z)}
converges to (5.3), is bounded and forms a uniformly equicontinuous family.
We first introduce some auxiliary results, which can be verified by applying Lemma A.4 in our theoretical framework through a similar proof of the
same statements in Bai and Silverstein (2004). First of all, we note that
(5.28) sup |Emn (z) − m(z)| → 0 and
z∈Cn
sup ||V −1 (z)|| < ∞,
n,z∈Cn
where V (z) = Emn (z)Σ + I. Then, for any nonrandom p × p matrix M ,
(5.29)
E|trD −1 (z)M − EtrD −1 (z)M |2 ≤ K||M ||2 .
Next, there exists a number θ ∈ (0, 1) such that for all n large enough
(5.30)
sup cn
z∈Cn
Z
(tEmn (z))2
dHp (t) < θ.
(1 + tEmn (z))2
Lastly, from (4.12) of Bai and Silverstein (2004) and (5.2) in Bai and Silverstein
(1998), we have that
(5.31)
Mn(2) (z) = −
1−
R
mF cn ,Hp (z)Qn (z)
cn Emn (z)m cn ,Hp (z)t2 dHp (t)
F
(1+tEmn (z))(1+tm cn ,Hp (z))
F
,
26
J. HU ET AL.
where
Qn (z) =n cn
Z
dHp (t)
+ zcn Emn (z)
1 + tEmn (z)
=nEβ1 (z) r1′ D1−1 (z)V −1 (z)r1 − n−1 EtrV −1 (z)ΣD −1 (z) .
(5.32)
From (5.30) and an analog inequality involving mF cn ,Hp (z), the denominator
of (5.31) is bounded away from zero. Therefore, we need only to study the
limit of Qn (z) for z ∈ Cn .
For simplicity, we suppress the variable z from expressions in the sequel
when there is no confusion. Let ̺1 := ̺1 (z) = r1′ D1−1 r1 − (1/n)EtrΣD1−1 .
By the equality
β1 = bn − bn β1 ̺1 = bn − b2n ̺1 + b2n β1 ̺21 ,
(5.33)
(1)
(2)
(3)
we have Qn = Qn + Qn + Qn , where
−1 −1
Q(1)
Σ − trV −1 ΣD −1 ),
n = bn E(trD1 V
2
′ −1 −1
Q(2)
r1 − n−1 trD1−1 V −1 Σ),
n = −nbn E̺1 (r1 D1 V
2
2
′ −1 −1
Q(3)
r1 − n−1 EtrV −1 ΣD −1 .
n = nbn Eβ1 ̺1 r1 D1 V
(1)
For Qn , apply (5.9) and (5.33) again,
EtrV −1 Σ(D1−1 − D −1 ) = Eβ1 r1′ D1−1 V −1 ΣD1−1 r1
=bn n−1 EtrD1−1 V −1 ΣD1−1 − bn Eβ1 ̺1 r1′ D1−1 V −1 ΣD1−1 r1 .
(5.34)
By Lemma (A.2), Hölder’s inequality and the fact that r1′ D1−1 V −1 ΣD1−1 r1 ,
bn and β1 are all bounded for z ∈ Cn , the second term in Equation (5.34)
(3)
is o(1). Analogously, we can get that Qn = o(1). Together with applying
(1)
Lemma A.1 to Qn , we finally obtain that
Eξ 4
2 −1
− 1 EtrD1−1 ΣEtrD1−1 V −1 Σ
Q n = − bn n
p(p + 2)
2Eξ 4
2 −1
− bn n
− 1 EtrD1−1 V −1 ΣD1−1 Σ + o(1)
p(p + 2)
2 (4)
2 (5)
:= − (τ − 2)c−1
n bn Qn − bn Qn + o(1),
(5.35)
(4)
(5)
where Qn = n−2 EtrD −1 ΣEtrD −1 V −1 Σ and Qn = n−1 EtrD −1 V −1 ΣD −1 Σ.
(5)
The limit of Qn can be obtained by a similar approach to deriving (4.13)(4.22) in Bai and Silverstein (2004). It turns out that
−1
Z
Z
t2 dHp (t)
cn
(tEmn )2 dHp (t)
(5.36) Q(5)
+ o(1).
=
1
−
c
n
n
z2
(1 + tEmn )3
(1 + tEmn )2
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
27
(4)
The quantity Qn is new under the elliptical model. To study its limit,
similar to (5.21), we approximate the matrix D −1 (z) by
L̃ = −zI + bn Σ.
Notice that
bn = Eβ1 + O(n−1/2 ) = −zEmn + O(n−1/2 ) → −zm,
as n → ∞. By (5.28), it follows that L̃ is nonsingular and ||L̃−1 || is bounded.
Then, analogous to (5.21)-(5.24), we have
(5.37)
D −1 − L̃−1 = bn R̃1 + R̃2 + R̃3 ,
where
(5.38)
(5.39)
|EtrR̃1 M | ≤ n1/2 K,
|EtrR̃2 M | ≤ n1/2 K(E||M ||4 )1/4 ,
|trR̃3 M | ≤ K(E||M ||2 )1/2
for any p × p nonrandom matrix M with bounded norm. From (5.37)-(5.38),
we have that
Z
tdHp (t)
cn
−1
−1
(5.40)
+ o(1),
n EtrD Σ = −
z
1 + tEmn
Z
tdHp (t)
cn
(5.41)
+ o(1),
n−1 EtrD −1 V −1 Σ = −
z
(1 + tEmn )2
Equations (5.40) and (5.41) imply that
Z
Z
tdHp (t)
c2n
tdHp (t)
(4)
+ o(1).
(5.42)
Qn = 2
z
1 + tEmn
(1 + tEmn )2
Combining (5.31), (5.35), (5.36) and (5.42), we finally get
Z
Z
tmF cn ,Hp dHp (t)
tm2n dHp (t)
(2)
Mn (z) = (τ − 2)cn
1 + tEmn
(1 + tEmn )2
−1
Z 2
Z
t mF cn ,Hp m2n dHp (t)
(tEmn )2 dHp (t)
+ cn
1 − cn
(1 + tEmn )3
(1 + tEmn )2
−1
Z
Emn mF cn ,Hp t2 dHp (t)
× 1 − cn
+ o(1)
(1 + tEmn )(1 + tmF cn ,Hp )
Z
Z
(m′ t)2 dH(t)
(zm + 1)m′ tdH(t)
+
c
,
→(τ − 2)
2
(1 + tm)
m(1 + tm)3
where m′ = m′ (z) denotes the derivative of m(z) with respective to z.
The boundedness and uniform equicontinuity for z ∈ Cn can be verified
directly following the arguments on Pages 592-593 of Bai and Silverstein
(2004). So we omit them here. Then, the proof of Theorem 2.2 is complete.
28
J. HU ET AL.
APPENDIX A: APPENDIX
These lemmas can be viewed as extensions of independent cases. Their
proofs are postponed to the supplementary material (Hu et al., 2017).
Lemma A.1. Let x = ξu where ξ and u are defined in Assumption (b).
Then, for any p × p complex matrices C and C̃,
(A.1)
Eξ 4
E(x′ Cx−trC)(x′ C̃x−trC̃) =
trCtrC̃ + trC C̃ ′ + trC C̃ −trCtrC̃.
p(p + 2)
Lemma A.2. Let x = ξu where ξ satisfies (5.8), independent of u ∼
U (S p−1 ), then for any p × p complex matrix C and q ≥ 2,
E|x′ Cx − trC|q ≤ K||C||q δnq−2 pq−1 ,
(A.2)
where K is a positive constant depending only on q.
√
Lemma A.3. Let r = ξu/ n where ξ satisfies (5.8), independent of
u ∼ U (S p−1 ). Then, for any nonrandom p × p matrix Ck , k = 1, . . . , q1 and
C̃l , l = 1, . . . , q2 , q1 , q2 ≥ 0,
!
q1
q2
q1
q2
Y
Y
Y
Y
−(1∧q2 ) (q2 −2)∨0
′
′
−1
≤ Kn
δn
||Ck ||
||C̃l ||,
E
r Ck r (r C̃l r − n trC̃l )
k=1
k=1
l=1
l=1
where K is a positive constant depending on q1 and q2 .
Lemma A.4.
Suppose (5.8) holds. Then, for any positive s,
P (||Bn || > ηr ) = o(n−s ),
√
whenever ηr > lim supp→∞ ||Σ||(1 + c)2 . If 0 < lim inf p→∞ λΣ
min I(0,1] (c)
then,
−s
n
P (λB
min < ηl ) = o(n ),
√ 2
whenever 0 < ηl < lim inf p→∞ λΣ
min I(0,1) (c)(1 − c) .
ACKNOWLEDGEMENT
The authors are grateful to Prof. Zhidong Bai for discussions on the innovation of high-dimensional elliptical models, and to Prof. Jianfeng Yao for
discussions on spiked covariance models. They also would like to thank the
editors and referees whose careful and detailed comments have led to great
improvements of the manuscript.
COVARIANCE MATRICES IN ELLIPTICAL DISTRIBUTIONS
29
REFERENCES
Anderson, T. W. (2003). An introduction to multivariate statistical analysis, 3rd. ed., John
Wiley & Sons, Hoboken, NJ.
Bai, Z. D., Jiang, D. D., Yao, J. F., and Zheng, S. R. (2009). Corrections to LRT on
large-dimensional covariance matrix by RMT. Ann. Statist., 37, 3822–3840.
Bai, Z. D., Hu, J., Pan, G. M., and Zhou, W. (2015). Convergence of the empirical spectral
distribution function of Beta matrices. Bernoulli, 21, 1538–1574.
Bai, Z. D. and Silverstein, J. W. (2004). CLT for linear spectral statistics of largedimensional sample covariance matrices. Ann. Probab., 32, 553–605.
Bai, Z. D. and Silverstein, J. W. (1998). No eigenvalues outside the support of the limiting
spectral distribution of large dimensional random matrices. Ann. Probab., 26, 316–345.
Bai, Z. D., Silverstein, J. W. (2010). Spectral analysis of large dimensional random matrices, 2nd. ed., Springer, New York.
Bai, Z. D., Zhou, W. (2008). Large sample covariance matrices without independence
structures in columns. Statist. Sinica, 18, 425–442.
Berthet, Q. and Rigollet, P. (2013). Optimal detection of sparse principal components in
high dimension. Ann. Statist. 41, 1780-1815.
Billingsley, P. (1968). Convergence of Probability Measures. Wiley, New York.
Billingsley, P. (1995). Probability and Measure, 3rd. ed., Wiley, New York.
Donoho, D. and Jin, J. S. (2015). Higher Criticism for Large-Scale Inference, Especially
for Rare and Weak Effects. Statistical Science, 30, 1–25.
El Karoui, N. (2009). Concentration of measure and spectra of random matrices: Applications to correlation matrices, elliptical distributions and beyond. Ann. Appl. Probab.,
19, 2362-2405.
Fang, K. T., Zhang, Y. T. (1990). Generalized multivariate analysis. Springer-Verlag,
Berlin; Science Press, Beijing.
Gao, J. T., Han, X., Pan, G. M., and Yang, Y. R. (2017). High dimensional correlation
matrices: the central limit theorem and its applications. J. Roy. Stat. Soc. Ser. B, 79,
677–693.
Gupta, A. K., Varga, T., Bodnar, T. (2013). Elliptically contoured models in statistics and
portfolio theory, 2nd. ed., Springer, New York.
Hu, J., Li, W.M., Liu, Z. and Zhou W. Supplementary material to “High-dimensional
covariance matrices in elliptical distributions with application to spherical test”.
John, S., (1972). The distribution of a statistic used for testing sphericity of normal distributions. Biometrika, 59, 169–173.
Jonsson, D. (1982). Some limit theorems for the eigenvalues of a sample covariance matrix.
J. Multivariate Anal., 12, 1–38.
Johnstone, I. M. (2001). On the distribution of the largest eigenvalue in principal components analysis. Ann. Statist., 29, 295–327.
Kelker, D. (1970). Distribution theory of spherical distributions and a location-scale parameter generalization. Sankhyā Ser. A, 32, 419–430.
Kotz, S. (1975). “Multivariate Distributions at a Cross-Road”, in Statistical Distributions
in Scientific Work, 1, 247–270.
Ledoit, O. and Wolf, M. (2002). Some hypothesis tests for the covariance matrix when the
dimension is large compared to the sample size. Ann. Statist., 30, 1081–1102.
Li, W. M. and Yao, J. F. (2017). On structure testing for component covariance matrices
of a high-dimensional mixture. J. Roy. Stat. Soc. Ser. B, to appear.
Onatski, A., Moreira, M. J., and Hallin, M. (2013). Asymptotic power of sphericity tests
for high-dimensional data. Ann. Statist., 41, 1204–1231.
30
J. HU ET AL.
Onatski, A., Moreira, M. J., and Hallin, M. (2014). Signal detection in high dimension:
The multispiked case. Ann. Statist., 42, 225–254.
Pan, G. M. and Zhou, W. (2008). Central limit theorem for signal-to-interference ratio of
reduced rank linear receiver. Ann. Appl. Probab., 18, 1232–1270.
Paindaveine, D. and Verdebout, T. (2016) On high-dimensional sign tests. Bernoulli, 22,
1745–1769.
Rama, C. (2001). Empirical properties of asset returns: stylized facts and statistical issues.
Quant. Finance, 1, 223-236.
Schott, J. R. (2007). A test for the equality of covariance matrices when the dimension is
large relative to the sample size. Comput. Statist. Data Anal., 51, 6535–6542.
Silverstein, J. W. (1995). Strong convergence of the empirical distribution of eigenvalues
of large-dimensional random matrices. J. Multivariate Anal., 55, 331–339.
Silverstein, J. W. and Bai, Z. D. (1995). On the empirical distribution of eigenvalues of a
class of large dimensional random matrices. J. Multivariate Anal., 54, 175–192.
Tian, X. T., Lu, Y. T., and Li, W. M. (2015). A robust test for sphericity of high dimensional covariance matrices. J. Multivariate Anal., 141, 217–227.
Wang, Q. W. and Yao, J. F. (2013). On the sphericity test with large-dimensional observations. Electronic Journal of Statistics, 7, 2164–2192.
Yang, Y. R. and Pan, G. M. (2015). Independence test for high dimensional data based
on regularized canonical correlation coefficients. Ann. Statist., 43, 467–500.
Yin, Y. Q., Bai, Z. D., and Krishnaiah, P. R. (1988). On limit of the largest eigenvalue
of the large dimensional sample covariance matrix. Probab. Theory Related Fields, 78,
509–521.
Zou, C. L., Peng, L. H., Feng, L., and Wang, Z. J. (2014). Multivariate-sign-based highdimensional tests for sphericity. Biometrika, 101, 229–236.
KLASMOE and School of Mathematics & Statistics School of Statistics and Management
Northeast Normal University
Shanghai University of Finance and Economics
Changchun China 130024
Shanghai China 200433
E-mail: [email protected]
E-mail: [email protected]
Faculty of Science and Technology Department of Statistics and Applied Probability
University of Macau
National University of Singapore
Macau China
Singapore 117546
E-mail: [email protected]
E-mail: [email protected]
| 10math.ST
|
arXiv:cs/0702131v1 [cs.CE] 22 Feb 2007
AICA: a New Pair Force Evaluation Method
for Parallel Molecular Dynamics in Arbitrary
Geometries
Graham B. Macpherson ∗ and Jason M. Reese
a Department
of Mechanical Engineering, University of Strathclyde,
Glasgow, G1 1XJ, UK
Abstract
A new algorithm for calculating intermolecular pair forces in Molecular Dynamics
(MD) simulations on a distributed parallel computer is presented. The Arbitrary
Interacting Cells Algorithm (AICA) is designed to operate on geometrical domains
defined by an unstructured, arbitrary polyhedral mesh, which has been spatially
decomposed into irregular portions for parallelisation. It is intended for nano scale
fluid mechanics simulation by MD in complex geometries, and to provide the MD
component of a hybrid MD/continuum simulation. AICA has been implemented
in the open-source computational toolbox OpenFOAM, and verified against a published MD code.
Key words: molecular dynamics, nano fluidics, hybrid simulation, intermolecular
force calculation, parallel computing
PACS: 31.15.Qg, 47.11.Mn
1
Introduction and Motivation
1.1 MD Simulation in Arbitrary Geometries
Simulations of nano scale liquid systems can provide insight into many naturallyoccurring phenomena, such as the action of proteins that mediate water transport across biological cell membranes [1]. They may also facilitate the design
∗ Corresponding author.
Email addresses: [email protected] (Graham B.
Macpherson), [email protected] (Jason M. Reese).
Preprint submitted to arXiv
6 December 2017
of future nano devices and materials (e.g. high-throughput, highly selective
filters or lab-on-a-chip components). The dynamics of these very small systems are dominated by surface interactions, due to their large surface area
to volume ratios. However, these surface effects are often too complex and
material-dependent to be treated by simple phenomenological parameters [2]
or by adding ‘equivalent’ fluxes at the boundary [3]. Direct simulation of the
fluid using molecular dynamics (MD) presents an opportunity to model these
phenomena with minimal simplifying assumptions.
Some MD fluid dynamics simulations have been reported [4–8], but MD is
prohibitively computationally costly for simulations of systems beyond a few
tens of nanometers in size. Fortunately, the molecular detail of the full flowfield that MD simulations provide is often unnecessary; in liquids, beyond 5–10
molecular diameters (. 3nm for water) from a solid surface the continuumfluid approximation is valid and the Navier-Stokes equations with bulk fluid
properties may be used [2, 9, 10]. Hybrid simulations have been proposed [11–
14] to simultaneously take advantage of the accuracy and detail provided by
MD in the regions that require it, and the computational speed of continuum
mechanics in the regions where it is applicable. An example application of
this technique is shown schematically in figure 1, where a complex molecule
is being electrokinetically transported into a nanochannel for separation and
identification [15]. Only the complex molecule, its immediately surrounding
solvent molecules, and selected near-wall regions require an MD treatment;
the remainder of the fluid (comprising the vast majority of the volume) may
be simulated by continuum mechanics. A hybrid simulation would allow the
effect of different complex molecules, solvent electrolyte composition, channel
geometry, surface coatings and electric field strengths to be analysed, at a
realistic computational cost.
In order to produce a useful, general simulation tool for hybrid simulations,
the MD component must be able to model complex geometrical domains. This
capability does not exist in currently available MD codes: domains are simple
shapes, usually with periodic boundaries. The most important, computationally demanding, and difficult aspect of any MD simulation is the calculation
of intermolecular forces. This paper describes an algorithm that is capable of
calculating pair force interactions in arbitrary, unstructured mesh geometries
that have been parallelised for distributed computing by spatial decomposition.
1.2 Neighbour Lists are Unsuitable
The conventional method of MD force evaluation in distributed parallel computation is to use the cells algorithm to build ‘neighbour lists’ for interacting
2
pairs with the ‘replicated molecule’ method of providing interactions across
periodic boundaries and interprocessor boundaries, where the system has been
spatially partitioned [16–18].
The spatial location of molecules in MD is dynamic, and hence not deducible
from the data structure that contains them. A neighbour list defines which
pairs of molecules are within a certain distance of each other, and as such need
to interact via intermolecular forces.
Neighbour lists are unsuitable when considering systems of arbitrary geometry,
that may have been divided into irregular and complex mesh segments using
standard mesh partitioning techniques (see, for example, figure 2) for two
reasons:
(1) Interprocessor molecule transfers: A molecule may cross an interprocessor boundary at any point in time (even part of the way through a
timestep), at which point it should be deleted from the processor it was
on and an equivalent molecule created on the processor on the other side
of the boundary. Given that neighbour lists are constructed as lists of
array indices, references or pointers to the molecule’s location in a data
structure, deleting a molecule would invalidate this location and require
searching to remove all mentions of it. Likewise, creating a molecule would
require the appropriate new pair interactions to be identified. Clearly this
is not practical. It is conventional to allow molecules to ‘stray’ outside of
the domain controlled by a processor and carry out interprocessor transfers (deletions and creations) during the next neighbour list rebuild. This
is only possible when the spatial region associated with a processor can
be simply defined by a function relating a position in space to a particular
cell on a particular processor (i.e. a uniform, structured mesh, representing a simple domain). In a geometry where the space in question is defined
by a collection of individual cells of arbitrary shape, this is not possible.
For example the location the molecule has ‘strayed’ to may be on the
other side of a solid wall on the neighbour processor, or across another
interprocessor boundary. For the molecule to end up unambiguously in
the correct place, interprocessor transfers must happen as molecules cross
a face.
(2) Spatially resolved flow properties: MD simulations used for flow
studies must be able to spatially resolve fluid mechanic and thermodynamic fields. This is achieved by accumulating and averaging measurements of the properties of molecules in individual cells of the same mesh
that defines the geometry. If a molecule is allowed to stray outside of
the domain controlled by a processor, as above, then it would not be unambiguous and automatic which cell’s measurement the molecule should
contribute to.
3
While both of these problems could be mitigated by, for example, working out
which cell a molecule outside the domain should be in on another processor and
sending its information, doing so would result in an inelegant and inflexible
arrangement with each additional simulation feature (i.e. a different measured
variable or class of intermolecular potential) requiring special treatment.
We have, therefore, developed a new algorithm which is of comparable computational cost to neighbour lists, but designed to be powerful and generic for
simulation in arbitrary geometries. As will be shown, neighbour lists also have
some unfavourable features that could be improved upon.
2
Arbitrary Interacting Cells Algorithm (AICA)
2.1 Replicated Molecule Periodicity and Parallelisation
When parallelising an MD calculation, the spatial domain of the simulation is
decomposed and each processor is given responsibility for a single region [16].
Molecules are, of course, able to cross the boundaries between these regions
and therefore need to be communicated from one processor to the next. Processors also communicate when carrying out intermolecular force calculations,
in which molecules close to processor boundaries need to be replicated on
their neighbours to provide interactions. This process is illustrated in figure 3.
Periodic boundaries also require information about molecules that are not adjacent physically in the domain (see figure 4), these required interactions can
also be constructed by creating copies of molecules outside the boundary.
It is possible to handle processor and periodic boundaries in exactly the same
way, since they have the same underlying objective: molecules near to the edge
of a region need to be copied either between processors or to other locations
on the same processor at every timestep to provide interactions. This is a
useful feature because decomposing a mesh for parallelisation will often turn
a periodic boundary into a processor boundary. The issue is how to efficiently
identify which molecules need to be copied, and to which location, because
this set continually changes as the molecules move.
2.2 Interacting Cell Identification
The new Arbitrary Interacting Cells Algorithm (AICA) we propose here is
an extension of the Conventional Cells Algorithm (CCA) [16]. In the CCA,
a simple (usually cuboid) simulation domain is subdivided into equally sized
4
cells. For computational and theoretical reasons, intermolecular potentials do
not extend to infinity; they are assigned a cut-off radius, rcut , beyond which
they are set to zero. The minimum dimension of the CCA cells must be greater
than rcut , so that all molecules in a particular cell interact with all other
molecules in their own cell and with those in their nearest neighbour cells (i.e.
those they share a face, edge or vertex with — 26 in 3D). AICA uses the same
type of mesh as would be used in Computational Fluid Dynamics (CFD) to
define the geometry of a region. There are no restrictions on cell size, shape
or connectivity. Each cell has a unique list of other cells that it is to interact
with, this list is known as the Direct Interaction List (DIL) for the Cell In
Question (CIQ). It is constructed by searching the mesh to create a set of cells
that have at least one vertex within a distance of rcut from any of the vertices
of the CIQ, see figure 5.
The DILs are established prior to the start of simulation and are valid throughout because the spatial relationship of the cells is fixed, whereas the set of
molecules they contain is dynamic. In a similar way to the CCA, at every
timestep a molecule in a particular cell calculates its interactions with the
other molecules in that cell and consults the cell’s DIL to find which other
cells contain molecules it should interact with. Information is required to be
maintained stating which cell a molecule is in — this is straightforward and
computationally cheap.
The construction of the DILs and accessing molecules is done in such a way as
to not double-count interactions, similarly to the CCA. For example, if cells
A and B need to interact, cell B will be on cell A’s DIL, but cell A will not
be on cell B’s DIL. When a molecule in cell A interacts with one in cell B
the molecule in cell B will receive the inverse of the force vector calculated to
be added to the molecule in cell A, because the pair forces are reciprocal, i.e.
fab = −fba .
It is possible in rare cases for small errors to be caused by this algorithm: small
slices of volume in cells may not be identified for interaction (see figure 6). The
errors introduced will be slight because the relative volume not accounted for
will be very small: in figure 6, only when molecules are in both the indicated
regions of each cell will either miss out on an interaction, and the intermolecular potential will be small at this distance. A small guard distance could be
added to rcut when constructing the DIL to reduce this error. This is less of a
problem in good quality meshes, e.g. hexahedral rather than tetrahedral cells.
2.2.1 Referred Molecules and Cells
Replicated molecule parallelisation and periodic boundaries are handled in the
same way using referred cells (see figure 5) and referred molecules.
5
Referred molecule: A copy of a real molecule that has been placed in a
region outside a periodic or processor boundary in order to provide the correct intermolecular interaction with molecules inside the domain. A referred
molecule holds only its own position and id (i.e. identification of which type
of molecule it is for heterogenous simulations). Referred molecules are created and discarded at each timestep, and do not report any information
back to their source molecules. Therefore if molecule j on processor 1 needs
to interact with molecule k on processor 2, a separate referred molecule will
be created on each processor.
Referred cell: Referred cells define a region of space and hold a collection
of referred molecules. Each referred cell knows
• which real cell in the mesh (on which processor) is its source;
• the required transformation to refer the positions and orientations of the
real molecules in the source cell to the referred location (see below for the
details of how cells and molecules are referred);
• the positions of all of its own vertices. These are the positions of the vertices of the source cell which have been transformed by the same referring
transform as the referred molecules it contains;
• which real cells are in range of this particular referred cell and hence
require intermolecular interactions to be calculated. This is constructed
once at the start of the simulation in the same way as the DIL for real-real
cell interactions — the vertices of the real cells in the portion of mesh on
the same processor as the referred cell are searched, those with at least
one vertex in range of any vertex of the referred cell are found.
2.2.2 Referring Transformations
A spatial transformation is required to refer a cell across a periodic or processor
boundary. Figure 7 shows the most general case of a cell being referred across
a separated, non-parallel boundary, where,
α0 = Cell with a face on one side of the boundary.
β 0 = Cell with a face on the other side of the boundary, coupled to α0
α1 = Cell α0 referred across the boundary.
C = Cell centre.
n = Face normal unit vector.
Sα0 β 0 = Cβ 0 − Cα0 . Position shift from Cα0 to Cβ 0 .
Rα0 β 0 = Tensor required to rotate −n̂α0 to n̂β 0 , given by,
6
Rα0 β 0
nα × nβ
= − (nα · nβ ) I + (1 + nα · nβ )
|nα × nβ |
!2
+ nα nβ − nβ nα ,
(1)
where I is the identity tensor. The absolute position, P′ , of a molecule has
been transformed across a boundary (with its containing cell) from its original
position, P is required. To derive the P → P′ transform, first the position of
P relative to centre of the coupled face on α0 is given as,
P − Cα0 .
(2)
Then this vector is rotated by the transformation tensor defined by the source
and destination coupled face normals,
Rα0 β 0 · (P − Cα0 ).
(3)
The rotated position is transformed back to global coordinates,
Cα0 + Rα0 β 0 · (P − Cα0 ),
(4)
and then shifted by the relative separation of the coupled face centres, i.e.
P′ = Cα0 + Sα0 β 0 + Rα0 β 0 · (P − Cα0 ),
= Cα0 + Cβ 0 − Cα0 + Rα0 β 0 · (P − Cα0 ),
= Cβ 0 − Rα0 β 0 · Cα0 + Rα0 β 0 · P.
(5)
The result is the same if the point is shifted by Sα0 β 0 prior to rotation by
Rα0 β 0 around Cβ 0 . The final position of P′ is the same if any coupled face
centre/face normal pair on the boundary is used. Therefore, all cells and all
molecules referred across a particular boundary may use the transformation
derived from one cell.
This result can be used to transform the positions of all of the molecules
in a source cell to their correct position in a referred cell. Molecules being
referred also operate on their own rotationally dependant properties (e.g. angular orientation) using the rotation tensor. This transformation is suitable
for multiple periodic boundaries and arbitrary mesh decompositions where existing referred cells must be re-referred by other boundaries. This creates the
7
cell relationships across edges, corners, and also on non-neighbouring processors. See Appendix B for an example. Cell re-referring is achieved by writing
equation (5) as a generic transform,
P′ = y + R · P,
(6)
y = Cβ 0 − Rα0 β 0 · Cα0 ,
R = Rα0 β 0 .
(7)
(8)
where,
If y and R are the transformations required to move P to P′ , then transforming
P′ to P′′ using y′ and R′ gives,
P′′ = y′ + R′ · P′ ,
= y′ + R′ · y + R′ · R · P,
= y′⋆ + R′⋆ · P,
(9)
reduced to the generic form by defining,
y′⋆ = y′ + R′ · y,
R′⋆ = R′ · R.
(10)
(11)
Further re-referrals can be reduced to a single transform in a predictable way,
with only a single vector/tensor pair required to be stored by the referred
cell at any stage. Any number of cell referring transformations may be made
sequentially, but the resulting transformation takes the initial source position (P) and directly transforms it to the final destination, regardless of, and
without having to intermediately visit, the other referred locations along the
way.
2.3 Intermolecular Force Calculation Algorithm Details
At the start of the simulation, the DIL for each cell are created, referred cells
are created and they determine which real cells they must supply interactions
8
to. Details of how these are constructed can be found in Appendix A. Cells
requiring to be referred need not be on processors sharing a boundary with
their destination processor, i.e. AICA must be able to identify interactions between non-neighbouring processors. However, when constructing the referred
cells, processors may only communicate across interprocessor faces, i.e. with
neighbours only.
The geometry of the system is defined by a 3D, static, unstructured mesh. A
cell’s position in the data structure, hence its index, does not imply anything
about its physical location or connectivity. Therefore, the ability to handle
arbitrary geometries comes from deducing locally which cells are in range to
interact. The same net force on any molecule will be calculated irrespective of
cell size and shape (except in rare instances, see section 2.2).
The total intermolecular force acting upon a molecule is calculated at each
timestep by considering two types of interactions. Real-Real interactions
occur between a molecule and others on the same processor. Real-Referred
interactions occur between real molecules and referred molecules arising from
the other side of a processor or periodic boundary. Referred molecules do
not need to calculate interactions between themselves, because every referred
molecule is a copy of a real molecule elsewhere, and as such will receive all of
its real-real and real-referred interactions in-situ.
2.3.1 Real-Real Molecule Force Calculation
At each time step, the real molecules on each processor calculate their intermolecular interactions as follows:
for all cells, denoted I {
for all molecules in cell I, denote mol k
{
for all cells in cell I’s DIL, denote cell J
{
for all molecules in cell J, denote mol l
{
calculate intermolecular force, Fkl,
between mol k and mol l.
Add Fkl to mol k, add Flk = -Fkl to mol l.
}
}
for all molecules in cell I, denote mol k’
{
if index of, or pointer to, mol k’ > mol k
9
then calculate intermolecular force, Fkk’,
between mol k and mol k’.
Add Fkk’ to mol k, add -Fkk’ to mol k’.
}
}
}
Comparison of the index, or pointer to molecules in the same cell, mol k’ > mol k,
means that interactions between molecules in the same cell are not double
counted and a molecule does not calculate an interaction with itself. The order of comparison is not important, as long as all pairs are covered, so using
mol k’ < mol k as a condition would work equally well.
2.3.2 Real-Referred Molecule Force Calculation
At each timestep all real cells which are the source cell of one or more referred cells send the position and id of all of the molecules they contain to
the appropriate referred cell on the appropriate processor. The destination
referred cells perform the appropriate position and orientation transformation when the molecules are received. These referred molecules calculate their
intermolecular interactions with real molecules as follows:
for all referred cells, denoted M {
for all referred molecules in referred cell M, denote mol q
{
for all real cells that referred cell M
interacts with, denote cell N
{
for all real molecules in cell N, denote mol p
{
calculate intermolecular force, Fpq,
between mol p and mol q.
Add Fpq to mol p.
}
}
}
}
2.4 Prediction of Computational Speed
The following analysis determines whether or not AICA is practical, by predicting its computational cost relative to that of the CCA and Neighbour List
Algorithm (NLA).
10
Assuming a system of approximately uniform density, the volume of space
around a particular molecule to be interrogated by a particular algorithm will
be directly proportional to the number of pair interactions to be evaluated,
which is itself directly proportional the computational cost per timestep. For
a given cut off radius, rcut , the volume associated with a particular algorithm
is:
Ideal: Evaluating the minimum possible number of pairs is equivalent to a
sphere of radius rcut ,
4 3
Videal = πrcut
.
(12)
3
However, it is not possible to design an algorithm that only evaluates the
ideal case because the relative spatial arrangement of molecules is not known
in advance from one timestep to the next.
CCA: The volume encompassed by the CCA depends on the cell size — it
is minimum when cells are cubic with side length rcut . The volume of the
cell containing the molecule in question, plus that of its 26 neighbours is
encompassed, giving
3
.
(13)
VCCA = 27rcut
NLA: The neighbour list for a particular molecule includes all molecule interaction pairs within a sphere of radius rcut plus a guard distance ∆R [16].
It is constructed using the CCA, where the cells must be at least rcut + ∆R
in size, and has a lifetime of L timesteps before invalidation. Therefore, each
timestep that the neighbour list is used for must absorb a fraction of the
CCA cost required to construct the list — the shorter the lifetime of the
neighbour list, the more costly the NLA becomes. Therefore,
4
27
VN LA = π (rcut + ∆R)3 +
(rcut + ∆R)3 .
(14)
3
L
AICA: The volume encompassed by AICA is the total volume of all cells
which have at least one vertex within rcut of any vertex of the cell containing
the molecule in question, plus the volume of this cell. This depends on the
local topology of the mesh. As a simple example, consider a uniform mesh
of cubes of side length w. A piecewise function, Nc (rcut /w) has been derived
to count the number of cubes, each of volume w 3 , that have at least one
corner within a distance of rcut /w of a corner of a central cube (which is
included). In this case,
VAICA = Nc (rcut /w)w 3.
(15)
These volumes are plotted on figure 8 with w/rcut as the independent variable,
lower values of w/rcut representing a finer mesh for a given cut-off radius. AICA
is costlier than the CCA or NLA in coarse meshes, however, when w . 0.4rcut
(see inset for detail) AICA has a similar computational cost to the NLA, for
11
the chosen representative values of L and ∆R. The discontinuities in VAICA
are due to the steps in Nc (rcut /w) which arise when a threshold is crossed and
another ‘layer’ of cubes is required.
While real meshes in complex geometries will not comprise uniform cubes, this
result demonstrates that AICA can be of comparable speed to the NLA, and
gives an indication of the required refinement. As the mesh becomes finer,
the set of cells selected to interact with the cell at the centre becomes an
increasingly accurate approximation of the ideal sphere, this can be seen in the
VAICA function, which asymptotically approaches Videal as w/rcut → 0. Note
that, in this example, AICA and the CCA are equivalent when w/ccut = 1
because their cells are the same size.
2.4.1 Neighbour List Lifetime Prediction
The computational cost of the neighbour list algorithm depends on the number
of timesteps a neighbour list remains valid for — a quantity whose dependence
on simulation properties can be predicted.
For a given number of molecules, N, of mass m, in equilibrium at a temperature T , there is a probability of 1/N that a molecule in the system will
be travelling with a velocity vN . It is therefore likely that at every timestep
there will be one molecule travelling at, or close to this velocity. Given that a
neighbour list is invalidated when the cumulative sum of maximum displacements in the system exceeds a fixed threshold (usually 0.5∆R, half the guard
radius [16]), finding the number of timesteps required for a molecule travelling
at vN to cover this distance gives an estimate of the lifetime of the list.
Finding vN by equating the Maxwellian velocity distribution to 1/N, where
T , m and vN are in reduced MD units [19] (see figure 9):
4π
m
2πT
3
2
2
vN
e
−mv 2
N
2T
=
1
.
N
(16)
The high speed solution is,
vN =
v
u
u 2T
u
t−
W−1 −
m
1
4N
s
2πT
,
m
(17)
where W−1 is the secondary real branch of the Lambert W function. Therefore,
L, the number of timesteps (of length ∆t) a neighbour list is valid for, is given
by
12
L=
0.5∆R
.
∆tvN
(18)
This lifetime reduces with increasing temperature and system size, as shown
in figure 10, resulting in a higher computational cost for the NLA. At higher
temperatures, vN is higher because the velocity distribution covers a higher
molecular speed range. In systems comprising more molecules, it is more probable that at least one molecule will be travelling at a given speed in the upper
region of the distribution, therefore vN is also higher.
The cost of AICA is not sensitive to either of these parameters, therefore
it becomes increasingly attractive in large systems (> 106 molecules) where
neighbour list lifetimes are low.
3
Implementation and Verification
AICA has been implemented in OpenFOAM [20, 21], which is an open source
C++ library intended for continuum mechanics simulation of user-defined
physics (primarily used for CFD) in arbitrary, unstructured geometries. The
AICA MD code has been built using OpenFOAM’s lagrangian particle tracking library, which has been used to track particles in solution, droplets in
combusting diesel sprays and to perform DSMC simulations [22].
The AICA algorithm must show that it produces the correct results, assessed
by momentum and energy conservation and the behaviour of average potential
and kinetic energy, in three ways:
(1) The code must produce the same results as another (validated) MD code.
For these purposes the other code is an in-house MD code which has been
validated against those supplied in [16].
(2) The results for the AICA MD code must be the same whether it is run on
one processor as a serial calculation, or run in parallel on many distributed
CPUs.
(3) The topology of the underlying mesh must not make any difference to
the results.
The serial/parallel test case for AICA is shown in figure 11. The test mesh is
decomposed into 24 portions, the three examples shown are not regular shapes
and the cells are nonuniform in size (the mesh is graded in each direction).
While this test mesh is structured, the algorithm makes no use of this fact,
so a mesh comprising cells of any size, shape or connectivity would produce
the same result. This case will therefore demonstrate the independence of the
AICA from the form of the geometry used.
13
Figures 12, 13 and 14 show the average kinetic (KE), potential (PE) and
total (TE) energy per molecule for a system containing 27000 Lennard Jones
molecules, starting in a simple cubic lattice at a temperature of 120K. No
equilibration or thermostat has been applied. Results for the in-house code,
the AICA code in serial on a PC and the AICA code decomposed and run
on 24 processors of a cluster show that the KE and PE graphs follow very
similar trends to the in-house code. The simulations do not follow exactly the
same trajectory because the velocities in the systems are initialised by random
numbers, which are different between the two codes. This is likely to account
for the slight (0.0025%) difference in average TE. The serial and parallel tests
use the same initial velocity configuration and each value for KE, PE and TE
are the same to better than 1 part in 1012 . Momentum is conserved in both
serial and parallel tests to approximately 1 part in 1016 .
Given that the serial and parallel results are exactly the same, the AICA
must be creating the correct referred cells and molecules, and calculating realreferred interactions perfectly, even with the relatively complex decomposed
portions of the mesh.
A comparison of how the computational speed changes with mesh refinement,
compared to the predictions in section 2.4, and further tests of how the AICA
performance scales in parallel will appear in a future publication.
4
Conclusions
Our Arbitrary Interacting Cell Algorithm (AICA) has been described, which
is a new way of calculating pair forces in a molecular dynamics simulation,
carried out on an arbitrary unstructured mesh that has been spatially decomposed to run in parallel. The algorithm is expected to achieve a similar
computational speed to the neighbour list method, but does not suffer from
the identified degradations in performance that neighbour lists experience in
large systems and at high temperatures. AICA has been implemented in the
OpenFOAM code, and produces the same results when a system is simulated
in serial, or in parallel on a mesh decomposed into irregular shapes on 24 processors. The exchange of intermolecular force information across interprocessor
boundaries must therefore be correct.
Currently this algorithm deals only with short ranged pair forces. Future developments will include rigid and flexible polyatomic molecules, and long-range
electrostatic forces. A hybrid MD/continuum implementation in OpenFOAM
is currently under development. OpenFOAM is ideal for incorporating polar
fluids and electrokinetic actuation, which feature in envisioned applications,
because it is able to simulate electromagnetic, magnetohydrodynamic and fluid
14
mechanic continuum fields on the same mesh that AICA operates with.
5
Acknowledgements
The authors would like to thank Chris Greenshields and Matthew Borg of
Strathclyde University (UK), and Henry Weller and Mattijs Janssens of OpenCFD
Ltd. (UK) for useful discussions. This work is funded in the UK by Strathclyde
University and the Miller Foundation, and through a Philip Leverhulme Prize
for JMR from the Leverhulme Trust.
15
References
[1] P. Agre, The aquaporin water channels, Proceedings of the American Thoracic
Society 3 (1) (2006) 5–13.
[2] J. Koplik, J. R. Banavar, Continuum deductions from molecular
hydrodynamics, Annual Review of Fluid Mechanics 27 (1995) 257–292.
[3] H. Brenner, V. Ganesan, Molecular wall effects: are conditions at a boundary
“boundary conditions”?, Physical Review E 61 (6) (2000) 6879–6897.
[4] D. Hirshfeld, D. Rapaport, Molecular dynamics simulation of Taylor-Couette
vortex formation, Physical Review Letters 80 (24) (1998) 5337–5340.
[5] D. Rapaport, E. Clementi, Eddy formation in obstructed fluid flow: a moleculardynamics study, Physical Review Letters 57 (6) (1986) 695–698.
[6] K. Travis, B. Todd, D. Evans, Poiseuille flow of molecular fluids, Physica A
240 (1-2) (1997) 315–27.
[7] K. Travis, B. Todd, D. Evans, Departure from Navier-Stokes hydrodynamics in
confined liquids, Physical Review E 55 (4) (1997) 4288–95.
[8] T. Qian, X.-P. Wang, Driven cavity flow: from molecular dynamics to continuum
hydrodynamics, Multiscale Modeling and Simulation 3 (4) (2005) 749–763.
[9] T. Becker, F. Mugele, Nanofluidics: viscous dissipation in layered liquid films,
Physical Review Letters 91 (2003) 166104.
[10] H. Okumura, D. M. Heyes, Comparisons between molecular dynamics and
hydrodynamics treatment of nonstationary thermal processes in a liquid,
Physical Review E 70 (6) (2004) 061206.
[11] R. Delgado-Buscalioni, P. V. Coveney, Hybrid molecular-continuum fluid
dynamics, Philosophical Transactions of the Royal Society London 362 (1821)
(2004) 1639–1654.
[12] G. Wagner, E. G. Flekkøy, Hybrid computations with flux exchange,
Philosophical Transactions of the Royal Society London 362 (1821) (2004) 1655–
1665.
[13] X. B. Nie, S. Y. Chen, W. E, M. O. Robbins, A continuum and molecular
dynamics hybrid method for micro- and nano-fluid flow, Journal of Fluid
Mechanics 500 (2004) 55–64.
[14] T. Werder, J. H. Walther, P. Koumoutsakos, Hybrid atomistic-continuum
method for the simulation of dense fluid flows, Journal of Computational
Physics 205 (1) (2005) 373–390.
[15] S. Pennathur, J. G. Santiago, Electrokinetic transport in nanochannels. 2.
experiments, Analytical Chemistry 77 (21) (2005) 6782–6789.
16
[16] D. C. Rapaport, The Art of Molecular Dynamics Simulation, 2nd ed.,
Cambridge University Press, 2004.
[17] W. Smith, Molecular dynamics on hypercube parallel computers, Computer
Physics Communications 62 (2–3) (1991) 229—248.
[18] D. Rapaport, Multi-million particle molecular dynamics. II. Design
considerations for distributed processing, Computer Physics Communications
62 (2–3) (1991) 217–228.
[19] M. Allen, D. Tildesley, Computer simulation of liquids, Oxford University Press,
1987.
[20] H. G. Weller, G. Tabora, H. Jasak, C. Fureby, A tensorial approach
to computational continuum mechanics using object-oriented techniques,
Computers in Physics 12 (1998) 620–631.
[21] OpenFOAM, www.openfoam.org.
[22] J. Allen, T. Hauser, foamDSMC: An object oriented parallel DSMC solver
for rarefied flow applications, in: 45th AIAA Aerospace Sciences Meeting and
Exhibit, Reno, Nevada, 2007.
17
A
Cell Interactions
Building interaction lists and identifying and creating referred cells is allowed
to be computationally expensive (within reason) because it will only happen
once at the beginning of the MD simulation. Accessing the information, however, must be as fast as possible because it will happen at every timestep.
Cell interactions are based on the largest cut-off radius of any intermolecular potential in a heterogenous system, r̂cut . It is possible to have different
interaction lists when potentials with substantially different cut-off radii are
used.
A.1
Building DILs
To create each real cell’s DIL, a non double counting loop through all cells is
used, where the cells are accessed by an index (i or j) running 0 → k − 1,
where there are k cells in the portion of mesh on the processor in question,
i.e.
for cells, denote cell i, starting i = 0, while i < k-1 {
for cells, denote cell j, starting j = i+1, while j < k
{
calculate magnitude of spatial separation of all
vertices of cell i with all vertices of cell j.
if any separation distance is less than the
maximum cut off radius, add cell j to cell i’s DIL.
}
}
For example, if there are k = 5 cells, the array index runs from 0 → 4:
i
j
0
1, 2, 3, 4
1
2, 3, 4
2
3, 4
3
4
All cell combinations have been created once, i.e. when i = 2, j = 1 would
produce the 2–1 combination, which is already identified as 1–2. This non
double counting procedure means that some cells will have few or no other
18
cells in their DIL, however, as long as the correct cell pair is created in one of
the cell’s DIL, this is not an issue. It is also designed so that a cell does not
end up on its own DIL, the interactions between molecules in the same cell
are handled separately.
A.2
Creating Referred Cells
Coupled patches are the basis of periodic and interprocessor communication
for creating referred cells. Patches, in general, are a collection of cell faces
representing a mesh boundary of some description — they may provide solid
surfaces, inlets, outlets, symmetry planes, periodic planes, or interprocessor
connections. Coupled patches provide two surfaces; whatever exits one enters
the other and vice versa. Two types are used in AICA:
• Periodic patches on a single processor are arranged into two halves, each
half representing one of the coupled periodic surfaces. When a molecule
crosses a face on one surface, it is wrapped around to the corresponding
position on the corresponding face on the other surface.
• Processor patches provide links between portions of the mesh on different
processors, one half of the patch is on each processor. When a molecule
crosses a face on one surface, it is moved to the corresponding position
on the corresponding face on the other surface, on the other processor.
Decomposing a mesh for parallelisation will often require a periodic patch
to be changed to a processor patch.
The surfaces of coupled patches may have any relative orientation, and may
be spatially separated as long as the face pairs on each surface correspond to
each other. This allows them to also be used as a ‘recycling’ surface in a flow
simulation — whatever exits from an outlet is reintroduced at an inlet, with
the option of scaling the molecule’s properties, e.g. temperature and pressure.
A.2.1 Create Patch Segments.
In the decomposed portion of the mesh on each processor, each processor and
periodic patch should be split into segments, such that:
• faces on a processor that were internal to the mesh prior to decomposition
end up on a segment;
• faces on a processor patch that were on separate periodic patches on the
undecomposed mesh end up on different segments. These segments are further split such that faces that were on different halves of the periodic patch
on the undecomposed mesh end up on different segments;
• faces on different halves of a periodic patch end up on different segments.
19
Each segment must produce a single vector/tensor transformation pair (see
section 2.2.2) which will be applied to all cells referred across it.
A.2.2 Cell Referring Iterations.
For each patch segment:
(1) Find all real and existing referred cells in range (with at least one vertex
within a distance of r̂cut ) of at least one vertex of the faces comprising
the patch segment.
(2) Refer or re-refer this set of cells across the boundary defined by the patch
segment using its transformation, see section 2.2.2. In the case of a segment of periodic boundary, this creates new referred cells on the same
processor. For a segment of a processor patch, these cells are communicated to, and created on the neighbouring processor. Before creating any
new referred cell a check is carried out to ensure that it is not
• a duplicate referred cell, one that has been created already by being
referred across a different segment;
• a referred cell trying to be duplicated on top of a real cell, i.e. a cell
being referred back on top of itself.
If the proposed cell for referral would create a duplicate of an existing
referred cell, or end up on top of a real cell, then it is not created. To be
a duplicate, the source processor, source cell and the vector part of the
transformation (see section 2.2.2) must be the same for the two cells (note,
the vector part of transformation for a real cell is zero by definition).
A single run of these steps will usually not produce all of the required referred
cells. They are repeated until no processor adds a referred cell in a complete
evaluation of all segments, meaning all possible interactions are accounted for.
In iterations after the first, in step 1 it is enough to only search for referred
cells in range of the faces on the patch segment, because the real cells will not
have changed, and would all be duplicates. The final configuration of referred
cells does not depend on the order of patch segment evaluation. Appendix B
contains an example of the results of the cell referring process.
B
Example Construction of Referred Cells
Figures 15 and 16 show the start and end points of the cell referring operation
on a mesh that has been decomposed to run in parallel. The example is in 2D
for clarity but the process is exactly the same in 3D.
In this example the number of referred cells created far exceeds the number
20
of real cells, which would lead to much costly interprocessor communication.
This is because the mesh has been made deliberately small relative to r̂cut to
demonstrate as many features of the algorithm as possible and to be practical
to understand. In realistic systems, the mesh portions would be significantly
bigger than r̂cut and the referred cells would form a relatively thin ‘skin’ around
each portion. Decomposition of the mesh should preferably be carried out to
minimise the number of cells that need to be referred, and to ensure that
the vast majority of the intermolecular interactions happen between real-real
molecule pairs; in this way the communication cost is minimised.
21
Solid
MD Region
Flow
Nanochannel
Complex
molecule
Fig. 1. Schematic of an application of a hybrid MD/continuum simulation: complex molecules being transported into a nano channel. Only the complex molecules,
regions near them and regions near solid surfaces need an MD treatment, the remaining volume can be simulated with continuum mechanics.
22
Fig. 2. Example of a molecular dynamics flow simulation in a complex nano channel.
The mesh is decomposed into five irregular portions, four of which are shown as
shaded blocks, the molecules contained in the fifth portion are shown.
23
A
r cut
Copied
molecules
Processor 1
Processor 2
Fig. 3. Domain decomposition. Molecule A must calculate intermolecular force contributions from all other molecules within its cut-off radius, rcut . When the domain
has been decomposed, some of these molecules may lie on a different processor. In
this case, copies of the appropriate molecules from processor 2 are made on processor
1.
24
B
r cut
Copied
molecules
Fig. 4. Periodic boundaries. Molecule B must calculate intermolecular force contributions from all other molecules within rcut . Some of those molecules may be on a
periodic image of the system, and as such, in implementation terms, will reside on
the other side of the domain. Serial calculations in simple geometries typically use
the minimum image convention [16, 19], but this is not suitable for parallelisation.
25
11
00
00
11
00
11
00
11
00
11
00
11
00
11
00
11
0000
1111
0000
1111
11
00
0000
1111
0000
1111
00
11
rcut
1111
0000
0000
1111
0000
1111
0000
1111
11
00
00
11
00
11
00
11
00
11
00
11
00
11
00
11
Fig. 5. Interacting cell identification using unstructured triangular cells to demonstrate insensitivity to mesh topology. The spatial domain (main square section comprising the real cells) is periodic top-bottom and left-right. Real cells within rcut
that interact with the CIQ (dark) are shaded in grey. The required referred cells
are hatched in alternate directions according to which boundary they have been referred across. Realistic systems would be significantly larger compared to rcut than
shown here.
26
B
A
rcut
Fig. 6. Errors caused by cut-off radius searching from vertices only. An arc of radius
rcut drawn from the indicated vertex on cell A intersects a face of cell B, therefore,
molecules near this vertex should interact with molecules in the shaded region.
Similarly, an arc of radius rcut drawn from the indicated face on cell B encompasses
the small shaded region of cell A. The cell searching algorithm will not, however,
identify cells A and B as needing to interact because none of their vertices lie within
rcut of each other.
27
α0
Cα0
P
Sα0 β 0
β0
−nα0
Cβ 0
α1
Rα0 β 0
nα0
P′
nβ 0
Fig. 7. A molecule at point P is referred across the boundary (heavy line) to P′ by
a transformation defined by the face center/face normal vectors of the faces of cells
α0 and β 0 on the boundary. The mesh is rigid, so all points of cell α1 have the same
relative spatial relationship as α0 . The vertices of referred cell α1 are calculated by
the same process.
28
90
V/rcut3 ∝ Computational Cost
80
70
20
15
60
10
50
0.2
40
0.3
0.4
30
VAICA
VCCA
VNLA, L = 5, ∆R = 0.16rcut
VNLA, L = 10, ∆R = 0.16rcut
Videal
20
10
0
0.2
0.4
0.6
w/rcut
0.8
1
Fig. 8. Predicted computational cost of new AICA using cubic cells, compared
with conventional algorithms. ∆R = 0.16rcut derived from typical values for liquid
density simulation rcut = 2.5, ∆R = 0.4 [16].
29
Probability
1/N
Velocity Magnitude
vN
Fig. 9. Maxwellian velocity distribution. Finding velocities corresponding to a probability of 1/N , where N is the number of molecules in the system. There are two
real, positive solutions — the high speed one is vN , the quantity of interest.
30
Predicted number of timesteps of neighbourList validity
12
T=1
T = 1.5
T = 2.5
11
10
9
8
7
6
5
4
102
103
104
105
Number of molecules, N (m = 1, ∆R = 0.4, ∆t = 0.005)
106
Fig. 10. Variation of estimated neighbour list lifetime with typical simulation parameters for a Lennard Jones fluid [16]. T, m, ∆R and ∆t are in reduced MD units.
31
Fig. 11. Mesh for the serial vs. parallel test; note the grading of cells. The serial test
uses the whole mesh (wireframe) and the parallel test decomposes the mesh into 24
portions (3 of which have been shown).
32
Average kinetic energy per molecule (reduced units)
1.5
1.4
1.407
1.3
1.404
1.2
1.401
1.1
1.398
1
1.395
80
0.9
100
120
140
24 processor parallel, cluster
Serial, PC
In house code
0.8
0.7
0
50
100
150
200
250
300
350
400
Timestep
Fig. 12. Verification of kinetic energy results. Serial and parallel results are identical,
see inset.
33
Average potential energy per molecule (reduced units)
-3.4
-4.123
-3.6
-4.126
-4.129
-3.8
-4.132
-4.135
-4
80
-4.2
100
120
140
24 processor parallel, cluster
Serial, PC
In house code
-4.4
0
50
100
150
200
250
300
350
400
Timestep
Fig. 13. Verification of potential energy results. Serial and parallel results are identical, see inset.
34
Average total energy per molecule (reduced units)
-2.7285
-2.73029
-2.729
-2.73031
-2.73033
-2.7295
-2.73035
-2.73037
-2.73
80
-2.7305
100
120
140
24 processor parallel, cluster
Serial, PC
In house code
-2.731
0
50
100
150
200
250
300
350
400
Timestep
Fig. 14. Verification of total energy results. Serial and parallel results are identical,
see inset.
35
Processor 0
Processor 1
Processor 2
Fig. 15. An arbitrary unstructured mesh, which is periodic top-bottom and left-right. It is decomposed into 3 irregular portions for parallel processing as marked.
36
Real cell
Referred cell
Source on processor 0
11111
00000
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
Referred cell
Source on processor 1
Referred cell
Source on processor 2
0000
1111
0000
00001111
1111
0000
1111
11111
00000
0000
1111
0000
1111
00000
11111
0000
1111
0000
1111
00001111
1111
0000
1111
00000
11111
0000
0000
1111
0000
1111
00000
11111
0000
1111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
00001111
1111
0000
00000
11111
0000
1111
1111
0000
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
Processor 0
1111
0000
0000
1111
00000
11111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
00000
11111
00000
11111
00000
11111
0000
1111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
00000
11111
Processor 1
0000
0000
1111
0000
111111111
0000
00000
11111
00000
11111
0000
1111
00000
11111
000001111
00001111
0000001111
111111
00001111
0000
1111
0000
1111
0000
1111
0000
1111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
000000
111111
0000
1111
0000
1111
0000
1111
00000
11111
00000
11111
000000
111111
0000
1111
00000
11111
00000
11111
0000
1111
00000
11111
0000
1111
0000001111
111111
0000
0000
1111
0000
1111
0000
1111
00000
11111
00000
11111
000000
111111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
000000
111111
000000
111111
0000
1111
000000
111111
0000
1111
0000
1111
0000
1111
0000
1111
00000
11111
00000
11111
00000
11111
0000
1111
00000
11111
00000
11111
000000
111111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
000000
111111
0000
1111
00000
11111
0000
1111
000000
111111
000000
111111
0000
1111
0000
1111
0000
1111
00000
11111
00000
11111
00000
11111
0000
1111
00000
11111
000000
111111
00000
11111
00000
11111
00000
11111
0000
1111
0000
1111
0000
1111
000000
111111
00000
11111
00000
11111
00000
11111
0000
1111
000000
111111
00000
11111
00000
11111
0000
1111
00000
11111
00000
11111
000000
111111
000000
111111
00000
11111
00000
11111
00000
11111
0000
1111
0000
1111
000000
111111
00000
11111
00000
11111
000000
111111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
000
111
00000
11111
00000
11111
00000
11111
000000
111111
00000
11111
0000
1111
0000
1111
00000
11111
00000
11111
00000
11111
111111
000000
000
111
00000
11111
00000
11111
00000
11111
00000
11111
0000
1111
1111
0000
000000
111111
0000
1111
00000
11111
00000
11111
00000
11111
000000
111111
000
111
00000
11111
00000
11111
00000
11111
0000
1111
000000
111111
00000
11111
00000
11111
000000
111111
00000
11111
000000
111111
000000
111111
0000
1111
000000
111111
00000
11111
000000
111111
000000
111111
0000
1111
00000
11111
000000
111111
000000
111111
0000
1111
000
111
00000
11111
000000
111111
000
111
00000
11111
000000
111111
000
111
00000
11111
000000
111111
000
111
00000
11111
000000
111111
00000
11111
000000
111111
00000
11111
000000
111111
00000
11111
00000
11111
0000001111
111111
0000
00000
11111
00000
11111
00000
11111
00000
11111
0000
1111
00000
11111
000000
111111
0000
1111
00000
11111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
00000
11111
000000
111111
000000
111111
0000
1111
00000
11111
00000
11111
00000
11111
00000
11111
000000
111111
0000
1111
00000
11111
00000
11111
0000
1111
00000
11111
000000
111111
0000
1111
0000
1111
000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
000000
111111
0000
1111
00000
11111
00000
11111
0000
1111
00000
11111
000000
111111
00000
11111
00000
11111
000000
111111
0000
1111
000000
111111
0000
1111
0000
1111
00000
11111
00000
11111
00000
11111
0000
1111
000000
111111
0000
1111
00000
11111
00000
11111
00000
11111
0000
1111
0000
1111
000000
111111
00000
11111
00000
11111
0000
1111
000000
111111
000
111
000000
111111
000000
111111
00000
11111
00000
11111
00000
11111
00000
11111
0000
1111
000000
111111
0000
1111
00000
11111
00000
11111
0000
1111
000000
111111
000
111
00000
11111
00000
11111
00000
11111
000000
111111
00000
11111
00000
11111
000000
111111
0000
1111
000
111
000
111
00000
11111
00000
11111
00000
11111
000000
111111
00000
11111
00000
11111
000000
111111
0000
1111
000
111
000
111
00000
11111
000000
111111
000000
111111
000000
111111
00000
11111
00000
11111
000
111
00000
11111
00000
11111
000000
111111
000000
111111
000000
111111
000
111
00000
11111
00000
11111
000000
111111
000000
111111
0000
1111
000000
111111
0000
1111
000
111
000
111
00000
11111
00000
11111
000000
000000
111111
000000
111111
0000111111
1111
0000
1111
000
111
00000
11111
00000
11111
000000
111111
0000
1111
00000
11111
000000
111111
0000
1111
000
111
00000
11111
0000
1111
00000
11111
000000
111111
0000
1111
00000
11111
000000
111111
000
111
Processor
2
00000
11111
0000
1111
0000
1111
00000
11111
000
111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
00001111
1111
0000
1111
000
111
0000
1111
0000
1111
00000
11111
0000
00000
11111
0000
1111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
00000
11111
00000
11111
0000
1111
0000
1111
00000
11111
0000
1111
00000
11111
0000
1111
Fig. 16. The final configuration of referred cells around the portions of the mesh in
figure 15 on each processor. A circle of radius r̂cut drawn from any vertex of a real
cell would be fully encompassed by other real or referred cells, thereby providing all
molecules in that cell with the appropriate intermolecular interactions, either across
a periodic boundary or from another processor. Note that many cells are referred
several times; for example, the source cell marked with ⋆ on processor 2 is referred
to processors 0 and 1 twice on each.
37
| 5cs.CE
|
Homological vanishing for the Steinberg representation
arXiv:1704.08344v3 [math.AT] 27 Nov 2017
Avner Ash
Andrew Putman∗
Steven V Sam†
November 20, 2017
Abstract
For a field k, we prove that the ith homology of the groups GLn (k), SLn (k), Sp2n (k),
SOn,n (k), and SOn,n+1 (k) with coefficients in their Steinberg representations vanish for
n ≥ 2i + 2.
1
Introduction
Let G be a connected reductive group over a field k. A basic geometric object associated to
G(k) is its Tits building. By definition, this is the simplicial complex TG (k) whose i-simplices
are increasing sequences
0 ( P0 ( · · · ( Pi ( G(k)
of parabolic k-subgroups of G(k). Letting r be the semisimple k-rank of G, the complex
TG (k) is (r −1)-dimensional, and the Solomon–Tits theorem [Br1, Theorem IV.5.2] says that
in fact TG (k) is homotopy equivalent to a wedge of (r − 1)-dimensional spheres. Letting R
be a commutative ring, the Steinberg representation of G(k) over R, denoted StG (k; R), is
e
H
r−1 (TG (k); R). This is one of the most important representations of G(k); for instance, if
G is any of the classical groups in Theorem 1.1 below (e.g. G = SLn ) and k is a finite field
of characteristic p, then StG (k; C) is the unique irreducible representation of G(k) whose
dimension is a positive power of p (see [MalZ], which proves this aside from three small
cases that must be checked by hand). See [H] for a survey of many results concerning the
Steinberg representation.
The twisted homology groups Hi (G(k); StG (k; R)) play an interesting role in algebraic
K-theory; see [Q, Theorem 3]. If G(k) is a finite group of Lie type, then StG (k; k) is a
projective G(k)-module (see [H]), and thus the homology groups Hi (G(k); StG (k; k)) all
vanish. However, it is definitely not the case that StG (k; R) is projective for a general
commutative ring R, and if k is an infinite field then StG (k; k) need not be projective.
Our main theorem says that nevertheless for the classical groups, the homology groups
Hi (G(k); StG (k; R)) always vanish in a stable range.
Theorem 1.1. Let Gn be either GLn , SLn , Sp2n , SOn,n , or SOn,n+1 . Then for all fields k
and all commutative rings R, we have Hi (Gn (k); StGn (k; R)) = 0 for n ≥ 2i+2. Furthermore,
there exists a surjection Hi (G2i (k); StG2i (k; R)) → Hi (G2i+1 (k); StG2i+1 (k; R)).
∗
†
AP is supported in part by NSF grants DMS-1255350 and DMS-1737434.
SS is supported in part by NSF grant DMS-1500069.
1
Remark 1.2. When char(k) = 2, the groups SOn,n (k) and SOn,n+1 (k) in Theorem 1.1 are
to be taken naively as the stabilizers of appropriate quadratic forms (see §2.1 below); we
ignore the Dickson invariant.
Theorem 1.1 (and its proof) is motivated by the following conjecture of Church–Farb–
Putman. Recall
that Borel–Serre [BoSe] proved that the virtual cohomological dimension of
n
SLn (Z) is 2 .
n
Conjecture 1.3 ([CFaP, Conjecture 2]). We have H( 2 )−i (SLn (Z); Q) = 0 for n ≥ i + 2.
In other words, the rational cohomology of SLn (Z) vanishes in codimension i as long as n is
sufficiently large. Conjecture 1.3 was proved for i = 0 by Lee–Szczarba [LSz] and for i = 1
by Church–Putman [CP]. It is open for i ≥ 2.
To connect Conjecture 1.3 to Theorem 1.1, recall that Borel–Serre [BoSe] proved that
SLn (Z) satisfies a version of Poincaré–Lefschetz duality called virtual Bieri–Eckmann duality.
This duality involves a “dualizing module” that measures the “homology at infinity”. In our
situation, that dualizing module is the Steinberg representation StSLn (Q; Q) and we have
n
H( 2 )−i (SLn (Z); Q) = Hi (SLn (Z); StSLn (Q; Q)).
Conjecture 1.3 is thus equivalent to the following conjecture, which resembles Theorem 1.1
for Gn = SLn .
Conjecture 1.4. We have Hi (SLn (Z); StSLn (Q; Q)) = 0 for n ≥ i + 2.
Remark 1.5. The proofs by Lee–Szczarba [LSz] and Church–Putman [CP] of special cases
of Conjecture 1.3 both start by translating things into the language of Conjecture 1.4.
We now briefly describe our proof of Theorem 1.1. As we will discuss in §3 below, there
is a natural inclusion StGn−1 (k; R) → StGn (k; R). This induces a stabilization map
Hi (Gn−1 (k); StGn−1 (k; R)) → Hi (Gn (k); StGn (k; R)).
(1.1)
We will show in §3 that to prove that Hi (Gn (k); StGn (k; R)) = 0 for large n, it is enough
to prove the seemingly weaker assertion that (1.1) is a surjection for large n. This idea was
first introduced by Church–Farb–Putman [CFaP] as a strategy for proving Conjecture 1.4.
It was also noticed by Ash in unpublished work.
The surjectivity of (1.1) is a weak form of homological stability. There is an enormous
literature on homological stability theorems. The basic technique underlying most results
in the subject goes back to unpublished work of Quillen. In [Dw], Dwyer used these ideas
to prove a twisted homological stability theorem for GLn (k) with quite general coefficient
systems. This work was later generalized by van der Kallen [VdK] and very recently by
Randal-Williams–Wahl [RaWiWa], whose results cover all the classical groups in Theorem
1.1. Unfortunately, the Steinberg representation does not satisfy the conditions in any of
these known theorems. Indeed, these theorems are general enough that if it did, then this
would quickly lead to a proof of Conjecture 1.4. Nevertheless, we are able to use some
delicate properties of the Steinberg representation to jury-rig the Quillen machine such that
it works to prove that (1.1) is surjective for large n.
2
Remark 1.6. Homological stability for a sequence of groups and homomorphisms X1 →
X2 → · · · states that the induced maps Hi (Xn ) → Hi (Xn+1 ) are isomorphisms for n ≫ 0.
L
Alternatively, we can think of each map as “multiplication by t” and give n Hi (Xn ) the
structure of a R[t]-module, where R denotes our coefficient ring. At least when R is a field,
this isomorphism would be a consequence of finite generation.
In our setting, with homology twisted by the Steinberg representation, one should instead
think of this map as “multiplication by t” where t is a generator for the exterior algebra in
one variable R[t]/t2 , so that the groups Hi being 0 for n ≫ 0 would again be a consequence
of finite generation. At least when k is a finite field of size q and R is the field of complex
numbers, this is consistent with the idea that GLn (Fq ) is a q-analogue of the symmetric
group and the Steinberg representation is the q-analogue of its sign representation, which is
made more precise via their connection to symmetric functions, see [Mac, §§I.7, IV.4].
Outline. We begin in §2 with some background and notation. Next, in §3 we reduce Theorem 1.1 to an appropriate homological stability theorem. We then prove a key isomorphism
in §4. We prove Theorem 1.1 in §5. This proof depends on a calculation which we perform
in §6.
Convention regarding the empty set. If X is the empty set and R is a commutative
e (X; R) = R. With this convention, if the semisimple k-rank of G is
ring, then we define H
−1
0, then StG (k; R) = R with the trivial G(k)-action.
Acknowledgments. The second author would like to thank Thomas Church and Benson
Farb for many inspiring conversations concerning Conjectures 1.3 and 1.4.
2
Background and notation
This section contains some background information and notation needed in the remainder
of the paper. It consists of two subsections: §2.1 introduces some distinguished parabolic
subgroups, and §2.2 gives some background about the Steinberg representations.
Throughout this section, k is a field and Gn is either GLn , SLn , Sp2n , SOn,n , or SOn,n+1 .
2.1
Parabolic and stabilizer subgroups
Our proof of Theorem 1.1 depends on a careful study of various subgroups of Gn (k). In
this section, we will introduce notation for these subgroups: a certain parabolic subgroup
PGℓn (k), its unipotent radical UGℓn (k), a Levi component LGℓn (k) of PGℓn (k), and another
subgroup FGℓn (k) that lies in PGℓn (k) and fixes certain vectors.
General and special linear groups. Assume first that Gn is either GLn or SLn . The
group Gn (k) thus acts on the vector space kn , and the k-parabolic subgroups of Gn (k) are
the stabilizers of flags of subspaces of kn . Let (~a1 , . . . ,~an ) be the standard basis for kn . For
3
1 ≤ ℓ ≤ n, the group PGℓn (k) is defined to be the Gn (k)-stabilizer of the flag
0 ( h~a1 , . . . ,~aℓ i.
The group UGℓn (k) is the subgroup of PGℓn (k) consisting of all M ∈ PGℓn (k) that act as the
identity on both
h~a1 , . . . ,~aℓ i
and
kn /h~a1 , . . . ,~aℓ i.
The group LGℓn (k) is defined to be the PGℓn (k)-stabilizer of the flag
0 ( h~aℓ+1 , . . . ,~an i.
If Gn = GLn then LGℓn (k) is the subgroup GLℓ (k) × GLn−ℓ (k) of Gn , while if Gn = SLn
then LGℓn (k) is the subgroup of GLℓ (k) × GLn−ℓ (k) consisting of matrices of determinant 1.
Finally, define
FGℓn (k) = {M ∈ Gn (k) | M(~aj ) = ~aj for 1 ≤ j ≤ ℓ}.
We thus have FGℓn (k) ⊂ PGℓn (k).
Symplectic groups. Now assume that Gn = Sp2n . Letting ω(·, ·) be the standard symplectic form on k2n , the group Gn (k) is the subgroup of GLn (k) consisting of elements
that preserve ω(·, ·). The k-parabolic subgroups of Gn (k) are the Gn (k)-stabilizers of flags
of isotropic subspaces of k2n , that is, subspaces on which ω(·, ·) vanishes identically. Let
(~a1 , . . . ,~an , ~b1 , . . . , ~bn ) be the standard symplectic basis for k2n , so
ω(~aj ,~aj ′ ) = ω(~bj , ~bj ′ ) = 0 and ω(~aj , ~bj ′ ) = δjj ′
for 1 ≤ j, j ′ ≤ n, where δjj ′ is the Kronecker delta function. For 1 ≤ ℓ ≤ n, the group
PGℓn (k) is defined to be the Gn (k)-stabilizer of the isotropic flag
0 ( h~a1 , . . . ,~aℓ i.
The group UGℓn (k) is the subgroup of PGℓn (k) consisting of all M ∈ PGℓn (k) that act as the
identity on both
h~a1 , . . . ,~aℓ i
and
h~a1 , . . . ,~aℓ i⊥ /h~a1 , . . . ,~aℓ i
= h~a1 , . . . ,~aℓ ,~aℓ+1 , ~bℓ+1 , . . . ,~an , ~bn i/h~a1 , . . . ,~aℓ i.
The group LGℓn (k) is defined to be the PGℓn (k)-stabilizer of the isotropic flag
0 ( h~b1 , . . . , ~bℓ i.
The group LGℓn (k) is thus isomorphic to GLℓ (k) × Gn−ℓ (k). Finally, define
FGℓn (k) = {M ∈ Gn (k) | M(~aj ) = ~aj for 1 ≤ j ≤ ℓ}.
We thus have FGℓn (k) ⊂ PGℓn (k).
4
Orthogonal groups. Finally, assume that Gn is either SOn,n or SOn,n+1 . For an appropriate m, the group Gn (k) is then the subgroup of SLm (k) consisting of elements that preserve
a quadratic form q(·) on km :
• If Gn = SOn,n , then let m = 2n and let (~a1 , . . . ,~an , ~b1 , . . . , ~bn ) be the standard basis
for km . The group Gn (k) is the SLm (k)-stabilizer of the quadratic form q(·) on km
defined via the formula
q
n
X
cj~aj
j=1
+ dj~bj
=
n
X
(cj , dj ∈ k).
cj d j
j=1
• If Gn = SOn,n+1, then let m = 2n + 1 and let (~a1 , . . . ,~an , ~b1 , . . . , ~bn , ~e) be the standard
basis for km . The group Gn (k) is the SLm (k)-stabilizer of the quadratic form q(·) on
km defined via the formula
q λ~e +
n
X
j=1
cj~aj + dj~bj = λ2 +
n
X
cj d j
(cj , dj , λ ∈ k).
j=1
In both cases, the k-parabolic subgroups of Gn (k) are the Gn (k)-stabilizers of flags of
isotropic subspaces of km , that is, subspaces on which q(·) vanishes identically. For 1 ≤ ℓ ≤ n
the group PGℓn (k) is defined to be the Gn (k)-stabilizer of the isotropic flag
0 ( h~a1 , . . . ,~aℓ i.
For Gn = SOn,n , the group UGℓn (k) is the subgroup of PGℓn (k) consisting of all M ∈ PGℓn (k)
that act as the identity on both
h~a1 , . . . ,~aℓ i
and
h~a1 , . . . ,~aℓ i⊥ /h~a1 , . . . ,~aℓ i
= h~a1 , . . . ,~aℓ ,~aℓ+1 , ~bℓ+1 , . . . ,~an , ~bn i/h~a1 , . . . ,~aℓ i,
while if Gn = SOn,n+1, then the group UGℓn (k) is the subgroup of PGℓn (k) consisting of all
M ∈ PGℓn (k) that act as the identity on both
h~a1 , . . . ,~aℓ i
and
h~a1 , . . . ,~aℓ i⊥ /h~a1 , . . . ,~aℓ i
= h~a1 , . . . ,~aℓ ,~aℓ+1 , ~bℓ+1 , . . . ,~an , ~bn , ~ei/h~a1 , . . . ,~aℓ i.
The group LGℓn (k) is defined to be the PGℓn (k)-stabilizer of the isotropic flag
0 ( h~b1 , . . . , ~bℓ i.
The group LGℓn (k) is thus isomorphic to GLℓ (k) × Gn−ℓ (k). Finally, define
FGℓn (k) = {M ∈ Gn (k) | M(~aj ) = ~aj for 1 ≤ j ≤ ℓ}.
We thus have FGℓn (k) ⊂ PGℓn (k).
5
2.2
Facts about the Steinberg representation
Let R be a commutative ring. The following theorem of Reeder [Re] will play an important
role in our proof of Theorem 1.1.
Theorem 2.1 ([Re, Proposition 1.1]). Let G be a connected reductive group defined over a
field k, let PG(k) be a k-parabolic subgroup of G(k), and let LG(k) be a Levi component of
PG(k). Then there exists an LG(k)-equivariant map
G(k)
StLG (k; R) −→ ResLG(k) StG (k; R)
such that the induced map
PG(k)
G(k)
IndLG(k) StLG (k; R) → ResPG(k) StG (k; R)
is an isomorphism.
Remark 2.2. The map in Theorem 2.1 is not unique; for instance, it can be post-composed
with any element of the unipotent radical of PG(k). The paper [Re] contains a specific
construction of this map, and whenever we refer to the map in Theorem 2.1 we mean the
one constructed in [Re].
We wish to apply this to the distinguished parabolic subgroups PGℓn (k) that we introduced in §2.1. To do this, we need to identify StLGℓn (k; R).
Lemma 2.3. Let Gn be either GLn , SLn , Sp2n , SOn,n , or SOn,n+1 . Then for all fields k and
all commutative rings R, we have
StLGℓn (k; R) = StGLℓ (k; R) ⊗ StGn−ℓ (k; R)
for 1 ≤ ℓ ≤ n.
Proof. For Gn 6= SLn , this follows from the decomposition LGℓn (k) = GLℓ (k) × Gn−ℓ (k). For
Gn = SLn , we instead have that LGℓn (k) is the subgroup of GLℓ (k) × GLn−ℓ (k) consisting of
matrices of determinant 1. The lemma in this case follows from two facts:
• There is a bijection between k-parabolic subgroups of GLn−ℓ (k) and SLn−ℓ (k), and thus
an SLn−ℓ (k)-equivariant isomorphism between StGLn−ℓ (k; R) and StSLn−ℓ (k; R).
• There is a bijection between k-parabolic subgroups of GLℓ (k) × GLn−ℓ (k) and LGℓn (k),
and thus an LGℓn (k)-equivariant isomorphism between StGLℓ (k; R) ⊗ StGLn−ℓ (k; R) and
StLGℓn (k; R).
Both of these bijections come from taking intersections.
These two results allow us to make the following definition.
Definition 2.4. Let Gn be either GLn , SLn , Sp2n , SOn,n , or SOn,n+1 . Also, let k be a field
and R be a commutative ring. For 1 ≤ ℓ ≤ n, the Reeder product map is the map
StGLℓ (k; R) ⊗ StGn−ℓ (k; R) −→ StGn (k; R)
obtained by combining Lemma 2.3 and Theorem 2.1.
6
Remark 2.5. Identifying StGLℓ (k; R) ⊗ StGn−ℓ (k; R) with its image in StGn (k; R) under the
Reeder product map, one way of viewing Theorem 2.1 is that it asserts that
StGn (k; R) =
M
u · (StGLℓ (k; R) ⊗ StGn−ℓ (k; R)).
u∈UGℓn (k)
We will also need the following lemma, which is precisely the case i = 0 of Theorem 1.1.
It generalizes [LSz, Theorem 4.1]. Recall that if G is a group and M is a G-module, then the
coinvariants MG are the largest quotient of M on which G acts trivially. The coinvariants
MG are isomorphic to H0 (G; M).
Lemma 2.6. Let Gn be either GLn , SLn , Sp2n , SOn,n , or SOn,n+1 . Then for all fields k and
all commutative rings R, we have (StGn (k; R))Gn (k;R) = 0 for n ≥ 2.
Proof. Theorem 2.1 implies that
Gn (k)
PG2 (k)
IndLG2n(k) StGL2 (k; R) ⊗ StGn−2 (k; R) ∼
= ResPG
2 (k) StGn (k; R).
n
n
It is thus enough to prove that
(StGL2 (k; R) ⊗ StGn−2 (k; R))LG2n (k) = 0.
Whatever Gn is, the group LG2n (k) contains the subgroup SL2 (k) × 1. It is thus enough to
prove that
(StGL2 (k; R))SL2 (k) = 0.
This is an easy exercise using the fact that
e (T (k); R) = H
e (P1 (k); Q),
StGL2 (k; R) = H
0
2
0
where P1 (k) is the projective line over k, regarded as a discrete set of points. For details, see
[LSz, Theorem 4.1].
3
Reduction to stability
Let Gn be either GLn , SLn , Sp2n , SOn,n , or SOn,n+1. Let k be a field and R be a commutative
ring. In this section, we reduce Theorem 1.1 to an appropriate homological stability theorem.
Fix some i ≥ 0 and some n ≥ 2. The stabilization map for Hi (Gn−1 (k); StGn−1 (k; R)) is
the map
(3.1)
Hi (Gn−1 (k); StGn−1 (k; R)) → Hi (Gn (k); StGn (k; R))
induced by the following two maps:
• The group homomorphism Gn−1 (k) → Gn (k) obtained as follows. The group LG1n (k)
is a subgroup of GL1 (k) × Gn−1 (k) that contains the subgroup 1 × Gn−1 (k). In fact,
LG1n (k) = GL1 (k) × Gn−1 (k) except when Gn = SLn . We can thus define a homomorphism Gn−1 (k) → Gn (k) via the composition
Gn−1 (k) = 1 × Gn−1 (k) ֒→ LG1n (k) ֒→ Gn (k).
7
• The map StGn−1 (k; R) → StGn (k; R) that equals the composition
StGn−1 (k; R) ∼
= R ⊗ StGn−1 (k; R) ∼
= StGL1 (k; R) ⊗ StGn−1 (k; R) → StGn (k; R),
where the final arrow is the Reeder product map. Here we are using the convention
regarding the empty set discussed at the end of the introduction which implies that
StGL1 (k; R) = R; this convention is compatible with Theorem 2.1.
The main result of this section is then as follows.
Lemma 3.1. Let Gn be either GLn , SLn , Sp2n , SOn,n , or SOn,n+1 . Let k be a field and let
R be a commutative ring. Assume that the stabilization map (3.1) is a surjection for n ≥ N.
Then Hi (Gn (k); StGn (k; R)) = 0 for n ≥ N + 1.
Proof. Consider n ≥ N + 1. By assumption, the map
Hi (Gn−2 (k); StGn−2 (k; R)) → Hi (Gn (k); StGn (k; R))
(3.2)
obtained by iterating the stabilization map twice is surjective. It is thus enough to show
that the image of this map is 0. We can factor this map as
Hi (Gn−2 (k); StGn−2 (k; R)) → Hi (Gn−2 (k); StGL2 (k; R) ⊗ StGn−2 (k; R))
→ Hi (Gn (k); StGn (k; R)).
Regard SL2 (k) as a subgroup of Gn (k) via the composition
SL2 (k) = SL2 (k) × 1 ֒→ LG2n (k) ֒→ Gn (k).
The subgroup SL2 (k) of Gn (k) commutes with the image of Gn−2 (k) in Gn (k) under the
map used to define (3.2). Inner automorphisms act trivially on homology, even with twisted
coefficients; see [Br2, Proposition III.8.1]. It follows that to show that the image of (3.2) is
0, it is enough to prove that
(StGL2 (k; R) ⊗ StGn−2 (k; R))SL2 (k) = 0.
This is equivalent to
(StGL2 (k; R))SL2 (k) = 0,
which is one case of Lemma 2.6.
4
The stabilizer subgroups
This section constructs an isomorphism (Lemma 4.1 below) that will play a fundamental
role in our proof of Theorem 1.1.
Let Gn be either GLn , SLn , Sp2n , SOn,n , or SOn,n+1. Let k be a field and R be a
commutative ring. Fix some 1 ≤ ℓ ≤ n. There is a map
G (k)
Hi (1 × Gn−ℓ (k); StGLℓ (k; R) ⊗ StGn−ℓ (k; R)) → Hi (FGℓn (k); ResFGn ℓ (k) StGn (k; R)).
n
induced by the following two maps:
8
(4.1)
• The inclusion map 1 × Gn−ℓ (k) → FGℓn (k). Here we are regarding FGℓn (k) as a
subgroup of PGℓn (k) that contains 1 × Gn−ℓ (k) ⊂ LGℓn (k) ⊂ PGℓn (k).
• The Reeder product map StGLℓ (k; R) ⊗ StGn−ℓ (k; R) → StGn (k; R).
Our main result is as follows.
Lemma 4.1. Let k be a field, let R be a commutative ring, let 1 ≤ ℓ ≤ n, and let i ≥ 0.
Then the map (4.1) is an isomorphism.
Proof. Shapiro’s Lemma [Br2, Proposition III.6.2] gives an isomorphism
Hi (1 × Gn−ℓ (k); StGLℓ (k; R) ⊗ StGn−ℓ (k; R))
FGℓ (k)
∼
= Hi (FGℓn (k); Ind1×Gn n−ℓ (k) StGLℓ (k; R) ⊗ StGn−ℓ (k; R)).
Below we will prove that there is an isomorphism
FGℓ (k)
G (k)
Ind1×Gnn−ℓ (k) StGLℓ (k; R) ⊗ StGn−ℓ (k; R) ∼
= ResFGn ℓ (k) StGn (k; R).
n
(4.2)
of FGℓn (k)-representations. Combined with the above, this will yield an isomorphism between
the left and right hand sides of (4.1) which is easily seen to be the map in (4.1).
It remains to construct the isomorphism (4.2). Since FGℓn (k) ⊂ PGℓn (k), we can restrict
the isomorphism given by Theorem 2.1 and Lemma 2.3 to obtain an isomorphism
G (k)
PGℓ (k)
PGℓ (k)
ResFGℓn(k) IndLGℓn(k) StGLℓ (k; R) ⊗ StGn−ℓ (k; R) ∼
= ResFGn ℓ (k) StGn (k; R).
n
n
n
(4.3)
The unipotent radical of PGℓn (k) is contained in FGℓn (k). This implies that there is a
single (FGℓn (k), LGℓn (k))-double coset in PGℓn (k). Also, FGℓn (k) ∩ LGℓn (k) = Gn−ℓ (k). The
double coset formula [Br2, Proposition III.5.6b] therefore implies that the left side of (4.3)
is canonically isomorphic to
FGℓ (k)
LGℓ (k)
FGℓ (k)
Ind1×Gn n−ℓ (k) Res1×Gnn−ℓ (k) StGLℓ (k; R)⊗StGn−ℓ (k; R) = Ind1×Gnn−ℓ (k) StGLℓ (k; R)⊗StGn−ℓ (k; R),
as desired.
The following alternate version of Lemma 4.1 will be useful.
Corollary 4.2. Let k be a field, let R be a commutative ring, let 1 ≤ ℓ ≤ n, and let i ≥ 0.
Then there exists an isomorphism
G (k)
StGLℓ (k; R) ⊗ Hi (Gn−ℓ (k); StGn−ℓ (k; R)) ∼
= Hi (FGℓn (k); ResFGn ℓ (k) StGn (k; R)).
n
Proof. Since StGLℓ (k; R) is a free R-module, we have
StGLℓ (k; R) ⊗ Hi (Gn−ℓ (k); StGn−ℓ (k; R)) ∼
= Hi (1 × Gn−ℓ (k); StGLℓ (k; R) ⊗ StGn−ℓ (k; R)).
The corollary now follows from Lemma 4.1.
We will also need an explicit inverse
G (k)
Hi (FGℓn (k); ResFGn ℓ (k) StGn (k; R)) → Hi (1 × Gn−ℓ (k); StGLℓ (k; R) ⊗ StGn−ℓ (k; R))
n
to the isomorphism (4.1). The map (4.4) will be induced by the following two maps:
9
(4.4)
• The homomorphism FGℓn (k) → 1 × Gn−ℓ (k) obtained by restricting the projection
PGℓn (k) → LGℓn (k) to FGℓn (k).
• The map
StGn (k; R) → StGLℓ (k; R) ⊗ StGn−ℓ (k; R)
which equals the composition
StGn (k; R) =
M
u · StGLℓ (k; R) ⊗ StGn−ℓ (k; R)
u∈UGℓn (k)
→ StGLℓ (k; R) ⊗ StGn−ℓ (k; R),
where the first equality comes from a combination of Theorem 2.1 and Lemma 2.3
(see Remark 2.5) and the last arrow takes u · x ∈ u · (StGLℓ (k; R) ⊗ StGn−ℓ (k; R)) to
x ∈ StGLℓ (k; R) ⊗ StGn−ℓ (k; R). We will call this map the Reeder projection map.
It is clear that these define a map of the form (4.4). The following lemma says that this is
an inverse to (4.1).
Lemma 4.3. Let k be a field, let R be a commutative ring, let 1 ≤ ℓ ≤ n, and let i ≥ 0.
Then the map (4.4) is an inverse to the map (4.1).
Proof. Immediate from the fact that the compositions
1 × Gn−ℓ (k) → FGℓn (k) → 1 × Gn−ℓ (k)
and
StGLℓ (k; R) ⊗ StGn−ℓ (k; R) → StGn (k; R) → StGLℓ (k; R) ⊗ StGn−ℓ (k; R)
of the maps used to define (4.1) and (4.4) equal the identity.
5
Vanishing
This section is devoted to the proof of Theorem 1.1. The actual proof is in §5.3. This is
preceded by two sections of preliminaries.
5.1
Equivariant homology
In our proof of Theorem 1.1, we will need some basic facts about equivariant homology. A
basic reference is [Br2, Chapter VII.7].
Let G be a group, let X be a semisimplicial set on which G acts, let R be a ring, and
let M be an R[G]-module. Let EG be a contractible semisimplicial set on which G acts
freely and let BG = EG/G, so BG is a classifying space for G. Denote by EG ×G X the
quotient of EG × X by the diagonal action of G. This is known as the Borel construction.
The homotopy type of EG ×G X does not depend on the choice of EG. The projection
EG ×G X → EG/G = BG induces a homomorphism π1 (EG ×G X) → π1 (BG) = G. Via
this homomorphism, we can regard M as a local coefficient system on EG ×G X. The
G-equivariant homology groups of X with coefficients in M, denoted HG
∗ (X; M), are the
homology groups of EG ×G X with respect to the local coefficient system M.
10
Lemma 5.1. If X is ℓ-connected, then the above map EG ×G X → EG/G = BG induces
G
∼
an isomorphism HG
i (X; M) = Hi (G; M) for 0 ≤ i ≤ ℓ and a surjection Hℓ+1 (X; M) →
Hℓ+1 (X; M).
Proof. The group G acts freely on EG × X and EG × X is ℓ-connected. Viewing EG × X
as a CW-complex, we can make EG × X contractible by adding cells of dimension at least
(ℓ + 2). We conclude that there exists a classifying space for G whose (ℓ + 1)-skeleton equals
the (ℓ + 1)-skeleton of EG ×G X. The lemma follows.
Our main tool for understanding HG
∗ (X; M) is the following spectral sequence, which is
constructed in [Br2, Equation VII.7.7].
Lemma 5.2. For all p ≥ 0, let Σp be a set containing exactly one representative for each
orbit of the action of G on the p-simplices of X. For σ ∈ Σp , let Gσ be the stabilizer of σ.
Then there is a first quadrant spectral sequence
1
Ep,q
=
M
G
Hq (Gσ ; ResG
Gσ M) =⇒ Hp+q (X; M).
σ∈Σp
Remark 5.3. In [Br2, Equation VII.7.7], the action of Gσ on M is twisted by an “orientation
character”; however, this is unnecessary in our situation, since we are working with semisimplicial sets rather than ordinary simplicial complexes (the point being that in the geometric
realization, the setwise stabilizer of a simplex stabilizes the simplex pointwise).
5.2
Complexes of partial bases
Let k be a field and let Gn be either GLn , SLn , Sp2n , SOn,n , or SOn,n+1 . To prove Theorem
1.1, we will need to construct a highly connected space CGn (k) on which Gn (k) acts. The
definition of this complex is as follows.
• If Gn = GLn or Gn = SLn , then define CGn (k) to be the complex of partial bases for
kn , i.e. the semisimplicial complex whose ℓ-simplices are ordered sequences [~v0 , . . . , ~vℓ ]
of linearly independent elements of kn .
• If Gn = Sp2n or Gn = SOn,n or Gn = SOn,n+1 and km is the vector space upon which
Gn (k) acts (so m is either 2n or 2n + 1), then define CGn (k) to be the complex of
partial isotropic bases for km , i.e. the semisimplicial complex whose ℓ-simplices are
ordered sequences [~v0 , . . . , ~vℓ ] of linearly independent elements of km that span an
isotropic subspace.
The following theorem summarizes the properties of CGn (k).
Theorem 5.4. Let k be a field and let Gn be either GLn , SLn , Sp2n , SOn,n , or SOn,n+1 . The
following then hold.
1. The group Gn (k) acts transitively on the ℓ-cells of CGn (k) for all 0 ≤ ℓ < n − 1.
2. The space CGn (k) is f (n)-connected where f (n) is given by:
(a) f (n) = n − 2 if Gn is either GLn or SLn ,
(b) f (n) =
n−3
2
if Gn is either Sp2n , SOn,n , or SOn,n+1.
11
Proof. The first assertion is well known (and also holds for ℓ = n−1 except when Gn = SLn ).
As for the second, Maazen proved in his thesis [Maa] that CGn (k) is (n − 2)-connected for
Gn = GLn and Gn = SLn . See [VdK] for a published proof of a more general result. Friedrich
proved in [Fr, Theorem 3.23] that CGn (k) is n−3
-connected for Gn = Sp2n and Gn = SOn,n
2
and Gn = SOn,n+1 (for Sp2n and SOn,n , this was proven earlier in [MiVdK, Theorem 7.3]).
To apply the cited result of Friedrich to our situation, we need the fact that the unitary
stable rank of a field is 1 (see, e.g., [MiVdK, Example 6.5]).
5.3
The proof of Theorem 1.1
Let us first recall the statement of the theorem. Let Gn be either GLn , SLn , Sp2n , SOn,n ,
or SOn,n+1 . Also, let k be a field and R be a commutative ring. Our goal is to prove that
Hi (Gn (k); StGn (k; R)) = 0 for n ≥ 2i + 2 and that there exists a surjection
Hi (G2i (k); StG2i (k; R)) → Hi (G2i+1 (k); StG2i+1 (k; R)).
(5.1)
Of course, this surjection will be induced by the stabilization map defined in §3.
The proof is by induction on i. We begin with the base case i = 0. Lemma 2.6 says
that H0 (Gn (k); StGn (k; R)) = 0 for n ≥ 2, so we only need to show that the map (5.1) is a
surjection for i = 0. For the domain, G0 (k) is the trivial group. By our convention regarding
the empty set discussed at the end of the introduction, we thus have StG0 (k; R) = R, and
hence Hi (G0 (k); StG0 (k; R)) = R. To simplify the codomain, we have several cases.
• G1 = GL1 or G1 = SO1,1 . In fact, these groups are isomorphic and are commutative,
so StG1 (k; R) = R in these cases and (5.1) is an isomorphism.
• G1 = SL1 . The group SL1 is the trivial group and thus StG1 (k; R) = R and (5.1) is an
isomorphism.
• G1 = Sp2 ∼
= SL2 or G1 = SO2,1 ∼
= PSL2 . These groups have isomorphic Steinberg
representations and the action of SL2 (k) on StSL2 (k; R) factors through PSL2 (k). This
case thus follows from Lemma 2.6, which says that H0 (SL2 (k); StSL2 (k; R)) = 0.
This completes the base case.
Assume now that i > 0 and that the desired result is true for all smaller values of i. We
will prove that the stabilization map
Hi (Gn−1 (k); StGn−1 (k; R)) → Hi (Gn (k); StGn (k; R))
(5.2)
is surjective for n ≥ 2i + 1. Lemma 3.1 will then imply that Hi (Gn (k); StGn (k; R)) = 0 for
n ≥ 2i + 2, and the theorem will follow.
Fix some n ≥ 2i + 1 and let km be the standard vector space representation of Gn (k) (so
m is either n, 2n, or 2n + 1). Let {~a1 , . . . ,~an } be the vectors in km such that
FGℓn (k) = {M ∈ Gn (k) | M(~aj ) = ~aj for 1 ≤ j ≤ ℓ}
for 1 ≤ ℓ ≤ n. Combining the second conclusion of Theorem 5.4 with Lemma 5.1, we have
a surjection
G (k)
(5.3)
Hi n (CGn (k); StGn (k; R)) → Hi (Gn (k); StGn (k; R)).
G (k)
We will analyze Hi n (CGn (k); StGn (k; R)) using the spectral sequence from Lemma 5.2.
To calculate its E 1 -page, observe that the first conclusion of Theorem 5.4 says that Gn (k)
12
acts transitively on the p-simplices of CGn (k) for 0 ≤ p < n − 1. The stabilizer of the
(p − 1)-simplex [~a1 , . . . ,~ap ] is FGpn (k), so the spectral sequence in Lemma 5.2 has
G (k)
n
1
Ep,q
= Hq (FGp+1
n (k); ResFGp+1 (k) StGn (k; R)),
n
(5.4)
for 0 ≤ p < n − 1.
We will prove that all of the terms on the p + q = i line of the E ∞ -page of our spectral
∞
sequence vanish except for possibly the term E0,i
. To do this, consider p, q ≥ 0 with p + q = i
and p ≥ 1. The case p = 1 and n = 2i + 1 is exceptional and must be treated separately.
To avoid getting bogged down here, we postpone this calculation until §6 below, where it
appears as Lemma 6.1.1
We thus can assume that either p ≥ 2 or that n ≥ 2i + 2. Since n ≥ 2i + 1, we certainly
1
have p < n − 1, so Ep,q
is in the regime where the above description of the E 1 -page holds.
Applying Corollary 4.2 to (5.4), we see that
1
Ep,q
= StGLp+1 (k; R) ⊗ Hq (Gn−p−1(k); StGn−p−1 (k; R)).
To see that this vanishes, it is enough to show that Hq (Gn−p−1(k); StGn−p−1 (k; R)) = 0. This
is a consequence of our inductive hypothesis; to see that it applies, observe that if p ≥ 2
then
n − p − 1 ≥ (2i + 1) − p − 1 = 2(p + q) − p = 2q + p ≥ 2q + 2,
while if p = 1 and n ≥ 2i + 2 then
n − p − 1 ≥ (2i + 2) − 1 − 1 = 2(p + q) = 2q + 2.
1
∞
This implies that Ep,q
= 0, and thus that Ep,q
= 0.
The p + q = i line of the E ∞ -page of our spectral sequence thus only has a single
∞
potentially nonzero entry, namely E0,i
, and this is a quotient of
G (k)
1
E0,i
= Hi (FG1n (k); ResFGn 1 (k) StGn (k; R)).
n
G (k)
This entry thus surjects onto Hi n
tion (5.3), we obtain a surjection
(CGn (k); StGn (k; R)). Combining this with the surjec-
G (k)
Hi (FG1n (k); ResFGn 1 (k) StGn (k; R)) −→ Hi (Gn (k); StGn (k; R)).
n
(5.5)
Examining the construction of our spectral sequence in [Br2, Chapter VII.7], it is easy to
see that this comes from the map induced by the inclusion FG1n (k) ֒→ Gn (k). Combining
(5.5) with the isomorphism
∼
G (k)
=
Hi (Gn−1 (k); StGn−1 (k; R)) −
→
Hi (FG1n (k); ResFGn 1 (k) StGn (k; R))
n
given by the ℓ = 1 case of Corollary 4.2, we conclude that (5.2) is a surjection, as desired.
1
This exceptional case could be avoided at the cost of only proving that Hi (Gn ; StGn (k; R)) = 0 for
n ≥ 3i + 2 instead of for n ≥ 2i + 2.
13
6
Killing the exceptional term in the spectral sequence
This section is devoted to proving the vanishing result postponed from the proof of Theorem
1.1 in §5.3. The notation in this section is thus identical to that in §5.3:
• Gn is either GLn , SLn , Sp2n , SOn,n , or SOn,n+1 .
• k is a field and R is a commutative ring.
• i > 0 and n = 2i + 1 (the only case that remained in that section).
• km is the standard vector space representation of Gn (k) (so m is either n, 2n, or 2n+1).
• {~a1 , . . . ,~an } is the set of vectors in km such that
FGℓn (k) = {M ∈ Gn (k) | M(~aj ) = ~aj for 1 ≤ j ≤ ℓ}
for 1 ≤ ℓ ≤ n.
G (k)
r
• Ep,q
is the spectral sequence from Lemma 5.2 converging to Hi n (CGn (k); StGn (k; R)).
What we must prove is as follows.
Lemma 6.1. Let the notation be as above, and assume that the stabilization map
Hi−1 (Gn−3 (k); StGn−3 (k; R)) → Hi−1 (Gn−2 (k); StGn−2 (k; R))
(6.1)
1
1
∞
is surjective. Then the differential E2,i−1
→ E1,i−1
is surjective, and thus E1,i−1
= 0.
The proof of Lemma 6.1 is divided into five sections:
1
1
• In §6.1, we give an explicit form for the differential E2,i−1
→ E1,i−1
.
• In §6.2, we translate that explicit form into one involving the stabilization map (6.1).
• In §6.3, we summarize what remains to be proved.
• In §6.4, we give some needed background information about apartments.
• In §6.5, we finish off the proof of Lemma 6.1.
6.1
Identifying the differential
1
The notation is as in the beginning of §6. In this section, we identify the differential E2,i−1
→
1
1
E1,i−1 . Since n = 2i + 1 and i > 0, we have 1 < n − 1, so E1,i−1 is as described in (5.4), i.e.
G (k)
1
∼
E1,i−1
= Hi−1 (FG2n (k); ResFGn 2n (k) StGn (k; R)).
1
If i = 1, then we do not have 2 < n − 1, so in this case E2,i−1
is not as described in (5.4). The
issue is that Gn (k) might not act transitively on the 2-simplices of CGn (k) (this is actually
1
only a problem for Gn = SLn ). However, for all values of i it is still the case that E2,i−1
contains
G (k)
Hi−1 (FG3n (k); ResFGn 3 (k) StGn (k; R))
n
1
1
as a summand. The restriction of the differential E2,i−1
→ E1,i−1
to this summand is a map
G (k)
G (k)
∂ : Hi−1 (FG3n (k); ResFGn 3 (k) StGn (k; R)) −→ Hi−1 (FG2n (k); ResFGn 2 (k) StGn (k; R)).
n
n
To prove Lemma 6.1, it is enough to prove that ∂ is surjective.
14
(6.2)
We can describe ∂ using the recipe described in [Br2, Chapter VII.8]. Recall that FG3n (k)
is the Gn (k)-stabilizer of the ordered sequence of vectors σ = [~a1 ,~a2 ,~a3 ]. For 1 ≤ m ≤ 3, let
σm be the ordered sequence obtained by deleting ~am from σ and let (Gn (k))σm denote the
Gn (k)-stabilizer of σm . We then have ∂ = ∂1 − ∂2 + ∂3 , where ∂m is the composition
G (k)
Hi−1 (FG3n (k); ResFGn 3 (k) StGn (k; R))
n
′
∂m
G (k)
−→ Hi−1 ((Gn (k))σm ; Res(Gnn (k))σm StGn (k; R))
∂ ′′
G (k)
m
−→
Hi−1 (FG2n (k); ResFGn 2 (k) StGn (k; R))
n
of the following two maps.
′
• ∂m
is the map induced by the inclusion FG3n (k) ֒→ (Gn (k))σm .
• Define κm ∈ Gn (k) as follows. First, κ3 = id. For m ∈ {1, 2}, we do the following.
– If Gn = GLn or Gn = SLn , then κm ∈ SLn (k) is the map kn → kn that takes ~am
to ~a3 , takes ~a3 to −~am , and fixes all the other basis vectors.
– If Gn = Sp2n or Gn = SOn,n or Gn = SOn,n+1, then κm ∈ Gn (k) is the map km →
km defined as follows. Let ~b1 , . . . , ~bn be the standard basis vectors for km that pair
with the ~aj (there is one additional standard basis vector if Gn = SOn,n+1 ). Then
κm takes ~am to ~a3 , takes ~a3 to −~am , takes ~bm to ~b3 , takes ~b3 to −~bm , and fixes all
the other basis vectors.
′′
Then ∂m
is induced by the map (Gn (k))σm → FG2n (k) that takes g ∈ (Gn (k))σm to
κm gκ−1
m and the map StGn (k; R) → StGn (k; R) that takes x ∈ StGn (k; R) to κm (x) ∈
StGn (k; R). We remark that easier choices of κm (without the signs) could be used for
Gn 6= SLn , but we chose the ones above to make our later formulas more uniform.
This is summarized in the following lemma.
Lemma 6.2. Let the notation be as above. Then the map ∂ in (6.2) equals ∂1 − ∂2 + ∂3 ,
where ∂m is induced by the map FG3n (k) → FG2n (k) defined via the formula
(g ∈ FG3n (k))
g 7→ κm gκ−1
m
and the map StGn (k; R) → StGn (k; R) defined via the formula
x 7→ κm (x)
6.2
(x ∈ StGn (k; R)).
Bringing in the stabilization map
The notation is as in the beginning of §6. Fix some 1 ≤ m ≤ 3, and let ∂m and κm be as in
Lemma 6.2. Applying the isomorphism in Corollary 4.2 to the domain and codomain of ∂m ,
we obtain a homomorphism
∂bm : StGL3 (k; R) ⊗ Hi−1 (Gn−3 (k); StGn−3 (k; R))
→ StGL2 (k; R) ⊗ Hi−1 (Gn−2 (k); StGn−2 (k; R)).
Our goal in this section is to prove that ∂bm is the tensor product of the stabilization map
Hi−1 (Gn−3 (k); StGn−3 (k; R)) → Hi−1 (Gn−2 (k); StGn−2 (k; R))
15
(6.3)
with the map ζm : StGL3 (k; R) → StGL2 (k; R) defined as follows. Let {~a1 ,~a2 ,~a3 } be the
b 3 = id ∈ SL3 (k), and for m ∈ {1, 2} let κ
b m ∈ SL3 (k) be the
standard basis for k3 . Let κ
element that takes ~am to ~a3 , takes ~a3 to −~am , and fixes all the other basis vectors. Then ζm
is the composition
b
κm
StGL3 (k; R) −→ StGL2 (k; R) ⊗ StGL1 (k; R) ∼
StGL3 (k; R) −→
= StGL2 (k; R),
where the second arrow is the Reeder projection map (see §4) and the final isomorphism
comes from the fact that StGL1 (k; R) = R.
The main result of this section is then as follows.
Lemma 6.3. Let the notation be as above. Then ∂bm is the tensor product of ζm with the
stabilization map (6.3).
Proof. By construction, ∂bm equals the composition
StGL3 (k; R) ⊗ Hi−1 (Gn−3 (k); StGn−3 (k; R))
∼
=
−→
Hi−1 (1 × Gn−3 (k); StGL3 (k; R) ⊗ StGn−3 (k; R))
∼
=
−→
Hi−1 (FG3n (k); StGn (k; R))
∂
m
−→
Hi−1 (FG2n (k); StGn (k; R))
∼
=
−→
Hi−1 (1 × Gn−2 (k); StGL2 (k; R) ⊗ StGn−2 (k; R))
∼
=
−→
StGL2 (k; R) ⊗ Hi−1 (Gn−2 (k); StGn−2 (k; R))
where the various maps are as follows:
• The first and last arrows use the fact that StGL3 (k; R) and StGL2 (k; R) are free Rmodules (cf. the proof of Corollary 4.2).
• The second arrow is the map described in Lemma 4.1, that is, the map induced by
the inclusion 1 × Gn−3 (k) ֒→ FG3n (k) and the Reeder product map StGL3 (k; R) ⊗
StGn−3 (k; R) → StGn (k; R).
• The third arrow is the map ∂m described in Lemma 6.2, that is, the map induced by
the map FG3n (k) → FG2n (k) given by conjugation by κm and the map StGn (k; R) →
StGn (k; R) induced by κm .
• The fourth arrow is the map described in Lemma 4.3, that is, the map induced by
the projection FG2n (k) → 1 × Gn−2 (k) together with the Reeder projection map
StGn (k; R) → StGL2 (k; R) ⊗ StGn−2 (k; R).
We must show that this composition equals the indicated tensor product of maps. This will
take some work.
Define Ψ to be the composition
StGL3 (k; R) ⊗ StGn−3 (k; R) −→ StGn (k; R)
κ
m
−→
StGn (k; R)
−→ StGL2 (k; R) ⊗ StGn−2 (k; R),
16
where the first map is the Reeder product map and the last map is the Reeder projection
map. Also, define Φ to be the composition
StGL3 (k; R) ⊗ StGn−3 (k; R) −→ StGL2 (k; R) ⊗ StGL1 (k; R) ⊗ StGn−3 (k; R)
−→ StGL2 (k; R) ⊗ StGn−2 (k; R),
where the maps are as follows:
• The first map is the tensor product of the Reeder projection map StGL3 (k; R) →
StGL2 (k; R) ⊗ StGL1 (k; R) and the identity map StGn−3 (k; R) → StGn−3 (k; R).
• The second map is the tensor product of the identity map StGL2 (k; R) → StGL2 (k; R)
and the Reeder product map StGL1 (k; R) ⊗ StGn−3 (k; R) → StGn−2 (k; R).
b m ⊗ id).
By the above, it is enough to prove that Ψ = Φ ◦ (κ
′
Define Ψ to be the composition
StGL3 (k; R) ⊗ StGn−3 (k; R) −→ StGn (k; R)
−→ StGL2 (k; R) ⊗ StGn−2 (k; R),
where the first map is the Reeder product map and the second map is the Reeder projection
b m ⊗ id). We thus see that it is enough to
map. From its definition, we see that Ψ = Ψ′ ◦ (κ
′
prove that Ψ = Φ.
Define U = UGL23 (k) to be the unipotent radical of the parabolic subgroup PGL23 (k)
of GL3 (k) (despite the bad notation, this is not the projective general linear group). Using
Theorem 2.1 as in Remark 2.5, we see that
StGL3 (k; R) =
M
u · (StGL2 (k; R) ⊗ StGL1 (k; R)) .
u∈U
Consider u ∈ U and x ∈ StGL2 (k; R) and y ∈ StGL1 (k; R) and z ∈ StGn−3 (k; R). Examining
the definition of Ψ′ , we see that
Ψ′ ((u · (x ⊗ y)) ⊗ z) = x ⊗ (y ⊗ z) ,
where y ⊗ z ∈ StGL1 (k; R) ⊗ StGn−3 (k; R) is identified with an element of StGn−2 (k; R) using
the Reeder product map. But this equals Φ ((u · (x ⊗ y)) ⊗ z), as desired.
6.3
Summary of where we are
The notation is as in the beginning of §6. Recall that Lemma 6.1 asserts that the differential
1
1
b m be as in §6.2. Define
E2,i−1
→ E1,i−1
is surjective. Let ∂ be as in §6.1. Also, let ζm and κ
ζ : StGL3 (k; R) → StGL2 (k; R)
via the formula ζ = ζ1 − ζ2 + ζ3 . Combining Lemmas 6.2 and 6.3, we see that to prove
Lemma 6.1, it is enough to show that the map
StGL3 (k; R) ⊗ Hi−1 (Gn−3 (k); StGn−3 (k; R)) → StGL2 (k; R) ⊗ Hi−1 (Gn−2 (k); StGn−2 (k; R))
obtained as the tensor product of ζ and the stabilization map
Hi−1 (Gn−3 (k); StGn−3 (k; R)) → Hi−1 (Gn−2 (k); StGn−2 (k; R))
is surjective. One of the assumptions in Lemma 6.1 is that this stabilization map is surjective.
To prove that lemma, it is thus enough to prove the following.
Lemma 6.4. Let the notation be as above. Then ζ is surjective.
17
v0
v2
v3
−
B0
v1
v3
v0
+
B1
v2
v0
v1
−
B2
v3
v1
= v
1
B3
v0
v3
v2
v2
Figure 1: As we illustrate here in the case n = 3, the apartment classes corresponding to the Bi
can be placed on the boundary of an n-dimensional simplex such that their simplices cancel in pairs.
In the picture, the vertices labeled with the vectors ~vi are taken to the lines spanned by the ~vi while
the unlabeled vertices are taken to the 2-dimensional subspaces spanned by the vectors on their two
neighbors.
6.4
Apartments
Before we prove Lemma 6.4, we need to discuss some background material on the Steinberg
representation. Unlike the previous sections, in this section n ≥ 1 is arbitrary. Recall that
e
StGLn (k; R) = H
n−2 (TGLn (k); R), where TGLn (k) is the Tits building associated to GLn (k).
This building can be described as the simplicial complex whose r-simplices are flags
0 ( V0 ( · · · ( Vr ( k n
of nonzero proper subspaces of kn .
The Solomon–Tits theorem [Br1, Theorem IV.5.2] says that the R-module StGLn (k; R)
is spanned by apartment classes, which are defined as follows. Consider an n × n matrix B
with entries in kn none of whose columns are identically 0. Let (~v1 , . . . , ~vn ) be the columns
of B. Let Sn be the simplicial complex whose r-simplices are chains
0 ( I0 ( · · · ( Ir ( {1, . . . , n}.
The complex Sn is isomorphic to the barycentric subdivision of the boundary of an (n − 1)simplex; in particular, Sn is homeomorphic to an (n − 2)-sphere. There is a simplicial map
f : Sn → TGLn (k) defined via the formula
f (I) = h~vi | i ∈ Ii
(∅ ( I ( {1, . . . , n}).
The apartment class corresponding to B, denoted B , is the image of the fundamental class
e
e
e
[Sn ] ∈ H
n−2 (Sn ; R) = R under the map f∗ : Hn−2 (Sn ; R) → Hn−2 (TGLn (k); R) = StGLn (k; R).
Remark 6.5. We have B
invertible.
= 0 if the ~vi do not form a basis for kn , i.e. if B is not
Permuting the columns of B changes B by the sign of the permutation, and multiplying
a column of B by a nonzero scalar does not change B . The apartment classes also satisfy
the following more interesting relation.
18
Lemma 6.6. Let k be a field, let R be a commutative ring, and let n ≥ 2. Let B be an
n × (n + 1)-matrix with entries in k. Assume that none of the columns of B are identically
0. Ordering the columns of B from 0 to n, for 0 ≤ m ≤ n let Bm be the result of deleting
the mth column from B. Then B0 − B1 + B2 − · · · + (−1)n Bn = 0.
Proof. The simplices forming the apartment classes Bi cancel in pairs; see Figure 1.
The Solomon–Tits theorem [Br1, Theorem IV.5.2] gives the following basis for StGLn (k; R).
Theorem 6.7 (Solomon–Tits). Let k be a field, let R be a commutative ring, and let n ≥ 1.
Then StGLn (k; R) is a free R-module on the basis consisting of all B such that B is an
upper unitriangular matrix in GLn (k).
6.5
The proof of Lemma 6.4
We finally prove Lemma 6.4, which as discussed in §6.3 suffices to prove Lemma 6.1. First,
b m be as in §6.2. Define
we recall its statement. For 1 ≤ m ≤ 3, let ζm and κ
ζ : StGL3 (k; R) → StGL2 (k; R)
via the formula ζ = ζ1 − ζ2 + ζ3 . Our goal is to prove that ζ is surjective.
Before we do that, we introduce some formulas. Let π : StGL3 (k; R) → StGL2 (k; R) be
the composition
∼
=
StGL2 (k; R),
StGL3 (k; R) −→ StGL2 (k; R) ⊗ StGL1 (k; R) −→
where the first arrow is the Reeder projection map and the second arrow comes from the
fact that StGL1 (k; R) = R. From its definition, we see that
1 x y
1 x
π( 0 1 z ) =
0 1
0 0 1
for all x, y, z ∈ k. What is more, for all 3×3 matrices B none of whose columns are identically
0 we have
b 1 (B)) − π(κ
b 2 (B)) + π(κ
b 3 (B)).
ζ( B ) = π(κ
b m act on B via matrix multiplication.
Here the κ
We now turn to proving that ζ is surjective. Consider a ∈ k, and set
Aa =
1 a
.
0 1
19
By Theorem 6.7, it is enough to prove that Aa ∈ Im(ζ). We have
1 a 0
1 a 0
0 0 −1
1 a 0
ζ( 0 1 0 ) = π( 0 1 0 ) − π( 0 0 −1 ) + π( 0 1 0 )
0 0 1
0 1 0
1 a 0
0 0 1
−1 0 0
1 0 a
1 a 0
= π(− 0 1 0 ) − π(− 0 −1 0 ) + π( 0 1 0 )
0 a 1
0 0 1
0 0 1
1 a 0
1 0 a
1 0 0
= π(− 0 1 0 ) − π(− 0 1 0 ) + π( 0 1 0 )
0 0 1
0 0 1
0 a 1
(6.4)
1 0 0
1 a
1 0
,
+
= −π( 0 1 0 ) +
0 1
0 1
0 a 1
where the second equality uses the fact that permuting the columns of a matrix changes the
associated apartment by the sign of the permutation and the third equality uses the fact
that multiplying a column by a nonzero scalar does not change the associated apartment.
If a = 0, then the right hand side of (6.4) simplifies to
−
1 0
1 0
1 0
1 0
= A0 ,
=
+
+
0 1
0 1
0 1
0 1
so A0 ∈ Im(ζ). Assume now that a 6= 0. Plugging the matrix
1 0 0 0
0 1 1 0
0 0 a 1
into Lemma 6.6, we get the relation
1 0 0
1 0 0
1 0 0
0 0 0
0= 1 1 0 − 0 1 0 + 0 1 0 − 0 1 1
0 0 a
0 0 1
0 a 1
0 a 1
1 0 0
1 0 0
1 0 0
= 0 − 0 1 0 + 0 1 0 − 0 1 a−1 ,
0 a 1
0 0 1
0 0 1
where the equality uses the fact that the columns of the first matrix are not linearly independent and the fact that multiplying a column of a matrix by a nonzero scalar does not change
the associated apartment. Plugging this relation into (6.4), we see that the right hand side
of (6.4) equals
1 0 0
1 0 0
1 a
1 0
+
− π( 0 1 0 ) − π( 0 1 a−1 )
+
0 1
0 1
0 0 1
0 0 1
=−
1 0
1 0
1 0
1 a
+
+
+
0 1
0 1
0 1
0 1
=A0 + Aa .
20
Since we have already seen that A0 ∈ Im(ζ), we deduce that Aa ∈ Im(ζ), as desired.
References
[BoSe]
[Br1]
[Br2]
[CFaP]
[CP]
[Dw]
[Fr]
[H]
[LSz]
[Maa]
[Mac]
[MiVdK]
[MalZ]
[Q]
[RaWiWa]
[Re]
[VdK]
A. Borel and J.-P. Serre, Corners and arithmetic groups, Comment. Math. Helv. 48 (1973),
436–491.
K. S. Brown, Buildings, Springer, New York, 1989.
K. S. Brown, Cohomology of groups, corrected reprint of the 1982 original, Graduate Texts in
Mathematics, 87, Springer, New York, 1994.
T. Church, B. Farb and A. Putman, A stability conjecture for the unstable cohomology of
SLn Z, mapping class groups, and Aut(Fn ), in Algebraic topology: applications and new
directions, 55–70, Contemp. Math., 620, Amer. Math. Soc., Providence, RI,
arXiv:1208.3216v3.
T. Church and A. Putman, The codimension-one cohomology of SLn Z, Geom. Topol., to
appear, arXiv:1507.06306v3.
W. G. Dwyer, Twisted homological stability for general linear groups, Ann. of Math. (2) 111
(1980), no. 2, 239–251.
N. Friedrich, Homological Stability of automorphism groups of quadratic modules and
manifolds, preprint 2016, arXiv:1612.04584v2.
J. E. Humphreys, The Steinberg representation, Bull. Amer. Math. Soc. (N.S.) 16 (1987),
no. 2, 247–263.
R. Lee and R. H. Szczarba, On the homology and cohomology of congruence subgroups,
Invent. Math. 33 (1976), no. 1, 15–53.
H. Maazen, Homology stability for the general linear group, thesis, University of Utrecht,
1979.
I. G. Macdonald, Symmetric Functions and Hall Polynomials, second edition, Oxford
Mathematical Monographs, Oxford, 1995.
B. Mirzaii, W. van der Kallen, Homology stability for unitary groups, Doc. Math. 7 (2002),
143–166.
G. Malle and A. Zalesski, Prime power degree representations of quasi-simple groups, Archiv.
Math. 77 (2001), 461–468.
D. Quillen, Finite generation of the groups Ki of rings of algebraic integers, in Algebraic
K-theory, I: Higher K-theories (Proc. Conf., Battelle Memorial Inst., Seattle, Wash., 1972),
179–198. Lecture Notes in Math., 341, Springer, Berlin.
O. Randal-Williams and N. Wahl, Homological stability for automorphism groups, preprint
2015, arXiv:1409.3541v3.
M. Reeder, The Steinberg module and the cohomology of arithmetic groups, J. Algebra 141
(1991), no. 2, 287–315.
W. van der Kallen, Homology stability for linear groups, Invent. Math. 60 (1980), no. 3,
269–295.
Avner Ash
Department of Mathematics
Boston College
Chestnut Hill, MA 02467-3806
[email protected]
Andrew Putman
Department of Mathematics
University of Notre Dame
279 Hurley Hall
Notre Dame, IN 46556
[email protected]
21
Steven V Sam
Department of Mathematics
University of Wisconsin
480 Lincoln Dr.
Madison, WI 53706-1325
[email protected]
| 4math.GR
|
A STUDY ON SPLAY TREES∗
LUÍS M. S. RUSSO†
arXiv:1511.03148v1 [cs.DS] 10 Nov 2015
Abstract. We study the dynamic optimality conjecture, which predicts that splay trees are a
form of universally efficient binary search tree, for any access sequence. We reduce this claim to a
regular access bound, which seems plausible and might be easier to prove. This approach may be
useful to establish dynamic optimality1 .
Key words. Data Structures, Binary Search Trees, Splay Trees, Dynamic Optimality, Competitive Analysis, Amortized Analysis
AMS subject classifications. 68P05, 68P10, 05C05, 94A17, 68Q25, 68P20, 68W27, 68W40,
68Q25
1. Introduction.
Binary search trees (BSTs) are ubiquitous in computer science. Their relevance
is well established, both in theory and in practise. Figure 1 illustrates this data
structure. The numbers in the nodes represent the keys. If read left to right, the
keys form an ordered sequence, shown bellow the tree. This property can be used
to efficiently determine if a given number exists in the tree. If we want to check
if the tree contains the number 0.55 we can start by comparing this value with the
one at the root, 0.56. Since the number we are searching for is larger than 0.56, the
search continues in the left sub-tree, i.e., it proceeds to the node containing 0.40, thus
discarding all the sub-tree to the right. It is the process of eliminating a large portion
of its search space that makes BSTs efficient.
0.56
0.40
0.58
0.35
0.21
0.51
0.87
0.41
0.52
0.85
0.21 < 0.35 < 0.40 < 0.41 < 0.51 < 0.52 < 0.56 < 0.58 < 0.85 < 0.87
Fig. 1. Binary search tree containing keys from [0, 1].
As the search goes through the tree it will eventually reach the node containing
0.52, at which point it becomes apparent that 0.55 is not present in the structure. In
fact this search concludes that no number in the open interval ]0.52, 0.56[ exists in the
∗ This work was supported by national funds through Fundação para a Ciência e a Tecnologia
(FCT) with reference UID/CEC/50021/2013 and DataStorm EXCL/EEI-ESS/0257/2012.
† INESC-ID and the Department of Computer Science and Engineering, Instituto Superior
Técnico, Universidade de Lisboa ([email protected]).
1 This draft version, as it exists immediately prior to editing and production by the Publisher
will be posted on the noncommercial pre-print server arXiv.org, in accordance with SIAM’s Consent
to Publish.
1
2
TEX PRODUCTION
tree. This is the only valid conclusion. Even though the search passed through nodes
0.40 and 0.51 it would not be valid to conclude that no number in ]0.40, 0.51[ exists,
because 0.41 is on the tree. Hence whenever a search fails we still obtain information
from the data structure. In particular we obtained the numbers in the tree that are
the successor (0.56) and the predecessor (0.52) of 0.55.
Consider how the shape of the tree affects the performance of the queries. If we
are computing a single, isolated, query, we may want to guarantee that, at each node,
the size of the left and right sub-trees is roughly the same. Note that we never know
which sub-tree a given search will choose. If both sub-trees have the same size at least
half the search space is discarded at each step. Alternatively we could impose that
the tree does not contain long branches, since a search might have to traverse all of
it.
Either way, some policy must be used to main a tree shape that reduces the search
time. Determining “good” tree shapes is a complex task, specially when we consider
a sequence of queries, instead of a single query. Moreover, must of the time, this
sequence is not known a priori. If a sequence of a couple million queries searches for
0.21 in half of them and only once for 0.85, then it might be better to keep 0.21 at
the root and leave 0.85 on the longest existing branch. We would expect such a shape
to reduce the overall time. Hence a balanced tree is important if we need to answer
several essentially distinct queries. On the other hand if the queries are strongly
biased towards a specific region of the keys maintaining certain branches shorter than
the rest might be necessary.
Several strategies are known to maintain efficient BST shapes, see Knuth [18] for
an introduction to the subject. In this paper we focus on the approach used by splay
trees [24]. Our goal is to show that this approach is optimal, in the sense that no
other strategy can be asymptotically faster than splay trees. Notice that splay trees
dynamically alter their structure as they process queries. Hence we assume that any
other BST can do the same. Figure 2 illustrates the tree that results from accessing
0.21 on a splay tree.
Our main contribution is the following:
• We, almost, show that splay trees are dynamically optimal. This means that
no other BST is asymptotically faster than splay trees, no matter what is
the structure of the query sequence or the re-shaping policy it uses. This
property was conjectured to be true, over 30 years ago [24].
Our proof depends on a regular access property, which we deem plausible and
conjecture to be true, see Section 3. At this time we do not have a proof of
this property.
2. The Problem.
The introduction exemplified how a BST can be used to maintain a finite set K
of, keys, elements of a universe set X. In our example X is Q. Our goal is to maintain
a data structure such that given a fraction x from Q we can determine if x is in K,
or not. We refer to these operations as queries. A query is successful it x does belong
to K and unsuccessful otherwise.
Several efficient data structures exist for this problem, depending on which resources are critical and on which extra operations are necessary. For our example we
also want map behaviour. This means that the elements of K contain information.
Whenever x does exist in K we also want to access the associated information. The
two major classes of data structures that can be used in this scenario are trees and
hashes. The set Q is, in general, referred to as the key universe and the associated
3
USING SIAM’S LATEX MACROS
0.21
0.56
0.35
0.58
0.40
0.87
0.51
0.41
0.85
0.52
0.21 < 0.35 < 0.40 < 0.41 < 0.51 < 0.52 < 0.56 < 0.58 < 0.85 < 0.87
Fig. 2. After splaying 0.21.
y
x
A
x
C
Zig
y
A
B
B
C
A < x < B < y < C
A < x < B < y < C
Fig. 3. Zig operation.
information is referred to as the value set. In this paper we omit this latter set.
Contrary to hashes, BSTs rely on the order relation among the elements of the
universe. In Q we have for example that (1/4) < (1/2). Hence BSTs can efficiently
determine successors and predecessors. The successor of x in K is min{y ∈ C|x < y}.
To simplify the analysis we do not consider unsuccessful queries, i.e., we assume that
all the queries find elements in K. An unsuccessful query for x can be modelled by
two successful ones, one for the successor and one for the predecessor. In general
BSTs support inserting and removing elements from K, but we also do not study
those operations.
The re-structuring policy of splay trees consists in moving the accessed nodes, and
the nodes in the respective path, upwards towards the root. The precise operations
are shown in Figures 3, 4 and 5, where the node containing x is the one being accessed.
The configuration before the access is represented on the left and the structure after
the access on the right. Accesses when x is on the right sub-tree are obtained by
symmetry.
To compare splay trees with any other BST we assume that the other BST works
in the following way. Besides the tree itself there is a cursor that moves between
nodes. An algorithm on a BST may perform any of the following operations:
4
TEX PRODUCTION
z
y
x
A
x
D
y
A
Zig Zig
C
z
B
B
C
A < x < B < y < C < z < D
D
A < x < B < y < C < z < D
Fig. 4. Zig Zig operation.
x
z
y
y
D
Zig Zag
x
A
A
z
B
C
D
A < y < B < x < C < z < D
B
C
A < y < B < x < C < z < D
Fig. 5. Zig Zag operation.
Compare the key at the cursor with the current search value.
Move the cursor to an adjacent node, left child, right child or parent.
Rotate the node upwards. The Zig operation is a rotation of node x, (Figure 3).
To perform a search in a BST the cursor starts at the root and performs a sequence of
the previous operations. The result of the search is the node for which a comparison
with x was equal. This node may, or may not, be the last one in the sequence. At the
end of a search the cursor must return to the root, those moves are accounted for.
In this model we allow for constant extra information at each node, such as colour
for red-black trees or height for AVL trees. The intention is to forbid dynamic memory
blocks, which could be used to implement hashing. Hence only a fixed constant
amount of information can be added to a node. Other BST models exist, that can be
shown to be within a constant factor of the one we consider [29].
Among all BSTs we focus on the one that achieves the best performance for
a query sequence. This means choosing the tree after knowing the complete query
sequence, in other words, offline. We refer to this optimal tree as T .
We do not count compare operations and therefore assume that they cost 0. For
the optimal tree we count the number of moves and rotations, but for the splay tree
we only count the number of moves. This avoids having to carry around a factor of 2
or 3, which would result from counting rotations and or comparisons.
3. Analysis Overview.
In this section we overview the main techniques in the analysis and describe our
new potential function.
USING SIAM’S LATEX MACROS
5
Amortized analysis. The analysis is amortized, meaning that the time a splay operation takes is not accounted per si, but considering preceding and succeeding
operations [28]. Therefore the amount of time an operation requires can be
partially shifted to some other operation in the sequence. The resulting time
is known as the amortized time. The most common way to transform the
total sum into an equivalent telescopic sum is to associate a “potential” value
to each state of the data structure D, the value Φ(D). The amortized cost
of an operation is then defined as ĉ = c + Φ(D′ ) − Φ(D), were c is the actual
cost and Φ(D′ ) the potential after the operation is performed. Summing the
previous equations over all m queries, and using
the fact that
P
it is a telescopic
P
sum, we obtain a global relation 1≤i≤m ci =
1≤i≤m ĉi −Φ(Dm ). In this
equation Φ(Dm ) represents the potential after the sequence, which might be
a negative number. The value Φ(D0 ) represents the potential at the beginning and it was omitted because it is assumed to be 0. This is a classical
tool in the analysis of splay trees. We present a new potential function. This
function is chosen so that the structure of the optimal tree gets transferred
into the splay tree.
Restricted Accesses. After bounding the amortized cost of accesses in S we proceed
to amortize the cost of rotations in T . Such rotations alter Φ(D) and therefore
most be accounted for. To obtain a constant bound on this variation we
introduce extra, organizing splays, in S. The amortized time to compute
these extra splays will depend on the depth of the node we are rotating in
T . Therefore if T performs two rotations in a row we need to count the node
depth twice. To obtain a valid bound we force T to move its cursor back to
the root after performing a rotation. We also further impose that T may not
access nodes of depth 3 or more. Although these restrictions seem harsh we
show that, with a linear slowdown factor, it is possible to force the optimal
sequence to respect them, Lemma 4.3.
Organizing Splays. Splay trees do not know the full sequence of queries in advance,
fortunately the analysis does. This means that in the analysis it is possible to
verify if the structure of the splay tree is adequate for the upcoming queries.
We optimize the tree structure by adding extra organizing splays. These
splays introduce the logic dependency on a regular access.
The regular access statement is the following:
Conjecture 3.1 (Regular Access). Let c1 +. . .+cm be the total cost of a sequence
of m splay operations and c′1 + . . . + c′m+e the cost of performing the same sequence
of splay operations, augmented with e extra splays anywhere. Then c1 + . . . + cm =
O(c′1 + . . . + c′m+e ).
This conjecture basically states that computing more splays takes longer. We
assumed this property in our reasoning, until it became evident that it was not trivial
to establish. Our proof relies precisely on this property because the extra splays are
used to optimize the tree for future accesses.
We will now define the potential function Φ and give an overview of the results
that we prove in Section 4. T denotes the optimal tree. To every node v of T we
assign a weight wT (v) = 1/4dT (v) , where dT (v) is the depth of v, i.e., the distance to
the root. The root itself has depth 0, its children have depth 1 and so on. For each
node we add up all the weights of its sub-tree, including wT (v) itself, the resulting
sum is denoted sT (v). The rank rT (v) is computed as log(sT (v)). The potential of
6
TEX PRODUCTION
tree T is given by the following sum:
P (T ) =
X
rT (v)
v∈T
Let us consider the trees in Figure 6. We have that wT (a) = wT (c) = 1/16,
wT (b) = wT (e) = 1/4 and wT (d) = 1. The resulting r values are rT (a) = rT (c) =
log(1/42 ) = −4, rT (e) = −2, rT (b) = log(1/4 + 1/8) = log(3/8) and rT (d) = log(1 +
1/2 + 1/8) = log(13/8). Hence the total potential of T is P (T ) = rT (b) + rT (d) − 10.
We use S to denote the splay tree. There is a one to one relation between the
nodes of these trees, because both trees share the same key values. For any node v, of
T , its correspondent in S is f (v), by correspondent we mean that v and f (v) store the
same key value. This correspondence is used to transfer the weight values from T to
S, more precisely set w(f (v)) = wT (v). We omit the S subscript to simplify notation.
The values of w, s and r that do not have a subscript refer to S. Moreover we also
omit the function f and assume instead that v in S means f (v). Hence the previous
relation will be written as w(v) = wT (v). The values s(v) and P (S) are computed as
before. In general s(v) and sT (v) are not, necessarily, equal. See Figure 6. Likewise
P (T ) and P (S) are also not, necessarily, equal. In fact we use Φ(D) = P (S) − P (T )
as our potential function.
In our example we have that r(a) = rT (a), r(c) = rT (c), r(e) = rT (e), r(b) =
rT (d) and r(d) = log(1 + 1/4 + 1/16) = log(21/16). Therefore Φ(D) = log(21/16) −
log(3/8) = log(7/2).
In the analysis we assume that S and T take turns. First S searches for x and T
remains idle. Then T searches for the same x, while S remains idle. Hence whenever
the structure of S changes P (S) changes, but P (T ) remains constant. However when
the structure of T changes, both P (T ) and P (S) change.
The structure of our argument is as follows:
• We show that −n < Φ(Dm ), Lemma 4.2, where n is the number of keys.
• We show that the amortized cost to splay a node v, in S, is at most O(1 +
dT (v)), Lemma 4.3.
• We show that whenever there is a rotation at node v in T we have ∆Φ = O(1),
by splaying at most 3 nodes, whose depth in T is at most dT (v), Lemma 4.5.
• We combine these results with a restricted optimal sequence, Lemma 4.3, and
obtain an optimality result, Theorem 4.6.
Thus we establish that if the optimal tree, with n nodes, needs R rotations and M
moves to process a given query sequence then splay trees require O(n + R + M ) time,
provided they have the regular access property.
4. The Details.
In this section we fill in the details that are summarized in the previous section.
This section is divided into three parts. We start by studying the properties of our
potential function Φ. We then shift the focus to T and explain how to restrict the
general access sequence to ensure that it respects certain desirable conditions. The
last part presents bounds for the amortized costs of accessing elements in S and
rotating nodes in T . We then combine these results to obtain an optimality result.
4.1. Properties of Φ.
Let us start by analyzing the potential function Φ(D). We assume that the initial
configuration of S and T is the same. Therefore P (S) = P (T ) and Φ(D0 ) = 0.
USING SIAM’S LATEX MACROS
13/8
T
d
3/8
1/4
b
e
1/16
1/16
a
w:
7
1
16
c
1
4
1
16
(1)
1
4
f
c
e
1/4
1/16
a
d
21/16
1/16
b
13/8
S
Fig. 6. Optimal tree T on top, splay tree T in the bottom, upside down. Example of computing
Φ. The f function is illustrated by the dashed arrows. The weight values w are show in between the
two trees. Inside the black rectangles we show the s values associated with the nodes.
T
T′
S
n
m
e
R
M
R′ , M ′
v, u
t
dT (v)
w(v)
s(v)
r(v) = log(s(v))
wT (v), sT (v), rT (v)
Φ(D)
The optimal binary search tree.
A restricted version of T .
The splay tree.
Number of nodes in T and S
Number of queries in a sequence.
Number of added organizing splays.
Number of rotations in T
Number of cursor moves in T
Same as R and M , but in T ′
Nodes
The root node
The depth of node v in T
Weight of node v
Sum of the weights, in the sub-tree bellow v
Rank for node v
Weight, sum and rank in T
Potential value for a given state of S and T
Fig. 7. Symbol table.
8
TEX PRODUCTION
Given that we are using base 2 logarithms it would be natural to use base (1/2)
powers for the weights w and wT . Using (1/4) instead causes the amortized access
cost to be twice as big. On the other hand we obtain some handy properties.
Lemma 4.1. Let v be a node. The following bounds hold:
(4.1)
0 ≤ w(v)
for any v in T or S.
(4.2)
(4.3)
w(v) ≤ 1
w(v) ≤ s(v)
for any v in T or S.
for any v in T or S.
(4.4)
(4.5)
sT (v) < 2 × wT (v)
s(v) < 2
only for v in T .
for any v in T or S
s(t′ ) = sT (t)
(4.6)
for the root t′ of S and t of T
Proof. Inequalities (4.1) and (4.2), follow directly from our definition of weights.
Inequality (4.3), follows from the definition of s and Inequality (4.1).
Inequality (4.4) use a geometric series. For the nodes v in T the value sT (v)
is a sum of powers of 1/4, which depends on the topology of tree. If T is a single
branch then sT (v) < wT (v)(1 + (1/4) + (1/42 ) + . . . + (1/4j ) + . . .) < (4wT (v)/3).
If T is perfectly balanced, this bound becomes even larger. In that case sT (v) <
wT (v)(1 + 2 × (1/4) + 22 × (1/42 ) + . . . + 2j × (1/4j ) + . . .) < 2 × wT (v). In this case
the bound comes from the geometric series of (1/2). Moreover the remaining cases
will also be bounded by the geometric series of (1/2).
Inequality (4.5) holds for T , because of inequalities (4.2) and (4.4). For S note
that in general T cannot have more than 2d nodes with depth d and consequently S
cannot have more that 2d nodes with weight (1/4d), hence we obtain the geometric
series of (1/2) again.
For equality (4.6) notice that the sums at the root contain all the weight values.
Since these values are mapped from T to S they are globally the same and therefore
so is their sum.
In particular from Inequality (4.5) we deduce that r(t) ≤ 1, for the root t of S.
Using potential functions that can assume negative values implies that the value
of Φ(Dm ) becomes a term in the total time. Hence we bound this value.
Lemma 4.2. The bound −n < Φ(D) holds for any configuration of S and T ,
where n is the number of keys.
Proof. The following derivation establishes the result.
(4.7)
P (S) − P (T ) = Φ(D)
X
(4.8)
log(w(v)) − P (T ) ≤ P (S) − P (T )
v∈S
(4.9)
X
v∈S
(4.10)
log(w(v)) −
X
log(2wT (v)) <
v∈T
log(w(v)) − P (T )
X
log(w(v)) −
v∈S
−
X
v∈T
(4.11)
X
1=
v∈S
−n = −
X
X
log(2wT (v))
v∈T
1
v∈T
To obtain Inequality (4.8) use Inequality (4.3), for all the terms of the sum in P (S).
For Inequality (4.9) use Inequality (4.4) and apply logarithms and change signals.
USING SIAM’S LATEX MACROS
9
Equation (4.10) follows from the same argument as Equation (4.6) and the fact that
log(2wT (v)) = 1 + log(wT (v)).
4.2. Restricting T .
Competing directly with T is fairly hard, partially because the sequence of accesses in T is completely free. For our purposes we need that the accesses in T respect
the following properties:
• When a node v is visited by the cursor or involved in a rotation its depth is
less than 3, i.e., dT (v) < 3.
• After a rotation the cursor moves back to the root.
We refer to a sequence of nodes that respects these conditions as a restricted sequence.
Let us show that given any access sequence of visited nodes in T it is possible to
simulate it in another tree T ′ , so that the accesses in T ′ are restricted. By simulation
we mean that the sequence visited by the cursor of T is a sub-sequence of the one
visited by the cursor of T ′ . Naturally the simulation will be slower than the original
sequence.
Lemma 4.3. Any sequence of nodes visited by the cursor of T , consisting of M
moves and R rotations, can be simulated by a restricted sequence of cursor moves on
a tree T ′ with 4M + 3R moves and 2M + R rotations.
Proof. Figure 8 illustrates the operations that we consider in this Lemma. Besides
the same keys as T , the tree T ′ contains two extra keys, min and max, that are,
respectively, smaller and larger than all the other keys in T . The simulation works
by keeping the node at the cursor of T in the root of T ′ .
We assume that the initial state of T and T ′ is almost alike. The same key is
stored a the root. The children of T ′ are min and max. The left sub-tree of T is
stored in the right child of min and the right sub-tree of T is stored in the left child
of max. Shown at the top in Figure 8.
Whenever the cursor of T moves down the corresponding node is moved up to
the root of T ′ , the process is a ZigZag operation on T ′ , or ZagZig depending on which
grand-child. Note that between the rotations of the ZigZag operation the cursor
must return to the root to comply with the second condition of restricted sequences.
These movements are underlined in the example. Hence every downward move in T
originates 4 moves and 2 rotations in T ′ . This process transforms the trees in the top
of Figure 8, to the trees in the middle. In this example the sequence of operations is:
moveTo(min), moveTo(x), rotate(), moveTo(y), moveTo(x) rotate().
The cursor may also move upwards on the tree, hence reversing the previous move.
In this case the sequence of operations would be: moveTo(y),rotate(), moveTo(x),
moveTo(min), rotate(), moveTo(y). This requires 4 moves and 2 rotations. Therefore a move in T originates 4 moves and 2 rotations in T ′ .
In this example the initial move was to y, because y is the child in T ′ that
corresponds to he parent in T . To determine this property we could store, in T ′ the
depth of the nodes in T . Only for the nodes in the leftmost and rightmost branches of
T ′ , otherwise updated values would be hard to maintain. However it is not necessary
to do so. Recall that T ′ is simulating T and T is the optimal tree, which does not
need to consult the keys to know which move to perform, it only needs to behave as
a BST. Therefore T ′ also does not need extra information.
Whenever a rotation is performed in T the structure of T ′ must be adapted
accordingly. Recall that a rotation on node v means that v is moved upwards. In
Figure 8 the transition from the middle to the bottom shows how a rotation alters
the structure of T ′ . In this case the sequence of moves is: moveTo(y), moveTo(max),
10
TEX PRODUCTION
T
T′
y
y
x
A
min
z
B
C
max
x
D
A
y
min
z
B
B
C
D
x
x
A
z
C
D
y
A
max
B
z
C
D
x
x
min
y
A
z
B
C
max
y
A
D
z
B
C
D
Fig. 8. Simulation of T (left) by T ′ (right). Initial configuration on top. The middle configuration represents the result of moving the cursor from node y to node x. The bottom configuration
shows the result of performing a rotation on x, while in the middle configuration.
rotate(), moveTo(x). Hence a rotation originates 3 moves and 1 rotation in T ′ . The
general procedure is to move to y because it is the child of T ′ that corresponds to the
parent of x in T , and move again in the same direction, in this case to max. This
node is rotated upwards and the procedure finishes by returning to the cursor back
to the root.
Note that in Figure 8 the cursor of T is drawn close to the root. In general the
structure of the upward path from the cursor of T to the root gets splitted into the
leftmost and rightmost branches of T ′ , but the update procedure is essentially as
explained.
USING SIAM’S LATEX MACROS
11
In the following results we continue to use T , instead of T ′ , which simplifies the
analysis. The tree T ′ is used only in Theorem 4.6.
4.3. Amortized Costs of S and T .
Let us now bound the amortized time of splaying a node of S.
Lemma 4.4. Splaying a node v takes at most 4 + 6dT (v), amortized time, where
dT (v) is the depth of the corresponding node in T .
Proof. The following derivation establishes this bound:
(4.12)
ĉi ≤ 1 + 3[r(t) − r(v)]
(4.13)
(4.14)
< 1 + 3[1 − r(v)]
= 4 − 3r(v)
(4.15)
≤ 4 − 3 log(w(v))
(4.16)
(4.17)
= 4 − 3 log(1/4dT (v) )
= 4 + 6dT (v)
Inequality (4.12) is the classic amortized access Lemma 5.1, see Section 5. Inequality (4.13) follows from (4.5) and recalling that r(t) = log(s(t)). Inequality (4.15)
results from applying logarithms and switching signs to (4.3). The remaining equations use the definition of weight w and simplify the result.
This Lemma shows that the amortized time to splay a node is slightly more than
6 times the time that it is necessary to access the corresponding node in T . Let us
focus on the amortized cost of accessing nodes in T . Whenever T accesses a node
there is no real cost for S. However if those accesses involve rotations in T then the
value of Φ changes, which will constitute an amortized cost that S most account for.
Lemma 4.5. Whenever a node v of T , with dT (v) < 3, gets rotated the bound
∆Φ ≤ 11 + log(11) holds, after, at most, 3 nodes of S are splayed. Each splayed node
v ∗ has dT (v ∗ ) ≤ dT (v).
Proof. Recall that in a rotation the node v moves upwards on the tree. We
consider the cases when dT (v) = 1 and dT (v) = 2.
The proof is illustrated in Figure 9, which shows the situation when dT (v) = 1,
but from which we can infer properties that apply to all cases.
The optimal tree T is represented on top and the splay tree S in the bottom,
upside down. To simplify let us assume that these trees are only slightly different.
Meaning that they share some structure. The nodes in sub-trees A, B and C are the
same, but their shape is not necessarily equal. For example the descendants of node
u are not, necessarily, the same.
A fundamental observation for this proof is that almost all the nodes, in the trees,
contribute 0 to the value ∆Φ. The only nodes that effectively contribute to ∆Φ belong
to, at least, one the following categories:
1. Nodes that contain descendants from more than one of the sets A, B, C,
either in S or T or both. Examples of these nodes are the ones containing x
and y. Moreover some nodes a ∈ A, b ∈ B and c ∈ C may appear in S with
this property.
2. Nodes for which the set of descendants is altered by the rotation in T . Only
nodes x and y.
Hence we are claiming that the nodes in A, B and C do not contribute to ∆Φ.
Consider a node u, contained in the sub-tree A. Let wT (u) be the weight of u before
the rotation and wT′ (u) be the weight after the rotation. Likewise we consider the sT ,
s′T , rT , rT′ values and the values s, s′ , r, r′ , over S.
12
TEX PRODUCTION
y
T
x
x
C
y
A
u
B
A
rotate(x)
B
C
u
u
u
A
A
B
C
x
S
y
B
C
x
y
Fig. 9. Variation of Φ with rotation on T.
Let us account for the contribution of u to ∆Φ, i.e., r′ (u) − r(u) + rT (u) − rT′ (u) =
log([s′ (u)/s(u)] × [sT (u)/s′T (u)]). Notice that w′ (u)/w(u) = 4, because the depth of
u decreases in T . Likewise s′ (u)/s(u) = 4, because the same happens for all the
elements in A. On the other hand sT (u)/s′T (u) = 1/4 because the order of the factors
is reversed. Therefore the previous value is log(4/4) = 0. A similar reasoning holds
for the elements in C. The nodes in B do not suffer potential variations and therefore
contribute 0 to the overall variation.
In this scenario the only nodes that contribute to ∆Φ are the ones that contain
descendants from more than one of the A, B, C sets. In our simplification only x
and y. However in the general case the nodes in sub-trees A, B and C have different
relations in S and in T . Hence we need to be more precise as to what these nodes
should be. The nodes in A are the descendants of x that are smaller than x. The
nodes in B are the descendants of both x and y. The nodes in y are the descendants
of y that are larger than y.
To force these sets of nodes to be consistent between S and T we splay node x
and node y in S, in this order. Now A and C contain the same nodes in S and T , but
it may happen that some ZigZig operation pulls a node b upwards from B, so that
it splits B and is in between the nodes x and y. In this case we must also count the
contribution from b.
Using an argument similar to the one above we can guarantee that s′ (b)/s(b) ≤ 4
and sT (b)/s′T (b) = 1 and therefore r′ (b) − r(b) + rT (b) − rT′ (b) ≤ 2 + 0 = 2.
For the node containing x we also have that s′ (x)/s(x) ≤ 4, but sT (x)/s′T (x) is
trickier, because s′T (x) now includes the extra terms related to wT′ (y) and s′T (C), these
values are non-negative and therefore decrease the fraction sT (x)/s′T (x). Therefore
the bound of 4 holds true. Hence for the node containing x we count another 4 units,
as r′ (x) − r(x) + rT (x) − rT′ (x) ≤ 2 + 2 = 4.
USING SIAM’S LATEX MACROS
13
For the node containing y we also have that s′ (y)/s(y) ≤ 4 but sT (y)/s′T (y) is
again trickier. The following derivation obtains a bound.
(4.18)
(4.19)
(4.20)
(4.21)
(4.22)
(4.23)
(4.24)
wT (y) + sT (C) + sT (B) + wT (x) + sT (A)
s′T (y)
wT (y) + sT (C) + sT (B) wT (x) + sT (A)
+
=
s′T (y)
s′T (y)
wT (x) + sT (A)
≤4+
s′T (y)
wT (x) + sT (A)
≤4+
1/4
(1/4) + sT (A)
=4+
1/4
(1/4) + (1/8)
≤4+
1/4
≤ 4 + 1.5
sT (y)/s′T (y) =
Equations (4.18) and (4.19) are simple manipulations. Inequality (4.20) is our
general bound of 4, for the fraction that does not change the descendants. Inequality (4.3) yields 1/4 = wT′ (y) ≤ s′T (y), which we use to establish Inequality (4.21).
Inequality (4.23) follows from Inequality (4.4) that yields sT (A) ≤ 2 × 1/16 = 1/8.
Therefore the total for y is r′ (y) − r(y) + rT (y) − rT′ (y) ≤ 2 + log(5.5) = 1 + log(11).
Summing up, in this case we splayed two nodes and have to account for b, x and
y, therefore ∆Φ ≤ 2 + 4 + (1 + log(11)) = 7 + log(11).
Let us consider the case dT (v) = 2. This means that there is a node vr which is
an ancestor of y, in T . Let us assume that vr is to the right of C. The case where
the node is to the left of A is easier. It may happen that vr is inside C in S, in that
case we splay vr . This splay operation may in turn split C by pulling up a node c.
In total we splayed 3 nodes and have 5 nodes that may contribute to ∆Φ, the nodes
b and c, that got pulled up, and the nodes x, y, vr .
For node c we have a bound of 4, using the argument above, which holds because
the descendants of c in T do not change. Hence r′ (c)−r(c)+rT (c)−rT′ (c) ≤ 2+2 = 4.
Note that vr = t is the root of both S and T . Therefore we can can bound its
variation by 0, instead of our pessimistic 4 value. For this node we have r′ (t) = rT′ (t)
and r(t) = rT (t), by Equation (4.6). Hence in this case the total bound is ∆Φ ≤
2 + 4 + 4 + (1 + log(11)) + 0 = 11 + log(11)
Notice that all this splaying may change the ancestor relation between x and y
in S. This may happen when there is no node b and we are pulling a node vℓ , to the
left of A. This relation in S does not change the argument above, as both x and y
are being pessimistically bounded in S. The crucial consequence of the extra splays
is that the nodes in A, B and C are properly encapsulated, and the number of nodes
whose descendants belong to more than one of these sets is limited.
We can now prove our optimality result.
Theorem 4.6. Consider a sequence of m queries, to a splay tree S with n + 2
keys, for which, a similar, optimal tree T uses R rotations and M cursor movements.
Provided that splay trees have regular access, then the total number of operations
performed by S is O(n + R + M ).
14
TEX PRODUCTION
Proof. The proof is given by the following deduction:
X
1≤i≤m
(4.25)
(4.26)
(4.27)
(4.28)
ci − k(n + 2) ≤ k
X
1≤i≤m+e
ℓ + c′i − k(n + 2)
= k (m + e)ℓ − (n + 2) +
≤ k (m + e)ℓ + Φ(Dm ) +
X
1≤i≤m+e
X
1≤i≤m+e
= k (m + e)ℓ + Φ(Dm ) − Φ(D0 ) +
= k (m + e)ℓ +
X
1≤i≤m+e
c′i
c′i
X
1≤i≤m+e
c′i
ĉ′i
(4.29)
(4.30)
≤ k [(m + e)ℓ + (11 + log 11)R′ + (1 + 3)(4 + 6M ′ )]
≤ k [(m + 3R′ )ℓ + (11 + log 11)R′ + (1 + 3)(4 + 6M ′ )]
(4.31)
(4.32)
= k [mℓ + 16 + (11 + 3ℓ + log 11)R′ + 24M ′ ]
= k [mℓ + 16 + (11 + 3ℓ + log 11)(2M + R) + 24(4M + 3R)]
(4.33)
= k [mℓ + 16 + (118 + 6ℓ + 2 log 11)M + (83 + 3ℓ + log 11)R]
The above deduction uses the restricted sequence over T ′ , instead of the original
optimal sequence over T . We point out which of the results in the previous lemmas
apply to T ′ . The values R′ and M ′ refer to the number of rotations and cursor
movements in T ′ .
We aim to bound the value on the left. The first inequality is our regular access
conjecture. In here we are using explicit constants, k and ℓ, instead of the notation O,
to determine the hidden factors. Equation (4.25) rearranges the sums. Equation (4.26)
follows from Lemma 4.2. In Equation (4.27) we add the value Φ(D0 ) = 0, because
we are assuming that the initial structure of S and T ′ is the same. Equation (4.28)
follows from the telescopic nature of the definition of amortized costs.
Inequality (4.29) follows from Lemmas 4.4 and 4.5. From Lemma 4.4 we conclude
that the sum of the ĉ′i that corresponds to splay operations that answer queries is at
most (4 + 6M ′ ), because T ′ must also move its cursor to those nodes and between
any such nodes the cursor returns to the root. Therefore for T ′ to answer a query it
most perform at least dT (v) cursor movements, where v is the node containing the
key that corresponds to the query.
To bound the sum of the ĉ′i that corresponds to the costs to update S, due to
rotations in T ′ , we use Lemma 4.5. Note that we can apply this Lemma because of the
first property of restricted sequences. The fixed cost yields the term (11 + log 11)R′ .
The amortized cost of splaying, at most, 3 nodes of S is bounded by the term 3(4 +
6M ′ ). This bound holds because the second property of restricted sequences implies
that T ′ must perform at least dT (v) cursor movements before rotating node v. This
property states that after each rotation the cursor of T ′ returns to the root, see
Section 4.2.
USING SIAM’S LATEX MACROS
15
In Equation (4.30) we use the fact that e ≤ 3R′ , i.e., we use at most 3 extra
splays for each rotation in T ′ . Equation (4.31) results by rearranging terms. In
Equation (4.32) we use the bounds from Lemma 4.3. Equation (4.33) obtains the
final bound, by rearranging terms.
Let us now survey related work, so that we can discuss our result in context.
5. Related Work.
This section describes splay trees and some previous results. These trees were
proposed by Sleator and Tarjan [24] in 1985. An extensive up to date survey on
dynamic optimally was given by Iacono [16].
There are several restricted optimally results for BSTs. Assuming
Pm that the optimal BST is static than the best performance is the entropy O( i=1 fi log(m/fi )),
where m is the total number of queries and fi is the number of queries for the ith key. Knuth presented the first algorithm to determine such an optimal tree [18].
This dynamic programming algorithm requires O(n2 ) time to determine the optimal
tree. Kurt Mehlhorn obtained a faster, O(n) time algorithm, which approximates the
optimal tree [21]. In the context of this paper this result would be referred to as
statically optimal, because we are hiding factors into the O notation. Splay trees also
obtain this kind of optimality and moreover do not need to know the fi values. Static
optimality becomes the best upper bound when the keys in the query sequence are
independent of each other.
We are only interested in approximating dynamic optimality, i.e., using O notation
and being a factor away from the optimal value, because it is expected that the
exact problem is NP-Complete [7]. Even in this case the only BST that achieves
the dynamic optimality bound, assumes free rotations and takes exponential time to
select operations [3].
The good performance properties of splay trees are due to the access Lemma,
which was establish in the paper on splay trees [24]. Here we repeat the original
argument and highlight the fact that it follows from Jenssen’s Inequality [17].
Lemma 5.1 (Amortized Access). Splaying a node v to the root t takes at most
1 + 3[r(t) − r(v)], amortized time.
Proof. Recall Figures 3, 4 and 5. When v is at the root, the bound is trivial.
Hence let us focus on the last operation that is used to access v, which contains x in
our figures.
Zig In this case x is a child of y, which is at the root. The amortized cost is computed
as follows:
(5.1)
(5.2)
(5.3)
ĉZig = 1 + r′ (x) + r′ (y) − r(x) − r(y)
≤ 1 + r′ (x) − r(x)
≤ 1 + 3[r′ (x) − r(x)]
The only nodes that change rank are x and y. The Inequality (5.1) follows
from the fact that r′ (y) ≤ r(y). On the other hand Inequality (5.2) is true
because r(x) ≤ r′ (x).
ZigZig In this case x, y and z change rank
(5.4)
ĉZigZig = 2 + r′ (x) + r′ (y) + r′ (z) − r(x) − r(y) − r(z)
(5.5)
(5.6)
= 2 + r′ (y) + r′ (z) − r(x) − r(y)
≤ 2 + r′ (x) + r′ (z) − 2r(x)
(5.7)
≤ 3[r′ (x) − r(x)]
16
TEX PRODUCTION
Equation (5.5) follows from the fact that r′ (x) = r(z). Inequality (5.6) is
true because r′ (y) ≤ r′ (x) and −r(y) ≤ −r(x). Inequality (5.7) reduces to
proving that (r(x) + r′ (z))/2 ≤ r′ (x) − 1, this relation is known as Jensen’s
inequality and it holds because the log function is concave. More explicitly
the relation is the following:
log(s(x)) + log(s′ (z))
s(x) + s′ (z)
≤ log
2
2
ZigZag In this case x, y and z change rank.
(5.8)
ĉZigZag = 2 + r′ (x) + r′ (y) + r′ (z) − r(x) − r(y) − r(z)
(5.9)
(5.10)
= 2 + r′ (y) + r′ (z) − r(x) − r(y)
≤ 2 + r′ (y) + r′ (z) − 2r(x)
(5.11)
(5.12)
≤ 2[r′ (x) − r(x)]
≤ 3[r′ (x) − r(x)]
Equation (5.9) follows from the fact that r′ (x) = r(z). Inequality (5.10) is
true because −r(y) ≤ −r(x). Inequality (5.11) is also Jensen’s inequality, in
this case it reduces to (r′ (y) + r′ (z))/2 ≤ r′ (x) − 1. The Inequality (5.12) is
true because r(x) ≤ r′ (x).
Notice that in any access to S there is at most 1 Zig operation and several ZigZig and
ZigZag. This justifies why it is important to omit the term 1 in the bound for the
ZigZig and ZigZag operations. The overall bound is obtained by summing the bounds
of the respective operations. This sum telescopes to the expression in the Lemma.
A recent detailed study of this Lemma is available [4]. An important consequence
of the splay operation is that most of the nodes in the splayed branch are moved
upwards, i.e., their depth gets reduced, essentially in half. A detailed study on depth
reduction was given by Subramanian [25].
Several good performance theorem follow from this Lemma:
Static Optimality, obtaining the entropy bound we mentioned in the beginning of
the Section.
Static Finger, the performance of splay trees can be made to made dependent on the
position of a given element of the key set X. Let f be the position of a certain
element x in the ordered set X and |xi − f | the distance between that element
and the element xi , the i-th element in the query sequence. Then
Pm processing
a sequence of queries with a splay tree takes O(n log n + m + i=1 log(|xi −
f | + 1)). In fact splay trees have an even better better performance because
the chosen element at position f can be made dynamic [6, 5], but this result
requires a longer analysis.
Working Set, meaning that recently accessed are quicker to access again. Let t(x)
be the number of different items accessed before accessing x since the last time
x was accessed, or since the beginning of the sequence if x was never accessed
before. The the total time to process a sequence of queries is O(n log n + m +
P
m
i=1 log(t(i) + 1)). An important consequence of this bound is that if the
keys are assigned arbitrarily (but consistently) to unordered data, splay trees
behave as dynamically optimal [15].
Unified bound, combining the best performance of the 3 previous bounds. Dedicated structures where designed to achieve an even better bound which uses
USING SIAM’S LATEX MACROS
17
the dynamic finger bound [2], instead of the static version. Generalized versions of this bound, with tighter values for structured sequences where also
proposed [14].
Another important performance bound is the sequential access, or scanning, bound,
which states that accessing the n keys sequentially takes only O(n) time. This was
initially established by Tarjan [27] with a factor of 9, the current best factor of 4.5 [11]
was proven by Elmasry.
Besides dynamic optimality Sleator and Tarjan [24] proposed the traversal conjecture, which remains unproven. It is similar to the sequential access but the key
order is taken to be the preorder of another BST.
Other open conjectures on splay trees include the Deque conjecture, which claims
that splay trees can be used to implement a deque, with O(1) amortized time per
operation [26, 23]. The split conjecture claims that deleting all the nodes of a splay
tree takes O(n) time, in whatever order [20].
Studying binary search trees from a geometric point of view has yielded several
important results [7]. It presented an online algorithm, which may yield dynamically
optimal BSTs. This algorithm is referred to as greedy, and it was originally proposed as an offline algorithm [19, 22]. The first O(log n) performance bound, of this
algorithm, was established by Fox [12].
The same authors proposed Tango trees which are O(log log n) competitive to
the optimal BST [8]. This was the first structure to obtain a proven non-trivial
competitive ratio. The worst case performance of these trees was further improved [10]
to O(log n) per operation. The O(log log n) ratio was also proved for the chain-splay
variation of splaying [13].
Another alternative to try and obtain, proven, dynamic optimality consists in
combining BSTs. In [9] the authors show that given any constant number of online
BST algorithms (subject to certain technical restrictions), there is an online BST
algorithm that performs asymptotically as well on any sequence as the best input
BST algorithm.
Iacono [16] recently proposed another approach, the Weighted Majority Algorithm. This approach is proven to yield a dynamically optimal BST, provided any
such data structure exists.
A considerable amount of research as also gone into the study of lower bounds, i.e.,
formulating expressions that are smaller than R + M , i.e., the amount of operations
required by T . The first lower bounds where established by Wilber [29]. These
bounds where further improved by the geometric view of BSTs [7]. In this view
accesses to nodes in a tree are drawn as points in the plane. One coordinate stores
the ordered keys and the other coordinate represents time. We can then consider
rectangles using these points as corners. Two rectangles rectangles are independent
if their corners are outside their interception or in the boundary of the interception.
An access sequence in a BST must respect any independent set of rectangles. Hence
the maximum independent set of rectangles yields a lower bound for the number of
visited nodes in a BST.
The alternation lower and the funnel bounds can also be explained with the
geometric view of BSTs. Moreover this latter bound is also related to the number of
turns that it is necessary to move a key to root, with a procedure proposed by Allen
and Munro [1].
6. Conclusions and Further Work.
In this paper we studied the dynamic optimality conjecture of splay trees, which
18
TEX PRODUCTION
was proposed by the authors of these trees [24] and has stood for more than 30 years.
During this time a vast amount of results and approaches have been proposed for
this problem and related conjectures. Note that a proof of dynamic optimality would
establish other conjectures.
All this research has resulted in new data structures, several of which where
proven to be (log log n) competitive. The results we present are a step forward. We
reduced dynamic optimality to the regular access conjecture. This conjecture seems
very plausible. On the other hand, if it is false that would also be a remarkable
property.
In the future we plan to research this conjecture. An exhaustive literature review
may provide the necessary tools to solve it. We also plan to investigate if our potential
function can be applied to the analysis of other dynamically optimal candidates,
namely the greedy algorithm, combining BSTs or the weighted majority algorithm.
REFERENCES
[1] B. Allen and J. I. Munro, Self-organizing binary search trees, J. ACM, 25 (1978), pp. 526–
535.
[2] M. Badoiu, R. Cole, E. D. Demaine, and J. Iacono, A unified access bound on comparisonbased dynamic dictionaries, Theor. Comput. Sci., 382 (2007), pp. 86–96.
[3] A. Blum, S. Chawla, and A. Kalai, Static optimality and dynamic search-optimality in lists
and trees, Algorithmica, 36 (2003), pp. 249–260.
[4] P. Chalermsook, M. Goswami, L. Kozma, K. Mehlhorn, and T. Saranurak, Self-adjusting
binary search trees: What makes them tick?, in Algorithms - ESA 2015 - 23rd Annual
European Symposium, Patras, Greece, September 14-16, 2015, Proceedings, vol. 9294 of
LNCS, Springer, 2015, pp. 300–312.
[5] R. Cole, On the dynamic finger conjecture for splay trees. part II: The proof, SIAM Journal
on Computing, 30 (2000), pp. 44–85.
[6] R. Cole, B. Mishra, J. Schmidt, and A. Siegel, On the dynamic finger conjecture for splay
trees. part I: Splay sorting log n-block sequences, SIAM Journal on Computing, 30 (2000),
pp. 1–43.
[7] E. D. Demaine, D. Harmon, J. Iacono, D. Kane, and M. Pǎtraşcu, The geometry of
binary search trees, in Proceedings of the 20th Annual ACM-SIAM Symposium on Discrete
Algorithms (SODA 2009), New York, New York, January 4–6 2009, pp. 496–505.
[8] E. D. Demaine, D. Harmon, J. Iacono, and M. Pǎtraşcu, Dynamic optimality—almost,
SIAM Journal on Computing, 37 (2007), pp. 240–251. Special issue of selected papers from
the 45th Annual IEEE Symposium on Foundations of Computer Science.
[9] E. D. Demaine, J. Iacono, S. Langerman, and Ö. Özkan, Combining binary search trees, in
Proceedings of the 40th International Colloquium on Automata, Languages and Programming (ICALP 2013), Riga, Latvia, July 8–12 2013, pp. 388–399.
[10] J. Derryberry and D. D. Sleator, Skip-splay: Toward achieving the unified bound in the
BST model, in Algorithms and Data Structures, 11th International Symposium, WADS
2009, Banff, Canada, August 21-23, 2009. Proceedings, vol. 5664 of LNCS, Springer, 2009,
pp. 194–205.
[11] A. Elmasry, On the sequential access theorem and deque conjecture for splay trees, Theoretical
Computer Science, 314 (2004), pp. 459 – 466.
[12] K. Fox, Upper bounds for maximally greedy binary search trees, in Algorithms and Data Structures - 12th International Symposium, WADS 2011, New York, NY, USA, August 15-17,
2011. Proceedings, vol. 6844 of LNCS, Springer, 2011, pp. 411–422.
[13] G. F. Georgakopoulos, Chain-splay trees, or, how to achieve and prove -competitiveness by
splaying, Information Processing Letters, 106 (2008), pp. 37 – 43.
[14] J. Howat, J. Iacono, and P. Morin, The fresh-finger property, CoRR, abs/1302.6914 (2013).
[15] J. Iacono, Key-independent optimality, Algorithmica, 42 (2005), pp. 3–10.
[16] J. Iacono, In pursuit of the dynamic optimality conjecture, in Space-Efficient Data Structures,
Streams, and Algorithms, vol. 8066 of LNCS, Springer Berlin Heidelberg, 2013, pp. 236–
250.
[17] J. Jensen, Sur les fonctions convexes et les inégalités entre les valeurs moyennes, Acta Mathematica, 30 (1906), pp. 175–193.
USING SIAM’S LATEX MACROS
19
[18] D. E. Knuth, The Art of Computer Programming, Volume 3: (2Nd Ed.) Sorting and Searching, Addison Wesley Longman Publishing Co., Inc., Redwood City, CA, USA, 1998.
[19] J. M. Lucas, Canonical forms for competitive binary search tree algorithms, Rutgers University, Department of Computer Science, Laboratory for Computer Science Research, 1988.
[20] J. M. Lucas, On the competitiveness of splay trees: Relations to the union-find problem,
On-line Algorithms, 7 (1992), pp. 95–124.
[21] K. Mehlhorn, Nearly optimal binary search trees, Acta Informatica, 5 (1975), pp. 287–295.
[22] J. I. Munro, On the competitiveness of linear search, in Algorithms-ESA 2000, Springer, 2000,
pp. 338–345.
[23] S. Pettie, Splay trees, davenport-schinzel sequences, and the deque conjecture, CoRR,
abs/0707.2160 (2007).
[24] D. D. Sleator and R. E. Tarjan, Self-adjusting binary search trees, J. ACM, 32 (1985),
pp. 652–686.
[25] A. Subramanian, An explanation of splaying, J. Algorithms, 20 (1996), pp. 512–525.
[26] R. Sundar, On the deque conjecture for the splay algorithm, Combinatorica, 12 (1992), pp. 95–
124.
[27] R. Tarjan, Sequential access in splay trees takes linear time, Combinatorica, 5 (1985), pp. 367–
378.
[28] R. E. Tarjan, Amortized computational complexity, SIAM Journal on Algebraic Discrete Methods, 6 (1985), pp. 306–318.
[29] R. E. Wilber, Lower bounds for accessing binary search trees with rotations, SIAM J. Comput.,
18 (1989), pp. 56–67.
| 8cs.DS
|
LogLog-Beta and More: A New Algorithm for Cardinality Estimation Based on LogLog Counting
Jason Qin, Denys Kim, Yumei Tung
The AOLP Core Data Service, AOL, 22000 AOL Way
Dulles, VA 20163
E-mail: j [email protected]
Abstract—The information presented in this paper defines
LogLog-Beta (LogLog-β). LogLog-β is a new algorithm for
estimating cardinalities based on LogLog counting. The new
algorithm uses only one formula and needs no additional bias
corrections for the entire range of cardinalities, therefore, it is
more efficient and simpler to implement. Our simulations
show that the accuracy provided by the new algorithm is as
good as or better than the accuracy provided by either of
HyperLogLog or HyperLogLog++. In addition to LogLog-β
we also provide another one-formula estimator for
cardinalities based on order statistics, a modification of an
algorithm described in Lumbroso’s work.
Index Terms— Data analysis, Approximation algorithms,
Data Mining.
I.
Introduction
Cardinality estimation of a multiset has numerous
applications in data management, data analysis, and data
services, see papers [2][3][7][11][14][15], etc. Over the years,
many cardinality estimation algorithms have been proposed
and studied, especially the ones related to probabilistic
counting [1][4][5][6][8][9][11][10][13][16][17]. In this paper
we present a new algorithm related to a particular probabilistic
counting family called LogLog Counting. LogLog Counting
has been long studied: see the earlier ideas in Alon and Matias
and Szegedy[1] and Flajolet[8], the basic LogLog and
SuperLogLog algorithms in Duran and Flajolet[6], the popular
HyperLogLog in Flajolet et al. [9], and the recent
improvement HyperLogLog++ in Heule et al. [11].
The HyperLoglog algorithm in Flajolet et al. [9] serves as
the foundation for our study. Our new algorithm, LogLog-β,
can be regarded as a modification of the HyperLogLog. In this
introduction we will briefly present the HyperLogLog
algorithm. The remainder of the paper is organized as follows:
section II details the new algorithm LogLog-β, section III
gives some accuracy tests on LogLog-β, and section IV
provides another one-formula only cardinality estimator and
some further discussions on the two new algorithms.
HyperLogLog algorithm in Flajolet et al. [9] has five major
components: data randomization by a hash function, stochastic
averaging and register vector generation, the raw estimation
formula, the Linear Counting in Whang and Vander-Zanden
and Taylor [16], and bias corrections.
Let S be a multiset of data, M be a vector of size m with
m = 2p with p = 4, 5, … , h be a hash function where the
hash value is a 32 -bit array, let ρ be a function that maps a
bit array to its number of leading zeros plus one, and αm be a
parameter dependent of m (see [9]).
HyperLogLog Algorithm:
1) Initialize M :
M [i] = 0 for i = 0, 1, …, m − 1
2) Generate M :
For each s∈S , let i be the integer formed by the left p
bits of h(s) and w be the right 32 − p bits of h(s) , set
M [i] = max(M [i], ρ(w))
3) Apply the raw formula:
2
E = m−1αmm−M[i]
∑ 2
(1)
i=0
4) Apply Linear Counting if E≤ 52 m :
E = mlog mz
where z is the number of elements of M with value 0 .
1 32
5) Apply correction if E > 30
2 :
E =− 232log (1 − 2E32 )
When using a 64-bit hash function, step 5 in above
algorithm is no longer needed as long as the cardinality does
not near 264 . In this paper we use 64-bit hash function in
algorithms and simulations.
The accuracy of cardinality estimation provided by
formula (1) gets better when cardinality becomes larger
(proved by Flajolet et al. [9]), but the accuracy from Linear
Counting degenerates as cardinality increases. Therefore, loss
of accuracy around switching point 5m
2 , defined in [9], is
inevitable. A major improvement in HyperLogLog++[11] (in
addition to using 64-bit hash and sparse representation) is
provided a bias correction table for a range of cardinalities to
boost the accuracy around the switching point. The bias
correction table and correction range are m dependent and
determined empirically based on the estimates offered by
formula (1). In this paper HyperLogLog++ refers to the
implementation without sparse representation.
The new algorithm, LogLog-β, uses only one formula and
needs neither bias corrections nor Linear Counting, and
therefore implementation is simplified. Our simulations show
the accuracy provided by the new algorithm is as good as or
better than the accuracy provided by either of HyperLogLog
or HyperLogLog++ (without sparse representation).
1
II.
LogLog-β algorithm
In [9] Flajolet et al. proved asymptotically that the raw
formula (1) of HyperLogLog has a very low standard error (
1.04
). However the formula fails to provide good cardinality
√m
estimation for small or pre-asymptotic cardinalities. The main
idea of the new algorithm, LogLog-β, is to modify the raw
formula and to make it applicable to the entire range of
cardinalities, from very small to very large.
A. LogLog-β algorithm
LogLog-β Algorithm:
1) Initialize M :
Set M [i] = 0 for i = 0, 1, …, m − 1
2) Generate M :
For each s∈S , let i be the integer formed by the left p
bits of h(s) and w be the right 64-p bits of h(s) , set
M [i] = max (M [i], ρ(w) )
3) Apply the formula:
E = αmm(m−z)
(2)
m−1
−M[i]
β(m,z)+ ∑ 2
i=0
where z is the number of elements of M with value 0 and
β(m, z) is a function of m and z .
Clearly, formula (2) is a modification of formula (1) which
replaces the numerator αmm2 with αmm(m − z) and adds a
term β(m, z) to the denominator. Furthermore, formula (2) and
(1) are identical when z = 0 and β(m, 0) = 0 .
B. Function β(m, z)
The function β(m, z) is a kind of bias minimizer and
should go to 0 when z goes to 0. Therefore, for very large
cardinalities, formula (2) is almost the same as formula (1).
The ultimate goal is to find β(m, z) such that formula (2)
could yield highly accurate estimations for the entire range of
cardinalities.
There are many ways to choose β(m, z), for example
β(m, z) = β0(m)z , or β(m, z) = β0(m)z + β1(m)z2 + …. For the
sake of convenience of discussion in this paper we limit
β(m, z) to the following form:
β(m, z) = β0(m)z + β1(m)zl + β2(m)zl2 + … + βk(m)zlk
where zl = log (z + 1) , k≥0 , and β0(m) , β1(m) ,…, βk(m) are
m dependent constants. Clearly β(m, z) = 0 when z = 0 . An
implementation, based on Horner's rule, β(m, z) could be
evaluated by a total of (k + 1) multiplications and k additions
when zl is provided.
C. Constants β0(m) , β1(m) ,…, βk(m)
For given m , k , and a data set of cardinality c , under the
best circumstances we expect β(m, z) = β̂(m, z) where
m−1
β̂(m, z) = αmm(m−z)
− ∑ 2−M [i]
c
i=0
If we randomly generate data sets with given cardinalities
c1 , c2 , …, and cn (from very small to very large) and
compute z and β̂(m, z) for each ci , then by solving a least
2
square problem min ||β(m, z) − β̂(m, z)||2 we shall be able to
determine β0(m) , β1(m) ,…, βk(m) . In practice we pick
cardinalities ci such that c1 < c2 < … < cn , equally distanced,
with n≫k and z = 0 for some of the large cardinalities.
Furthermore for each given ci , we compute the means of z
and β̂(m, z) over many randomly generated data sets, then use
the means to solve the least square problem.
In our study we used MURMUR3-64 as the hash function
and java.util.Random for generating data sets with given
cardinalities. For example for p = 14 , m = 214 , we computed a
few β(m, z) :
β(m, z) = − 0.309142z + 13.733192zl − 8.636985zl2 + 1.328973zl3
β(m, z) = − 0.350308z + 3.949176zl − 6.403082zl2
+ 3.289908zl3 − 0.643495zl4 + 0.051568zl5
The number of terms of β(m, z) should be used is
determined by accuracy requirement: the larger k the better
accuracy the algorithm provides. However, we cannot reach
arbitrary accuracy by simply increasing k : the optimal
accuracy that can be achieved is dictated by m , the size of
vector M . In our study we found 3 to 7 to be a reasonable
range for k .
III.
Accuracy tests of LogLog-β
As long as β(m, z) is properly determined, the LogLog-β
algorithm can provide estimation for different ranges of
cardinalities with accuracy comparable to HyperLogLog++.
Namely, formula (2) can perform as well as Linear Counting
for small cardinalities and HyperLogLog Raw with bias
correction for large and very large cardinalities.
For completeness, we should determine β(m, z) and
perform accuracy tests for each precision p ( m = 2p ).
However, since the processes are quite similar, this paper only
shows the tests for precision p = 14 ( m = 214 = 16384 ). For
very large cardinalities the estimations by HyperLogLog,
LogLog-β, and HyperLogLog++ are literally the same
therefore cardinalities are limited to ≤200, 000 in accuracy
tests.
To determine β(m, z) for p = 14 , we computed the means
of z and β̂(m, z) over 100 (the more the better) randomly
generated data sets per cardinality for cardinalities from
1, 000 to 170, 000 in increments of 1, 000 . With k = 7 ( 8
terms) the function β(214, z) is
− 0.370393914z + 0.070471823zl
+ 0.17393686zl2 + 0.16339839zl3
− 0.09237745zl4 + 0.03738027zl5
− 0.005384159zl6 + 0.00042419zl7
2
and it is used for p = 14 in all the accuracy tests in this study.
By using above β(m, z) and any of the hash functions
MURMUR3, MD5, and SHA, we found the test results of
LogLog-β to be almost the same. Specifically, the function
β(m, z) was calculated using one hash function could be used
in LogLog-β in place of any other hash function as long as
their hash values are uniformly distributed and well
randomized. We chose MMURMUR3 for this study because
of its performance.
Both tests on the mean of relative errors and the mean of
absolute values of relative errors for 500 randomly generated
datasets per cardinality show that LogLog-β provides slightly
better estimations than HyperLogLog and HyperLogLog++,
especially for the cardinalities in the range of 10, 000 to
85, 000 , see Fig. 1 and Fig. 2. Please note that LogLog-β has
better performance in terms of accuracy and stability over
Linear Counting for almost all the small to mid-range
cardinalities. Figure 3 shows the relative error of one
randomly generated dataset per cardinality.
Fig.1. The mean of relative errors of cardinality estimations for 500 randomly
generated datasets per cardinality (the x -axis). Tested cardinalities are from
500 to 200, 000 in every 500 .
Figure 4, 5, and 6 show the empirical histograms of
cardinality estimations for 500 randomly generated datasets
per cardinality with cardinality being 1,000; 50,000; and
100,000
respectively.
Both
HyperLogLog
and
HyperLogLog++ use the same formulas in Fig. 4 and 6: Linear
Counting for cardinality = 1, 000 and formula (1) (with a
minor bias correction for HyperLogLog++) for cardinality=
100, 000 . Therefore, the histograms corresponding to
HyperLogLog and HyperLogLog++ are almost identical in
Fig. 4 and 6. In both figures LogLog-β shows comparable or
slightly better behaviors. In Figure 5 HyperLogLog,
HyperLogLog++, and LogLog-β show different behaviors
since HyperLogLog uses formula (1) and HyperLogLog++
uses both formula (1) and bias correction.
Fig. 2. The mean of abs(relative errors) of cardinality estimations for 500
randomly generated datasets per cardinality (the x -axis). Tested cardinalities
are from 500 to 200, 000 in every 500 .
Fig. 3. The relative error of the cardinality estimation of one randomly
generated dataset per cardinality (the x -axis). Tested cardinalities are from
500 to 200, 000 in every 500 .
Fig. 4. The histogram of the cardinality estimations of 500 randomly
generated datasets for cardinality 1, 000 .
3
Fig. 5. The histogram of the cardinality estimations of 500 randomly
generated datasets for cardinality 50, 000 .
IV.
requires no regeneration of vectors M and it can be applied
directly to the existing vectors. Therefore, switching to
LogLog-β requires a very small effort and produces
comparable or better results.
The idea of replacing m2 by m(m − z) can be applied to
other cardinality estimation algorithms as well. For example,
we apply this idea to an unbiased optimal cardinality
estimation algorithm proposed and studied in Lumbroso[13].
The algorithm, like HyperLogLog, performs well for very
large cardinalities but depends on Linear Counting and bias
corrections for small and pre-asymptotical cardinalities. The
core algorithm in Lumbroso[13], without Linear Counting and
bias correction, is described as follows:
Let S , M , m , and p be the same as described in
HyperLogLog algorithm, but h be a hash function with
hash value in interval (0, 1),
More discussions
The formula (2) of algorithm LogLog-β requires two major
changes compared to HyperLogLog Raw (1), and these two
changes make the formula capable of covering the entire range
of cardinalities with comparable or better accuracy. Replacing
m2 with m(m − z), makes no difference for the very large
cardinalities but has significant impact on the estimates of
small and mid-range cardinalities. It forces the estimates to
roughly align with the real cardinalities. The other change,
adding a term β (m, z) to the denominator, further corrects the
remaining error/bias. The combination of the two changes
makes formula (2) work more accurately for the entire range
of cardinalities. Both changes are essential, and the second
change could have many variations and forms.
Lumbroso's Core Algorithm:
1) Initialize M :
Set M [i] = 1 for i = 1, …, m
2) Generate M :
For each s∈S , set y = h(s) , i = ⌊ym⌋ + 1 , and
M [i] = min (M [i], ym − ⌊ym⌋)
3) Apply the core formula:
E = m(m−1)
m
∑ M [i]
(3)
i=1
It is proved by Lumbroso[13] that the above algorithm
yields optimal estimation for very large cardinalities. To make
it work for small or mid-range cardinalities we substitute 1 in
formula (3) with z which results in a new formula:
E = m(m−z)
m
∑ M [i]
(4)
i=1
Fig. 6. The histogram of the cardinality estimations of 500 randomly
generated datasets for cardinality 100, 000 .
where z is the number of M [i] with value 1. There is no need
to add a correction term to the denominator in this case. The
estimations provided by formula (4) are strikingly accurate for
all ranges of cardinalities, especially for the small and
mid-range cardinalities. In Figure 7 we show some test results
on this new algorithm, called MMV (Mean of Minimum
Values).
Formula (2) of LogLog-β and formula (4) of MMV,
derived from different approaches (counting vs. ordering), are
deeply related (detailed study is in a separate paper). The term
2−M [i] of (2) could be regarded as an approximation of the
term M [i + 1] of (4). For a given m both algorithms provide
comparable accuracies, but LogLog-β requires less memory.
In terms of implementation and computation complexity,
LogLog-β offers more advantages: one formula for the entire
range of cardinalities, a much simpler process flow, and no
bias correction and lookup tables. Furthermore, for any
existing
implementations
of
HyperLogLog
or
HyperLogLog++ (without sparse representation), LogLog-β
4
Fig. 7. Cardinality estimations of 500 randomly generated datasets for each
cardinality in every 500 from 500 to 200, 000 .
References
[1] N. Alon, Y. Matias, and M. Szegedy, “The space
complexity of approximating the frequency moments,” in
Journal of Computer and System Sciences 58 (1999),
137-147.
[2] S. Acharya, P. B. Gibbons, V. Poosala, and S.
Ramaswamy. “The aqua approximate query answering
system,” in SIGMOD, pages 574-576, 1999.
[3] P. Brown, P. J. Haas, J. Myllymaki, H. Pirahesh, B.
Reinwald, and Y. Sismanis. “Toward automated
large-scale information integration and discovery,” in
Data Management in a Connected World, pages 161-180,
2005.
[4] K. Beyer, P. J. Haas, B. Reinwald, Y. Sismanis, and R.
Gemulla. “On synopses for distinct-value estimation
under multiset operations,” in Proceedings of the 2007
ACM SIGMOD international conference on Management
of data, pages 199-210. ACM, 2007.
[5] R. Cohen, L. Katzir, A. Yehezkel. “A Unified scheme for
generalizing cardinality estimators to sum aggregation,”
in Inf. Process. Lett.(2014).
[6] M. Durand and P. Flajolet, “LogLog counting of large
cardinalities,” in Annual European Symposium on
Algorithms (ESA03) (2003), G. Di Battista and U. Zwick,
Eds., vol. 2832 of Lecture Notes in Computer Science, pp.
605-617.
[7] C. Estan, G. Varghese, and M. Fisk. “Bitmap algorithms
for counting active flows on high-speed links,” in
IEEE/ACM Trans. Netw., 14(5):925-937, Oct. 2006.
[8] P. Flajolet and G. Martin. “Probabilistic counting
algorithms for data base applications,” in Journal of
Computer and System Sciences, 31(2):182-209, 1985.
[9] P. Flajolet, É. Fusy, O. Gandouet, and F. Meunier.
“Hyperloglog: The analysis of a near-optimal cardinality
estimation algorithm,” in Analysis of Algorithms
(AOFA), pages 127-146, 2007.
[10] S. Ganguly. “Counting distinct items over update
streams,”
in
Theoretical
Computer
Science,
378(3):211-222, 2007.
[11] S. Heule, M. Nunkesser, and A Hall. “HyperLogLog in
Practice: Algorithmic Engineering of a State of The Art
Cardinality Estimation Algorithm,” Google, Inc.
[12] D. M. Kane, J. Nelson, and D. P. Woodruff. “An optimal
algorithm for the distinct elements problem,” in PODS,
pages 41-52, 2010.
[13] J. Lumbroso. “An optimal cardinality estimation
algorithm based on order statistics and its full analysis,”
in DMTCS Proceedings 01 (2010), 489-504.
[14] C. R. Palmer, P. B. Gibbons, and C. Faloutsos. “Anf: a
fast and scalable tool for data mining in massive graphs,”
in SIGKDD, pages 81-90, 2002.
[15] P. G. Selinger, M. M. Astrahan, D. D. Chamberlin, R. A.
Lorie, and T. G. Price. “Access path selection in a
relational database management system,” in SIGMOD,
pages 23-34, 1979.
[16] K.-Y. Whang, B. Vander-Zanden, and H. Taylor. “A
linear-time probabilistic counting algorithm for database
applications,” in ACM Transactions on Database Systems
15, 2 (1990), 208-229.
[17] Z. Bar-Yossef, T. Jayram, R. Kumar, D. Sivakumar, and
L. Trevisan. “Counting distinct elements in a data
stream,” in Randomization and Approximation
Techniques in Computer Science, pages 1-10. Springer,
2002.
5
| 8cs.DS
|
THE SALLY MODULES OF IDEALS: A SURVEY
arXiv:1612.06261v1 [math.AC] 19 Dec 2016
WOLMER V VASCONCELOS
Abstract. The Sally module of a Rees algebra B relative to one of its Rees subalgebras A is a construct that can be used as a mediator for the trade-off of cohomological
(e.g. depth) information between B and the corresponding associated graded ring
for several types of filtrations. While originally devised to deal with filtrations of
finite colength, here we treat aspects of these developments for filtrations in higher
dimensions as well.
Key Words and Phrases: Castelnuovo regularity, extremal Rees algebra, Hilbert function,
Rees algebra, relation type, Sally module.
Contents
1. Introduction
2. General Properties of Sally Modules
2.1. Dimension, depth and reduction number
2.2. Veronese subrings
2.3. The fiber of the Sally module
3. Dimension, Multiplicity and Euler Characteristics
3.1. The Hilbert series and dimension
3.2. Cohen–Macaulayness of the Sally module
3.3. Change of rings for Sally modules
3.4. Special fiber and reduction number
3.5. Multiplicity and reduction number
3.6. Sally questions
4. Equimultiple Ideals
4.1. Extended Hilbert function
4.2. Ideals of positive dimension
4.3. Cohen-Macaulayness
4.4. Ideals of positive dimension: non Cohen-Macaulay rings
4.5. j-multiplicities
5. Ideals of Dimension One
5.1. Generic complete intersections
5.2. Dimensions two and three
6. This and That
6.1. Multiplicity one
6.2. Normal filtration
6.3. Monomial ideals
7. Current issues
8. List of definitions and notation
References
1
2
3
3
5
5
5
6
10
11
12
13
14
14
15
16
17
18
19
20
20
24
24
25
25
26
28
28
29
2
WOLMER V VASCONCELOS
1. Introduction
The aim of this survey is to review some of the properties of Sally modules of ideals.
For a Noetherian local ring (R, m) and an m-primary ideal I, these structures were
introduced in [73] to give a quick access to some results of Judith Sally (see Section 3.6)
connecting the depth properties of associated graded rings grI (R) to extremal values
of their multiplicity e0 (I). They have since been used to refine this connection taking
into account the roles of the higher Hilbert coefficients e1 (I) and e2 (I). Here we intend
to visit the literature to discuss a representative set of these results with emphasis on
some of the techniques. The more novel material are sought-after generalizations to a
more general theory that applies to non Cohen-Macaulay rings or to filtrations defined
by ideals of positive dimension. The latter are rather experimental in character.
Let (R, m) be Noetherian local ring of dimension d > 0. Sally modules are defined
in the context of Rees algebras and some of their finite extensions. One of its main
classes arises as follows. Let J be an ideal of R and let F = {Ij , j ≥ 0}, I0 = R,
be a multiplicative
filtration of ideals. Consider the Rees algebras A = R[Jt] and
P
j
B =
j≥0 Ij t = R ⊕ B+ . Suppose that B is a finite extension of A. The Sally
module SB/A of B relative to A is defined by the exact sequence of finitely generated
A–modules
(1)
0 → I1 A −→ B+ [+1] =
M
Ij tj−1 −→ SB/A =
j≥1
M
Ij+1 /I1 J j → 0.
j≥1
The assumption is that Ij+1 = JIj for j ≫ 1, so that in particular J is a reduction of
I1 .
Some of the problems about these structures concern the following issues:
• [Reduction number] The reduction number of B relative to A is defined to be
X
ABj }.
r(B/A) = inf{n | B =
j≤n
•
•
•
•
•
•
We observe that SB/A = 0 if and only if r(B/A) ≤ 1. This condition is often
referred as minimal multiplicity.
[Dimension] What are the possible values of dim SB/A ? Note that for each
x ∈ J, localizing at Rx gives that SA (B)x = 0, so that SA (B) is annihilated
by a power of J and therefore dim SA (B) ≤ dim R.
[Multiplicity] What are the possible values of multiplicity of SB/A espressed in
Hilbert coefficients?
[Cohen–Macaulayness] When is SB/A Cohen–Macaulay?
[Regularity] There are known relationships ([71]) between the Castelnuovo regularity of the Rees algebra and that of the associated graded ring. We want to
relate/extend comparisons to the Castelnuovo regularity of SB/A (if R is not
Cohen-Macaulay).
We would like to deal with these issues by relating them to some Hilbert functions associated to gr(B). One additional difficulty lies with the character of
the ring R itself. This arises because the coefficients of these functions may be
positive or negative, often depending on whether R is Cohen–Macaulay or not.
[Existence] Which modules are Sally modules?
THE SALLY MODULES OF IDEALS: A SURVEY
3
• [Independence] How independent of J are the properties of SJ (I)?
• [a-invariant] Review its role
• [Equations of the ideal] What are the relationships between SJ (I) and the
defining equations of R[IT]? What if SJ (I) = 0?
• [Non-primary ideals] Equimultiple, generically complete intersection, monomial
Acknowledgments: We are thankful to several authors whose work motivated the
writing of this survey. We apologize for unfinished discussions and invite readers to
comment and suggestions. We are particularly grateful to Alberto Corso, Laura Ghezzi,
Jooyoun Hong, Shiro Goto, Maria Evelina Rossi, Maria Vaz Pinto, Judy Sally, Aron
Simis and Giuseppe Valla.
2. General Properties of Sally Modules
2.1. Dimension, depth and reduction number. One of its most useful properties
as a conduit of cohomological information between a Rees algebra and the associated
graded ring is the following ([73]). Throughout this section, (R, m) is a Noetherian
local ring of dimension d and I is an m-primary ideal. Whenever required will assume
that R has an infinite residue field.
Cohen–Macaulay rings offer the most opportunities to exploit this notion. In the
following (R, m) is a Cohen–Macaulay local ring, I is an m-primary ideal, J or Q denote
minimal reductions of I, and A and B are Rees algebras as above.
Proposition 2.1. Suppose R is a Cohen–Macaulay local ring of dimension d ≥ 1. If
J is an ideal generated by a system of parameters then the following hold:
(1) If SB/A = 0, then gr(B) is Cohen–Macaulay.
(2) If SB/A 6= 0, then mR[JT ] is its only associated prime.
(3) If SB/A 6= 0 then dim SB/A = d.
(4) SB/A is Cohen–Macaulay if and only if depth gr(B) ≥ d − 1.
(5) If depth gr(B) < d then depth SB/A = depth gr(B) + 1.
(6) If SB/A 6= 0, then reg(SB/A ) = reg(gr(B)) = reg(B).
(7) Furthermore if SB/A is Cohen-Macaulay, then reg(SB/A ) = reg(B) = rA (B) +
1.
Proof. The assertions follow from computing the cohomology (see [43], [45], [71]) with
respect to the maximal homogeneous ideal of A in the following exact sequences of
finitely generated A–modules:
0 → I1 A∗ −→ A∗ −→ grJ (R) ⊗ R/I1 ≃ R/I1 [T1 , . . . , Tr ]∗ → 0,
0 → B+ −→ B −→ R∗ → 0,
0 → I1 A∗ −→ B+ [+1] −→ SB/A → 0,
0 → B+ [+1] −→ B −→ gr(B) → 0,
where the modules tagged with ∗ are Cohen–Macaulay. Along with the shifts, they
allow for a measure of pivoting in all the depth calculations (and their grading in the
calculation of Castelnuovo regularity).
For the proof of (2) and (3), in the third sequence above,
0 → I1 A −→ B+ [+1] −→ SB/A → 0,
4
WOLMER V VASCONCELOS
I1 A is a maximal Cohen–Macaulay A-module, while B+ is torsion-free over A, that
is, it has the condition S1 of Serre. It follows that if SA (B) 6= 0 then SA (B) has
the condition S1 of Serre. Therefore p = mA is its only associated, in particular
dim SA (B) = d.
Additionally, if d ≥ 2 and B has the condition S2 of Serre then so does B+ , and
consequently SA (B) has the condition S2 as well.
✷
One of our goals is to examine which parts of this formalism survive in case J is an
ideal of positive dimension, and/or R is not Cohen–Macaulay.
Corollary 2.2. ([27, Proposition 2.2]) Let A, B and p be as above. Then
s0 (SA (B)) = λ(SA (B)p ).
Furthermore, if s0 (SA (B)) = 1, then mSA (B) = 0.
Proof. The general assertion is clear since p is the unique associated associated prime
of SA (B) and A/p has multiplicity 1. On the other hand, if λ(SA (B)p ) = 1, we must
have pSA (B)p = 0 and therefore pSA (B) = 0 since this module is either trivial or p
would be its only associated prime, in which case s0 (SA (B)) ≥ 2.
✷
Corollary 2.3. Let (R, m) be a Noetherian local ring, let I be an m-primary ideal
and J one of its minimal reductions. If for some integer s ≥ 1, JI s = IJ s , then
dim SJ (I) = 0. Moreover if R is Cohen–Macaulay then I 2 = JI.
Proof. The assertion means that the Sally module components I n+1 /J n I = 0 for n ≫ 0.
Let n be an integer of the form n = a(s − 1) + 1 ≥ rJ (I). Then
I n+1 = JI a(s−1)+1 = JI s I (a−1)(s−1) = J s−1 JI (a−1)(s−1)+1 = · · · = J a(s−1)+1 I = IJ n .
The last assertion follows from Proposition 2.1(3).
Alternatively, the assumption is that SJ (I) is annihilated by a power of (JT). Since
by construction SJ (I) is also annihilated by a power of J, the assertion follows.
✷
Example 2.4. Let us consider some examples.
• R = k[x, y]/x(x, y)2 , J = (y), I = (x, y). Then I 2 = (x2 , xy, y 2 ) 6= (xy, y 2 ) =
JI mod x(x, y)2 , but I 2 J = J 2 I.
• For positive integers d > r > 1, give an example of a ring with dim R = d with
a Sally module with dim SJ (I) = r.
• Adding variables to J and I, what happens to the Sally modules? Set S′ = R[x],
L = (J, x), Q = (I, x). Then SQ (L) = SJ (I)[x].
• Let B = k[x, y, z], I = (x, y, z)3 and the system of parameters (x3 , y 3 , z 3 ).
View I as the ideal of the subring B0 of B generated by the forms of degree
≥ 3. Now set Q0 = (x3 , y 3 , z 3 )B0 and Q = Q0 B. We have the equality
SQ0 (I) = SQ (I) 6= 0. It follows that although B0 is not Cohen-Macaulay, the
Sally module SQ0 (I) is Cohen-Macaulay.
• Let (R, m) be a Cohen–Macaulay local ring and let p be an ideal of codimension
g such that Rp is not a regular local ring. If J = (a1 , . . . , ag ) ⊂ p is a regular
sequence, then for the link I = J : p, I 2 = JI ([9, Theorem 2.3]). In particular
S(J; I) = 0.
THE SALLY MODULES OF IDEALS: A SURVEY
5
2.2. Veronese subrings. Let I be an ideal with a reduction J = (a1 , . . . , ad ), I r+1 =
JI r . For q > 0, Jq = (aq1 , . . . , aqd ) is a reduction of I q , then ([3], [44, Corollary 2.6], [53,
Theorem 3.10])
r(I q ) ≤ max⌈1 +
r(I) − 1
, 2⌉,
q
where ⌈ ⌉ is the integral part function. In particular, for q ≫ 0, the Sally module
SB(q) /A(q) is a standard graded module.
Proposition 2.5. If R is Cohen-Macaulay then for q larger than the postulation number of grI (R), e0 (A(q) ; B(q) ) = e2 (I).
For a proof see Corollary 3.10(b). It would be interesting to compare two such Sally
modules with respect to their dimensions, multiplicities and depth.
2.3. The fiber of the Sally module. The Sally module SB/A is a module over
A = R[JT]. An important role here is that of the module
FA (B) = SB/A ⊗A (A/(JT)) =
r
M
In+1 /JIn ,
n≥1
which we call the its special fiber or simply the Sally fiber of SB/A . Here are some
properties of FA (B):
(1) There is a natural action of B on FA (B): For x = bs Ts ∈ Is Ts ,
x · (In+1 /JIn ) = xIn+1 + JIn+s /JIn+s ⊂ In+s+1 /JIn+s .
For the filtration In = I n , I r+1 ⊂ JI r , in particular r = rA (B) ≤ λ(FA (B))+1.
In the case of equality I n+1 /JI n ≃ R/m for all 1 ≤ n ≤ r − 1.
We conjecture that in this case all I n are Ratliff-Rush closed! Check in [13].
This can be further refined by considering the action of Is Ts on FA (B)⊗R/m =
F and benefit from Nakayama’s Lemma: rA (B) ≤ λ(F ) + 1.
n its Ratliff-Rush closure. There is a natural
(6) Let I n be an I-adic filtration and If
mapping of Sally modules
(2)
(3)
(4)
(5)
e n−1 .
I n /IJ n−1 7→ Ifn /IJ
If I = Ie this mapping is an embedding that is an isomorphism for n ≫ 0.
(7) In this case the Hilbert polynomials agree so that in particular s0 (SB/A ) =
).
s0 (SB/A
e
3. Dimension, Multiplicity and Euler Characteristics
We begin to develop the basic metrics to examine Sally modules with a focus on the
determination of multiplicities.
6
WOLMER V VASCONCELOS
3.1. The Hilbert series and dimension. A basic calculation refines the relationship
between the multiplicity and the dimension of the Sally module.
Proposition 3.1. Let (R, m) be a Noetherian local ring of dimension d, I an mprimary ideal and Q one of its minimal reductions. Then dim SQ (I) = d if and only if
sQ (I) 6= 0.
More generally we have observed that dim SQ (I) ≤ d. This means that its Hilbert
polynomial
d−1
X
n+i−1
i
n+1
n
(−1) si
λ(I
/IQ ) =
,
d−i−1
0
where we denote sQ (I) = s0 . Finally observe
Proposition 3.2. If SQ (I) 6= 0 then
dim SQ (I) = d − min{i | si 6= 0}, or 0.
Proposition 3.3. Let (R, m) be a Noetherian local ring of dimension d, I an mprimary ideal and Q one of its minimal reductions. Suppose that the corresponding
Sally module SQ (I) has dimension d. Then its multiplicity satisfies
sQ (I) ≤ e1 (I) − e1 (Q) − e0 (I) + λ(R/I).
Proof. Let SQ (I) =
exact sequences
L
n≥1 I
n+1 /IQn
be the Sally module of I relative to Q. From the
0 → I n+1 /IQn −→ R/IQn −→ R/I n+1 → 0
0 →
Qn /IQn
−→ R/IQn −→
R/Qn
→ 0
we obtain that
λ(I n+1 /IQn ) = −λ(R/I n+1 ) + λ(R/Qn ) + λ(Qn /IQn ).
(2)
For n ≫ 0, we have
n+d−1
λ(I
/IQ ) = sQ (I)
+ lower degree terms
d−1
n+d
n+d−1
n+1
λ(R/I
) = e0 (I)
− e1 (I)
+ lower degree terms.
d
d−1
n+1
n
Since e0 (Q) = e0 (I), for n ≫ 0 we have
n+d−1
n+d−2
λ(R/Qn ) = e0 (Q)
− e1 (Q)
+ lower degree terms
d
d−1
n+d
n+d−1
n+d−1
n+d−2
= e0 (Q)
−
− e1 (Q)
−
+ ···
d
d−1
d−1
d−2
n+d
n+d−1
n+d−1
= e0 (I)
− e0 (I)
− e1 (Q)
+ lower degree terms
d
d−1
d−1
THE SALLY MODULES OF IDEALS: A SURVEY
7
Let G = grQ (R) ⊗ R/I. Then for n ≫ 0 we have
n+d−1
n
n
λ(Q /IQ ) = e0 (G)
+ lower degree terms.
d−1
It follows that
sQ (I) = e1 (I) − e0 (I) − e1 (Q) + e0 (G).
But G is a homomorphic image of a polynomial ring R/I[T1 , . . . , Td ], so that e0 (G) ≤
λ(R/I), and we obtain the desired upper bound for sQ (I).
✷
In a minor refinement, write (2) as
(3) λ(I n+1 /IQn ) = −λ(R/I n+1 ) + λ(R/Qn+1 ) − λ(Qn /Qn+1 ) + λ(Qn /IQn ).
We now match the Hilbert coefficients:
0 = e0 (I) − e0 (Q)
sQ (I) = s0 = e1 (I) − e1 (Q) + e0 (Q) + e0 (G)
..
.
si = ei+1 (I) − ei+1 (Q) + ei (Q) + ei (G), i ≥ 1.
To apply Proposition 3.2 is simpler when ei (G) = 0 for i ≥ 1–and we treat two cases
here. Another simplification is when the values of ei (Q), for i ≥ 1, are independent of
I (Buchsbaum rings, to check).
This highlights the fact that, unlike the Cohen–Macaulay case, several properties of
SQ (I) [dimension, multiplicity] may depend on Q. Meanwhile we have
Corollary 3.4. Let (R, m) be a Cohen-Macaulay local ring of dimension d ≥ 1, let I
be an m-primary ideal and let Q be one of its minimal reductions generated by a regular
sequence. Then
(1) The Hilbert function of SQ (I) is independent of Q.
(2) If d = 1, rQ (I) is independent of Q. More generally, the assertion holds if
h(t)
depth grI (R) ≥ d − 1. More precisely, if (1−t)
d is the Hilbert series of SQ (I),
then rQ (I) = deg h(t).
Proof. (1) follows directly from the formula
λ(I n+1 /IQn ) = −λ(R/I n+1 ) + λ(R/Qn ) + λ(Qn /IQn ),
as the last two terms are respectively
d+n−1
λ(R/Q ) = e0 (I)
,
d
d+n−1
n
n
λ(Q /IQ ) = λ(R/I)
.
d−1
n
(2) Let us consider a general observation first. From the definition of SQ (I),
0 → IR[QT] −→ IR[IT] −→ SQ (I) → 0,
8
WOLMER V VASCONCELOS
we see that SQ (I) is a module over R[QT] and as such has the condition S1 of Serre.
Considering the natural presentation of the polynomial ring A = R[t1 , . . . , td ] →
R[QT], we view SQ (I) as a finite A-module.
Of course the action of the variables depend on Q as does L the annihilator of SQ (I)
as an R-module. Nevertheless as an A/LA-module SQ (I) keeps the condition S1 and
therefore there is a form of degree 1 that induces an injective homomorphism on SQ (I).
Consequently its Hilbert function is monotonically increasing.
Let d = 1 now and suppose SQ (I) 6= 0 [otherwise there is no difficulty]: With
A = R[t] the action of
t : Sn−1 7→ Sn
n+1
n
is surjective if I
= QI thus n ≥ rQ (I). Since this means that λ(I n /IQn−1 ) =
n+1
n
λ(I
/IQ ), a condition that is independent of Q, This proves the assertion.
SimiIar observations can be made if depth grI (R) ≥ d−1 since this means that SQ (I)
is Cohen-Macaulay for each Q and therefore the reduction number rQ (I) is given by
the degree of the h-polynomial minus one.
✷
A general result that also addresses the case when SQ (I) is Cohen-Macaulay is [70,
Theorem 1.2].
Corollary 3.5. If R is Cohen-Macaulay then
sQ (I) = e1 (I) − e0 (I) + λ(R/I),
in particular the dimension, the depth and the multiplicity of SQ (I) are independent of
Q.
Proof. The assertions about the dimension and the depth follow from Proposition 2.1.
✷
Remark 3.6. Suppose F = {In } is a filtration
Lso that the corresponding Rees algebra
B is finite over A = R[QT]. Let SQ (B) = n≥1 In+1 /I1 Qn be the associated Sally
module. The same calculation yield the formulas
sQ (F) ≤ e1 (F) − e1 (Q) − e0 (Q) + λ(R/I1 ),
and if R is Cohen–Macaulay
sQ (F) = e1 (F) − e0 (Q) + λ(R/I1 ).
Let us consider a mild generalization of these formulas.
Corollary 3.7. Let (R, m) be a Noetherian local ring of dimension d, I an m-primary
ideal and Q one of its minimal reductions. Suppose that dim SQ (I) = d. Let ϕ is the
matrix of syzygies of Q. If Q is generated by a d-sequence and I1 (ϕ) ⊂ I, then
sQ (I) = e1 (I) − e1 (Q) − e0 (I) + λ(R/I).
Proof. Let G = grQ (R) ⊗ R/I. By the proof of Theorem 3.3, we have
sQ (I) = e1 (I) − e0 (I) − e1 (Q) + e0 (G).
Set S = R[T1 , . . . , Td ]. Since Q is generated by a d-sequence, the approximation
complex
0 → Hd (Q) ⊗ S[−d] → · · · → H1 (Q) ⊗ S[−1] → H0 (Q) ⊗ S → grQ (R) → 0
THE SALLY MODULES OF IDEALS: A SURVEY
9
is acyclic ([33, Theorem 5.6]). By tensoring this complex by R/I, we get the exact
sequence
φ
H1 (Q) ⊗ R/I ⊗ S[−1] −→ R/I ⊗ S −→ G → 0.
Since I1 (ϕ) ⊂ I, the mapping φ is trivial so that e0 (G) = λ(R/I).
✷
Remark 3.8. One instance when I1 (ϕ) ⊂ I occurs if R is unmixed and I is integrally
closed, according to [54, Proposition 3.1].
Corollary 3.9. Suppose (R, m) is an unmixed Buchsbaum local ring of dimension d.
Let I be an m-primary integrally closed ideal and Q is one of its minimal reductions.
If dim SQ (I) = d then the dimension and multiplicity of the Sally modules of I are
independent of the chosen minimal reduction.
Proof. By assumption SQ (I) 6= 0. Since e1 (Q) is independent of Q for Buchsbaum rings
([30]), SQ (I) is unchanged across all minimal reductions of I. In particular dim SQ (I)
is also unchanged.
✷
The following classical inequalities for the Hilbert coefficients of primary ideals arises
from this formula (see [13] for a detailed discussion).
Corollary 3.10. Let (R, m) be a Cohen–Macaulay local ring of dimension d and let I
be a m-primary ideal. Then
(a) Northcott inequality [50] e1 (I) − e0 (I) + λ(R/I) ≥ 0.
(b) Narita inequality [49]
e2 (I) ≥ 0.
Proof. For (a), one invokes the fact that for R Cohen-Macaulay, e1 (I) − e0 (I) + λ(R/I)
is the multiplicity of SQ (I). For (b), we may assume that d = 2 and using a trick of J.
Lipman to evaluate this multiplicity on I q for q larger than the postulation number of
the Hilbert function of grI (R).
We first consider the case for d = 2 and without great loss of generality take In = I n .
Let
n+1
n
n
λ(R/I ) = e0
− e1
+ e2 , n ≫ 0
2
1
be the Hilbert polynomial of I. Let q be an integer greater than the postulation number
for this function. We have
n+1
n
qn
q n
λ(R/I ) = λ(R/(I ) ) = f0
− f1
+ f2
2
1
qn + 1
qn
= e0
− e1
+ e2 , n ≫ 0.
2
1
Comparing these two polynomials in n, gives
f0 = q 2 e0
1
1
f1 = qe1 + q 2 e0 − qe0
2
2
f2 = e2 .
Since the multiplicity s0 (I q ) of the Sally module of I q is given by
s0 (I q ) = f1 − f0 + λ(R/I q ),
10
WOLMER V VASCONCELOS
which by the equalities above yields e2 (I) = s0 (I q ), the multiplicity of the Sally module
of I q . Thus e2 (I) is nonnegative and vanishes if and only if the reduction number of I q
is less than or equal to 1. In dimension d > 2, we reduce R and I modulo a superficial
element.
✷
3.1.1. Maximal ideals. The dimension of Sally modules over non-Cohen-Macaulay local
rings is hard to ascertain. We state a result of A. Corso ([7]) for maximal ideals.
Theorem 3.11. [7, Theorem 2.1] Let (R, m) be a local Noetherian ring of dimension
d > 0 with infinite residue field, and let Q be a minimal reduction of m.
(a) The Sally module SQ (m) has dimension d if and only if e1 (m)−e0 (m)−e1 (Q)+1
is strictly positive. Otherwise, its dimension is d−inf{i | 1 ≤ i ≤ d−1, ei+1 (m)−
ei (Q) − ei+1 (Q) 6= 0} or 0.
(b) If R is a Buchsbaum ring, the Sally module SQ (m) has dimension d or 0.
Remark 3.12. For m-primary ideals I 6= m, a similar description is not known, even
for Buchsbaum rings.
3.2. Cohen–Macaulayness of the Sally module. As a motivation, it is worthwhile
to point out one useful aspect of these modules. Under the conditions of (2.1), if I1
is m–primary there is a numerical test for the Cohen-Macaulayness of SB/A ([42]; see
also [39]). It does not always require that R be Cohen-Macaulay.
Theorem 3.13 (Huckaba test). If R is Cohen-Macaulay, the Sally module SB/A is
Cohen–Macaulay if and only if
X
e1 (gr(B)) =
λ(Ij /QIj−1 ).
j≥1
Proof. Its brief proof, as in [39], goes as follows. First note that SB/A is an A–module
annihilated by a power of Q, from which it follows that the generators of QT is a
system of parameters. In the case of an m–primary ideal Q, the interpretation of this
formula is
e0 (SB/A ) = λ(SB/A ⊗ R[QT]/(QT)).
In other words, χ1 (QT; SB/A ) = 0, and therefore SB/A is Cohen–Macaulay by Serre’s
Theorem ([4, Theorem 4.7.10]) on Euler’s characteristics.
✷
Here is an important case that does not use this test.
Theorem 3.14. [55, Corollary 3.8]. Let (R, m) be a Cohen–Macaulay local ring of
dimension d. and let I be an m-primary ideal. If
e0 (I) = λ(I/I 2 ) + (1 − d)λ(R/I) + 1
then depth (G) ≥ d − 1. In particular all the Sally modules of I are Cohen–Macaulay.
Remark 3.15. The equations of the ideal I ([72]) are the the ideal of relations of a
presentation
ϕ
0 → L −→ C = R[T1 , . . . , Tn ] −→ R[IT] → 0.
L
L = j≥1 Lj is a graded ideal whose properties are independent of the presentation.
For instance, the relation type of I, rel(I), is the maximum degree of a minimal set of
homogeneous generators of L. If R is a Cohen–Macaulay local ring and I in m-primary,
THE SALLY MODULES OF IDEALS: A SURVEY
11
then the Sally module SJ (I) impacts the relation type in a number of ways, for instance
if SJ (I) is Cohen–Macaulay then rel(I) ≤ rJ (I) + 1, according to [70, Theorem 1.2].
Several questions can be raised if I is not m-primary. We will attempt to do this
later.
3.2.1. Dimension one. Let (R, m) be a Cohen-Macaulay local ring of dimension one.
By Proposition 2.1(iv), for each m-primary ideal I the Sally module S = SQ (I) is
Cohen-Macaulay, and therefore by Theorem 3.13
X
e1 (I) =
λ(Ij /QIj−1 ).
j≥1
Proposition 3.16. Let (R, m) be a Cohen–Macaulay local ring of dimension 1 and let
I be an m-primary ideal. If Q = (a) is a minimal reduction of I then the corresponding
Sally module S = SQ (I) has the following properties. Setting L = ann(I/Q),
(i) S is Cohen-Macaulay.
(ii) If s = r(I) ≥ 2, then
e1 (I) = λ(R/L) +
s−1
X
λ(I j+1 /aI j ).
j=1
(iii) If ν(I) = 2, then
e1 (I)
.
λ(R/L)
(iv) If ν(I) = r(I) = 2, then e1 (I) ≤ 2 · λ(R/L).
r(I) ≥
Proof. (i), (ii): Since depth grI (R) ≥ d − 1, S is Cohen-Macaulay.
(iii): If I = (a, b), I j+1 /QI j is cyclic and annihilated by L, so
e1 (I) ≤ r(I) · λ(R/L).
(iv): is a special case of (iii).
✷
3.3. Change of rings for Sally modules. Let (R, m) be a Noetherian local ring of
dimension d ≥ 1 and let f : R → S be a finite morphism. Suppose S is a Cohen–
Macaulay ring locally of dimension d. For each m-primary ideal I with a minimal
reduction Q, for each maximal ideal M of S the ideal ISM is MSM -primary and QSM
is a minimal reduction. We define the Sally module of IS relative to QS in the same
manner,
M
SQS (IS) =
I n+1 S/IQn S.
n≥1
SQS (IS) is a finitely generated graded grQ (R)-module. Its localization at M gives
SQSM (ISM ).
Consider the natural mapping
M
M
ϕf : S ⊗R SQ (I) =
I n+1 /IQn ⊗R S 7→ SQS (IS) =
I n+1 S/IQn S.
n≥1
n≥1
12
WOLMER V VASCONCELOS
ϕf is a graded surjection from S ⊗R SQ (I) onto SQS (IS), which combined with a
presentation Rb → S → 0, b = ν(S), gives rise to a homogeneous surjection of graded
modules
[SQ (I)]b 7→ SQS (IS).
The following takes information from this construction into Proposition 2.1. It is a
factor in our estimation of several reduction number calculations, according to [23].
Theorem 3.17 (Change of Rings Theorem). Let R, S, I, Q be as above. If S is CohenMacaulay then
(1) If dim SQ (I) < d, then SQS (IS) = 0.
(2) If s0 (IS) 6= 0, then dim SQ (I) = d.
(3) If dim SQ (I) = d, then s0 (IS) ≤ b · s0 (I).
We note that s0 (IS) is a multiplicity relative to R. It is a positive summation of
local multiplicities, so each of these is also bounded.
Proof. All the assertions follow from the surjection [SQ (I)]b → SQS (IS) and the vanishing property of Sally modules over Cohen-Macaulay rings (Proposition 2.1.2).
✷
Corollary 3.18. s0 (IS) ≤ b · s0 (I).
Remark 3.19. It would be good to prove that the condition (3.17.2) is actually an
equivalence.
Question 3.20. Let S = k[x1 , . . . , xd ] and let G be a finite group of k-automorphisms
of S. Set R = SG .
[True or false?] For any ideal I of R of finite colength SQ (I) is [locally] either 0
or has dimension d.
If Q is a minimal reduction of I, QS is a minimal reduction of IS. Suppose I 2 S =
QIS.
Extra remarks on change of rings. Suppose (R, d, m, I, Q) is as above and ϕ : R → S
is an injective finite morphism.
• We would like to prove dim SQ (I) = dim SQS (IS). Set C = S[QST] and
consider the sequence of natural surjective homomophisms of B modules
S ⊗R SQ (I) = S ⊗R B ⊗B SQ (I) → C ⊗B SQ (I) → SQS (IS).
• Claim 1. dim C ⊗B SQ (I) = dim SQ (I): Let P be an associated prime ideal
of SQ (I), as a B-module, with dim B/P = dim SQ (I). Localize at P . Let
L ⊂ P BP be a system of parameters contained in the annihilator of SQ (I)P .
It follows easily that CP /LCP ⊗B SQ (I)P 6= 0, and therefore C ⊗B SQ (I) has
dimension dim B/P , as desired.
• Claim 2. We will argue that dim SQ (I) = dim SQS (IS), as B-modules.
3.4. Special fiber and reduction number. Following [12] we study relationships
between r(I) and the multiplicity f0 (I) of the special fiber F (I) = R[IT] ⊗ R/m of the
Rees algebra of I. Let (R, m) be a Cohen–Macaulay local ring of dimension d ≥ 1, and
let I be an m-primary ideal, let Q be a minimal reduction of I and let SQ (I) denote the
Sally module of I relative to Q. Suppose I is minimally generated by s elements. We
THE SALLY MODULES OF IDEALS: A SURVEY
13
can write I = (Q, b1 , . . . , bs−d ). Consider the exact sequence introduced in [12, Proof
of 2.1]
ϕ
R[QT] ⊕ R[QT][−1]s−d −→ R[IT] −→ SQ (I)[−1] → 0,
P
where ϕ is the map defined by ϕ(f, a1 , . . . , as−d ) = f + j ai bj T. Tensoring this
sequence with R/m yields the exact sequence
(4)
ϕ
F (Q) ⊕ F (Q)[−1]s−d −→ F (I) −→ SQ (I)[−1] ⊗ R/m → 0.
We make some observations about the relationship between s0 (I) and rQ (I) arising
from this complex in special cases:
• The exact sequence gives
f0 (I) ≤ deg SQ (I) + ν(I) − d + 1 ≤ deg SQ (I) + ν(I) − d + 1.
• If mSQ (I) = 0, f0 (I) ≤ s0 (I) + ν(I) − d + 1. If mSQ (I) 6= 0, dim(mSQ (I)) = d
since mR[QT] is the only associated prime of SQ (I). Therefore deg SQ (I) <
deg SQ (I) and thus f0 (I) ≤ s0 (I) + ν(I) − d.
• Suppose I is an almost complete intersection. If mSQ (I) = 0, we are in a case
similar to that treated in [23, Corollary 3.8]. Then F (I) is a hypersurface ring
and rQ (I) + 1 = f0 (I) = s0 (I) + 2.
To benefit from these bounds, we need information about F (I). For example, if I and
Q are given by forms of the same degree, F (I) is an integral domain. Such conditions
could be used in generalized versions of the classical Cayley-Hamilton theorem:
Theorem 3.21. Let (R, m) be a Noetherian local ring, I an ideal and Q one of its
minimal reductions. Then
(1) rQ (I) ≤ νF (Q) (F (I)) − 1.
(2) If F (I) satisfies the condition S1 of Serre, rQ (I) ≤ f0 (I) − 1.
Proof. (1) is the standard Cayley-Hamilton theorem. The condition in (2) means that
F (I) is a torsionfree graded module of rank f0 (I) over the polynomial ring F (Q). One
then invokes [74, Proposition 9].
✷
3.5. Multiplicity and reduction number. The following result of M. E. Rossi ([56])
is the main motivation of our discussion here.
Theorem 3.22. ([56, Corollary 1.5]) If (R, m) is a Cohen-Macaulay local ring of dimension at most 2, then for any m-primary ideal I with a minimal reduction Q,
rQ (I) ≤ e1 (I) − e0 (I) + λ(R/I) + 1.
Given that e1 (I) − e0 (I) + λ(R/I) is the multiplicity s0 (Q; I) of the Sally module
SQ (I), we still consider:
(1) If R is Cohen-Macaulay, [56] raised the question on whether in all dimensions
rQ (I) ≤ s0 (Q; I) + 1,
a fact observed in numerous cases.
14
WOLMER V VASCONCELOS
(2) If R is not Cohen-Macaulay the expression for the multiplicity of SQ (I) is
different [as seen in some cases above]. Nevertheless we do not have a failure
for this bound for rQ (I).
(3) If R is not Cohen-Macaulay this formula may require the addition, or multiplication, by a constant that expresses some defficiency d(R) of R being nonCohen-Macaulay, such as
• rQ (I) ≤ d(R) · s0 (Q; I), or
• rQ (I) ≤ s0 (Q; I) + d(R).
3.6. Sally questions. Let (R, m) be a Cohen–Macaulay local ring of dimension d.
Denote the embedding dimension of R by n = ν(m) and e = e0 (m) its multiplicity.
Sally series of questions grew out of an elementary observation of S. Abhyankar ([1]):
n ≤ e + d − 1.
This is proved by first assuming, harmlessly, that the residue field of R is infinite
and modding out R by a minimal reduction Q of m. The inequality results from
λ(m/Q) ≥ ν(m/Q), the fact that Q is generated by a subset of a minimal generating
set of m and the interpretation of the multiplicity e as λ(R/Q).
The task undertaken by Sally in a series of papers ([63, 64, 65, 66, 67]) was to examine
the impact of extremal values of this inequality on the properties of the associated
graded ring G = grm (R). Thus in [63] it is proved that if n = e + d − 1, then G
is Cohen–Macaulay. In the next case, if n = e + d − 2 or n = e + d − 3, G is still
Cohen–Macaulay if R is Gorenstein.
If n = e + d − 2 numerous cases showed that G is not always Cohen–Macaulay but
nevertheless that still depth G ≥ d − 1. That depth G ≥ d − 1 for all Cohen–Macaulay
rings coalesced into a question that became known as the Sally Conjecture. It was
independently resolved in the affirmative by M. E. Rossi and G. Valla ([59]) and H.-J.
Wang ([79]).
It should be pointed out that these developments have been greatly enhanced for
more general filtrations in [62]. Another aspect muted here, for which this author
apologizes, is the lack of a discussion of the Hilbert of function of G, which was of great
interest to the authors mentioned above.
4. Equimultiple Ideals
Let F be a multiplicative filtration and let A = R[QT] be a reduction of B. If Q is
not m–primary, but still a complete intersemection, that is I1 is an equimultiple ideal,
we would need an extended notion of Hilbert polynomial. Of course this requires an
understanding of the relationship between the coefficients and the cohomology of SB/A .
Let us outline two ways that might work out. The first uses an extended degree function
([76]), the other a generalized multiplicity. They are both focused on determining the
multiplicity of SB/A .
THE SALLY MODULES OF IDEALS: A SURVEY
15
4.1. Extended Hilbert function. We are just going to replace lengths by multiplicities. Let F = {In , I0 = R} be a filtration of ideals. Let deg(·) be a multiplicity
function, possibly a Hilbert–Samuel multiplicity [usually denoted by e0 (I; ·)] Define
n 7→ deg(R/In ),
and set
H(R; t) =
X
deg(R/In )tn ,
H(gr(B); t) =
n≥1
X
deg(In /In+1 )tn .
n≥0
Proposition 4.1. Let Q be an ideal of codimension r. Then H(R; t) is a rational
function of degree r + 1.
Proof. Let p1 , . . . , ps be the minimal prime ideals of Q. Since Q is a reduction of
I1 , they are also the minimal primes of height r of the ideals In . By the elementary
additivity formula for multiplicities,
X
deg(R/In ) =
(5)
deg(R/pk )λ((R/In )pk ).
1≤k≤s
The assertion follows because λ((R/In )pk ) is the Hilbert function of the localization of
the filtration {In } in the local ring Rpk of dimension r.
✷
We note that when written out, the series
X
hk (t)
,
λ((In /In+1 )pk )tn =
(1 − t)r
hk (1) > 0.
n≥0
We collect the first Hilbert coefficients of these localizations as
X
E0 (gr(B)) =
deg(R/pk )e0 (gr(Bpk ))
1≤k≤s
E1 (gr(B)) =
X
deg(R/pk )e1 (gr(Bpk ))
1≤k≤s
which we make use of to write the corresponding Hilbert polynomial as
n+r−1
n+r−2
E0 (gr(B))
− E1 (gr(B))
+ lower terms.
r−1
r−2
The coefficients are integers and are called the extended Hilbert coefficients of gr(B).
Now we use this technique to determine some properties of the corresponding Sally
module SB/A in case Q is a complete intersection of codimension r and In = I n
where I is an unmixed ideal. Note that if I is m-primary, E0 (gr(B)) = e0 (I) and
E1 (gr(B)) = e1 (I).
Proposition 4.2. Let R be a Cohen–Macaulay local ring of dimension d ≥ 1. If Q
is a complete intersection and I is unmixed then SB/A = 0 or dim SB/A = d. In the
latter case SB/A has the S1 property of Serre.
Proof. Start by noticing that R[Qt] ⊗ R/I = R[Qt] ⊗ R/Q ⊗ R/I is a polynomial ring
over R/I, and therefore has the condition S1 since I is unmixed. This means that
IR[Qt] is a divisorial ideal of the Cohen-Macaulay ring A = R[Qt]. If SB/A 6= 0 and
P ⊂ A is one of its associated primes of codimension at least 2, the exact sequence
0 → Hom(A/P, IA) = 0 −→ Hom(A/P, IB) = 0 −→ Hom(A/P, SB/A ) −→ Ext1A (A/P, IA)
16
WOLMER V VASCONCELOS
leads to a contradiction since the last module vanishes as IA has the condition S2 of
Serre.
✷
4.2. Ideals of positive dimension. Let I be of dimension one and examine how
SB/A is Cohen-Macaulay. [Set J = Q] In the exact sequence
0 → IR[QT ] −→ IR[IT ] −→ SQ (I) → 0,
SQ (I) 6= 0 is Cohen-Macaulay means that depth IR[IT ] ≥ d, which from the basic
sequences mean that depth R[IT ] ≥ d also [recall IR[QT ] is MCM R[QT ]-module. A
maximal sop for SQ (I) is (QT, a), a a sop for R/Q. Reduction mod QT gives rise to
the exact sequence
0 → IA ⊗A A/(QT ) −→ IR[IT ] ⊗A A/(QT ) −→ SQ (I) ⊗A A/(QT ) → 0.
The last module is
FQ (I) =
r
M
I n+1 /QI n ,
r = rQ (I)
n=1
must be Cohen-Macaulay over R/Q. In particular
Proposition 4.3. rQ (I) ≤ deg FQ (I).
Corollary 4.4. If SQ (I) is Cohen-Macaulay then deg SQ (I) = deg FQ (I).
This gives a curious fact [probably trivial!]
Proposition 4.5. For all n, I n /QI n−1 is a torsion free R/Q-module.
Let us examine the possible extension to the case of an ideal Cohen-Macaulay ideal
J of codimension two and deviation one, generically a complete intersection. This will
imply that A = R[Jt] is Cohen-Macaulay.
P
The extended Hilbert series of SB/A , n≥2 deg(I n /IJ n−1 )tn , can be related to that
of gr(B) by comparison of the exact sequences
0 → I n /IJ n−1 −→ R/IJ n−1 −→ R/I n → 0,
and
0 → J n−1 /IJ n−1 = J n−1 /J n ⊗ R/I −→ R/IJ n−1 −→ R/J n−1 → 0
of modules of dimension d − r.
These sequences give rise to the equalities
deg(I n /IJ n−1 ) = deg(R/IJ n−1 ) − deg(R/I n )
deg(R/IJ n−1 ) = deg(J n−1 /IJ n−1 ) + deg(R/J n−1 ).
They give for n ≥ 2,
deg(I n /IJ n−1 ) = deg(J n−1 /IJ n−1 ) + deg(R/J n−1 ) − deg(R/I n )
n+r−2
n+r−2
= deg(R/I)
+ deg(R/J)
r−1
r
n+r−1
n+r−1
− [E0 (gr(B))
− E1 (gr(B))
+ lower terms].
r
r−1
THE SALLY MODULES OF IDEALS: A SURVEY
17
In particular
E0 (gr(B)) = deg(R/J)
E1 (gr(B)) ≥ 0.
We can also define an extended Hilbert function for the Sally module SB/A and the
corresponding Hilbert polynomial
n+r−1
n+r−2
n
n−1
deg(I /IJ
) = E0 (SB/A )
− E1 (SB/A )
+ lower terms.
r−1
r−2
The coefficient E0 (SB/A ) ≥ 0 but may vanish, that is this Hilbert polynomial could be
of degree < r.
Proposition 4.6. Let J be an unmixed ideal generated by a d-sequence with a CohenMacaulay Rees algebra. Then
(1) E0 (SB/A ) = E1 (gr(B)) − deg(I/J).
(2) E0 (SB/A ) = 0 if and only if dim SB/A < d.
(3) If dim SB/A = d, E0 (SB/A ) = deg SB/A .
Proof. To treat the vanishing of E0 (SB/A ), from the expression for the Hilbert polynomial of the filtration I n ,
X
E0 (SB/A ) = E1 (gr(B)) − deg(I/J) =
[e1 (gr(Bpk )) − λ(Ipk /Jpk )] deg(R/pk )
k
=
X
deg SBpk /Apk deg(R/pk ).
k
Therefore E0 (SB/A ) = 0 if and only if the Sally modules at the localizations Rpk all
vanish according to Proposition 2.1.
✷
Let us interpret these coefficients.
Proposition 4.7. E0 (SB/A ) = deg SB/A .
Proof. This follows from (5) and the general associativity formula for multiplicities
([48, Theorem 24.7]).
✷
4.3. Cohen-Macaulayness. This allows to state the Cohen–Macaulay test for SB/A
as follows:
Theorem 4.8 (CM test for Sally modules). Suppose J is a complete intersection. Let
a = {a1 , . . . , ad−r } be a multiplicity set of parameters for R/J, that is deg R/J =
λ(R/(J, a)). If I1 is a Cohen–Macaulay ideal and SB/A 6= 0 then
X
E1 (gr(B)) ≤
λ(Ij /((a)Ij + JIj−1 )),
j≥1
with equality if and only if SB/A is Cohen–Macaulay.
It is important to keep in mind that Jt, a is a system of parameters for both gr(B)
and SB/A . To explain this formulation, just note that the right-hand side is the sum of
the multiplicity of the Cohen–Macaulay R/J–module I1 /J (recall that I1 is a Cohen–
Macaulay ideal) plus λ(SB/A ⊗ A/(a, Jt)). One then invokes Serre’s Theorem.
18
WOLMER V VASCONCELOS
To make it effective requires information about E1 (gr(B)), which is not an easy task.
We will make observations about the following topics:
Proposition 4.9. With B as above, let C be its integral closure. Then
(1) B satisfies the condition R1 of Serre if and only if
E1 (gr(B)) = E1 (gr(C)).
(2) If J is m–primary, these conditions are equivalent to
deg F (B) = deg F (C).
4.4. Ideals of positive dimension: non Cohen-Macaulay rings. Let (R, m) be a
Noetherian local ring of dimension d > 0 that satisfies the condition Sr of Serre, r ≥ 1.
For the study of Sally modules of ideals I with dim R/I > 0 we need an understanding
of the Rees algebras of some of their reductions. Let us consider some special cases
of equimultiple ideals. We make use of the techniques of [34, Theorems 6.1, 9.1 and
10.1]).
(1) Q = (x1 , . . . , xq ) is a partial system of parameters and a d-sequence. Since Q is
generated by a d-sequence, the approximation complex (set S = R[T1 , . . . , Tq ])
0 → Hq (Q) ⊗ S[−d] → · · · → H1 (Q) ⊗ S[−1] → H0 (Q) ⊗ S → grQ (R) → 0
is acyclic ([33, Theorem 5.6]). As R has the condition Sr , Hj (Q) = 0 for
j > q − r, which gives the estimate
depth grQ (R) ≥ q − (q − r) = r.
(2) By tensoring this complex by R/I, we get the exact sequence
φ
H1 (Q) ⊗ R/I ⊗ S[−1] −→ R/I ⊗ S −→ G = grQ (R) ⊗ R/I → 0.
If I1 (ϕ) ⊂ I, the mapping φ is trivial so that G = R/I ⊗ S, which gives
depth G ≥ q.
(3) Now we look at the approximation complex
0 → Zq (Q) ⊗ S[−q] → · · · → Z1 (Q) ⊗ S[−1] → Z0 (Q) ⊗ S → R[QT] → 0,
which is also acyclic. Here the Zi are the modules of cycles of the Koszul
complex of Q. We have
Zq =
depth Zq−1 ≥
..
.
depth Zi ≥
depth Z0 ≥
0
r
r − i + 1,
r.
i>q−r
Depth chasing the exact sequence yields
depth R[QT] ≥ r + 1,
in all cases. This means that R[QT] satisfies Sr+1 .
THE SALLY MODULES OF IDEALS: A SURVEY
19
(4) Suppose Q is a minimal reduction of I and consider the corresponding Sally
module
0 → IR[QT] −→ IR[IT] −→ SQ (I) → 0.
To get information about depth IR[QT], the exact sequence
0 → IR[QT] −→ R[QT] −→ G → 0,
and (1) and (2) gives SQ (I) has the condition S1 or is zero.
(5) If R[IT] has the condition S2 then so does SQ (I).
(6) We need a version of Proposition 4.6 for ideals such as Q [which may be mixed].
(7) To approach a calculation of s0 (SQ (I)), consider the exact sequences
0 → I n+1 /IQn −→ R/IQn −→ R/I n+1 → 0,
and
0 → Qn /IQn = Sn ⊗ R/I −→ R/IQn −→ R/Qn → 0
of modules of dimension ≤ 1. Applying the functor Hm (·) to these sequences,
we get
0 → I n+1 /IQn −→ H0m (R/IQn ) −→ H0m (R/I n+1 ) → 0,
and
0 → Sn ⊗ H0m (R/I) −→ H0m (R/IQn ) −→ H0m (R/Qn ) −→ Sn ⊗ H1m (R/I).
4.5. j-multiplicities. Let (R, m) be a Noetherian local ring, A a standard graded Ralgebra and M a finitely generated graded A-module. We can attach to H = H0m (M )
a Hilbert function
n 7→ λ(Hn ),
which we call the j–transform of M .
Hilbert coefficients of Achilles-Manaresi polynomials. The corresponding Hilbert series
and Hilbert polynomial will be still written as P (M ; t) and H(M ; t). We use a different
notation for the coefficients of these functions:
r−1
X
t+r−i−1
i
(−1) ji (M )
,
H(M ; t) =
r−i−1
r = ℓ(M ).
i=0
If r = 1 this polynomial does not provide for j1 (M ), so we use instead the function
X
n 7→
λ(Hk ).
k≤n
The coefficients ji (M ) are integers but unlike the usual case of an Artinian local R
it is very hard to calculate being less directly related to M . In addition, some general
relationships that are known to exist between the standard coefficients e0 , e1 , e2 , for
instance, are not known.
We illustrate one of these issues with a series of questions. Let R be a Noetherian
local ring and let I = (x1 , . . . , xr ), r ≤ d = dim R, be an ideal generated by a partial
system of parameters. Let G be the associated graded ring of I, G = grI (R). The
module H = H0m (G) has dimension ≤ r.
20
WOLMER V VASCONCELOS
5. Ideals of Dimension One
Our data is now the following kind:
• R is Cohen–Macaulay of dimension d ≥ 1
• I1 is a Cohen-Macaulay equimultiple ideal of dimension 1, or
• J is an aci and I is an ideal such that IR(J) is CM–HOW TO FIND SUCH?
• Setting up the Hilbert function in case of low reduction number
• Role of the Sally fiber
Here is a specific problem: Suppose R = k[x, y, z] and J = (a, b) is a homogeneous complete intersection. We want to find I ⊂ J¯ such that IR[JT] is
Cohen-Macaulay. For example, what if I is the integral closure itself?
Of course, if R[IT] is Cohen-Macaulay [which is the case if J is monomial],
then all powers I n are Cohen–Macaulay, by standard factors: grI (R) is Cohen–
Macaulay of dimension 3 which the ideal mgrI (R) has height 1 since the analytic
spread is 2 and thus m is not an associated prime of any component I n /I n+1 ,
then by induction get that m is not associated to any R/I n .
If I is Cohen-Macaulay, get IR[JT] CM and the Sally module behaves as in
the dimension zero case [m primary ideal].
The more interesting case is that of an aci J with I CM. In this case R[JT]
is still CM, but not sure IR[JT] is CM. Need examples.
5.1. Generic complete intersections. Let R be a Cohen-Macaulay local ring of
dimension d > 1 and I a Cohen-Macaulay ideal with dim R/I ≥ 1. We assume that I
is not equimultiple. One special target is the set of prime ideals of dimension one of a
regular local ring.
Let I be an ideal that is a complete intersection on the punctured spectrum. If Q
is a minimal reduction the Sally module SQ (I) has finite support and therefore has an
ordinary Hilbert function. We do not know well how to express its Hilbert coefficients
and use them to determine the properties of SQ (I).
Let us assume dim R/I = 1, and let Q be a minimal reduction. We examine the
Sally module SQ (I) under he assumption that I is generically a complete intersection.
In this case Q is generated by a d-sequence and the approximation complex
0 → H1 (Q) ⊗ S[−1] −→ H0 (Q) ⊗ S −→ grQ (R) → 0
is exact (see [34, Theorem 9.1, 10.1]). Furthermore, by [34, Theorem 6.1], both grQ (R)
and R[QT] are Cohen–Macaulay. Note that the Koszul homology module H1 (Q) is a
Cohen-Macaulay module of dimension 1 and I · H1 (Q) = 0 since Ip = Qp at any of its
minimal primes and Ip is generated by a regular sequence. It follows that if we reduce
this complex by R/I we get the complex
(6)
0 → H1 (Q) ⊗ S[−1] −→ R/I ⊗ S −→ grQ (R) ⊗ R/I → 0,
that localization at the primes p shows it is acyclic.
Proposition 5.1. Under these conditions we have:
(1) grQ (R) and grQ (R) ⊗ R/I are Cohen-Macaulay.
THE SALLY MODULES OF IDEALS: A SURVEY
21
(2) If the Sally module is nonzero then dim SQ (I) = d and has the condition S1 of
Serre.
Proof. The first assertion follows from the sequence (6), as the other modules are
Cohen-Macaulay of dimension d + 1.
Now consider the Sally module
0 → IR[QT] −→ IR[IT] −→ SQ (I) → 0.
Since IR[QT] is a maximal Cohen-Macaulay module, SQ (I) = 0 or has the condition
S1 of Serre.
✷
Remark 5.2. With greater generality, suppose R is a Gorenstein local ring and I is
a Cohen-Macaulay ideal of codimension r and analytic deviation 1, that is a minimal
reduction Q is generated by r + 1 elements. If I is a complete intersection on the
punctured spectrum the same assertions about SQ (I) will hold.
These Sally modules have properties akin to those of m-primary ideals.
Corollary 5.3. Let R and I be as above. If dim SA (B) = d, then
(1) p = mA is the only associated prime of SA (B).
(2) The multiplicity of SA (B) satisfies
e0 (SQ (I)) = λ(SQ (I)p ) ≤
r−1
X
λ(I n+1 /QI n ),
n=1
with equality if it is Cohen–Macaulay.
Proof. Since SQ (I)q = 0 for q 6= m, SQ (I) is annihilated by a power of m. As it has the
condition S1 , p is indeed its unique associated prime. As a consequence (QT) gives a
system of parameters for SQ (I). In particular
e0 (SQ (I)) ≤
r−1
X
λ(I n+1 /QI n ),
n=1
by Serre’s formula on Euler characteristics.
✷
Unfortunately neither the multiplicity nor the reduction number have estimates.
Reduction number: Stuff below is not entirely well. Will return.
To approach a calculation of s0 (SQ (I)), consider the exact sequences
(7)
(8)
0 → I n+1 /IQn −→ R/IQn −→ R/I n+1 → 0,
0 → Qn /IQn = Sn ⊗ R/I −→ R/IQn −→ R/Qn → 0
of modules of dimension ≤ 1. Applying the functor Hm (·) to these sequences, we get
(9)
0 → I n+1 /IQn = H0m (I n+1 /IQn ) −→ H0m (R/IQn ) −→ H0m (R/I n+1 ) → 0,
(10)
0 → Sn ⊗ H0m (R/I) = 0 −→ H0m (R/IQn ) −→ H0m (R/Qn ) −→ Sn ⊗ H1m (R/I).
22
WOLMER V VASCONCELOS
We need to know about the growth of
s(n) = λ(I n+1 /IQn )
a(n) = λ(H0m (R/Qn ))
b(n) = λ(H0m (R/Qn+1 ))
c(n) = λ(H0m (R/I n+1 ))
d(n) = λ(H0m (R/IQn ))
e(n) = λ(H0m (I n+1 /Qn+1 ))
f (n) = λ(H0m (Qn /Qn+1 ))
g(n) = λ(H0m (Qn /IQn ))
by making use of the approximation complexes. These are for n ≫ 0 Hilbert polynomials of degree < d and their coefficients denoted in the usual manner. Let us see what
numerical data can be extracted from them.
(1) By local duality we have Hom(H1m (R/I), E) = R/I, where E is the injective
envelope of R/m.
(2) The first two sequences yield
s0 = d0 − c0 < d0
except if
and
d0 ≤ a0 ,
and therefore
s 0 ≤ a0 .
(3) From the exact sequence
0 → H1 (Q) ⊗ S[−1] −→ H0 (Q) ⊗ S −→ grQ (R) → 0,
we have
H0m (H1 (Q))⊗S[−1] = 0 → H0m (H0 (Q))⊗S = I/Q⊗S → H0m (grQ (R)) → H1m (H1 (Q))⊗S[−1].
From which we have
f0 ≥ λ(I/Q).
(4) Meanwhile from
0 → Qn /Qn+1 −→ R/Qn+1 −→ R/Qn → 0
we have
a 0 + f 0 ≥ b0 .
(5) Let us exploit the coarse inequality s0 ≤ a0 in some special cases. Suppose R
is a two-dimensional Gorenstein local ring. In this case we have the Z-complex
ϕ
0 → Z1 ⊗ Sn−1 −→ Z0 ⊗ Sn −→ Qn → 0,
which dualizing gives
ϕ∗
0 → (Qn )∗ −→ Z0∗ ⊗ S∗n −→ Z1∗ ⊗ S∗n−1 −→ Ext1 (Qn , R) = Ext2 (R/Qn , R) → 0.
THE SALLY MODULES OF IDEALS: A SURVEY
23
Note how ϕ acts: If Q = (a, b) and (r, s) ∈ Z1 , that is ra + sb = 0, then
ϕ(r, s) = rT1 + sT2 , and thus
ϕ∗ (T∗1 )((r, s)h(T1 , T2 )) = T∗1 (rT1 + sT2 )h(T1 , T2 ) = rh(T1 , T2 ),
and similar actions. This shows that
coker ϕ = R/I1 (ϕ) ⊗ Sn−1 .
(6) To sum up
s0 ≤ a0 = λ(R/I1 (ϕ)).
(7) This is incomplete: Must use the sequence 0 → B1 −→ Z1 −→ H1 → 0,
dualize to get 0 → Z1∗ −→ B1∗ −→ R/I → 0
Special cases. Suppose ν(I) = d + 1 and consider the sequence (see (4))
ϕ
F (Q) ⊕ F (Q)[−1] −→ F (I) −→ SQ (I)[−1] ⊗ R/m → 0.
If mSQ (I) = 0, we examine the image of ϕ as in [23, Proposition 3.1]. We want to
argue that ϕ is injective. It suffices to show that its image C in F (I) has rank 2. We
begin by observing that F (Q) injects into F (I), so consider the exact sequence
0 → F (Q) −→ C −→ D → 0,
where
D=
M
(IQn−1 + mI n )/(Qn + mI n ).
n≥1
We examine its Hilbert function, λ((IQn−1 + mI n )/(Qn + mI n )) for n >> 0 to show
that dim D = d. First note that by assumption mSQ (I) = 0, that is mI n ⊂ IQn−1 for
all n ≥ 2. Of course if I = (Q, a) it suffices to assume n = 2. We want to argue that
mI n = mIQn−1 . This is embedded in the proof of the following ([23, Proposition 3.1]):
Proposition 5.4. Let (R, m) be a Noetherian local ring with infinite residue field,
I = (Q, a) is an ideal and Q is one of its minimal reductions. If for some integer n,
λ(I n /QI n−1 ) = 1, then r(I) ≤ nν(m) − 1.
Proof. Since Q is a minimal reduction of I, it is generated by a subset of the minimal
set of generators of I so if ν(Q) = r the expected number of generators of I n is r+n
n .
A lesser value for ν(I n ) would imply by [17, Theorem 1] that I n = JI n−1 for some
minimal reduction J of I, and therefore r(I) ≤ n − 1.
n
n−1 ) = 1, mI n ⊂ QI n−1 . Moreover, we
Suppose that ν(I n ) = r+n
n . Since λ(I /QI
have
mQI n−1 ⊂ mI n ⊂ QI n−1 ⊂ I n .
Note that
λ(mI n /mQI n−1 ) = λ(I n /mQI n−1 )−λ(I n /mI n ) = λ(I n /QI n−1 )+ν(QI n−1 )−ν(I n ) = 0.
It follows that mI n = mQI n−1 . From the Cayley-Hamilton theorem we have
(I n )s = (QI n−1 )(I n )s−1 ,
where s = ν(m). In particular I ns = QI ns−1 , as desired.
✷
24
WOLMER V VASCONCELOS
We return to the calculation of the dimension of D. Since mI n = mQn−1 I, we have
λ((IQn−1 + mI n )/(Qn + mI n )) = λ(Qn−1 /(Qn + mI n )) − λ(Qn−1 /(IQn−1 + mI n ))
= λ((Qn−1 /Qn ) ⊗ R/(Q + mI)) − λ((Qn−1 /Qn ) ⊗ R/I)
n+d−2
= λ(I/(Q + mI))
,
d−1
which shows that D has dimension d and multiplicity λ(I/(Q + mI)) = 1. This concludes the proof that
f0 (I) = 2 + s0 (SQ (I)).
We are now ready for the main application of these calculations.
Theorem 5.5. Let (R, m) be a Cohen-Macaulay local ring of dimension d and I a
Cohen-Macaulay ideal of dimension 1. Suppose ν(I) = d + 1 and I is generically a
complete intersection. Let Q be a minimal reduction of I and SQ (I) the associated
Sally module. If mSQ (I) = 0 then F (I) is Cohen-Macaulay and depth SQ (I) ≥ d − 1.
Proof. If the mapping ϕ is injective, in the exact sequence
0 → F (Q) ⊕ F (Q)[−1] −→ F (I) −→ SQ (I) → 0,
F (I) is torsionfree over F (Q) since SQ (I) is also torsionfree over F (Q). The CohenMacaulayness of F (I) then follows as in Theorem 3.21(2).
✷
Remark 5.6. If ν(I) ≥ d + 2,
• From mQI = mI 2 , we still have rQ (I) ≤ 2 · ν(m) − 1
• For ν(I) = d + 2 say, we have the complex
F (Q) ⊕ F (Q)[−1]2 −→ F (I) −→ SQ (I) → 0,
but how to determine the kernel?
5.2. Dimensions two and three. We will begin by illustrating with elementary examples.
Example 5.7. Let R = k[x, y, z] and I = (x2 y, y 2 z, z 2 x, xzy). I is Cohen-Macaulay
and generically a complete intersection. Q = (x2 y − xyz, y 2 z, z 2 x − xyz) is a minimal
reduction of I and λ(I 2 /QI) = 1. Then rQ (I) ≤ 2 (actually rQ (I) = 2).
Example 5.8. Let R be a regular local ring of dimension 3 and I a prime ideal of
codimension two. Let Q be a minimal reduction of I.
What can be said?
• We may assume ν(Q) = 3, as otherwise I is a complete intersection.
6. This and That
We will report briefly on the structure of Sally modules of small rank.
THE SALLY MODULES OF IDEALS: A SURVEY
25
6.1. Multiplicity one. [13], [7], [27], [28], [29], [78]
Theorem 6.1. [28, Theorem 1.2] Let (R, m) be a Cohen-Macaulay local ring of dimension d. If I is a m–primary ideal with a minimal reduction Q then the following three
conditions are equivalent to each other.
(1) e1 = e0 − λ(R/I) + 1.
(2) mS = (0) and rankF (Q) (S) = 1.
(3) S ≃ (X1 , X2 , . . . , Xc )F(Q) as graded R(Q)–modules for some 0 < c ≤ d, where
{Xi }1≤i≤c are linearly independent linear forms of the polynomial ring F(Q).
When this is the case, c = λ(I 2 /QI) and I 3 = QI 2 , and the following assertions hold
true.
(i) depth G ≥ d − c and depth R(Q) S = d − c + 1.
(ii) depth G = d − c if c ≥ 2.
(iii) Suppose c < d. Then
n+d
n+d−1
n+d−c−1
n+1
λ(R/I
) = e0
− e1
+
d
d−1
d−c−1
for all n ≥ 0. Hence
ei =
0
(−1)c+1
if i 6= c + 1
if i = c + 1
for 2 ≤ i ≤ d.
(iv) Suppose c = d. Then
λ(R/I
n+1
) = e0
n+d
n+d−1
− e1
d
d−1
for all n ≥ 1. Hence ei = 0 for 2 ≤ i ≤ d.
This result is very similar to the assumptions and the consequences of the Sally
conjecture.
Corollary 6.2. In addition, if I is an almost complete intersection then the Sally
module is Cohen-Macaulay.
6.2. Normal filtration. [10], [51], [52]
Theorem 6.3. [52, Lemma 2.3] Let (R, m) be an analytically unramified local ring
satisfying Serre conditionLS2 and let F = {In } be a filtration of ideals such that
n
codim I1 ≥ 2 and B =
n≥0 In T is Noetherian. Then the integral closure of B
in R[T] has the condition S2 .
Theorem 6.4. [10, Theorem 2.5] Let (R, m) be an analytically umramified CohenMacaulay local ring of dimension d and infinite residue field and let I be an m primary
ideal. If F = {In = I n } is the corresponding normal filtration, denote by B its Rees
algebra and by S its Sally module. Then
(1) If S 6= 0 then dim S = d.
(2) If S 6= 0 then depth gr(B) ≥ depth S − 1.
(3) If s0 (S) = 1 then S is Cohen-Macaulay.
26
WOLMER V VASCONCELOS
Theorem 6.5. [51, Theorem 1.2] Let (A, m) be a Cohen-Macaulay local ring and I an
m–primary ideal. Let Q be a minimal reduction of I and let
R = A[It], T = A[Qt], G = grI (R), C = (I 2 R/I 2 T )(−1), B = T /mT.
Assume that I is integrally closed. Then the following conditions are equivalent:
(1) e1 (I) = e0 (I) − λA (A/I) + λA (I 2 /QI) + 1,
(2) mC = (0) and rankB (C) = 1.
(3) C ≃ (X1 , X2 , . . . , Xc )B(−1) as graded T –modules for some 1 ≤ c ≤ d, where
X1 , . . . , Xc are linearly independent linear forms of the polynomial ring B.
When this is the case, c = λA (I 3 /QI 2 ) and I 4 = QI 3 , and the following assertions
hold true:
(i) depth G ≥ d − c and depth T C = d − c + 1.
(ii) depth G = d − c if c ≥ 2.
(iii) Suppose c = 1 < d. Then HPI (n) = λA (A/I n+1 ) for all n ≥ 0 and
e1 (I) − e0 (I) + λA (A/I) + 1 if i = 2
1
if i = 3 and d ≥ 3
ei (I) =
0
if 4 ≤ i ≤ d.
(iv) Suppose 2 ≤ c < d. Then HPI (n) = λA (A/I n+1 ) for all n ≥ 0 and
e1 (I) − e0 (I) + λA (A/I) if i = 2
0
if i 6= c + 1, c + 2, 3 ≤ i ≤ d
ei (I) =
(−1)c+1
if i = c + 1, c + 2, 3 ≤ i ≤ d.
(v) Suppose c = d. Then HPI (n) = λA (A/I n+1 ) for all n ≥ 2 and
e1 (I) − e0 (I) + λA (A/I) if i = 2, and d ≥ 2
ei (I) =
0
if 3 ≤ i ≤ d
(vi) The Hilbert series HSI (z) is given by
HSI (z) =
λA (A/I) + (e0 (I) − λA (A/I) − λA (I 2 /QI) − 1)z + (λA (I 2 /QI) + 1)z 2 + (1 − z)c+1 z
.
(1 − z)d
6.3. Monomial ideals. Sally modules of monomial ideals are hard to construct in a
manner that inherit the monomial structure. More precisely if I is a monomial ideal of
R = k[x1 , . . . , xd ], its minimal reductions Q used in the definition of the Sally modules
of I are rarely monomial. One should keep in mind that the point of using Q was to
guarantee a good platform in R[QT] from which to examine R[IT]. Of course this
requirement may be satisfied in other cases, that is without having for Q a minimal
reduction.
Let us illustrate with a special class of ideals. Monomial ideals of finite colength
which are almost complete intersections have a very simple description. Let R =
k[x1 , . . . , xd ] be a polynomial ring over a (possibly infinite) field and let J and I be
R–ideals such that
J = (xa11 , . . . , xadd ) ⊂ (J, xb11 · · · xbdd ) = I.
This is the general form of almost complete intersections of R generated by monomials.
X bi
< 1. This inequality ensures
Perhaps the most interesting cases are those where
ai
THE SALLY MODULES OF IDEALS: A SURVEY
27
that J is not a reduction of I. Let
a
d−1
− xadd , xb1d · · · xbdd ).
Q = (xa11 − xadd , . . . , xd−1
Note that Q is a reduction of I: It is enough to observe that
a
a
d−1
d−1
) ∈ QI d−1 ,
+ xd−1
(xadd )d = xadd (xadd − xa11 + xa11 ) · · · (xadd − xd−1
in particular rQ (I) ≤ d − 1.
Let now B be the Rees algebra of the integral closure filtration {I¯n }. If A0 = R[QT]
and A = R[IT], either of these defines a Sally module, SA0 (B) and SA (B), the latter
carrying a monomial structure. While A0 is always Cohen–Macaulay, A is still often
very amenable.
Conjecture 6.6. Let I be a monomial ideal of k[x1 , . . . , xn ]. If I is an almost complete
intersection of finite colength its Rees algebra R[IT] is almost Cohen–Macaulay.
For d = 2 this comes from [57]. An important special case was settled in [6, Theorem
2.5]:
Theorem 6.7. If ai = a and bi = b for all i, then R[IT] is almost Cohen-Macaulay.
We will recall the artesanal method of [39] and examine in detail the case a = b = c =
n ≥ 3 and α = β = γ = 1. We want to argue that R[IT] is almost Cohen–Macaulay. To
benefit from the monomial generators in using Macaulay2 we set I = (xyz, xn , y n , z n ).
Setting B = R[u, T1 , T2 , T3 ], we claim that
L = (z n−1 u−xyT3 , y n−1 u−xzT2 , xn−1 u−yzT1 , z n T2 −y n T3 , z n T1 −xn T3 , y n T1 −xn T2 ,
y n−2 z n−2 u2 −x2 T2 T3 , xn−2 z n−2 u2 −y 2 T1 T3 , xn−2 y n−2 u2 −z 2 T1 T2 , xn−3 y n−3 z n−3 u3 −T1 T2 T3 ).
We also want to show that these ideals define an almost Cohen–Macaulay Rees algebra.
There is a natural specialization argument. Let X, Y and Z be new indeterminates
and let B0 = B[X, Y, Z]. In this ring define the ideal L0 obtained by replacing in the
list above of generators of L, xn−3 by X and accordingly xn−2 by xX, and so on; carry
out similar actions on the other variables:
L0 = (z 2 Zu−xyT3 , y 2 Y u−xzT2 , x2 Xu−yzT1 , z 3 ZT2 −y 3 Y T3 , z 3 ZT1 −x3 XT3 , y 3 Y T1 −x3 XT2 ,
yzY Zu2 − x2 T2 T3 , xzXZu2 − y 2 T1 T3 , xyXY u2 − z 2 T1 T2 , XY Zu3 − T1 T2 T3 ).
Invoking Macaulay2 gives a (non-minimal) projective resolution
φ4
φ3
φ2
φ1
10
22
0 → B40 −→ B17
0 −→ B0 −→ B0 −→ B0 −→ B0 /L0 → 0.
We claim that the specialization X → xn−3 , Y → y n−3 , Z → z n−3 gives a projective
resolution of L.
(1) Call L′ the result of the specialization in B. We argue that L′ = L.
(2) Inspection of the Fitting ideal F of φ4 shows that it contains (x3 , y 3 , z 3 , u3 , T1 T2 T3 ).
From standard theory, the radicals of the Fitting ideals of φ2 and φ2 contain
L0 , and therefore the radicals of the Fitting ideals of these mappings after
specialization will contain the ideal (L1 ) of B, as L1 ⊂ L′ .
28
WOLMER V VASCONCELOS
(3) Because (L1 ) has codimension 3, by the acyclicity theorem ([4, 1.4.13]) the
complex gives a projective resolution of L′ . Furthermore, as proj. dim B/L′ ≤
4, L′ has no associated primes of codimension ≥ 5. Meanwhile the Fitting ideal
of φ4 having codimension ≥ 5, forbids the existence of associated primes of
codimension 4. Thus L′ is unmixed.
(4) Finally, if (L1 ) ⊂ L′ , as L′ is unmixed its associated primes are minimal primes
of (L1 ), but by [39, Proposition 2.5(iii)], there are just two such, mB and L.
Since L′ 6⊂ mB, L is its unique associated prime. Localizing at L gives the
equality of L′ and L since L is a primary component of (L1 ).
Let us sum up this discussion:
Proposition 6.8. The Rees algebra of I(n, n, n, 1, 1, 1), n ≥ 3, is almost Cohen–
Macaulay.
Corollary 6.9. e1 (I(n, n, n, 1, 1, 1)) = 3(n + 1).
Proof. Follows easily since e0 (I) = 3n2 , the colengths of the monomial ideals I and
I1 (φ) directly calculated and rJ (I) = 2 so that
e1 (I) = λ(I/J) + λ(I 2 /JI) = λ(I/J) + [λ(I/J) − λ(R/I1 (φ))] = (3n − 1) + 4.
In particular we have s0 (I) = 4 for the multiplicity of the Sally module SQ (I).
✷
7. Current issues
We leave this as a reminder of unfinished business.
• Veronese relations
• Reduction mod superficial elements
• j-coefficients of Sally modules
• SQ (I) versus SQ (I)
8. List of definitions and notation
Throughout (R, m) is a Noetherian local ring of dimension d > 0.
• Multiplicative filtration of ideals: a sequence of ideals F = {Ij , j ≥ 0} such
that I0 = R, Ij ⊂ Ij−1 and Ij Ik ⊂ Ij+k
P
• Rees algebra of a multiplicatice filtration: B = j≥0 Ij Tj = R ⊕ B+ ⊂ R[T]
L
• Associated graded ring of F: grF (R) = j≥0 Ij /Ij+1
• Reductions of a Rees algebra: a Rees subalgebra A = R[QT] ⊂ B such that B
is finite over A
• Reduction number: The reduction number of B relative to a Rees subalgebra
A is
X
r(B/A) = inf{n | B =
ABj }
j≤n
• Special fiber of a Rees algebra: F (B) = B ⊗R R/m
• Sally module: The Sally module SB/A of B relative to the Rees subalgebra A =
R[QT] is the A-module is defined by the exact sequence of finitely generated
A–modules
29
THE SALLY MODULES OF IDEALS: A SURVEY
0 → I1 A −→ B+ [+1] =
M
Ij tj−1 −→ SB/A =
j≥1
M
Ij+1 /I1 Qj → 0
j≥1
• Special fiber of a Sally module: F (SA (B)) = SA (B)⊗A A/(QT) =
r = r(B/A)
Lr−1
j=1 Ij+1 /I1 Q
References
[1] S. Abhyankar, Local rings of high embedding dimension, American J. Math. 89 (1967), 1073–1079.
[2] R. Achilles and M. Manaresi, Multiplicity for ideals of maximal analytic spread and intersection
theory, J. Math. Kyoto Univ. 33 (1993), 1029-1046.
[3] J. Backelin and R. Fröberg, Koszul algebras, Veronese subrings and rings with linear resolutions,
Rev. Roumaine Math. Pure Appl. 30 (1985), 85–97.
[4] W. Bruns and J. Herzog, Cohen–Macaulay Rings, Cambridge University Press, 1993.
[5] W. Bruns and G. Restuccia, Canonical modules of Rees algebras, J. Pure & Applied Algebra 201
(2008), 189–203.
[6] R. Burity, A. Simis and S. O. Tohaneanu, On a conjecture of Vasconcelos via Sylvester forms, J.
Symbolic Computation, to appear.
[7] A. Corso, Sally modules of m-primary ideals in local rings, Comm. Algebra 37 (2009), 4503–4515.
[8] A. Corso, C. Huneke and W. V. Vasconcelos, On the integral closure of ideals, Manuscripta Math.
95 (1998), 331–347.
[9] A. Corso and C. Polini, Links of prime ideals and their Rees algebras, J. Algebra 178 (1995),
224–238.
[10] A. Corso, C. Polini and M. E. Rossi, Bounds on the normal Hilbert coefficients, Preprint, 2014,
arXiv:1410.4233v1 [math.AC].
[11] A. Corso, C. Polini and W. V. Vasconcelos, Links of prime ideals, Math. Proc. Camb. Phil. Soc.
115 (1994), 431–436.
[12] A. Corso, C. Polini and W. V. Vasconcelos, Multiplicity of the special fiber of blowups, Math.
Proc. Camb. Phil. Soc. 140 (2006), 207–219.
[13] A. Corso, C. Polini and M. Vaz Pinto, Sally modules and associated graded rings, Comm. Algebra
26 (1998), 2689–2708.
[14] T. Cortadellas and C. D’Andrea, Rational plane curves parametrizable by conics, arXiv:1104.2782
[math.AG].
[15] D. Cox, J. W. Hoffman and H. Wang, Syzygies and the Rees algebra, J. Pure & Applied Algebra
212 (2008) 1787–1796.
[16] A. V. Doria, H. Hassanzadeh and A. Simis, A characteristic free criterion of birationality, Advances
in Math. 230 (2012), 390–413.
[17] P. Eakin and A. Sathaye, Prestable ideals, J. Algebra 41 (1976), 439–454.
[18] D. Eisenbud, Commutative Algebra with a view toward Algebraic Geometry, Springer, 1995.
[19] C. Escobar, J. Martı́nez-Bernal and R. H. Villarreal, Relative volumes and minors in monomial
subrings, Linear Algebra Appl. 374 (2003), 275–290.
[20] H. Flenner, L. O’Carroll and W. Vogel, Joins and Intersections, Springer Monographs in Mathematics, New York, 1999.
[21] L. Ghezzi, S. Goto, J. Hong, K. Ozeki, T.T. Phuong and W. V. Vasconcelos, Cohen–Macaulayness
versus the vanishing of the first Hilbert coefficient of parameter ideals, J. London Math. Soc. 81
(2010), 679–695.
[22] L. Ghezzi, S. Goto, J. Hong and W. V. Vasconcelos, Variation of Hilbert coefficients, Proc. Amer.
Math. Soc. 141 (2013), 3037–3048.
[23] L. Ghezzi, S. Goto, J. Hong and W. V. Vasconcelos, Sally modules and the reduction numbers of
ideals, Nagoya Math. J., to appear.
[24] L. Ghezzi, J. Hong and W. V. Vasconcelos, The signature of the Chern coefficients of local rings,
Math. Research Letters 16 (2009), 279–289.
[25] S. Goto, J. Hong and M. Mandal, The positivity of the first coefficients of normal Hilbert polynomials, Proc. Amer. Math. Soc. 139 (2011), 2399–2406.
j,
30
WOLMER V VASCONCELOS
[26] S. Goto, J. Hong and W. V. Vasconcelos, Hilbert polynomials of j-transforms, Math. Proc. Cambridge Math. Soc., to appear.
[27] S. Goto, K. Nishida and K. Ozeki, Sally modules of rank one, Michigan Math. J. 57 (2008),
359–381.
[28] S. Goto, K. Nishida and K. Ozeki, The structure of Sally modules of rank one, Math. Research
Letters 15 (2008), 881–892.
[29] S. Goto and K. Ozeki, The structure of Sally modules – towards a theory of non-Cohen-Macaulay
cases, J. Algebra 324 (2010), 2129–2165.
[30] S. Goto and K. Ozeki, Buchsbaumness in local rings possessing constant first Hilbert coefficients
of parameters, Nagoya Math. J. 199 (2010), 95–105.
[31] S. Goto and Y. Shimoda, Rees algebras of Cohen–Macaulay local rings, in Commutative Algebra,
Lect. Notes in Pure and Applied Math. 68, Marcel Dekker, New York, 1982, 201–231.
[32] D. Grayson and M. Stillman, Macaulay 2, a software system for research in algebraic geometry,
2006. Available at http://www.math.uiuc.edu/Macaulay2/.
[33] J. Herzog, A. Simis and W. V. Vasconcelos, Approximation complexes of blowing-up rings, J.
Algebra 74 (1982), 466–493.
[34] J. Herzog, A. Simis and W. V. Vasconcelos, Approximation complexes of blowing-up rings.II, J.
Algebra 82 (1983), 53–83.
[35] J. Herzog, A. Simis and W. V. Vasconcelos, Koszul homology and blowing-up rings, in Commutative Algebra, Proceedings: Trento 1981 (S. Greco and G. Valla, Eds.), Lecture Notes in Pure and
Applied Mathematics 84, Marcel Dekker, New York, 1983, 79–169.
[36] J. Herzog, N. V. Trung and B. Ulrich, On the multiplicity of blow-up rings of ideals generated by
d–sequences, J. Pure & Applied Algebra 80 (1992), 273–297.
[37] J. Hong, A. Simis and W. V. Vasconcelos, The homology of two-dimensional elimination, J.
Symbolic Computation 43 (2008), 275–292.
[38] J. Hong, A. Simis and W. V. Vasconcelos, On the equations of almost complete intersections, Bull.
Braz. Math. Soc. 43 (2012), 171-199.
[39] J. Hong, A. Simis and W. V. Vasconcelos, Extremal Rees algebras, J. Commutative Algebra 5
(2013), 231–267.
[40] J. Hong, A. Simis and W. V. Vasconcelos, Ideals generated by quadrics, J. Algebra 423 (2015),
177–189.
[41] S. Huckaba, Reduction number for ideals of higher analytic spread, Math. Proc. Camb. Phil. Soc.
102 (1987), 49–57.
[42] S. Huckaba, A d-dimensional extension of a lemma of Huneke’s and formulas for the Hilbert
coefficients, Proc. Amer. Math. Soc. 124 (1996), 1393-1401.
[43] C. Huneke, On the associated graded ring of an ideal, Illinois J. Math. 26 (1982), 121-137.
[44] B. Johnston and D. Katz, On the relation type of large powers of an ideal, Mathematika 41 (1994),
209–214.
[45] B. Johnson and D. Katz, Castelnuovo regularity and graded rings associated to an ideal, Proc.
Amer. Math. Soc. 123 (1995), 727-734.
[46] S. Morey and B. Ulrich, Rees algebras of ideals with low codimension, Proc. Amer. Math. Soc.
124 (1996), 3653–3661.
[47] F. Muiños and F. Planas-Vilanova, The equations of Rees algebras of equimultiple ideals of deviation one, Proc. Amer. Math. Soc. 141 (2013), 1241–1254.
[48] M. Nagata, Local Rings, Interscience, New York, 1962.
[49] M. Narita, A note on the coefficients of Hilbert characteristic functions in semi-regular local rings,
Proc. Cambridge Philos. Soc. 59 (1963), 269–275.
[50] D. G. Northcott, A note on the coefficients of the abstract Hilbert function, J. London Math. Soc.
35 (1960), 209–214.
[51] K. Ozeki and M. E. Rossi, The structure of the Sally module of integrally closed ideals,
arXiv:1510.08292.v1 [math.AC].
[52] T. T. Phuong, Normal Sally modules of rank one, arXiv:1506.05210v1 [math.AC].
[53] F. Planas-Vilanova, On the module of effective relations of a standard algebra, Mathematical
Proceedings Cambridge Math. Soc. 124 (1998), 215–229.
THE SALLY MODULES OF IDEALS: A SURVEY
31
[54] L. J. Ratliff, Jr., Local quasi-unmixed Noetherian rings and ideals of the principal class, Pacific
J. Math. 52 (1974), 185-205.
[55] M. E. Rossi, Primary ideals with good associated graded ring, J. Pure & Applied Algebra 145
(2000), 75–90.
[56] M. E. Rossi, A bound on the reduction number of a primary ideal, Proc. Amer. Math. Soc. 128
(2000), 1325–1332.
[57] M. E. Rossi and I. Swanson, Notes on the behavior of the Ratliff-Rush filtration, Commutative
Algebra (Grenoble/Lyon, 2001), 313–328, Contemp. Math., 331, Amer. Math. Soc., Providence
RI, 2003.
[58] M. E. Rossi and G. Valla, Cohen–Macaulay local rings of embedding dimension e + d − 3, Proc.
London Math. Soc. 80 (2000), 107–126.
[59] M. E. Rossi and G. Valla, On a conjecture of J. Sally, Comm. Algebra 24 (1996), 4249–4261.
[60] M. E. Rossi and G. Valla, Cohen-Macaulay local rings of dimension two and an extended version
of a conjecture of J. Sally, J. Pure & Applied Algebra 122 (1997) 293–311.
[61] M. E. Rossi and G. Valla, Stretched m-primary ideals, Beiträge zur Algebra und Geometrie, Contributions to Algebra and Geometry 42 (2001), 103-122.
[62] M. E. Rossi and G. Valla, Hilbert Functions of Filtered Modules, Lecture Notes of the Unione
Matematica Italiana 9, Springer, 2010.
[63] J. D. Sally, On the associated graded ring of a local Cohen–Macaulay ring, J. Math. Kyoto Univ.
17 (1977), 19–21.
[64] J. D. Sally, Cohen–Macaulay local rings of maximal embedding dimension, J. Algebra 56 (1979),
168–183.
[65] J. D. Sally, Good embedding dimensions for Gorenstein singularities, Math. Ann. 249 (1980),
95–106.
[66] J. D. Sally, Tangent cones at Gorenstein singularities, Compositio Math. 40 (1980), 167–175.
[67] J. D. Sally, Numbers of Generators of Ideals in Local Rings, Lecture Notes in Pure and Applied
Mathematics 36, Marcel Dekker, New York, 1978.
[68] A. Simis, Cremona transformations and some related algebras, J. Algebra 280 (2004), 162–179.
[69] J. Stückrad and W. Vogel, Buchsbaum Rings and Applications, VEB Deutscher Verlag der Wissenschaften, Berlin, 1986.
[70] N. V. Trung, Reduction exponent and degree bound for the defining equations of graded rings,
Proc. Amer. Math. Soc. 101 (1987), 222–236.
[71] N. V. Trung, The Castelnuovo regularity of the Rees algebra and associated graded ring, Trans.
Amer. Math. Soc. 350 (1998), 2813–2832.
[72] W. V. Vasconcelos, The equations of Rees algebras, J. reine angew. Math. 418 (1991), 189–218.
[73] W. V. Vasconcelos, Hilbert functions, analytic spread and Koszul homology, A. M. S. Contemporary Mathematics 159 (1994), 401–422.
[74] W. V. Vasconcelos, The reduction number of an algebra, Compositio Math. 104 (1996), 189–197.
[75] W. V. Vasconcelos, Arithmetic of Blowup Algebras, London Math. Soc., Lecture Note Series 195,
Cambridge University Press, 1994.
[76] W. V. Vasconcelos, Integral Closure, Springer Monographs in Mathematics, New York, 2005.
[77] W. V. Vasconcelos, The Chern coefficients of local rings, Michigan Math. J. 57 (2008), 725–744.
[78] M. Vaz Pinto, Hilbert functions and Sally modules, J. Algebra 192 (1996), 504–523.
[79] H.-J. Wang, On Cohen–Macaulay local rings with embedding dimension e + d − 2, J. Algebra 190
(1997), 226-240.
| 0math.AC
|
On Coordinate Minimization of Convex Piecewise-Affine Functions
Tomáš Werner
Dept. of Cybernetics, Faculty of Electrical Engineering, Czech Technical Univ. in Prague
Research Report CTU–CMP–2017–05
arXiv:1709.04989v1 [math.OC] 14 Sep 2017
September 14, 2017
Abstract
A popular class of algorithms to optimize the dual LP relaxation of the discrete energy minimization
problem (a.k.a. MAP inference in graphical models or valued constraint satisfaction) are convergent
message-passing algorithms, such as max-sum diffusion, TRW-S, MPLP and SRMP. These algorithms
are successful in practice, despite the fact that they are a version of coordinate minimization applied
to a convex piecewise-affine function, which is not guaranteed to converge to a global minimizer. These
algorithms converge only to a local minimizer, characterized by local consistency known from constraint
programming. We generalize max-sum diffusion to a version of coordinate minimization applicable to
an arbitrary convex piecewise-affine function, which converges to a local consistency condition. This
condition can be seen as the sign relaxation of the global optimality condition.
1
Introduction
Coordinate minimization is an iterative method for unconstrained optimization, which in every iteration
minimizes the objective function over a single variable while keeping the remaining variables fixed. For
some functions, this simple method converges to a global minimum. This class includes differentiable
convex functions [1, §2.7] and convex functions whose non-differentiable part is separable [25]. For general
non-differentiable convex functions, the method need not converge to a global minimum.
For large-scale non-differentiable convex optimization, coordinate minimization may be an acceptable
option despite its inability to find a global minimum. An example is dual LP relaxations of some NPhard combinatorial optimization problems, such as discrete energy minimization [24, 8] (also known as
MAP inference in graphical models [27] or valued constraint satisfaction [17, 26]) and also some other
problems [23]. This dual LP relaxation leads to the unconstrained minimization of a convex piecewiseaffine function. A number of algorithms for solving this dual LP relaxation have been proposed. One
class of algorithms, sometimes referred to as convergent message passing [27, 5, 6, 16], consists of various
forms of (block-)coordinate minimization applied to various forms of the dual. Examples are max-sum
diffusion [13, 20, 28, 29], TRW-S [9], MPLP [5, 22], and SRMP [10]. Besides these, several large-scale
convex optimization methods converging to a global minimum have been applied to the problem, such as
subgradient methods [21, 11], bundle methods [18], alternating direction method of multipliers [15], and
adaptive diminishing smoothing [19]. For practical problems from computer vision, it was observed [8]
that convergent message-passing methods converge faster than these global methods and their fixed points
are often not far from global minima, especially for sparse instances.
We ask whether convergent message-passing algorithms can be reformulated to become applicable to
a wider class of non-differentiable convex functions than those arising from the above dual LP relaxations.
In other words, whether these algorithms can be studied independently of any combinatorial optimization
problems. In this paper, we make a step in this direction and generalize max-sum diffusion to an algorithm
applicable to an arbitrary convex piecewise-affine function.
Consider an objective function in the form of a pointwise maximum of affine functions. In this case,
univariate minimizers in each iteration of coordinate minimization need not be unique and therefore some
rule must be designed to choose a unique minimizer. We show that for a certain natural deterministic
rule, fixed points of coordinate minimization can be poor. Imitating max-sum diffusion, we propose
a better rule which we call the unique rule: in each iteration, minimize the maximum of only those
1
affine functions that depend on the current variable. With this rule, univariate minimizers are unique
and fixed points of the algorithm satisfy a natural condition, the sign relaxation of the true optimality
condition. This can be seen as a local consistency as used in constraint programming [2], which is known
to characterize fixed points of many algorithms to solve the dual LP relaxation, namely message-passing
algorithms [28, 29, 9, 10] and the algorithm [12, 3].
Little is known theoretically about convergence properties of message-passing algorithms. Although
max-sum diffusion and TRW-S were always observed to converge to a fixed point, this was never proved.
For finite-valued energy terms, it has been proved that every accumulation point of TRW-S and SRMP
satisfy local consistency [9, 10]. Under the same assumption, a stronger result has been proved for
max-sum diffusion [20]: a quantity that measures how much the local consistency condition is violated
converges to zero (note, this is still a weaker result that convergence to a fixed point). Nothing is known
theoretically about convergence rates of max-sum diffusion, TRW-S, MPLP and SRMP.
We revisit the proof from [20] to show that, under a certain assumption, during coordinate minimization with the unique rule the above quantity converges to zero. In contrast to max-sum diffusion, we
show there are objective functions for which the algorithm has no fixed point or even no point satisfying
sign consistency.
2
Minimizing a Pointwise Maximum of Affine Functions
Consider a function f : Rn → R that is a pointwise maximum of affine functions, i.e.,
f (x) = max(aTi x + bi )
i∈[m]
(1)
where a1 , . . . , am ∈ Rn and b1 , . . . , bm ∈ R, and we denote [m] = {1, . . . , m}. For brevity, we will further
on abuse symbols ‘max’ and ‘argmax’ and denote, for y ∈ Rm ,
max y = max yi ,
(2a)
i∈[m]
argmax y = argmax yi = { i ∈ [m] | yi = max y }.
(2b)
i∈[m]
Now function (1) can be written simply as
f (x) = max(Ax + b),
(3)
where matrix A = (aij ) ∈ Rm×n has rows aT1 , . . . , aTm and b = (b1 , . . . , bm ) ∈ Rm .
The well-known condition for a global minimum of a convex function f : Rn → R is that 0 ∈ ∂f (x).
For function (1), the subdifferential reads
∂f (x) = conv{ ai | i ∈ argmax(Ax + b) }.
(4)
In more detail, the following holds:
Proposition 1. For function f given by (1) it holds:
• f is bounded from below iff 0 ∈ conv{a1 , . . . , am }.
• x is a minimizer of f iff 0 ∈ conv{ ai | i ∈ argmax(Ax + b) }.
• If f is bounded from below, it has a minimizer x such that 0 ∈ rint conv{ ai | i ∈ argmax(Ax + b) }.
Proof. The minimization of function (1) can be written as the linear program
min{ z | Ax + b ≤ z, x ∈ Rn , z ∈ R }.
(5)
The dual to this linear program reads
max{ bT λ | λ ∈ Rm , λ ≥ 0, 1T λ = 1, AT λ = 0 }.
(6)
The primal (5) is always feasible. By strong duality, f is bounded from below iff the dual (6) is feasible.
By complementary slackness, x and λ are both optimal iff λi = 0 for every i ∈
/ argmax(Ax + b). If f is
2
bounded from below, by strict complementary
slackness there exist optimal x and λ such that λi > 0 iff
P
i ∈ argmax(Ax + b). Note that AT λ = i∈[m] λi ai . Note that, for any k ∈ N and a1 , . . . , ak ∈ Rn ,
0 ∈ conv{a1 , . . . , ak }
0 ∈ rint conv{a1 , . . . , ak }
⇐⇒
⇐⇒
∃λ ∈ Rk : λi ≥ 0,
k
∃λ ∈ R : λi > 0,
P
P
i
λi = 1,
i
λi = 1,
P
P
i
λi ai = 0,
(7a)
i
λi ai = 0.
(7b)
All three claims are now obvious.
Proposition 2. A non-empty set { Ax | x ∈ Rn , Ax ≤ b } is bounded iff 0 ∈ rint conv{a1 , . . . , am }.
Proof. As the set is non-empty, it is bounded iff the linear program
min{ cT Ax | x ∈ Rn , Ax ≤ b }
is bounded for every c ∈ Rm . This holds iff the dual linear program
max{ −bT λ | AT (λ + c) = 0, λ ≥ 0 }
is feasible for every c ∈ Rm . We show that
∀c ∈ Rm ∃λ ≥ 0: AT (λ + c) = 0
⇐⇒
∃λ′ > 0: AT λ′ = 0.
(8)
• To prove ⇒, take c = 1 and let λ ≥ 0 be such that AT (λ + c) = 0. Then set λ′ = λ + c = λ + 1.
• To prove ⇐, set λ = αλ′ − c where α > 0 is arbitrary such that λ ≥ 0. Such α clearly exists.
Note that λ′ can be multiplied by a positive scale to satisfy 1T λ′ = 1. By (7b), the right-hand statement
in (8) is thus equivalent to 0 ∈ rint conv{a1 , . . . , am }.
3
Sign Relaxation of the Optimality Condition
We said that deciding if x ∈ Rn is a minimizer of function (1) requires deciding if the convex hull of a
finite set of vectors contains the origin. Deciding this condition for large problems may be infeasible. We
define the sign relaxation of this condition which is cheaper to decide, obtained by considering only the
signs of the vectors’ components and dropping their magnitudes.
To describe the key idea, suppose that the convex hull of some vectors a1 , . . . , am ∈ Rn contains the
origin, 0 ∈ conv{a1 , . . . , am }. That is, there are numbers λ1 , . . . , λm ≥ 0 such that
X
λi = 1,
(9a)
i∈[m]
X
λi aij = 0,
j ∈ [n].
(9b)
i∈[m]
Let us relax these conditions, considering only the signs σi = sgn λi ∈ {0, 1} and sij = sgn aij ∈ {−1, 0, 1}.
Equality (9a) implies that at least one of the numbers σ1 , . . . , σm equals 1. For each j ∈ [n], equality (9b)
implies that the numbers σ1 s1j , . . . , σm smj either are all zero, or some are positive and some negative.
One way to write this is as follows:
∃i ∈ [m]: σi = 1,
′
(∃i ∈ [m]: σi sij = −1) ⇔ (∃i ∈ [m]: σi′ si′ j = 1),
(10a)
j ∈ [n].
(10b)
Thus, it is necessary for 0 ∈ conv{a1 , . . . , am } that there exist some σ1 , . . . , σm ∈ {0, 1} satisfying (10).
This can be seen as a constraint satisfaction problem (CSP) [14, 4] with m binary variables and
n + 1 constraints. This particular CSP class can be solved by enforcing (generalized) arc consistency [2].
Suppose that for some j, the signs s1j , . . . , smj are, say, all non-negative and some of them is positive.
Then constraint (10b) enforces that for all i for which sij 6= 0 we have σi = 0. Repeating this for various
coordinates j progressively sets some σi to zero. If finally σi = 0 for all i ∈ [n], constraint (10) is violated
and the CSP has no solution. Otherwise, the CSP has a solution.
3
Since we believe that the described concept of sign relaxation might have a wider applicability in nondifferential convex optimization, we further develop it in more detail and greater generality than is needed
in this paper. Namely, we consider the sign relaxation of the condition 0 ∈ conv X for any (not necessarily
finite) set X ⊆ Rn . This is straightforward because the sign relaxation of this condition depends only
on the set { sgn a | a ∈ X } ⊆ {−1, 0, 1}n which is finite. Here, for a vector a = (a1 , . . . , an ) ∈ Rn , we
denoted sgn a = (sgn a1 , . . . , sgn an ) ∈ {−1, 0, 1}n .
Theorem 3. Let X ⊆ Rn and a ∈ conv X. Then a ∈ rint conv(X ∩ F ) where F is the intersection of all
faces of conv X that contain1 a.
Proof. It can be shown that for any face F of conv X it holds that2 conv(X ∩ F ) = F . So it suffices to
show that a ∈ rint F . For contradiction, suppose a is a relative boundary point of F . But every relative
boundary point of a face is contained in some subface of the face, so F cannot be the intersection of all
faces of conv X containing a.
Definition 1. A set S ⊆ {−1, 0, 1}n is consistent in coordinate j ∈ [n] if it holds that
∃s ∈ S: sj = −1
⇐⇒
∃t ∈ S: tj = 1.
(11)
Set S is consistent if it is consistent in every coordinate j ∈ [n].
In particular, note that the sets ∅ and {0} (where 0 denotes the vector with n zeros) are consistent.
Theorem 4. Let X ⊆ Rn . If 0 ∈ rint conv X, then the set { sgn a | a ∈ X } is consistent.
Proof. Let 0 ∈ rint conv X. Then the projection of the set rint conv X onto each coordinate axis is either
the set {0} or an interval containing zero as its interior point. Noting that projections commute with the
convex hull operator, this means for every j ∈ [n] we have that
∃a ∈ X: aj < 0
⇐⇒
∃b ∈ X: bj > 0.
(12)
This is equivalent to condition (11) for the set { sgn a | a ∈ X }.
Theorem 5. Let X ⊆ Rn . If 0 ∈ conv X, then the set { sgn a | a ∈ X } has a non-empty consistent
subset.
Proof. If 0 ∈ conv X, then by Theorem 3 there is Y ⊆ X such that 0 ∈ rint conv Y . Thus, by Theorem 4,
the set { sgn a | a ∈ Y } is consistent. This set is non-empty and it is a subset of { sgn a | a ∈ X }.
Theorem 5 gives a necessary condition for a set X ⊆ Rn to satisfy 0 ∈ conv X. This is what we called
the sign relaxation of 0 ∈ conv X. This condition is of course not sufficient, e.g., for X = {(1, −2), (−2, 1)}
the set { sgn a | a ∈ X } = {(1, −1), (−1, 1)} is consistent but 0 ∈
/ conv X.
Next we develop the concept of consistency closure, analogous to arc consistency closure [2].
Theorem 6. Let the sets S, T ⊆ {−1, 0, 1}n be consistent. Then the set S ∪ T is consistent.
Proof. Let s ∈ S ∪ T , so s ∈ S or s ∈ T . Suppose s ∈ S. Suppose for some j ∈ [n] we have, say, sj < 0.
As S is consistent, there is t ∈ S such that sj > 0. But t ∈ S ∪ T .
Definition 2. The consistency closure of a set S ⊆ {−1, 0, 1}n is the greatest (with respect to partial
ordering by inclusion) consistent subset of S, i.e., it is the union of all consistent subsets of S. We will
denote it by cons S.
Theorem 7. The cons operator satisfies the axioms of a closure operator, i.e., it is
• intensive (cons S ⊆ S),
1
More precisely, it can be shown that Y = X ∩ F is the greatest (with respect to partial ordering by inclusion) subset
of X such that a ∈ rint conv Y .
2
We omit the proof of this claim. The proof is obvious if X is finite and hence conv X is a convex polytope. For infinite X,
recall [7] that a face of a convex set C is a convex set F ⊆ C such that every line segment from C whose relative interior
has a non-empty intersection with F is contained in F .
4
• idempotent (cons cons S = cons S),
• non-increasing (S ⊆ T =⇒ cons S ⊆ cons T ).
Proof. Intensivity and idempotency are immediate from Definition 2. To show the non-increasing property, let S ⊆ T . By intensivity, we have cons S ⊆ T . But cons S is consistent, thus by Definition 2 it
must be a subset of cons T .
Definition 3. Enforcing consistency of a set S ⊆ {−1, 0, 1}n is the algorithm that repeats the following
iteration:
1. choose an arbitrary coordinate j ∈ [n] in which S is not consistent,
2. remove from S all elements s such that sj 6= 0.
If S is consistent, the algorithm stops.
Theorem 8. Enforcing consistency of any set S ⊆ {−1, 0, 1}n yields cons S.
Proof. The algorithm creates a sequence of sets S = S0 ⊃ S1 ⊃ · · · ⊃ SK , where SK is consistent and
K is the number of iterations. This sequence is given recurrently by Sk = Sk−1 \ { s ∈ Sk−1 | sjk 6= 0 }
where each jk ∈ [n] is such that Sk−1 is not consistent in coordinate jk .
Let a set T ⊆ S be consistent. We show by induction that T ⊆ Sk for all k. Assume T ⊆ Sk−1 .
Since Sk−1 is not consistent in coordinate jk , it follows from Definition 1 that sjk = 0 for every s ∈ T .
Therefore T ⊆ Sk−1 \ { s ∈ Sk−1 | sjk 6= 0 } = Sk .
We have shown that every consistent subset of S is contained in SK , which is itself consistent. This
means that SK = cons S.
To conclude, the sign relaxation of the condition 0 ∈ conv X is that cons{ sgn a | a ∈ X } =
6 ∅.
4
Coordinate Minimization
Coordinate minimization of a function f : Rn → R is a method that, starting from an initial point
x = (x1 , . . . , xn ), repeats the following iteration:
1. choose j ∈ [n],
2. choose x∗j ∈ argmin f (x),
xj ∈R
x∗j .
3. set xj ←
Since the choices in the first two steps are not specified, this does not define a single algorithm but rather
a class of algorithms (a ‘method’).
Further on, we focus on applying coordinate minimization to functions of the form (1).
For functions of the form (1), in Step 2 of coordinate minimization the univariate minimizer x∗j is in
general not unique. Therefore, some rule must be adopted to choose a unique minimizer. Let us restrict
ourselves to deterministic rules. We show that some rules can behave poorly. Consider the following
rule: choose the element of the set argminxj ∈R f (x) that is nearest to xprev
, the j-th variable from the
j
3
previous iteration. We refer to this as the proximal rule .
Example 1. Consider the function
f (x1 , x2 , x3 ) = max{ x2 − x3 , x3 − x1 , x1 − x2 },
(13)
which has minimum value 0, attained for any x1 = x2 = x3 . The point x = (x1 , x2 , x3 ) = (2, 1, 0) is
fixed for coordinate minimization with the proximal rule, with value f (x) = max{1, −2, 1} = 1. The
subdifferential at this point is ∂f (x) = conv X where X = {(0, 1, −1), (1, −1, 0)}. Using the algorithm
from Definition 3 we find that cons{ sgn a | a ∈ X } = cons{(0, 1, −1), (1, −1, 0)} = ∅. Thus, point x does
not satisfy the sign relaxation of the condition 0 ∈ ∂f (x).
3
This can be indeed seen as a proximal regularization of coordinate minimization: instead of function f we minimize
the function g(x, y) = f (x) + µkx − yk2 for a small µ > 0. For this function, the univariate minimization would read
argminxj ∈R (f (x1 , . . . , xn )+µ(xj −xprev
)2 ), which has the unique minimizer given by the proximal rule. Cf. [1, Exercise 2.7.2].
j
5
Imitating max-sum diffusion, we propose a better rule, which we call the unique rule: in Step 2
of coordinate minimization, rather than minimizing the maximum of all affine functions, minimize the
maximum of only those affine functions that depend on variable xj . That is, minimize the function
φj (x) = max (aTi x + bi )
i|aij 6=0
(14)
where maxi|aij 6=0 denotes maximization over all i ∈ [m] such that aij 6= 0. Further on, we will assume
that the set { sgn ai | i ∈ [m] } is consistent. Under this assumption, for every j ∈ [n] and every
x1 , . . . , xj−1 , xj+1 , . . . , xn ∈ R, the univariate function xj 7→ φj (x) has exactly one minimizer, x∗j . This
minimizer is the unique solution of the equation
max (aTi x + bi ) = max (aTi x + bi ).
i|aij <0
i|aij >0
(15)
To summarize, the iteration of the algorithm for coordinate j ∈ [n] adjusts variable xj to satisfy (15),
keeping the other variables unchanged. A fixed point of the algorithm4 is a point x ∈ Rn that satisfies (15)
for all j ∈ [n].
For functions in the form (1), coordinate minimization can be equivalently formulated in terms of the
values y = Ax + b of the affine functions, i.e., instead of updating the numbers x1 , . . . , xn we update the
numbers y1 , . . . , ym , while x1 , . . . , xn are no longer explicitly kept. We first set y = Ax + b, where x is
the initial point, and then repeat the following iteration:
1. choose j ∈ [n],
2. choose x∗j ∈ argmin max(aij xj + yi ),
xj ∈R
i∈[m]
aij x∗j
3. set yi ←
+ yi for all i ∈ [m].
Using the unique rule, in Step 2 we need to find the unique minimizer x∗j of the univariate function
max (aij xj + yi ), which is the solution of the equation
i|aij 6=0
max (aij xj + yi ) = max (aij xj + yi ).
(16)
max yi = max yi .
(17)
i|aij <0
i|aij >0
After the iteration, we thus have
i|aij <0
i|aij >0
A fixed point of the algorithm is a point y ∈ { A(x + x′ ) + b | x′ ∈ Rn } that satisfies (17) for all j ∈ [n].
We introduce the following notations:
• Mapping pj : Rm → Rm denotes the action of Steps 2 and 3 of coordinate minimization with the
unique rule, formulated in terms of y. That is, for y ∈ Rm and j ∈ [n], pj (y) is computed as follows:
find the solution x∗j of equation (16) and then set yi ← aij x∗j + yi for all i ∈ [m].
• Further on, we assume that coordinates j in Step 1 of coordinate minimization are visited in the
cyclic order. Let p = pn ◦ pn−1 ◦ · · · ◦ p2 ◦ p1 denote the action of the algorithm for one cycle.
• For k ∈ N, let pk = p ◦ · · · ◦ p (k-times) denote the action of the algorithm for k cycles.
In this notation, y ∈ Rm is a fixed point of the algorithm iff pj (y) = y for every j ∈ [n], which holds iff
p(y) = y.
Next we give several examples of the algorithm’s behavior.
Example 2. Recall that coordinate minimization is not guaranteed to find a global minimum of a
function of the form (1) because it is not differentiable. An example is the function
f (x1 , x2 ) = max{ x1 − 2x2 , x2 − 2x1 },
which is unbounded but any point x1 = x2 is fixed for coordinate minimization. At any such point we have
0∈
/ ∂f (x1 , x2 ) = conv X where X = {(1, −2), (−2, 1)}. The set { sgn a | a ∈ X } = {(1, −1), (−1, 1)}
is consistent. There is no difference between the proximal rule and unique rule, because univariate
minimizers in Step 2 are unique. Coordinate minimization converges in one iteration.
4
Note, such a point is a Nash equilibrium for penalty functions φ1 , . . . , φn : Rn → R.
6
The unique rule is not worse than the proximal rule, in the sense that every fixed point of coordinate
minimization with the unique rule is also a fixed point of coordinate minimization with the proximal rule.
The following example shows that the unique rule is in fact strictly better.
Example 3. Let us return to Example 1 and apply coordinate minimization with the unique rule to
function (13). For j = 1, equation (16) reads −x1 + y2 = x1 + y3 with the solution x∗1 = (y2 − y3 )/2, thus
y is updated as y2 ← −x∗1 + y2 = (y2 + y3 )/2 and y3 ← x∗1 + y3 = (y2 + y3 )/2. Similarly we derive the
iterations for j = 2, 3. In summary, we have
p1 (y1 , y2 , y3 ) = ( y1 , (y2 + y3 )/2, (y2 + y3 )/2 ),
p2 (y1 , y2 , y3 ) = ( (y3 + y1 )/2, y2 , (y3 + y1 )/2 ),
p3 (y1 , y2 , y3 ) = ( (y1 + y2 )/2, (y1 + y2 )/2, y3 ).
We see that in every iteration, the algorithm takes a pair of the numbers y1 , y2 , y3 and replaces both of
them with their average. For any initial y1 , y2 , y3 , the sequence (pk (y1 , y2 , y3 ))k∈N converges to the point
with y1 = y2 = y3 = (y1 + y2 + y3 )/3, i.e., to a minimizer of function (13).
The next example shows that there are functions for which coordinate minimization with the unique
rule has no fixed point and even no point satisfying the sign-relaxed optimality condition.
Example 4. Let
f (x1 , x2 , x3 ) = max{ x1 − x2 − x3 , x1 + 4, x1 + x2 + x3 , −x1 + x2 + 2 }.
(18)
This function is not bounded from below. System (15), defining the fixed point condition, reads
max{ x1 − x2 − x3 , x1 + 4, x1 + x2 + x3 } = −x1 + x2 + 2
max{ x1 + x2 + x3 , −x1 + x2 + 2 } = x1 − x2 − x3
x1 + x2 + x3 = x1 − x2 − x3
The third equation implies x2 + x3 = 0, thus the system simplifies to
max{ x1 , x1 + 4, x1 } = −x1 + x2 + 2
max{ x1 , −x1 + x2 + 2 } = x1
The first equation simplifies to x1 + 4 = −x1 + x2 + 2, hence x2 = 2x1 + 2. Plugging this to the second
equation gives a contradiction. This shows that the algorithm has no fixed point.
We shall show that there is even no point x ∈ Rn such that cons{ sgn ai | i ∈ argmax(Ax + b) } =
6 ∅.
For contradiction, suppose it is so. That is, there is x ∈ Rn and I ⊆ [m] such that the set { sgn ai | i ∈ I }
is consistent and ∅ =
6 I ⊆ argmax(Ax + b). It can be checked that the only non-empty subset of [m] for
which the set { sgn ai | i ∈ I } is consistent is I = [m]. But there is no x such that argmax(Ax + b) = [m],
i.e., at which all four affine functions are active.
How will the algorithm behave in this case? For the initial point (x1 , x2 , x3 ) = (0, 0, 0), the first three
iterations of the algorithm are
j
x1 x2 x3 y 1 y 2 y 3 y 4
0
0
0
0
4
0
2
1 −1 0
0 −1 3 −1 3
2 −1 −2 0
1
3 −3 1
3 −1 −2 2 −1 3 −1 1
The resulting values of y are the initial values minus one. Every later cycle will again decrease them by
one, therefore algorithm therefore will decrease y unboundedly.
Example 5. Consider the function
f (x1 , x2 , x3 ) = max{ x1 − x2 − x3 , x1 + 4, x1 + x2 + x3 , −x1 + x2 + 2, 0 },
(19)
which differs from (18) only by the extra zero function. As shown in the previous example, the algorithm
has no fixed point and the first four components of the vector y = (y1 , y2 , y3 , y4 , y5 ) will diverge. Since after
a few iterations y1 , y2 , y3 , y4 become all negative but y5 remains zero, the set cons{ sgn ai | i ∈ argmax y }
becomes non-empty.
7
The following theorem shows that every fixed point of coordinate minimization with the unique rule
satisfies the sign relaxation of the condition 0 ∈ ∂f (x).
Theorem 9. If (17) holds for every j ∈ [n], then the set { sgn ai | i ∈ argmax y } is consistent.
Proof. For every j ∈ [n], (17) implies that
∃i ∈ argmax y: aij < 0
⇐⇒
∃i′ ∈ argmax y: ai′ j > 0.
(20)
Indeed, if the common value of both sides in (17) is equal to [less than] max y, both sides of (20) are true
[false]. Condition (20) is equivalent to condition (11) for the set { sgn ai | i ∈ argmax y }.
For an initial point y ∈ Rm , consider the sequence of vectors (pk (y))k∈N . Although we believe that,
under some reasonably weak assumptions, this sequence converges to a fixed point, we are not able to
prove this. Following [20], we formulate and prove a weaker result.
For ǫ ≥ 0, let
argmaxǫ y = { i ∈ [m] | yi + ǫ ≥ max y }
(21)
denote the set of ǫ-maximal components of a vector y ∈ Rm . We now define function e: Rm → R+ by
e(y) = inf{ ǫ ≥ 0 | cons{ sgn ai | i ∈ argmaxǫ y } =
6 ∅ }.
(22)
This function measures how much point y violates the condition5 cons{ sgn ai | i ∈ argmax y } =
6 ∅, which
is the sign relaxation of the condition 0 ∈ conv{ ai | i ∈ argmax y } = ∂f (x).
Theorem 10. Let 0 ∈ rint conv{a1 , . . . , am }. Let y ∈ Rm . Then lim e(pk (y)) = 0.
k→∞
Proof. In the appendix.
4.1
Sum of Maxima of Affine Functions
We have applied coordinate minimization to pointwise maximum of affine functions (1). But what if we
want to apply it to a convex piecewise-affine function in the form of a sum of pointwise maxima of affine
function, i.e.,
X
f (x) =
max(aTij x + bij ),
(23)
j
i
where aij ∈ Rn and bij ∈ R. It turns out that minimizing function (23) can be easily transformed to
minimizing a function (1). One way to do that is as follows.
Theorem 11. For every α1 , . . . , αm ∈ R,
1 X
αi =
min
max(αi + ui ).
u1 +···+um =0 i∈[m]
m
(24)
i∈[m]
Proof. Clearly, at optimum all expressions under the maximum
will have
P
P the same value. Let us denote
this common value by b = αi + ui , so αi = b − ui . Now i αi = mb − i ui = mb.
Using Theorem 11, minimizing function (23) can be transformed to minimizing the function
g(x, u) = max(aTij x + bij + ui )
i,j
P
subject to i ui = 0. Minimization over u can be done in closed form, interlacing iterations of coordinate
minimization over x.
5
In [20], ‘consistency’ means cons{ sgn ai | i ∈ argmax y } =
6 ∅, rather than consistency in the sense of our Definition 1.
8
A
Proof of Theorem 10
A.1
Properties of the Algorithm
Proposition 12. For every y ∈ Rm and j ∈ [n] we have max pj (y) ≤ max y.
Proof. This just says that coordinate minimization never increase the objective function.
For y ∈ Rm and z ∈ R, we denote
I(y, z) = { i ∈ [m] | yi ≥ z }.
(25)
In particular, note that argmax y = I(y, max y) and argmaxǫ y = I(y, max y − ǫ).
Proposition 13. Let y ∈ Rm and j ∈ [n].
• If the set { sgn ai | i ∈ I(y, max y) } is consistent in coordinate j, then I(pj (y), max y) = I(y, max y).
• Otherwise, I(pj (y), max y) = I(y, max y) \ { i | aij 6= 0 }.
Proof. Denote y ′ = pj (y) and
α− = max yi ,
i|aij <0
α+ = max yi ,
α′ = max yi′ = max yi′ = max yi′ .
i|aij >0
i|aij <0
i|aij 6=0
i|aij >0
It follows from the definition of pj that yi′ = yi for all i with aij = 0 and that α′ lies between α− and α+ .
By Definition 1, the set { sgn ai | i ∈ I(y, max y) } is consistent in coordinate j iff
∃i ∈ I(y, max y): aij < 0
⇐⇒
∃i′ ∈ I(y, max y): ai′ j > 0.
(26)
Using the above observations, we see that:
• If both sides of (26) are true, then α− = α+ = α′ , hence y ′ = y, hence I(y ′ , max y) = I(y, max y).
• If both sides of (26) are false, then α− , α+ , α′ < max y. Hence yi , yi′ < max y for all i with aij 6= 0.
Hence I(y ′ , max y) = I(y, max y).
• If, say, the LHE of (26) is true and the RHS is false, then α− = max y and α+ , α′ < max y. Hence
yi′ < max y for all i with aij 6= 0. Hence I(y ′ , max y) = I(y, max y) \ { i | aij 6= 0 }.
Proposition 14. For every y ∈ Rm and every k ≥ m we have
{ sgn ai | i ∈ I(pk (y), max y) } = cons{ sgn ai | i ∈ I(y, max y) }.
Proof. Proposition 13 shows that the algorithm in fact enforces consistency (see Definition 3) of the set
{ sgn ai | i ∈ I(pk (y), max y) }. Moreover, in every m iterations each coordinate is visited and hence
the set shrinks unless it is already consistent. As the set initially has no more than m elements, after
m applications of p the set stops shrinking and becomes consistent.
Proposition 15. Let y ∈ Rm .
(a) If cons{ sgn ai | i ∈ argmax y } =
6 ∅, then for every k ∈ N we have max pk (y) = max y.
(b) If cons{ sgn ai | i ∈ argmax y } =
6 ∅, then the set { sgn ai | i ∈ argmax pm (y) } is consistent.
(c) If cons{ sgn ai | i ∈ argmax y } = ∅, then max pm (y) < max y.
Proof. By Proposition 12, for every k ∈ N we have max pk (y) ≤ max y. By (25) we have:
• If I(pk (y), max y) 6= ∅, then max pk (y) = max y.
• If I(pk (y), max y) = ∅, then max pk (y) < max y.
Noting that argmax y = I(y, max y), the claims now follow from Proposition 14.
9
A.2
Continuity and Boundedness
Proposition 16. Let a1 , . . . , am ∈ R be such that { sgn ai | i ∈ [m] } = {−1, 1}. Then the function
ξ: Rm → R given by
ξ(y) = ξ(y1 , . . . , ym ) = argmin max(ai x + yi )
(27)
x∈R
i∈[m]
is continuous.
Proof. On a neighborhood of any point y ∈ Rm , function ξ depends only on the coordinates for which
the affine functions ai x + yi are active at y. Moreover, we can move the minimum to the origin without
loss of generality. Therefore, to show that function (27) is continuous on Rm , it suffices to show that the
function of the form (27) is continuous at the point y = 0.
For any y ∈ Rm , ξ(y) is the x-coordinate of the intersection of the graphs of two affine functions
ai x + yi and aj x + yj , one with negative and one with positive slope. Thus, ξ(y) = (yi − yj )/(aj − ai ) for
some (i, j) such that ai < 0 and aj > 0. Therefore,
kyk∞ ≤ δ
=⇒
|yi − yj |
2
≤ δ max max
.
a
−
a
a
−
ai
i|ai <0 j|aj >0 j
i|ai <0 j|aj >0 j
i
|ξ(y)| ≤ max max
This shows that function ξ is continuous at y = 0.
Proposition 17. For every j ∈ [n], the map pj is continuous.
Proof. Map pj is continuous because it is a composition of function ξ from Proposition 16 and the affine
map yi 7→ aij xj + yi (as given by Step 3 of coordinate minimization).
Proposition 18. Let 0 ∈ rint conv{a1 , . . . , am }. Let y ∈ Rn . Then the sequence (pk (y))k∈N is bounded.
Proof. By Proposition 12, for every k we have pk (y) ≤ max y. The claim now follows from Proposition 2.
Proposition 19. Let y, y ′ ∈ Rm , ǫ ≥ 0, δ ≥ 0, and ky − y ′ k∞ ≤ δ. Then argmaxǫ y ⊆ argmaxǫ+2δ y ′ .
Proof. Let ky − y ′ k∞ ≤ δ, i.e., −δ ≤ yi − yi′ ≤ δ for every i. This implies −δ ≤ max y − max y ′ ≤ δ. By
these inequalities, for every i we have the implication
max y − yi ≤ ǫ
=⇒
max y ′ − yi′ ≤ ǫ + 2δ.
(28)
By (21), this means that argmaxǫ y ⊆ argmaxǫ+2δ y ′ .
Proposition 20. The function e is continuous.
Proof. Let ky − y ′ k∞ ≤ δ. By (22), the set { sgn ai | i ∈ argmaxe(y) y } has a non-empty consistent subset.
By Proposition 19, the set { sgn ai | i ∈ argmaxe(y)+2δ y ′ } has the same consistent subset, therefore
e(y ′ ) ≤ e(y) + 2δ. Similarly we prove that e(y) ≤ e(y ′ ) + 2δ. Thus |e(y) − e(y ′ )| ≤ 2δ.
Proposition 21. Let cons{sgn ai | i ∈ [m]} =
6 ∅. Then for every y ∈ Rn we have e(y) ≤ max y − min y.
Proof. Let ǫ = max y −min y. Then clearly argmaxǫ y = [m], hence cons{ sgn ai | i ∈ argmaxǫ y } =
6 ∅.
A.3
Convergence
Using the preparations from §A.1 and §A.2, we now prove the main convergence result. We will do it by
reformulating [20, Theorem 1].
Let q = pm denote m cycles of coordinate minimization. In this section, yk will denote a vector
from Rm , rather than the k-th component of a vector y ∈ Rm . Sequences such as (yk )k∈N will be denoted
in short as (yk ).
Recall that an accumulation point of a sequence is the limit point of its convergent subsequence.
Theorem 22. Let q: Rm → Rm be continuous. Let y ∈ Rm . Let the sequence (max q k (y)) be convergent.
Then every accumulation point y ∗ of the sequence (q k (y)) satisfies max q(y ∗ ) = max y ∗ .
10
Proof. For brevity, denote yk = q k (y). Let y ∗ be an accumulation point of (yk ), thus
lim yk(l) = y ∗
l→∞
(29)
for some strictly increasing function k: N → N. Applying the continuous map q to equality (29) yields
lim q(yk(l) ) = lim yk(l)+1 = q(y ∗ ),
l→∞
l→∞
(30)
where we used that q(yk(l) ) = yk(l)+1 . Now
max y ∗ = lim max yk(l) = lim max yk = lim max yk(l)+1 = max q(y ∗ ).
l→∞
k→∞
l→∞
(31)
The first and last equality follow from applying the continuous function max : Rm → R (defined by (2a)) to
equalities (29) and (30). The second and third equality hold because the sequence (max yk ) is convergent
and thus every its subsequence converges to the same number.
The following fact is well-known from analysis:
Proposition 23. Let (ak ) be a bounded sequence. If every convergent subsequence of (ak ) converges to
a point a, then the sequence (ak ) converges to a.
Proof. Suppose (ak ) does not converge to a. Then for some ǫ > 0, for every k0 there is k > k0 such
that kak − ak > ǫ. So there is a subsequence (bk ) such that kbk − ak > ǫ for all k. As (bk ) is bounded,
by Bolzano-Weierstrass it has a convergent subsequence, (ck ). But (ck ) clearly cannot converge to a, a
contradiction.
Theorem 24. Let q: Rm → Rm and e: Rm → R be continuous such that, for every y ∈ Rm :
1. max q(y) ≤ max y,
2. max q(y) = max y implies e(y) = 0,
3. the sequences (q k (y)), (max q k (y)) and (e(q k (y))) are bounded.
Then for every y ∈ Rm we have lim e(q k (y)) = 0.
k→∞
Proof. Denote yk = q k (y). The sequence (max yk ) is bounded and non-increasing, therefore convergent.
By Theorem 22, every accumulation point y ∗ of (yk ) satisfies max q(y ∗ ) = max y ∗ . This implies e(y ∗ ) = 0.
By Proposition 23, now it suffices to show that every convergent subsequence of the sequence (e(yk ))
converges to 0. So let (zk ) be a subsequence of (yk ) such that limk→∞ e(zk ) = e∗ .
• If (zk ) is convergent, then y ∗ = limk→∞ zk is an accumulation point of (yk ), therefore e(y ∗ ) = 0.
Applying the continuous function e to this limit yields e(y ∗ ) = limk→∞ e(zk ) = e∗ = 0.
• If (zk ) is not convergent, by Bolzano-Weierstrass it has a convergent subsequence, (wk ). As (wk ) is
also a subsequence of (yk ), by the above reasoning we have limk→∞ e(wk ) = 0. But because
limk→∞ e(zk ) = e∗ , every subsequence of (e(zk )) converges to e∗ . As (wk ) is a subsequence of (zk ),
this implies e∗ = 0.
Theorem 24 implies Theorem 10. Indeed, map q is continuous because it is a composition of maps pj
which are continuous by Proposition 17. Function e is continuous by Proposition 20. Condition 1 holds
by Proposition 12 and Condition 2 by Proposition 15(c). The sequences (q k (y)) and (max q k (y)) are
bounded by Proposition 18. The sequence (e(q k (y))) is bounded by Proposition 21.
We remark that Theorem 24 has a wider applicability, to prove convergence to local consistency for
several other message-passing algorithms. For that, the functions max, q and e need to be replaced by
appropriate functions in these algorithms and they must satisfy the assumptions of he theorem.
Acknowledgement
This work has been supported by the Czech Science Foundation grant 16-05872S.
11
References
[1] D. P. Bertsekas. Nonlinear Programming. Athena Scientific, Belmont, MA, 2nd edition, 1999.
[2] C. Bessiere. Constraint propagation. In Handbook of Constraint Programming, chapter 3. Elsevier,
2006.
[3] M. C. Cooper, S. de Givry, M. Sanchez, T. Schiex, M. Zytnicki, and T. Werner. Soft arc consistency
revisited. Artificial Intelligence, 174(7-8):449–478, 2010.
[4] E. Freuder and A. K. Mackworth. Constraint satisfaction: An emerging paradigm. In Handbook of
Constraint Programming, chapter 2. Elsevier, 2006.
[5] A. Globerson and T. Jaakkola. Fixing max-product: Convergent message passing algorithms for
MAP LP-relaxations. In Neural Information Processing Systems, pages 553–560, 2008.
[6] T. Hazan and A. Shashua. Convergent message-passing algorithms for inference over general graphs
with convex free energies. In Conf. on Uncertainty in Artificial Intelligence, pages 264–273, 2008.
[7] J. Hiriart-Urruty and C. Lemaréchal. Fundamentals of Convex Analysis. Grundlehren Text Editions.
Springer, 2004.
[8] J. H. Kappes, B. Andres, F. A. Hamprecht, C. Schnörr, S. Nowozin, D. Batra, S. Kim, B. X.
Kausler, T. Kröger, J. Lellmann, N. Komodakis, B. Savchynskyy, and C. Rother. A comparative
study of modern inference techniques for structured discrete energy minimization problems. Intl. J.
of Computer Vision, 115(2):155–184, 2015.
[9] V. Kolmogorov. Convergent tree-reweighted message passing for energy minimization. IEEE Trans.
Pattern Analysis and Machine Intelligence, 28(10):1568–1583, 2006.
[10] V. Kolmogorov. A new look at reweighted message passing. IEEE Trans. on Pattern Analysis and
Machine Intelligence, 37(5), May 2015.
[11] N. Komodakis, N. Paragios, and G. Tziritas. MRF energy minimization and beyond via dual
decomposition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(3):531–552,
2011.
[12] V. K. Koval and M. I. Schlesinger. Dvumernoe programmirovanie v zadachakh analiza izobrazheniy
(Two-dimensional programming in image analysis problems). Automatics and Telemechanics, 8:149–
168, 1976. In Russian.
[13] V. A. Kovalevsky and V. K. Koval. A diffusion algorithm for decreasing the energy of the max-sum
labeling problem. Glushkov Institute of Cybernetics, Kiev, USSR. Unpublished, approx. 1975.
[14] A. Mackworth. Constraint satisfaction. In Encyclopaedia of Artificial Intelligence, pages 285–292.
Wiley, 1991.
[15] A. L. Martins, M. A. T. Figueiredo, P. M. Q. Aguiar, N. A. Smith, and E. P. Xing. An augmented
Lagrangian approach to constrained MAP inference. In Intl. Conf. on Machine Learning, pages
169–176, 2011.
[16] T. Meltzer, A. Globerson, and Y. Weiss. Convergent message passing algorithms: a unifying view.
In Conf. on Uncertainty in Artificial Intelligence, pages 393–401, 2009.
[17] P. Meseguer, F. Rossi, and T. Schiex. Soft constraints. In Handbook of Constraint Programming,
chapter 9. Elsevier, 2006.
[18] B. Savchynskyy. A bundle approach to efficient MAP-inference by Lagrangian relaxation. In IEEE
Conf. on Computer Vision and Pattern Recognition, pages 1688–1695, 2012.
12
[19] B. Savchynskyy, S. Schmidt, J. H. Kappes, and C. Schn orr. Efficient MRF energy minimization via
adaptive diminishing smoothing. In Conf. on Uncertainty in Artificial Intelligence, pages 746–755,
2012.
[20] M. I. Schlesinger and K. Antoniuk. Diffusion algorithms and structural recognition optimization
problems. Cybernetics and Systems Analysis, 47:175–192, 2011.
[21] M. I. Schlesinger and V. V. Giginjak. Solving (max,+) problems of structural pattern recognition using equivalent transformations. Upravlyayushchie Sistemy i Mashiny (Control Systems and
Machines), Kiev, Naukova Dumka, 1 and 2, 2007. In Russian, English translation available on www.
[22] D. Sontag, D. K. Choe, and Y. Li. Efficiently searching for frustrated cycles in MAP inference. In
Conf. on Uncertainty in Artificial Intelligence, pages 795–804, Corvallis, Oregon, 2012.
[23] P. Swoboda, J. Kuske, and B. Savchynskyy. A dual ascent framework for lagrangean decomposition
of combinatorial problems. In The IEEE Conference on Computer Vision and Pattern Recognition,
July 2017.
[24] R. Szeliski, R. Zabih, D. Scharstein, O. Veksler, V. Kolmogorov, A. Agarwal, M. Tappen, and
C. Rother. A comparative study of energy minimization methods for Markov random fields. In Eur.
Conf. on Computer Vision, pages II: 16–29, 2006.
[25] P. Tseng. Convergence of a block coordinate descent method for nondifferentiable minimization. J.
Optim. Theory Appl., 109(3):475–494, June 2001.
[26] S. Živný. The Complexity of Valued Constraint Satisfaction Problems. Cognitive Technologies.
Springer, 2012.
[27] M. J. Wainwright and M. I. Jordan. Graphical models, exponential families, and variational inference.
Foundations and Trends in Machine Learning, 1(1-2):1–305, 2008.
[28] T. Werner. A linear programming approach to max-sum problem: A review. IEEE Trans. Pattern
Analysis and Machine Intelligence, 29(7):1165–1179, July 2007.
[29] T. Werner. Revisiting the linear programming relaxation approach to Gibbs energy minimization and weighted constraint satisfaction. IEEE Trans. Pattern Analysis and Machine Intelligence,
32(8):1474–1488, August 2010.
13
| 1cs.CV
|
Local Linearizability∗
Andreas Haas1 , Thomas A. Henzinger2 , Andreas Holzer3 ,
Christoph M. Kirsch4 , Michael Lippautz1 , Hannes Payer1 ,
Ali Sezgin5 , Ana Sokolova4 , and Helmut Veith6,7
arXiv:1502.07118v4 [cs.PL] 24 Jun 2016
1
2
3
4
5
6
7
Google Inc.
IST Austria, Austria
University of Toronto, Canada
University of Salzburg, Austria
University of Cambridge, UK
Vienna University of Technology, Austria
Forever in our hearts
Abstract
The semantics of concurrent data structures is usually given by a sequential specification and a
consistency condition. Linearizability is the most popular consistency condition due to its simplicity and general applicability. Nevertheless, for applications that do not require all guarantees
offered by linearizability, recent research has focused on improving performance and scalability
of concurrent data structures by relaxing their semantics.
In this paper, we present local linearizability, a relaxed consistency condition that is applicable
to container-type concurrent data structures like pools, queues, and stacks. While linearizability
requires that the effect of each operation is observed by all threads at the same time, local
linearizability only requires that for each thread T, the effects of its local insertion operations and
the effects of those removal operations that remove values inserted by T are observed by all threads
at the same time. We investigate theoretical and practical properties of local linearizability and
its relationship to many existing consistency conditions. We present a generic implementation
method for locally linearizable data structures that uses existing linearizable data structures as
building blocks. Our implementations show performance and scalability improvements over the
original building blocks and outperform the fastest existing container-type implementations.
1998 ACM Subject Classification D.3.1 [Programming Languages]: Formal Definitions and
Theory—Semantics; E.1 [Data Structures]: Lists, stacks, and queues; D.1.3 [Software]: Programming Techniques—Concurrent Programming
Keywords and phrases (concurrent) data structures, relaxed semantics, linearizability
1
Introduction
Concurrent data structures are pervasive all along the software stack, from operating system
code to application software and beyond. Both correctness and performance are imperative
for concurrent data structure implementations. Correctness is usually specified by relating concurrent executions, admitted by the implementation, with sequential executions,
admitted by the sequential version of the data structure. The latter form the sequential
specification of the data structure. This relationship is formally captured by consistency
conditions, such as linearizability, sequential consistency, or quiescent consistency [25].
Linearizability [26] is the most accepted consistency condition for concurrent data structures due to its simplicity and general applicability. It guarantees that the effects of all
∗
This paper is an extended version of [17]
licensed under Creative Commons License CC-BY
Leibniz International Proceedings in Informatics
Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany
XX:2
Local Linearizability
operations by all threads are observed consistently. This global visibility requirement imposes the need of extensive synchronization among threads which may in turn jeopardize
performance and scalability. In order to enhance performance and scalability of implementations, recent research has explored relaxed sequential specifications [23, 40, 2], resulting in
well-performing implementations of concurrent data structures [2, 18, 23, 28, 38, 6]. Except
for [27], the space of alternative consistency conditions that relax linearizability has been
left unexplored to a large extent. In this paper, we explore (part of) this gap by investigating local linearizability, a novel consistency condition that is applicable to a large class
of concurrent data structures that we call container-type data structures, or containers for
short. Containers include pools, queues, and stacks. A fine-grained spectrum of consistency conditions enables us to describe the semantics of concurrent implementations more
precisely, e.g., we show in our appendix that work stealing queues [35] which could only be
proven to be linearizable wrt pool are actually locally linearizable wrt double-ended queue.
Local linearizability is a (thread-)local consistency
condition that guarantees that insertions per thread are
enq(1)
deq(2)
T1
observed consistently. While linearizability requires
enq(2)
deq(1)
T2
a consistent view over all insertions, we only require
that projections of the global history—so called thread- The thread-induced history of thread T1
induced histories—are linearizable. The induced his- is enclosed by a dashed line while the
thread-induced history of thread T2 is
tory of a thread T is a projection of a program exe- enclosed by a solid line.
cution to the insert-operations in T combined with all
remove-operations that remove values inserted by T irFigure 1 Local Linearizability
respective of whether they happen in T or not. Then,
the program execution is locally linearizable iff each thread-induced history is linearizable.
Consider the example (sequential) history depicted in Figure 1. It is not linearizable wrt a
queue since the values are not dequeued in the same order as they were enqueued. However,
each thread-induced history is linearizable wrt a queue and, therefore, the overall execution
is locally linearizable wrt a queue. In contrast to semantic relaxations based on relaxing
sequential semantics such as [23, 2], local linearizability coincides with sequential correctness for single-threaded histories, i.e., a single-threaded and, therefore, sequential history is
locally linearizable wrt a given sequential specification if and only if it is admitted by the
sequential specification.
Local linearizability is to linearizability what coherence is to sequential consistency. Coherence [22], which is almost universally accepted as the absolute minimum that a shared
memory system should satisfy, is the requirement that there exists a unique global order per
shared memory location. Thus, while all accesses by all threads to a given memory location
have to conform to a unique order, consistent with program order, the relative ordering of
accesses to multiple memory locations do not have to be the same. In other words, coherence
is sequential consistency per memory location. Similarly, local linearizability is linearizability per local history. In our view, local linearizability offers enough consistency for the
correctness of many applications as it is the local view of the client that often matters. For
example, in a locally linearizable queue each client (thread) has the impression of using a
perfect queue—no reordering will ever be observed among the values inserted by a single
thread. Such guarantees suffice for many e-commerce and cloud applications. Implementations of locally linearizable data structures have been successfully applied for managing
free lists in the design of the fast and scalable memory allocator scalloc [5]. Moreover, except for fairness, locally linearizable queues guarantee all properties required from Dispatch
Queues [1], a common concurrency programming mechanism on mobile devices.
A. Haas et al.
In this paper, we study theoretical and practical properties of local linearizability. Local
linearizability is compositional—a history over multiple concurrent objects is locally linearizable iff all per-object histories are locally linearizable (see Thm. 12) and locally linearizable
container-type data structures, including queues and stacks, admit only “sane” behaviours—
no duplicated values, no values returned from thin air, and no values lost (see Prop. 4). Local
linearizability is a weakening of linearizability for a natural class of data structures including
pools, queues, and stacks (see Sec. 4). We compare local linearizability to linearizability,
sequential, and quiescent consistency, and to many shared-memory consistency conditions.
Finally, local linearizability leads to new efficient implementations. We present a generic
implementation scheme that, given a linearizable implementation of a sequential specification S, produces an implementation that is locally linearizable wrt S (see Sec. 6). Our
implementations show dramatic improvements in performance and scalability. In most cases
the locally linearizable implementations scale almost linearly and even outperform state-ofthe-art pool implementations. We produced locally linearizable variants of state-of-the-art
concurrent queues and stacks, as well as of the relaxed data structures from [23, 28]. The
latter are relaxed in two dimensions: they are locally linearizable (the consistency condition is relaxed) and are out-of-order-relaxed (the sequential specification is relaxed). The
speedup of the locally linearizable implementation to the fastest linearizable queue (LCRQ)
and stack (TS Stack) implementation at 80 threads is 2.77 and 2.64, respectively. Verification of local linearizability, i.e. proving correctness, for each of our new locally linearizable
implementations is immediate, given that the starting implementations are linearizable.
2
Semantics of Concurrent Objects
The common approach to define the semantics of an implementation of a concurrent data
structure is (1) to specify a set of valid sequential behaviors—the sequential specification, and
(2) to relate the admissible concurrent executions to sequential executions specified by the
sequential specification—via the consistency condition. That means that an implementation
of a concurrent data structure actually corresponds to several sequential data structures, and
vice versa, depending on the consistency condition used. A (sequential) data structure D is
an object with a set of method calls Σ. We assume that method calls include parameters,
i.e., input and output values from a given set of values. The sequential specification S of D
is a prefix-closed subset of Σ∗ . The elements of S are called D-valid sequences. For ease of
presentation, we assume that each value in a data structure can be inserted and removed at
most once. This is without loss of generality, as we may see the set of values as consisting
of pairs of elements (core values) and version numbers, i.e. V = E × N. Note that this
is a technical assumption that only makes the presentation and the proofs simpler, it is
not needed and not done in locally linearizable implementations. While elements may be
inserted and removed multiple times, the version numbers provide uniqueness of values. Our
assumption ensures that whenever a sequence s is part of a sequential specification S, then,
each method call in s appears exactly once. An additional core value, that is not an element,
is empty. It is returned by remove method calls that do not find an element to return. We
denote by Emp the set of values that are versions of empty, i.e., Emp = {empty} × N.
I Definition 1 (Appears-before Order, Appears-in Relation). Given a sequence s ∈ Σ∗ in which
each method call appears exactly once, we denote by ≺s the total appears-before order over
method calls in s. Given a method call m ∈ Σ, we write m ∈ s for m appears in s.
Throughout the paper, we will use pool, queue, and stack as typical examples of containers. We specify their sequential specifications in an axiomatic way [24], i.e., as sets of
XX:3
XX:4
Local Linearizability
(1)
(2)
(3)
(4)
(5)
∀i, j ∈ {1, . . . , n}. s = m1 . . . mn ∧ mi = mj ⇒ i = j
∀x ∈ V. r(x) ∈ s ⇒ i(x) ∈ s ∧ i(x) ≺s r(x)
∀e ∈ Emp. ∀x ∈ V. i(x) ≺s r(e) ⇒ r(x) ≺s r(e)
∀x, y ∈ V. i(x) ≺s i(y) ∧ r(y) ∈ s ⇒ r(x) ∈ s ∧ r(x) ≺s r(y)
∀x, y ∈ V. i(x) ≺s i(y) ≺s r(x) ⇒ r(y) ∈ s ∧ r(y) ≺s r(x)
Table 1 The pool axioms (1), (2), (3); the queue order axiom (4); the stack order axiom (5)
axioms that exactly define the valid sequences.
I Definition 2 (Pool, Queue, & Stack). A pool, queue, and stack with values in a set V
have the sets of methods ΣP = {ins(x), rem(x) | x ∈ V } ∪ {rem(e) | e ∈ Emp}, ΣQ =
{enq(x), deq(x) | x ∈ V } ∪ {deq(e) | e ∈ Emp}, and ΣS = {push(x), pop(x) | x ∈ V } ∪
{pop(e) | e ∈ Emp}, respectively. We denote the sequential specification of a pool by SP ,
the sequential specification of a queue by SQ , and the sequential specification of a stack
by SS . A sequence s ∈ Σ∗P belongs to SP iff it satisfies axioms (1) - (3) in Table 1—the
pool axioms—when instantiating i() with ins() and r() with rem(). We keep axiom (1)
for completeness, although it is subsumed by our assumption that each value is inserted
and removed at most once. Specification SQ contains all sequences s that satisfy the pool
axioms and axiom (4)—the queue order axiom—after instantiating i() with enq() and r()
with deq(). Finally, SS contains all sequences s that satisfy the pool axioms and axiom
(5)—the stack order axiom—after instantiating i() with push() and r() with pop().
We represent concurrent executions via concurrent histories. An example history is
shown in Figure 1. Each thread executes a sequence of method calls from Σ; method
calls executed by different threads may overlap (which does not happen in Figure 1). The
real-time duration of method calls is irrelevant for the semantics of concurrent objects;
all that matters is whether method calls overlap. Given this abstraction, a concurrent
history is fully determined by a sequence of invocation and response events of method calls.
We distinguish method invocation and response events by augmenting the alphabet. Let
Σi = {mi | m ∈ Σ} and Σr = {mr | m ∈ Σ} denote the sets of method-invocation events
and method-response events, respectively, for the method calls in Σ. Moreover, let I be the
set of thread identifiers. Let ΣIi = {mki | m ∈ Σ, k ∈ I} and ΣIr = {mkr | m ∈ Σ, k ∈ I}
denote the sets of method-invocation and -response events augmented with identifiers of
executing threads. For example, mki is the invocation of method call m by thread k. Before
we proceed, we mention a standard notion that we will need in several occasions.
I Definition 3 (Projection). Let s be a sequence over alphabet Σ and M ⊆ Σ. By s|M
we denote the projection of s on the symbols in M , i.e., the sequence obtained from s by
removing all symbols that are not in M .
I Definition 4 (History). A (concurrent) history h is a sequence in (ΣIi ∪ ΣIr )∗ where (1) no
invocation or response event appears more than once, i.e., if h = m1 . . . mn and mh = mk∗ (x)
and mj = ml∗ (x), for ∗ ∈ {i, r}, then h = j and k = l, and (2) if a response event mkr appears
in h, then the corresponding invocation event mki also appears in h and mi ≺h mr .
I Example 5. A queue history (left) and its formal representation as a sequence (right):
T1
T2
enq(2)
deq(1)
enq(1)
enq(2)1i enq(1)2i enq(2)1r deq(1)1i enq(1)2r deq(1)1r
A. Haas et al.
A history is sequential if every response event is immediately preceded by its matching invocation event and vice versa. Hence, we may ignore thread identifiers and identify
a sequential history with a sequence in Σ∗ , e.g., enq(1)enq(2)deq(2)deq(1) identifies the
sequential history in Figure 1.
A history h is well-formed if h|k is sequential for every thread identifier k ∈ I where h|k
denotes the projection of h on the set {mki | m ∈ Σ} ∪ {mkr | m ∈ Σ} of events that are local
to thread k. From now on we will use the term history for well-formed history. Also, we
may omit thread identifiers if they are not essential in a discussion.
A history h determines a partial order on its set of method calls, the precedence order:
I Definition 6 (Appears-in Relation, Precedence Order). The set of method calls of a history h
is M (h) = {m | mi ∈ h}. A method call m appears in h, notation m ∈ h, if m ∈ M (h).
The precedence order for h is the partial order <h such that, for m, n ∈ h, we have that
m <h n iff mr ≺h ni . By <kh we denote <h|k , the subset of the precedence order that relates
pairs of method calls of thread k, i.e., the program order of thread k.
We can characterize a sequential history as a history whose precedence order is total. In
particular, the precedence order <s of a sequential history s coincides with its appears-before
order ≺s . The total order for history s in Fig. 1 is enq(1) <s enq(2) <s deq(2) <s deq(1).
I Definition 7 (Projection to a set of method calls). Let h be a history, M ⊆ Σ, MiI = {mki |
m ∈ M, k ∈ I}, and MrI = {mkr | m ∈ M, k ∈ I}. Then, we write h|M for h|(MiI ∪ MrI ).
Note that h|M inherits h’s precedence order: m <h|M n ⇔ m ∈ M ∧ n ∈ M ∧ m <h n
A history h is complete if the response of every invocation event in h appears in h. Given
a history h, Complete(h) denotes the set of all completions of h, i.e., the set of all complete
histories that are obtained from h by appending missing response events and/or removing
pending invocation events. Note that Complete(h) = {h} iff h is a complete history.
A concurrent data structure D over a set of methods Σ is a (prefix-closed) set of concurrent histories over Σ. A history may involve several concurrent objects. Let O be a set of
concurrent objects with individual sets of method calls Σq and sequential specifications Sq
for each object q ∈ O. A history h over O is a history over the (disjoint) union of method
S
calls of all objects in O, i.e., it has a set of method calls q∈O {q.m | m ∈ Σq }. The added
prefix q. ensures that the union is disjoint. The projection of h to an object q ∈ O, denoted
by h|q, is the history with a set of method calls Σq obtained by removing the prefix q. in
every method call in h|{q.m | m ∈ Σq }.
I Definition 8 (Linearizability [26]). A history h is linearizable wrt the sequential specification S if there is a sequential history s ∈ S and a completion hc ∈ Complete(h) such
that (1) s is a permutation of hc , and (2) s preserves the precedence order of hc , i.e., if
m <hc n, then m <s n. We refer to s as a linearization of h. A concurrent data structure
D is linearizable wrt S if every history h of D is linearizable wrt S. A history h over a set
of concurrent objects O is linearizable wrt the sequential specifications Sq for q ∈ O if there
exists a linearization s of h such that s|q ∈ Sq for each object q ∈ O.
3
Local Linearizability
Local linearizability is applicable to containers whose set of method calls is a disjoint union
Σ = Ins ∪ Rem ∪ DOb ∪ SOb of insertion method calls Ins, removal method calls Rem, dataobservation method calls DOb, and (global) shape-observation method calls SOb. Insertions
(removals) insert (remove) a single value in the data set V or empty; data observations return
XX:5
XX:6
Local Linearizability
a single value in V ; shape observations return a value (not necessarily in V ) that provides
information on the shape of the state, for example, the size of a data structure. Examples
of data observations are head(x) (queue), top(x) (stack), and peek(x) (pool). Examples of
shape observations are empty(b) that returns true if the data structure is empty and false
otherwise, and size(n) that returns the number of elements in the data structure.
Even though we refrain from formal definitions, we want to stress that a valid sequence
of a container remains valid after deleting observer method calls:
S | (Ins ∪ Rem) ⊆ S.
(1)
There are also containers with multiple insert/remove methods, e.g., a double-ended queue
(deque) is a container with insert-left, insert-right, remove-left, and remove-right methods,
to which local linearizability is also applicable. However, local linearizability requires that
each method call is either an insertion, or a removal, or an observation. As a consequence,
set is not a container according to our definition, as in a set ins(x) acts as a global observer
first, checking whether (some version of) x is already in the set, and if not inserts x. Also
hash tables are not containers for a similar reason.
Note that the arity of each method call in a container being one excludes data structures
like snapshot objects. It is possible to deal with higher arities in a fairly natural way,
however, at the cost of complicated presentation. We chose to present local linearizability
on simple containers only. We present the definition of local linearizability without shape
observations here and discuss shape observations in Appendix A.
I Definition 9 (In- and out-methods). Let h be a container history. For each thread T
we define two subsets of the methods in h, called in-methods IT and out-methods OT of
thread T , respectively:
IT = {m | m ∈ M (h|T ) ∩ Ins}
OT =
{m(a) ∈ M (h) ∩ Rem | ins(a) ∈ IT } ∪ {m(e) ∈ M (h) ∩ Rem | e ∈ Emp}
∪ {m(a) ∈ M (h) ∩ DOb | ins(a) ∈ IT }.
Hence, the in-methods for thread T are all insertions performed by T . The out-methods
are all removals and data observers that return values inserted by T . Removals that remove
the value empty are also automatically added to the out-methods of T as any thread (and
hence also T ) could be the cause of “inserting” empty. This way, removals of empty serve as
means for global synchronization. Without them each thread could perform all its operations
locally without ever communicating with the other threads. Note that the out-methods OT
of thread T need not be performed by T , but they return values that are inserted by T .
I Definition 10 (Thread-induced History). Let h be a history. The thread-induced history
hT is the projection of h to the in- and out-methods of thread T , i.e., hT = h| (IT ∪ OT ).
I Definition 11 (Local Linearizability). A history h is locally linearizable wrt a sequential
specification S if (1) each thread-induced history hT is linearizable wrt S, and (2) the
thread-induced histories hT form a decomposition of h, i.e., m ∈ h ⇒ m ∈ hT for some
thread T . A data structure D is locally linearizable wrt S if every history h of D is locally
linearizable wrt S. A history h over a set of concurrent objects O is locally linearizable wrt
the sequential specifications Sq for q ∈ O if each thread-induced history is linearizable over O
and the thread-induced histories form a decomposition of h, i.e., q.m ∈ h ⇒ q.m ∈ hT for
some thread T .
Local linearizability is sequentially correct, i.e., a single-threaded (necessarily sequential)
history h is locally linearizable wrt a sequential specification S iff h ∈ S. Like linearizabil-
A. Haas et al.
XX:7
ity [25], local linearizability is compositional. The complete proof of the following theorem
and missing or extended proofs of all following properties can be found in Appendix B.
I Theorem 12 (Compositionality). A history h over a set of objects O with sequential specifications Sq for q ∈ O is locally linearizable iff h|q is locally linearizable wrt Sq for every q ∈ O.
Proof (Sketch). The property follows from the compositionality of linearizability and the
fact that (h|q)T = hT |q for every thread T and object q.
J
The Choices Made. Splitting a global history into subhistories and requiring consistency
for each of them is central to local linearizability. While this is common in shared-memory
consistency conditions [22, 31, 32, 3, 16, 4, 20], our study of local linearizability is a first
step in exploring subhistory-based consistency conditions for concurrent objects.
We chose thread-induced subhistories since thread-locality reduces contention in concurrent objects and is known to lead to high performance as confirmed by our experiments. To
assign method calls to thread-induced histories, we took a data-centric point of view by (1)
associating data values to threads, and (2) gathering all method calls that insert/return a
data value into the subhistory of the associated thread (Def. 9). We associate data values to
the thread that inserts them. One can think of alternative approaches, for example, associate with a thread the values that it removed. In our view, the advantages of our choice are
clear: First, by assigning inserted values to threads, every value in the history is assigned
to some thread. In contrast, in the alternative approach, it is not clear where to assign the
values that are inserted but not removed. Second, assigning inserted values to the inserting
thread enables eager removals and ensures progress in locally linearizable data structures.
In the alternative approach, it seems like the semantics of removing empty should be local.
An orthogonal issue is to assign values from shape observations to threads. In Appendix A, we discuss two meaningful approaches and show how local linearizability can be
extended towards shape and data observations that appear in insertion operations of sets.
Finally, we have to choose a consistency condition required for each of the subhistories.
We chose linearizability as it is the best (strong) consistency condition for concurrent objects.
4
Local Linearizability vs. Linearizability
We now investigate the connection between local linearizability and linearizability.
I Proposition 1 (Lin 1). In general, linearizability does not imply local linearizability.
Proof. We provide an example of a data structure that is linearizable but not locally
linearizable. Consider a sequential specification SNearlyQ which behaves like a queue except when the first two insertions were performed without a removal in between—then
the first two elements are removed out of order. Formally, s ∈ SNearlyQ iff (1) s =
s1 enq(a)enq(b)s2 deq(b)s3 deq(a)s4 where s1 enq(a)enq(b)s2 deq(a)s3 deq(b)s4 ∈ SQ and s1 ∈
{deq(e) | e ∈ Emp}∗ for some a, b ∈ V , or (2) s ∈ SQ and s 6= s1 enq(a)enq(b)s2 for
s1 ∈ {deq(e) | e ∈ Emp}∗ and a, b ∈ V . The example below is linearizable wrt SNearlyQ .
However, T1 ’s induced history enq(1)enq(2)deq(1)deq(2) is not.
T1
T2
enq(1)
enq(2) deq(3)
enq(3)
deq(2)
deq(1)
J
The following condition on a data structure specification is sufficient for linearizability
to imply local linearizability and is satisfied, e.g., by pool, queue, and stack.
XX:8
Local Linearizability
T1
T2
i(1)
r(empty)
i(2)
r(2)
r(1)
Figure 2 LL, not SC (Pool, Queue, Stack)
T1
T2
i(1)
r(1)
r(empty)
Figure 3 SC, not LL (Pool, Queue, Stack)
I Definition 13 (Closure under Data-Projection). A seq. specification S over Σ is closed
under data-projection 1 iff for all s ∈ S and all V 0 ⊆ V , s|{m(x) ∈ Σ | x ∈ V 0 ∪ Emp} ∈ S.
For s = enq(1)enq(3)enq(2)deq(3)deq(1)deq(2) we have s ∈ SNearlyQ , but s|{enq(x), deq(x) |
x ∈ {1, 2} ∪ Emp} ∈
/ SNearlyQ , i.e., SNearlyQ is not closed under data-projection.
I Proposition 2 (Lin 2). Linearizability implies local linearizability for sequential specifications that are closed under data-projection.
Proof (Sketch). The property follows from Definition 13 and Equation (1).
J
There exist corner cases where local linearizability coincides with linearizability, e.g., for
S = ∅ or S = Σ∗ , or for single-producer/multiple-consumer histories.
We now turn our attention to pool, queue, and stack.
I Proposition 3. The seq. specifications SP , SQ , and SS are closed under data-projection.
Proof (Sketch). Let s ∈ SP , V 0 ⊆ V , and let s0 = s| ({ins(x), rem(x) | x ∈ V 0 ∪ Emp}).
Then, it suffices to check that all axioms for pool (Definition 2 and Table 1) hold for s0 . J
I Theorem 14 (Pool & Queue & Stack, Lin). For pool, queue, and stack, local linearizability
is (strictly) weaker than linearizability.
Proof. Linearizability implies local linearizability for pool, queue, and stack as a consequence of Proposition 2 and Proposition 3. The history in Figure 2 is locally linearizable
but not linearizable wrt pool, queue and stack (after suitable renaming of method calls). J
Although local linearizability wrt a pool does not imply linearizability wrt a pool (Theorem 14), it still guarantees several properties that ensure sane behavior as stated next.
I Proposition 4 (LocLin Pool). Let h be a locally linearizable history wrt a pool. Then:
1. No value is duplicated, i.e., every remove method appears in h at most once.
2. No out-of-thin-air values, i.e., ∀x ∈ V. rem(x) ∈ h ⇒ ins(x) ∈ h ∧ rem(x)6<h ins(x).
3. No value is lost, i.e., ∀x ∈ V. ∀e ∈ Emp. rem(e) <h rem(x) ⇒ ins(x) 6<h rem(e) and
∀x ∈ V. ∀e ∈ Emp. ins(x) <h rem(e) ⇒ rem(x) ∈ h ∧ rem(e)6<h rem(x).
Proof. By direct unfolding of the definitions.
J
Note that if a history h is linearizable wrt a pool, then all of the three stated properties
hold, as a consequence of linearizability and the definition of SP .
5
Local Linearizability vs. Other Relaxed Consistency Conditions
We compare local linearizability with other classical consistency conditions to better understand its guarantees and implications.
1
The same notion has been used in [7] under the name closure under projection.
A. Haas et al.
XX:9
Sequential Consistency (SC). A history h is sequentially consistent [25, 30] wrt a sequential specification S, if there exists a sequential history s ∈ S and a completion hc ∈
Complete(h) such that (1) s is a permutation of hc , and (2) s preserves each thread’s program order, i.e., if m <Th n, for some thread T , then m <s n. We refer to s as a sequential
witness of h. A data structure D is sequentially consistent wrt S if every history h of D is
sequentially consistent wrt S.
Sequential consistency is a useful consistency condition for shared memory but it is not
really suitable for data structures as it allows for behavior that excludes any coordination
between threads [39]: an implementation of a data structure in which every thread uses a
dedicated copy of a sequential data structure without any synchronization is sequentially
consistent. A sequentially consistent queue might always return empty in one (consumer)
thread as the point in time of the operation can be moved, e.g., see Figure 3. In a producerconsumer scenario such a queue might end up with some threads not doing any work.
I Theorem 15 (Pool, Queue & Stack, SC). For pool, queue, and stack, local linearizability
is incomparable to sequential consistency.
J
Figures 2 and 3 give example histories that show the statement of Theorem 15. In contrast
to local linearizability, sequential consistency is not compositional [25].
(Quantitative) Quiescent Consistency (QC & QQC). Like linearizability and sequential
consistency, quiescent consistency [13, 25] also requires the existence of a sequential history,
a quiescent witness, that satisfies the sequential specification. All three consistency conditions impose an order on the method calls of a concurrent history that a witness has to
preserve. Quiescent consistency uses the concept of quiescent states to relax the requirement
of preserving the precedence order imposed by linearizability. A quiescent state is a point
in a history at which there are no pending invocation events (all invoked method calls have
already responded). In a quiescent witness, a method call m has to appear before a method
call n if and only if there is a quiescent state between m and n. Method calls between
two consecutive quiescent states can be ordered arbitrarily. Quantitative quiescent consistency [27] refines quiescent consistency by bounding the number of reorderings of operations
between two quiescent states based on the concurrent behavior between these two states.
The next result about quiescent consistency for pool is needed to establish the connection
between quiescent consistency and local linearizability.
I Proposition 5. A pool history h satisfying 1.-3. of Prop. 4 is quiescently consistent.
J
From Prop. 4 and 5 follows that local linearizability implies quiescent consistency for pool.
I Theorem 16 (Pool, Queue & Stack, QC). For pool, local linearizability is (strictly) stronger
than quiescent consistency. For queue and stack, local linearizability is incomparable to
quiescent consistency.
J
Local linearizability also does not imply the stronger condition of quantitative quiescent consistency. Like local linearizability, quiescent consistency and quantitative quiescent
consistency are compositional [25, 27]. For details, please see Appendix D.
Consistency Conditions for Distributed Shared Memory. There is extensive research on
consistency conditions for distributed shared memory [3, 4, 8, 16, 20, 22, 30, 31, 32]. In
Appendix E, we compare local linearizability against coherence, PRAM consistency, processor consistency, causal consistency, and local consistency. All these conditions split a
history into subhistories and require consistency of the subhistories. For our comparison,
XX:10
Local Linearizability
T1
T2
ins(1)
ins(2) head(2) head(1) head(2) head(1)
Figure 4 Problematic shared-memory history.
we first define a sequential specification SM for a single memory location. We assume that
each memory location is preinitialized with a value vinit ∈ V . A read-operation returns
the value of the last write-operation that was performed on the memory location or vinit if
there was no write-operation. We denote write-operations by ins and read-operations by
head. Formally, we define SM as SM = {head(vinit )}? · {ins(v)head(v)i | i ≥ 0, v ∈ V }? .
Note that read-operations are data observations and the same value can be read multiple
times. For brevity, we only consider histories that involve a single memory location. In the
following, we summarize our comparison. For details, please see Appendix E.
While local linearizability is well-suited for concurrent data structures, this is not necessarily true for the mentioned shared-memory consistency conditions. On the other hand,
local linearizability appears to be problematic for shared memory. Consider the locally
linearizable history in Figure 4. There, the read values oscillate between different values
that were written by different threads. Therefore, local linearizability does not imply any
of the shared-memory consistency conditions. In Appendix E, we further show that local
linearizability is incomparable to all considered shared-memory conditions.
6
Locally Linearizable Implementations
In this section, we focus on locally linearizable data structure implementations that are generic as follows: Choose a linearizable implementation of a data structure Φ wrt a sequential
specification SΦ , and we turn it into a (distributed) data structure called LLD Φ that is
locally linearizable wrt SΦ . An LLD implementation takes several copies of Φ (that we call
backends) and assigns to each thread T a backend ΦT . Then, when thread T inserts an
element into LLD Φ, the element is inserted into ΦT , and when an arbitrary thread removes
an element from LLD Φ, the element is removed from some ΦT eagerly, i.e., if no element is
found in the attempted backend ΦT the search for an element continues through all other
backends. If no element is found in one round through the backends, then we return empty.
I Proposition 6 (LLD correctness). Let Φ be a data structure implementation that is linearizable wrt a sequential specification SΦ . Then LLD Φ is locally linearizable wrt SΦ .
Proof. Let h be a history of LLD Φ. The crucial observation is that each thread-induced
history hT is a backend history of ΦT and hence linearizable wrt SΦ .
J
Any number of copies (backends) is allowed in this generic implementation of LLD Φ.
If we take just one copy, we end up with a linearizable implementation. Also, any way of
choosing a backend for removals is fine. However, both the number of backends and the
backend selection strategy upon removals affect the performance significantly. In our LLD
Φ implementations we use one backend per thread, resulting in no contention on insertions,
and always attempt a local remove first. If this does not return an element, then we continue
a search through all other backends starting from a randomly chosen backend.
LLD Φ is an implementation closely related to Distributed Queues (DQs) [18]. A DQ
is a (linearizable) pool that is organized as a single segment of length ` holding ` backends.
DQs come in different flavours depending on how insert and remove methods are distributed
A. Haas et al.
across the segment when accessing backends. No DQ variant in [18] follows the LLD approach described above. Moreover, while DQ algorithms are implemented for a fixed number
of backends, LLD Φ implementations manage a segment of variable size, one backend per
(active) thread. Note that the strategy of selecting backends in the LLD Φ implementations is similar to other work in work stealing [35]. However, in contrast to this work
our data structures neither duplicate nor lose elements. LLD (stack) implementations have
been successfully applied for managing free lists in the fast and scalable memory allocator
scalloc [5]. The guarantees provided by local linearizability are not needed for the correctness of scalloc, i.e., the free lists could also use a weak pool (pool without a linearizable
emptiness check). However, the LLD stack implementations provide good caching behavior
when threads operate on their local stacks whereas a weak pool would potentially negatively
impact performance.
We have implemented LLD variants of strict and relaxed queue and stack implementations. None of our implementations involves observation methods, but the LLD algorithm
can easily be extended to support observation methods. For details, please see App. F.4.
Finally, let us note that we have also experimented with other locally linearizable implementations that lacked the genericity of the LLD implementations, and whose performance
evaluation did not show promising results (see App. G). As shown in Sec. 4, a locally linearizable pool is not a linearizable pool, i.e., it lacks a linearizable emptiness check. Indeed,
LLD implementations do not provide a linearizable emptiness check, despite of eager removes. We provide LL+ D Φ, a variant of LLD Φ, that provides a linearizable emptiness
check under mild conditions on the starting implementation Φ (see App. F.1 for details).
Experimental Evaluation. All experiments ran on a uniform memory architecture (UMA)
machine with four 10-core 2GHz Intel Xeon E7-4850 processors supporting two hardware
threads (hyperthreads) per core, 128GB of main memory, and Linux kernel version 3.8.0.
We also ran the experiments without hyper-threading resulting in no noticeable difference.
The CPU governor has been disabled. All measurements were obtained from the artifactevaluated Scal benchmarking framework [12, 19, 11], where you can also find the code of
all involved data structures. Scal uses preallocated memory (without freeing it) to avoid
memory management artifacts. For all measurements we report the arithmetic mean and
the 95% confidence interval (sample size=10, corrected sample standard deviation).
In our experiments, we consider the linearizable queues Michael-Scott queue (MS) [34]
and LCRQ [36] (improved version [37]), the linearizable stacks Treiber stack (Treiber) [42]
and TS stack [14], the k-out-of-order relaxed k-FIFO queue [28] and k-Stack [23] and linearizable well-performing pools based on distributed queues using random balancing [18] (1-RA
DQ for queue, and 1-RA DS for stack). For each of these implementations (but the pools)
we provide LLD variants (LLD LCRQ, LLD TS stack, LLD k-FIFO, and LLD k-Stack) and,
when possible, LL+ D variants (LL+ D MS queue and LL+ D Treiber stack). Making the
pools locally linearizable is not promising as they are already distributed. Whenever LL+ D
is achievable for a data structure implementation Φ we present only results for LL+ D Φ as,
in our workloads, LLD Φ and LL+ D Φ implementations perform with no visible difference.
We evaluate the data structures on a Scal producer-consumer benchmark where each
producer and consumer is configured to execute 106 operations. To control contention, we
add a busy wait of 5µs between operations. This is important as too high contention results in measuring hardware or operating system (e.g., scheduling) artifacts. The number of
threads ranges between 2 and 80 (number of hardware threads) half of which are producers
and half consumers. To relate performance and scalability we report the number of data
XX:11
million operations per sec (more is better)
Local Linearizability
26
24
22
20
18
16
14
12
10
8
6
4
2
0
MS
LCRQ
k-FIFO
LL+D MS
LLD LCRQ
LLD k-FIFO
1-RA DQ
2
10
20
30
40
50
60
70
80
number of threads
“queue-like” data structures
million operations per sec (more is better)
XX:12
26
24
22
20
18
16
14
12
10
8
6
4
2
0
Treiber
TS Stack
k-Stack
LL+D Treiber
LLD TS Stack
LLD k-Stack
1-RA DS
2
10
20
30
40
50
60
70
80
number of threads
“stack-like” data structures
Figure 5 Performance and scalability of producer-consumer microbenchmarks with an increasing
number of threads on a 40-core (2 hyperthreads per core) machine
structure operations per second. Data structures that require parameters to be set are configured to allow maximum parallelism for the producer-consumer workload with 80 threads.
This results in k = 80 for all k-FIFO and k-Stack variants (40 producers and 40 consumers
in parallel on a single segment), p = 80 for 1-RA-DQ and 1-RA-DS (40 producers and
40 consumers in parallel on different backends). The TS Stack algorithm also needs to be
configured with a delay parameter. We use optimal delay (7µs) for the TS Stack and zero
delay for the LLD TS Stack, as delays degrade the performance of the LLD implementation.
Figure 5 shows the results of the producer-consumer benchmarks. Similar to experiments performed elsewhere [14, 23, 28, 36] the well-known algorithms MS and Treiber do
not scale for 10 or more threads. The state-of-the-art linearizable queue and stack algorithms
LCRQ and TS-interval Stack either perform competitively with their k-out-of-order relaxed
counter parts k-FIFO and k-Stack or even outperform and outscale them. For any implementation Φ, LLD Φ and LL+ D Φ (when available) perform and scale significantly better
than Φ does, even slightly better than the state-of-the-art pool that we compare to. The best
improvement show LLD variants of MS queue and Treiber stack. The speedup of the locally
linearizable implementation to the fastest linearizable queue (LCRQ) and stack (TS Stack)
A. Haas et al.
implementation at 80 threads is 2.77 and 2.64, respectively. The performance degradation
for LCRQ between 30 and 70 threads aligns with the performance of fetch-and-inc—the
CPU instruction that atomically retrieves and modifies the contents of a memory location—
on the benchmarking machine, which is different on the original benchmarking machine [36].
LCRQ uses fetch-and-inc as its key atomic instruction.
7
Conclusion & Future Work
Local linearizability splits a history into a set of thread-induced histories and requires consistency of all such. This yields an intuitive consistency condition for concurrent objects
that enables new data structure implementations with superior performance and scalability.
Local linearizability has desirable properties like compositionality and well-behavedness for
container-type data structures. As future work, it is interesting to investigate the guarantees
that local linearizability provides to client programs along the line of [15].
Acknowledgments
This work has been supported by the National Research Network RiSE on Rigorous Systems
Engineering (Austrian Science Fund (FWF): S11402-N23, S11403-N23, S11404-N23, S11411N23), a Google PhD Fellowship, an Erwin Schrödinger Fellowship (Austrian Science Fund
(FWF): J3696-N26), EPSRC grants EP/H005633/1 and EP/K008528/1, the Vienna Science
and Technology Fund (WWTF) trough grant PROSEED, the European Research Council
(ERC) under grant 267989 (QUAREM) and by the Austrian Science Fund (FWF) under
grant Z211-N23 (Wittgenstein Award).
References
1
2
3
4
5
6
7
8
9
10
URL:
https://developer.apple.com/library/ios/documentation/General/
Conceptual/ConcurrencyProgrammingGuide/OperationQueues/OperationQueues.
html.
Y. Afek, G. Korland, and E. Yanovsky. Quasi-Linearizability: Relaxed Consistency for
Improved Concurrency. In OPODIS, pages 395–410, 2010.
M. Ahamad, R.A. Bazzi, R. John, P. Kohli, and G. Neiger. The Power of Processor
Consistency. In SPAA, pages 251–260, 1993.
M. Ahamad, G. Neiger, J.E. Burns, P. Kohli, and P.W. Hutto. Causal memory: definitions,
implementation, and programming. Distributed Computing, 9(1):37–49, 1995.
M. Aigner, C. M. Kirsch, M. Lippautz, and A. Sokolova. Fast, multicore-scalable, lowfragmentation memory allocation through large virtual memory and global data structures.
In OOPSLA, pages 451–469, 2015.
D. Alistarh, J. Kopinsky, J. Li, and N. Shavit. The SprayList: A Scalable Relaxed Priority
Queue. In PPoPP, pages 11–20, 2015.
A. Bouajjani, M. Emmi, C. Enea, and J. Hamza. On Reducing Linearizability to State
Reachability. In ICALP, pages 95–107, 2015.
S. Burckhardt, A. Gotsman, H. Yang, and M. Zawirski. Replicated Data Types: Specification, Verification, Optimality. In POPL, pages 271–284, 2014.
A. Cerone, A. Gotsman, and H. Yang. Parameterised Linearisability. In ICALP, pages
98–109, 2014.
S. Chakraborty, T. A. Henzinger, A. Sezgin, and V. Vafeiadis. Aspect-Oriented Linearizability Proofs. Logical Methods in Computer Science, 11(1:20):1–33, 2015.
XX:13
XX:14
Local Linearizability
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
POPL 2015 Artifact Evaluation Committee. POPL 2015 Artifact Evaluation. Accessed on
01/14/2015. URL: http://popl15-aec.cs.umass.edu/home/.
Computational Systems Group, University of Salzburg. Scal: High-Performance MulticoreScalable Computing. URL: http://scal.cs.uni-salzburg.at.
J. Derrick, B. Dongol, G. Schellhorn, B. Tofan, O. Travkin, and H. Wehrheim. Quiescent
Consistency: Defining and Verifying Relaxed Linearizability. In FM, pages 200–214, 2014.
M. Dodds, A. Haas, and C.M. Kirsch. A Scalable, Correct Time-Stamped Stack. In POPL,
pages 233–246, 2015.
I. Filipovic, P.W. O’Hearn, N. Rinetzky, and H. Yang. Abstraction for concurrent objects.
Theor. Comput. Sci., 411(51-52):4379–4398, 2010.
J.R. Goodman. Cache consistency and sequential consistency. University of WisconsinMadison, Computer Sciences Department, 1991.
A. Haas, T.A. Henzinger, A. Holzer, C.M. Kirsch, M. Lippautz, H. Payer, A. Sezgin,
A. Sokolova, and H. Veith. Local Linearizability for Concurrent Container-Type Data
Structures. In CONCUR, 2016.
A. Haas, T.A. Henzinger, C.M. Kirsch, M. Lippautz, H. Payer, A. Sezgin, and A. Sokolova.
Distributed Queues in Shared Memory: Multicore Performance and Scalability through
Quantitative Relaxation. In CF, 2013.
A. Haas, T. Hütter, C.M. Kirsch, M. Lippautz, M. Preishuber, and A. Sokolova. Scal: A
Benchmarking Suite for Concurrent Data Structures. In NETYS, pages 1–14, 2015.
A. Heddaya and H. Sinha. Coherence, Non-coherence and Local Consistency in Distributed
Shared Memory for Parallel Computing. Technical report, Computer Science Department,
Boston University, 1992.
S. Heller, M. Herlihy, V. Luchangco, M. Moir, W.N. Scherer, and N. Shavit. A Lazy
Concurrent List-based Set Algorithm. In OPODIS, 2005.
J.L. Hennessy and D.A. Patterson. Computer Architecture, Fifth Edition: A Quantitative
Approach. Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 5th edition, 2011.
T.A. Henzinger, C.M. Kirsch, H. Payer, A. Sezgin, and A. Sokolova. Quantitative relaxation
of concurrent data structures. In POPL, pages 317–328, 2013.
T.A. Henzinger, A. Sezgin, and V. Vafeiadis. Aspect-Oriented Linearizability Proofs. In
CONCUR, pages 242–256, 2013.
M. Herlihy and N. Shavit. The Art of Multiprocessor Programming. Morgan Kaufmann
Publishers Inc., San Francisco, CA, USA, 2008.
M. Herlihy and J.M. Wing. Linearizability: A Correctness Condition for Concurrent Objects. ACM Trans. Program. Lang. Syst., 12(3):463–492, 1990.
R. Jagadeesan and J. Riely. Between Linearizability and Quiescent Consistency - Quantitative Quiescent Consistency. In ICALP, pages 220–231, 2014.
C.M. Kirsch, M. Lippautz, and H. Payer. Fast and Scalable, Lock-free k-FIFO Queues. In
PaCT, pages 208–223, 2013.
A. Kogan and E. Petrank. Wait-free queues with multiple enqueuers and dequeuers. In
PPoPP, pages 223–234, 2011.
L. Lamport. How to Make a Multiprocessor Computer That Correctly Executes Multiprocess Programs. IEEE Trans. Comput., 28(9):690–691, September 1979.
R.J. Lipton and J.S. Sandberg. PRAM: A Scalable Shared Memory. Technical Report Nr.
180, Princeton University, Department of Computer Science, 1988.
R.J. Lipton and J.S. Sandberg. Oblivious memory computer networking, September 28
1993. CA Patent 1,322,609.
M.M. Michael. Hazard Pointers: Safe Memory Reclamation for Lock-Free Objects. IEEE
Trans. Parallel Distrib. Syst., 15(6):491–504, 2004.
A. Haas et al.
34
35
36
37
38
39
40
41
42
M.M. Michael and M.L. Scott. Simple, fast, and practical non-blocking and blocking
concurrent queue algorithms. In PODC, pages 267–275, 1996.
M.M. Michael, M.T. Vechev, and V.A. Saraswat. Idempotent Work Stealing. In PPoPP,
pages 45–54, 2009.
A. Morrison and Y. Afek. Fast Concurrent Queues for x86 Processors. In PPoPP, pages
103–112, 2013.
Multicore Computing Group, Tel Aviv University. Fast Concurrent Queues for x86 Processors. Accessed on 01/28/2015. URL: http://mcg.cs.tau.ac.il/projects/lcrq/.
H. Rihani, P. Sanders, and R. Dementiev. MultiQueues: Simpler, Faster, and Better
Relaxed Concurrent Priority Queues. CoRR, 2014. arXiv:1411.1209.
A. Sezgin. Sequential Consistency and Concurrent Data Structures. CoRR, abs/1506.04910,
2015.
N. Shavit. Data Structures in the Multicore Age. CACM, 54(3):76–84, March 2011.
R.C. Steinke and J.G. Nutt. A Unified Theory of Shared Memory Consistency. J. ACM,
51(5):800–849, September 2004.
R.K. Treiber. Systems Programming: Coping with Parallelism. Technical Report RJ-5118,
IBM Research Center, 1986.
XX:15
XX:16
Local Linearizability
A
Local Linearizability with Shape Observers
There are two possible ways to deal with shape observers: treat them locally, in the threadinduced history of the performing thread, or treat them globally. While a local treatment
is immediate and natural to a local consistency condition, a global treatment requires care.
We present both solutions next.
I Definition 17 (Local Linearizability LSO). A history h is locally linearizable with local
shape observers (LSO) wrt a sequential specification S if it is locally linearizable according
to Definition 11 with the difference that the in-methods (Definition 9) also contain all shape
observers performed by thread T , i.e., IT = {m | m ∈ M (h|T ) ∩ (Ins ∪ SOb)}.
Global observations require more notation and auxiliary notions. Let sj for j ∈ J be
a collection of sequences over alphabet Σ with pairwise disjoint sets of symbols M (sj ). A
S
sequence s is an interleaving of sj for j ∈ J if M (s) = j M (sj ) and s|M (sj ) = sj for all
Q
j ∈ J. We write j sj for the set of all interleavings of sj with j ∈ J.
Given a history h and a method call m ∈ h, we write h≤m for the (incomplete) history
that is the prefix of h up to and without mr , the response event of m. Hence, h≤m contains
all invocation and response events of h that appear before mr .
I Definition 18. Let S denote the sequential specification of a container D. A shape
observer m in a history h has a witness if there exists a sequence s ∈ Σ∗ such that sm ∈ S
Q
and s ∈ T sT for some sT that is a linearization of the thread-induced history (h≤m )T .
Informally, the above definition states that a global shape observer m must be justified
by a (global) witness. Such a global witness is a sequence that (1) when extended by m
belongs to the sequential specification, and (2) is an interleaving of linearizations of the
thread-induced histories up to m.
I Definition 19 (Local Linearizability GSO). A history h is locally linearizable with global
shape observers (GSO) wrt a sequential specification S if it is locally linearizable and each
shape observer m ∈ SOb has a witness.
We illustrate the difference in the local vs. the global approach for shape observers with
the following example.
I Example 20. Consider the following queue history with global observer size()
T1
T2
enq(1)
enq(2)
deq(1)
size(n)
where n is just a placeholder for a concrete natural number. For n = 0, the history h is
locally linearizable LSO, but not locally linearizable GSO. For n = 1, the history h is locally
linearizable GSO, but not locally linearizable LSO.
Global observers and non-disjoint operations are expected to have negative impact on
performance. If one cares for global consistency, local linearizability is not the consistency
condition to be used. The restriction to containers and disjoint operations specifies, in an
informal way, the minimal requirements for local consistency to be acceptable.
Neither sets nor maps are containers according to our definition. However, it is possible to
extend our treatment to sets and maps similar to our treatment of global observers. Locally
A. Haas et al.
XX:17
linearizable sets and maps will be weaker than their linearizable counterparts, but, due to
the tight coupling between mutator and observer effects, the gain in performance is unlikely
to be as substantial as the one observed in other data structures. The technicalities needed
to extend local linearizability to sets and maps would complicate the theoretical development
without considerable benefits and we, therefore, excluded such data structures.
B
Additional Results and Proofs
Theorem 12 (Compositionality). A history h over a set of objects O with sequential
specifications Sq for q ∈ O is locally linearizable if and only if h|q is locally linearizable with
respect to Sq for every q ∈ O.
Proof. The property follows from the compositionality of linearizability and the fact that
(h|q)T = hT |q for every thread T and object q. Assume that h over O is locally linearizable. This means that all thread-induced histories hT over O are linearizable. Hence, since
linearizability is compositional, for each object q ∈ O the history hT |q is linearizable with
respect to Sq . Now from (h|q)T = hT |q we have that for every object q the history (h|q)T
is linearizable for every thread T .
Similarly, assume that for every object q ∈ O the history h|q is locally linearizable. Then,
for every q, (h|q)T = hT |q is linearizable for every thread T . From the compositionality of
linearizability, hT is linearizable for every thread T . This proves that h is locally linearizable.
J
Proposition 2 (Lin vs. LocLin 2).
Linearizability implies local linearizability for
sequential specifications that are closed under data-projection.
Proof. Assume we are given a history h which is linearizable with respect to a sequential
specification S that is closed under data-projection. Further assume that, without loss of
generality, h is complete. Then there exists a sequential history s ∈ S such that (1) s is a
permutation of h, and (2) if m <h n, then also m <s n. Given a thread T , consider the
thread-induced history hT and let sT = s| (IT ∪ OT ). Then, sT is a permutation of hT since
hT and sT consist of the same events. Furthermore, sT ∈ S since S is closed under dataprojection and since Equation (1) holds for containers. Finally, we have for each m ∈ hT
and n ∈ hT that, if m <hT n, then also m <sT n since m <h n and therefore m <s n which
implies m <sT n. Thereby, we have shown that hT is linearizable with respect to S, for an
arbitrary thread T . Hence h is locally linearizable with respect to S.
J
Proposition 3 (Data-Projection Closedness).
queue, and stack are closed under data-projection.
The sequential specifications of pool,
Proof. Let s ∈ SP , V 0 ⊆ V , and let
s0 = s| ({ins(x), rem(x) | x ∈ V 0 ∪ Emp}) .
Then, it suffices to check that all axioms for pool (Definition 2 and Table 1) hold for s0 .
Clearly, all methods in s0 appear at most once, as they do so in s. If rem(x) ∈ s0 , then
rem(x) ∈ s and, since s ∈ SP , ins(x) ≺s rem(x). But then also rem(x) ∈ s0 and hence
ins(x) ≺s0 rem(x). Finally, if ins(x) ≺s0 rem(e) for e ∈ Emp, then ins(x) ≺s rem(e) implying
that rem(x) ∈ s and rem(x) ≺s rem(e). But then rem(x) ∈ s0 as well and rem(x) ≺s0 rem(e).
This shows that SP is closed under data-projection.
XX:18
Local Linearizability
Assume now that s ∈ SQ and s0 is as before (with enq() and deq() for ins() and rem(),
respectively). Then, as SP is closed under data-projection, s0 satisfies the pool axioms.
Moreover, the queue-order axiom (Definition 2 and Table 1) also holds: Assume enq(x) ≺s0
enq(y) and deq(y) ∈ s0 . Then enq(x) ≺s enq(y) and deq(y) ∈ s. Since s ∈ SQ we get
deq(x) ∈ s and deq(x) ≺s deq(y). But this means deq(x) ∈ s0 and deq(x) ≺s0 deq(y).
Hence, SQ is closed under data-projection.
Finally, if s ∈ SS and s0 is as before (with push() and pop() for ins() and rem(),
respectively), we need to check that the stack-order axiom (Definition 2 and Table 1) holds.
Assume push(x) ≺s0 push(y) ≺s0 pop(x). This implies push(x) ≺s push(y) ≺s pop(x)
and since s ∈ SS we get pop(y) ∈ s and pop(y) ≺s pop(x). But then pop(y) ∈ s0 and
pop(y) ≺s0 pop(x). So, SS is closed under data-projection.
J
Proposition 4 (LocLin Pool). Let h be a locally linearizable history wrt a pool. Then:
1. No value is duplicated, i.e., every remove method appears in h at most once.
2. There are no out-of-thin-air values, i.e.,
∀x ∈ V. rem(x) ∈ h ⇒ ins(x) ∈ h ∧ rem(x)6<h ins(x).
3. No value is lost, i.e., ∀x ∈ V. ∀e ∈ Emp. ins(x) <h rem(e) ⇒ rem(x) ∈ h ∧ rem(e)6<h rem(x)
and ∀x ∈ V. ∀e ∈ Emp. rem(e) <h rem(x) ⇒ ins(x) 6<h rem(e).
Proof. Note that if a history h is linearizable wrt a pool, then all of the three stated
properties hold, as a consequence of linearizability and the definition of SP . Now assume
that h is locally linearizable wrt a pool.
If rem(x) appears twice in h, then it also appears twice in some thread-induced history
hT contradicting that hT is linearizable with respect to a pool. This shows that no value is
duplicated.
If rem(x) ∈ h, then rem(x) ∈ hT for some T and, since hT is linearizable with respect to
a pool, ins(x) ∈ hT and rem(x)6<hT ins(x). This yields ins(x) ∈ h and rem(x)6<h ins(x).
Hence, there are no thin-air values.
Finally, if rem(e) ∈ h for e ∈ Emp then rem(e) ∈ hT for all T . Let ins(x) <h rem(e) and
let T 0 be such that ins(x) ∈ hT 0 . Then ins(x) <hT 0 rem(e) and since hT 0 is linearizable
with respect to a pool, rem(x) ∈ hT 0 and rem(e)6<hT 0 rem(x). This yields rem(x) ∈ h and
rem(e)6<h rem(x). Similarly, the other condition holds. Hence, no value is lost.
J
Theorem 25 (Queue Local Linearizability). A queue concurrent history h is locally
linearizable with respect to the queue sequential specification SQ if and only if
1. h is locally linearizable with respect to the pool sequential specification SP , and
2. ∀x, y ∈ V. ∀i. enq(x) <ih enq(y) ∧ deq(y) ∈ h ⇒ deq(x) ∈ h ∧ deq(y) 6<h deq(x).
Proof. Assume h is locally linearizable with respect to SQ . Since SQ ⊆ SP (with suitably
renamed method calls), h is locally linearizable with respect to SP . Moreover, since all hi
are linearizable with respect to SQ , by Theorem 24, for all i we have ∀x, y ∈ V. enq(x) <hi
enq(y) ∧ deq(y) ∈ hi ⇒ deq(x) ∈ hi ∧ deq(y) 6<hi deq(x).
Assume x, y ∈ V are such that enq(x) <ih enq(y) and deq(y) ∈ h. Then enq(x) <hi
enq(y) and deq(y) ∈ hi so deq(x) ∈ hi and deq(y) 6<hi deq(x). This implies deq(x) ∈ h
and deq(y) 6<h deq(x).
For the opposite, assume that conditions 1. and 2. hold for a history h. We need to
show that (1) hi form a decomposition of h, which is clear for a queue, and (2) each hi is
linearizable with respect to SQ .
A. Haas et al.
XX:19
By 1., each hi is linearizable with respect to a pool. Assume enq(x) <hi enq(y) and
deq(y) ∈ hi . Then enq(x) <ih enq(y) ∧ deq(y) ∈ h and hence by 2., deq(x) ∈ h ∧
deq(y) 6<h deq(x). Again, as enq(x), deq(x) ∈ hi we get deq(x) ∈ hi ∧ deq(y) 6<hi deq(x).
According to Theorem 24 this is enough to conclude that each hi is linearizable with respect
to SQ .
J
Theorem 15 (Pool, Queue, & Stack, SC). For pool, queue, and stack, local linearizability is incomparable to sequential consistency.
Proof. The following histories, when instantiating i() with ins(), enq(), and push(), respectively, and instantiating r() with rem(), deq(), and pop(), respectively, are sequentially
consistent but not locally linearizable wrt pool, queue and stack:
(a) Pool:
T1
i(1)
r(1)
r(empty)
T2
(b) Queue:
r(2)
T1
T2
i(1)
i(2)
i(1)
i(2)
r(1)
(c) Stack:
T1
r(1)
T2
r(2)
History (a) is already not locally linearizable wrt pool, queue, and stack, respectively,
histories (b) and (c) provide interesting examples. The history in Figure 2 is locally linearizable but not sequentially consistent wrt a pool. The following histories are locally
linearizable but not sequentially consistent wrt a queue and a stack, respectively:
(d) Queue:
T1
T2
i(1)
i(2)
i(3)
r(1)
r(2)
i(4)
r(4)
r(3)
The two thread-induced histories i(1)i(2)i(3)r(1)r(2)r(3) and i(4)r(4) are both linearizable with respect to a queue. However, the overall history has no sequential witness and is
therefore not sequentially consistent: To maintain the queue behavior, the order of operations r(1) and r(2) cannot be changed. However, this implies that the value 3 instead of
the value 4 would have to be removed directly after i(4).
(e) Stack:
XX:20
Local Linearizability
qi = ins(xi,1 )rem(xi,1 ) . . . ins(xi,p )rem(xi,p )rem(xi,p+1 )rem(xi,q )rem(empty)ri ins(xi,q+1 ) . . . ins(xi,m ).
Figure 6 Sequential history qi .
T1
i(1)
i(2)
r(2)
T2
i(3)
r(1)
r(3)
The two thread-induced histories i(1)i(2)r(2)r(1) and i(3)r(3) are both linearizable with
respect to a stack. The operations i(2) and r(2) prevent the reordering of operations
i(1) and i(3). Therefore, the overall history has no sequential witness and hence it is not
sequentially consistent.
J
Proposition 5 (Pool, QC). Let h be a pool history in which no data is duplicated, no
thin-air values are returned, and no data is lost, i.e., h satisfies 1.-3. of Proposition 4.
Then h is quiescently consistent.
Proof. Assume h is a pool history that satisfies 1.-3. of Proposition 4. Let h1 , . . . , hn
be histories that form a sequential decomposition of h. That is h = h1 · · · hn and the
only quiescent states in any hi are at the beginning and at the end of it. Note that this
decomposition has nothing to do with a thread-local decomposition. Let Mi = Mhi be the
set of methods of hi , for i ∈ {1, . . . , n}. Note that the sanity conditions 1.-3. ensure that
none of the following two situations can happen:
rem(x) ∈ Mi , ins(x) ∈ Mj , j > i,
ins(x) ∈ Mi , rem(empty) ∈ Mj , rem(x) ∈ Mk , k > j > i,
Let Vi = {xi,1 , . . . , xi,m } denote the set of values in Mi ordered in a way that there is a p
and q such that
ins(xi,j ), rem(xi,j ) ∈ Mi for j ≤ p;
rem(xi,j ) ∈ Mi for j > p, j ≤ q; and
ins(xi,j ) ∈ Mi for j > q.
Moreover, let ri be the number of occurrences of rem(empty) in hi .
We now construct a sequential history for h, which has the form q = q1 · · · qn where
each sequential history qi is a permutation of Mi shown in Figure 6. Using the observations
above, it is easy to check that q is indeed a quiescent witness for h.
J
Theorem 16 (Pool, Queue, & Stack, QC). For pool, local linearizability is stronger than
quiescent consistency. For queue and stack, local linearizability is incomparable to quiescent
consistency.
Proof. The following histories are quiescently consistent but not locally linearizable wrt
pool, queue, and stack, respectively:
(a) Pool:
rem(empty)
T1
T2
ins(1)
rem(empty)
rem(1)
A. Haas et al.
XX:21
(b) Queue:
enq(1)
T1
T2
enq(2)
enq(3)
deq(3)
deq(2)
(c) Stack:
push(1)
T1
T2
push(2) push(3)
pop(2)
pop(3)
In all three histories, the only quiescent states are before and after the longest operation.
Therefore, all operations in thread T2 can be reordered arbitrarily, in particular in a way
such that they satisfy the sequential specification of the respective concurrent data structure. However, each of the thread-induced histories for thread T2 are not linearizable with
respect to pool, queue, and stack, respectively. Therefore, none of these histories is locally
linearizable. Also here history (a) suffices.
On the other hand, the following histories are not quiescently consistent but locally
linearizable wrt queue, and stack, respectively:
(d) Queue:
T1
enq(1)
deq(2)
enq(2)
T2
deq(1)
(e) Stack:
T1
T2
push(1)
pop(1)
push(2)
pop(2)
In histories (d) and (e), between each two operations, the concurrent data structure
is in a quiescent state. Therefore, none of the operations can be reordered and, hence, no
sequential witness exists. However, all thread-induced histories are linearable and, therefore,
the overall histories are locally linearizable. In particular, on a history where each pair of
operations is separated by a quiescent state, i.e., there is no overlap of operations, a quiescent
consistent data structure behaves as it would be linearizable with respect to its sequential
specification and we see the same semantic differences to local linearizability as we see
between linearizability and local linearizability.
J
C
Case Study: Work Stealing Queues
Consider a data structure D which admits two operation types: ins(x), which inserts the
element x into the container, and rem(), which returns and removes an element from the
container. Now imagine that the implementation uses a Work Stealing Queue (WSQ) [35].
Every thread T that uses D has its unique designated buffer QT in the WSQ. Whenever
thread T calls ins(x), x is appended to the tail of QT . When T calls rem(), WSQ first
XX:22
Local Linearizability
T1
T2
enq(3)
enq(2) enq(1)
deq(1) deq(2) deq(3)
Figure 7 History that is QQC but not LL.
checks whether QT is non-empty; if it is, then it returns the element at the tail of QT (LIFO
semantics) and removes it. Otherwise, it chooses some other QT 0 and tries to return an
element from that buffer. But any time a different thread’s buffer is checked, the element to
be removed is taken from the head (FIFO semantics). If T and T 0 are both trying to access
the same buffer at the same time, then usual synchronization measures are taken to ensure
that exactly one thread removes one element.
Given this implementation, the developer of D wants to write a specification for the
potential users of D. Since D is essentially a collection of deques, the developer is tempted
to state that D is a deque with a particular consistency condition. However, D is not a
linearizable deque because ins(x) by T followed by ins(y) by T 0 followed by rem() returns
either x or y depending on whether T or T 0 calls it; i.e. rem() has ambiguous semantics.
D can be seen as a sequentially consistent (SC) deque but then D does not allow many
behaviors that an SC deque would allow; i.e. SC does not capture the behaviors of D
tightly. Relaxed sequential specifications will not work either since D does converge to
sequential semantics (of a LIFO stack) when a single thread uses it. In short, the developer
will fail to capture the semantics of D in a satisfactory manner.
D on the other hand is a locally linearizable deque in which rem() by T from QT 0 is
treated as FIFO removal whenever T 6= T 0 and as LIFO removal whenever T = T 0 . In other
words, local linearizability provides a succinct and clean representation of a well-known
implementation framework (WSQ) hiding away implementation details. Compare this with
the fact that even though WSQ has a queue in it, to argue its correctness it is proved to be
a linearizable pool even though it has stronger semantics than a pool; i.e. linearizable pool
semantics is too weak for D. Observe also that since what we have described in the example
is essentially providing the illusion of using a monolithic structure which is implemented
in terms of distributed components (shared memory is typically implemented on message
passing), we expect local linearizability to be widely applicable.
D
Quiescent Consistency & Quantitative Quiescent Consistency
Without going into the details of the definition of quantitative quiescent consistency we give
a history in Figure 7 that is quantitatively quiescently consistent but not locally linearizable
wrt a queue. Quantitative quiescent consistency allows to reorder the two insert-operations
in thread T2 and thereby violates local linearizability.
E
Consistency Conditions for Distributed Shared Memory
Decomposition per
LL
thread
Coherence
PRAM
#SHs
Write-Operations Ih (i)
Read-Operations Oh (i)
CCfSH
LoD
n
{ins(v) ∈ h|Ti | v ∈ V }
{head(vinit ) ∈ h}
∪ {head(v) ∈ h | ins(v) ∈ Ih (i)}
Lin.
no
memory location
k
{ins(v) ∈ h | v ∈ V }
{head(v) ∈ h | v ∈ V }
SC
yes
thread
n
{ins(v) ∈ h | v ∈ V }
{head(v) ∈ h|Ti | v ∈ V }
SC
yes
a
yes
PC
thread
n
{ins(v) ∈ h | v ∈ V }
{head(v) ∈ h|Ti | v ∈ V }
SC
CC
thread
n
{ins(v) ∈ h | v ∈ V }
{head(v) ∈ h|Ti | v ∈ V }
SCb
yes
{head(v) ∈ h|Ti | v ∈ V }
c
yes
LC
thread & memory location
n·k
{ins(v) ∈ h|Ti | v ∈ V }
∪ {ins(v) ∈ h | head(v) ∈ h|Ti }
SC
A. Haas et al.
Consistency
Condition
SCa : SC and ins-operations are in the same order for each witness.
SCb : SC and ins-operations are ordered by the transitive closure of the thread program orders and write-read pairs.
SCc : SC and ins-operations from threads other than Ti can be reordered even if they are from the same thread and only logical contradictions
in the local history are considered for consistency.
n: number of threads, k: number of memory locations,
#SHs: number of subhistories, CCfSH: consistency condition for subhistories, LoD: loss of data
Table 2 Comparison of consistency conditions for a single distributed shared memory location, i.e., k = 1
XX:23
XX:24
Local Linearizability
In Table 2 we compare local linearizability (LL) against the consistency conditions coherence [3], pipelined RAM (PRAM) consistency [31, 32, 41, 4], processor consistency (PC) [3,
16], causal consistency (CC) [4], and local consistency (LC) [20]. Local linearizability shares
with all these consistency conditions the idea of decomposing a concurrent history into
several subhistories.
Coherence projects a concurrent history to the operations on a single memory location
and each resulting history has to be sequentially consistent. Since sequential consistency is
not compositional, coherence does not imply sequential consistency for the overall history [3]
whereas local linearizability for each single memory location implies local linearizability for
the overall history.
In contrast to coherence and local consistency, local linearizability, PRAM consistency,
PC, and CC all decompose the history into per-thread subhistories, i.e., if there are n threads
then these conditions consider n subhistories and need n sequential witnesses. Coherence
requires one witness per memory location and local consistency requires one witness per
thread and memory location.
For determining the subhistory for a thread Ti , coherence, PRAM consistency, PC, and
CC consider all write-operations in a given history, i.e., Ih (i) = {ins(v) ∈ h | v ∈ V }. In
contrast, local linearizability only considers the write-operations in thread Ti , i.e., Ih (i) =
{ins(v) ∈ h|Ti | v ∈ V } and local consistency considers all write-operations in thread Ti as
well as all write-operations whose values are read in thread Ti , i.e., Ih (i) = {ins(v) ∈ h|Ti |
v ∈ V } ∪ {ins(v) ∈ h | head(v) ∈ h|Ti }. Regarding read-operations, PRAM consistency,
PC, CC, and LC consider only the read-operations in thread Ti . Coherence considers all
read-operations in a given history and local linearizability only considers read-operations
that read the initial value vinit and read-operations that read values that were written by a
write-operation in thread Ti . Reading the initial value is analogous to returning empty in a
data structure.
Local linearizability requires that each subhistory, i.e., thread-induced history, is linearizable with respect to the sequential specification under consideration. In contrast, coherence,
PRAM consistency, PC, CC, and LC require that each subhistory is sequentially consistent
(or a variant thereof) with respect to the sequential specification. However, the variants
of sequential consistency that are used by these consistency conditions are vulnerable to
a loss of data as discussed in Section 5 and, therefore, make these consistency conditions
unsuitable for concurrent data structures.
When considering PRAM consistency, the sequentialization of the write-operations of different threads might be observed differently by different threads, e.g., a thread T1 might observe all write operations of thread T2 before the write operations of thread T3 but a thread T4
might observe all write operations of T3 before the write operations of T2 . In contrast, threadinduced histories as defined by local linearizability do not involve write-operations from other
threads but involve (some) read-operations performed by other threads. Like PRAM consistency, processor consistency requires for each thread Ti that the read- and write-operations
performed by Ti are seen in Ti ’s program order and that the write-operations performed by
other threads are seen in their respective program order. Furthermore, processor consistency
also requires that two write-operations to the same memory location appear in the same order in each sequential witness of each thread even if they are from different threads [3, 16].
This additional condition makes processor consistency strictly stronger than PRAM consistency [3]. This condition also creates a similar effect as the consideration of read-operations
in different threads when forming the thread-induced history in local linearizability. Causal
consistency considers a causal order instead of the thread program orders alone. Like local
A. Haas et al.
XX:25
` = 4, v = 4
0
`−1
0
backend
` = 5, v = 5
`−1
s:
s:
n:
` = 4, v = 4
max
`−1
0
s:
ni :
ni :
(b) Add new node ni
(c) Adjust `, then v
alive
(a) Initial state
Figure 8 Segment modifications throughout announce_thread().
linearizability, causal consistency matches write-read pairs across different threads. In particular, the causal order is the transitive closure of the thread program orders and write-read
pairs. By considering the causal order, writes from different threads can become ordered
which is not the case for local linearizability.
F
LLD and LL+ D Implementation Details
As already mentioned, each thread inserts elements into a local backend and removes elements either from its local backend (preferred) or from other backends (fall-back) accessed
through a single segment (thread-indexed array), effectively managing single-producer/multipleconsumer backends for a varying number of threads.
The segment is dynamic in length (with a predefined maximum). A slot in this segment
refers to a node that consists of a backend and a flag indicating whether the corresponding thread is alive or has terminated. Similar to other work [2, 21] the flag is used for
logically removing the node from the segment (it stays in the segment until its backend is
empty). Additionally, a (global) version number keeps track of all changes in the segment.
The algorithm is divided into two parts: (1) maintaining the segment, and (2) adding and
removing elements to backends.
In the following we refer to the segment as s, a thread’s Ti local node as ni , the version
number of the segment as v and the current length of the segment as `. The range of indices
r is then defined as 0 ≤ r < `.
For maintaining the segment we provide two methods announce_thread() and
cleanup_thread(node) that are used to add and remove nodes to the segment. Upon
removal of a node the segment is also compacted, i.e., the hole that is created by removing
a node pointer is filled with the last node pointer in the segment. As nodes are added and
removed the length of the segment ` and thus the range of valid indices i of the segment,
0 ≤ i < `, is updated. All changes to the segment involve incrementing the version number.
More detailed, the operations for maintaining the segment and compacting it as nodes
are cleaned up are:
announce_thread(): Allocates a node for the thread as follows: searches for an existing
node of a terminated thread and reuses it if it finds one; otherwise it creates a new node,
adds the node to s, and adjusts `. In both cases it then increments v and returns the
node. The creation of new node is illustrated in Figure 8.
cleanup_thread(Node n): Searches for the node n in s using linear search. If it finds n
at slot j, it copies the pointer of s[` − 1] to s[j], decrements `, increments v, and resets
s[`] to null using the new `. If n is not found, then a concurrent thread has already
performed the cleanup and the operation just returns. Figure 9 illustrates an example
XX:26
Local Linearizability
` = 5, v = 5
0
` = 5, v = 5
`−1
s:
` = 4, v = 6
`−1
0
0
s:
`−1
s:
backend
n:
dead
n:
backend
m:
m:
dead
(a) n at s[0] is empty and
dead
(b) Write s[` − 1] into s[0]
Adjust ` then v
Figure 9 Segment modifications throughout cleanup_thread(Node n).
where initially ` = 5, the thread owning the node at s[0] is dead and the corresponding
backend is empty.
Note that updating the segment state is only needed when threads are joining or when
backends of terminated threads become empty. We consider both scenarios as infrequent and
implement the corresponding operations using locks. Alternatively those operations can be
implemented using helping approaches, similar to wait-free algorithms [29]. Also note that
although operations on segments are protected by locks, partial changes can be observed,
e.g., a remove operation (as defined below) can observe a segment in an intermediate state
with two pointers pointing to a node during cleanup. The invariant is that no change can
destroy the integrity of the segment within the valid range, i.e., all slots within the range
either point to a valid node or nothing (null).
The actual algorithm for adding and removing elements is then defined as follows:
ins(): Upon first insertion, a thread Ti gets assigned a node ni (containing backend bi )
using announce_thread(). The element is then inserted into bi . Subsequent insertions
from this thread will use ni throughout the lifetime of the thread.
rem(): The remove operation consists of two parts: (a) finding and removing an element
and (b) cleaning up nodes of terminated threads. For (a) a thread Ti tries to get an
element from its own backend in ni . If ni does not exist (because the thread has not
yet performed a single ins() operation) or the corresponding backend is empty, then a
different node n is selected randomly within the valid range. If the backend contained
in n is empty, the operation scans all other nodes’ backends in linear fashion. However,
if the version number changed during the round of scanning through all backends, the
operation is restarted immediately. Note that since ` is dynamic a remove operation may
operate on a range that is no longer valid. Checking the version number ensures that the
operation is restarted in such a case. For (b) a thread calls cleanup_thread(n) upon
encountering a node n that has its alive-flag set to false (dead) and contains an empty
backend. A cleanup also triggers a restart of the remove operation.
terminate(): Upon termination a thread Ti changes the alive flag of ni to false (dead).
Dynamic memory used for nodes is susceptible to the ABA problem and requires proper
handling to free memory. Our implementations use 16-bit ABA counters to avoid the ABA
problem and refrain from freeing memory. Hazard pointers [33] can be used for solving the
ABA problem as well as for freeing memory.
F.1
LL+D: LLD with Linearizable Emptiness Check
We call a data structure implementation Φ stateful if the remove methods of Φ can be
modified to return a so-called state that changes upon an insert or a remove of an element,
A. Haas et al.
but does not change between two removes that return empty unless an element has been
inserted in the data structure in the meantime. For stateful implementations Φ we can create
the locally linearizable version with linearizable emptiness check LL+ D Φ. Michael-Scott
queue [34] and Treiber stack [42] are stateful implementations, whereas LCRQ [36] is not.
Also TS stack [14], and k-FIFO [28] and k-Stack [23] are stateful implementations, but the
notion of a state in these data structures is huge making it unsuitable for LL+ D.
For LL+ D implementations, linearizable emptiness checks are achieved via an atomic
snapshot [25], just like for DQs. A detailed description of the LLD and LL+ D implementations, as well as the pseudo code, can be found in the appendix. Here, we only present the
results of the experimental performance evaluation.
F.2
Correctness of LL+ D
I Proposition 7 (LLD and LL+ D). Let Φ be a stateful data structure implementation that is
linearizable with respect to a sequential specification SΦ . Then LL+ D Φ is linearizable with
respect to a pool.
Proof. Proving that LL+ D Φ is linearizable with respect to pool, in particular that it has
a linearizable emptiness check, follows the proof for DQ in general, see [18]: The emptiness
check is performed by creating an atomic snapshot [25] of the states of all backends (stored
in the states array) using the first loop (lines 28-43). If the atomic snapshot is valid
(checked via the second loop, lines 46-52, in particular line 48) and all backends are empty
in this atomic snapshot, then there existed a point in time during the creation of the atomic
snapshot where all backends were indeed empty.
Notice that since the segment is dynamic in length it can happen that some backends
are not contained in the atomic snapshot. To guarantee that no elements are missed in the
emptiness check the atomic snapshot is extended by the version number v of the segment. If
a new backend is added to the segment during the generation of the atomic snapshot, then
the version number is increased and the atomic snapshot becomes invalid (line 45).
The linearization point of the remove operation that returns empty is inbetween the two
loops (the last remove attempt of the first loop) if the version check and second loop go
through.
J
F.3
LLD Pseudo Code
All implementations use the interfaces depicted in Listing 1. For simplicity, the interface
only mentions pool, queue, and stack. The highlighted code refers to linearizable emptiness
check, i.e., it is only part of the LL+ D implementations: Methods retrieving elements (e.g.
rem) are assumed (or modified when possible) to also return a State object that uniquely
identifies the state of the data structure with respect to methods inserting elements (e.g.
ins). The same state can be accessed via the get_state() observer method.
Listing 2 illustrates the pseudo-code for maintaining the segment. The backend on line 2
can either be declared as Stack or Queue as defined in Listing 1 (or any other linearizable
data structure).
Listing 3 shows the pseudo-code for LL+ D. When removing the highlighted code, we
obtain the code for LLD. Each thread maintains its own backend, enclosed in a thread-local
node (line 3), for insertion. The local backend is always accessed through get_local_node
(line 5). This method also makes sure that a thread is announced (line 7) upon first insertion and acquires a node. An ins() operation then always uses a thread’s local backend
XX:27
XX:28
Local Linearizability
1
2
3
4
5
Pool {
< Element , State > rem () ;
v o i d ins ( Element e ) ;
State get_state();
}
6
7
8
9
10
11
12
Queue : Pool {
< Element , State > dequeue () ;
v o i d enqueue ( Element e ) ;
v o i d ins ( Element e ) = > enqueue ( e ) ;
< Element , State > rem () = > dequeue () ;
}
13
14
15
16
17
18
19
Stack : Pool {
< Element , State > pop () ;
v o i d push ( Element e ) ;
v o i d ins ( Element e ) = > push ( e ) ;
< Element , State > rem () = > pop () ;
}
Listing 1 Pool, queue, and stack interfaces
(line 13 and 14) for insertion. For removing an element in rem(), a thread tries to remove an
element from its local backend first (line 19-23). If no element can be found, all backends
in the valid range are searched in a linear fashion, starting from a random index. The
highlighted code (lines 46-52) illustrates checking the atomic snapshot for LL+ D.
F.4
LLD with Observer Methods
We have implemented LLD variants of (strict and relaxed) queue and stack implementations.
None of our LLD implementations involves observer methods, but the LLD algorithm can
easily be extended to support observer methods:
A data observer on LLD Φ (independently of which thread performs it) amounts to a
data observer on any ΦT .
A local shape observer on LLD Φ performed by thread T executes the shape observer
on ΦT .
A global shape observer on LLD Φ executes the shape observer on each backend ΦT and
produces an aggregate value.
G
Additional Implementations
We now present and evaluate additional algorithms that provide locally linearizable variants
of queues and stacks, obtained by modifying relaxed k-out of order queues and stacks [23,
28] in a way that makes them sequentially correct. We have also tried another generic
implementation, related to the construction in [9], that implements a flat-combining wrapper
with sequential (to be precise, single-producer multiple-consumer) backends. In our initial
experiments the performance of such an implementation was not particularly promising.
G.1
Locally Linearizable k-FIFO Queue and k-Stack
k-FIFO queues [28] and k-Stacks [23] are relaxed queues and stacks based on lists of segments
where each segment holds k slots for elements, effectively allowing reorderings of elements of
up to k − 1. The list of segments is implemented by a variant of Michael-Scott queue [34] for
A. Haas et al.
1
2
3
4
Node {
Pool backend ;
Bool alive ;
}
XX:29
// Any linearizable data structure .
5
6
7
8
9
Segment {
Node nodes [ MAX_THREADS ];
Int l = 0;
Int version = 0;
10
// Returns all indexes between 0 and l ( exclusive ) in random order .
[ Int ] range () ;
11
12
13
// Announces a node in the buffer , effectively adding it to nodes_ ,
// adjusting l , and changing the version .
Node a nn ou nc e _t hr ea d () {
segment_lock () ; // Protecting against concurrent announce or cleanup operations .
Node n = find_ dead_nod e () ;
i f n == null {
n = Node ( b : Backend () ) ;
nodes [ l ] = n ;
l ++;
}
n . alive = true ;
version ++;
segme nt_unloc k () ;
r e t u r n n;
}
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
// Removes a node from the buffer , effectively removing it from nodes_ ,
// adjusting l , and changing the version .
v o i d clea nup_thre ad ( Node n , Int old_version ) {
segment_lock () ; // Protecting against concurrent announce or cleanup operations .
<j , error > = f i n d _ n o d e _ i n _ s e g m e n t ( n ) ;
i f error || n . alive || old_version != version {
segm ent_unlo ck () ;
return ;
}
nodes [ j ] = nodes [l -1];
l - -;
version ++;
nodes [ l ] = null ;
segme nt_unloc k () ;
}
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
}
Listing 2 Node and segment structure for LLD and LL+ D (queue or stack)
k-FIFO and a variant of Treiber stack [42] for k-Stack. Insert and remove methods operate
on the segments ignoring any order of elements within the same segment. Segments used
for insertion and removal are identified by insertion and removal pointers, respectively.
For queues, elements are removed from the oldest segment and inserted into the mostrecent not-full segment. Upon trying to remove an element from an empty segment the
segment is removed and the removal pointer advanced to the next segment. Upon trying to
insert an element into a full segment a new segment is appended and the insertion pointer
is advanced to this new segment. Similarly (but different) for a stack, removal and insertion
operate on the most-recent segment, i.e., removal and insertion pointer are synonyms and
identify the same segment at all times. Again, upon trying to remove an element from
an empty segment the segment is removed and the removal pointer advanced to the next
segment. Upon trying to insert an element into a full segment a new segment is prepended
and the insertion pointer is set to this new segment.
k-FIFO queues and k-Stacks are relaxed queues and stacks that are: (1) linearizable with
respect to k-out-of-order queue and stack [23], respectively; (2) linearizable with respect to
a pool [23, 28]; (3) not locally linearizable with respect to queue and stack, respectively, for
k ≥ 1 since reordering elements that are inserted in the same segment (even sequentially by
XX:30
Local Linearizability
1
2
3
DynamicLocallyLinearizableDQ {
Segment s ;
thread_local Node local_node ;
4
Node get_loca l_node ( Bool c r e a t e _ i f_ a b s e n t ) {
i f ( c r e a t e _ i f _ a b s e n t ) && ( local_node == null ) {
local_node = s . an n ou nc e_ t hr ea d () ;
}
r e t u r n local_node ;
}
5
6
7
8
9
10
11
v o i d ins ( Element e ) {
n = get_loca l_node ( cr e a t e _ i f _ a b s e n t : true ) ;
n . backend . ins ( e ) ;
}
12
13
14
15
16
Element rem () {
// Fast path of retrieving an element from the thread - local backend .
n = get_loca l_node ( cr e a t e _ i f _ a b s e n t : false ) ;
i f n != null {
<e , state > = n . backend . rem () ;
i f e != null { r e t u r n e ; }
}
w h i l e true {
retry = false ;
old_version = s . version ;
range = s . range () ;
f o r i in range {
n = s . nodes [ i ];
i f old_version != s . version {
retry = true ; b r e a k ; }
Bool alive = n . alive ;
<e , state > = n . backend . rem () ;
i f e == null {
states[i] = state;
i f ! alive {
s . c leanup_ thread (n , old_version ) ;
retry = true ; b r e a k ;
}
} else {
r e t u r n e;
}
}
i f retry { c o n t i n u e ; }
i f old_version != s . version { c o n t i n u e ; }
f o r i in range {
n = s . nodes [ i ];
i f n == null || n . backend . get_state () != states [ i ] {
retry = true ;
break ;
}
}
i f retry { c o n t i n u e ; }
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
r e t u r n null ; // Empty case .
55
}
56
}
57
58
// Called upon thread termination .
v o i d terminate () {
n = get_loca l_node ( c r e a t e _ i f _ a b s e n t : false ) ;
i f n != null { n . alive = false ; }
}
59
60
61
62
63
64
}
Listing 3 LLD and LL+ D (queue and stack)
a single thread) is allowed, see the histories (b) and (c) in the proof of Theorem 15; and (4)
not sequentially consistent with respect to queue and stack, as shown by the histories (d)
and (e) in the proof of Theorem 15 that are k-FIFO and k-Stack histories, respectively, for
k ≥ 1.
A. Haas et al.
XX:31
We now present LL k-FIFO and LL k-Stack, modifications of k-FIFO and k-Stack, that
enforce local linearizability by ensuring that no thread inserts more than once in a single
segment. Assuming that segments are unique (by tagging pointers), LL k-FIFO remembers
the last used insertion pointer per thread. For LL k-Stack the situation is more subtle
as (due to the stack semantics) segments can be reached multiple times for insertion and
removal. Figure 10 illustrates an example where the top segment of a k-Stack is reached
multiple times by the same thread (T1 ). Since in the general case all segments could be
reached multiple times by a single thread it is required to maintain the full history of each
thread’s insertions. Assuming the maximum number of threads is known in advance, a
bitmap is used to maintain the information in which segment a thread has already pushed a
value. One can similarly implement a locally linearizable version of the Segment Queue [2].
T1
ins(1)
ins(3)
ins(4)
ins(2)
T2
rem(3)
rem(2)
2
⊥
1
1
T1 segment state
insert/remove
⊥
⊥
2
⊥
2
⊥
4
⊥
1
1
3
1
1
⊥
Figure 10 LL k-Stack run (k = 2). T1 can only insert in uncolored segments and needs to
prepend a new segment (for insertion) otherwise.
G.1.1
k-FIFO Queue and LL k-FIFO Queue Pseudo Code.
Listing 4 shows the pseudo code for LL k-FIFO queue. Again we highlight the code we
added to the original pseudo code [28]. Similar to the locally linearizable k-Stack each
thread inserts at most one element into a segment. However, in the k-FIFO queue we do
not need flags in each segment to achieve this property. It is sufficient to remember the last
segment used for insertion for each thread (set_last_tail; line 15). For each enqueue the
algorithm checks whether the executing thread has already used this segment for enqueueing
an element (get_last_tail; line 5). If the segment has already been used, the thread tries
to append a new segment (effectively adding a new tail).
G.1.2
Correctness Proof of LL k-FIFO Queue.
Having Theorem 25, the proof of correctness of LL k-FIFO queue is easy.
I Theorem 21 (Correctness of LL k-FIFO). LL k-FIFO queue presented in Listing 5 is locally
linearizable.
Proof. Using Theorem 25, as a first proof obligation we have to show that any history h of
the LL k-FIFO queue is locally linearizable with respect to the pool sequential specification
SP . This proof is analogous to the proof that any history of the LL k-Stack is locally
linearizable with respect to the pool sequential specification SP , and is therefore postponed
until the corresponding LL k-Stack theorem.
What remains to show is that
∀x, y ∈ V. ∀i. enq(x) <ih enq(y) ∧ deq(y) ∈ h ⇒ deq(x) ∈ h ∧ deq(y) 6<h deq(x)
XX:32
Local Linearizability
1
2
3
4
5
6
7
LocallyLinearizableKFIFOQueue {
enqueue ( item ) :
w h i l e true :
tail_old = get_tail () ;
i f get_last_tail ( thread_id ) == tail_old :
advance_tail ( tail_old , k ) ;
c o n t i n u e ; // Restart while loop .
head_old = get_head () ;
item_old , index = f in d _e mp ty _ sl ot ( tail_old , k ) ;
i f tail_old == get_tail () :
i f item_old . value == EMPTY :
item_new = atomic_value ( item , item_old . version + 1) ;
i f CAS (& tail_old - > segment [ index ] , item_old , item_new ) :
i f committed ( tail_old , item_new , index ) :
8
9
10
11
12
13
14
set_last_tail ( thread_id , tail_old ) ;
15
r e t u r n true ;
else:
advance_tail ( tail_old , k ) ;
16
17
18
19
bool committed ( tail_old , item_new , index ) :
i f tail_old - > segment [ index ] != item_new :
r e t u r n true ;
head_current = get_head () ;
tail_current = get_tail () ;
item_empty = atomic_value ( EMPTY , item_new . version + 1) ;
i f i n _ q u e u e _ a f t e r _ h e a d ( tail_old , tail_current , head_current ) :
r e t u r n true ;
e l s e i f not_in_queue ( tail_old , tail_current , head_current ) :
i f ! CAS (& tail_old - > segment [ index ] , item_new , item_empty ) :
r e t u r n true ;
e l s e : // in queue at head
head_new = atomic_value ( head_current . value , head_current . version + 1) ;
i f CAS (& head , head_current , head_new ) :
r e t u r n true ;
i f ! CAS (& tail_old - > segment [ index ] , item_new , item_empty ) :
r e t u r n true ;
r e t u r n false ;
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
item dequeue () :
w h i l e true :
head_old = get_head () ;
item_old , index = find_item ( head_old , k ) ;
tail_old = get_tail () ;
i f head_old == get_head () :
i f item_old . value != EMPTY :
i f head_old . value == tail_old . value :
advance_tail ( tail_old , k ) ;
item_empty = atomic_value ( EMPTY , item_old . version + 1) ;
i f CAS (& head_old [ index ] , item_old , item_empty ) :
r e t u r n item_old . value ;
else:
i f head_old . value == tail_old . value && tail_old . value == get_tail () :
r e t u r n null ;
advance_head ( head_old , k ) ;
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
}
Listing 4 Locally Linearizable k-FIFO Queue
Assume enq(x) <ih enq(y). This means that x and y were enqueued by the same thread
i and therefore inserted into different segments. Moreover, the segment of x is closer to
the head of the list than the segment of y. A deq(y) method call can remove y only if the
segment of y is the head segment. The segment of y can only become the head segment if
all segments closer to the head of the list get empty. This means that also the segment of x
has to become empty. Therefore there has to exist a deq(x) method call which removes x
from the segment, and deq(x) 6<h deq(y).
J
A. Haas et al.
G.1.3
k-Stack and LL k-Stack Pseudo Code.
Listing 5 shows the pseudo code for LL k-Stack. The highlighted code is the code we
added to the original pseudo code [23] to achieve local linearizability. The difference to
the original algorithm is that a thread inserts at most one element into a segment. To
achieve this property each segment in the k-stack contains a flag per thread which is set
when an element is inserted into the segment (mark_segment_as_used; line 14 and line 60).
If a thread encounters a segment where its flag is already set, the thread does not insert
its element into that segment but tries to prepend a new segment (is_segment_marked;
line 50). Otherwise the element is inserted into the existing segment and the flag of the
thread in that segment is set.
1
2
LocallyLinearizableKStack {
SegmentPtr top ;
3
4
5
6
v o i d init () :
new_ksegment = calloc ( s i z e o f ( ksegment ) ) ;
top = atomic_value ( new_ksegment , 0) ;
7
8
9
10
11
12
13
14
15
16
17
bool t r y _ a d d _ n e w _ k s e g m e n t ( top_old , item ) :
i f top_old == top :
new_ksegment = calloc ( s i z e o f ( ksegment ) ) ;
new_ksegment - > next = top_old ;
new_ksegment - > s [0] = atomic_value ( item , 0) ; // Use first slot for item .
top_new = atomic_value ( new_ksegment , top_old . ver +1) ;
m a r k _ s e g m e n t _ a s _ u s e d ( top_new ) ;
i f CAS (& top , top_old , top_new ) :
r e t u r n true ;
r e t u r n false ;
18
19
20
21
22
23
24
25
26
27
v o i d t r y _ r e m o v e _ k s e g m e n t ( top_old ) :
i f top_old == top :
i f top_old - > next != null :
a t o m i c _ i n c r e me n t (& top_old - > remove ) ;
i f empty ( top_old ) :
top_new = atomic_value ( top_old - > next , top_old . ver +1) ;
i f CAS (& top , top_old , top_new ) :
return ;
a t o m i c _ d e c r e me n t (& top_old - > remove ) ;
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
bool committed ( top_old , item_new , index ) :
i f top_old - > s [ index ] != item_new :
r e t u r n true ;
e l s e i f top_old - > remove == 0:
r e t u r n true ;
e l s e : // top_old - > remove >= 1
item_empty = atomic_value ( EMPTY , item_new . ver +1) ;
i f top_old != top :
i f ! CAS (& top_old - > s [ index ] , item_new , item_empty ) :
r e t u r n true ;
else:
top_new = atomic_value ( top_old . val , top_old . ver +1) ;
i f CAS (& top , top_old , top_new ) :
r e t u r n true ;
i f ! CAS (& top_old - > s [ index ] , item_new , item_empty ) :
r e t u r n true ;
r e t u r n false ;
46
47
48
49
50
51
52
53
54
55
56
57
58
59
v o i d push ( item ) :
w h i l e true :
top_old = top ;
i f s e g m e n t _ i s _ m a r k e d ( top_old ) :
i f t r y _ a d d _ n e w _ k s e g m e n t ( top_old , item ) ;
r e t u r n true ;
c o n t i n u e ; // Restart while loop .
item_old , index = f i nd _e mp t y_ sl ot ( top_old ) ;
i f top_old == top :
i f item_old . val == EMPTY :
item_new = atomic_value ( item , item_old . ver +1) ;
i f CAS (& top_old - > s [ index ] , item_old , item_new ) :
i f committed ( top_old , item_new , index ) :
XX:33
XX:34
Local Linearizability
m a r k _ s e g m e n t _ a s _ u s e d ( old_top ) ;
60
r e t u r n true ;
61
else:
i f t r y _ a d d _ n e w _ k s e g m e n t ( top_old , item ) :
r e t u r n true ;
62
63
64
65
item pop () :
w h i l e true :
top_old = top ;
item_old , index = find_item ( top_old ) ;
i f top_old == top :
i f item_old . val != EMPTY :
item_empty = atomic_value ( EMPTY , item_old . ver +1) ;
i f CAS (& top_old - > s [ index ] , item_old , item_empty ) :
r e t u r n item_old . val ;
else:
i f only_ksegment ( top_old ) :
i f empty ( top_old ) :
i f top_old == top :
r e t u r n null ;
else:
t r y _ r e m o v e _ k s e g m e n t ( top_old ) ;
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
}
Listing 5 Locally Linearizable k-Stack
G.1.4
Correctness Proof of LL k-Stack.
The local linearizability proof of LL k-Stack is more involved, but very interesting. We use
a theorem from the published artifact of [14], which has been mechanically proved in the
Isabelle HOL theorem prover.
I Theorem 22 (Empty Returns for Stack). Let h be a history, and let h0 be the projection of
h to Σ \ pop(empty). If h is linearizable with respect to the sequential specification SP of a
pool (see Definition 2), and h0 is linearizable with respect to the sequential specification SS
of a stack (see Definition 2), then h is linearizable with respect to SS .
Proof. Here we repeat the key insights of the proof and leave out technical details. A
complete and mechanized version of the proof is available in the published artifact of [14].
As h is linearizable with respect to SP , and h0 is linearizable with respect to SS , there
exists a sequential history s ∈ SP such that s is a linearization of h, and there exists
a sequential history s0 ∈ SS such that s0 is a linearization of h0 . We show that we can
construct a sequential history t ∈ SS such that t is a linearization of h.
The linearization t is constructed as follows: the position of pop(empty) in s is preserved in t. This means for any method call m ∈ s that if pop(empty) ≺s m, then also
pop(empty) ≺t m, and if m ≺s pop(empty), then also m ≺t pop(empty). Moreover, if two
method calls m, n ∈ s are ordered as m ≺s pop(empty) ≺s n and therefore by transitivity it
holds that m ≺s n, then also m ≺t n.
For all other method calls the order of s0 is preserved. This means for any two method
calls m, n ∈ s with m ≺s0 n, that if for all pop(empty) it holds that pop(empty) ≺s m if and
only if pop(empty) ≺s n, then m ≺t n.
By construction, the history t is sequential and a permutation of h. Next we show that
t is a linearization of h by showing that t preserves the precedence order of h. Also by
construction, it holds that if m ≺t n for any two method calls m, n ∈ t, then also either
m ≺s n or m ≺s0 n. Both s and s0 are linearizations of h and h0 , respectively. Therefore
it cannot be for any m, n with m <h n that n ≺s m or n ≺s0 m, it can also not be that
n ≺t m. Since t is sequential, this means that t preserves the precedence order of h.
Next we show that t ∈ SP according to Definition 2:
A. Haas et al.
(1) Every method call, but pop(empty), appears in s at most once: This is guaranteed since
t is a permutation of s, and s ∈ SP .
(2) If pop(x) appears in t, then also push(x) does and push(x) ≺t pop(x): again, since
t is a permutation of s and s ∈ SP , if pop(x) ∈ t, then also push(x) ∈ t. Since
push(x) ≺s pop(x) and push(x) ≺s0 pop(x) (because both s and s0 are in SP ) it also
holds that push(x) ≺t pop(x), as we argued already above.
(3) ∀x ∈ V. push(x) ≺t pop(empty) ⇒ pop(x) ≺t pop(empty): this property is satisfied
trivially as all pop(empty) operations are ordered the same in t as in s, and s ∈ SP .
It only remains to check that all elements are removed in a stack fashion. We have to
show the following:
∀x, y ∈ V. push(x) ≺t push(y) ≺t pop(x) ⇒ pop(y) ∈ t ∧ pop(y) ≺t pop(x)
First we show that if push(x) ≺t push(y) ≺t pop(x), then also push(x) ≺s0 push(y) ≺s0
pop(x). We do this by showing that there cannot exist a pop(empty) such that push(x) ≺t
pop(empty) ≺t push(y) or push(y) ≺t pop(empty) ≺t pop(x).
Assume, towards a contradiction, push(x) ≺t pop(empty) ≺t push(y). By the transitivity of ≺t this implies that push(x) ≺t pop(empty) ≺t pop(x), which contradicts our
observation above that t ∈ SP . Therefore push(x) ≺t pop(empty) ≺t push(y) is not possible, and for the same reason also push(y) ≺t pop(empty) ≺t pop(x) is not possible.
Now, as s0 ∈ SS and push(x) ≺s0 push(y) ≺s0 pop(x), there has to exist a pop(y) ∈ s0 with
pop(y) ≺s0 pop(x). For the same reason as above it cannot be that pop(x) ≺s pop(empty) ≺s
pop(y). Therefore pop(y) and pop(x) are ordered in t the same as in s0 , i.e. pop(y) ≺t pop(x),
and therefore t ∈ SS .
J
I Theorem 23 (Correctness of LL k-Stack). The LL k-Stack algorithm presented in Listing 5
is locally linearizable.
Proof. We have to show that every history h of LL k-Stack is locally linearizable with
respect to the sequential specification SS defined in Definition 2. This means that we have
to show that every thread-induced history hi of h is linearizable with respect to SS for any
thread i.
Having Theorem 22 we only have to show that hi is linearizable with respect to the
sequential specification SP of a pool (defined in Definition 2), and that h0 i , the projection
of hi to Σ \ pop(empty), is linearizable with respect to the sequential specification SS of a
stack.
We start with the proof that hi is linearizable with respect to SP . We construct a
sequential history si from hi by identifying the linearization points of the push and pop
method calls of the LL k-Stack. This means that two method calls m, n are ordered in si ,
m ≺si n if the linearization point of m is executed before the linearization point of n in hi .
The linearization point of push method calls is either the successful insertion of a new
segment in line 15, or the last successful CAS which writes the element into a segment slot
in line 58. The linearization point of pop method calls is the successful CAS which removes
an element from its segment slot in line 73.
For the linearization point of pop(empty) we take the linearization point of the call to
empty in line 77. The empty method creates an atomic snapshot [25] of the top segment.
This atomic snapshot is the state of the top segment at some point (i.e. linearization point
of empty) within the execution of empty. If empty returns true, then there exists no element
in the atomic snapshot of the segment.
XX:35
XX:36
Local Linearizability
Next we show that si is in SP as defined in Definition 2.
(1) Since there exists exactly one linearization point per method call, every method call,
but rem(empty), appears in si at most once.
(2) If pop(x) appears in si , then it reads x in a slot of the top segment before its linearization
point. Since only push method calls write their elements into segment slots, there has to
exist a push(x) which wrote x into that slot. Therefore the linearization point of push(x)
is always before the linearization point of pop(x), and therefore push(x) ≺si pop(x).
(3) Segments are only removed from the list of segments when they become empty. The call
to committed guarantees that elements are not inserted into segments which are about
to be removed.
A pop method calls empty only if there is a single segment left in the LL k-Stack and
no element was found in that segment in find_item.
Now assume a push(x) method call inserts an element x which is missed by find_item.
If push(x) wrote x into a segment before the linearization point of pop(empty) and
the segment was not the last segment, then the top segment changed since pop(empty)
searched for an element and therefore the check in line 70 would fail. If push(x) wrote x
into the last segment of the LL k-Stack, then a pop(x) method call removed x from the
segment because otherwise x would be in the atomic snapshot of empty and therefore
empty would return false. Therefore, if push(x) ≺si pop(empty), then also pop(x) ≺si
pop(empty).
Therefore si is in the sequential specification SP of a pool.
Next we show that h0 i is linearizable with respect to SS . We construct again a sequential
history s0 i from h0 i by identifying the linearization points of the push and pop method calls
of LL k-Stack.
The linearization point of the push operations is the successful insertion of a new segment
in line 15 if it is executed, or the reading of the empty slot (line 54) in the last (and therefore
successful) iteration of the main loop. The linearization point of a pop operation is the
reading of a non-empty slot (line 69) in the last (and therefore successful) iteration of the
main loop. There do not exist any pop(empty) method calls in s0 i . Since we assume a
sequentially consistent memory model, these read operations define a total order on the LL
k-Stack method calls in h0 i .
First we show that s0 i is in the sequential specification SP of a pool as defined in Definition 2.
(1) Since there exists exactly one linearization point per method call, every method call
appears in s0 i at most once.
(2) If pop(x) appears in s0 i , then it read x in a slot of the top segment at its linearization
point. Since only push operations write their elements into segment slots, there has
to exist a push(x) which wrote x into that slot. The linearization point of push(x) is
always before x is written into a segment slot. Therefore push(x) ≺s0 i pop(x).
(3) Since there exist no pop(empty) operations in s0 i the third pool condition is trivially
correct.
Next we show that s0 i also provides a stack order, which means that we have to show
that
∀x, y ∈ V. push(x) ≺s0 i push(y) ≺s0 i pop(x) ⇒ pop(y) ∈ s0 i ∧ pop(y) ≺s0 i pop(x).
We start by observing some invariants.
A. Haas et al.
1. A thread never inserts elements into the same segment twice. This is guaranteed by the
call to segment_is_marked.
2. Between the linearization point of a push and the time it writes its element into a segment
the segment the element gets written into is not removed: if the push operation inserts
a new segment this is trivially correct. If the push operation writes the element into an
existing segment, then the call to committed in line 59 guarantees that the segment was
not removed.
3. At the time of the linearization point of the pop, which is the time when the pop reads
the non-empty slot (line 69) in the last (and therefore successful) iteration, the pop reads
the non-empty slot from the top segment. This is guaranteed by the check in line 70.
Now assume there exist the operations push(x), push(y) and pop(x) in s0 i and
push(x) ≺s0 i push(y) ≺s0 i pop(x). Since push(x) and push(y) are both in s0 i , this means
that both operations are executed by the same thread. Therefore, according to Invariant 1.,
x and y get inserted into different segments, with the segment y on top of the segment of x.
The linearization point of pop(x) cannot be before y is written into its segment because
according to Invariant 2. the segment y gets inserted into does not get removed between the
linearization point of push(y) and the time y is written into the segment. With Invariant
3. this means that x is unaccessible for pop(x) before y gets written into a segment. Also
because of the third invariant the top segment changes between the insertion of y and the
linearization point of pop(x).
Next we observe that as long as y is not removed, no segment below the segment of y can
become the top segment. Therefore for the segment of x to become the top segment so that
pop(x) can remove it, y has to be removed first. Only a pop(y) can remove y, and therefore
there exists a pop(y) and the linearization point of pop(y) is before the linearization point
of pop(x).
Hence s0i is in the sequential specification of a stack. Using Theorem 22 this means that
LL k-Stack in listing 5 is locally-linearizable with respect to the sequential specification of
a stack.
J
H
Additional Experiments
We also evaluate the implementations on another Scal workload, the sequential alternating
workload. However, we note that in this workload in the locally linearizable implementations
threads only access their local backends, so no wonder they perform perfectly well.
Mixed Workload. In order to evaluate the performance and scalability of mixed workloads, i.e., workloads where threads produce and consume values, we exercise the so-called
sequential alternating workload in Scal. Each thread is configured to execute 106 pairs of
insert and remove operations, i.e., each insert operation is followed by a remove operation.
As in the producer-consumer workload, the contention is controlled by adding a busy wait
of 5µs. The number of threads is configured to range between 1 and 80. Again we report
the number of data structure operations per second.
Data structures that require parameters to be set are configured like in the producerconsumer benchmark. Figure 11 shows the results of the mixed workload benchmark for all
considered data structures.
The MS queue and Treiber stack do not perform and scale for more than 10 threads.
As in the producer-consumer benchmark, LCRQ and TS Stack either perform competitively
XX:37
32
28
24
20
16
12
8
4
0
1
10
20
30
40
50
60
70
80
million operations per sec (more is better)
Local Linearizability
million operations per sec (more is better)
XX:38
32
28
24
20
16
12
8
4
0
1
10
20
30
40
50
60
70
80
number of threads
number of threads
MS
LCRQ
k-FIFO
LL+D MS
LLD LCRQ
LLD k-FIFO
1-RA DQ
Queues, LL queues, and “queue-like” pools
Treiber
TS Stack
k-Stack
LL+D Treiber
LLD TS Stack
LLD k-Stack
1-RA DS
Stacks, LL stacks, and “stack-like” pools
Figure 11 Performance and scalability of sequential alternating microbenchmarks with an increasing number of threads on a 40-core (2 hyperthreads per core) machine
with their k-out-of-order relaxed counter parts k-FIFO and k-Stack or even outperform and
outscale them (in the case of LCRQ, that even outperforms the pool).
LL+ D MS queue, LLD LCRQ, and LL+ D Treiber stack perform very well and scale
(nearly) linearly in the number of threads. A surprising result is that LLD k-FIFO performs
poorly in this experiment. The reason is that k-FIFO performs poorly when it is almost
empty, and in this experiment each backend instance of LLD k-FIFO contains at most one
element at any point in time. The k-Stack performs better on a nearly-empty state. The
benefit of trying to perform a local operation first in the LLD algorithms is visible when
comparing to 1-RA DQ and DS that do not utilize a local fast path.
I
Verifying Local Linearizability
In general, verifying local linearizability amounts to verifying linearizability for a set of
smaller histories. This might enable verification in a modular/compositional way. Aside
from this, it is important to mention (again) that for our locally linearizable data structures
in Section 6 built from linearizable building blocks, the correctness proofs are straightforward
assuming the building blocks are proven to be linearizable. In addition, for queue we can
state an “axiomatic" verification theorem for local linearizability in the style of [24, 10],
whose main theorem we recall next (with a slight reformulation).
I Theorem 24 (Queue Linearizability). A queue concurrent history h is linearizable wrt the
queue sequential specification SQ if and only if
1. h is linearizable wrt the pool sequential specification SP (with suitable renaming of method
calls), and
2. ∀x, y ∈ V. enq(x) <h enq(y) ∧ deq(y) ∈ h ⇒ deq(x) ∈ h ∧ deq(y) 6<h deq(x). J
We note that an analogous change to the axioms in the sequential specification of a pool
and a stack does not lead to a characterisation of linearizability for pools and stacks, cf. [14].
An axiomatic characterisation of linearizability for pools and stacks would involve an infinite
number of axioms/infinite axioms, due to the need to prohibit infinitely many problematic
shapes, cf. [7].
We are now able to state the queue-local-linearizability-verification result.
A. Haas et al.
I Theorem 25 (Queue Local Linearizability). A queue concurrent history h is locally linearizable wrt the queue sequential specification SQ if and only if
1. h is locally linearizable wrt the pool sequential specification SP (after suitable renaming
of method calls), and
2. ∀x, y ∈ V. ∀T. enq(x) <Th enq(y) ∧ deq(y) ∈ h ⇒ deq(x) ∈ h ∧ deq(y) 6<h
deq(x).
J
XX:39
| 8cs.DS
|
Inference for Continuous Time Random
Maxima with Heavy-Tailed Waiting Times
arXiv:1802.05218v2 [math.ST] 17 Feb 2018
Katharina Hees ∗
University of Dortmund
and
Smarak Nayak †
National Australia Bank
and
Peter Straka ‡
UNSW Sydney
February 20, 2018
Abstract
In many complex systems of interest, inter-arrival times between events such as
earthquakes, trades and neuron voltages have a heavy-tailed distribution. The set of
event times is fractal-like, being dense in some time windows and empty in others,
a phenomenon dubbed “bursty” in the physics literature. Renewal processes with
heavy-tailed waiting times reproduce this bursty behaviour.
This article develops an inference method for “Continuous Time Random Maxima” (also called “Max-renewal processes”), which assume i.i.d. magnitudes at the
renewal events and model the largest cumulative magnitude. For high thresholds
and infinite-mean waiting times, we show that the times between threshold crossings
are Mittag-Leffler distributed, i.e. form a fractional Poisson Process. Exceedances of
thresholds are known to be Generalized Pareto distributed, according to the Peaks
Over Threshold approach. We model threshold crossing times and threshold exceedances jointly and provide graphical means of estimating model parameters. We
show that these methods yield meaningful insights on real-world datasets.
Keywords: heavy tails, renewal process, extreme value theory, peaks over threshold
∗
Katharina Hees ([email protected]) was supported by the DAAD co-financed by
the German Federal Ministry of Education and Research (BMBF). The authors would like to thank Peter
Scheffler for insights on stochastic process limits for CTRMs, and Gurtek Gill who helped create the
MittagLeffleR R-package.
†
([email protected])
‡
Corresponding author ([email protected]). Peter Straka was supported by the Discovery Early
Career Research Award DE160101147 on the Project “Predicting Extremes when Events Occur in Bursts”
by the Australian Research Council.
1
1
Introduction
Time series displaying temporally inhomogeneous behaviour have received strong interest in
the recent statistical physics literature (Barabási 2005, Oliveira & Barabási 2005, Vasquez
et al. 2006, Vazquez et al. 2007, Omi & Shinomoto 2011, Min et al. 2011, Karsai et al.
2011, Bagrow & Brockmann 2013), and have been observed in the context of earthquakes,
sunspots, neuronal activity and human communication (see Karsai et al. 2012, Vajna et al.
2013, and Meerschaert & Stoev (2008) for a list of references). Such time series exhibit high
activity in some ‘bursty’ intervals, which alternate with other, quiet intervals. Although
several mechanisms are plausible explanations for bursty behaviour (most prominently selfexciting point processes (Hawkes 1971)), there seems to be one salient feature which very
typically indicates the departure from temporal homogeneity: a heavy-tailed distribution
of waiting times (Vasquez et al. 2006, Karsai et al. 2012, Vajna et al. 2013). As we show
below in simulations, a simple renewal process with heavy-tailed waiting times can capture
this type of dynamics. For many systems, the renewal property is appropriate; a simple
test of the absence of correlations in a succession of waiting times can be undertaken by
randomly reshuffling the waiting times (Karsai et al. 2012).
Often a magnitude, or mark can be assigned to each event in the renewal process,
such as for earthquakes, sun flares or neuron voltages. Extreme value theory provides
models for the distribution of the events with the largest magnitude. A commonly made
assumption is that times between events are either fixed or light-tailed, and this entails
that the times at which a (high) threshold is crossed form a Poisson process (Beirlant
et al. 2006). In the heavy-tailed waiting time scenario, as we will see, threshold crossing
times form a fractional Poisson process, which generalizes the standard Poisson process.
Maxima of events which occur according to a renewal process with heavy-tailed waiting
times have been studied under the names “Continuous Time Random Maxima process”
(CTRM) (Benson et al. 2007, Meerschaert & Stoev 2008, Hees & Scheffler 2016, 2017),
“Max-Renewal process” (Silvestrov 2002, Silvestrov & Teugels 2004, Basrak & Špoljarić
2014), and “Shock process” (Esary & Marshall 1973, Shanthikumar & Sumita 1983, 1984,
1985, Anderson 1987, Gut 1999). The existing literature focuses on probabilistic results
surrounding these models. In this work, however, we introduce a method of inference for
2
x
J7
J4
J5
J2
J3
J8
J1
W1
J9
J6
W2
W3
W4
W5
W6 W7
W8
W9
t
Figure 1: A CTRM sample path.
this type of model, a problem which has seemingly received little attention so far, even by
the statistical community.
We review the CTRM, or max-renewal model, in Section 2, and derive a scaling limit
theorem for inter-exceedance times in Sections 3 & 4. A statistical procedure to estimate model parameters via stability plots is outlined in Section 5, and diagnostic plots
for model criticism are discussed in Section 6. In Section 7, we give case studies for simulated data as well as for two real-world datasets, with a discussion in Section 8. All
statistical computations have been done in R, and all code, data, as well as the source
code for this manuscript has been organized into an R software package, available at
https://github.com/strakaps/CTRM.
2
CTRMs
Continuous Time Random Maxima (CTRM), or Max-Renewal processes, track the largest
event in a renewal sequence of events (Basrak & Špoljarić 2014, Meerschaert & Stoev 2008,
Hees & Scheffler 2016):
Definition (CTRM): Let (W, J), (W1 , J1 ), (W2 , J2 ), . . . be i.i.d. pairs of random variables, where the Wk > 0 are interpreted as the waiting times or inter-arrival times and
3
Jk ∈ [xL , xR ] as the event magnitudes or marks (xL ∈ [−∞, +∞), xR ∈ (−∞, +∞]).
The renewal process associated with the Wk is
N (t) = max{n : W1 + . . . + Wn ≤ t},
and the running maximum is M (n) :=
Wn
k=1
Jk . Then the process
N (t)
V (t) = M (N (t)) =
_
Jk ,
t ≥ 0.
k=1
is called a CTRM (Continuous Time Random Maxima) process. If W and J are
independent, the CTRM is said to be uncoupled. We let M (0) = xL .
Figure 1 shows a CTRM sample path. The alternative case where one considers the maximum of the first (N (t) + 1)’s event is called an OCTRM process (Overshooting Continuous
Time Random Maxima; or Model II by Shanthikumar & Sumita (1983)). The conceptual
difference between CTRM and OCTRM is the order of waiting times and events, which is
W1 , J1 , W2 , J2 , .. for the CTRM and J1 , W1 , J2 , W2 , ... for the OCTRM. In the uncoupled
case, unsurprisingly CTRM and OCTRM are asymptotically equivalent. In the coupled
case, however, their distributions can be dramatically different (Hees & Scheffler 2016).
The main purpose of this article is to introduce and discuss a method for fitting data
to an uncoupled CTRM. The following two quantities are our main ingredients.
Definition: The random variables
X(`) = V (T (`)) − `,
T (`) = inf{t : V (t) > `},
` ∈ (xL , xR )
are called the exceedance resp. time until exceedance of level `.
POT-like approach. Since we assume the uncoupled case, X(`) and T (`) are independent. To see this, note that X(`) is, in distribution, simply equal to J|J > `, independent of
any waiting times Wk . For our inference on the CTRM model, we will extract realizations
of X(`) and T (`) from data by considering thresholds ` ranging over some interval [`0 , xR ).
(For instance, let `0 equal the 95-th percentile of J.) Write X(`, i) for the i-th exceedance
of level `, and T (`, i) for the inter-arrival time between X(`, i−1) and X(`, i). Thus for any
4
8
0
2
4
6
magnitudes
0
20000
40000
60000
80000
times
Figure 2: Exceedances (red) and Times until Exceedance (durations between blue crosses)
for a given threshold ` (dashed line).
threshold ` ∈ [`0 , xR ), we construct sequences {T (`, i)}i and {X(`, i)}i . These sequences
are i.i.d. since the underlying Wk and Jk are i.i.d., and independent because Wk and Jk are
uncoupled. Extracting the exceedances X(`) for inference is a standard procedure known as
the Peaks Over Threshold (POT) approach in extreme value theory (Beirlant et al. 2006).
Figure 2 illustrates the exceedances X(`, i) in red and the times until exceedance T (`, i) as
the preceding intervals marked by blue crosses.
A standard result from extreme value theory provides the asymptotic distribution of
X(`) as ` ↑ xR . For later use, we also include the scaling limit of the cumulative maximum
M (n).
Theorem GEV (e.g. Chapter 2 & Section 5.3, Beirlant et al. (2006)): Suppose the
distribution of Jk is continuous. Then there exist non-decreasing norming functions
a(c) and d(c) such that as c → ∞, the running maximum1 M (c) converges weakly to
1
We extend M (n) continuously to [0, +∞) via M (c) := M (bcc) where bcc is the greatest integer not
greater than c.
5
a Generalized Extreme Value Distribution with shape parameter ξ:
d
[M (c) − d(c)]/a(c) → A,
P(A ≤ z) = G(z|ξ) = exp −[1 + ξz]−1/ξ ,
(1)
defined for all z such that 1 + ξz > 0. This definition extends continuously to ξ = 0
via
G(z|0) = exp(− exp(−x)).
For µ and σ > 0, we write GEV(ξ, µ, σ) for the probability distribution of the random
variable σA + µ.
Moreover, asymptotically as ` ↑ xR ,
P(J − ` > y|J > `) ∼ 1 − H(y),
where
1 − (1 + ξy/σ̃)−1/ξ , y ∈ (0, ∞) if ξ > 0,
H(y) = 1 − exp(−y/σ̃),
y ∈ (0, ∞) if ξ = 0,
1 − (1 + ξy/σ̃)−1/ξ , y ∈ (0, −σ/ξ) if ξ < 0
and where σ̃ := σ + ξ(` − µ). The CDF H(y) is said to belong to the Generalised
Pareto family GP (ξ, σ̄).
3
3.1
CTRM Scaling limits
Waiting times
We will assume that the waiting times Wk have infinite mean, i.e. a tail parameter β ∈ (0, 1)
and regularly varying tail probabilities: P(W > t) ∼ L(t)t−β as t ↑ ∞ for some slowly
varying function L(t). As usually, f (t) ∼ g(t) means that the quotient of f (t) and g(t)
converges to 1. The law of the W then lies in the domain of attraction of a stable law, in
the sense that the distributional limit
W1 + . . . + Wn d
−→ D,
b(n)
6
n→∞
(2)
exists, for a regularly varying scaling function
b ∈ RV∞ (1/β)
(3)
(See e.g. Section 2.9.2 in Beirlant et al. 2006, for a short introduction to regular variation).
The limit is then a positively skewed stable distribution, whose scale parameter is 1 if b(n) is
chosen accordingly (see e.g. Section 3.7 in Meerschaert & Sikorskii 2012). This distribution
is most concisely defined via its Laplace transform
E[exp(−sD)] = exp(−sβ ).
(4)
Moreover, the following functional limit theorem holds (see e.g. Remark 4.17 in Meerschaert
& Sikorskii 2012):
d
(W1 + . . . + Wbctc )/b(c) −→ D(t),
c→∞
with convergence in the Skorokhod J1 topology. The limit D(t) is a stable subordinator,
i.e. an increasing Lvy process with Laplace transform exp(−tsβ ).
It is well known (see e.g. Meerschaert & Scheffler 2004) that the renewal process then
satisfies the functional limit
d
N (ct)/b̃(c) → E(t) = inf{r : D(r) > t},
c→∞
for a scaling function b̃(c) ∈ RV∞ (β) which is asymptotically inverse to b(c), in the sense
of Seneta (1976) (p.20):
b(b̃(c)) ∼ c ∼ b̃(b(c)).
(also see Prop 4.15 in Meerschaert & Sikorskii 2012). The limit process E(t) is called the
inverse stable subordinator (Meerschaert & Straka 2013).
3.2
Magnitudes and Extreme Value Distributions
The extremal limit theorem then allows for an extension to functional (i.e. pathwise) limits:
assume that the norming sequences a(c) and d(c) are as in Theorem GEV. Then
d
[M (ct) − d(c)]/a(c) → A(t),
7
c → ∞.
Here convergence holds weakly in Skorokhod’s J1 topology, and the limit process A(t) is
an extremal process, with finite-dimensional distributions given by
P(A(ti ) ≤ xi , 1 ≤ i ≤ d) = FA (∧di=1 xi )t1 FA (∧di=2 xi )t2 −t1 . . . FA (xd )td −td−1 ,
(5)
where FA (x) is a GEV distribution (see e.g. Resnick 2013).
3.3
CTRM limit
The CTRM V (t) results from the running maximum M (n) via a time change by the renewal
process n = N (t). This is also reflected in its corresponding limit process:
Theorem MS1 (Meerschaert & Stoev 2008): The CTRM process V (t) = M (N (t))
satisfies the following functional scaling limit in the Skorokhod J1 topology:
d
[V (ct) − d(b̃(c))]/a(b̃(c)) → A(E(t)),
c → ∞.
The distribution of the hitting time of a level `∗ by the limit process A(E(t)) is also
known:
Theorem MS2 (Meerschaert & Stoev 2008): Let FA be the CDF of a GEV distribution, and let A(t) be the extremal process corresponding to FA via (5). For a given
threshold `∗ in the support of FA , the hitting time
B(`∗ ) = inf{t : A(E(t)) > `∗ }
is equal in distribution to (− log FA (`∗ ))−1/β X 1/β D, where X ∼ Exp(1) and D (defined in (4)) are independent.
The Mittag-Leffler distribution with parameter β ∈ (0, 1] is (except for β = 1) a heavytailed positive distribution. For σ > 0, we write ML(β, σ) for the distribution of σY , where
Y has Laplace transform E[exp(−sY )] = 1/(1 + sβ ). Due to e.g. Theorem 19.1 in Haubold
et al. (2011), the exceedance time from Theorem MS2 can be written as
B(`∗ ) ∼ ML β, (− log FA (`∗ ))−1/β
8
(6)
−0.05
−0.15
Shape
1.0 1.2 1.4 1.6
Modified Scale
1.5
2.0
2.5
3.0
1.5
Threshold
2.0
2.5
3.0
Threshold
Figure 3: Stability plot for the shape and scale parameters of the generalized Pareto distribution. Here, the magnitudes are standard Gumbel, hence ξ = 0 and σ = 1. Plots were
generated using the POT package in R (Ribatet & Dutang 2016).
4
Scaling limit of Exceedance Times
In classical extreme value theory, there are two main approaches for the inference on GEV
distributions: Block Maxima and Peaks-Over-Threshold (POT). Block Maxima subdivide
the time line into ‘blocks’ of equal length, and the maximum of each block is taken as
realization of the random variable Mn . This approach is not appropriate for CTRM data,
because maxima in same-sized blocks will not be i.i.d. As mentioned in Section 2, the
POT approach does extract i.i.d. X(`) (and T (`)), and can hence be directly used for
the inference on the extreme values of Jk . Figure 3 shows an example for the so-called
“threshold-selection” plots for data drawn from the standard Gumbel distribution. For a
range of thresholds, maximum likelihood estimators of the shape and scale parameters in the
GP distribution are computed and plotted. Note that the scale parameter is transformed
so that it appears independent of the threshold (see e.g. Section 4.3.4 in Coles 2001).
For the remainder of this article, we focus on statistical inference for the inter-exceedance
9
times T (`) of the threshold crossings. In our setting, the inter-exceedance times follow a
fractional Poisson process instead of a standard Poisson process, which is how our approach
generalizes the standard POT approach.
Theorem: Let b(c) be the scaling function (3), and let ` be a threshold from the support
[xL , xR ] of J. Furthermore let p := P(J > `). Under the assumption (2) and the
assumptions of the GEV Theorem, we have
T (`)/b(1/p) → ML(β, 1) as ` ↑ xR .
(7)
Proof: Let the scaling functions a(c) and d(c) be as in Section 3.2, and assume WLOG
` = a(n)`∗ + d(n) for some arbitrary but fixed `∗ , out of the support of A. We first note
that due to (1)
d
FJ (a(c)`∗ + d(c))c → FA (`∗ ), as c → ∞.
(8)
Taking the logarithm on each side and using the relationship z − 1 ∼ log(z) as z → 1, we
get
d
c(1 − FJ (a(c)`∗ + d(c))) → − log FA (`∗ ).
(9)
Hence we have
d
c · p → − log FA (`∗ )
(10)
as ` ↑ xR . Now we get
P [T (`)/b(1/p) > t] = P[V (b(1/p)t) ≤ `] = P[V (b(1/p)) ≤ a(c)`∗ + d(c)]
V (b(1/p)/b(c) · b(c) · t) − d(c)
∗
≤`
=P
a(c)
(11)
d
→ P(A(E(− log FA (l))−1/β t) ≤ `∗ ),
where the convergence follows due to Theorem MS1 (with c replaced by b(c)) and the fact
that
b(1/p)
b(1/cp · c)
=
∼ (cp)−1/β → (− log FA (l))−1/β as c → ∞.
b(c)
b(c)
(12)
In (12) we used (10) and b ∈ RV (1/β), which means that b(λc)/b(c) → λ1/β as c → ∞
uniformly in λ on each compact subset of (0, ∞). The limit in (11) is equal to
P (− log(FA (`∗ )))1/β B(`∗ ) > t .
10
Due to Theorem MS2 and (6), we know that (− log FA (`∗ ))1/β B(`∗ ) is M L(β, 1) distributed,
and hence the theorem follows.
Remark: By the above theorem, the inter-exceedance times are i.i.d. and approximately
Mittag-Leffler distributed, which means that the exceedance times form a fractional
Poisson process (Laskin 2003). Since the Mittag-Leffler distribution nests the exponential distribution (the special case where β = 1), the fractional Poisson process
generalizes the standard Poisson process. As the threshold height increases, the
threshold crossing events are effectively “thinned out”, and the thinned (or rarefied)
processes retain the characteristics of a fractional Poisson process (Gorenflo 2010),
with a growing scale parameter b(1/p).
Remark: A fractional Poisson process can be represented as a standard Poisson process
N (t), time-changed by an inverse stable subordinator E(t) (Meerschaert et al. 2011).
This pattern can be made sense of in our context: The exceedance times for a standard
extreme value process A(t) form a standard Poisson process N (t). Accordingly, the
exceedance times of the CTRM limit A(E(t)) must be N (E(t)), a fractional Poisson
process.
5
Inference on Exceedance times
Equation (7) implies that for large ` we may approximate the distribution of T (`) with
an ML(β, b(1/p)) distribution. Building on the POT (Peaks Over Threshold) method, we
propose the following estimation procedure for the distribution of T (`):
1. For varying thresholds ` ↑ xR , extract datasets of exceedance times {T (`, i)}i .
2. Fit a Mittag-Leffler distribution to each dataset, which results in the `-dependent
estimates {β̂(`)}` and {σ̂(`)}` .
3. Plot ` vs. β̂(`). As ` increases towards xR , β̂(`) stabilizes around a constant β̂. Use β̂
as an estimate for the tail parameter β of the Mittag-Leffler distribution of exceedance
times.
11
4. Approximate p ≈ |{k : Jk > `}|/n. Recall that b(c) is regularly varying with parameter 1/β, and hence has the representation b(c) = L(c)c1/β for some slowly varying function L(c). Assuming that the variation of L(c) is negligible, we hence plot
` vs. p1/β̂ σ̂(`). Again as ` increases towards xR , p1/β̂ σ̂(`) is expected to stabilize
around a constant σ̂0 . We then use p−1/β̂ σ̂0 as an estimate of the scale parameter of
the Mittag-Leffler distribution of exceedance times for the level `.
The above approach, though theoretically sound, benefits from the following practical
adjustments:
• We choose ` from the order statistics, i.e. ` is the k-th largest of the observations Xj ,
where k runs from kmin , kmin + 1, . . . , kmax . The datasets are then of length k − 1.
• We use k rather than ` for the horizontal axis of our plots.
• In Step 4, rather than plotting p1/β̂ σ̂(`) we plot k 1/β̂ σ̂(`). This changes σ̂0 by the
multiplicative constant n1/β̂ , but has the advantage that σ̂0 does not change if one
pre-processes the data by removing all observations below a certain threshold.
The estimates β̂ and σ̂0 give an estimate of the distribution of exceedance times, dependent on the threshold `:
T (`) ∼ ML(β̂, k −1/β̂ σ̂0 ).
For quick estimates of the Mittag-Leffler parameters the method of log-transformed moments by Cahoy (2013) was used.
6
Checking Model Assumptions
The CTRM model is based on three main assumptions, which are repeated below. For
each assumption, we suggest one means of checking if it holds:
I.i.d.: The pairs (Wi , Ji ) are i.i.d. An indication if this is true is given by an autocorrelation plot for the logarithms (to ensure finite moments) of the two time series.
Uncoupled: Each Wi is independent of each Ji . We propose an empirical copula plot to
check for any dependence.
12
Heavy tails: The waiting times are heavy-tailed with tail parameter β ∈ (0, 1). A Hill
plot can confirm this assumption.
The motivation behind a CTRM model is to fit extreme observations and their interarrival times. Observations below a certain threshold may be interpreted as noise and
discarded. Since the signal rather than the noise needs to satisfy our modelling assumptions,
we suggest that the above model assumptions are checked after the noisy small observations
are discarded.
7
Application to data
7.1
Simulated Data
To test our inference method, we have simulated 10000 independent waiting time and
magnitude pairs (Wk , Jk ). In order to have exact analytical values available for β and
σ0 , a distribution for Wk needs to be chosen for which b(n) from (2) is known. If we
d
choose Wk = D, where D is as in (4), then due to the stability property we have the
d
equality of distribution W1 + . . . + Wn = b(n)D, for b(n) = n1/β . Using the parametrisation
of Samorodnitsky & Taqqu (1994), a few lines of calculation (see e.g. the vignette on
parametrisation in Gill & Straka (2017)) show that D must have the stable distribution
Sβ (cos(πβ/2)1/β , +1, 0), which is implemented in the R package stabledist by Wuertz
et al. (2016).
By (7), the distribution of T (`) is approximately ML(β, p−1/β ) = ML(β, k −1/β n1/β ),
which means σ0 = n1/β . The distribution of Jk is irrelevant for the inference on β and σ0
(we have chosen unit exponential random variables). Figure 4 displays plots of β̂(`) and
σ̂(`) vs. k, the number of exceedances as ` traverses the largest observations Ji . Dotted
lines show 95% confidence intervals, which are derived from the asymptotic normality of the
log-moments estimators (Cahoy 2013) and the δ-method (Gill & Straka 2017). The dashed
lines show the actual values of β resp. σ0 , showing that our inference method effectively
identifies the right parameters.
13
ML scale
0
100000
scale parameter
1.0
0.5
0.0
tail parameter
1.5
ML tail
0
100
300
500
0
exceedances
100
300
500
exceedances
Figure 4: Tail and scale estimates for simulated data, with waiting times drawn from the
stable distribution Sβ (cos(πβ/2)1/β , +1, 0) with β = 0.8. Dashed lines are 95% confidence
intervals, dotted lines are the known theoretical values (0.8 and 100001/0.8 ).
14
1.0e+08
magnitudes
0.0e+00
2011
2012
2013
2014
2015
2016
times
Figure 5: Time Series of ‘Bitcoin days destroyed’, a measure for the traded volume of
bitcoins.
7.2
Bitcoin trading volumes
We study the daily trade volumes of Bitcoin since 2009, provided by Blockchain and downloaded from Data Market2 . Although the observations are regular (daily), extremes in
the trade volumes occur in bursts, which can be seen over timescales of months and years
(Figure 5).
Figure 6 shows plots identifying the candidates β = 0.9 and σ0 = 2000 for the tail and
scale parameters of the exceedance time distribution. That is, exceedances of a threshold
` at the k-th largest observed magnitude are distributed as
T (`) = ML(0.9, 2000 ∗ k −1/0.9 ).
A QQ plot (Figure 10) of the exceedance times at the level of the 200th largest event lends
credence to a heavy-tailed distribution with tail parameter near the value 1.
15
scale parameter
1.0
0.5
0
0.0
tail parameter
2000 4000 6000
ML scale
1.5
ML tail
0
50 100
200
0
50 100
exceedances
200
exceedances
100000 200000
0
magnitudes
Figure 6: Tail and scale plot of the bitcoin data.
Jan
Mrz
Mai
Jul
Sep
times
Figure 7: The solar flare time series data.
16
Nov
Jan
ML scale
0
50
100
150
4e+07
0e+00
scale parameter
0.0
0.5
1.0
tail parameter
1.5
ML tail
0
50
exceedances
100
150
exceedances
Figure 8: Stability plots for the tail and scale parameter of the Mittag-Leffler distribution
of the Solar Flare dataset.
7.3
Solar Flare Data
The “complete Hard X Ray Burst Spectrometer event list”
3
Dennis et al. (1991) is a
comprehensive reference for all measurements of the Hard X Ray Burst Spectrometer on
NASA’s Solar Maximum Mission from the time of launch on Feb 14, 1980 to the end of
the mission in Dec 1989. 12,776 events were detected, with the “vast majority being solar
flares”. The list includes the start time, peak time, duration, and peak rate of each event.
We have used “start time” as the variable for event times, and “peak rate” as the variable
for event magnitudes (Figure 7).
Figure 8 shows the fitted tail and scale parameters of the Mittag-Leffler distribution.
The dotted line for the tail parameter is at β = 0.85, yielding the plot of (transformed)
scale estimates σ̂0 = k 1/0.85 σ̂(`). We suggest the estimate σ0 = 2.5 × 107 seconds ≈ 289
days. The exceedance times of a threshold ` as high as the k-measurement are hence
2
3
https://datamarket.com/data/list/?q=provider:blockchain
https://umbra.nascom.nasa.gov/smm/hxrbs.html
17
Mittag-Leffler distributed as
T (`) ∼ ML(0.85, 289k −1/0.85 days).
8
Discussion
Comparing plots of the simulated data in Figure 2 with plots of the empirical data 5
and 7, it can be seen that the empirical data show long time intervals without any event
occurring, whereas for real-world data events appear to happen continuously. This seeming
discrepancy vanishes, however, when all but the kmax largest observations are discarded.
We repeat that the CTRM model focuses on extreme events and interprets smaller events
as noise.
The CTRM model, moreover, assumes a “pure” power-law for the event inter-arrival
times. Real datasets, however, are often exponentially tapered or truncated, see e.g. the
discussion of tempered power-law distributions by Meerschaert et al. (2012) and the truncated Pareto distribution by Aban et al. (2006). This means that the fit to a power-law
distribution may seem adequate for medium-size time scales, but that at very large timescales the power-law character of the distribution can weaken or disappear. This behaviour
seems to appear in both the Bitcoin data and the Solar-Flare data: the Hill plot for the
tail parameter is increasing towards lighter tails as the threshold reaches very high values.
This may mean that the CTRM model may overestimate the heaviness of the tail of the
waiting time distribution; or that a CTRM base on a tempered power-law renewal process
might be a more realistic model.
9
Conclusion
We have proposed the CTRM model for the extremes of “bursty” events. Burstiness is a
phenomenon of intermittency which has received a lot of attention in the recent statistical
physics literature. The CTRM model is a straightforward generalization of the POT (Peaks
over Threshold) model, for which the inter-arrival times between threshold crossings are
exponentially distributed. If the Mittag-Leffler tail parameter of the CTRM equals 1, then
18
the CTRM model reproduces this behaviour exactly, showing that the CTRM model nests
the POT model.
The scale parameter σ of the inter-event distribution follows a power-law σ ∝ p−1/β ,
where β ∈ (0, 1) is the tail parameter of the inter-event distribution and p the probability
that an event crosses a given threshold. From this power law, we have constructed “stability
plots”, from which estimates of β and σ can be read off.
We have demonstrated the applicability of the CTRM model with two real-world
datasets. Although the CTRM model possibly overestimates the heaviness of the tail
of the inter-arrival distribution, it clearly captures that the inter-arrival times of threshold crossings scale non-linearly with the treshold crossing probability, and follow a clearly
non-Poissonian behaviour. Thus we have shown that the CTRM model is a useful tool
for the modelling of extremes of bursty events, which requires very little computation
and which allows for straightforward ways of checking and criticizing model assumptions.
The analyses from this paper are readily reproduced via the R package CTRM, available at
https://github.com/strakaps/CTRM.
A
Diagnostic Plots
19
0.0
0.4
ACF
0.4
0.0
0
5
10
15
20
25
0
5
10
15
20
25
CrossCor (Exc & Exc Time)
Mittag−Leffler QQ Plot
Population Quantiles
−0.05
ACF
−20
−10
0
10
1e+00
Lag
0.05 0.10
Lag
1e−04
ACF
0.8
ACF (Exceedances)
0.8
ACF (Exceedance Times)
20
1
Lag
10
100 1000
Sample Quantiles
Threshold
Emp. Copula (Exc & Exc Time)
Hill Plot
12.400
50.500
0.444
1.0
0.2
0.0
0.6
0.4
y
alpha
0.8
1.4
375.000
0.0
0.2
0.4
0.6
0.8
1.0
15
x
123
231
338
Order Statistics
Figure 9: Diagnostic plots for the simulated data.
20
446
0.0
0
5
10
15
20
0
5
10
15
20
CrossCor (Exc & Exc Time)
Mittag−Leffler QQ Plot
−10
0
10
1e−03
0.10
0.00
−20
1e+01
Lag
Population Quantiles
Lag
−0.10
ACF
0.4
ACF
0.4
0.0
ACF
0.8
ACF (Exceedances)
0.8
ACF (Exceedance Times)
20
1
Lag
2
5
10
20
50
Sample Quantiles
Threshold
Hill2 Plot1
27
6
15
79
1
3
2
0.0
1
0.4
y
alpha
0.8
4
Emp. Copula (Exc & Exc Time)
0.2
0.4
0.6
0.8
1.0
x
142
205
Order Statistics
Figure 10: Diagnostic plots for the Bitcoin data.
21
268
0.2
−0.2
−0.2
0
5
10
15
20
0
5
10
15
20
CrossCor (Exc & Exc Time)
Mittag−Leffler QQ Plot
−5
0
5
1e−01
1e−03
Population Quantiles
0.1
0.0
−15
1e+01
Lag
0.2
Lag
−0.1
ACF
0.6
ACF
0.6
0.2
ACF
1.0
ACF (Exceedances)
1.0
ACF (Exceedance Times)
10 15
20
Lag
100
500
5000
50000
Sample Quantiles
Threshold
Emp. Copula (Exc & Exc Time)
Hill Plot
318.0
1440.0
22.7
1.0
0.0
0.5
0.4
y
alpha
1.5
0.8
2.0
8390.0
0.0
0.2
0.4
0.6
0.8
1.0
15
x
45
75
105
Order Statistics
Figure 11: Diagnostic plots for the Solar Flare data.
22
135
References
Aban, I. B., Meerschaert, M. M. & Panorska, A. K. (2006), ‘Parameter estimation for the
truncated pareto distribution’, J. Am. Stat. Assoc. 101(473), 270–277.
URL: http://projecteuclid.org/euclid.ss/1340110872
Anderson, K. K. (1987), ‘Limit Theorems for General Shock Models with Infinite Mean
Intershock Times’, J. Appl. Probab. 24(2), 449–456.
URL: http://www.jstor.org/stable/3214268
Bagrow, J. P. & Brockmann, D. (2013), ‘Natural emergence of clusters and bursts in
network evolution’, Phys. Rev. X 3(2), 1–6.
Barabási, A. L. (2005), ‘The origin of bursts and heavy tails in human dynamics’, Nature
435(May), 207–211.
Basrak, B. & Špoljarić, D. (2014), ‘Extremal behaviour of random variables observed in
renewal times’.
URL: http://arxiv.org/abs/1406.6208
Beirlant, J., Goegebeur, Y., Segers, J. & Teugels, J. (2006), Statistics of extremes: theory
and applications, John Wiley & Sons.
Benson,
of
D.
extreme
A.,
Schumer,
events
with
R.
&
Meerschaert,
power-law
interarrival
M.
M.
times’,
(2007),
‘Recurrence
Geophys. Res. Lett.
34(l16404), DOI:10.1029/2007GL030767.
URL: http://www.agu.org/pubs/crossref/2007/2007GL030767.shtml
Cahoy, D. O. (2013), ‘Estimation of Mittag-Leffler Parameters’, Commun. Stat. - Simul.
Comput. 42(2), 303–315.
URL:
https://www.tandfonline.com/doi/full/10.1080/03610918.2011.640094
http://www.tandfonline.com/doi/abs/10.1080/03610918.2011.640094
Coles, S. (2001), An Introduction to Statistical Modelling of Extreme Values, SpringerVerlag, London.
23
Dennis, B. R., Orwig, L. E., Kennard, G. S., Labow, G. J., Schwartz, R. A., Shaver, A. R.
& Tolbert, A. K. (1991), ‘The complete Hard X Ray Burst Spectrometer event list, 19801989’.
URL: https://umbra.nascom.nasa.gov/smm/hxrbs.html
Esary, J. D. & Marshall, A. W. (1973), ‘Shock Models and Wear Processes’.
Gill, G. & Straka, P. (2017), MittagLeffleR: Using the Mittag-Leffler distributions in R.
URL: https://github.com/strakaps/MittagLeffler/
Gorenflo, R. (2010), ‘Mittag-Leffler Waiting Time, Power Laws, Rarefaction, Continuous
Time Random Walk, Diffusion Limit’, p. 28.
URL: http://arxiv.org/abs/1004.4413
Gut, A. (1999), ‘Extreme Shock Models’, Extremes (1983), 295–307.
URL: http://link.springer.com/article/10.1023/A:1009959004020
Haubold, H., Mathai, A. M. & Saxena, R. K. (2011), ‘Mittag-Leffler Functions and Their
Applications’, J. Appl. Math. 2011, 1–51.
URL: http://www.hindawi.com/journals/jam/2011/298628/
Hawkes, A. G. (1971), ‘Point spectra of some mutually exciting point processes’, J. R. Stat.
Soc. Ser. B Stat. Methodol. pp. 438–443.
Hees, K. & Scheffler, H.-P. (2016), ‘On joint sum/max stability and sum/max domains of
attraction’, pp. 1–31.
URL: http://arxiv.org/abs/1606.03109
Hees, K. & Scheffler, H.-P. (2017), ‘Coupled Continuous Time Random Maxima’, Extremes
(June), 1–24.
URL: http://link.springer.com/10.1007/s10687-017-0304-6
Karsai, M., Kaski, K., Barabási, A. L. & Kertész, J. (2012), ‘Universal features of correlated
bursty behaviour’, Sci. Rep. 2.
24
Karsai, M., Kivelä, M., Pan, R. K., Kaski, K., Kertész, J., Barabási, A. L. & Saramäki, J.
(2011), ‘Small but slow world: How network topology and burstiness slow down spreading’, Phys. Rev. E - Stat. Nonlinear, Soft Matter Phys. 83, 1–4.
Laskin, N. (2003), ‘Fractional Poisson process’, Commun. Nonlinear Sci. Numer. Simul.
8(3-4), 201–213.
Meerschaert, M. M., Nane, E. & Vellaisamy, P. (2011), ‘The fractional Poisson process and
the inverse stable subordinator’, Electron. J. Probab. 16, 1600–1620.
Meerschaert, M. M., Roy, P. & Shao, Q. (2012), ‘Parameter estimation for exponentially
tempered power law distributions’, Commun. Stat. - Theory Methods 41(10), 1839–1856.
Meerschaert, M. M. & Scheffler, H.-P. (2004), ‘Limit Theorems for Continuous-Time Random Walks with Infinite Mean Waiting Times’, J. Appl. Probab. 41(3), 623–638.
Meerschaert, M. M. & Sikorskii, A. (2012), Stochastic models for fractional calculus, Vol. 43,
Walter de Gruyter.
Meerschaert, M. M. & Stoev, S. A. (2008), ‘Extremal limit theorems for observations
separated by random power law waiting times’, J. Stat. Plan. Inference 139(7), 2175–
2188.
URL: http://linkinghub.elsevier.com/retrieve/pii/S0378375808003984
Meerschaert, M. M. & Straka, P. (2013), ‘Inverse Stable Subordinators’, Math. Model. Nat.
Phenom. 8(2), 1–16.
URL: http://www.mmnp-journal.org/10.1051/mmnp/20138201
Min, B., Goh, K. I. & Vazquez, A. (2011), ‘Spreading dynamics following bursty human
activity patterns’, Phys. Rev. E - Stat. Nonlinear, Soft Matter Phys. 83(3), 2–5.
Oliveira, J. & Barabási, A. L. (2005), ‘Darwin and Einstein correspondence patterns’,
Nature 437(October), 1251.
Omi, T. & Shinomoto, S. (2011), ‘Optimizing Time Histograms for Non-Poissonian Spike
Trains’, Neural Comput. 23(12), 3125–3144.
25
Resnick, S. I. (2013), Extreme values, regular variation and point processes, Springer.
Ribatet, M. & Dutang, C. (2016), POT: Generalized Pareto Distribution and Peaks Over
Threshold. R package version 1.1-6.
URL: https://CRAN.R-project.org/package=POT
Samorodnitsky, G. & Taqqu, M. S. (1994), Stable Non-Gaussian Random Processes:
Stochastic Models with Infinite Variance, Stochastic Modeling, Chapman Hall, London.
Seneta, E. (1976), Regularly Varying Functions, Vol. 508 of Lecture Notes in Mathematics,
Springer-Verlag, Berlin.
Shanthikumar, J. G. & Sumita, U. (1983), ‘General shock models associated with correlated
renewal sequences’, J. Appl. Probab. 20(3), 600–614.
URL: http://www.jstor.org/stable/3213896
Shanthikumar, J. G. & Sumita, U. (1984), ‘Distribution Properties of the System Failure
Time in a General Shock Model’, Adv. Appl. Probab. 16(2), 363–377.
URL: http://www.jstor.org/stable/1427074
Shanthikumar, J. G. & Sumita, U. (1985), ‘A class of correlated cumulative shock models’,
Adv. Appl. Probab. 17(2), 347–366.
URL: http://www.jstor.org/stable/1427145
Silvestrov, D. S. (2002), Limit Theorems for Randomly Stopped Stochastic Processes,
Springer (Berlin, Heidelberg).
Silvestrov, D. S. & Teugels, J. L. (2004), ‘Limit theorems for mixed max-sum processes
with renewal stopping’, Ann. Appl. Probab. 14(4), 1838–1868.
URL: http://projecteuclid.org/Dienst/getRecord?id=euclid.aoap/1099674080/
Vajna, S., Tóth, B. & Kertész, J. (2013), ‘Modelling bursty time series’, New J. Phys.
15(10), 103023.
URL: http://stacks.iop.org/1367-2630/15/i=10/a=103023?key=crossref.eac401be0d0a5b8ecbf34a51a9f1
26
Vasquez, a., Oliveira, J. G., Dezso, Z., Goh, K.-I., Kondor, I. & Barabási, A. L. (2006),
‘Modeling bursts and heavy tails in human dynamics’, Phys. Rev. E 73, 361271–3612718.
Vazquez, A., Rácz, B., Lukács, A. & Barabási, A. L. (2007), ‘Impact of non-poissonian
activity patterns on spreading processes’, Phys. Rev. Lett. 98(APRIL), 1–4.
Wuertz, D., Maechler, M. & core team members., R. (2016), stabledist: Stable Distribution
Functions. R package version 0.7-1.
URL: https://CRAN.R-project.org/package=stabledist
27
| 10math.ST
|
This paper has been accepted for publication in forthcoming issue of IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. The
content is final but has NOT been proof-read. This is an author copy for personal record only.
SRE: Semantic Rules Engine For the Industrial InternetOf-Things Gateways
Charbel El Kaed, Member, IEEE, Imran Khan, Member, IEEE, Andre Van Den Berg, Hicham Hossayni and Christophe SaintMarcel
Abstract—The Advent of the Internet-of-Things (IoT) paradigm has
brought opportunities to solve many real-world problems. Energy
management, for example has attracted huge interest from academia,
industries, governments and regulatory bodies. It involves collecting
energy usage data, analyzing it, and optimizing the energy consumption
by applying control strategies. However, in industrial environments,
performing such optimization is not trivial. The changes in business
rules, process control, and customer requirements make it much more
challenging. In this paper, a Semantic Rules Engine (SRE) for industrial
gateways is presented that allows implementing dynamic and flexible
rule-based control strategies. It is simple, expressive, and allows
managing rules on-the-fly without causing any service interruption.
Additionally, it can handle semantic queries and provide results by
inferring additional knowledge from previously defined concepts in
ontologies. SRE has been validated and tested on different hardware
platforms and in commercial products. Performance evaluations are
also presented to validate its conformance to the customer
requirements.
Index Terms—Internet-of-Things; Semantic Web; Industrial Control;
Rules; Industrial Internet-of-Things; Embedded Systems
I. INTRODUCTION
The Internet-of-Things (IoT) is expected to interconnect
sensors, devices, gateways, and objects on a massive scale [1] to
solve many challenges. Such interconnectivity will play an
important part to build powerful industrial systems [2] that are
more energy efficient with lower costs while contributing to a
better environment. In addition, the Semantic Web technology
[3] is getting more popular in the development of rich and
interactive applications [4]. In industrial environments and
automation domains, the Semantic Web has been used to solve
data interoperability issues [5], [6] and to provide context aware
applications and services [7].
In industrial facilities, black-box devices (gateways or data
loggers) are typically used to collect data from variety of devices
like; sensors, actuators, machines, plants, processes, and systems.
Typically these gateways, sample, collect, and push data to a
remote platform for further analysis and may also send
notifications for certain events.
In this era of inter-connectivity and optimization, users and
managers desire some sort of flexible control over their
installations. Due to ever changing business requirements and the
nature of IoT applications and services, it has become
The authors are with IoT & Digital Transformation, Schneider Electric,
Boston One Campus, Andover, MA, 01810, USA and 38TEC, Grenoble, 38000,
France. (e-mail: {charbel.el-kaed, imran2.khan}@schneiderelectric.com,
[email protected], {hicham.hossayni,
christophe.saint-marcel}@non.schneider-electric.com)
difficult to preempt the requirements of the users from
manufacturer perspective. More dynamic solutions are required
that by design offer more flexibility and better control to theses
gateways users. Further more, issues like low bandwidth, latency,
and security in the industrial domain are additional factors to
consider [8]. Hence, easy-to-use solutions located closer to the
devices are preferred but issues like device and protocol
heterogeneity, multi-vendor solutions, and variety of data
models make it a challenging task.
In this paper, we present Semantic Rules Engine (SRE) which
consists of two parts: a Rules Engine and a Semantic Engine. The
Rules Engine provides a simple and effective way to deploy a
control mechanism (as rules) on gateways. These rules are
expressed in a simple scripting language and can be modified and
uploaded at run-time without disrupting the operation of the
gateways. The Semantic Engine provides absolute abstraction
from the heterogeneity of devices, protocols, data and any
topological changes. It leverages devices meta-data and enables
the retrieval of contextual information using semantic queries.
Inspired from our previous work [9], we leverage a modular
approach with a set of common and domain specific ontologies
across our enterprise. For instance, a domain ontology capturing
one of our industrial automation contexts, is used to annotate the
devices with their contextual information, thereby promoting
data interoperability and its understanding to users and
applications.
The following are key differentiating factors of SRE. 1) The rules
rely on devices’ annotations and are not tied to the devices ID or
any other binding. Hence, they are flexible and independent of
topology changes and assigning unique identifiers does not affect
their execution. 2) Rules themselves use semantic queries to
retrieve contextual information in order to perform
actions/eventing functions. Thus, higher-level concepts like,
location and measurement type can be used to compose logic in
rules to achieve a desired behavior from the gateway. 3) We have
developed a simple reasoning module in SRE to answer ondemand queries. For example, a device’s location is in a room A,
and such room is located on a floor within a building A. The
reasoning feature will deduce that the device’s location is also in
building A. This is powerful when aggregating measurements or
searching for specific device types. 4) SRE provides complete lifecycle management of its rules, they can be installed, activated,
modified at run-time, deactivated, and uninstalled. 5) SRE
supports cooperation between rules to share functions which are
This paper has been accepted for publication in forthcoming issue of IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. The
content is final but has NOT been proof-read. This is an author copy for personal record only.
needed by many rules thereby avoiding redundancy. 6) SRE uses the industrial process. One simple solution is to install light
the concept of settable parameters in rules. These parameters indicators, along the industrial process, to inform the local
can be updated on-the-fly without disturbing the execution of the personnel about situations like energy wastage (Orange light ON),
rules.
process fault (Red light ON) or normal operation (Green light ON).
The rest of the paper is organized as follows: Section II presents The audit team along with the facility manager proposed a set of
the background, motivating scenario, requirements and the rules to create this solution.
related works. Section III describes the proposed solution while
section IV details the implementation and evaluation. Section V
outlines lessons learned and future work while Section VI
concludes the paper.
II. BACKGROUND, REQUIREMENTS, AND RELATED WORK
This section presents the background, motivating scenario,
requirements and a review of the related works.
A. Background
Energy consuming devices have evolved from just being able to
switch ON or OFF to be more sophisticated for control and
monitoring purposes. However, these devices have mainly
worked in isolation, only taking into account their own
measurements without any sort of intelligence or learning. With
the advent of IoT paradigm, the “smart devices” of today are
much more advanced than their predecessors. They can
communicate with each other using various protocols and can
push data of their connected sensors and actuators to a remote
cloud platform where it can be analyzed.
However, relying on a cloud platform to remotely control
devices is sometimes not desirable for various reasons ranging
from the privacy concerns to latency issues and to minimize the
downtime or to enable a facility to operate even without the
cloud connectivity. More over, once the data has been analyzed
in the cloud, normally there is no clear defined mechanism to
deploy/modify the behavior of the devices, such as changing the
push interval of the data or to select additional data points from
a specific set of devices from a particular location.
Thus, there is need to be able to create control strategy (i.e.
rules) and execute them on the gateways for more localized
control. These rules can range from simply sending an alarm when
the temperature in an area reaches a certain threshold, to a more
complex rule to control a liquid flow within an underground
mining facility [10].
B. Motivating Scenario and Requirements
In one of our industrial manufacturing facility, an energy
monitoring solution has been deployed. The solution consists of
an IoT gateway and a set of wireless sensors deployed along the
industrial manufacturing machines. The IoT gateway collects data
from the sensors and pushes it to the remote cloud platform for
visualization and analysis, as shown in Fig. 1.
In our continuous efforts to maintain green facilities, an
internal energy audit concluded the following: most of the energy
waste is occurring in the semi-automated industrial process.
Energy can be saved by engaging our local personnel working on
Fig. 1. Our Industrial Facility Use Case
For example, in Fig. 1, when the conveyor is ON while the mold
and the furnace are OFF, energy is being wasted. However, for
preheating purposes, it is acceptable for a short period of time
for the furnace alone to be ON.
From this simple scenario, it is clear that the facility manager
needs a solution that gives her more flexibility to tackle new
situations that were not envisioned during the initial setup. The
solution needs to be dynamic, flexible and should not involve any
reprogramming or hard reset of the devices. It should also be
usable by non-IT experts.
There are two possible solutions : One is to use a tailored
solution to solve one problem only and when new requirements
arise, reprogram everything again. This option does not scale well
and is not suitable to solve new problems or implement new
business rules efficiently. The second option is to have a more
flexible solution that is easier to implement, does not require
reprogramming and supports quick implementation of new
business rules. Thus, the solution requires the following
characteristics: 1) Expressing requirements and conditions in an
explicit manner to control the devices and to send alerts. 2)
Deployment strategy of these requirements. 3) Efficient
execution of these requirements. 4) Support semantics to query
status and details of the devices by location, type, connectivity,
and other concepts. 5) Additionally, the solution should be
suitable for different types of devices (gateways, edge, fog nodes,
and Raspberry Pi), easily integrate-able, modular in design to
implement new features, and be independent of device types,
protocols or application domain constraints.
This paper has been accepted for publication in forthcoming issue of IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. The
content is final but has NOT been proof-read. This is an author copy for personal record only.
In literature, we found several Rules Engines and Semantic of a rule. In fact, our early investigation started with drools [22],
Engines, we discuss some of them in the following subsection.
however, due to its rigid dependency on Java and its packages, we
opted for a Lua based approach which is light weight and portable
C. Related Work
even on smaller footprint gateways.
The work in [11] proposes LINC, a coordination programming
On the Semantic Web side, the authors in [24] discuss a unified
environment to execute Rules efficiently. LINC uses different semantic engine for IoT and Smart Cities to tackle data
Linda-like tuple-spaces. The global tuple space is split into ”bags”. interoperability and scalability among others. However, their
Each bag can either encapsulates a database, service (SOAP or work does not provide any details regarding the implementation
RPC), event system or hardware like a sensor or an actuator. A or technologies used to address the identified issues. A Semantic
coordination language is used to specify Rules with statements gateway is proposed in [25] to tackle the industrial field service
and has operation types get, put and read to interact with the use-case where service engineers rely on a plethora of tools to
”bags” and identify their content using keys or identifiers. track and solve the issues of the products which may have been
Dependence on identifiers and keys makes LINC unsuitable for installed decades ago. The semantic aspect in this work only deals
dynamic environment because rules will need updates as well. with the mapping of the SOAP bindings to the networking
This is not the case in SRE, which remains immune to such platform. A gateway architecture is presented in [26] that
changes thanks to the use of semantic queries. Also, compared to abstracts the network services and translates network services to
SRE, LINC does not provide any query mechanism except the keys a standard DPWS interface but there is no notion of queries or
to retrieve objects from bags. Therefore it is not possible to look rules.
up various things on the gateway based on their contextual
The authors in [27], [28] present a semantic smart gateway
information such as their measurements types and/or physical framework to achieve device interoperability. Features such as
locations. Moreover, a simple rule example as described in [11] ‘on-the-fly’ ontology learning and ontology alignment are
does not seem trivial for a developer compared to Lua in SRE provided. Similarly, our previous work[6] proposes a
following an EventCondition-Action paradigm. It is also not clear semiautomatic approach to dynamically generate and install
if LINC uses higher-level of abstraction for Rules similar to SRE.
proxies on the gateway from ontology alignments to achieve
Authors in [12] propose a Rules Engine to efficiently handle the semantic interoperability. However, these related approaches do
high data throughput and large set of rules in the context of smart not fulfill our requirements of query and control using rules. In
building systems. Due to high frequency of data, it is difficult to our approach, instead of relying on ontology alignment, we opted
apply corresponding Rules from a large set, for example to handle for the commissioning phase described in our previous work
urgent events. The sensor and actuator data are used to find SQenIoT [9] to semantically annotate data sources from
atomic events then rules are used to construct the minimum ontologies when the gateway is commissioned.
prefect hash table to filter meaningless atomic events. Using a
III. PROPOSED SOLUTION
feedback mechanism the rule matching overhead is further
By carefully considering requirements, constraints and
reduced. The main strengths of this approach are scalability and
rules conflict management with the help of the user. However, in viewpoints of the stakeholders we concluded that our proposed
their approach, a rule is described in an XML data structure solution will have the following features: Flexibility: design rules
limiting the expressiveness of the rule, e.g., it is not clear how a to express, deploy, and execute user requirements quickly with
loop is handled or whether atomic conditions with logical time and cost savings and avoiding software development lifeoperations can be expressed to trigger an action. In addition, each cycle for every requirement. Deployment: install rules on the
rule is tied to a sensor identifier along with a room and a user gateways either locally or remotely through a cloud application.
identifier making it very rigid when a device is replaced or its Interoperability: leverage the ontology concepts to provide
physical location changes. SRE does not have these drawbacks consistent, reusable, and shareable view about the data.
because it relies on the Semantic Query Engine to handle dynamic Semantic Query: rely on a natural language-like grammar to
environments and the Lua language constructs for better retrieve information from the gateway based on annotated data
liberating rules from static unique identifiers. Modularity: to
expressiveness.
There are several other works that address the same problem easily integrate new functionalities and to scale. Finally, the
using approaches such as programming models like [13] and [14], solution should provide dedicated rules execution environment.
After a careful study of the state-of-the-art and available open
middleware like [15], coordination languages like [16] and [17]
source
rules engines like [22], [23], we decided to design the
and the solutions based on tuplespace like [18], [19], [20] and
Semantic
Rules Engine (SRE) to fulfill the required features
[21]. Most of these works provide complex solutions that are
tailored for a particular scenario or are difficult to use in medium mentioned earlier. The SRE comprises of two parts: the Rules
to small gateways. Also some of these works are prior to the IoT Engine and an existing work Semantic Engine [9]. The overall
design architecture of SRE is shown in Fig. 2. The components are
paradigm and do not consider the challenges it introduces.
Other existing Rules Engines such as [22], [23] are designed to detailed in the following subsections.
run with an underlying Java Virtual Machine with a clear
dependency on java packages as shown in the import declaration
This paper has been accepted for publication in forthcoming issue of IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. The
content is final but has NOT been proof-read. This is an author copy for personal record only.
A. Rules Engine
a sensor or measure and is capable of handling changes in the
The Rules Engine provides gateways with the decentralized topology (e.g. device joining/leaving/being replaced).
The interaction between the Rules Engine and SQenIoT is a
intelligence and the ability to control or query the connected
devices. Figure 3 shows the components of the Rules Engine. A major differentiating factor from other solutions. In fact, existing
Rules Manager interacts with remote applications, receives and Rules Engines rely on a unique identifier to address a sensor,
manages the rule files. Once started, the Rules Manager loads the actuator, or measurement. Since the identifier is hardcoded in
execution environment used to execute the rules. An execution the rule, it requires manual update when, for example, a faulty
environment provides a set of functions, for example to handle sensor is replaced by a new one. On the other hand, by relying on
subscriptions, timers, and interactions with other components contextual information and meta-data (tags), the rule will still be
like SQenIoT and the Cloud Connectivity Agent. Once the valid even when the sensor is replaced since the meta-data will
execution environment is setup, the Rules Manager handles the remain the same for the new sensor.
life-cycle of the rules (e.g. install, start, stop or delete). Each rule
Overall, with this capability, SRE offers a flexible solution where
is executed in isolation to avoid conflicts with other rules and the same rules can be reused in different products and solutions
resources access issues.
as long as higher-level concepts (like location) are reused. This
provides time and cost savings as well as same level of
functionality across different domains and platforms.
Fig. 2. Design Architecture of Semantic Rules Engine
B. Semantic Engine: SQenIoT
To provide the rules with the ability to integrate semantic
queries, we reused existing work SQenIoT [9]. An API is provided
with SQenIoT which makes it possible for SQenIoT to access the
device meta-data. The values of the meta-data tags can be
injected during the commissioning phase manually or
automatically as described in [9]. By combining Rules Engine with
SQenIoT, rules are able to execute semantic queries and use the
returned results. For example, a rule can delegate the aggregation
of temperature of a given floor to SQenIoT through a query. Then,
the returned results will be used in another query to trigger an
action or an event.
The formulated query will rely on annotated data in the
gateway such as location (i.e. a floor) and measurement (i.e.
temperature) tags as means to get the list of temperature sensors
to get the average temperature of a given floor. Once received by
SQenIoT, it gets the values of the temperature sensors of that
particular floor only and returns the average temperature values.
The Rules Engine can then compare the result with a given
threshold and execute the action (a notification, an alarm, or an
actuation) according to the logic specified in the rule. This
semantic query feature decouples the rule from any identifier of
Fig. 3. SRE Implementation Architecture
C. Rule Design
The users’ requirements, regarding control or information are
expressed in the form of rules. Therefore the rules are designed
to enable the following:
1) Event-Condition-Action (ECA) paradigm on the resources
(i.e. devices/sensors/actuators) and their capabilities
(measurement, points).
2) Publish/Subscribe mechanism on a resource or
measurement with a related condition.
3) Support life-cycle management process (install, start, stop,
and uninstall).
4) Orchestration and cooperation between rules which allow
delegation and reuse.
These features are elaborated as follows:
This paper has been accepted for publication in forthcoming issue of IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. The
content is final but has NOT been proof-read. This is an author copy for personal record only.
1) ECA Design: The condition part of a rule is designed to be the unique identifier of the sensor and its resource, for example,
flexible. It takes only one resource or capability into account. To SRE.subscribe(@exist[Sensor11]
evaluate multiple resources and their capabilities, a rule can have == True, HandleDynamicSensorCallback)1. The second mechanism
multiple conditions combined by logical operators. The following goes through SQenIoT with a semantic query based on the sensor
syntax enables to look up ”SensorA” and its measurement value metadata instead of its unique identifier, for example,
”Temperature”:
SRE.query("subscribe catalog:sensor and location:Room1",
[SensorA] Temperature
HandleDynamicSensorCallback).
Using this definition, multiple conditions can be combined
3)
Rule Life-cycle Design: One of the requirements for the
together as follows and are also shown in Fig.4
rules is to be able to manage them easily. For example, a rule that
lowers the temperature of a hot water geyser might only be
[MultiSensorA]Temp > 25◦ C AND
started on weekends and stopped on weekdays. We designed
[DoorSensorB]isOpen == True
life-cycle management of rules, as shown in Fig.5, to easily
The supported logical operators are AND and OR, whereas the
manage them using a Graphical User Interface (GUI).
allowed comparators are: =,<,==,≤,≥,<>,>.
4)
Orchestration and Cooperation: To avoid rule conflicts
The following types of conditions are defined to evaluate
and to enable reuse, orchestration is primordial. Consider for
resources and their capabilities.
example, an eventing function which sends an notification with a
A) Evaluator: Evaluates the measurement of a resource and its
description to a remote monitoring center, such function can be
capability: [SensorA]Temp>25◦ C It is useful when a
implemented once and reused by different rules. Therefore, SRE
capability reaches a threshold or is equal (or not) to a
enables rules cooperation and orchestration.
specific value.
5)
Rule Persistence: Rules are persisted on the device along
with their states (installed, started, stopped) and the recent
values of their settable parameters. When a rule is created, it is
first encoded in Base64 then the resulting JSON file is digitally
signed (for security) and compressed as ZIP file. SRE receives this
ZIP file and validates its signature. If valid, the ZIP file is stored
locally and the rule is extracted and loaded in the SRE for
installation. Additionally, SRE saves the state of the rules locally
in a small database with details such as name of the rules, list of
Fig. 4. Example of Condition Consisting of Two Basic Condition Blocks
settable parameters, and their most recent values. If device
reboots, SRE first validates the local ZIP files, extracts the rules
and, by using the details in the database, restores them to their
B) Existence: This condition is used to check if a resource exist,
last execution state along with the last stored values of the
for
example,
in
a
dynamic
context,
parameters.
appearance/disappearance of a device. @exist[SensorA]
Finally, although it is not the focus of this paper, we did
== True
consider the security aspects to allow safe deployment of rules in
C) EventOnChange: A granular subscription on the change of our gateways for execution. Only rules, signed with corporate PKI,
state/value of a capability (e.g. window is open or close) is are executed by SRE. Every received rule is first validated to check
of interest.
its origin and authenticity and then installed in the gateway,
@change[DoorSensor1]State == True
otherwise it is discarded.
D) Increment/Decrement: Enables the subscription on the
increase/decrease
of
the
capability
value.
IV. IMPLEMENTATION & EVALUATION
@incr[SensorA]Temp == True (@decr)
In this Section, we first present the SRE implementation
The Existence, EventOnChange, and Increment/Decrement choices and then the setup. Later, performance metrics and the
operations are coupled with a callback function. The logical results are discussed. The section concludes with the
operators AND and OR can be combined to specify powerful presentation of real use-cases where SRE is being used.
conditions in the rules. By relying on SQenIoT, users do not need
to know the name of the resource and capability they are
A. Implementation Choices
interested in, instead they can use higher-level concepts like
1) Rules Engine: is expected to run on embedded devices as
location, device type, and physical quantity as described in [9].
2)
Publish/Subscribe Mechanism: There are two well as resourceful devices. There were quite a few choices like
mechanisms to express publish/subscribe requests. The first one JESS, JBoss Drools [22], Lua and OpenRules. However, Lua [29]
is identifier dependent, which means that the rule must contain was selected due to the following factors: 1) Developed for
embedded devices but it is also usable for resourceful devices. 2)
1
https://github.com/InnovationSE/SRE/blob/master/TII/LuaAlertEx.lua
This paper has been accepted for publication in forthcoming issue of IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. The
content is final but has NOT been proof-read. This is an author copy for personal record only.
Open source, has a strong user base, and its interpreters are the account by the rule in its next execution cycle. Several
available in several languages including Java and C for integration semantic queries can be combined together to compose complex
within different platforms. 3) Easy to use syntax. 4) rules. Lua also supports userdefined functions so complex
Expressiveness of basic and complex rules. 5) Multiple execution computations can be performed in these functions for more
environments to run rules independently. The Lua interpreter can advanced control features.
exist in a separate process on the gateway and therefore has less
2) SQenIoT: The implementation is shown in Fig. 3. Like the
interference with the main framework and can be shutdown and Rules Engine, it is also developed in both Java and C languages. It
restarted independently. The design of Rules Engine is shown in exposes a Query Interface through which rules can send semantic
Fig. 3. A rule is basically a Lua script executed by the Lua queries. The necessary functions are provided by the Engine API
interpreter. Both Java and C version of Rules Engine were module. For semantic support SQenIoT uses open source libraries,
developed for different hardware platforms offering same namely OWL API2 and Redland RDF Libraries3 in Java and C version
functionality. Using the Engine API, rules can interact with respectively.
functions to handle subscriptions and timers. The execution
environment is created by the Lua interpreter and ensures that
multiple rules co-exist without any issues. Also, rules can call Java B. Implementation Setup
methods or C functions for more complex operations.
SRE Java version is deployed on a commercial off-theshelf
An example of a Lua rule is shown in Listing 1, it regulates
industrial Gateway which has Modbus 4, Ethernet, WiFi, ZigBee,
the luminosity levels of the lights in a given site based on the
and GPRS interfaces to connect to devices and a remote cloud
average luminosity level. Lines 10, 11, 15, and 16 of Listing 1
platform. The gateway uses Linux OS with IBM J9 [30] as Java
demonstrates the expressivity of SRE through semantic
Virtual Machine (JVM) on top. J9 is certified for embedded
queries. Lines 10-11 shows a semantic query which delegates
systems in industrial usage. The ProSyst5 implementation of Open
to SQenIoT, the average luminosity value from the Sensors
Service Gateway initiative (OSGi) framework 6 is used over J9 for
located in Site1. This query makes it very easy to retrieve the
modular design. SRE is implemented as one of the OSGi
aggregated value from the sensors without a reference to any
component.
fixed identifier or other bindings. In addition, the inference is
SRE C version is implemented for our in-house hardware
requested through the @ symbol in the query. To calculate
platform based on dual core Cortex A9 chip clocked at 900Mhz
with 1GB RAM. It also runs Linux 7 for embedded systems. The
connectivity options include, ZigBee, Ethernet, Bluetooth and WiFi. SRE is developed as a POSIX compliant library that can be
imported by client applications.
Fig. 5. Lifecycle of a Rule
1
2
3
4
function LightControl.init()
engine.timer("LightControl", "Control", 500,2000,-1)
defaultThreshold = 600
end
6 function LightControl.Control()
7 threshold = engine.getRuleSetting("LightControl","threshold") 8 if( ˜ threshold)
the inference, SQenIoT, relies on the annotated data and the
location ontology created during the commissioning phase which
contains concepts and relationships between Site1 and all
locations in it (e.g. floors, rooms). The Sensors are not directly
tagged with location:Site1, they were annotated with their
locations (mainly rooms and open spaces). SQenIoT relies on the
ontology to infer and retrieve the sensors located in the site. This
feature is a major differentiating factor of SRE from other
solutions.
If the average luminosity is below a certain threshold, then the
light actuators are queried (lines 15-16) and a new value is set in
lines 18-22 of the Listing. Another interesting feature is on line 7
which allows a rule to get threshold value dynamically. A new
threshold value can be provided on the fly and it will be taken into
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
threshold = defaultThreshold
averageLuminosityOfSite1_query =
"Avg variable usage:LuminositySensor and @loc:Site1"
averageLuminosityOfSite1_result
= engine.query(averageLuminosityOfSite1_query)
if (averageLuminosityOfSite1_result < threashold)
light_actuators_devices =
"Search Device usage:LightActuator and @loc:Site1"
lightActuators = engine.query(light_actuators_devices)
for i = 1, #lightActuators, 1 do
lux = engine.getCapability(sensors[i],"LuminositySetPoint")
differenceToSet = threshold - lux.value
if(differenceToSet > 0)
engine.setValue(lux, lux.value+differenceToSet)
end
end
end ...
2
5
3
6
http://owlapi.sourceforge.net
http://librdf.org
4 http://www.modbus.org
http://www.prosyst.com
https://www.osgi.org
7 http://www.windriver.com/products/linux
This paper has been accepted for publication in forthcoming issue of IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. The
content is final but has NOT been proof-read. This is an author copy for personal record only.
Listing 1: Lua Rule Example
C. Performance Metrics
The performance of SRE is assessed in terms of Rule Execution
Time (RET), Rule Latency Time (RLT) & Rule Memory Usage
(RMU). RET and RLT are measured in milliseconds.
RET measures the time needed to execute rules. It includes the
time needed to parse the rule condition, its evaluation, and the
execution of an action. To evaluate RET multiple sets with
different number of rules are defined. Each rule defines a
condition on 5 smart devices. The condition part results in a true
evaluation each time to ensure that the action (i.e. toggle the
state of the actuator and send message to cloud) is executed.
RLT measures the average latency between the occurrence of
an event and the execution of the corresponding rule. For this we
simulated occupant comfort scenario in three multi-story
residential buildings to control the heater and air conditioner
based on temperature. Each building has ten floors, which have
three rooms each. Each room has a temperature sensor. In order
to test the robustness of SRE we stress-tested it using a Lua script8
to control the heater and air conditioner in each room when it’s
too cold or too hot respectively. SRE evaluates and executes rules
in a sequential manner so it is important to determine how much
latency is introduced when multiple rules are enabled. We define
different sets of rules where each rule monitors a temperature
sensor. Then, periodically and simultaneously, multiple events
are generated on all temperature sensors to activate
corresponding rules.
RMU measures the amount of memory needed to manage
certain number of rules, it is measured in Kilobytes.
D. Results
For RET, recall that the action part of the rules consists of
changing the value of a device and sending a message to the
Fig. 7. Rules Latency Time Evaluation
cloud. Figure 6 shows the results of the RET evaluation, it shows
that only 12ms are necessary to evaluate and execute 100 rules,
while 56ms are needed to evaluate 500 rules. We can conclude
that on average the evaluation and execution of each rule takes
about 0.12 ms. RET shows that the Rules Engine is suitable to
execute large number of rules, however in real situations,
executing 100s of rules will not be common.
Figure 7 shows the results of RLT evaluation. It shows the
averages of latency time of different scenarios. Each scenario
consists of the execution during 2 minutes of a specific number of
rules between (1, 10, 30, 50, 70, 90, 100, 200, 300, 400), and the
generation of a number of events for each rule among (5, 10, 11,
15, 20, 30, 100, 200) events per second. It is found that for all sets
of rules, the delay is negligible when the number of
events/second are less than or equal to 11. After this latency
becomes noticeable and exceeds 80ms for 200 events.
Figure 8 shows RMU, i.e. amount of memory needed by SRE to
manage different number of rules. We found that even for large
number of active rules, the amount of memory remains
acceptable for embedded systems (less than 2400KB).
We find that SRE is suitable for embedded as well as resourceful
devices. Its code size is also manageable, as Java version is
≈1.3MB and C version is less than 1MB in size.
Fig. 6. Rules Execution Time
Fig. 8. Memory Usage
8
https://github.com/InnovationSE/SRE/blob/master/TII/SreEvalRule.lua
This paper has been accepted for publication in forthcoming issue of IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. The
content is final but has NOT been proof-read. This is an author copy for personal record only.
E. SRE Deployment in Real Industrial Environments
to write rules but we still need to abstract them with GUI to make
We have also tested SRE to solve our real business use cases. it easier for non-technical users. The solution should consider
The Java version has been tested in an industrial environment to QoE and the human experience when interacting with technology
monitor the production lines within a factory to ensure safety of [37].
VI. CONCLUSION
the production lines. The setup is shown in Fig. 9. The rules in SRE
receive measurements from Zigbee Green Power modules and
This paper presented SRE with control and querying capability
then generate alerts, as blinking light, to alert the on-site on top of existing gateways and data logger devices. The Rules
supervisor. The same SRE version is used in a different facility Engine allows different stakeholders to define rules to control and
where a rule is used to detect the tripping event of a circuit monitor devices. Rule-based solutions bring flexibility to quickly
breaker, read specific values from Modbus registers to determine address business requirements without going through the
the causes of the trip and send this data to the remote monitoring complex implementation and installation process. There is a
platform without any human intervention. In another complete life-cycle management of rules as well as corporation
deployment, SRE Java C version resides in a gateway and uses between rules to enable reuse of functionality. The Semantic
rules to, for example, stop sending the data to the cloud when Engine, SQenIoT, allows access to device data using semantic tags.
the luminosity is zero in order to save cost on the 3G network.
The cooperation between the Rules Engine and SQenIoT is what
In terms of added value of SRE, we can envision an application sets this work apart from existing solutions. The rules are defined
that receives data from multiple SRE deployments and then using semantic tags so that they do refer to higher level concepts
displays comparison/bench-marking of energy, occupancy, and not specific devices. The advantage of this is that rules remain
efficiency, and other dimensions. Different widgets can be used valid even when devices are added or removed.
to send semantic queries to SRE and display the results in an
SRE solved simple yet important requirements related to cost
appropriate manner.
and time savings. Instead of implementing custom functions for
Similarly, an analytics application can get aggregated data from each of our end clients by going through typical software
SRE by sending a semantic query. As part of machine learning development process, SRE makes our industrial gateways
process, semantic queries can be used to subscribe to the data customizable with little effort using rules and semantic queries.
from machines for a specific duration which can be then fed into The work described in this paper has been implemented, tested
an algorithm to learn about the machine failures with real values. and validated in real industrial settings. Several work items have
In fact, with SRE rules, a complex machine learning strategy can been identified and will be used to extend this work.
be developed and executed at the Edge without depending on
input from the Cloud.
We can see that the rule-based approach in SRE makes it much
easier to handle different requirements. There is no need to code
and deploy a new solution where requirements change. The rules
can be easily managed thanks to our rule life-cycle approach. Our
experience shows that SRE is a general purpose solution to solve
different problems in different domains.
V. LESSONS LEARNED AND FUTURE WORK
During this work we learned several lessons and plan to tackle
them in future works. The first lesson is that when there are
multiple rules executing concurrently, conflicts may occur.
However, detecting conflicts in rules and solving them is not
trivial. The works such as [12], and BZR language [31] and [32] can
be explored to find a solution. The second lesson is about using
properly aligned ontologies, for the meta-data of the devices.
This raises ontology governance issue in large enterprises. There
is a tremendous scope of using ontologies in industrial context.
We can find early initiative like SAREF ontology [33] and its
proposed extensions, Haystack Tagging Ontology [34] and IEEE
Ontologies for Robotics and Automation [35].
The third lesson is about efficient inference engines usable on
gateways and edge devices. Localized inference capabilities can
be particularly useful for emerging trends like local analytics [36].
The fourth and final lesson is the need for a userfriendly way to
write rules and semantic queries. Developers can easily use Lua
Fig. 9. SRE Deployment in Industrial Environment
This paper has been accepted for publication in forthcoming issue of IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS. The
content is final but has NOT been proof-read. This is an author copy for personal record only.
[24] A. Gyrard and M. Serrano, “A unified semantic engine for internet of things
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
L. Atzori, A. Iera, and G. Morabito, “The internet of things: A survey,”
Comput. Netw., vol. 54, no. 15, pp. 2787–2805, Oct. 2010.
L. D. Xu, W. He, and S. Li, “Internet of things in industries: A survey,” IEEE
Transactions on Industrial Informatics, vol. 10, no. 4, pp. 2233– 2243, Nov
2014.
J. Ye et al., “Semantic web technologies in pervasive computing: A survey
and research roadmap,” Pervasive and Mobile Computing, 2015.
I. Khan et al., “A data annotation architecture for semantic applications in
virtualized wireless sensor networks,” in Integrated Network Management
(IM), 2015 IFIP/IEEE International Symposium on, 2015.
E. Kharlamov et al., “How semantic technologies can enhance data access at
siemens energy,” in The Semantic Web–ISWC, 2014.
C. E. Kaed, Y. Denneulin, and F. G. Ottogalli, “Dynamic service adaptation for
plug and play device interoperability,” in Proceedings of the 7th
International Conference on Network and Services Management.
International Federation for Information Processing, 2011.
S. N. Han, G. M. Lee, and N. Crespi, “Semantic context-aware service
composition for building automation system,” IEEE Transactions on
Industrial Informatics, vol. 10, no. 1, pp. 752–761, Feb 2014.
R. Want, B. N. Schilit, and S. Jenson, “Enabling the internet of things,”
Computer, vol. 48, no. 1, pp. 28–35, Jan 2015.
C. E. Kaed, I. Khan, H. Hossayni, and P. Nappey, “Sqenlot: Semantic query
engine for industrial internet-of-things gateways,” in 2016 IEEE 3rd World
Forum on Internet of Things (WF-IoT), Dec 2016, pp. 204– 209.
P. Nappey, C. E. Kaed, A. W. Colombo, J. Eliasson, A. Kruglyak, R. Kyusakov,
C. Hbner, T. Bangemann, and O. Carlsson, “Migration of a legacy plant
lubrication system to soa,” in IECON 2013 - 39th Annual Conference of the
IEEE Industrial Electronics Society, 2013.
M. Louvel and F. Pacull, “Linc: A compact yet powerful coordination
environment,” in International Conference on Coordination Languages and
Models, 2014.
Y. Sun, T. Y. Wu, G. Zhao, and M. Guizani, “Efficient rule engine for smart
building systems,” IEEE Transactions on Computers, June 2015.
J. Barbosa, F. Dillenburg, G. Lermen, A. Garzao, C. Costa, and J. Rosa,˜
“Towards a programming model for context-aware applications,” Computer
Languages, Systems & Structures, 2012.
E. Kuhn, S. Craß, G. Joskowicz, A. Marek, and T. Scheller,¨
Peer-Based
Programming Model for Coordination Patterns, 2013.
C. Julien and G. C. Roman, “Egospaces: facilitating rapid development of
context-aware mobile applications,” Transactions on Software Engineering,
2006.
M. Viroli, D. Pianini, and J. Beal, “Linda in space-time: An adaptive
coordination model for mobile ad-hoc environments,” in Proceedings of the
14th International Conference on Coordination Models and Languages, ser.
COORDINATION’12, 2012.
Y. Zheng, A. T. S. Chan, and G. Ngai, “Mcl: A mobigate coordination language
for highly adaptive and reconfigurable mobile middleware: Experiences with
auto-adaptive and reconfigurable systems,” Softw. Pract. Exper., 2006.
G. Cabri, L. Leonardi, and F. Zambonelli, “Mars: a programmable
coordination architecture for mobile agents,” IEEE Internet Computing, vol.
4, no. 4, pp. 26–35, Jul 2000.
A. L. Murphy, G. P. Picco, and G.-C. Roman, “Lime: A coordination model and
middleware supporting mobility of hosts and agents,” ACM Trans. Softw.
Eng. Methodol., vol. 15, no. 3, pp. 279–328, Jul. 2006.
M. Mamei and F. Zambonelli, “Programming pervasive and mobile
computing applications: The tota approach,” ACM Trans. Softw. Eng.
Methodol., vol. 18, no. 4, pp. 15:1–15:56, Jul. 2009.
M. Viroli, M. Casadei, S. Montagna, and F. Zambonelli, “Spatial coordination
of pervasive services through chemical-inspired tuple spaces,” ACM Trans.
Auton. Adapt. Syst., 2011.
M. Proctor, Drools: A Rule Engine for Complex Event Processing, 2012.
S. Kuboi, K. Baba, S. Takano, and K. Murakami, “An evaluation of a complex
event processing engine,” in 2014 IIAI 3rd International Conference on
Advanced Applied Informatics, Aug 2014, pp. 190–193.
[25]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
and smart cities: From sensor data to end-users applications,” in 2015 IEEE
International Conference on Data Science and Data Intensive Systems, 2015.
T. Riedel et al., “Using web service gateways and code generation for
sustainable iot system development,” in Internet of Things (IOT), 2010.
F. Jammes and H. Smit, “Service-oriented paradigms in industrial
automation,” Industrial informatics, IEEE transactions on, 2005.
K. Kotis et al., “Semantic interoperability on the web of things: the semantic
smart gateway framework,” in Complex, Intelligent and Software Intensive
Systems, Sixth International Conference on, 2012.
K. Kotis and A. Katasonov, “Semantic interoperability on the internet of
things: The semantic smart gateway framework,” Int. J. Distrib. Syst.
Technol., 2013.
R. Ierusalimschy, L. H. De Figueiredo, and W. Celes Filho, “Lua-an extensible
extension language,” Softw., Pract. Exper., 1996.
N. Grcevski, A. Kielstra, K. Stoodley, M. G. Stoodley, and V. Sundaresan, “Java
just-in-time compiler and virtual machine improvements for server and
middleware applications.” in 3rd Conference on Virtual Machine Research
And Technology Symposium, 2004.
M. Zhao, G. Privat, E. Rutten, and H. Alla, “Discrete control for the internet
of things and smart environments,” in Presented as part of the 8th
International Workshop on Feedback Computing, 2013.
G. Delaval, E. Rutten, and H. Marchand, “Integrating discrete controller´
synthesis into a reactive programming language compiler,” Discrete Event
Dynamic Systems, vol. 23, no. 4, pp. 385–418, 2013.
L. Daniele, F. den Hartog, and J. Roes, “Created in close interaction with the
industry: the smart appliances reference (saref) ontology,” in International
Workshop Formal Ontologies Meet Industries, 2015.
V. Charpenay, S. Kbisch, D. Anicic, and H. Kosch, “An ontology design pattern
for iot device tagging systems,” in 2015 5th International Conference on the
Internet of Things (IOT), Oct 2015, pp. 138–145.
C. Schlenoff, E. Prestes, R. Madhavan, P. Goncalves, H. Li, S. Balakirsky, T.
Kramer, and E. Miguelanez, “An ieee standard ontology for robotics and
automation,” in 2012 IEEE/RSJ International Conference on Intelligent
Robots and Systems. IEEE, 2012, pp. 1337–1342.
M. Satyanarayanan, P. Simoens, Y. Xiao, P. Pillai, Z. Chen, K. Ha, W. Hu, and
B. Amos, “Edge analytics in the internet of things,” IEEE Pervasive
Computing, vol. 14, no. 2, pp. 24–31, Apr 2015.
K. U. R. Laghari, K. Connelly, and N. Crespi, “Toward total quality of
experience: A qoe model in a communication ecosystem,” IEEE
Communications Magazine, vol. 50, no. 4, pp. 58–65, April 2012.
| 2cs.AI
|
1
Full-Duplex Relaying with Improper Gaussian
Signaling over Nakagami-m Fading Channels
arXiv:1704.06640v2 [cs.IT] 26 Sep 2017
Mohamed Gaafar, Student Member, IEEE, Mohammad Galal Khafagy, Senior Member, IEEE, Osama Amin,
Senior Member, IEEE, Rafael F. Schaefer, Senior Member, IEEE and Mohamed-Slim Alouini, Fellow, IEEE
Abstract—We study the potential employment of improper
Gaussian signaling (IGS) in full-duplex relaying (FDR) with
non-negligible residual self-interference (RSI) under Nakagamim fading. IGS is recently shown to outperform traditional proper
Gaussian signaling (PGS) in several interference-limited settings.
In this work, IGS is employed as an attempt to alleviate RSI.
We use two performance metrics, namely, the outage probability
and the ergodic rate. First, we provide upper and lower bounds
for the system performance in terms of the relay transmit power
and circularity coefficient, a measure of the signal impropriety.
Then, we numerically optimize the relay signal parameters based
only on the channel statistics to improve the system performance.
Based on the analysis, IGS allows FDR to operate even with high
RSI. The results show that IGS can leverage higher power budgets
to enhance the performance, meanwhile it relieves RSI impact
via tuning the signal impropriety. Interestingly, one-dimensional
optimization of the circularity coefficient, with maximum relay
power, offers a similar performance as the joint optimization,
which reduces the optimization complexity. From a throughput
standpoint, it is shown that IGS-FDR can outperform not only
PGS-FDR, but also half-duplex relaying with/without maximum
ratio combining over certain regions of the target source rate.
Keywords—improper Gaussian signaling, asymmetric complex
signaling, interference mitigation, full-duplex relay, residual selfinterference, outage probability, ergodic rate, coordinate descent.
I. I NTRODUCTION
Contrary to a long-held acceptance that radio front-ends
cannot simultaneously transmit and receive, a truly promising
potential for full-duplex communications has been shown by
recent hardware developments [1], [2]. Indeed, by multiplexing
inbound and outbound traffic over the same channel resource, a
full-duplex radio can recover the spectral efficiency loss known
to be encountered by its half-duplex counterpart. Performance
merits of full-duplex radio have been recently investigated in
different communication settings, including full-duplex bidirectional communication, full-duplex base stations, and fullduplex relaying (FDR) [3], with the latter being the focus of
M. Gaafar and R. F. Schaefer are with the Information Theory and Applications Chair, Technische Universität Berlin, Germany. Email: {mohamed.gaafar,
rafael.schaefer}@tu-berlin.de.
O. Amin and M.-S. Alouini are with the Computer, Electrical, and
Mathematical Sciences and Engineering (CEMSE) Divison, King Abdullah
University of Science and Technology (KAUST), Thuwal, Makkah Province,
Saudi Arabia. Email: {osama.amin, slim.alouini}@kaust.edu.sa.
M. G. Khafagy was with the CEMSE Division, KAUST, Thuwal, Makkah
Province, Saudi Arabia. Currently, he is with the Computer Science and
Engineering Department, College of Engineering, Qatar University, Doha,
Qatar. Email: [email protected].
this work. These merits have qualified full-duplex communication to be considered as a candidate technology for future
fifth generation (5G) wireless networks [4].
FDR allows a relay node to listen to an information source
and simultaneously forward to its intended destination. Theoretically, this simultaneous transmission/reception doubles the
spectral efficiency in the relay channel. However, in practice,
this comes at the expense of a self-interference level introduced
at the receiver of the relay node from its own transmitter. Even
with the application of advanced self-interference isolation and
cancellation techniques, a level of residual self-interference
(RSI) persists. Such a persistent RSI link and the means
to mitigate it represent the main challenge in full-duplex
communications, especially with the fact that its adverse effect
can typically be an increasing function of the relay power.
Therefore, increasing the relay power no longer guarantees an
enhanced end-to-end performance. For instance, by increasing
the relay power in a fixed-rate transmission scheme, the relay
may forward more reliably to the destination in the second
hop. However, it also increases the RSI level which negatively
affects the reliability in the first hop. Hence, higher relay power
budgets cannot be always utilized beyond a certain threshold.
Consequently, employing interference mitigation strategies in
FDR is crucial to attain a satisfactory performance of fullduplex transmissions.
Improper Gaussian signaling (IGS) has been first introduced
in [5], where higher degrees of freedom for the 3-user singleinput single-output (SISO) interference channel (IC) were
shown to be achievable. This comes in contrary to other
communication settings with interference-free channels where
proper Gaussian signaling (PGS) is the optimal choice. PGS
assumes the zero-mean complex Gaussian transmit signal to
be statistically circularly symmetric with uncorrelated real and
imaginary components. On the other hand, IGS is a class of
signals where circularity and uncorrelatedness conditions can
be relaxed [6]. The results in [5] motivate the need to further
study the potential gains of IGS in communication scenarios
where interference imposes a noticeable limitation.
IGS has been recently adopted to improve the performance of different interference-limited communication settings, namely, SISO-IC [7], [8], multiple-input single-output
(MISO)-IC [9], multiple-input multiple-output (MIMO)-IC
[10], [11], Z-IC [12]–[14], MIMO X-channel [15], interference
broadcast channels [16], cognitive radio channels using underlay [17]–[21] and overlay [22] communication paradigms,
alternate (two-path, virtual full-duplex) relaying [23], symbol
error rate reduction [24] and asymmetric hardware distor-
2
tions [25].
The potential gains of IGS have been also recently studied
in [26] for the MIMO relay channel when a partial decodeand-forward strategy is adopted. In such a relaying strategy,
the relay only decodes a part of the message, while the rest of
the message is treated as an additional interference term. It was
shown in [26] that PGS is optimal within the class of Gaussian
signals. However, the work in [26] assumed an ideal FDR,
where the self-interference imposed by the relay’s transmitter
on its own receiver is perfectly canceled. In [27], a simplified
model for FDR with IGS was considered, where the transmit
signal only occupies the real dimension of the two-dimensional
complex signal space. Also, all the communicating nodes in
[27] are assumed to perfectly share the instantaneous Rayleighfading channel coefficients. For such a simplified model, it was
shown that IGS was able to effectively improve the achievable
rate by eliminating the RSI via its alignment in only one
orthogonal signal space dimension, and decoding the desired
signal from the other. The more general problem, however,
remains more challenging with further interesting aspects to
investigate, where complex transmit signals are employed and
only channel statistics are made available at the transmitter
side.
In this work1 , we study the potential performance merits of
employing IGS in FDR systems with non-negligible RSI over
Nakagami-m fading channels. The main contributions of this
paper can be summarized as follows:
• We first derive the exact end-to-end outage probability
of the FDR system adopting IGS at the relay over
Nakagami-m fading as a function of the relay’s transmit
power and circularity coefficient in an integral form.
Further, we derive a lower bound of the end-to-end
outage probability in closed form. Moreover, we provide
simpler forms of the derived expressions for Rayleigh
fading and derive an upper bound for the end-to-end
outage probability based on some convexity properties
that are also proven herein. Based on this upper bound,
we present an asymptotic analysis that shows the benefits
which can be reaped from IGS in FDR systems.
• We derive an integral form of the exact end-to-end
ergodic rate of the IGS-FDR system as a function of the
relay’s transmit power and circularity coefficient over
Nakagami-m fading. In addition, we derive an upper
bound for the end-to-end ergodic rate. We also present
a lower bound for the end-to-end ergodic rate over
Rayleigh fading.
• We prove unimodality properties of the presented endto-end outage probability upper bound in the relay’s
transmit power and circularity coefficient over Rayleigh
fading, which allow for efficient numerical optimization
1 While this work was in progress, preliminary results have been accepted
and presented in IEEE ICC’16 [28]. In this work, we consider a more
generalized framework under Nakagami-m fading, which includes the work in
[28] as a special case. We present herein the derived expressions for another
performance metric, the ergodic rate, to evaluate the performance of IGS in
FDR. Furthermore, we provide more detailed analysis and insights addressing
the outage performance. Finally, more numerical results are presented to verify
the mathematical analysis.
using standard tools. Also, we design the relay’s transmit signal characteristics, represented in the power and
circularity coefficient, by a coordinate descent algorithm
based on the outage upper bound over Rayleigh fading.
• Finally, with the aid of numerical results, we first validate the derived bounds for the outage probability and
ergodic rate. Next, we discuss the benefits that can be
reaped by employing IGS in FDR in comparison to PGS,
and also relative to half-duplex relaying (HDR), under
different system and fading parameters. We show that
IGS-FDR is not only able to outperform PGS-FDR in
terms of the end-to-end throughput, but also that of HDR
even with maximum ratio combining (MRC) over certain
ranges of the targeted source rate. Finally, we show that
one may not seriously compromise the benefits attained
via the joint tuning of the power and circularity, and still
attain a close end-to-end performance merits through a
simple one-dimensional tuning of the circularity coefficient.
The rest of the paper is organized as follows. In Section II, we
describe the adopted FDR system model. Section III derives
the outage probability of the FDR system over Nakagami-m
fading, while Section IV deals with the ergodic rate of the
system. The design of the signal characteristics for the relay
transmissions based on the outage probability over Rayleigh
fading is presented in Section V. We validate the performance
of IGS in FDR systems in Section VI through numerical
simulations. Finally, we conclude the paper in Section VII.
Notation and Special Functions: Throughout the rest of
the paper, we use |·| to denote the absolute value operation.
P{A} denotes the probability of occurrence of an event A. The
operator E{·} is used to denote the statistical expectation, with
the mean of a random variable (RV) X defined as X̄ = E{X}.
n
k is the binomial coefficient and min (x, y) is the minimum
of the two quantities x and y. Throughout the analysis,
R ∞ a−1we−tuse
the following special functions [29]. Γ(a) =
e dt
R ∞0 t
denotes the gamma function and Γ(a, x) = x ta−1 e−t dt is
R ∞ −xt
the upper incomplete gamma function. En (x) = 1 e tn dt
R ∞ a−1
b−a−1 −zt
1
and U (a, b, z) = Γ(a)
t
(t + 1)
e dt , a, z > 0
0
are the exponential integral and the Tricomi’s confluent hypergeometric functions, respectively. When n = 1, we use
E1 (x) = −Ei(−x),
R ∞ −t x > 0, where the exponential integral
Ei(x) = − −x e t dt. Also, we use Ξn (x) = ex En (x) to
combine the exponential function and integral.
II.
S YSTEM M ODEL
We consider the communication setting depicted in Fig. 1,
where a source (S) intends to communicate with a distant
destination (D). The direct S − D link is assumed of a relatively weak gain due to path loss and shadowing effects.
Accordingly, a full-duplex relay (R) is utilized to assist the
end-to-end communication and extend the coverage. FDR
can offer higher spectral efficiency when compared to HDR.
However, FDR in practice suffers from an RSI level which
imposes an additional communication challenge. In addition,
3
Desired Signal
Interference
hsr
S
B. Signal Model
hrr
hrd
R
D
hsd
Fig. 1: A full-duplex cooperative setting in coverage extension
scenarios.
the received signal component via the S − D link is assumed
to be weak and hence, it is considered as interference at
the destination for simpler decoding purposes as commonly
assumed in the literature [30], [31]. Thus, the FDR system
under consideration suffers from two interference sources; the
RSI at the relay, and the direct S − D link signal received at
the destination.
A. Channel Model
The fading coefficient of the i − j link is denoted by hij ,
for i ∈ {s, r} and j ∈ {r, d}, where s, r and d refer to the
source, relay, and destination nodes, respectively. Moreover,
the i − j link gain is denoted by gij = |hij |2 . All channels are
assumed to follow a block fading model, where hij remains
constant over one block, and varies independently from one
block to another following a Nakagami-m fading distribution
with a shaping parameter mij and average power E |hij |2 =
πij . Accordingly, the channel gain gij is a gamma RV with
π
a shaping parameter2 mij and a scale parameter θij = mijij .
Furthermore, it should be noted that the probability density
function (PDF) of gij is defined as
−
x
xmij −1 e θij
fgij (x; mij , θij ) =
m , x ≥ 0·
Γ(mij )θij ij
(1)
In the special case mij = 1, Rayleigh fading is obtained. All
channel fading gains are assumed to be mutually independent.
The relay operates in a full-duplex mode where simultaneous listening/forwarding is allowed with an introduced level of
loopback interference. The link coefficient hrr 3 is assumed to
represent the RSI after undergoing all possible isolation and
cancellation techniques [30], [31], [35]. The source and the
relay powers are denoted by Ps and Pr , respectively, where
both are restricted to a maximum allowable value of Pmax .
Also, nr and nd denote the circularly-symmetric complex
additive white Gaussian noise (AWGN) components at the
relay and the destination, with variance σr2 and σd2 , respectively.
Without loss of generality, we assume that σr2 = σd2 = 1.
2 In
this paper, we assume only integer shape parameter. Typical analysis
involving wireless communications over Nakagami-m channels uses values of
m = 1 up to m = 4 to keep the fading effects [32].
3 Depending on the communication setting, transmit power ranges, and the
employed RSI cancellation techniques, the average RSI link gain is typically
assumed to range from 3 to 10/15 dB above the noise floor [2, Section IV.B],
[33, Chapter 2], [34].
The transmit signals at the source and the relay at time t are
denoted by xs [t] and xr [t], respectively. PGS is adopted at the
source 4 which transmits with its power budget, Ps = Pmax .
On the other hand, with the availability of statistical transmit
channel state information (CSI) at the relay, zero-mean IGS is
adopted at the relay in order to mitigate the non-negligible RSI
at its receiver. The degree of impropriety of xr [t] is measured
based on the following definitions.
Definition 1. [36] The variance and pseudo-variance of the
2
relay’s transmit signal, xr [t], are given by σx2 = E{|xr | } and
2
2
σ̃x = E{xr }, respectively.
Definition 2. A signal is called proper if it has a zero pseudovariance σ̃x2 , while an improper signal has a non-zero σ̃x2 .
Definition 3. [17] A circularity coefficient is a measure of
the degree of impropriety of the signal xr [t], which is given as
Cx = σ̃x2 /σx2 .
Following from Definitions 2 and 3, the circularity coefficient
satisfies 0 ≤ Cx ≤ 1. In particular, Cx = 0 implies a proper
signal, while Cx = 1 implies a maximally improper signal.
The received signals at the relay and the destination at time
t are given, respectively, by
p
p
yr [t] =
Ps hsr xs [t] + Pr hrr xr [t] + nr [t],
(2)
p
p
yd [t] =
Pr hrd xr [t] + Ps hsd xs [t] + nd [t]·
(3)
The relay is assumed to adopt a decode-and-forward relaying
strategy, where it does not transmit any message of its own,
but forwards the regenerated source message after decoding.
Due to the source and relay asynchronous transmissions, the
signal transmitted by the relay (source) is considered as an
additional noise term at the relay (destination) in the decoding
stage as commonly treated in the related literature [30], [31].
C. Achievable Rates with IGS
From the adopted signal model in (2) and (3), each transmit
signal, i.e., from the source and the relay transmitter, is
considered as a desired signal at one receiver while treated
as interference at the other. Hence, the rate expressions for
the first and second hops have the same form of those of a
two-user IC.
Lemma 1. [37] The achievable rate of a single link that is
subjected to interference-plus-noise z when x is transmitted
and observed as y is expressed as
2!
σy4 − σ̃y2
1
R = log2
·
(4)
2
2
σz4 − |σ̃z2 |
As a result of using IGS, while treating the interference
as a Gaussian noise, the achievable rates of the FDR system
4 For the ease of exposition, we use PGS at the source. This assumption
can be further justified since the use of IGS at the source calls for a joint
source/relay improper signal optimization. Although an IGS source is expected
to offer further gains, such a joint optimization turns out to be mathematically
involved and intractable.
4
are analyzed based on Lemma 1. First, the achievable rate
supported by the S − R link can be expressed as
!
1 − Cy2r
1
Ps gsr
, (5)
Rsr (Pr , Cx ) = log2 1 +
+ log2
Pr grr + 1
2
1 − CI2r
where Cyr and CIr are the circularity coefficients of the received
and interference-plus-noise signals at the relay, respectively.
Hence, (5) can be simplified as
!
2
2
1
(Ps gsr + Pr grr + 1) −(Pr grr Cx )
Rsr (Pr , Cx ) = log2
· (6)
2
2
2
(Pr grr + 1) −(Pr grr Cx )
Similarly, the achievable rate supported by the R − D link is
given by
!
2
2
1
(Pr grd +Ps gsd + 1) −(Pr grd Cx )
Rrd (Pr , Cx ) = log2
· (7)
2
2
(Ps gsd + 1)
One can notice that if Cx = 0, we obtain the well known
achievable rates of PGS. From (6) and (7), by adopting IGS
at the relay transmitter, the rate of the S − R hop improves
while the rate of the R − D hop deteriorates which creates a
trade-off that can be optimized to yield a better performance
than PGS.
III.
O UTAGE P ERFORMANCE
In this section, we derive the end-to-end outage probability
of the FDR system depicted in Fig. 1 under the assumption of
Nakagami-m fading while the relay transmits improper signals.
Moreover, we obtain simpler forms for the Rayleigh fading
case. The end-to-end outage probability is given by
PE−E
=
1 − P sr P rd ,
(8)
where Psr and Prd denote the outage probability in the S − R
and the R − D links, respectively, while P ij = 1 − Pij . In
what follows, we derive the outage probability expressions in
the individual links, i.e., Psr and Prd .
A. Outage Probability of S − R Link
Let r (bits/sec/Hz) denote the target rate of the S − R link,
then its outage probability is defined as
Psr (Pr , Cx ) = P {Rsr (Pr , Cx ) < r} ·
(9)
The outage probability of the S − R can be calculated from
Lemma 2.
Lemma 2. The exact outage probability of the S − R link with
a target rate r is given by
1
Psr (Pr , Cx ) = 1 −
mrr
Γ (mrr ) Γ (msr ) θrr
Z∞
x
Pr xCx
(Pr x + 1)
Ψr
e− θrr dx,
× xmrr −1 Γ msr ,
Ps θsr
Pr x + 1
0
where Ψr (x) =
p
(10)
1 + γ (1 − x2 ) − 1 and γ = 22r − 1.
Proof: By substituting (6) into (9), we get
n
2
Psr (Pr , Cx ) = P Ps2 gsr
+ 2Ps gsr (Pr grr + 1)
o
2
2
− γ (Pr grr + 1) − (Pr grr Cx ) < 0 ·
(11)
The expression inside the probability expression is a quadratic
function in gsr . Hence, the conditional outage probability given
grr is equivalent to integrating the gamma PDF of gsr over the
region in which the quadratic function is less than 0 which
gives
◦
Zgsr
x
1
Psr (Pr , Cx |grr ) =
xmsr −1 e− θsr dx
msr
Γ (msr ) θsr
0
◦
gsr
Γ msr , θsr
=1−
,
(12)
Γ (msr )
Cx
◦
where gsr
= (Pr gPrrs +1) Ψr PPrrggrrrr+1
represents the positive
root of the inequality inside the probability sign. Therefore,
by averaging over the statistics of grr in (12), we obtain
Pr grr Cx
r grr +1)
Γ msr , (PP
Ψ
r
θ
P
g
+1
s sr
r rr
Psr (Pr , Cx ) = 1 − Egrr
,
Γ (msr )
(13)
which directly yields the result in (10).
Unfortunately, there is no closed form expression for the
integral in Lemma 2. In the following, we derive a lower bound
on the S − R outage probability for Nakagami-m fading, in
addition to an upper bound for the special Rayleigh fading
case.
1) Lower Bound: A lower bound on Psr is provided in the
following lemma.
Lemma 3. The S − R link outage probability can be lowerbounded as
Ψr (Cx )
e− Ps θsr
Psr ≥ 1 −
mrr
Γ (mrr ) θrr
m
Ψr (Cx )
k
mX
m
sr −1 X
P
Γ
(k
+
m
)
rr
r
Ps θsr
m
×
k+mrr
k
Pr Ψr (Cx )
m=0 k=0
Γ (m + 1)
+ θ1rr
Ps θsr
∆
LB
= Psr
(Pr , Cx ) ·
(14)
Proof: The S − R outage probability expression in
rx
Lemma 2 can be lower-bounded by replacing PP
Cx
r x+1
by Cx . This can be easily seen as Ψr (x) and Γ (., x) are
monotonically decreasing in x. Then, we use the series
expansion of the upper incomplete gamma function [38,
8.352-2] as
5
Then, the outage probability of the R − D link can be obtained
from the following result.
(Pr x + 1) Ψr (Cx )
Γ msr ,
Ps θsr
= Γ (msr ) e−
(Pr x+1)Ψr (Cx )
Ps θsr
mX
sr −1
(Pr x+1)Ψr (Cx )
Ps θsr
m=0
Γ (m + 1)
Theorem 1. The R − D link outage probability with a target
rate of r is expressed as
m
· (15)
Hence, the integral can be rewritten as
Ψ (C )
r x
mX
rd −1
1
e− Ps θsr
(Pr , Cx ) = 1 −
mrr
Γ (mrr ) θrr m=0 Γ (m + 1)
mZ ∞
Pr Ψr (Cx )
1
Ψr (Cx )
m
xmrr −1 (Pr x + 1) e−( Ps θsr + θrr )x dx·
×
Ps θsr
0
(16)
m
P
m
k
m
By using the binomial theorem (Pr x + 1) =
k (Pr x) ,
LB
Psr
k=0
the result is obtained by solving the integral in (16) by [38,
3.381-4].
2) Upper Bound over Rayleigh Fading: In the case of
Rayleigh fading, the S − R hop outage probability is simplified
as
o
n (Pr grr +1)
Pr grr Cx
P
(P , C ) = 1 − E
e− Ps πsr Ψr ( Pr grr +1 ) · (17)
sr,RF
r
x
grr
Unfortunately, there is no closed-form expression for this
expectation except at Cx = 0, which gives the known PGS
outage probability given as
− Psηπsr
Ps πsr e
,
(18)
Ps πsr + Pr πrr η
where η = 2r − 1. Otherwise, we resort to obtain an upper
bound as follows.
Psr,RF (Pr , 0) = 1 −
−
Ψr (Cx )
e Pr θrd (1−Cx2 )
Prd (Pr , Cx ) = 1 −
msd
Γ (msd ) θsd
m
Ψr (Cx )
k
mX
m
rd −1 X
P
Γ
(k
+
m
)
2
sd
s
Pr θrd (1−Cx )
m
×
k+msd ·
k
1
r (Cx )
m=0 k=0
Γ (m + 1) PrPθsrdΨ(1−C
2) + θ
sd
x
(21)
Proof: The proof follows similar steps as in Lemma 2
and the S − R outage probability lower bound over Rayleigh
fading in Lemma 3.
For Rayleigh fading, the outage probability of the R − D
link can be obtained from the following corollary.
Corollary 1. The outage probability of the R − D link, over
Rayleigh fading, is expressed as
Prd,RF (Pr , Cx ) = 1 −
e
−
Ψr (Cx )
2)
Pr πrd (1−Cx
(Cx )
Ps πsd Pr πΨrdr (1−C
2) + 1
(22)
·
x
From Corollary 1, it can be noticed that, for the PGS case,
i.e., Cx = 0, Eq. (22) yields the known expression for PGS
[30] as
−
η
Pr πrd e Pr πrd
·
(23)
Pr πrd + Ps πsd η
Also, for the maximally improper case, i.e., Cx = 1, it yields
Prd,RF (Pr , 0) = 1 −
− 2Prγπ
Prd,RF (Pr , 1) = lim Prd,RF (Pr , Cx ) = 1 −
Cx →1
Proposition 1. The exponential term inside the expectation
operator in (17) is convex in grr .
e
γPs πsd
2Pr πrd
rd
·
+1
(24)
Proof: The proof is given in Appendix VII.
Therefore, an upper bound on the S − R link outage probability is given as follows.
C. End-to-End Outage Probability
The exact end-to-end outage performance can be obtained
from the following theorem.
Lemma 4. The S − R link outage probability, over Rayleigh
fading, is upper-bounded by
Theorem 2. The exact end-to-end outage probability with a
target rate r can be numerically calculated from
Psr,RF (Pr , Cx ) ≤ 1 − e−
where α =
Pr πrr +1
Ps πsr Ψr (αCx )
∆
UB
= Psr,RF
(Pr , Cx ) ,
(19)
Pr πrr
Pr πrr +1 .
Proof: First, we follow Proposition 1, then by applying
the Jensen’s inequality to the expectation in (17), we obtain
the given outage probability upper bound of the S − R link.
−
0
B. Outage Probability of R − D Link
Similarly, let r (bits/sec/Hz) denote the target rate of the
R − D link, then its outage probability is defined as
Prd (Pr , Cx ) = P {Rrd (Pr , Cx ) < r} ·
(20)
Ψr (Cx )
e Pr θrd (1−Cx2 )
PE−E = 1 −
msd mrr
Γ (msd )Γ (mrr ) Γ (msr ) θsd
θrr
m
Ψr (Cx )
k
mX
m
rd −1 X
P
Γ
(k
+
m
)
2
sd
s
Pr θrd (1−Cx )
m
×
k+msd
k
Ps Ψr (Cx )
m=0 k=0
Γ (m + 1) Pr θrd (1−C 2 ) + θ1sd
x
Z∞
x
(Pr x + 1)
Pr xCx
mrr −1
× x
Γ msr ,
Ψr
e− θrr dx·
Ps θsr
Pr x + 1
(25)
Proof: Based on the derived expressions of the outage
probability for S − R and R − D links from Lemma 2 and
Theorem 1, respectively, and by direct substitution in (8), we
obtain the result.
6
Unfortunately, there is no closed form solution for the
end-to-end outage probability in Theorem 2. Therefore, we
resort to obtain an expression for the lower bound of the
end-to-end outage probability in the Nakagami-m fading, in
addition to an upper bound expression for the special case of
Rayleigh fading. These bounds have a significantly reduced
computational complexity and an enhanced numerical stability
than the exact expression in (25).
1) Lower Bound: The following theorem provides a lower
bound for the end-to-end outage probability.
Theorem 3. The end-to-end outage probability can be lowerbounded as
PE−E (Pr , Cx ) ≥ 1 −
e
−Ψr (Cx )
1
Ps θsr
+
1
2)
Pr θrd (1−Cx
0
Prk−m Γ(k+mrr )Γ(k0 +msd )Ψrm+m (Cx )
0
0
m θ m0 (1−C 2 )m
Psm−k Γ(m+1)Γ(m0 +1)θsr
x
rd
Pr Ψr (Cx )
Ps θsr
+
1
θrr
k+mrr
Ps Ψr (Cx )
2)
Pr θrd (1−Cx
+
1
θsd
k0 +msd
∆
LB
= PE−E
(Pr , Cx ) ·
(26)
Proof: The result follows directly from (8), Lemma 3 and
Theorem 1.
Corollary 2. The lower bound in Theorem 3 at Cx = 0 reduces
to the exact expression of the end-to-end outage probability
for the PGS case. Specifically, for the PGS case over Rayleigh
fading, we have the exact expression for the end-to-end outage
probability as
−η
1
+
1
Ps πsr
Pr πrd
Ps Pr πsr πrd e
PE−E,RF (Pr , 0) = 1 −
·
(Ps πsr + Pr πrr η) (Pr πrd + Ps πsd η)
(27)
Proof: For Cx = 0, the S − R outage probability lower
bound in Lemma 3 reduces to the exact expression and then
the result follows directly.
2) Upper Bound over Rayleigh Fading: From a system
design prospective, it is typically more beneficial to have an
upper bound than a lower bound for the end-to-end outage
probability. Therefore, we state next the upper bound of the
end-to-end outage probability over Rayleigh fading in Theorem
4.
Theorem 4. The end-to-end outage probability over Rayleigh
fading can be upper-bounded by
PE−E,RF (Pr , Cx ) ≤ 1 −
e
−
Ψr (Cx )
Pr πrr +1
2 ) + Ps πsr Ψr (αCx )
Pr πrd (1−Cx
(Cx )
Ps πsd Pr πΨrdr (1−C
2) + 1
x
∆
UB
= PE−E,RF
(Pr , Cx ) ·
Corollary 3. When the relay node uses maximally IGS, the
end-to-end outage probability, over Rayleigh fading, can be
upper-bounded by
−
UB
lim PE−E,RF
=1−
2Pr πrd e
γ
2Pr πrd
+
(Pr πrr +1)
Ψr (αCx )
Ps πsr
2Pr πrd + γPs πsd
Cx →1
·
(29)
Theorem 5. In the limiting case where πrr → ∞ over
Rayleigh fading with a fixed relay transmit power Pr ,
the exact end-to-end outage probability for the PGS case
PE−E,RF (Pr , 0) → 1, while the upper bound for the
end-to-end outage probability for the maximally IGS case
UB
(Pr , 1) → K, where
PE−E,RF
m =0 k=0 k =0
×
Asymptotic Analysis: For maximally IGS, we obtain the
upper bound of the end-to-end outage probability from the
following corollary.
In order to evaluate the performance of the end-to-end
outage probability upper bound with respect to RSI when using
maximally IGS at the relay transmitter, we state the following
theorem.
msd mrr
Γ (msd ) Γ (mrr ) θsd
θrr
0
mX
m X
m
sr −1 mX
rd −1 X
m m0
×
k
k0
0
0
m=0
0
Proof: It follows directly from (8), Lemma 4 and Corollary 1.
(28)
−
γ
+
γ
2Pr πrd e 2Pr πrd Ps πsr
K =1−
·
2Pr πrd + γPs πsd
(30)
Proof: The result is obtained from Corollary 2, 3 and
taking the limit at πrr → ∞.
Interestingly, different from the PGS case, the maximally
IGS introduces immunity against high RSI and achieves less
outage probability with a constant upper bound (30), which
depends on the quality of both S − R and R − D links, in
addition to the target rate.
IV.
E RGODIC R ATE P ERFORMANCE
In this section, we evaluate the ergodic rate performance of
the canonical cooperative setting depicted in Fig. 1 when IGS
is allowed at the relay under Nakagami-m fading. We also
present simplified expressions for Rayleigh fading.
The end-to-end ergodic rate of the FDR system can be
calculated as
Z∞
RE−E = E {min (Rsr , Rrd )} = P {min {Rsr , Rrd } ≥ r} dr·
0
(31)
The last complementary cumulative distribution function
(CDF) integral form for the statistical expectation follows from
the fact that min {Rsr , Rrd } is a non-negative RV. Hence the
exact end-to-end ergodic rate can be numerically computed
from the following result.
Theorem 6. The exact end-to-end ergodic rate can be numer-
7
partial fraction decomposition, I can be expressed as
" 2 Z∞
m+m0
X
1
λi Ψr
e−ΩΨr
dΨr
I=
i
log (2) i=1
Ψ + 1 + (−1) C
ically calculated from
mrd −1 m
X X
m P k Γ (k + m )
sd
s
RE−E =
k
Γ (m + 1)
m=0 k=0
m
Ψr (Cx )
(P x+1)
Pr xCx
Z∞Z∞ xmrr −1 Γ msr , Pr θ
Ψr P
2)
Pr θrd (1−Cx
s sr
r x+1
×
k+m
0
0
−
×e
m
mrr
Γ (msd ) Γ (mrr ) Γ (msr ) θsdsd θrr
Ψr (Cx )
2
Pr θrd 1−Cx
(
Ps Ψr (Cx )
2)
Pr θrd (1−Cx
+
1
θsd
)
rr
sd
(32)
dxdr·
Proof: The result is obtained directly from (31). The
expression of the integrand represents the complement of the
end-to-end outage probability which is obtained in Theorem 2.
Unfortunately, this double integral5 does not have a closed
form solution in the general case. In what follows, we alternatively derive an upper bound on the end-to-end ergodic
rate performance in the Nakagami-m fading. Additionally, we
obtain a lower bound in the Rayleigh fading special case. The
derived bounds are computationally simpler and numerically
more stable than the exact form in (32).
1) Upper Bound: An upper bound for the end-to-end ergodic rate is presented in the following result.
Theorem 7. The end-to-end ergodic rate can be upperbounded as (33)
Proof: We define an upper bound for the end-to-end
ergodic rate of the FDR system as
Z∞
RE−E ≤
∆
LB
1−PE−E
(Pr , Cx |r ) dr = RUB
E−E (Pr , Cx ) , (34)
0
LB
is the lower bound of the outage probability
where PE−E
which is obtained in Theorem 3, which depends on the target
rate r. By substituting (26) in (34), we get the following
integral expression for the upper bound of the end-to-end
ergodic rate.
RUB
E−E
1
=
m
mrr
Γ (msd ) Γ (mrr ) θsdsd θrr
mX
m X
m0
sr −1 mX
rd −1 X
m=0
m0 =0 k=0 k0 =0
Ps
m (θ (1 − C 2 ))
Γ (m + 1) Γ (m0 + 1) θsr
rd
x
−Ψr (Cx )
Z∞
e
×
0
|
Pr Ψr (Cx )
Ps θsr
0
Ps Ψr
2)
Pr θrd (1−Cx
+
1
θsd
#
l dΨr , (36)
2) Lower Bound over Rayleigh Fading: Similar to the
discussion in the previous section, providing a lower bound
for the ergodic rate offers further insights regarding the least
performance merits IGS is able to offer relative to PGS.
A lower bound for the ergodic rate of the FDR system
over Rayleigh fading channels is presented in the following
theorem.
Theorem 8. The end-to-end ergodic rate, over Rayleigh fading, can be lower-bounded by
Pr πrd 1 − Cx2
RE−E,RF ≥
Ps πsd log (2)
i
i
2
Pr πrr 1 + (−1) αCx
1 + (−1) αCx
X
×
κi Ξ1
+
αPs πsr
Pr πrd (1 − Cx2 )
i=1
(37)
+ κ3 Ξ1
p2r πrd πrr 1 − Cx2
αPs2 πsd πsr
+
1
Ps πsd
!!
(38)
m0
0.5Ps πsd
and
i
2 )−P π
Pr πrd (1−Cx
s sd (1+(−1) αCx )
2
Ps πsd (Ps πsd −Pr πrd (1−Cx ))
2 )−P π (1−αC ))(P π (1−C 2 )−P π (1+αC )) ·
(Pr πrd (1−Cx
s sd
x
r rd
s sd
x
x
where κi =
1
+
2
Pr θrd 1−Cx
) (Ψ (C ))m+m0
r
x
k0 +msd dr ·
k+mrr
P
Ψ
(C
)
1
s r
x
+ θ1
+
2)
θsd
Pr θrd (1−Cx
rr
{z
}
(
I
e−ΩΨr
Remark 1. Following Corollary 2, the upper bound in Theorem 7 at Cx = 0 reduces to the exact expression of the endto-end ergodic rate for the PGS case.
∆
!
1
Ps θsr
m+m0
ξl Ψr
which can be solved, using [38, Eq. 3.353-5, 8.352-5] and [39,
Eq. 6.5.9] for the first integral, and [39, Eq. 2.3.6-9] for the
other integrals.
= RLB
E−E,RF ,
0
m−k0
k0X
+msd Z∞
l=1
k0
Prk−m Γ (k + mrr ) Γ (k0 + msd )
×
+
mm0
k
x
∞
0
k+m
Xrr Z ζj Ψm+m e−ΩΨr
r
+
j dΨr
Pr Ψr
1
j=1 0
+
Ps θsr
θrr
!
+ θx
r
0
(35)
By employing a change of variables in the integral I and using
κ3 =
Proof: First, we define a lower bound on the ergodic
rate as
Z∞
LB
UB
RE−E,RF = 1 − PE−E,RF
(Pr , Cx |r ) dr,
(39)
0
5 The
direct numerical computation of this double integral needs a careful
selection of the upper limit value to avoid overflow.
UB
where PE−E,RF
is the outage probability upper bound in the
Rayleigh fading case which is stated in Theorem 4. Hence,
8
0
RE−E
mX
m X
m
sr −1 mX
rd −1 X
1
≤
m
mrr
log (2) m=0 0
Γ (msd ) Γ (mrr ) θsdsd θrr
k=0 0
m =0
×
2
X
!
!
0
Prk−m Γ (k + mrr ) Γ (k0 + msd )
m
m0
0
m (θ (1 − C 2 ))m
k
k0 Psm−k0 Γ (m + 1) Γ (m0 + 1) θsr
rd
x
k =0
0
Di Ω−(m+m ) Γ m + m0 + 1 Ξm+m0 +1
1 + (−1)i Cx Ω
i=1
k+m
Xrr
+
0
j
ζj Pr−j Psj θsr
Ω(j−1)−(m+m ) Γ m + m0 + 1 U
j, j − m + m0 ,
j=1
k0 +msd
X
+
l
ξl Ps−l Prl θrd
1−
l
0
Cx2 Ω(l−1)−(m+m ) Γ
0
where Ω =
+
Pr θrd 1 − Cx2
1
+
l, l − m + m ,
Ps θsd
Ps2 θsr θsd
m+m +1 U
λi =
lim
Ψr →−(1+(−1)i Cx )
2
Q
i=1
1+(−1)i Cx
(Ψr +(
lim
))
Pr Ψr
Ps θsr
, ξl =
rr
(
2
Pr θrd 1−Cx
)
0
dk +msd −l
k0 +msd −l
dΨr
!k0 +m
P s Ψr
(
2
Pr θrd 1−Cx
)
+ θ1
sd
Z∞
=
e
−
Ψr (Cx )
2)
Pr πrd (1−Cx
(a)
dr (40)
x
Z∞ −
e
≥
(Pr πrr +1)
Ψr (αCx )
+ P
s πsr
(Cx )
Ps πsd Pr πΨrdr (1−C
2) + 1
0
0
Ψr (αCx )
(Pr πrr +1)
Ψr (αCx )
2) +
Ps πsr
Pr πrd (1−Cx
r (αCx )
Ps πsd PrΨ
πrd (1−C 2 ) + 1
×e
−
Pr πrr
αPs πsr
(
2
Pr θrd 1−Cx
)
+ θ1
sd
sd
F (Ψr )
dr,
sd −l
)
(k0 +msd −l)!
sd
I MPROPER S IGNALING O PTIMIZATION
In this part, for Rayleigh fading, we optimize the parameters of the IGS transmit signal to minimize the end-to-end
outage probability upper bound given some boundaries for
the optimization variables. First, we individually optimize the
relay power/circularity coefficient, assuming the other variable
is kept fixed. Second, we jointly optimize the power and
circularity.
x
where (a) is obtained by replacing every Ψr (Cx ) by Ψr (αCx )
as it can be readily verified that Ψr (x) is a monotonically
decreasing function in x and α < 1. Then, by changing
of variables and performing partial fraction decomposition,
we get
Pr πrd 1 − Cx2
LB
RE−E,RF =
Ps πsd log (2)
Z∞ X
2
κi
κ3
+
×
2)
Pr πrd (1−Cx
i
Ψ
+
1
+
(−1)
αC
Ψ
+
i=1
r
x
r
Ps πsd
0
Ps Ψr
·
V.
!k0 +m
(
rr
!k0 +m
Ps θsd
Ps
2
Pr θrd 1−Cx
(Ψr +1)
k+m
+ θ1
Ωr →−
we get
RLB
E−E,RF
∆
= RUB
E−E ,
Ψr + 1 + (−1)i Cx F (Ψr ),
k+m
rr
Pr Ψr
dk+mrr −j
lim
F (Ψr )
+ θ1
k+mrr −j
Ps θsr
rr
P θ
Ωr →− s sr dΨr
Pr θrr
k+m −j
rr
Pr
(k+mrr −j)!
Ps θsr
and F (Ψr ) =
!!
(33)
1
,
2)
Pr θrd (1−Cx
ζj =
0
l=1
1
Ps θsr
1
Ps θsr
+ 2
Pr θrr
Pr θrd θrr (1 − Cx2 )
1
+
2)
Pr πrd (1−Cx
Ψr
dΨr ,
(41)
which can be solved using [38, Eq. 3.352-4] to give the lower
bound.
A. Individual Power and Circularity Coefficient Optimization
In order to investigate the merits of IGS over conventional
PGS in FDR channels, we aim at finding the optimal circularity coefficient value that minimizes the end-to-end outage
probability upper bound. Specifically, we aim at solving the
following optimization problem:
min
Cx
s.t.
UB
PE−E,RF
(Pr , Cx )
0 ≤ Cx ≤ 1·
(42)
In order to solve the optimization problem, we analyze the conUB
vexity properties of the objective function PE−E,RF
(Pr , Cx ) in
(28). In general, the function is found to be non-convex due
to the indefinite sign of the second derivative. However, other
desirable properties that allow us to find the global optimal
point are presented in the following theorem.
9
Theorem 9. The upper bound of the end-to-end outage
probability, over Rayleigh fading, is either a monotonic or
a unimodal function in each of its variables, Pr and Cx ,
individually.
Proof: The proof is provided in Appendix VII.
Since monotonicity and unimodality are special cases of
quasi-convexity, such a result allows for the use of quasiconvex optimization algorithms. For instance, the optimal Cx
can be numerically obtained using the well-known bisection
method operating on its derivative given in Appendix B. Similar properties are shown for the individual power optimization
problem.
B. Joint Power and Circularity Coefficient Optimization
The power optimization problem in PGS is formulated as
min
PE−E,RF (Pr , 0)
s.t.
0 < Pr ≤ Pmax ·
Pr
(43)
Also, for the IGS case, the joint problem is given as follows:
min
UB
PE−E,RF
(Pr , Cx )
s.t.
0 < Pr ≤ Pmax ,
0 ≤ Cx ≤ 1·
Pr ,Cx
(44)
It can be readily verified that the PGS outage probability
in (27) is non-convex in Pr . However, it is shown in [33,
Appendix 3.A] that the interior of the function is unimodal,
and hence quasi-convex, in Pr following similar footsteps of
the proof in Theorem 9, i.e., via the Descartes rule of signs
applied to the derivative of the outage probability function.
Hence, the bisection method can be used to locate the global
optimum.
The second problem is a minimization of a non-convex function with simple box constraints. Thus, one may try to solve
it numerically by, for example, the gradient projected method
or the projected Newton’s method without any guarantee to
converge to an optimal solution [40]. For exact performance
analysis purposes, we find the optimal solution via a fine grid
search. Moreover, in Theorem 9, we proved that the objective
function is either a monotonic or unimodal in each of the
optimization variables individually over the interior of the
constraint set. This property motivates us to use a coordinate
descent (alternating optimization) method based on a twodimensional bisection algorithm as in [41]. Fortunately, as
it will be noticed in the numerical results section, it always
converges numerically to the optimal solution obtained by grid
search.
VI. N UMERICAL R ESULTS
We numerically evaluate the benefits that can be reaped by
employing IGS in FDR systems. Throughout the following,
we compare the performance of IGS to that of PGS as a
benchmark in terms of the outage probability and ergodic
rate metrics. For the adaptive design of the FDR system
based on the outage probability over Rayleigh fading, we use
the algorithms discussed in Section V. Specifically, for PGS,
we show the unoptimized performance with maximum power
allocation (MPA), alongside that with optimized relay power
using the bisection algorithm (BA) in addition to a fine grid
search (GS) for verification purposes. On the other hand, the
IGS outage performance is shown via two expressions, namely,
a) the derived upper bound (UB) in Theorem 4 and, b) the exact
end-to-end expression involving the numerical computation
of the integral in (17). The IGS optimization involves two
variables; Pr and Cx . Hence, we consider two cases for IGS
in the presented figures, namely, i) one-dimensional (1D) optimization over Cx while adopting maximum power allocation
for Pr , and ii) joint Pr /Cx two-dimensional (2D) optimization.
The optimization is done for the two aforementioned cases
using both BA and GS, with the prefixes 1D and 2D to
distinguish between them.
Also, for performance analysis purposes, we optimize the
outage probability and the ergodic rate performances of the
FDR system over Nakagami-m fading based on a fine 1D and
2D grid search (GS). We use the simulation parameters in
Table I unless otherwise stated.
A. Outage Probability Performance
Here, we evaluate the proposed lower bound of the end-toend outage probability for m ≥ 1 and the upper bound for
Rayleigh fading. We use these bounds to optimize the outage
performance. Finally, we compare the throughput of the IGSFDR to that of PGS-FDR and HDR.
1) Performance of Proposed Outage Bounds: We evaluate
the performance of the proposed outage probability bounds
comparing to the exact integral form expressions which are
computed numerically. For this purpose, we plot the end-toend outage probability versus Pr and Cx in Fig. 2 for (a) πrr =
0 dB and (b) πrr = 10 dB. It is clear from the figure that the
outage performance improves by increasing m as the fading
effect becomes less severe in the relayed path. Also, we can
see the tightness of the outage probability bounds for different
values of Pr and Cx for the two πrr values. For a fixed Pr while
increasing Cx , the outage performance depends on the RSI.
Specifically, in our simulation setup, it is increasing for πrr =
0 dB and decreasing for πrr = 10 dB. Further, for a fixed Cx
while increasing Pr , in both RSI values, the outage improves
till a certain relay power, at which it starts to deteriorate.
2) Outage Probability Optimization: We evaluate here the
performance of the optimized end-to-end outage probability
with respect to several system and fading parameters based on
the aforementioned optimization algorithms.
Effect of RSI: In Fig. 3, we plot the outage probability versus
πrr for m = 2 based on a fine 1D and 2D GS. As shown,
one can observe that at lower values of the RSI, the IGS
solution reduces to PGS since the RSI is low and the relay can
use more power without deteriorating the S − R link qualityof-service. As πrr increases, the outage performance of the
PGS is significantly deteriorated. On the other hand, the IGS
design saturates at a fixed level. However, this level depends
on the target rate and the S − R and R − D link conditions
which can be clearly noticed from the outage performance at
the two values of π. Moreover, a very interesting observation
10
Parameter
πsr = πrd = π
πsd
Cx
mrr = msd
Value
20 dB
3 dB
0.9
1
Parameter
πrr
Ps = Pr = Pmax
r
msr = mrd
Value
10 dB
1W
1 bits/sec/Hz
m ∈ {1, 2, 3}
TABLE I: Simulation Parameters.
10 -1
10 0
Outage Probability
Outage Probability
m=1
m=1
10
10
m=2
-2
-3
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
10
10 -2
m=2
10
1
-1
-3
0
0.1
0.2
0.3
0.4
Cx
0.7
0.8
0.9
1
IGS-UB
IGS-Exact
IGS-LB
10 0
m=1
m=1
Outage Probability
Outage Probability
0.6
Cx
IGS-UB
IGS-Exact
IGS-LB
10 0
0.5
10 -1
10 -2
10 -3
10 -1
10 -2
m=2
m=2
10 -4
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Pr (W )
(a) πrr = 0 dB
10 -3
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5
Pr (W )
(b) πrr = 10 dB
Fig. 2: Outage prob. vs. Pr [lower] and Cx [upper], for different πrr and m.
in Fig. 3 is that the individual optimization of the signal
asymmetry gives nearly the same performance as that via the
joint optimization of Pr and Cx . Hence, one can resort to the
simpler 1D optimization on Cx to reduce the complexity of the
adaptive FDR system design.
Effect of Allowable Relay Power Budget: In Fig. 4, we
study the outage probability versus the available power budget
at the relay. For FDR with PGS, and specifically when the
relay transmits with its maximum power, the outage probability
performance is known to be enhanced by increasing the
allowable power only till a breakeven point as shown. This
point is where the increasing adverse effect of RSI on the first
hop due to higher relay power starts to exceed any performance
returns due to the higher reliability of the second hop. After
such a point, any increase in the relay power causes a steady
increase in the end-to-end outage probability. If relay power
optimization is allowed in PGS, the performance can at best
be kept constant after this breakeven point by clipping the
transmit power level, rendering any further increase in the
power budget unutilized. On the other hand, the performance
trend is different when IGS is adopted at the relay node.
Indeed, by optimizing the relay’s circularity coefficient, the
outage probability performance continues its decreasing trend.
It is also observed that, unlike in PGS, the relay tends to
use its maximum power in IGS when joint power/circularity
optimization is considered. For high power budgets, the optimal circularity coefficient value approaches unity, denoting
a maximally improper signal that tends to allocate most of
its power in only one dimension of the complex signal space.
This renders the worst case scenario to have the remaining
orthogonal signal space dimension as self-interference-free.
The decreasing trend of the outage probability in IGS, however,
still shows diminishing returns due to the outage performance
bottleneck in the first hop, which is primarily influenced by
the first hop gain, πsr .
Effect of Average S − R Link Gain: In Fig. 5, we plot the
outage probability versus πsr for different source target rates.
First, communication fails at low πsr values due to the first
hop bottleneck, causing the outage probability of both PGS
and IGS to start close to unity. As πsr increases, using IGS
enables the relay to utilize more power relative to PGS to
boost the outage performance. At the end, when πsr reaches
a significantly higher value than the RSI, the first hop no
longer operates in the interference-limited regime, and hence,
the IGS merits become less significant relative to PGS. Finally,
as shown, the merits of IGS over PGS are more clear as the
target rate decreases. In this case, the rate requirements in
the first hop become less stringent, allowing IGS to utilize
higher transmit power relative to PGS and yielding a better
performance.
11
100
10 0
10 -1
10 -2
10-1
10 -3
0
2
4
6
8
10
12
14
16
18
20
Fig. 3: Optimized outage prob. vs. πrr for different π and
m = 2.
0
1
2
3
4
5
Fig. 4: Outage prob. performance vs. relay power budget, for
m = 1.
100
100
10
-1
10
-2
10-1
100
10-2
10-1
10-2
10-3
0.1
0
5
10
15
20
25
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
30
Fig. 5: Outage prob. performance vs. πsr for different target
rates and m = 1.
Fig. 6: Outage prob. performance vs. normalized relay location,
for m = 1 and r = 0.5 bits/sec/Hz.
Effect of Relay Location: We study the relative relay
location impact on the end-to-end outage performance for
πrr ∈ {0, 15} dB. The relay location in Fig. 6 is defined
as the normalized distance of S − R link, Lsr with respect
to the distance of S − D link, Lsd . When the relay location is
closer to the source, the S − R link gain is very strong relative
to the RSI. In such a relatively self-interference-free scenario,
the IGS solution reduces as expected to the PGS solution. As
the relay moves towards the destination, the relative adverse
effect of RSI increases, causing the first hop to operate in
the interference-limited regime. In such a regime, the benefits
of IGS start to show up in mitigating the adverse effect
of the RSI by tuning the signal impropriety. This gives the
performance improvement in the second hop, due to the higher
R − D link gain, a better opportunity to enhance the end-toend performance. When the relay is too close to the destination,
the RSI effect significantly decreases due to the very low relay
power required for successful communication, yielding similar
IGS/PGS performance. It is clear that the benefits of IGS are
noticeable only when the RSI link effect is non-negligible.
When πrr = 0 dB, i.e., at the noise level, IGS yields the PGS
solution.
3) Throughput Comparison to HDR: It is essential to compare the performance of FDR system with IGS not only
to PGS but also to HDR. For this purpose, we consider
the optimized network throughput versus the target rate r
12
for (a) m = 2 and (b) m = 1 in Fig. 7. First, the
optimized throughput can be computed directly from the
optimized end-to-end outage probability, with
a target rate
∗
∗
r, as TE−E
(Pr , Cx ) = r 1 − PE−E
(Pr , Cx ) . For the outage expressions of the HDR, we consider two protocols; 1)
simple multi-hop decode-and-forward (MHDF) HDR, where
the destination distills the desired information only from the
relayed path, and 2) HDR with MRC, where the destination
combines the two time-orthogonal copies of the signal via the
direct and relayed paths, as given in [35, Table I]. It can be
seen from Fig. 7 that the target rate support set is divided into
three regions where one protocol outperforms the others; 1)
FDR with PGS, 2) FDR with IGS, and 3) HDR. As shown,
for very small (as well as for very high) target rates, optimized
IGS simply yields the PGS solution. As the rate requirement
gradually increases, IGS can perform better than PGS and
HDR by increasing the signal asymmetry as discussed earlier.
However, at higher rates, the RSI saturates the FDR receiver
and the performance deteriorates significantly, even with IGS
reaching the maximum impropriety, i.e., Cx = 1. At this point,
HDR with/without MRC starts to offer better throughput than
FDR. We can see that HDR with MRC is slightly better than
HDR without MRC as expected. Moreover, we can observe
from the figure that the FDR with PGS region is wider at
m = 2. This is caused since the fading is less severe in the
relayed path than at m = 1. Hence, the adverse effect of the
RSI on the first hop relatively decreases, leaving more room for
FDR with IGS to compete with HDR via circularity coefficient
tuning. Thus, for lower values of m, the performance of HDR
starts to crossover that of FDR at an earlier target rate cutoff
than for higher m.
B. Ergodic Rate Performance
We first evaluate the performance of the proposed upper
bound of the end-to-end ergodic rate for m ≥ 1 and the lower
bound for Rayleigh fading. Then, we analyze the optimized
end-to-end ergodic rate based on a fine 1D and 2D GS.
1) Performance of Proposed Ergodic Rate Bounds: We plot
the ergodic rate versus Pr and Cx for (a) low RSI: πrr = 0 dB
and (b) high RSI: πrr = 20 dB for different m in Fig. 8.
From the figure, one can notice the tightness of the lower and
upper bounds compared to the exact ergodic rate. However, the
lower bound over Rayleigh fading becomes loose at very high
values of Cx , i.e., Cx ≈ 1. As m increases, higher end-to-end
rate can be achieved as expected. Moreover, similar to Fig. 2,
for a fixed relay transmit power, the value of RSI affects the
ergodic rate performance. For instance, it is decreasing and
increasing for (a) and (b), respectively.
2) Ergodic Rate Optimization: Here, we aim at maximizing
the ergodic rate under the box constrains of Pr and Cx based
on a fine 1D and 2D GS. We plot the end-to-end ergodic rate
versus πrr for (a) m = 2 and (b) m = 1 in Fig. 9. It can
be noticed that IGS reduces to PGS at lower RSI, similar to
the previously noticed outage probability trend in Fig. 3. As
the RSI increases, the PGS performance degrades, while the
IGS attains nearly a fixed value. Although the gap between
the analytical bounds and the exact numerical value widens in
(b) to within 0.5 bits/sec/Hz as the RSI increases, the bounds
still exhibit the same trend as the exact solution, and hence,
they are able to reflect good design insights. For instance, a
good figure of the RSI level at which IGS starts to offer better
rates than PGS can be obtained to be within the two points
via the bounds. Finally, similar to Fig. 3, the optimizing only
the circularity coefficient performs nearly the same as the joint
optimization.
VII. C ONCLUSION
In this work, we study the potential merits of employing
improper Gaussian signaling (IGS) in full-duplex relaying
(FDR) with non-negligible residual self-interference (RSI) over
Nakagami-m fading channels. To analyze the benefits of IGS,
we present exact integral forms as well as analytical lower and
upper bound expressions for the end-to-end outage probability
and ergodic rate in terms of the relay’s transmit power and
circularity coefficient, which measures the degree of the signal
impropriety. In order to optimize the FDR system performance,
we numerically tune the relay power and circularity coefficient
based only on the relay’s knowledge of the channel statistics.
The findings of this work show that IGS yields promising performance merits over proper Gaussian signaling (PGS) in FDR
channels. Specifically, for strong RSI, IGS tends to leverage
higher powers to enhance the performance while alleviating
the RSI impact by tuning the relay’s circularity coefficient.
Further, we show that unlike PGS, as the RSI value increases,
IGS is able to maintain a fixed performance that depends
on the channel statistics and the target rate. It is also shown
that, from an end-to-end throughput standpoint, IGS-FDR can
outperform not only PGS-FDR but also half-duplex relaying
with/without maximum ratio combining over certain regions of
the target rate. Furthermore, numerical simulations show that
it is usually sufficient to individually optimize the circularity
coefficient to obtain nearly the same performance as the joint
optimization of the power and circularity. Future research
lines include obtaining efficient optimization algorithms for
the FDR system performance in the Nakagami-m case based
on the derived bounds. Moreover, by capitalizing on the results
reported herein, a more general multi-antenna setting with RSI
could be considered while adopting IGS at both the source and
relay. In this case, it would be interesting, yet challenging,
to investigate the IGS benefits via the efficient design of the
covariance/pseudo-covariance matrices of the transmit signals.
A PPENDIX A: P ROOF OF P ROPOSITION 1
We prove the convexity of the exponential term inside the
expectation operator in (17) by expressing it as e−f (grr ) . In
fact, one can show easily that f (grr ) can be written as
p
2 + Bg + C − (Dg + F ),
f (grr ) = Agrr
(45)
rr
rr
Pr 2 (1+γ(1−Cx 2 ))
r
where A =
, B = 2(1+γ)P
, C = P(1+γ)
2
2,
Ps 2 πsr 2
Ps 2 πsr 2
s πsr
Pr
1
D = Ps πsr , and F = Ps πsr are positive. Indeed, the second
derivative of f (grr ) with respect to grr is
∂ 2 f (grr )
4AC − B 2
=
≤ 0,
(46)
2
3/2
∂grr
4(C + grr (B + Agrr ))
13
2
2.5
FDR-PGS Region
FDR-PGS Region
HDR Region
HDR-Region
1.8
1.6
FDR-IGS
FDR-PGS
HDR-MRC
HDR
1.5
FDR-IGS
FDR-PGS
HDR-MRC
HDR
1.4
Throughput (bits/sec/Hz)
Throughput (bits/sec/Hz)
2
1
1.2
1
0.8
0.6
0.5
0.4
FDR-IGS
Region
0
0
1
FDR-IGS
Region
0.2
2
3
4
5
6
7
0
8
0
1
2
3
r (bits/sec/Hz)
4
5
6
7
8
r (bits/sec/Hz)
(a) m = 2
(b) m = 1
Ergodic Rate (bits/sec/Hz)
Ergodic Rate (bits/sec/Hz)
Fig. 7: Optimized throughput vs. r, for πrr = 15 dB.
5
4
3
m=3
2
m=1
1
3
m=3
2.5
m=1
2
1.5
1
0
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0
0.1
0.2
0.3
0.4
0.5
Cx
0.7
5
4
3
2
0.8
0.9
IGS-UB
IGS-Exact
IGS-LB
3
Ergodic Rate (bits/sec/Hz)
Ergodic Rate (bits/sec/Hz)
6
0.6
Cx
IGS-UB
IGS-Exact
IGS-LB
m=3
m=1
1
0
m=3
2.5
2
1.5
1
m=1
0.5
0
0
0.5
1
1.5
2
2.5
3
0
0.5
1
1.5
Pr (W )
2
2.5
3
Pr (W )
(a) πrr = 0 dB
(b) πrr = 20 dB
Fig. 8: Erg. rate vs. Pr [lower] and Cx [upper], for different πrr and m.
since 1 + γ(1 − Cx 2 ) ≤ 1 + γ for 0 ≤ Cx ≤ 1. Hence, f (grr )
is concave and e−f (grr ) is convex.
A PPENDIX B: P ROOF OF T HEOREM 9
s πsd
r πrr +1
where 0 ≤ x ≤ 1, a = Pr1πrd , b = PP
and c = P
Pr πrd .
s πsr
In the following, we analyze the stationary points of f (x) =
1 − f (x). Its derivative is given by
−a
df (x)
e
=x
dx
1) Unimodality in Cx : The outage upper bound in (28) as a
function of Cx , denoted here as x, is given on the form.
f (x) = 1 −
e
−a
Ψr (x)
(1−x2 )
−bΨr (αx)
Ψr (x)
c (1−x
2)
+1
,
(47)
where
Ψr (x)
−bΨr (αx)
1−x2
r (x)
c Ψ1−x
2 + 1
2 S(x),
(48)
14
5
4
4.5
3.5
4
3
Ergodic Rate (bits/s/Hz)
Ergodic Rate (bits/s/Hz)
3.5
3
2.5
IGS-UB-2DGS
IGS-UB-1DGS
IGS-Exact-2DGS
IGS-Exact-1DGS
PGS-GS
PGS-MPA
2
1.5
1
2.5
2
IGS-UB-2DGS
IGS-UB-1DGS
IGS-Exact-2DGS
IGS-Exact-1DGS
IGS-LB-2DGS
IGS-LB-1DGS
PGS-GS
PGS-MPA
1.5
1
0.5
0.5
0
0
0
5
10
15
20
25
30
0
5
10
15
πrr (dB)
20
25
30
πrr (dB)
(a) m = 2
(b) m = 1
Fig. 9: Optimized ergodic rate vs. πrr , for different m.
"
a 2Ψr (x) + γ x2 − 1
Ψr (x)
S(x) = c
+
1
2
1 − x2
(Ψr (x) + 1) (1 − x2 )
#
bγα2
γc
2cΨr (x)
+
+
−
·
Ψr (αx) + 1
(Ψr (x) + 1) (1 − x2 ) (1 − x2 )2
(49)
From the given form, and in addition to the roots of S(x),
(x)
it is clear that dfdx
admits only a zero at x = 0. Now, we
investigate the roots for S(x), and use the change of variables,
z = Ψr (x) + 2. Hence, 1 − x2 = z(z−2)
. After substitution
γ
and some manipulations,√S(z) is hence given for our region
of interest, 2 ≤ z ≤ 1 + 1 + γ, by
γ
−aγ 2
bγα2
γ2c
S(z) = c + 1
+
−
·
z
z 2 (z − 1) Ψr (αx) + 1
z 2 (z − 1)
(50)
Since 0 < α < 1, we know that 1 − α2 x2 ≥ 1 − x2 . Hence,
Ψr (αx)+1 ≥ Ψr (x)+1 = z−1. Let Ψr (αx)+1 = tz (z−1),
where tz ≥ 1. Therefore,
(cγ + z)(−aγ 2 tz + bγα2 z 2 ) − γ 2 ctz z
· (51)
S(z) =
tz z 3 (z − 1)
The numerator is a cubic polynomial in z which is given by
T (z) = bα2 γz 3 + bα2 cγ 2 z 2 − (a + c)γ 2 tz z − acγ 3 tz · To find
the number of positive roots for T (z), we use the Descartes
rule of signs [42]. Specifically, for the sequence formed by
the descending order of the cubic equation coefficients, i.e., the
sequence {bα2 γ, bα2 cγ 2 , −(a+c)γ 2 tz , −acγ 3 tz }, the number
of sign changes is only one. For our real cubic polynomial, this
determines the number of positive roots to be exactly one
√ root.
Hence, in the positive region of interest, 2 ≤ z ≤ 1 + 1 + γ,
either one or no feasible roots exist for T (z), and hence for
S(z). This shows that f (x) is either monotonic or unimodal
due to the existence of one root at maximum in its interior.
2) Unimodality in Pr : The outage probability upper bound
in (28) as a function of Pr , also denoted here as x, is given
by
a
x +bx
−
f (x) = 1 − c
e
d
x
|
,
+1
{z }
(52)
D(x)
Ψr (αCx )
Ψr (Cx )
Ψr (αCx )
, c = e− Ps πsr , and
2 ) , b = πrr P π
πrd (1−Cx
s sr
Ps πsd Ψr (Cx )
2 ) . In what follows, we analyze the stationary
πrd (1−Cx
where a =
d=
points of the interior of the function D(x). Differentiating
D(x) with respect to x yields
S(x)
z
}|
{
dD
dx + (a − bx2 )(x + d) −( xa +bx)
=
e
.
(53)
dx
x(x + d)2
Now, the stationary points of D(x) originate from the roots of
the cubic polynomial S(x) = −bx3 − bdx2 + (a + d)x + ad.
Again, there exists only one sign change in the coefficients of
the cubic equation, and hence, only one root exists for S(x).
Therefore, in the positive region of interest, i.e., 0 < Pr <
Pmax , D(x) and hence f (x), are either monotonic or unimodal
in x.
R EFERENCES
[1]
M. Jain, J. Choi, T. Kim, D. Bharadia, S. Seth, K. Srinivasan, P. Levis,
S. Katti, and P. Sinha, “Practical, real-time, full duplex wireless,” in
Proc. ACM MobiCom’11, Las Vegas, NV, Sep. 2011.
15
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
M. Duarte, C. Dick, and A. Sabharwal, “Experiment-driven characterization of full-duplex wireless systems,” IEEE Trans. Wireless Commun.,
vol. 11, no. 12, pp. 4296–4307, Dec. 2012.
A. Sabharwal, P. Schniter, D. Guo, D. W. Bliss, S. Rangarajan, and
R. Wichman, “In-band full-duplex wireless: Challenges and opportunities,” IEEE J. Sel. Areas Commun., vol. 32, no. 9, pp. 1637–1652, Sep.
2014.
S. Hong, J. Brand, J. Choi, M. Jain, J. Mehlman, S. Katti, and P. Levis,
“Applications of self-interference cancellation in 5G and beyond,” IEEE
Commun. Mag., vol. 52, no. 2, pp. 114–121, Feb. 2014.
V. R. Cadambe, S. A. Jafar, and C. Wang, “Interference alignment with
asymmetric complex signaling: settling the Høst-Madsen-Nosratinia
conjecture,” IEEE Trans. Inf. Theory, vol. 56, no. 9, pp. 4552–4565,
Sep. 2010.
P. J. Schreier and L. L. Scharf, Statistical signal processing of complexvalued data: the theory of improper and noncircular signals. Cambridge University Press, 2010.
Z. Ho and E. Jorswieck, “Improper Gaussian signaling on the two-user
SISO interference channel,” IEEE Trans. Wireless Commun., vol. 11,
no. 9, pp. 3194–3203, Sep. 2012.
C. Lameiro and I. Santamaria, “Degrees-of-freedom for the 4-user SISO
interference channel with improper signaling,” in Proc. IEEE ICC’13,
Budapest, Hungary, Jun. 2013.
Y. Zeng, R. Zhang, E. Gunawan, and Y. Guan, “Optimized transmission
with improper Gaussian signaling in the K-user MISO interference
channel,” IEEE Trans. Wireless Commun., vol. 12, no. 12, pp. 6303–
6313, Dec. 2013.
C. Wang, T. Gou, and S. A. Jafar, “On optimality of linear
interference alignment for the three-user MIMO interference channel
with constant channel coefficients,” 2011. [Online]. Available:
http://escholarship.org/uc/item/6t14c361
S. Lagen, A. Agustin, and J. Vidal, “Coexisting linear and widely linear
transceivers in the MIMO interference channel,” IEEE Trans. Signal
Process., vol. 64, no. 3, pp. 652–664, Feb. 2016.
E. Kurniawan and S. Sun, “Improper Gaussian signaling scheme for the
Z-interference channel,” IEEE Trans. Wireless Commun., vol. 14, no. 7,
pp. 3912–3923, Jul. 2015.
S. Lagen, A. Agustin, and J. Vidal, “On the superiority of improper
Gaussian signaling in wireless interference MIMO scenarios,” IEEE
Trans. Commun., vol. 64, no. 8, pp. 3350–3368, Aug. 2016.
C. Lameiro, I. Santamara, and P. J. Schreier, “Rate region boundary of
the siso z-interference channel with improper signaling,” IEEE Trans.
Commun., vol. 65, no. 3, pp. 1022–1034, Mar. 2017.
L. Yang and W. Zhang, “Interference alignment with asymmetric
complex signaling on MIMO X channels,” IEEE Trans. Commun.,
vol. 62, no. 10, pp. 3560–3570, Oct. 2014.
H. Y. Shin, S. H. Park, H. Park, and I. Lee, “A new approach
of interference alignment through asymmetric complex signaling and
multiuser diversity,” IEEE Trans. Wireless Commun., vol. 11, no. 3, pp.
880–884, Mar. 2012.
C. Lameiro, I. Santamaria, and P. Schreier, “Benefits of improper
signaling for underlay cognitive radio,” IEEE Wireless Commun. Lett.,
vol. 4, no. 1, pp. 22–25, Feb. 2015.
M. Gaafar, O. Amin, W. Abediseid, and M.-S. Alouini, “Spectrum
sharing opportunities of full-duplex systems using improper Gaussian
signaling,” in IEEE PIMRC’15, Hong Kong, Aug. 2015.
C. Lameiro, I. Santamaria, W. Utschick, and P. J. Schreier, “Maximally
improper interference in underlay cognitive radio networks,” in Proc.
IEEE ICASSP’16, Shanghai, China, Mar. 2016.
O. Amin, W. Abediseid, and M.-S. Alouini, “Underlay cognitive radio
systems with improper Gaussian signaling: Outage performance analysis,” IEEE Trans. Wireless Commun., vol. 15, no. 7, Jul. 2016.
M. Gaafar, O. Amin, W. Abediseid, and M.-S. Alouini, “Underlay
spectrum sharing techniques with in-band full-duplex systems using
improper Gaussian signaling,” IEEE Trans. Wireless Commun., vol. 16,
no. 1, pp. 235–249, Jan. 2017.
[22]
O. Amin, W. Abediseid, and M.-S. Alouini, “Overlay spectrum sharing
using improper Gaussian signaling,” IEEE J. Sel. Areas Commun.,
vol. 35, no. 1, pp. 50–62, Jan. 2017.
[23]
M. Gaafar, O. Amin, A. Ikhlef, A. Chaaban, and M.-S. Alouini, “On
alternate relaying with improper Gaussian signaling,” IEEE Commun.
Lett., vol. 20, no. 8, pp. 1683–1686, Aug. 2016.
[24]
H. D. Nguyen, R. Zhang, and S. Sun, “Improper signaling for symbol
error rate minimization in K-user interference channel,” IEEE Trans.
Commun., vol. 63, no. 3, pp. 857–869, Mar. 2015.
[25]
S. Javed, O. Amin, S. S. Ikki, and M. S. Alouini, “Asymmetric hardware
distortions in receive diversity systems: Outage performance analysis,”
IEEE Access, vol. 5, pp. 4492–4504, Feb. 2017.
[26]
C. Hellings, L. Gerdes, L. Weiland, and W. Utschick, “On optimal
Gaussian signaling in MIMO relay channels with partial decode-andforward,” IEEE Trans. Signal Process., vol. 62, no. 12, pp. 3153–3164,
Jun. 2014.
[27]
C. Kim, E.-R. Jeong, Y. Sung, and Y. H. Lee, “Asymmetric complex
signaling for full-duplex decode-and-forward relay channels,” in Proc.
Int. Conf. on ICT Convergence (ICTC), Jeju Island, South Korea, Oct.
2012.
[28]
M. Gaafar, M. G. Khafagy, O. Amin, and M.-S. Alouini, “Improper
Gaussian signaling in full-duplex relay channels with residual selfinterference,” in Proc. IEEE ICC’16, Kuala Lumpur, Malaysia, May
2016.
[29]
M. Abramowitz and I. A. Stegun, Handbook of mathematical functions.
10th Printing, Dover Publications, Dec. 1972.
[30]
T. Kwon, S. Lim, S. Choi, and D. Hong, “Optimal duplex mode for DF
relay in terms of the outage probability,” IEEE Trans. Veh. Technol.,
vol. 59, no. 7, pp. 3628–3634, Sep. 2010.
[31]
T. Riihonen, S. Werner, and R. Wichman, “Hybrid full-duplex/halfduplex relaying with transmit power adaptation,” IEEE Trans. Wireless
Commun., vol. 10, no. 9, pp. 3074–3085, Sep. 2011.
[32]
M. K. Simon and M.-S. Alouini, Digital communication over fading
channels. Wiley-Interscience, 2005, vol. 95.
[33]
M. G. Khafagy, “On the performance of in-band full-duplex cooperative
communications,” Ph.D. dissertation, KAUST, Jun. 2016. [Online].
Available: http://hdl.handle.net/10754/617879
[34]
D. Bharadia, E. McMilin, and S. Katti, “Full duplex radios,” in Proc.
ACM SIGCOMM’13. New York, NY, USA: ACM, 2013, pp. 375–386.
[Online]. Available: http://doi.acm.org/10.1145/2486001.2486033
[35]
M. G. Khafagy, A. Ismail, M.-S. Alouini, and S. Assa, “Efficient
cooperative protocols for full-duplex relaying over nakagami- m fading
channels,” IEEE Trans. Wireless Commun., vol. 14, no. 6, pp. 3456–
3470, Jun. 2015.
[36]
F. D. Neeser and J. L. Massey, “Proper complex random processes with
applications to information theory,” IEEE Trans. Inf. Theory, vol. 39,
no. 4, pp. 1293–1302, Jul. 1993.
[37]
Y. Zeng, C. M. Yetis, E. Gunawan, Y. L. Guan, and R. Zhang, “Transmit
optimization with improper Gaussian signaling for interference channels,” IEEE Trans. Signal Process., vol. 61, no. 11, pp. 2899–2913, Jun.
2013.
[38]
I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series, and
Products, Seventh Edition. Academic Press, 2007.
[39]
A. Prudnikov, Y. A. Brychkov, and O. Marichev, Integrals and Series,
vol. I: Elementary Functions. Gordon and Breach Science Publishers,
1998.
[40]
D. P. Bertsekas, Nonlinear Programming.
[41]
M. Fadel, A. El-Keyi, and A. Sultan, “QOS-constrained multiuser peerto-peer amplify-and-forward relay beamforming,” IEEE Trans. Signal
Process., vol. 60, no. 3, pp. 1397–1408, Mar. 2012.
[42] V. V. Prasolov, Polynomials.
2009.
Athena Scientific, 1999.
Springer Science & Business Media,
| 7cs.IT
|
C OMPACT N EURAL N ETWORKS BASED ON THE
M ULTISCALE E NTANGLEMENT R ENORMALIZATION
A NSATZ
Andrew Hallam1 , Edward Grant2 , Vid Stojevic1,3 , Simone Severini2,4 , Andrew G. Green5
arXiv:1711.03357v2 [cs.NE] 3 Apr 2018
1
Department of Physics & Astronomy, University College London
Department of Computer Science, University College London
3
GTN Ltd.
4
Institute of Natural Sciences, Shanghai Jiao Tong University
5
London Centre for Nanotechnology, University College London
2
A BSTRACT
This paper demonstrates a method for tensorizing neural networks based upon
an efficient way of approximating scale invariant quantum states, the Multi-scale
Entanglement Renormalization Ansatz (MERA). We employ MERA as a replacement for the fully connected layers in a convolutional neural network and test this
implementation on the CIFAR-10 and CIFAR-100 datasets. The proposed method
outperforms factorization using tensor trains, providing greater compression for
the same level of accuracy and greater accuracy for the same level of compression. We demonstrate MERA layers with 14000 times fewer parameters and a
reduction in accuracy of less than 1% compared to the equivalent fully connected
layers, scaling like O(N ).
1
I NTRODUCTION
The curse of dimensionality is a major bottleneck in machine learning, stemming from the exponential growth of variables with the number of modes in a data set (Cichocki et al. (2016)). Typically
state-of-the-art convolutional neural networks have millions or billions of parameters. However,
previous work has demonstrated that representations stored in the network parameters can be highly
compressed without significant reduction in network performance (Novikov et al. (2015), Garipov
et al. (2016), ?). Determining the best network architecture for a given task remains an open problem.
Descriptions of quantum mechanical systems raise a similar challenge; representing n d-dimensional
particles requires a rank-n tensor whose memory cost scales as dn . Indeed, it was the promise of
harnessing this that led Richard Feynman (Feynman (1982)) to suggest the possibility of quantum
computation. In the absence of a quantum computer, however, one must use compressed representations of quantum states.
A level of compression can be achieved by factorizing the tensorial description of the quantum
wavefunction. Many such factorizations are possible, the optimal structure of the factorization being
determined by the structure of correlations in the quantum system being studied. A revolution in
quantum mechanics was made by realizing that the best way to characterize the distribution of
correlations and information in a state is by a quantity known as entanglement – loosely the mutual
quantum information between partitions of a quantum system (Eisert et al. (2010)).
This has led to many successful applications of tensorial approaches to problems in solid state
physics and quantum chemistry over the past 25 years (Orús (2014), Kin-Lic Chan et al. (2007)).
Intriguing ideas have also emerged over the past few years attempting to bridge the successes of
neural networks in machine learning with those of tensorial methods in quantum physics, both at
a fundamental level (Lin et al. (2017), Mehta & Schwab (2014)), and as a practical tool for network design (Levine et al. (2017)). Recent work has suggested that entanglement itself is a useful
quantifier of the performance of neural networks (Levine et al. (2017), Liu et al. (2017))
1
The simplest factorization employed in quantum systems is known as the matrix product state (Orús
(2014)). In essence, it expresses the locality of information in certain quantum states. It has already
been adopted to replace expensive linear layers in neural networks – in which context it has been
independently termed tensor trains (Oseledets (2011)). This led to substantial compression of neural
networks with only a small reduction in the accuracy (Novikov et al. (2015), Garipov et al. (2016)).
Here we use a different tensor factorization – known as the Multi-scale Entanglement Renormalization Ansatz (MERA) – that encodes information in a hierarchical manner (Vidal (2008)). MERA
works through a process of coarse graining or renormalization. There have been a number of papers
looking at the relationship between renormalization and deep learning. MERA is a concrete realization of such a renormalization procedure (Vidal (2009)) and so possesses a multi-scale structure that
one might anticipate in complex data. A number of works have utilized tree tensor network models
that possess a similar hierarchical structure. However, they do not include the disentangler tensors
that are essential if each layer of the MERA is to capture correlations on different length scales (Liu
et al. (2017)).
In this work we employ MERA as a replacement for linear layers in a neural network used to
classify the CIFAR-10 and CIFAR-100 datasets. Our results show that this performs better than the
tensor train decomposition of the same linear layer, and gives better accuracy for the same level
of compression and better compression for the same level of accuracy. In Section 2 we introduce
factorizations of fully connected linear layers, starting with the tensor train factorization followed
by a tree-like factorization and finally the MERA factorization. In Section 3 we discuss how this is
employed as a replacement for a fully connected linear layer in deep learning networks. Section 4
gives our main results and we note connections with the existing literature in Section 5. Finally, in
Section 6 we discuss some potential developments of the work.
2
T ENSOR FACTORIZATION OF L INEAR L AYERS
a)
i1
b)
i2
iN
j1
j2
jN
i1
i2
j1
c)
α1
α2
αN
j2
jN
i1
j1
j2
j3
j4
j5
j6
j7
j8
i2
i3
i4
i5
i6
i7
i8
d)
iN
i1
j1
j2
j3
j4
j5
j6
j7
j8
i2
i3
i4
i5
i6
i7
i8
Figure 1: Schematic diagrams of various tensor factorizations of linear layers. a) a general linear
layer, b) its tensor train factorization. The squares represent smaller tensors. Connections represent
contractions as indicated in Eq.(1). c) Tree network factorization. d) MERA factorization.
In this report we have replaced the linear layers of the standard neural network with tensorial MERA
layers. The first step in achieving this involves expressing a linear layer as a tensor. This can
be accomplished by taking a matrix W and reshaping it to be a higher dimensional array. For
2
example, suppose W is dn by dn dimensional with components WAB . It can be transformed into
a rank 2n tensor by mapping A to n elements A → i1 , i2 , ..., in and B to another n elements
B → j1 , j2 , ..., jn . In this case each of the elements of the new tensor will be of size d.
,i2 ,...,in
. It is important to note
Figure 1a gives a graphical representation of this rank 2n tensor Wji11,j
2 ,...,jn
that in this representation, the lines represent the indices of the tensors rather than weights. Figure
1b illustrates the tensor train decomposition of W. This consists of writing the larger tensor as the
contraction of a train of smaller tensors:X
αn−1 ,in
,i2 ,...,in
1 ,i1
.
(1)
=
Aij11 ,α1 Aα
Wji11,j
j1 ,α2 · · · Ajn
2 ,...,jn
α1 ,α2 ,...,αn−1
In the tensor graphical notation, closed legs represent indices being summed over and free legs
represent indices that aren’t being summed over. For example, in equation 1 the αi indices are being
summed over and in Figure 1b the αi lines are connected to tensors at both ends.
If each index runs over values from 1 to d, this represents an exponential reduction from d2n parameters to n(Dd)2 , where the indices α run over values from 1 to D (known as the bond order
or Schmidt rank in the quantum context). As noted above, this type of tensor factorization works
well in physics when the information has a local structure (Eisert et al. (2010), Verstraete & Cirac
(2006)); tensor trains capture correlations effectively up to length scales of order log D (Schollwöck
(2011)). This means that while useful for many tasks, the learned representations will be highly
local. Tensors at either end of a tensor train decomposition of a linear layer will not be strongly
correlated with one another.
A hierarchically structured tensor network can better represent correlations across the linear layer.
The tree tensor network shown in Figure 1c represents one possible hierarchical factorization. Each
element of this network is a rank 4 tensor. The two tensors on the top left would have the form
Mji11,i,α21 and Nij32,i,α4 2 . The in elements being represented by the lines on the left of the figure, the
jn elements represented by the dotted lines on the right of the figure and the αn lines being those
connected with the tensor immediately to the right of M and N .
Reading from left to right Figure 1c can be interpreted as follows: the tree-like connectivity imbues
the network with a causal structure whereby a given linear element and its outputs are influenced by
inputs in a region determined by its height in the tree.
For example, the rightmost element in Figure 1c is influenced by all of the inputs, whereas the top
element in the middle column is influenced by inputs i1 to i4 . Elements other than the rightmost
tensor have one dashed output (that connects directly to the overall output) and one solid output
(that ties it to the branching tree structure). These dashed lines are controlled by representations
occurring on a particular scale in the data.
Notice that removing these dashed lines, the network has a true tree structure and represents a coarse
graining or renormalization of the network. In this case, the linear elements are the isometries of the
original MERA’s definition (Vidal (2008; 2009)).
The simple tree network, which has been studied before in the context of neural networks, has
a major deficiency. At each branching, it partitions the system in two, so that in extremis, the
correlations between neighbouring inputs – for example i4 and i5 in Figure 1c – are only controlled
by the element at the end of the network. Requiring the higher elements in the tree-structure to
capture correlations between neighbouring inputs restricts their ability to describe the longer length
scale correlations you would hope to capture by using a hierarchical structure.
The MERA (Vidal (2009)) factorization was introduced in order to solve this problem. As can be
seen in Figure 1d it adds an additional set of rank 4 tensors called disentanglers. The MERA is
,β2
constructed by taking a tree network and placing one of these rank 4 tensors Dγβ11,γ
such that its
2
right-going legs β1 and β2 connect to two adjacent tensors of the tree network. For example, if we
consider the top left-most disentangler in Figure 1d it has elements Diβ21,i,β32 and connects to the tree
j ,α
j ,α
elements M0 i11,β11 and N 0 β22 ,i42 with β1 and β2 then being summed over.
The role of the disentanglers is to cause all correlations on the same length scale to be treated
similarly. For example, correlations between any two neighbouring input indices in and in+1 will
be captured by either the first row of tree elements or the disentanglers. This allows the later elements
in the network to work at capturing longer range correlations.
3
In summary, a rank-N MERA layer can be constructed in the following manner:
1. Create a tree tensor layer. For example, an N = 2τ tree can be constructed from 2τ −1
,β2
rank-4 tree tensors Mβγ11,γ
in the first layer, followed by 2τ −2 tree tensors in the second
2
layer until after τ layers there is only a single tree tensor.
,β2
2. A set of disentanglers are introduced. These are rank-4 tensors Dγβ11,γ
which are placed
2
such that every disentangler is contracted with two neighbouring tree tensors in an upcoming layer of the tree tensor.
3
E XPERIMENTS & NETWORK STRUCTURE
We have considered the performance of a neural network with the two penultimate fully connected
layers of the model replaced with MERA layers, similar to the Novikov et al. (2015) study of compression of fully connected layers using tensor trains. We have quantified the performance of the
MERA layer through comparisons with two other classes of networks: fully connected layers with
varying numbers of nodes and tensor train layers with varying internal dimension. The three types
of network are otherwise identical.
The networks consisted of three sets of two convolutional layers each followed by max pooling
layers with 3 × 3 kernels and stride 2. The convolutional kernels were 3 × 3. There were 64
channels in all of the convolutional layers except for the input, which had three channels, and the
last convolutional layer, which had 256 channels. The final convolutional layer was followed by two
more hidden layers, these were either fully connected, MERA layers or TT-layers depending upon
the network. The first of these layers was of size 4096 × x, the second is of size x × 64. For the
MERA and TT networks, these layers were 4096 × 4096 and 4096 × 64.
The final layer had 10 or 100 nodes corresponding to the image classes in CIFAR-10 and CIFAR100. Leaky rectified linear units (LReLU) were used on all layers except the final layer, with leak =
0.2 (Maas et al. (2013)).
During training, nodes in the final convolutional layer and the two first fully connected layers were
dropped with probability 0.5. The penultimate convolutional layer nodes were dropped with probability 0.2 (Srivastava et al. (2014)). Batch-normalization was used on all layers after dropout and
max pooling (Ioffe & Szegedy (2015)). We did not use bias units.
Gaussian weight initialization was employed in the fully connected models with standard deviation
equal to √n1in , where nin was the number of inputs (He et al. (2015)).
In this report we considered networks with two varieties of fully-connected layers. The first of these
networks had a 4096 × 4096 fully connected layer followed by one which was 4096 × 64; this
network was used as a benchmark against which the other models could be compared. The second
network instead had a 4096 × n fully connected layer followed by a n × 64 layer where n = 5
for the CIFAR-10 network and n = 10 for the CIFAR-100 network. We trained these network to
compare the MERA and tensor train layers to a fully connected model with a comparable number
of parameters, in order to evaluate how detrimental naive compression is to accuracy.
A schematic of the two MERA layers can be found in Figure 2. The input to the first MERA layer
was reshaped in to a rank-12 tensor with each index being dimension 2, as described in Section 2.
The MERA layer was then constructed from a set of rank-4 tensors using the method described in
Section 2.
The first MERA layer works as follows: It contains a column of 6 rank-4 tree elements, followed by
3 tree elements and finally a single tree element. 5 disentanglers are placed before the first column
of tree elements and 2 more disentanglers are placed before the second column of tree elements.
The second MERA layer has an identical structure to the first MERA layer, one of the outputs of the
first set of tree elements is fixed. As a result the output of the second MERA layer is 64 nodes.
MERA weights were initialized using elements of randomized orthogonal matrices (Saxe et al.
(2013)). The tensors themselves were constructed by reshaping these matrices, as described in Section 2. The random orthogonal matrix was constructed using the method of Stewart (?, ?). Starting
from a random n−1×n−1 dimensional orthogonal matrix, a random n×n dimensional orthogonal
4
CONV OUT
PREDICTION
Figure 2: A schematic of the MERA layers of the model. The small rectangles represent linear
elements to factorize a general linear layer. White rectangles represent disentanglers. Red rectangles
represent tree elements. Solid black lines connecting nodes represent tensor contraction and dashed
lines with arrow heads represent the nonlinearities being applied. Dashed lines ending in a circle
represent fixed outputs.
matrix can be constructed by taking a randomly distributed n-dimensional vector, constructing its
Householder transformation, and then applying the n − 1 dimensional matrix to this vector.
Finally, a network with its fully connected layers replaced with a tensor train decomposition was
trained in order to provide a comparison with the MERA layers. The tensor train layers were constructed as described in Section 2 with the internal dimension being fixed at D = 3. In the second
tensor train layer, half of the output indices were fixed to match the second MERA layer.
We tested performance on the CIFAR-10 and CIFAR-100 datasets. We used 45, 000 images for
training, 5, 000 for validation and 10, 000 for testing. Each training batch consisted of 50 images.
Training data was augmented by randomly flipping and translating the input images by up to 4
pixels. Translated images were padded with zeros. All images were normalized by dividing by 255
and subtracting the mean pixels value from the training set.
Validation and test set accuracy was recorded every 500 iterations and training was stopped when
validation accuracy did not improve for 10 successive tests. The network was trained using backpropagation and the Adam optimizer, with initial learning rate 0.001 (Kingma & Ba (2014)) and
a softmax-cross-entropy objective. The test set accuracy for the model with the highest validation
set accuracy was recorded. Each network was trained 10 times with a different random weight
initialization.
The networks were implemented in Tensorflow r1.3 and trained on NVIDIA Titan Xp and 1080ti
GPUs.
4
E XPERIMENTAL RESULTS
In Table 1 we compare the different models described in section 3 trained on the CIFAR-10 dataset.
The compression rate stated is with respect to the number of parameters used in the fully-connected
benchmark model, FC-1.
When comparing the MERA network to the fully connected model, FC-1 we see a considerable
drop in the number of parameters required with only a modest drop in the accuracy of the network.
MERA compresses the fully connected layers by a factor of 14, 000 with a drop in the accuracy of
only 0.4%. We do not attempt to compress the convolutional layers in this work so in the MERA
network the vast majority of the parameters are used in the convolutional layers which are identical
to the fully connected model.
How significant is the MERA network structure we have chosen to the results obtained? To test this
we compare the MERA results obtained to the fully connected model with many fewer parameters
in the fully connected layers, FC-2. Despite having around 20 times more parameters in the fully
5
connected layer than the MERA model, the MERA model significantly out performs FC-2, with a
1.2% drop in the accuracy of FC-2 compared to MERA.
The MERA network also compares favourably to a tensor train network. In this case, the two networks have a comparable number of parameters but the MERA appears to achieve a higher accuracy
than the tensor train network in this case.
Results for the CIFAR-100 model can be seen in Table 2. While none of the networks are as accurate
as the benchmark case, the MERA network continues to outperform the tensor train and ablated fully
connected network. However, the reduction in accuracy compared to the fully connected network is
larger than for the CIFAR-10 dataset.
In addition to the degree of compression achieved by these networks, we also address the time to
optimize. There is evidently a degree of compromise required here: the number of multiplications
required to apply a MERA layer scales with the input size N and bond order D as N log2 D . The
equivalent scaling for a tensor train and fully connected layer are N D2 and N 2 , respectively. This
is reflected in the times taken to optimize these networks. Note however, that MERA can accommodate correlations at all scales of its input even at low bond order, whereas tensor trains require a
bond order that scales exponentially with the length scale of correlation (Orús (2014)). MERA is,
therefore, expected to scale better for very large data sets than either tensor trains or fully connected
layers.
Table 1: The CIFAR-10 experimental results for the different models. FC1 was the fully-connected
model and FC2 was the fully-connected model with severely reduced number of parameters in the
fully-connected layers. MERA are the result for the MERA inspired network. Finally TT is the
tensor train model with the internal dimension being 3.
Network
Parameters
(Total)
17,336,640
Compression
(FC layer)
1
Compression
(Total)
1
Accuracy
FC-1
Parameters
(FC Layer)
17,040,000
88.9
Standard
Deviation
0.2
FC-2
21,440
318,080
795
54.5
86.5
0.8
MERA
1192
297,832
14,295
58.21
88.5
0.1
TT
1312
297,952
12,987
58.19
87.9
0.2
Table 2: The CIFAR-100 experimental results for the different models. FC1 was the fully-connected
model and FC2 was the fully-connected model with severely reduced number of parameters in the
fully-connected layers. MERA are the result for the MERA inspired network. Finally TT is the
tensor train model with the internal dimension being 3.
Network
5
Parameters
(Total)
17,342,400
Compression
(FC Layer)
1
Compression
(Total)
1
Accuracy
FC-1
Parameters
(FC Layer)
17,045,760
61.8
Standard
Deviation
0.7
FC-2
48,000
344,640
355
50.3
53.4
0.6
MERA
6952
303,592
2451
57.12
58.4
0.6
TT
7072
303,712
2410
57.10
57.9
0.6
R ELATED W ORK
Given how memory intensive deep neural networks typically are, substantial effort has been made to
reduce number of parameters these networks require without significantly reducing their accuracy.
Some of these have taken a similar approach to the MERA network described above, using tensor
decompositions of the fully connected layers.
6
These include the tensor train models of Novikov et al. (2015) and Garipov et al. (2016). Here we
have found replacing a fully connected linear layer with a MERA factorization resulted in superior
accuracy for a comparable number of parameters.
More directly related to this MERA model are a number of tree tensor network models (Liu et al.
(2017), Levine et al. (2017)). As Section 2 explained, tree tensor networks inconsistently capture
correlations on the same length scale, this is the reason for the introduction of disentanglers. Tree
tensors do not possess these and we expect them to struggle to capture long range correlations as
effectively as MERA.
A MERA works through a process of coarse graining or renormalization. There have been a number of other papers looking at the relationship between renormalization and deep learning. Lin
et al. (2017) argue that the effectiveness of deep neural networks should be thought of in terms of
renormalization and Mehta & Schwab (2014) demonstrate an exact mapping between the variational
renormalization group and restricted Boltzmann machines. In this report we have taken a different
approach: only the fully connected layers of the network were replaced with MERA layers.
6
D ISCUSSION
We have shown that replacing the fully connected layers of a deep neural network with layers based
upon the multi-scale entanglement renormalization ansatz can generate significant efficiency gains
with only small reduction in accuracy. When applied to the CIFAR-10 data we found the fully
connected layers can be replaced with MERA layers with 14, 000 times less parameters with a
reduction in the accuracy of less than 1%. The model significantly outperformed compact fully
connected layers with 70 − 100 times as many parameters. Moreover, it outperformed a similar
replacement of the fully connected layers with tensor trains, both in terms of accuracy for a given
compression and compression for a given accuracy. While the MERA layer resulted in a larger
accuracy drop in the CIFAR-100 case, it still outperformed a comparable tensor train network.
An added advantage — not explored here — is that a factorized layer can potentially handle much
larger input data sets, thus enabling entirely new types of computation. Correlations across these
large inputs can be handled much more efficiently by MERA than by tensor trains. Moreover, a
compressed network may provide a convenient way to avoid over-fitting of large data sets. The
compression achieved by networks with these factorized layers comes at a cost. They can take
longer to train than networks containing the large fully connected layers due to the number of tensor
contractions required to apply the factorized layer.
Our results suggest several immediate directions for future inquiry. Firstly, there are some questions
about how to improve the existing model. For example, before the MERA layer is used the input is
reshaped into a rank-12 tensor. There isn’t a well defined method for how to perform this reshaping
optimally and some experimentation is necessary. The best way to initialize the MERA layers is
also still an open question.
The results presented here are a promising first step for using MERA in a more fundamental way.
Since MERA can be viewed as a coarse graining procedure (as explained in Section 2), and image
data is often well represented in a hierarchical manner, one possibility would be to simply train a
two-dimensional MERA directly on an image dataset, with no reference to a neural network. In
Stoudenmire & Schwab (2016) a similar idea was explored with matrix product states being trained
directly on MNIST. An alternative possibility would be the replacement of just the convolutional
layers of the network with a two-dimensional MERA. Both of these approaches would be closer
in spirit to the fundamental ideas about the relationships between quantum physics and machine
learning proposed in Lin et al. (2017) and Mehta & Schwab (2014).
Additionally, there has been some work using entanglement measures to explore how correlations
are distributed in deep neural networks, and then utilizing these in order to optimize the design of
networks (Liu et al. (2017), Levine et al. (2017)). It would be intriguing to explore such ideas using
MERA, for example by using the concrete MERA model explored in this paper, or one of the more
ambitious possibilities mentioned above.
We end by noting two facts: any variational approximation to a quantum wavefunction can be used
to construct a replacement for linear layers of a network. There are many examples and each may
7
have its sphere of useful application. Moreover, quantum computers of the type being developed
currently by several groups are precisely described by a type of tensor network (a finite-depth circuit
- and one that may very soon be too large to manipulate classically) and could be used as direct
replacement for linear layers in a hybrid quantum/classical neural computation scheme.
ACKNOWLEDGMENTS
This work was supported by the Engineering and Physical Sciences Research Council [grant number
EP/P510270/1]. The Titan Xp used for this research was donated by the NVIDIA Corporation. We
would like to thank Miles Stoudenmire for many enlightening discussions.
R EFERENCES
A. Cichocki, N. Lee, I. V. Oseledets, A.-H. Phan, Q. Zhao, and D. Mandic. Low-Rank Tensor
Networks for Dimensionality Reduction and Large-Scale Optimization Problems: Perspectives
and Challenges PART 1. ArXiv e-prints arXiv:1609.00893, September 2016.
J. Eisert, M. Cramer, and M. B. Plenio. Colloquium: Area laws for the entanglement entropy.
Reviews of Modern Physics, 82:277–306, January 2010. doi: 10.1103/RevModPhys.82.277.
R. P. Feynman. Simulating physics with computers. International Journal of Theoretical Physics,
21(6):467–488, 1982.
Timur Garipov, Dmitry Podoprikhin, Alexander Novikov, and Dmitry P. Vetrov. Ultimate tensorization: compressing convolutional and FC layers alike. CoRR, abs/1611.03214, 2016. URL
http://arxiv.org/abs/1611.03214.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing
human-level performance on imagenet classification. In Proceedings of the IEEE international
conference on computer vision, pp. 1026–1034, 2015.
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by
reducing internal covariate shift. In International Conference on Machine Learning, pp. 448–456,
2015.
G. Kin-Lic Chan, J. J. Dorando, D. Ghosh, J. Hachmann, E. Neuscamman, H. Wang, and T. Yanai.
An Introduction to the Density Matrix Renormalization Group Ansatz in Quantum Chemistry.
ArXiv e-prints arXiv:0711.1398, November 2007.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Yoav Levine, David Yakira, Nadav Cohen, and Amnon Shashua. Deep learning and quantum entanglement: Fundamental connections with implications to network design. CoRR, abs/1704.01552,
2017. URL http://arxiv.org/abs/1704.01552.
H. W. Lin, M. Tegmark, and D. Rolnick. Why Does Deep and Cheap Learning Work So Well? Journal of Statistical Physics, 168:1223–1247, September 2017. doi: 10.1007/s10955-017-1836-5.
D. Liu, S.-J. Ran, P. Wittek, C. Peng, R. Blázquez Garcı́a, G. Su, and M. Lewenstein. Machine
Learning by Two-Dimensional Hierarchical Tensor Networks: A Quantum Information Theoretic
Perspective on Deep Architectures. ArXiv e-prints arXiv:1710.04833, October 2017.
Andrew L Maas, Awni Y Hannun, and Andrew Y Ng. Rectifier nonlinearities improve neural network acoustic models. In Proc. ICML, volume 30, 2013.
P. Mehta and D. J. Schwab. An exact mapping between the Variational Renormalization Group and
Deep Learning. ArXiv e-prints arXiv:1410.3831, October 2014.
Alexander Novikov, Dmitry Podoprikhin, Anton Osokin, and Dmitry P. Vetrov. Tensorizing neural
networks. CoRR, abs/1509.06569, 2015. URL http://arxiv.org/abs/1509.06569.
8
R. Orús. A practical introduction to tensor networks: Matrix product states and projected entangled
pair states. Annals of Physics, 349:117–158, October 2014. doi: 10.1016/j.aop.2014.06.013.
I. V. Oseledets. Tensor-train decomposition. SIAM J. Sci. Comput., 33(5):2295–2317, September
2011. ISSN 1064-8275. doi: 10.1137/090752286. URL http://dx.doi.org/10.1137/
090752286.
A. M. Saxe, J. L. McClelland, and S. Ganguli. Exact solutions to the nonlinear dynamics of learning
in deep linear neural networks. ArXiv e-prints arXiv:1312.6120, December 2013.
U. Schollwöck. The density-matrix renormalization group in the age of matrix product states. Annals
of Physics, 326:96–192, January 2011. doi: 10.1016/j.aop.2010.09.012.
Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov.
Dropout: a simple way to prevent neural networks from overfitting. Journal of machine learning
research, 15(1):1929–1958, 2014.
Miles Stoudenmire and D. J. Schwab. Supervised Learning with Quantum-Inspired Tensor Networks. ArXiv e-prints arXiv:1605.05775, May 2016.
F. Verstraete and J. I. Cirac. Matrix product states represent ground states faithfully. Physical Review
B: Condensed Matter and Materials, 73(9):094423, March 2006. doi: 10.1103/PhysRevB.73.
094423.
G. Vidal. Class of Quantum Many-Body States That Can Be Efficiently Simulated. Physical Review
Letters, 101(11):110501, September 2008. doi: 10.1103/PhysRevLett.101.110501.
G. Vidal. Entanglement Renormalization: an introduction. ArXiv e-prints arXiv:0912.1651, December 2009.
9
| 9cs.NE
|
Implementing an Agent Trade Server
Magnus Boman
arXiv:cs/0307064v1 [cs.CE] 29 Jul 2003
Swedish Institute of Computer Science (SICS) AB and The Royal Institute of
Technology, Kista, Sweden
Anna Sandin
Swedish Institute of Computer Science (SICS) AB, Kista, Sweden
Abstract
An experimental server for stock trading autonomous agents is presented and made
available, together with an agent shell for swift development. The server, written
in Java, was implemented as proof-of-concept for an agent trade server for a real
financial exchange.
Key words: Trading agent, Agent server, Financial exchange, Complex order,
Agent programming
1
Introduction
There are always investors seeking to place orders more complicated than
can be accepted by the software of a financial exchange (F/X) system, regardless of its current level of sophistication. To some of these investors, the
agent metaphor is a means to implementing combinatorial, temporal, contingent, or otherwise complex orders. In general, a trading agent is a piece
of encapsulated software that codes the preferences of its owner. In theoretical research, trading agents have been used in idealized games [1], artificial
markets [2], and competitions such as the Trading Agent Competition [3]
(see http://www.sics.se/tac/). To economists, trading agents are speculating noise traders (traders with non-rational expectations and potentially
zero intelligence) [4], often subjected to the Efficient Market Hypothesis and
the Rational Expectations Hypothesis [5], since the empirical evidence against
the accompanying assumptions are directed almost exclusively towards human
traders [6]. In the system of trading agents we consider, the agents interact
with the same order book as the human traders (cf. [7]), resulting in a system only slightly more complex than an F/X is today, but a system that is
Preprint submitted to Elsevier Science
14 September 2017
extremely difficult to analyze (cf., e.g., [8]). From an academic perspective,
the introduction of agents in the actual exchange of real stocks is a formidable
challenge in that it prompts research in real-time control, agent programming,
decision analysis, user interaction, privacy, and security. Its theoretical models must go beyond agent-based computational finance [9] and agent-based
computational economics [10], and already simulation studies are highly complex [11] [12] [13] [14]. From a commercial perspective, the concept is promising
in that it enables implementation of new services in the service portfolio of
the F/X. A vast range of commercial efforts have helped renew and refine
the concept of an electronic F/X, including ambitious liquidity-moving approaches, such as OptiMark [15] [16] [17]. This fact notwithstanding, there
has been no technical platform for real stock trading agent development and
no special-purpose server generally available. We rectify this matter by making available our documented Java code [18] for such a platform. We have also
implemented an experimental server for supporting agent trading of stocks,
an Agent Trade Server (ATS). The conceptual development [19] was initiated
in 2001 (cf. Fig. 1) in cooperation with OM, the world’s largest supplier of
software to stock exchanges [20], and OM has filed a U.S. patent on the ATS
concept [21]. The ATS was built not with commercial operation in mind, but
as proof-of-concept. That said, OM’s trading package SAXESS and its third
party APIs provided for much inspiration. We have not designed the ATS
for any particular market design. Instead, all agents must abide by the rules
of the F/X. The submarket emerging from ATS operations will depend on
the agents’ business logic, including their decision-analytic [22] and normative
modeling [23] of the market, which we have previously investigated. We have
no illusion of ATS traffic accounting for more than one per cent of any stock
exchange before 2010, and whether or not investors will find the ATS attractive will to a large extent depend on pricing and other commercial or political
concerns.
The following section briefly describes the role that software, and agent software in particular, plays at an electronic F/X. In Section 3, the architecture
of our ATS is explained. In Section 4, the issue of logging and other forms of
keeping history is highlighted. The various business roles, i.e. those interested
in the logs, are also discussed. We end with our conclusions and directions for
future research.
2
Development of Financial Exchange Software
Trading of stocks is normally done through a broker, as shown in Fig. 2. The
trader can be a professional investor, a daytrader, or any other person using
stocks as a means to invest. The broker could be a bank, an Internet broker,
or some other financial institution. The figure is simplified: orders can also be
2
Fig. 1. The four phases of the ATS implementation project, with the most fundamental at the bottom, and with all four still running. The Gantt schema also indicates
key breakthroughs in each of the four phases. The second conceptual design will
hopefully be made at OM, or at some other ITC provider for F/X software, while in
itself an interesting academic exercise. The first ATS was made publicly available in
January 2003, but the permission to run agents on the ATS was restricted to select
agent developers at SICS and at two Swedish universities. These developers came
up with a first shell in January, which was refined over the following months, and
sophisticated agents are now under development. As a result of a small empirical
study, not reported on here, the first user service—a trade notification service for
smart phones—was recently implemented.
placed by walking into a bank and filling out a form, IP phones are no requirement, the computer used is often a PDA, etc. Each broker communicates with
the core of the F/X via a gateway, over leased lines (at Stockholmsbörsen, the
Swedish stock exchange, backed up by an ISDN connection, as well as OMnet,
a fibre-optic MetroLAN option). In Fig. 2, the dissemination of information
from the F/X to the outside world is not shown. Suffice to say here that stock
exchange dissemination is an elaborate system of latencies and routing, in order to secure fairness, and that such information is an important part of the
feedback from core to broker, and from broker to trader.
Suppliers of software to an electronic exchange system are basically of two
kinds: those that provide software serving the core (the order book), and those
that provide software for those interacting with the core (the traders). In most
countries, the software supplier in the first category is also competing in the
second category. In Sweden, for example, OM provides all of the software for
the core, and they also provide the SAXESS and CLICK (for the derivatives
3
Fig. 2. A schematic view of the current trade information flow. For simplicity, dissemination procedures are not shown.
market, which we will neglect entirely at this stage) software trading packages,
as well as the SECUR software clearing and back office package. The bulk of
the software aimed at traders and brokers is however developed by third party
software houses. We would like to complicate Fig. 2 slightly (see Fig. 3) by
considering individuals as service provisioners [24], thus introducing a third
kind of code developer. We anticipate that end-user code development in the
form of trading agents will be an increasingly important part of the future
F/X service portfolio.
The third party code pertains to the trading applications used. In order to
use the SAXESS system today, the broker needs a member SAXESS trade
application. Each such application must be certified, to guarantee smooth operation alongside all other running applications. The OM application offered
is SAXESS Trade, an NT-based client-server solution with the trading data
being maintained on an SQL server. There are currently about two dozen certified alternatives. Each application connects to a SAXESS trade server, which
is in turn connected to the core. The transaction protocol used is XTP (open
eXchange Transaction Protocol) 2.40, an OSI Layer 7 protocol. The Session
Layer protocol used is XMP (eXchange Message Protocol). The transactions
involving the core are governed by a C program, essentially matching the stack
of sell orders with the stack of buy orders. It is important to note that the
matching is deterministic. The ATS concept is not a suggested alternative
4
Fig. 3. Three kinds of code development.
to the SAXESS (or any other) trade server currently in use, but a complement. In Fig. 4, the ordinary electronic trading is represented by SAXESS.
Upward arrows denote the placing of orders while downward arrows denote
feedback. Each trade application uses zero or more distributed instances of
the SAXESS Trade Server, a freeware available to all registered traders. In
practice, the ATS will also contain an instance of the SAXESS Trade Server,
since the latter contains the latest API for communicating with the core. All
communication with the core goes via the Front End (FE). The wall indicates
the physical and logical boundary between the market owner and the outside
world, while the barbed wire surrounding the core illustrates a further level of
protection. The shadow backup system is not shown.
Note that brokers may still be used, for example for legal reasons, when delegating trading to an agent. While delegating the right to place orders in itself
does not introduce non-determinism in the marketplace, there is nothing that
per se stops an agent developer to introduce non-determinism in the business
logic of the agent. Security control of agent code is done at the time of agent
certification and testing, and also online in the ATS.
In SAXESS, each trade application must also connect to a dedicated Distributed Dissemination Server (DDS). With an ATS running together with a
SAXESS trade server, the owner of the F/X has the option to either route all
traffic feedback through the existing DDS channels, or to complement the
5
Fig. 4. The ATS as a complement to ordinary electronic trading.
6
DDS with ATS traffic feedback. In particular, the agents running on the
ATS could subscribe to ATS traffic feedback of various sorts, and at variable cost. Even non-professional traders are no strangers to varying levels of
sophistication of the individually specified trading constraints built on feedback. The stop loss service offered by Internet brokers (such as Avanza; see
http://www.avanza.com) is a case in point.
The ATS concept provides the third party software houses with a business opportunity, viz. to provide code for the agents trading on the ATS. This involves
coding investor preferences and encapsulating these into agents, whereafter the
agents are placed on the ATS, without disclosure. Preference elicitation is as
always difficult, but suffice to say here that investors will come in many varieties, from those that have developed advanced agents on their own, to those
that have no clue on agent programming. A template agent can be coded
based on the investor filling out a fairly simple questionnaire. The size and
structure of such template agents is simple, and the first agent shell for our
ATS has already been written [25]. After the agent is placed on the ATS,
there remains the challenge of providing software services for notification and
(depending on security issues) possibly for termination of agents. Investors
no doubt require interfaces for a variety of modalities, and will in the future
include roaming [26].
3
An Agent Trade Server Architecture
We developed the ATS in-house at SICS, in order to allow for free experimentation with agent code and simultaneously adjusting and adding code to the
server. A jar-file containing an agent can be run locally or remotely. Each trading agent executes in its own thread. The three server packages implemented
(cf. Fig. 5) are described in turn below. The choice of Java was motivated,
among several factors, by smooth compatibility with small devices, such as
Java-enabled cell phones. While there is no theoretical reason to limit any
sandbox environment for agent development to only one language, Java currently seems the most practical first choice. For instance, Java’s policy files are
useful for defining the boundary between ATS and agent, and java.security
allows for basic signing of agent code.
3.1 se.sics.ats.core
This package represents the interface towards the agents, and is distributed to
agent developers. Initially, the list of agent developers will most likely overlap
considerably with the list of certified SAXESS members, and the principles
7
Fig. 5. The three ATS packages implemented.
for testing similar. The next step is then to automate the certification process
so that individual agent developers can submit agents to the ATS. Since there
are legal and practical reasons for the market owner not to certify every single
developer, there is room here for traditional as well as agent brokerage. The
class AgentComponent specifies all ATS interfacing demands on the agent. For
instance, the agent must have functionality for starting and stopping. The class
AgentContext is the agent’s handle on the server. AgentContext is inspired
by and adheres to the SAXESS Trade API for third party code development,
based on Microsoft’s Component Object Model (COM).
3.2 se.sics.ats.reference
This package contains the implementation of the ATS itself. It contains log
handlers and socket tranceivers.
8
3.3 se.sics.ats.data
This package represents the interface towards the stock exchange. This is
where the ATS will connect to the core of the F/X. For now, it contains
a parser which, like the connection, can easily be replaced by another. It
now parses data disseminated through the WWW (more specifically, it parses
http://www.stockholmsborsen.se). The ATS thus uses real stock exchange
data, albeit with a delay of up to 15 minutes. We have limited its capability
at the outset to the most traded stocks on the Stockholmsbörsen A-list. The
package also simulates an order book.
4
Agent Shell Execution
We can now specify the “Agent” rectangle from Fig. 5 above by explaining the
agent shell. We first turn to some more abstract aspects of ATS management,
however.
4.1 Roles
In our proof-of-concept implementation work, we have felt a need for a division
of labor between agent developers and ourselves: the ATS developers. In order
to leave the core software unchanged—a requirement from OM, and most likely
from any market owner—we also need to define the role of ATS administrator.
The regulation of trade in SAXESS is today done through surveillance of the
Trading Engine (SaxView), as well as through the official trading control and
supervisory functions units. These units operate on the core and will be used
also for agent trading. In addition, the ATS administrator must store and sign
certificates for authenticated agent code developers.
In addition, there are new business roles in the future scenario of commercially available ATS services, perhaps including dedicated agent brokers. We
have enabled this new business role by implementing an object relating agent
brokers and their agents. This object maintains references, deducts brokerage
fees, and supplies data for notification services. For security reasons, the agent
broker object is placed in the same execution environment as the ATS. The
only agent manipulation granted the object is the killing of an agent thread,
should this become necessary.
9
Fig. 6. Excerpt from a log of Daytrader, a simple agent running on the ATS.
10
4.2 Logs
The ATS administrator must monitor the actions of all agents on the ATS,
and activity logs are therefore needed. These logs are likely to be processed
only automatically, but in case of system failure or suspected illicit agent action, manual inspection may be required. For this reason, the responsibility
for logging should not be distributed, but be the responsibility of the ATS
administrator. Each agent should be allowed read and write access to its own
action logs, since logs provide interesting information on top of the generally available trade statistics. Logged information could be used for machine
learning, which could occur offline or online, depending on the agent business
logic. An excerpt from a log of a simple agent, to be discussed further in the
next subsection, is shown in Fig. 6. The full source code of the agent, as well
as the full log, has been published [25]. Since all stakeholders are potentially
interested in parsing the log, it is in XML format. For instance, the ATS administrator might like to find the agent responsible for some illegal action on
the ATS. Analogously, an agent broker might seek to attribute errors to the
ATS implementation, using logs as evidence. The error log of Java exceptions
and String objects is kept separate from all other logged data.
4.3 se.sics.ats.agent
The structurally designed agent shell packages are intended as support for ATS
agent development, and their JavaDoc is publicly available [27]. We give an
abstract overview here of the low coupling design, so that the role of the shell
in our implementation of the ATS is made clear. Agents extend the abstract
TradeAgent class (see Fig. 7), which handles all ATS interaction, including the
final methods initialize and start, invoked by the ATS [18]. The agent
(SubAgent) gets information about stocks through various get methods, or by
subscribing to stock changes (cf. Fig. 6). The latter alternative is the one used
by most non-artificial stock traders today, and so subscription procedures and
costs have already been carefully designed at most exchanges.
Agents place orders using the Order object, and its interface OrderListener [18].
An order is created by calling the method createOrder. The OrderListener
then implements the methods orderCancelled and orderClosed, which result in account updates and logging, as demonstrated in Fig. 6. The account
and portfolio maintenance should in any commercial implementation of the
ATS be handled by standard F/X middleware. Similarly, the AgentBroker
has only minimal functionality in this version of the ATS.
11
Fig. 7. The agent shell package structure.
5
Conclusions and Further Research
We have described a proof-of-concept implementation of a running agent trade
server, which we believe serves as a blueprint for future implementations intended for live use on a real financial exchange. We have also constructed and
described an agent shell, currently in use for server implementation of agents
of increasing complexity. These agent developers are also providing the first
anecdotal evidence of the quality of our implementation. We intend to engage more developers before a formal evaluation is made. Our future work (cf.
Fig. 1) includes demonstrating that the size of an agent with a sophisticated
business logic is manageable. We have also implemented various demonstrators for notifying services, to be used for presenting agent logs to the agent
brokers and agent owners [28]. While professional trading is a highly collaborative activity [29], the agents in our set-up model each other only in a weak
sense, based solely on their bidding. There are several multi-agent system aspects that deserve attention, such as multiple agent submission for teamwork,
or even for manipulative purposes. Finally, the effects on the entire financial
system as a result of wide adoption of agent trade is under investigation. This
work is being done together with the Stockholm School of Economics and OM.
Acknowledgments
The authors would like to thank Jesper Johansson and Michael Poijes, who
wrote the first agent shell and some of the server code, and kindly provided
12
log data. David Lybäck and Ulf Essler contributed with important comments
and fruitful discussions. The Vinnova project TAP on accessible autonomous
software provided the authors with the time required for this study.
References
[1] J. Rust, J. H. Miller, R. Palmer, Characterizing effective trading strategies–
insights from a computerized double auction tournament, Economic Dynamics
and Control 18 (1994) 61–96.
[2] R. G. Palmer, W. B. Arthur, J. H. Holland, B. LeBaron, P. Tayler, Artificial
economic life—a simple model of a stockmarket, Physica D 75 (1–3) (1994)
264–274.
[3] A. Greenwald, The 2002 trading agent competition—an overview of agent
strategies, AI Magazine 24 (1) (2003) 83–91.
[4] J. B. D. Long, A. Shleifer, L. H. Summers, R. Waldmann, The survival of noise
traders in financial markets, Journal of Business 64 (1991) 1–19.
[5] T. Lux, M. Ausloos, Market fluctuations I: Scaling, multi-scaling and their
possible origins, in: A. Bunde, J. Kropp, H.-J. Schellnhuber (Eds.), Science
of Disaster—Scaling Laws Governing Weather, Body, Stock-Market Dynamics,
Springer-Verlag, 2002, pp. 373–409.
[6] T. Lux, Herd behaviour, bubbles and crashes, The Economic Journal 105 (1995)
881–896.
[7] M. Boman, S. Johansson, D. Lybäck, Parrondo strategies for artificial traders,
in: N. Zhong, J. Liu, S. Ohsuga, J. Bradshaw (Eds.), Intelligent Agent
Technology, World Scientific, Singapore, 2001, pp. 150–159.
[8] P. Jefferies, M. Hart, P. Hui, N. Johnson, From market games to real-world
markets, European Physical Journal B 20 (1) (2001) 493–501.
[9] B. LeBaron, Agent-based computational finance–suggested readings and early
research, Economic Dynamics and Control 24 (2000) 679–702.
[10] L. Tesfatsion, Agent-based computational economics–growing economies from
the bottom up, Artificial Life 8 (1) (2002) 55–82.
[11] M. Lettau, Explaining the facts with adaptive agents–the case of mutual fund
flows, Journal of Economic Dynamics and Control 21 (1997) 1117–1147.
[12] T. Lux, M. Marchesi, Scaling and criticality in a stochastic multi-agent model
of a financial market, Nature 397 (1999) 498–500.
[13] K. Bertels, M. Boman, Agent-based social simulation in markets, Electronic
Commerce Research 1 (1–2) (2001) 149–158.
13
[14] J. D. Farmer, S. Joshi, The price dynamics of common trading strategies,
Journal of Economic Behavior and Organization 49 (2002) 149–171.
[15] W. A. Lupien, J. T. Rickard, Crossing Network Utilizing Optimal Mutual
Satisfaction Density Profile, uS Patent No. 5689652 (1997).
[16] E. K. Clemons, B. W. Weber, Restructuring institutional block trading—an
overview of the OptiMark system, in: Proc HICSS 6—Organizational Systems
and Technology, IEEE Computer Society, 1998, pp. 301–310.
[17] C. Gerber, J. Teich, H. Wallenius, J. Wallenius, A simulation of OptiMark’s
electronic matching algorithm and its simple variations of instutional block
trading, Decision Support Systems In press.
[18] A.
Sandin,
AgentTradeServer
javaDoc,
http://www.sics.se/~sandin/ats/doc/ (2002).
Available
at
[19] D. Lybäck, M. Boman, Agent trade servers in financial exchange systems, ACM
Transactions on Internet Technology 3 (4), in press.
[20] R. Sales, First Europe, now the world, Wall Street and Technology online,
available
at http://www.wallstreetandtech.com/story/inDepth/WST20010816S0009
(August 2001).
[21] An Automated Semi-Deterministic Trading System, filed to US Patent and
Trademark Office (March 2002).
[22] L. Ekenberg, M. Danielson, M. Boman, Imposing security constraints on agentbased decision support, Decision Support Systems 20 (1) (1997) 3–15.
[23] M. Boman, Norms in artificial decision making, Artificial Intelligence and Law
7 (1999) 17–35.
[24] F. Espinoza, Individual service provisioning, Ph.D. thesis, Dept of Computer
and Systems Sciences, Stockholm Univ (2003).
[25] J. Johansson, M. Poijes, Agent shell for stock market systems, Master’s thesis,
Dept of Computer and Systems Sciences, Stockholm Univ (2003).
[26] M. Boman, M. Bylund, F. Espinoza, M. Danielson, D. Lybäck, Trading agents
for roaming users, presented at the Tokyo Mobile Roundtable, May 2002.
Available at http://www.arxiv.org/abs/cs.CE/0204056 (May 2002).
[27] J. Johansson, M. Poijes, DayTrader agent javaDoc, Available
http://www.dsv.su.se/~mich-poi/trade_agent/javadoc/ (2003).
at
[28] S. Nylander, M. Bylund, M. Boman, Mobile access to real-time information—
the case of autonomous stock brokering, submitted. (2003).
[29] C. Heath, M. Jirotka, P. Luff, J. Hindmarsh, Unpacking collaboration—the
interactional organisation of trading in a city dealing room, in: G. D. Michelis,
C. Simone, K. Schmidt (Eds.), Proc ECSCW’93, Kluwer Academic, 1993, pp.
155–170.
14
| 5cs.CE
|
Real-Time Stochastic Optimal Control for Multi-agent Quadrotor Systems
Vicenç Gómez1 , Sep Thijssen2 , Andrew Symington3 , Stephen Hailes4 , Hilbert J. Kappen2
arXiv:1502.04548v5 [cs.SY] 10 Mar 2016
2
1
Universitat Pompeu Fabra. Barcelona, Spain [email protected]
Radboud University Nijmegen, the Netherlands {s.thijssen,b.kappen}@donders.ru.nl
3
University of California Los Angeles, USA [email protected]
4
University College London, United Kingdom [email protected]
Abstract
This paper presents a novel method for controlling teams
of unmanned aerial vehicles using Stochastic Optimal Control (SOC) theory. The approach consists of a centralized
high-level planner that computes optimal state trajectories as
velocity sequences, and a platform-specific low-level controller which ensures that these velocity sequences are met.
The planning task is expressed as a centralized path-integral
control problem, for which optimal control computation corresponds to a probabilistic inference problem that can be
solved by efficient sampling methods. Through simulation
we show that our SOC approach (a) has significant benefits
compared to deterministic control and other SOC methods
in multimodal problems with noise-dependent optimal solutions, (b) is capable of controlling a large number of platforms
in real-time, and (c) yields collective emergent behaviour in
the form of flight formations. Finally, we show that our approach works for real platforms, by controlling a team of
three quadrotors in outdoor conditions.
1
Introduction
The recent surge in autonomous Unmanned Aerial Vehicle
(UAV) research has been driven by the ease with which platforms can now be acquired, evolving legislation that regulates their use, and the broad range of applications enabled by both individual platforms and cooperative swarms.
Example applications include automated delivery systems,
monitoring and surveillance, target tracking, disaster management and navigation in areas inaccessible to humans.
Quadrotors are a natural choice for an experimental platform, as they provide a safe, highly-agile and inexpensive
means by which to evaluate UAV controllers. Figure 1 shows
a 3D model of one such quadrotor, the Ascending Technologies Pelican. Quadrotors have non-linear dynamics and are
naturally unstable, making control a non-trivial problem.
Stochastic optimal control (SOC) provides a promising
theoretical framework for achieving autonomous control of
quadrotor systems. In contrast to deterministic control, SOC
directly captures the uncertainty typically present in noisy
environments and leads to solutions that qualitatively depend on the level of uncertainty (Kappen 2005). However,
with the exception of the simple Linear Quadratic Gaussian
Copyright c 2016, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
Figure 1: Control hierarchy: The path-integral controller
(1) calculates target velocities/heights for each quadrotor.
These are converted to roll, pitch, throttle and yaw rates by
a platform-specific Velocity Height PID controller (2). This
control is in turn passed to the platform’s flight control system (3), and converted to relative motor speed changes.
case, for which a closed form solution exists, solving the
SOC problem requires solving the Hamilton Jacobi Bellman
(HJB) equations. These equations are generally intractable,
and so the SOC problem remains an open challenge.
In such a complex setting, a hierarchical approach is usually taken and the control problem is reduced to follow a
state-trajectory (or a set of way points) designed by hand
or computed offline using trajectory planning algorithms
(Kendoul 2012). While the planning step typically involves
a low-dimensional state representation, the control methods
use a detailed complex state representation of the UAV. Examples of control methods for trajectory tracking are the
Proportional Integral Derivative or the Linear-Quadratic regulator.
A generic class of SOC problems was introduced in Kappen; Todorov (2005; 2006) for which the controls and the
cost function are restricted in a way that makes the HJB
equation linear and therefore more efficiently solvable. This
class of problems is known as path integral (PI) control,
linearly-solvable controlled diffusions or Kullback-Leibler
control, and it has lead to successful robotic applications,
e.g. (Kinjo, Uchibe, and Doya 2013; Rombokas et al. 2012;
Theodorou, Buchli, and Schaal 2010). A particularly interesting feature of this class of problems is that the computation of optimal control is an inference problem with a solution given in terms of the passive dynamics. In a multiagent system, where the agents follow independent passive
dynamics, such a feature can be exploited using approximate inference methods such as variational approximations
or belief propagation (Kappen, Gómez, and Opper 2012;
Van Den Broek, Wiegerinck, and Kappen 2008).
In this paper, we show how PI control can be used for
solving motion planning tasks on a team of quadrotors in
real time. We combine periodic re-planning with receding
horizon, similarly to model predictive control, with efficient importance sampling. At a high level, each quadrotor is modelled as a point mass that follows simple double integrator dynamics. Low-level control is achieved using a standard Proportional Integral Derivative (PID) velocity controller that interacts with a real or simulated flight
control system. With this strategy we can scale PI control to
ten units in simulation. Although in principle there are no
further limits to experiments with actual platforms, our first
results with real quadrotors only include three units. To the
best of our knowledge this has been the first real-time implementation of PI control on an actual multi-agent system.
In the next section we describe related work. We introduce
our approach in Section 3 Results are shown on three different scenarios in Section 4 Finally, Section 5 concludes
this paper.
2
Related Work on UAV Planning and
Control
There is a large and growing body of literature related to
this topic. In this section, we highlight some of the most
related papers to the presented approach. An recent survey of control methods for general UAVs can be found
in Kendoul (2012).
Stochastic optimal control is mostly used for UAV control in its simplest form, assuming a linear model perturbed
by additive Gaussian noise and subject to quadratic costs
(LQG), e.g. (How et al. 2008). While LQG can successfully
perform simple actions like hovering, executing more complex actions requires considering additional corrections for
aerodynamic effects such as induced power or blade flapping (Hoffmann et al. 2011). These approaches are mainly
designed for accurate trajectory control and assume a given
desired state trajectory that the controller transforms into
motor commands.
Model Predictive Control (MPC) has been used optimize
trajectories in multi-agent UAV systems (Shim, Kim, and
Sastry 2003). MPC employs a model of the UAV and solves
an optimal control problem at time t and state x(t) over a
future horizon of a fixed number of time-steps. The first optimal move u∗ (t) is then applied and the rest of the optimal
sequence is discarded. The process is repeated again at time
t + 1. A quadratic cost function is typically used, but other
more complex functions exist.
MPC has mostly been used in indoor scenarios, where
high-precision motion capture systems are available. For instance, in Kushleyev et al. (2013) authors generate smooth
trajectories through known 3-D environments satisfying
specifications on intermediate waypoints and show remarkable success controlling a team of 20 quadrotors. Trajectory
optimization is translated to a relaxation of a mixed integer quadratic program problem with additional constraints
for collision avoidance, that can be solved efficiently in realtime. Examples that follow a similar methodology can be
found in Turpin, Michael, and Kumar; Augugliaro, Schoellig, and D’Andrea (2012; 2012). Similarly to our approach,
these methods use a simplified model of dynamics, either using the 3-D position and yaw angle Kushleyev et al.; Turpin,
Michael, and Kumar (2013; 2012) or the position and velocities as in Augugliaro, Schoellig, and D’Andrea (2012).
However, these approaches are inherently deterministic and
express the optimal control problem as a quadratic problem.
In our case, we solve an inference problem by sampling and
we do not require intermediate trajectory waypoints.
In outdoor conditions, motion capture is difficult and
Global Positioning System (GPS) is used instead. Existing
control approaches are typically either based on Reynolds
flocking (Bürkle, Segor, and Kollmann 2011; Hauert et al.
2011; Vásárhelyi et al. 2014; Reynolds 1987) or flight formation (Guerrero and Lozano 2012; Yu et al. 2013). In
Reynolds flocking, each agent is considered a point mass
that obeys simple and distributed rules: separate from neighbors, align with the average heading of neighbors and steer
towards neighborhood centroid to keep cohesion. Flight formation control is typically modeled using graphs, where every node is an agent that can exchange information with all
or several agents. Velocity and/or position coordination is
usually achieved using consensus algorithms.
The work in Quintero, Collins, and Hespanha (2013)
shares many similarities with our approach. Authors derive a
stochastic optimal control formulation of the flocking problem for fixed-wings UAVs. They take a leader-follower strategy, where the leader follows an arbitrary (predefined) policy that is learned offline and define the immediate cost as
a function of the distance and heading with respect to the
leader. Their method is demonstrated outdoors with 3 fixedwing UAVs in a distributed sensing task. As in this paper,
they formulate a SOC problem and perform MPC. However,
in our case we do not restrict to a leader-follower setup and
consider a more general class of SOC problems which can
include coordination and cooperation problems.
Planning approaches Within the planning community,
Bernardini, Fox, and Long (2014) consider search and tracking tasks, similar to one of our scenarios. Their approach
is different to ours, they formulate a planning problem that
uses used search patterns that must be selected and se-
quenced to maximise the probability of rediscovering the
target. Albore et al. (2015) and Chanel, Teichteil-Knigsbuch,
and Lesire (2013) consider a different problem: dynamic
data acquisition and environmental knowledge optimisation.
Both techniques use some form of replanning. While Albore et al. (2015) uses a Markov Random Field framework
to represent knowledge about the uncertain map and its quality, Chanel, Teichteil-Knigsbuch, and Lesire (2013) rely on
partially-observable MDPs. All these works consider a single UAV scenario and low-level control is either neglected
or deferred to a PID or waypoint controller.
bitrary functions, f is the drift in the uncontrolled dynamics
(including gravity, Coriolis and centripetal forces), and G
describes the effect of the control u into the state vector x.
A realization τ = x0:dt:T of the above equation is called
a (random) path. In order to describe a control problem we
define the cost that is attributed to a path (cost-to-go) by
Recent Progress in Path-Integral Control There has
been significant progress in PI control, both theoretically and
in applications. Most of existing methods use parametrized
policies to overcome the main limitations (see Section 3.1).
Examples can be found in Theodorou, Buchli, and Schaal;
Stulp and Sigaud; Gómez et al. (2010; 2012; 2014). In
these methods, the optimal control solution is restricted by
the class of parametrized policies and, more importantly,
it is computed offline. In Rawlik, Toussaint, and Vijayakumar (2013), authors propose to approximate the transformed
cost-to-go function using linear operators in a reproducing
kernel Hilbert space. Such an approach requires an analytical form of the PI embedding, which is difficult to obtain
in general. In Horowitz, Damle, and Burdick (2014), a lowrank tensor representation is used to represent the model dynamics, allowing to scale PI control up to a 12-dimensional
system. More recently, the issue of state-dependence of the
optimal control has been addressed (Thijssen and Kappen
2015), where a parametrized state-dependent feedback controller is derived for the PI control class.
Finally, model predictive PI control has been recently proposed for controlling a nano-quadrotor in indoor settings
in an obstacle avoidance task (Williams, Rombokas, and
Daniel 2014). In contrast to our approach, their method is
not hierachical and uses naive sampling, which makes it less
sample efficient. Additionally, the control cost term is neglected, which can have important implications in complex
tasks involving noise. The approach presented here scales
well to several UAVs in outdoor conditions and is illustrated
in tasks beyond obstacle avoidance navigation.
where rT (xT ) and rt (xt ) are arbitrary state cost terms at
end and intermediate times, respectively. R is the control
cost matrix. The general stochastic optimal control problem
is to minimize the expected cost-to-go w.r.t. the control
3
Path-Integral Control for Multi-UAV
planning
We first briefly review PI control theory. This is followed by
a description of the proposed method used to achieve motion
planning of multi-agent UAV systems using PI control.
3.1
Path-Integral Control
We consider continuous time stochastic control problems,
where the dynamics and cost are respectively linear and
quadratic in the control input, but arbitrary in the state.
More precisely, consider the following stochastic differential
equation of the state vector x ∈ Rn under controls u ∈ Rm
dx = f (x)dt + G(x)(udt + dξ),
(1)
where ξ is m−dimensional Wiener noise with covariance
Σu ∈ Rm×m and f (x) ∈ Rn and G(x) ∈ Rn×m are ar-
S(τ |x0 , u) = rT (xT )
X
1 >
+
rt (xt )dt + ut Rut dt,
2
(2)
t=0:dt:T −dt
u∗ = arg min E[S(τ |x0 , u)].
u
In general, such a minimization leads to the HamiltonJacobi-Bellman (HJB) equations, which are non-linear, second order partial differential equations. However, under the
following relation between the control cost and noise covariance Σu = λR−1 , the resulting equation is linear in the
exponentially transformed cost-to-go function. The solution
is given by the Feynman-Kac Formula, which expresses optimal control in terms of a Path-Integral, which can be interpreted as taking the expectation under the optimal path
distribution (Kappen 2005)
p∗ (τ |x0 ) ∝ p(τ |x0 , u) exp(−S(τ |x0 , u)/λ),
hu∗t (x0 )i = hut + (ξt+dt − ξt )/dti ,
(3)
(4)
where p(τ |x0 , u) denotes the probability of a (sub-optimal)
path under equation (1) and h·i denotes expectation over
paths distributed by p∗ .
The constraint Σu = λR−1 forces control and noise to act
in the same dimensions, but in an inverse relation. Thus, for
fixed λ, the larger the noise, the cheaper the control and viceversa. Parameter λ act as a temperature: higher values of λ
result in optimal solutions that are closer to the uncontrolled
process.
Equation (4) permits optimal control to be calculated by
probabilistic inference methods, e.g., Monte Carlo. An interesting fact is that equations (3, 4) hold for all controls
u. In particular, u can be chosen to reduce the variance in
the Monte Carlo computation of hu∗t (x0 )i which amounts
to importance sampling. This technique can drastically improve the sampling efficiency, which is crucial in high dimensional systems. Despite this improvement, direct application of PI control into real systems is limited because it is
not clear how to choose a proper importance sampling distribution. Furthermore, note that equation (4) yields the optimal control for all times t averaged over states. The result
is therefore an open-loop controller that neglects the statedependence of the control beyond the initial state.
3.2
Multi-UAV planning
The proposed architecture is composed of two main levels.
At the most abstract level, the UAV is modeled as a 2D
point-mass system that follows double integrator dynamics.
Algorithm 1 PI control for UAV motion planning
1: function PIC ONTROLLER(N, H, dt, rt (·), Σu , ut:dt:t+H )
2:
for k = 1, . . . , N do
3:
Sample paths τk = {xt:dt:t+H }k with Eq. (5)
4:
end for
5:
Compute Sk = S(τk |x0 , u) with Eq. (2)
6:
Store the noise realizations {ξt:dt:t+H P
}k
7:
Compute the weights: wk = e−Sk /λ / l e−Sl /λ
8:
for s = t : dt : t +P
H do
1
9:
u∗s = us + dt
k wk ({ξs+dt }k − {ξs }k )
10:
end for
11:
Return next desired velocity: vt+dt = vt + u∗t dt
and u∗t:dt:t+H for importance sampling at t + dt
12: end function
At the low-level, we use a detailed second order model that
we learn from real flight data (De Nardi and Holland 2008).
We use model predictive control combined with importance
sampling. There are two main benefits of using the proposed
approach: first, since the state is continuously updated, the
controller does not suffer from the problems caused by using an open-loop controller. Second, the control policy is not
restricted by any parametrization.
The two-level approach permits to transmit control signals from the high-level PI controller to the low-level control
system at a relatively low frequencies (we use 15Hz in this
work). Consequently, the PI controller has more time available for sampling a large number of trajectories, which is
critical to obtain good estimates of the control. The choice of
2D in the presented method is not a fundamental limitation,
as long as double-integrator dynamics is used. The control
hierarchy introduces additional model mismatch. However,
as we show in the results later, this mismatch is not critical
for obtaining good performance in real conditions.
Ignoring height, the state vector x is thus composed of the
East-North (EN) positions and EN velocities of each agent
i = 1, . . . , M as xi = [pi , vi ]> where pi , vi ∈ R2 . Similarly,
the control u consists of EN accelerations ui ∈ R2 . Equation (1) decouples between the agents and takes the linear
form
dxi = Axi dt + B(ui dt + dξi ),
0 1
0
A=
,
B=
.
0 0
1
(5)
Notice that although the dynamics is decoupled and linear,
the state cost rt (xt ) in equation (2) can be any arbitrary
function of all UAVs states. As a result, the optimal control will in general be a non-linear function that couples all
the states and thus hard to compute.
Given the current joint optimal action u∗t and velocity
vt , the expected velocity at the next time t0 is calculated as
vt0 = vt + (t0 − t)u∗t and passed to the low-level controller.
The final algorithm optionally keeps an importance-control
sequence ut:dt:t+H that is incrementally updated. We summarize the high-level controller in Algorithm 1.
The importance-control sequence ut:dt:t+H is initialized
using prior knowledge or with zeros otherwise. Noise is
Figure 2: The flight control system (FCS) is comprised of
two control loops: one for stabilization and the other for pose
control. A low-level controller interacts with the FCS over a
serial interface to stream measurements and issue control.
dimension-independent, i.e. Σu = σu2 Id. To measure sampling convergence, we define the Effective Sample Size
PN
(ESS) as ESS := 1/ k=1 wk2 , which is a quantity between
1 and N . Values of ESS close to one indicate an estimate
dominated by just one sample and a poor estimate of the
optimal control, whereas an ESS close to N indicates near
perfect sampling, which occurs when the importance- equals
the optimal-control function.
3.3
Low Level Control
>
The target velocity v = [vE vN ] is passed along with
a height p̂U to a Velocity-Height controller. This controller uses the current state estimate of the real quadrotor
>
y = [pE pN pU φ θ ψ u v w p q r] , where (pE , pN , pU )
and (φ, θ, ψ) denote navigation-frame position and orientation and (u, v, w), (p, q, r) denote body-frame and angular
velocities, respectively. It is composed of four independent
PID controllers for roll φ̂, pitch θ̂, throttle γ̂ and yaw rate r̂.
that send the commands to the flight control system (FCS)
to achieve v.
Figure 2 shows the details of the FCS. The control loop
runs at 1kHz fusing triaxial gyroscope, accelerometer and
magnetometer measurements. The accelerometer and magnetometer measurements are used to determine a reference
global orientation, which is in turn used to track the gyroscope bias. The difference between the desired and actual
angular rates are converted to motor speeds using the model
in Mahony, Kumar, and Corke (2012).
An outer pose control loop calculates the desired angular
rates based on the desired state. Orientation is obtained from
the inner control loop, while position and velocity are obtained by fusing GPS navigation fixes with barometric pressure (BAR) based altitude measurements. The radio transmitter (marked TX in the diagram) allows the operator to
switch quickly between autonomous and manual control of
a platform. There is also an acoustic alarm on the platform
itself, which warns the operator when the GPS signal is lost
or the battery is getting low. If the battery reaches a critical
level or communication with the transmitter is lost, the platform can be configured to land immediately or alternatively,
to fly back and land at its take-off point.
horizon problem. A key difference between iLQG and PI
control is that the linear-quadratic approximation is certainty
equivalent. Consequently, iLQG yields a noise independent
solution.
4.1
Figure 3: Drunken Quadrotor: a red target has to be reached
while avoiding obstacles. (Left) the shortest route is the optimal solution in the absence of noise. (Right) with control
noise, the optimal solution is to fly around the building.
3.4
Simulator Platform
We have developed an open-source framework called
CRATES1 . The framework is a implementation of QRSim
(De Nardi 2013; Symington et al. 2014) in Gazebo, which
uses Robot Operating System (ROS) for high-level control.
It permits high-level controllers to be platform-agnostic. It is
similar to the Hector Quadrotor project (Meyer et al. 2012)
with a formalized notion of a hardware abstraction layers.
The CRATES simulator propagates the quadrotor state
forward in time based on a second order model (De Nardi
and Holland 2008). The equations were learned from real
flight data and verified by expert domain knowledge. In addition to platform dynamics, CRATES also simulates various noise-perturbed sensors, wind shear and turbulence.
Orientation and barometric altitude errors follow zero-mean
Ornstein-Uhlenbeck processes, while GPS error is modeled
at the pseudo range level using trace data available from
the International GPS Service. In accordance with the Military Specification MIL-F-8785C, wind shear is modeled as
a function of altitude, while turbulence is modeled as a discrete implementation of the Dryden model. CRATES also
provides support for generating terrain from satellite images
and light detection and ranging (LIDAR) technology, and
reporting collisions between platforms and terrain.
4
Results
We now analyze the performance of the proposed approach
in three different tasks. We first show that, in the presence of
control noise, PI control is preferable over other approaches.
For clarity, this scenario is presented for one agent only. We
then consider two tasks involving several units: a flight formation task and a pursuit-evasion task.
We compare the PI control method described in Section
3.2 with iterative linear-quadratic Gaussian (iLQG) control
(Todorov and Li 2005). iLQG is a state-of-the-art method
based on differential dynamic programming, that iteratively
computes local linear-quadratic approximations to the finite
1
CRATES stands for ’Cognitive Robotics Architecture for
Tightly-Coupled Experiments and Simulation’. Available at
https://bitbucket.org/vicengomez/crates
Scenario I: Drunken Quadrotor
This scenario is inspired in Kappen (2005) and highlights the
benefits of SOC in a quadrotor task. The Drunken Quadrotor is a finite horizon task where a quadrotor has to reach a
target, while avoiding a building and a wall (figure 3). There
are two possible routes: a shorter one that passes through a
small gap between the wall and the building, and a longer
one that goes around the building. Unlike SOC, the deterministic optimal solution does not depend on the noise level
and will always take the shorter route. However, with added
noise, the risk of collision increases and thus the optimal
noisy control is to take the longer route.
This task can be alternatively addressed using other planning methods, such as the one proposed by Ono, Williams,
and Blackmore (2013), which allow for specification of
user’s acceptable levels of risk using chance constraints.
Here we focus on comparing deterministic and stochastic
optimal control for motion planning. The amount of noise
thus determines whether the optimal solution is go through
the risky path or the longer safer path.
The state cost in this problem consists of hard constraints
that assign infinite cost when either the wall or the building is hit. PI control deals with collisions by killing particles that hit the obstacles during Monte Carlo sampling.
For iLQG, the local approximations require a twice differentiable cost function. We resolved this issue by adding a
smooth obstacle-proximity penalty in the cost function. Although iLQG computes linear feedback, we tried to improve
it with a MPC scheme, similar as for PI control. Unfortunately, this leads to numerical instabilities in this task, since
the system disturbances tend to move the reference trajectory through a building when moving from one time step to
the next. For MPC with PI control we use a receding horizon
of three seconds and perform re-planning at a frequency of
15 Hz with N = 2000 sample paths. Both methods are initialized with ut = 0, ∀t. iLQG requires approximately 103
iterations to converge with a learning rate of 0.5%.
Figure 3 (left) shows an example of real trajectory computed for low control noise level, σu2 = 10−3 . To be able
to obtain such a trajectory we deactivate sensor uncertainties in accelerometer, gyroscope, orientation and altimeter.
External noise is thus limited to aerodynamic turbulences
only. In this case, both iLQG and PI solutions correspond to
the shortest path, i.e. go through the gap between the wall
and the building. Figure 3 (right) illustrates the solutions obtained for larger noise level σu2 = 1. While the optimal reference trajectory obtained by iLQG does not change, which
results in collision once the real noisy controller is executed
(left path), the PI control solution avoids the building and
takes the longer route (right path). Note that iLQG can find
both solutions depending on initialization. However, However, it will always choose the shortest route, regardless of
nearby obstacles. Also, note that the PI controlled unit takes
a longer route to reach the target. The reason is that the con-
1
Percent of crashes
90
Cost solution
80
70
60
50
0
iLQG
PI
0.8
0.6
0.4
0.2
0.05
0.1
0.15
Wind velocity (m/s)
0.2
0
0
0.05
0.1
0.15
Wind velocity (m/s)
0.2
Figure 4: Results: Drunken Quadrotor with wind: For different wind velocities and fixed control noise σu2 = 0.5. (Left) cost
of the obtained solutions and (Right) percentage of crashes using iLQG and PI.
trol cost R is set quite high in order to reach a good ESS.
Alternatively, if R is decreased, the optimal solution could
reach the target sooner, but at the cost of a decreased ESS.
This trade-off, which is inherent in PI control, can be resolved by incorporating feedback control in the importance
sampling, as presented in Thijssen and Kappen (2015).
We also consider more realistic conditions with noise not
limited to act in the control. Figure 4 (a,b) shows results in
the presence of wind and sensor uncertainty. Panel (a) shows
how the wind affects the quality of the solution, resulting in
an increase of the variance and the cost for stronger wind.
In all our tests, iLQG is not able to bring the quadrotor to
the other side. Panel (b) shows the percentage of crashes
using both methods. Crashes occur often using iLQG control and only occasionally using PI control. With stronger
wind, the iLQG controlled unit does occasionally not even
reach the corridor (the unit did not reach the other side but
did not crash either). This explains the difference in percentages of Panel (b). We conclude that for multi-modal tasks
(tasks where multiple solution trajectories exist), the proposed method is preferable to iLQG.
4.2
Scenario II: Holding Pattern
The second scenario addresses the problem of coordinating
agents to hold their position near a point of interest while
keeping a safe range of velocities and avoiding crashing into
each other. Such a problem arises for instance when multiple
aircraft need to land at the same location, and simultaneous
landing is not possible. The resulting flight formation has
been used frequently in the literature (Vásárhelyi et al. 2014;
How et al. 2008; Yu et al. 2013; Franchi et al. 2012), but
always with prior specification of the trajectories. We show
how this formation is obtained as the optimal solution of a
SOC problem.
Consider the following state cost (omitting time indexes)
rHP (x) =
M
X
exp (vi − vmax ) + exp (vmin − vi )
i=1
+ exp (k pi − d k2 ) +
M
X
Chit / k pi − pj k2
j>i
(6)
Figure 5: Holding pattern in the CRATES simulator. Ten
units coordinate their flight in a circular formation. In this
example, N = 104 samples, control noise is σu2 = 0.1 and
horizon H = 1 sec. Cost parameters are vmin = 1, vmax =
3, Chit = 20 and d = 7. Environmental noise and sensing
uncertainties are modeled using realistic parameter values.
where vmax and vmin denote the maximum and minimum velocities, respectively, d denotes penalty for deviation from
the origin and Chit is the penalty for collision risk of two
agents. k · k2 denotes `-2 norm.
The optimal solution for this problem is a circular flying
pattern where units fly equidistantly from each other. The
value of parameter d determines the radius and the average
velocities of the agents are determined from vmin and vmax .
Since the solution is symmetric with respect to the direction of rotation (clockwise or anti-clockwise), only when the
control is executed, a choice is made and the symmetry is
broken. Figure 5 shows a snapshot of a simulation after the
flight formation has been reached for a particular choice of
parameter values 2 . Since we use an uninformed initial control trajectory, there is a transient period during which the
agents organize to reach the optimal configuration. The coordinated circular pattern is obtained regardless of the initial
positions. This behavior is robust and obtained for a large
range of parameter values.
2
Supplementary video material is available at http://www.
mbfys.ru.nl/staff/v.gomez/uav.html
(a)
(d)
3
1.1
250
10
2
N=10
N=102
400
N=10
N=104
3
N=10
200
350
1.05
250
200
150
100
10
ESS
300
Cost iLQG/PI
2
state cost
state cost
(c)
(b)
450
1
1
10
0.95
150
100
0
50
100
time
150
50
2
4
6
8
number of quadrotors
10
0.9
−3
10
0
−2
−1
10
10
Noise level σ2
10
−3
10
u
−2
−1
10
10
Noise level σ2
0
10
u
Figure 6: Holding pattern: (a) evolution of the state cost for different number of samples N = 10, 102 , 103 . (b) scaling of the
method with the number of agents. For different control noise levels, (c) comparison between iLQG and PI control (ratios > 1
indicate better performance of PI over iLQG) and (d) Effective Sample Sizes. Errors bars correspond to ten different random
realizations.
Position for both UAVs
6
UAV0
UAV1
4
2
Meters North
Figure 6(a) shows immediate costs at different times. Cost
always decreases from the starting configuration until the
formation is reached. This value depends on several parameters. We report its dependence on the number N of sample
paths. For large N , the variances are small and the cost attains small values at convergence. Conversely, for small N ,
there is larger variance and the obtained dynamical configuration is less optimal (typically the distances between the
agents are not the same). During the formation of the pattern the controls are more expensive. For this particular task,
full convergence of the path integrals is not required, and the
formation can be achieved with a very small N .
Figure 6(b) illustrates how the method scales as the number of agents increases. We report averages over the mean
costs over 20 time-steps after one minute of flight. We varied M while fixing the rest of the parameters (the distance
d which was set equal to the number of agents in meters).
The small variance of the cost indicates that a stable formation is reached in all the cases. As expected, larger values of
N lead to smaller state cost configurations. For more than
ten UAVs, the simulator starts to have problems in this task
and occasional crashes may occur before the formation is
reached due to limited sample sizes. This limitation can be
addressed, for example, by using more processing power and
parallelization and it is left for future work.
We also compared our approach with iLQG in this scenario. Figure 6(c) shows the ratio of cost differences after
convergence of both solutions. Both use MPC, with a horizon of 2s and update frequency of 15Hz. Values above 1 indicate that PI control consistently outperforms iLQG in this
problem. Before convergence, we also found, as in the previous task, that iLQG resulted in occasional crashes while PI
control did not. The Effective Sample Size (ESS) is shown
in Figure 6(d). We observe that higher control noise levels
result in better exploration and thus better controls. We can
thus conclude that the proposed methodology is feasible for
coordinating a large team of quadrotors.
For this task, we performed experiments with the real platforms. Figure 7 shows real trajectories obtained in outdoor
0
−2
−4
−6
−5
−3
−1
1
Meters East
3
5
Figure 7: Resulting trajectories of a Holding Pattern experiment using two platforms in outdoors conditions.
conditions (see also the video that accompanies this paper
for an experiment with three platforms). Despite the presence of significant noise, the circular behavior was also obtained. In the real experiments, we used a Core i7 laptop with
8GB RAM as base station, which run its own ROS messaging core and forwarded messages to and from the platforms
over a IEEE 802.11 2.4GHz network. For safety reasons, the
quadrotors were flown at different altitudes.
4.3
Scenario III: Cat and Mouse
The final scenario that we consider is the cat and mouse scenario. In this task, a team of M quadrotors (the cats) has
to catch (get close to) another quadrotor (the mouse). The
mouse has autonomous dynamics: it tries to escape the cats
by moving at velocity inversely proportional to the distance
to the cats. More precisely, let pmouse denote the 2D position of the mouse, the velocity command for the mouse is
Mouse
Cats
Figure 8: Cat and mouse scenario: (Top-left) four cats and one mouse. (Top-right) for horizon time H = 2 seconds, the
four cats surround the mouse forever and keep rotation around it. (Bottom-left) for horizon time H = 1 seconds, the four
cats chase the mouse but (bottom-right) the mouse manages to escape. With these settings, the multi-agent system alternates
between these two dynamical states. Number of sample paths is N = 104 , noise level σu2 = 0.5. Other parameter values are
d = 30, vmin = 1, vmax = 4, vmin = 4 and vmax mouse = 3.
computed (omitting time indexes) as
M
vmouse = v
max
mouse k
X pi − pmouse
v
.
, where v =
v k2
k pi − pmouse k2
i=1
The parameter v max determines the maximum velocity of
mouse
the mouse. As state cost function we use equation (6) with
an additional penalty term that depends on the sum of the
distances to the mouse
M
X
k pi − pmouse k2 .
rCM (x) = rHP (x) +
i=1
This scenario leads to several interesting dynamical states.
For example, for a sufficiently large value of M , the mouse
always gets caught (if its initial position is not close to the
boundary, determined by d). The optimal control for the cats
consists in surrounding the mouse to prevent collision. Once
the mouse is surrounded, the cats keep rotating around it,
as in the previous scenario, but with the origin replaced by
the mouse position. The additional video shows examples
of other complex behaviors obtained for different parameter
settings. Figure 8 (top-right) illustrates this behavior.
The types of solution we observe are different for other
parameter values. For example, for M = 2 or a small time
horizon, e.g. H = 1, the dynamical state in which the cats
rotate around the mouse is not stable, and the mouse escapes.
This is displayed in Figure 8 (bottom panels) and better illustrated in the video provided as supplementary material.
We emphasize that these different behaviors are observed
for large uncertainty in the form of sensor noise and wind.
5
Conclusions
This paper presents a centralized, real-time stochastic optimal control algorithm for coordinating the actions of multiple autonomous vehicles in order to minimize a global cost
function. The high-level control task is expressed as a PathIntegral control problem that can be solved using efficient
sampling methods and real-time control is possible via the
use of re-planning and model predictive control. To the best
of our knowledge, this is the first real-time implementation
of Path-Integral control on an actual multi-agent system.
We have shown in a simple scenario (Drunken Quadrotor) that the proposed methodology is more convenient than
other approaches such as deterministic control or iLQG
for planning trajectories. In more complex scenarios such
as the Holding Pattern and the Cat and Mouse, the proposed methodology is also preferable and allows for realtime control. We observe multiple and complex group behavior emerging from the specified cost function. Our experimental framework CRATES has been a key development
that permitted a smooth transition from the theory to the real
quadrotor platforms, with literally no modification of the underlying control code. This gives evidence that the model
mismatch caused by the use of a control hierarchy is not
critical in normal outdoor conditions. Our current research
is addressing the following aspects:
Large scale parallel sampling− the presented method
can be easily parallelized, for instance, using graphics processing units, as in Williams, Rombokas, and Daniel (2014).
Although the tasks considered in this work did not required
more than 104 samples, we expect that this improvement
will significantly increase the number of application domains and system size.
Distributed control− we are exploring different distributed formulations that take better profit of the factorized
representation of the state cost. Note that the costs functions
considered in this work only require pairwise couplings of
the agents (to prevent collisions). However, full observability of the joint space is still required, which is not available
in a fully distributed approach.
References
[Albore et al.] Albore, A.; Peyrard, N.; Sabbadin, R.; and Teichteil Königsbuch, F. 2015. An online replanning approach for
crop fields mapping with autonomous UAVs. In International Conference on Automated Planning and Scheduling (ICAPS).
[Augugliaro, Schoellig, and D’Andrea] Augugliaro, F.; Schoellig,
A.; and D’Andrea, R. 2012. Generation of collision-free trajectories for a quadrocopter fleet: A sequential convex programming
approach. In Intelligent Robots and Systems (IROS), 1917–1922.
[Bernardini, Fox, and Long] Bernardini, S.; Fox, M.; and Long, D.
2014. Planning the behaviour of low-cost quadcopters for surveillance missions. In International Conference on Automated Planning and Scheduling (ICAPS).
[Bürkle, Segor, and Kollmann] Bürkle, A.; Segor, F.; and Kollmann, M. 2011. Towards autonomous micro UAV swarms. J.
Intell. Robot. Syst. 61(1-4):339–353.
[Chanel, Teichteil-Knigsbuch, and Lesire] Chanel, C. C.; TeichteilKnigsbuch, F.; and Lesire, C. 2013. Multi-target detection and
recognition by UAVs using online POMDPs. In International Conference on Automated Planning and Scheduling (ICAPS).
[De Nardi and Holland] De Nardi, R., and Holland, O. 2008. Coevolutionary modelling of a miniature rotorcraft. In 10th International Conference on Intelligent Autonomous Systems (IAS10), 364
– 373.
[De Nardi] De Nardi, R. 2013. The QRSim Quadrotors Simulator.
Technical Report RN/13/08, University College London.
[Franchi et al.] Franchi, A.; Masone, C.; Grabe, V.; Ryll, M.;
Bülthoff, H. H.; and Giordano, P. R. 2012. Modeling and control of UAV bearing-formations with bilateral high-level steering.
Int. J. Robot. Res. 0278364912462493.
[Gómez et al.] Gómez, V.; Kappen, H. J.; Peters, J.; and Neumann,
G. 2014. Policy search for path integral control. In European Conf.
on Machine Learning & Knowledge Discovery in Databases, 482–
497.
[Guerrero and Lozano] Guerrero, J., and Lozano, R. 2012. Flight
Formation Control. John Wiley & Sons.
[Hauert et al.] Hauert, S.; Leven, S.; Varga, M.; Ruini, F.; Cangelosi, A.; Zufferey, J.-C.; and Floreano, D. 2011. Reynolds flocking in reality with fixed-wing robots: Communication range vs.
maximum turning rate. In Intelligent Robots and Systems (IROS),
5015–5020.
[Hoffmann et al.] Hoffmann, G. M.; Huang, H.; Waslander, S. L.;
and Tomlin, C. J. 2011. Precision flight control for a multi-vehicle
quadrotor helicopter testbed. Control. Eng. Pract. 19(9):1023 –
1036.
[Horowitz, Damle, and Burdick] Horowitz, M. B.; Damle, A.; and
Burdick, J. W. 2014. Linear Hamilton Jacobi Bellman equations in
high dimensions. arXiv preprint arXiv:1404.1089.
[How et al.] How, J.; Bethke, B.; Frank, A.; Dale, D.; and Vian,
J. 2008. Real-time indoor autonomous vehicle test environment.
IEEE Contr. Syst. Mag. 28(2):51–64.
[Kappen, Gómez, and Opper] Kappen, H. J.; Gómez, V.; and Opper, M. 2012. Optimal control as a graphical model inference
problem. Mach. Learn. 87:159–182.
[Kappen] Kappen, H. J. 2005. Path integrals and symmetry breaking for optimal control theory. Journal of statistical mechanics:
theory and experiment 2005(11):P11011.
[Kendoul] Kendoul, F. 2012. Survey of advances in guidance, navigation, and control of unmanned rotorcraft systems. J. Field Robot.
29(2):315–378.
[Kinjo, Uchibe, and Doya] Kinjo, K.; Uchibe, E.; and Doya, K.
2013. Evaluation of linearly solvable Markov decision process with
dynamic model learning in a mobile robot navigation task. Front.
Neurorobot. 7:1–13.
[Kushleyev et al.] Kushleyev, A.; Mellinger, D.; Powers, C.; and
Kumar, V. 2013. Towards a swarm of agile micro quadrotors.
Auton. Robot. 35(4):287–300.
[Mahony, Kumar, and Corke] Mahony, R.; Kumar, V.; and Corke, P.
2012. Multirotor aerial vehicles: Modeling, estimation, and control
of quadrotor. IEEE Robotics & Automation Magazine 20–32.
[Meyer et al.] Meyer, J.; Sendobry, A.; Kohlbrecher, S.; and Klingauf, U. 2012. Comprehensive Simulation of Quadrotor UAVs
Using ROS and Gazebo. Lecture Notes in Computer Science
7628:400–411.
[Ono, Williams, and Blackmore] Ono, M.; Williams, B. C.; and
Blackmore, L. 2013. Probabilistic planning for continuous dynamic systems under bounded risk. J. Artif. Int. Res. 46(1):511–
577.
[Quintero, Collins, and Hespanha] Quintero, S.; Collins, G.; and
Hespanha, J. 2013. Flocking with fixed-wing UAVs for distributed
sensing: A stochastic optimal control approach. In American Control Conference (ACC), 2025–2031.
[Rawlik, Toussaint, and Vijayakumar] Rawlik, K.; Toussaint, M.;
and Vijayakumar, S. 2013. Path integral control by reproducing
kernel Hilbert space embedding. In Twenty-Third International
Joint Conference on Artificial Intelligence, 1628–1634. AAAI
Press.
[Reynolds] Reynolds, C. W. 1987. Flocks, herds and schools: A distributed behavioral model. SIGGRAPH Comput. Graph. 21(4):25–
34.
[Rombokas et al.] Rombokas, E.; Theodorou, E.; Malhotra, M.;
Todorov, E.; and Matsuoka, Y. 2012. Tendon-driven control of
biomechanical and robotic systems: A path integral reinforcement
learning approach. In International Conference on Robotics and
Automation, 208–214.
[Shim, Kim, and Sastry] Shim, D. H.; Kim, H. J.; and Sastry, S.
2003. Decentralized nonlinear model predictive control of multiple flying robots. In IEEE conference on Decision and control
(CDC), volume 4, 3621–3626.
[Stulp and Sigaud] Stulp, F., and Sigaud, O. 2012. Path integral
policy improvement with covariance matrix adaptation. In International Conference Machine Learning (ICML).
[Symington et al.] Symington, A.; De Nardi, R.; Julier, S.; and
Hailes, S. 2014. Simulating quadrotor UAVs in outdoor scenarios. In Intelligent Robots and Systems (IROS), 3382–3388.
[Theodorou, Buchli, and Schaal] Theodorou, E.; Buchli, J.; and
Schaal, S. 2010. A generalized path integral control approach to
reinforcement learning. J. Mach. Learn. Res. 11:3137–3181.
[Thijssen and Kappen] Thijssen, S., and Kappen, H. J. 2015.
Path integral control and state-dependent feedback. Phys. Rev. E
91:032104.
[Todorov and Li] Todorov, E., and Li, W. 2005. A generalized iterative LQG method for locally-optimal feedback control of constrained nonlinear stochastic systems. In American Control Conference (ACC), 300–306 vol. 1. IEEE.
[Todorov] Todorov, E. 2006. Linearly-solvable Markov decision
problems. In Advances in neural information processing systems
(NIPS), 1369–1376.
[Turpin, Michael, and Kumar] Turpin, M.; Michael, N.; and Kumar,
V. 2012. Decentralized formation control with variable shapes for
aerial robots. In International Conference on Robotics and Automation (ICRA), 23–30.
[Van Den Broek, Wiegerinck, and Kappen] Van Den Broek, B.;
Wiegerinck, W.; and Kappen, H. J. 2008. Graphical model inference in optimal control of stochastic multi-agent systems. J. Artif.
Intell. Res. 32:95–122.
[Vásárhelyi et al.] Vásárhelyi, G.; Virágh, C.; Somorjai, G.; Tarcai,
N.; Szorenyi, T.; Nepusz, T.; and Vicsek, T. 2014. Outdoor flocking
and formation flight with autonomous aerial robots. In Intelligent
Robots and Systems (IROS), 3866–3873.
[Williams, Rombokas, and Daniel] Williams, G.; Rombokas, E.;
and Daniel, T. 2014. GPU based path integral control with learned
dynamics. In Autonomously Learning Robots - NIPS Workshop.
[Yu et al.] Yu, B.; Dong, X.; Shi, Z.; and Zhong, Y. 2013. Formation
control for quadrotor swarm systems: Algorithms and experiments.
In Chinese Control Conference (CCC), 7099–7104.
| 3cs.SY
|
The gene family-free median of three
Daniel Doerr1 , Pedro Feijão2 , Metin Balaban1 , and Cedric Chauve3
arXiv:1606.05910v1 [cs.DS] 19 Jun 2016
1
3
School of Computer and Communication Sciences, EPFL, 1015 Lausanne,
Switzerland [email protected]
2
Faculty of Technology, Bielefeld University, 33615 Bielefeld, Germany
Department of Mathematics, Simon Fraser University, Burnaby BC, Canada
Abstract. The gene family-free framework for comparative genomics
aims at developing methods for gene order analysis that do not require
prior gene family assignment, but work directly on a sequence similarity multipartite graph. We present a model for constructing a median
of three genomes in this family-free setting, based on maximizing an
objective function that generalizes the classical breakpoint distance by
integrating sequence similarity in the score of a gene adjacency. We show
that the corresponding computational problem is MAX SNP-hard and
we present a 0-1 linear program for its exact solution. The result of our
FF-median program is a median genome with median genes associated
to extant genes, in which median adjacencies are assumed to define positional orthologs. We demonstrate through simulations and comparison
with the OMA orthology database that the herein presented method
is able compute accurate medians and positional orthologs for genomes
comparable in size of bacterial genomes.
1
Introduction
The prediction of evolutionary relationships between genomic sequences is a longstanding problem in computational biology. According to Fitch [9], two genomic
sequences are called homologous if they descended from a common ancestral sequence. Furthermore, Fitch identifies different events that give rise to a branching point in the phylogeny of homologous sequences, leading to the concepts
of orthologous genes (who descend from their last common ancestor through a
speciation) and paralogous genes (descending from their last common ancestor
through a speciation), that reach far beyond evolutionary genomics [10]. Until
quite recently, orthology and paralogy relationships were mostly inferred from
sequence similarity. However it is now well accepted that the syntenic context
can carry valuable evolutionary information, which has lead to the notion of
positional orthologs [6]. In the present work, we describe a method to compute
groups of likely orthologous genes for a group of three genomes, through a new
problem we introduce, the gene family-free median of three.
Most methods for detecting potential orthologous groups require a prior clustering of the genes of the considered genomes into homologous gene families,
defined as groups of genes assumed to originate from a single ancestral gene;
clustering protein sequences into families is already in itself a difficult problem.
Here, we follow the matching-based approach, framed within the gene familyfree principle, that embodies the idea to perform gene order analysis without the
prerequisite of gene family or homology assignments. Instead, we are given allagainst-all gene similarities through a symmetric and reflexive similarity measure σ : Σ×Σ → R≥0 over the universe of genes Σ [4]. We use sequence similarity
but other similarity measures can fit the previous definition. Gene family or homology assignments represent a particular subgroup of gene similarity functions
that require transitivity. Independent of the particular similarity measure σ, relations between genes imposed by σ are considered as candidates for homology
assignments. A gene family-free research program was outlined in [4] (see also [8])
and has so far been developed for the pairwise comparison of genomes [7,14,11]
and shown to be effective for orthology analysis [12].
In Section 2 we introduce a new genome median problem in the family-free
framework, that generalizes the traditional breakpoint median problem [19]. For
a group of three genomes, the input of the family-free median problem is a
tripartite similarity graph of pairwise gene similarities. Informally, a median of
three is defined as both a set of median genes – each defined by three extant
genes forming a clique in the similarity graph, scored according the edges of this
clique –, forming a set of median adjacencies, each supported by at least one
extant gene adjacency (hence any median gene belongs to at least one median
adjacency). A median is optimal if it maximizes the sum of the scores of its
median genes. Hence, the optimization criterion of this problem fully integrates
both sequence similarity and synteny conservation. In Section 3 we study its the
computational complexity and give an exact algorithm for its solution. We show
that our method can be used for positional ortholog prediction in simulated and
real data sets of bacterial genomes in Section 4.
2
The gene family-free median of three
Extant genomes, genes and adjacencies. In this work, a genome G is entirely
represented by a tuple G ≡ (C, A), where C denotes a non-empty set of unique
genes, and A is a set of adjacencies. Genes are represented by their extremities,
i.e., a gene g ≡ (g t , g h ), g ∈ C, consists of a head g h and a tail g t . Telomeres are
modeled explicitly, as special genes of C(G) with a single extremity, denoted by
“◦”. Extremities g a , ḡ b , a, b ∈ {h, t} of any two genes g, ḡ can form an adjacency.
In the following, we will conveniently use the notation C(G) and A(G) to denote
the set of genes and the set of adjacencies of genome G, respectively. We indicate
the presence of an adjacency {xa1 , xb2 } in an extant genome X by
(
1 if {xa1 , xb2 } ∈ A(X)
IX (xa1 , xb2 ) =
(1)
0 otherwise.
Given two genomes G and H and gene similarity measure σ, two adjacencies, {g1a , g2b } ∈ A(G) and {ha1 , hb2 } ∈ A(H) with a, b ∈ {h, t} are conserved iff
σ(g1 , h1 ) > 0 and σ(g2 , h2 ) > 0. We subsequently define the adjacency score of
2
any four extremities g a , hb , ic , j d , where a, b, c, d ∈ {h, t} and g, h, i, j ∈ Σ as the
geometric mean of their corresponding gene similarities:
p
s(g a , hb , ic , j d ) ≡ σ(g, h) · σ(i, j)
(2)
Median genome, genes and adjacencies. Informally, the family-free median problem asks for a fourth genome M that maximizes the sum of pairwise adjacency
scores to three given extant genomes G, H, and I. In doing so, the gene content
of the requested median M must first be defined: each gene m ∈ C(M ) must
be unambiguously associated with a triple of extant genes (g, h, i), g ∈ C(G),
h ∈ C(H), and i ∈ C(I). Moreover, we want to associate to a median gene m a
sequence similarity score (g, h, i) relatively to the three extant genes it is related
to. As the sequence of the median gene is obviously not available, we define this
score as the geometric mean of their pairwise similarities:
p
σ(g, m) = σ(h, m) = σ(i, m) ≡ 3 σ(g, h) · σ(g, i) · σ(h, i)
(3)
g1
i
(a)
?
σ(g, i)
σ(h, i)
m
?
g2
g3
G
(b)
H
?
g
σ(g, h)
g4
h3
h1
h2
I
h
i1
i2
i3
Fig. 1. (a) Illustration of the score of a candidate median gene. (b) Gene similarity
graph of three genomes G, H, and I. Colored components indicate candidate median
genes m1 = (g1 , h1 , i2 ), m2 = (g2 , h2 , i1 ), m3 = (g3 , h3 , i2 ), and m4 = (g4 , h3 , i3 ).
Median gene pairs m1 , m3 and m3 , m4 are conflicting.
In the following we make use of mapping πG (m) ≡ g, πH (m) ≡ h, and
πI (m) ≡ i to relate gene m with its extant counterparts. Two candidate median
genes or telomeres m1 and m2 are conflicting if m1 6= m2 and the intersection
between associated gene sets {πG (m1 ), πH (m1 ), πI (m1 )} and {πG (m2 ), πH (m2 ),
πI (m2 )} is non-empty. A set of candiate median genes or telomeres C is called
conflict-free if no two of its members m1 , m2 ∈ C are conflicting. This definition
trivially extends to the notion of a conflict-free median.
Problem 1 (FF-Median). Given three genomes G, H, and I, and gene similarity
measure σ, find a conflict-free median M , which maximizes the following formula:
X
X
s(ma1 , πX (m1 )a , mb2 , πX (m2 )b ),
F (M ) =
b
{ma
1 ,m2 }∈A(M)
X∈{G,H,I},
{πX (m1 )a ,πX (m2 )b }∈A(X)
(4)
where a, b ∈ {h, t} and s(·) is the adjacency score as defined by Equation (2).
3
Remark 1. The adjacency score for a median adjacency {ma1 , mb2 } with respect
to the corresponding potential extant adjacency {πX (m1 )a , πX (m2 )b }, where
{ma1 , mb2 } ∈ A(M ) and X ∈ {G, H, I}, can be entirely expressed in terms of
pairwise similarities between genes of extant genomes using Equation (3):
s(ma1 , πX (m1 )a , mb2 , πX (m2 )b ) =
s
6
Y
σ(πY (m1 ), πZ (m1 )) · σ(πY (m2 ), πZ (m2 ))
{Y,Z}⊂{G,H,I}
In the following, a median gene m and its extant counterparts (g, h, i) are
treated as equivalent. We denote the set of all candidate median genes by
Σ = {(g, h, i) | g ∈ C(G), h ∈ C(H), i ∈ C(I) : σ(g, h) · σ(g, i) · σ(h, i) > 0} . (5)
Each pair of median genes (g1 , h1 , i1 ), (g2 , h2 , i2 ) ∈ Σ and extremities a, b ∈
{h, t} give rise to a candidate median adjacency {(g1a , ha1 , ia1 ), (g2b , hb2 , ib2 )} if (g1a , ha1 , ia1 ) 6=
(g2b , hb2 , ib2 ), and (g1a , ha1 , ia1 ) and (g2b , hb2 , ib2 ) are non-conflicting. We denote the set
of all candidate median adjacencies and the set of all conserved (i.e. present in at
least one extant genome) candidate median adjacencies
by A = {{ma1 , ma2 } | m1 , m2 ∈
P
a
b
a
b
Σ , a, b ∈ {h, t}} and AC
=
{
{m
,
m
}
∈
A
|
1
2
X∈{G,H,I} IX (πX (m1 ) , πX (m2 ) ) ≥
1}, respectively.
Remark 2. A median gene can only belong to a median adjacency with nonzero adjacency score if all pairwise similarities of its corresponding extant genes
g, h, i are non-zero. Thus, the search for median genes can be limited to 3-cliques
(triangles) in the tripartite similarity graph.
Remark 3. The right-hand side of the above formula for the weight of an adjacency is independent of genome X. From Equation (4), an adjacency in median
M has only an impact in a solution to problem FF-Median if it participates in a
gene adjacency in at least one extant genome. So including in a median genome
median genes that do not belong to a candidate median adjacency in AC
do not
increase the objective function.
Related problems. The FF-median problem relates to previously studied gene
order evolution problems. It is a generalization of the tractable mixed multichromosomal median problem introduced in [19], that can indeed be defined as
an FF-median problem with a similarity graph composed of disjoint 3-cliques
and edges having all the same weight. The FF-median problem also bears similarity with methods aimed at detecting groups of orthologous genes based on
gene order evolution, especially the MultiMSOAR [18] algorithm, although other
method integrate synteny and sequence conservation for inferring orhogroups,
see [6]. Our approach differs first and foremost in its family-free principle (all
other methods require a prior gene family assignment). Compared to MultiMSOAR, the only other method that can handle more than two genomes with an
optimization criterion that considers gene order evolution, both MultiMSOAR
(for three genomes) and FF-median aim at computing a maximum weight tripartite matching. However we differ fundamentally from MultiMSOAR by the
4
full integration of sequence and synteny conservation into the objective function,
while MultiMSOAR proceeds first by computing pairwise orthology assignments
to define a multipartite graph.
3
Algorithmic and complexity results
We now describe our theoretical results: a NP-hardness proof, an exact Integer Linear Program (ILP), and an algorithm to detect local optimal structures.
Theorem 1. Problem FF-Median is MAX SNP-hard.
We describe the full hardness proof in Appendix A. It is based on a reduction
from the Maximum Independent Set for Graphs of Bounded Degree 3.
An exact ILP algorithm to problem FF-Median. We now present program FF-Median,
described by Algorithm 1, that exploits the specific properties of problem FFMedian to design an ILP using O(n5 ) variables and statements. Program FF-Median
makes use of two types of binary variables a and b as declared in domain specifications (D.01) and (D.02), that defines the set of median genes Σλ and of
candidate conserved median adjacencies AC
(Remark 3). The former variable
type indicates the presence or absence of candidate genes in an optimal median
M . The latter, variable type b, specifies if an adjacency between two gene extremities or telomeres is established in M . Constraint (C.01) ensures that M
is conflict-free, by demanding that each extant gene (or telomere) can be associated with at most one median gene (or telomere). Further, constraint (C.02)
dictates that a median adjacency can only be established between genes that
both are part of the median. Lastly, constraint (C.03) guarantees that each gene
extremity and telomere of the median participates in at most one adjacency.
Property 1. The size (i.e. number of variables and statements) of any ILP returned by program FF-Median is limited by O(n5 ) where n = max(|C(G)|, |C(H)|, |C(I)|).
Remark 4. The output of the algorithm FF-Median is a set of adjacencies between median genes that define a set of linear and/or circular orders, called
CARs (Contiguous Ancestral Regions), where linear segments are not capped
by telomeres. So formally the computed median might not be a valid genome.
However, as adding adjacencies that do not belong to AC
do not modify the
score of a given median, a set of median adjacencies can always be completed
into a valid genome by such adjacencies that join the linear segments together
and add telomeres. These extra adjacencies would not be supported by any extant genome and thus can be considered as dubious, and in our implementation,
we only return the median adjacencies computed by the ILP, i.e. a subset of AC
.
Remark 5. Following Remark 2, preprocessing the input extant genomes requires
to handle the extant genes that do not belong to at least one 3-clique in the
similarity graph. Such genes can not be part of any median. So one could decide
to leave them in the input, and the ILP can handle them and ensures they
5
Algorithm 1 Program FF-Median for three genomes (G, H, I)
Objective: Maximize
b a b
b(g1a , g2b , ha
1 , h2 , i1 , i2 )
X
q
6
b
a b
σ(g1 , h1 , i1 )σ(g2 , h2 , i2 ) (IG (g1a , g2b ) + IH (ha
1 , h2 ) + II (i1 , i2 ))
(g1 ,h1 ,i1 ),
(g2 ,h2 ,i2 )∈AC
,
a,b∈{h, t}
Constraints:
(C.01)
X
∀ g ′ ∈ C(G):
a(g, h, i) ≤ 1
(g,h,i)∈Σ , g=g ′
∀ h′ ∈ C(H):
X
a(g, h, i) ≤ 1
(g,h,i)∈Σ , h=h′
∀ i′ ∈ C(I):
X
a(g, h, i) ≤ 1
(g,h,i)∈Σ , i=i′
(C.02)
∀ (g1 , h1 , i1 ), (g2 , h2 , i2 ) ∈ Σ and ∀ a, b ∈ {h, t}:
2 · b(g1a , g2b , ha1 , hb2 , ia1 , ib2 ) ≤ a(g1 , h1 , i1 ) + a(g2 , h2 , i2 )
(C.03)
∀ (g1 , h1 , i1 ) ∈ Σ and ∀ a ∈ {h, t}:
X
b(g1a , g2b , ha1 , hb2 , ia1 , ib2 ) ≤ 1
(g2 ,h2 ,i2 )∈Σ , b∈{h, t}
Domains:
(D.01)
∀ (g, h, i) ∈ Σ : a(g, h, i) ∈ {0, 1}
(D.02)
∀ (g1 , h1 , i1 ), (g2 , h2 , i2 ) ∈ AC
and ∀ a, b ∈ {h, t}:
b(g1a , g2b , ha1 , hb2 , ia1 , ib2 ) ∈ {0, 1}
are never part of the output solution. However, discarding them from the extant
genomes can help recover adjacencies that have been disrupted by the insertion of
a mobile element for example, so in our implementation we follow this approach.
As discussed at the end of Section 2, the FF-median problem is a generalization of the mixed multichromosomal breakpoint median [19]. However, it was
shown in [19] that this breakpoint median problem can be solved in polynomial
time by a Maximum-Weight Matching (MWM) algorithm. This motivates the
results presented in the next paragraph that use a MWM algorithm to identify
optimal median substructures by focusing on conflict-free sets of median genes.
Finding local optimal segments. Tannier et al. [19] solve the mixed multichromosomal breakpoint median problem by transforming it into an MWM problem,
that we outline now. A graph is defined in which each extremity of a candidate
median gene and each telomere gives rise to a vertex. Any two vertices are connected by an edge, weighted according to the number of observed adjacencies
between the two gene extremities in extant genomes. Edges corresponding to
adjacencies between a gene extremity and telomeres are weighted only by half
6
as much. An MWM in this graph induces a set of adjacencies that defines an
optimal median.
We first describe how this approach applies to our problem. We define a graph
Γ (G, H, I, σ) constructed from an FF-Median instance (G, H, I, σ) that is similar
to that of Tannier et al., only deviating by defining vertices as candidate median
genes and weighting an edge between two candidate median gene extremities (or
telomeres) ma1 , mb2 , a, b ∈ {h, t}, by
w({ma1 , mb2 }) =
s
6
Y
σ(πX (m1 ), πY (m1 ))σ(πX (m2 ), πY (m2 ))
{X,Y }⊂{G,H,I}
·
X
(6)
a
b
IX (πX (m1 ) , πX (m2 ) ).
X∈{G,H,I}
We make first the following observation, where a conflict-free matching is
a matching that does not contain two conflicting vertices (candidate median
genes):
Observation 2 Any conflict-free matching in graph Γ (G, H, I, σ) of maximum
weight defines an optimal median.
We show now that we can define notions of sub-instances – of a full FFmedian instance – that contains no internal conflicts, for which applying the
MWM can allow to detect if the set of median genes defining the sub-instance
is part of at least one optimal FF-median. Let S be a set of candidate median
genes. An internal conflict is a conflict between two genes from S; an external
conflict is a conflict between a gene from S and a candidate median gene not
in S. We say that S is contiguous in extant genome X if the set πX (S) forms
a unique, contiguous, segment in X. We say that S is an internal-conflict free
segment (IC-free segment) if it contains no internal conflict and is contiguous in
all three extant genomes; this can be seen as the family-free equivalent of the
notion of common interval in permutations [3]. An IC-free segment is framed if
the extremities of the extant segments belong to the same two median genes,
with conserved relative orientations (the equivalent of a conserved interval ). An
IC-free segment is a run if the order of the extant genes is conserved in all three
extant genomes, up to a full reversal of the segment.
Intuitively, one can find an optimal solution to the sub-instance defined by
an IC-free segment, but it might not be part of an optimal median for the whole
instance due to side effects of the rest of the instance. So we need to adapt the
graph to which we apply an MWM algorithm to account for such side effects.
To do so, we define the potential of a candidate median gene m as
w({ma1 , mb }) + w({ma , mb2 }) .
∆(m) =
max
b
b
a
{{ma
1 ,m },{m ,m2 }}∈A
We then extend graph Γ =: (V, E) to graph Γ ′ := (V, E ′ ) by adding edges
between the extremities of each candidate median gene of an IC-free segment S,
7
i.e. E ′ = E ∪ {{mh , mt } | m ∈ S}. In the following we refer to these edges as
conflict edges. Let C(m) be the set of candidate median genes that are involved
in an (external) conflict with a given candidate median gene m of S, then the
conflict edge {mh , mt } ∈ E ′ is weighted by the maximum potential of a nonconflicting subset of C(m),
X
w′ ({mh , mt }) = max({
∆(m′ ) | C ′ ⊆ C(m) : C ′ is conflict-free}) .
m′ ∈C ′
A conflict-free matching in Γ ′ is a matching that does not contain a conflict
edge.
Lemma 1. Given an internal conflict-free segment S, any maximum weight
matching in graph Γ ′ (S) that is conflict-free defines a set of median genes and
adjacencies that belong to at least one optimal FF-median of the whole instance.
A proof is presented in Appendix B. Lemma 1 leads to a procedure (Algorithm 2) that iteratively identifies and tests IC-free segments in the FF-Median
instance. For each identified IC-free segment S an adjacency graph Γ ′ (S) is constructed and a maximum weight matching is computed (lines 2-3). If the resulting
matching is conflict free (line 4), adjacencies of IC-free segment S are reported
and S is removed from an FF-Median instance by masking its internal adjacencies and removing all candidate median genes (and consequently their associated
candidate median adjacencies) corresponding to external conflicts (lines 5-6). It
then follows immediately from Lemma 1 that the set median genes returned by
Algorithm 2 belongs to at least one optimal solution to the FF-median problem.
Algorithm 2 Algorithm ICF-SEG
Input: FF-Median instance (G, H, I, σ)
Output: Set of adjacencies AdjM that is part of a median M of (G, H, I, σ).
1: while there exists an unobserved IC-free conserved segment S in (G, H, I, σ) do
2:
Construct adjacency graph Γ ′ (S) of S
3:
Find maximum weight matching M ⊆ E(Γ ′ (S))
4:
if A(S) = M then
5:
Add A(S) to AdjM
6:
Remove S including external conflicts from (G, H, I, σ)
7:
end if
8: end while
4
Experimental results and discussion
Our algorithms have been implemented in Python and require CPLEX4 ; they
are freely available as part of the family-free genome comparison tool FFGC downloadable at http://bibiserv.cebitec.uni-bielefeld.de/ffgc.
4
http://www.ibm.com/software/integration/optimization/cplex-optimizer/
8
In subsequent analyses, gene similarities are based on local alignment hits
identified with BLASTP [2] on protein sequences using an e-value threshold
of 10−5 . In gene similarity graphs, we discard spurious edges by applying a
stringency filter proposed by Lechner et al. [13] that utilizes a local threshold
parameter f ∈ [0, 1] and BLAST bitscores: a BLAST hit from a gene g to h is
only retained if it is has a higher or equal score than f times the best BLAST
hit from h to any gene g ′ that is member of the same genome as g. In all our
experiments, we set f to 0.5. Edge weights of the gene similarity graph are then
calculated according to the relative reciprocal BLAST score (RRBS) [17]. Finally
we applied Algorithm ICF-SEG with conserved segments defined as runs.
For solving the FF-Median problem, we granted CPLEX two CPU cores,
4 GB memory, and a time limit of 3 hours per dataset.
In our experiments, we compare ourselves against the orthology prediction
tool MultiMSOAR [18]. This tool requires precomputed gene families, which we
constructed by following the workflow described in [18].
Evaluation on simulated data. We first evaluate our algorithms on simulated data
sets obtained by ALF [5]. The ALF simulator covers many aspects of genome
evolution from point mutations to global modifications. The latter includes two
types of genome rearrangements, as well as various options to customize the
process of gene family evolution. In our simulations, we mainly use standard
parameters suggested by the authors of ALF and we focus on three parameters
that primarily influence the outcome of gene family-free genome analysis: (i) the
rate of sequence evolution, (ii) the rate of genome rearrangements, and (iii) the
rate of gene duplications and losses. We keep all three rates constant, only varying the evolutionary distance between the generated extant genomes. We confine
our simulations to protein coding sequences. A comprehensive list of parameter
settings used in our simulations is shown in Table 2 in Appendix C. As root
genome in the simulations, we used the genomic sequence of an E. coli K-12
strain5 which comprises 4, 320 protein coding genes. We then generated 7 × 10
data sets with increasing evolutionary distance ranging from 10 to 130 percent
accepted mutations (PAM). Details about the generated data sets are shown in
Table 1 in Appendix C. Figure 2 (a) shows the outcome of our analysis with
respect to precision and recall6 of inferring positional orthologs. In all simulations, FF-Median generated no or very few false positives, leading to perfect or
near-perfect precision score, consistently outperforming MultiMSOAR. However,
since the objective of FF-Median only takes median genes into account that are
conserved by synteny, the increase in mutational changes over evolutionary time
causes a growing loss of syntenic context which results in a lower recall. Therefore, MultiMSOAR retains a better recall for larger evolutionary distances, while
FF-Median provides better results for more closely related genomes.
5
6
Accession no: NC 000913.2
precision: #true positives/(#true positives + #false positives), recall: #true positives/(#true positives + #false negatives)
9
real data: median statistics
ALF simulated datasets
10 5
1
0.98
10 4
0.96
10 3
#median genes
#cars
#circular cars
0.94
10 2
0.92
precision FF-Median
recall FF-Median
precision MultiMSOAR
recall MultiMSOAR
0.9
0.88
10
30
50
70
10 1
90
110
10 0
130
0
100
(a)
2500
400
OMA: MultiMSOAR
5000
agreeing
compatible
disagreeing
3000
300
(b)
OMA: FF-Median
3500
200
combinations of 3 input genomes
evolutionary distance (PAM)
agreeing
compatible
disagreeing
4000
2000
3000
1500
2000
1000
1000
500
0
0
100
200
300
400
combinations of 3 input genomes
0
0
100
200
300
400
500
combinations of 3 input genomes
(c)
(d)
Fig. 2. Top: (a) Precision and recall of FF-Median and MultiMSOAR in simulations;
(b) statistical assessment of CARs and median genes on real datasets. Bottom: agreement, compatiblity and disagreement of positional orthologs inferred by (c) FF-Median
and (d) MultiMSOAR with OMA database.
Evaluation on real data. We study 15 γ-proteobacterial genomes that span a
large taxonomic spectrum and are contained in the OMA database [1]. A complete list of species names is given in Appendix D. We obtained the genomic
sequences from the NCBI database and constructed for each combination of
three genomes a gene similarity graph following the same procedure as in the
simulated dataset. In 9 out of the 455 combinations of genomes the time limit
prohibited CPLEX from finding an optimal solution. However, in those cases
CPLEX was still able to find integer feasible suboptimal solutions. Figure 2 (b)
displays statistics of the real dataset. The number of candidate median genes
and adjacencies ranges from 442 to 18, 043 and 3, 164 to 2, 261, 716, respectively,
giving rise to up to 3, 227 median genes that are distributed on 5 to 91 CARs
per median. Some CARs are circular, indicating dubious conformations mostly
arising from tandem duplications, but the number of such cases were low (mean:
2.78, max: 13).
10
We observed that the gene families in the OMA database are clustered tightly
and therefore missing many true orthologies in the considered triples of genomes.
As a result, many of the orthologous groups inferred by FF-Median and MultiMSOAR fall into more than one gene family inferred by OMA. We therefore
evaluate our results by classifying the inferred orthologous groups into three
categories: An orthologous group agrees with OMA if its three genes are in the
same OMA group. It disagrees with OMA if extant genes x and y (of genomes
X and Y respectively) are in different OMA groups but the OMA group of x
contains another gene from genome Y . It is compatible with OMA if it neither
agrees nor disagrees with OMA. We measure the number of median genes as
well as the number orthologous groups of MultiMSOAR in each of the three
categories. Figure 2 (c) and (d) show the outcome this analysis. MultiMSOAR
is generally able to find more orthology relations in the dataset. This comes at
no surprise, as it is clear from the objective of problem FF-Median and from
the results of the simulated datasets that our method does not retain candidate
median genes which have lost their syntenic context, which happens in triples of
highly divergent genomes. The number of disagreeing orthologous groups that
disagree with OMA is comparably low for both FF-Median (mean: 35.16, var:
348) and MultiMSOAR (mean: 48.61, var: 348).
We then performed another analysis to assess the robustness of the positional
orthology predictions. To this end, we look at orthologous groups across multiple
datasets that share two extant genomes, but vary in the third. Given two genes,
x of genome X and y of genome Y , an orthologous group that contains x and
y is called robust if x and y occur in the same orthologous group, whatever
the third extant genome is. We computed the percentage of robust orthologous
groups for all gene pairs of genomes E. coli K-12 MG 1655 and S. enterica subsp.
enterica serovar Typhimurium str. 14028s in our dataset. The results indicate
that orthologous groups inferred by FF-Median are slightly more robust (95.61%)
than robust those by MultiMSOAR (91.77%). This is likely due to the strict
constraint of defining median adjacencies only from genes that participate in at
least one observed adjacency (Remark 4).
Overall, we can observe that FF-Median performed better than MultiMSOAR only for triples of closely related genomes – which is consistent with
our observation on simulated data – while being slightly more robust in general. This suggests FF-Median is an interesting alternative to identify higher
confidence positional orthologs, at the expense of a higher recall rate.
Future work. We first aim to investigate alternative methods to reduce the
computational load of Program FF-Median by identifying further strictly suboptimal and optimal substructures, which might require to understand better
the impact of internal conflicts within substructures defined by intervals in the
extant genomes. Without the need to modify drastically either the FF-median
problem definition or the ILP, one can think about more complex weighting
schemes for adjacencies that could account for known divergence time between
genomes or relaxed notion of adjacencies that would address the high recal rate
we observe in FF-Median. Within that regard, it would probably be interesting to combine this with the use of common intervals instead of runs to define
11
conflict-free sub-instances. Finally, ideal family-free analysis should take into account the effects of gene family evolution. However, the presented family-free
median model can only resolve certain cases of gene duplication. It is generally
susceptible to gene losses that occurred along the evolutionary paths between
the three extant genomes and their common ancestor. The definition of a familyfree median model that tolerates events of gene family evolution at a reasonable
computational cost is likely an interesting research avenue.
References
1. A M Altenhoff, N Skunca, N Glover, C-M Train, A Sueki, I Pilizota, K Gori,
B Tomiczek, S Müller, H Redestig, G H Gonnet, and C Dessimoz. The OMA
orthology database in 2015: function predictions, better plant support, synteny
view and other improvements. Nucleic Acids Res., 43(Database-Issue):240–249,
2015.
2. S F Altschul, W Gish, W Miller, E W Myers, and D J Lipman. Basic local alignment
search tool. J. Mol. Biol, 215(3):403–410, 1990.
3. A Bergeron, C Chauve, and Y Gingras. Formal models of gene clusters. In Bioinformatics algorithms: techniques and applications, pages 177–202. Wiley, 2008.
4. M D V Braga, C Chauve, D Doerr, K Jahn, J Stoye, A Thévenin, and R Wittler.
The potential of family-free genome comparison. In Models and Algorithms for
Genome Evolution, pages 287–323. Springer London, 2013.
5. D A Dalquen, M Anisimova, G H Gonnet, and C Dessimoz. Alf – a simulation
framework for genome evolution. Mol. Biol. Evol., 29(4):1115–1123, 2012.
6. C N Dewey. Positional orthology: Putting genomic evolutionary relationships into
context. Brief. Bioinformatics, 12(5):401–412, 2011.
7. D Doerr, A Thévenin, and J Stoye. Gene family assignment-free comparative
genomics. BMC Bioinformatics, 13(Suppl 19):S3, 2012.
8. D Dörr. Gene family-free genome comparison. PhD thesis, Universität Bielefeld,
Bielefeld, Germany, 2016.
9. W M Fitch. Homology a personal view on some of the problems. Trends Genet.,
16:227–231, 2000.
10. T Gabaldón and E V Koonin. Functional and evolutionary implications of gene
orthology. Nat. Rev. Genet., 14:360–366, 2013.
11. L A B Kowada, D Doerr, S Dantas, and J Stoye. New genome similarity measures
based on conserved gene adjacencies. In Proc. of RECOMB 2016, pages 204–224,
2016.
12. M Lechner, M Hernandez-Rosales, D Doerr, N Wieseke, A Thévenin, J Stoye,
R K Hartmann, S J Prohaska, and P F Stadler. Orthology detection combining
clustering and synteny for very large datasets. PLoS ONE, 9(8):e105015, 2014.
13. Marcus Lechner, Sven Findeiß, Lydia Steiner, Manja Marz, Peter F Stadler, and
Sonja J Prohaska. Proteinortho: Detection of (co-)orthologs in large-scale analysis.
BMC Bioinformatics, 12:124, 2011.
14. F V Martinez, P Feijão, M D V Braga, and J Stoye. On the family-free DCJ
distance and similarity. Algorithms Mol. Biol., 10:13, 2015.
15. J Misra and D Gries. A constructive proof of Vizing’s theorem. Inform. Process.
Lett., 41(3):131–133, 1992.
16. C H Papadimitriou and M Yannakakis. Optimization, approximation, and complexity classes. J. Comp. Sys. Sci., 43(3):425–440, 1991.
12
17. C Pesquita, D Faria, H Bastos, A E N Ferreira, A O Falcão, and F M Couto.
Metrics for GO based protein semantic similarity: a systematic evaluation. BMC
Bioinformatics, 9(Suppl 5):S4, 2008.
18. G Shi, M-C Peng, and Jiang T. Multimsoar 2.0: An accurate tool to identify
ortholog groups among multiple genomes. PLoS ONE, 6(6):e20892, 2011.
19. E Tannier, C Zheng, and D Sankoff. Multichromosomal median and halving problems under different genomic distances. BMC Bioinformatics, 10:120, 2009.
20. V G Vizing. On an estimate of the chromatic class of a p-graph. Diskret. Analiz
No., 3:25–30, 1964.
13
A
A.1
Hardness proof
Reduction
The maximum independent set problem for graphs bounded by node degree 3, denoted as MAX IS-3 is MAX SNP-hard [16]. The corresponding decision problem
can be informally stated as follows: Given a graph Λ bounded by degree 3 and
some number l ≥ 1, does there exists a set of vertices V ′ ⊆ V of size |V ′ | = l
whose induced subgraph is unconnected? In the following, we present a transformation scheme R to phrase Λ as FF-median instance R(Λ) = (G, H, I, σ) such
that the value F (M ) of a median M of R(Λ) is limited by F (M ) ≤ 2 · l + 3.
In doing so, we associate vertices of V with genes of extant genomes G, H and I.
In order to keep track of associated genes, we denote by function ξ(x) the list of
vertices associated with gene x. We further introduce two types of unassociated
genes, “∅” and “∗”, whose members are identified by subscript notation.
Transformation R:
1. Construct genome G such that for each vertex v ∈ V there exists two associated genes gv , ḡv ∈ C(G), i.e. ξ(gv ) = ξ(ḡv ) = v. Further, let each
gene pair gv , ḡv form a circular chromosome, giving rise to adjacency set
A(G) = {{ḡvh, gvt }, {ḡvh , gvt } | v ∈ V, gv , ḡv ∈ C(G)}.
2. For each edge (u, v) ∈ E construct a circular chromosome Xuv hosting two
genes xuv , x∅ ∈ C(Xuv ), with gene xuv being associated with both vertices u
and v and gene x∅ being unassociated. Further, let both genes form a circular
chromosome, giving rise to adjacency set A(Xuv ) = {{xhuv , xt∅ }, {xh∅ , xtuv }}.
3. Assign each chromosome constructed in the previous step either to genome
H or to genome I such that each vertex v ∈ V is associated with at most
two genes per genome.
4. Complete genomes H and I with additional circular chromosomes Xv where
C(Xv ) = {xv , x∅ } and A(Xv ) = {{xhv , xt∅ }, {xh∅ , xtv }} such that each vertex in
V is associated with exactly two genes per genome.
5. For each vertex v ∈ V , let g, ḡ ∈ C(G), h, h̄ ∈ C(H), and i, ī ∈ C(I) be the
pairs of genes associated with v, i.e. ξ(g) = ξ(ḡ) = ξ(h) ∩ ξ(i) = ξ(h̄) ∩ ξ(ī) =
v. Assign gene similarities σ(g, h) = σ(g, i) = σ(h, i) = 1 and σ(ḡ, h̄) =
σ(ḡ, ī) = σ(h̄, ī) = 1.
6. Add a copy of circular chromosome X∗ to each genome G, H, and I, where
C(X∗ ) = {x∗ , x̄∗ } and A(X∗ ) = {{xh∗ , x̄t∗ }, {x̄h∗ , xt∗ }}. Let g∗ , ḡ∗ ∈ C(G), h∗ , h̄∗ ∈
C(H), and i∗ , ī∗ ∈ C(I), set the gene similarity score between all pairs of
genes in {g∗ , h∗ , i∗ } and {ḡ∗ , h̄∗ , ī∗ } respectively, to 1. Lastly, set the gene
similarity score of all pairs of unassociated genes of type “∅” including genes
g∗ , ḡ∗ to 14 .
Except for step 3, none of the instructions of transformation scheme R are
computationally challenging. Note that in step 3 the demanded partitioning
of chromosomes into genomes H and I is always possible as consequence of
Vizing’s Theorem [20], by which every graph with maximum node degree d is
14
a
b
c
d
∗
ā
b̄
c̄
d̄
∗
ab
ād
cd̄
b̄
c̄
∗
∅
∅
∅
∅
∅
∗
bd
b̄c
a
ā
c̄
d̄
∗
∅
∅
∅
∅
∅
∅
∗
G
a
b
H
I
c
d
(a)
(b)
Fig. 3. (a) A simple graph bounded by degree three and (b) a corresponding FF-Median
instance constructed with transformation scheme R.
edge-colorable using at most d or d + 1 colors. Hence, using colors χ1 , χ2 , χ3 , χ4
with χ1 = χ2 ≡ I, χ3 = χ4 ≡ H and Misra and Gries’ algorithm [15], edges of
graph Λ = (E, V ) can be partitioned into two groups in O(|E||V |) time implying
an assignment to genomes H and I.
Example 1. Figure 3 (b) shows a FF-Median instance constructed with transformation scheme R from the simple graph depicted in Figure 3 (a). Gene similarities between genes are not shown, but can be derived from the genes’ labeling.
We structure our proof that the presented transformation is in fact a valid
mapping of an MAX IS-3 instance to an instance of FF-Median into three different lemmas:
Lemma 2. Given a median M of FF-Median instance R(Λ) = (G, H, I, σ),
(1) for each median gene (g, h, i) ∈ C(M ) where g, h, or i are associated with
vertices in V (Λ) holds ξ(g) = ξ(h) ∩ ξ(i) = v, v ∈ V (Λ); (2) there exist at most
two median genes whose corresponding extant genes are not associated to any
vertex in V (Λ).
Proof. Assume for contradiction that claim (1) does not hold. Then either ξ(g) 6=
ξ(h) ∩ ξ(i), or ξ(h) ∩ ξ(i) = ∅, both of which violate the constraint of establishing
gene similarities between associated genes that is given in step 5. For claim
(2), observe that the only unassociated genes in genome G are gene g∗ and ḡ∗
introduced in step 6, limiting the overall number of unassociated genes in any
median M .
⊓
⊔
Lemma 3. The conserved adjacency set of any median M of FF-Median int
h
t
G
h
stance R(Λ) = (G, H, I, σ) is of the form A(M )∩AC
= A (M )∪{{m∗ , m∗ }, {m∗ , m∗ }},
where the extant genes corresponding
h t to mC∗ and m∗ are all unassociated genes
of type “∗” and A(M )G
⊆
{m1 , m2 } ∈ A | ξ(πG (m1 )) = ξ(πG (m2 )) .
15
Proof. Observe that both candidate median adjacencies a∗ = {mh∗ , mt∗ } and
ā∗ = {mh∗ , mt∗ } are conserved in all three genomes, whereas all other conserved
candidate adjacencies between associated and unassociated genes can be at most
conserved in H and I. Establishing adjacencies a∗ , ā∗ gives rise to a cumulative
adjacency score of 6. Conversely, up to 4 non-conflicting adjacencies between
associated and unassociated genes can be established that are conserved in both
genomes H and I. However, since such adjacencies are only conserved between
unassociated genes of type “∅” whose gene similarities are set to 41 , the best
cumulative adjacency score can not exceed 4. Thus, adjacencies a∗ , ā∗ must be
contained in any median. Further, because of this and the fact that in both
genomes H and I, each gene associated with vertices of V (Λ) is only adjacent
to an unassociated gene, M cannot contain adjacencies that are conserved in
extant genomes other than genome G, which are the adjacencies of each gene
pair (gv , ḡv ) associated with the same vertex v ∈ V (Λ).
⊓
⊔
Lemma 4. Given FF-median instance R(Λ) = (G, H, I, σ), let mu , mv be any
pair of candidate median adjacencies of A whose corresponding extant genes
are associated to vertices u, v ∈ V (Λ), then mu , mv are conflicting if and only if
(u, v) ∈ E.
Proof. By construction in step 5 of transformation scheme R, each vertex v ∈ V
is associated with exactly two candidate median genes mv = (g, h, i), mv =
(ḡ, h̄, ī), mv , mv ∈ Σ , such that ξ(g) = ξ(h)∩ξ(i) = v and ξ(ḡ) = ξ(h̄)∩ξ(ī) = v.
Further, let u be another vertex of V (Λ), such that (u, v) ∈ E(Λ), and mu , mu
are its two corresponding candidate median genes. Then, by construction in step
2, there exists exactly one extant gene x with ξ(x) = uv (which, by assignment
in step 3, is either contained in genome H or I). Consequently, either mu is
in conflict with mv , or mu with mv , or mu with mv , or mu with mv . Recall
that by construction in step 2 in R and by Lemma 3, mu , mu and mv , mv form
conserved candidate adjacencies {mhu , mtu }, {mhu , mtu } and {mhv , mtv }, {mhv , mtv },
respectively. Clearly, independent of which of the candidate median gene pairs of
u and v are in conflict, both pairs of candidate median adjacencies are in conflict
with each other.
Now, let u, v be two vertices of V (Λ) such that edge (u, v) 6∈ E(Λ), then there
exists no gene x in extant genomes H and I with ξ(x) = uv. Even more, due
to Lemma 2, there cannot exist a candidate median gene (g, h, i) with {u, v} ⊆
ξ(g)∪ξ(h)∪ξ(i). Thus, the candidate median genes of u and v are not conflicting
and neither are their corresponding candidate median adjacencies.
⊓
⊔
We proceed to show that the given transformation scheme gives rise to an approximation preserving reduction known as L-reduction. An L-reduction reduces
a problem P to a problem Q by means of two polynomial-time computable transformation functions: A function f : P → Q′ ⊆ Q that maps each instance of
P onto an instance of Q, herein represented by transformation scheme R, and
a function g : Q′ → P to transform any feasible solution of an instance in Q′
to a feasible solution of an instance of P . Here, a feasible solution means any –
not necessarily optimal – solution that obeys the problem’s constraints. A feasible solution of FF-Median instance (G, H, I, σ) is an ancestral genome X where
16
C(X) ⊆ Σ and A(X) ⊆ A such that A(X) is conflict-free. We give the following transformation scheme to map ancestral genomes of an FF-Median instance
to solutions of an MAX IS-3 instance:
Transformation S: Given any ancestral genome X of R(Λ), return {ξ(πG (m1 )) |
{ma1 , mb2 } ∈ A(X) : IG (πG (m1 )a , πG (m2 )b ) = 1 and ξ(πG (m1 )) 6= ∅}.
We define score function s (X) ≡ 12 F (X) − 3 of an ancestral genome X. For
(R, S) to be an L-reduction the following two properties must hold for any given
MAX IS-3 instance (Λ, l): (1) There is some constant α such that for any median
M of the transformed FF-Median instance R(Λ) holds s (M ) ≤ α·l; (2) There is
some constant β such that for any ancestral genome X of R(Λ) holds l−|S(X)| ≤
β · |s (M ) − s (X)|. We proceed to proof the following lemma:
Lemma 5. (R, S) is an L-reduction of problem MAX IS-3 to problem FF-Median
with α = β = 1.
Proof. For any median M of FF-Median instance R(Λ), the number of conserved median adjacencies with correspondence to the same vertex of Λ is two,
giving rise a cumulative adjacency score of two. From Lemmata 3 and 4 immediately follows that any ancestral genome of R(Λ) that maximizes the number of
conserved adjacencies also maximizes the number of independent vertices in Λ.
Recall that the two conserved adjacencies between unassociated genes of type
“∗” (which are part of all medians) give rise to a cumulative adjacency score of
1
6, we conclude that |A(M ) ∩ AC
| − 2 = 2 F (M ) − 3 = s (M ) = l, thus α = 1.
Because l = sλ (M ), it remains to show that l−|S(X)| ≤ β|l−s (X)|. In a suboptimal ancestral genome of R(Λ), median genes with no association to vertices
of Λ can also contain extant genes of type “∅”. These unassociated median
genes can form “mixed” conserved adjacencies with genes that are associated
with vertices of Λ. Such mixed conserved adjacencies have no correspondence
to vertices in Λ and do not contribute to the transformed solution S(X) of an
ancestral genome X. Yet, as mentioned earlier, the cumulative adjacency score
of all mixed conserved adjacencies can not not exceed 4. Therefore it holds that
|S(X)| ≥ s (X) and we conclude β = 1.
⊓
⊔
B
Speeding up the search for a median
Proof of Lemma 1:
Proof. Given an IC-free segment S = {m1 , . . . , mk } of an FF-Median instance
(G, H, I, σ). Let M be a conflict-free matching in graph Γ ′ (S). Because M is
conflict-free and S contiguous in all three extant genomes, M must contain all
candidate median genes of S. Now, let M be a median such that S 6⊆ C(M ′ ).
Further, let C(m) be the set of candidate median genes that are
S involved in a
conflict with with a given median gene m of S and X = C(M ′ )∩( m∈S C(m)∪S).
Clearly, X 6= ∅ and for the contribution F (X) must hold F (X) ≥ F (S),
17
otherwise M ′ is not optimal since it is straightforward to construct a median
higher score which P
includes S. Clearly, the contribution F (X) to the median is
bounded by max({ m′ ∈C ′ ∆(m′ ) | C ′ ⊆ C(m) : C ′ is conflict-free}) + F (S).
But since S gives rise to a conflict-free matching with maximum score, also
median M ′′ with C(M ′′ ) = (C(M ′ ) \ X) ∪ C(S) and A(M ′′ ) = (A(M ′ ) \ A(X)) ∪
A(S)) must be an (optimal) median.
⊓
⊔
C
Simulated sequence evolution with ALF
PAM Genome Inversions Transpositions Duplications Losses
10
G
H
I
8.7
7.3
8.5
6.1
4.5
6.6
7.3
6.3
10.4
6.9
5.4
5.6
30
G
H
I
24.5
23.4
25.5
16.9
19.8
17.2
21.0
20.6
17.5
22.7
18.4
20.9
50
G
H
I
39.9
41.8
43.2
27.8
31.8
30.0
32.4
31.0
28.7
36.7
31.7
39.7
70
G
H
I
58.6
57.0
60.4
42.3
43.6
41.4
41.1
46.3
40.7
39.2
45.1
39.1
90
G
H
I
75.0
69.9
75.2
54.5
50.5
55.5
53.1
54.1
60.3
64.2
65.0
58.5
110
G
H
I
96.3
90.6
90.2
69.4
64.2
68.5
67.0
62.5
62.6
74.6
70.9
61.2
130
G
H
I
105.7
108.7
110.8
76.3
78.2
73.6
74.4
79.6
73.9
81.0
82.8
77.3
Table 1. Average benchmark data of seven evolutionary distances, each comprising
ten genomic datasets generated by ALF [5].
18
Parameter name Value
sequence evolution
substitution model WAG (amino acid substitution model)
insertion and deletion Zipfian distribution
insertion rate
maximum insertion length
rate variation among sites Γ -distribution
number of classes
rate of invariable sites
exponent c = 1.8214
0.0003
50
shape parameter a = 1
5
0.01
genome rearrangement
inversion rate
maximum inversion length
transposition rate
maximum transposition length
rate of inverted transposition
0.0004
100
0.0002
100
0.1
gene family evolution
gene duplication rate
max. no. of genes involved in one dupl.
probability of transposition after dupl.
fission/fusion after duplication
probability of rate change
rate change factor
probability of temporary rate change (duplicate)
temporary rate change factor (duplicate)
life of rate change (duplicate)
probability of temporary rate change (orig+duplicate)
temporary rate change factor (orig+duplicate)
life of rate change (orig+duplicate)
gene loss rate
maximum length of gene loss
gene fission/fusion rate
maximum number of fused genes
0.0001
5
0.5
0.1
0.2
0.9
0.5
1.5
10 PAM
0.3
1.2
10 PAM
0.0001
5
0.0
−
Table 2. Parameter settings for simulations generated by ALF [5].
D
Real genomes dataset
19
Genbank ID Name
U00096.3
AE004439.1
AE016853.1
AM039952.1
CP000266.1
CP000305.1
CP000569.1
CP000744.1
CP000766.3
CP000950.1
CP001120.1
CP001172.1
CP001363.1
FM180568.1
CP002086.1
Escherichia coli str. K-12 substr. MG1655, complete genome
Pasteurella multocida subsp. multocida str. Pm70, complete genome
Pseudomonas syringae pv. tomato str. DC3000, complete genome
Xanthomonas campestris pv. vesicatoria complete genome
Shigella flexneri 5 str. 8401, complete genome
Yersinia pestis Nepal516, complete genome
Actinobacillus pleuropneumoniae L20 serotype 5b complete genome
Pseudomonas aeruginosa PA7, complete genome
Rickettsia rickettsii str. Iowa, complete genome
Yersinia pseudotuberculosis YPIII, complete genome
Salmonella enterica subsp. enterica serovar Heidelberg str. SL476, complete genome
Acinetobacter baumannii AB307-0294, complete genome
Salmonella enterica subsp. enterica serovar Typhimurium str. 14028S, complete genome
Escherichia coli 0127:H6 E2348/69 complete genome, strain E2348/69
Nitrosococcus watsoni C-113, complete genome
Table 3. Dataset of genomes used in comparison with the OMA database.
20
| 8cs.DS
|
arXiv:1707.05163v2 [physics.ao-ph] 26 Sep 2017
Parcels v0.9: prototyping a Lagrangian Ocean Analysis framework
for the petascale age
Michael Lange1 and Erik van Sebille2,3
1
Grantham Institute & Department of Earth Science and Engineering, Imperial College London, UK
Institute for Marine and Atmospheric research Utrecht, Utrecht University, Utrecht, Netherlands
3
Grantham Institute & Department of Physics, Imperial College London, UK
2
Correspondence to: Erik van Sebille ([email protected])
Abstract.
As Ocean General Circulation Models (OGCMs) move into the petascale age, where the output of single simulations exceeds
petabytes of storage space, tools to analyse the output of these models will need to scale up too. Lagrangian Ocean Analysis,
where virtual particles are tracked through hydrodynamic fields, is an increasingly popular way to analyse OGCM output, by
mapping pathways and connectivity of biotic and abiotic particulates. However, the current software stack of Lagrangian Ocean
Analysis codes is not dynamic enough to cope with the increasing complexity, scale and need for customisation of use-cases.
Furthermore, most community codes are developed for stand-alone use, making it a nontrivial task to integrate virtual particles
at runtime of the OGCM. Here, we introduce the new Parcels code, which was designed from the ground up to be sufficiently
scalable to cope with petascale computing. We highlight its API design that combines flexibility and customisation with the
ability to optimise for HPC workflows, following the paradigm of domain-specific languages. Parcels is primarily written in
Python, utilising the wide range of tools available in the scientific Python ecosystem, while generating low-level C-code and
using Just-In-Time compilation for performance-critical computation. We show a worked-out example of its API, and validate
the accuracy of the code against seven idealised test cases. This version 0.9 of Parcels is focussed on laying out the API, with
future work concentrating on support for curvilinear grids, optimisation, efficiency and at-runtime coupling with OGCMs.
1
Introduction
Lagrangian Ocean Analysis, where virtual particles are tracked within the flow field of hydrodynamic models, has over the last
two decades increasingly been used by physical oceanographers and marine biologists alike (Van Sebille et al., submitted).The
particles can represent passive parcels of seawater (e.g. Döös, 1995; Blanke and Raynaud, 1997) or its constituent tracers such
as nutrients (e.g. Jönsson et al., 2011; Qin et al., 2016), as well as particulate matter such as microbes (e.g. Hellweger et al.,
2014; Doblin and van Sebille, 2016), larvae (e.g. Cowen et al., 2006; Paris et al., 2005; Teske et al., 2015; Cetina-Heredia et al.,
2015), pumice (e.g. Jutzeler et al., 2014), plastic litter (e.g. Lebreton et al., 2012), or icebergs (e.g. Marsh et al., 2015). The
trajectories of the virtual particles can be used to analyse the flow within Ocean General Circulation Models (OGCMs) and
other velocity fields for dispersion characteristics (e.g. Beron-Vera and LaCasce, 2016), Lagrangian Coherent Structures (e.g.
Haller, 2015), water mass pathways and transit times (e.g. Rühs et al., 2013), Lagrangian streamfunctions (e.g. Döös et al.,
1
2008) and biological connectivity between regions (e.g. Kool et al., 2013). See Van Sebille et al. (submitted) for an extensive
review on Lagrangian Ocean Analysis.
There are currently three main community codes available to calculate the trajectories of virtual particles in Ocean General
Circulation Models: Ariane (Blanke and Raynaud, 1997), TRACMASS (Döös et al., 2013; Döös et al., 2017), and the Connectivity Modeling System (CMS, Paris et al., 2013). These codes, being open-source and having excellent support teams, have
served the wider community very well over the past decades. However, it is not clear that these three codes will be able to
scale up easily to the petascale age of computing, where particle trajectory codes will need to be able to deal with potentially
petabytes of hydrodynamic field data and gigabytes of particle trajectory data. Exploring advanced optimization strategies to
overcome these big-data challenges, such as coupled (online) execution with the host OGCM or reducing the volume of hydrodynamic data by selectively filtering data regions based on particle locations, will require a flexible execution model that can
dynamically be adapted to complement the respective data and execution formats of various host OGCMs.
Furthermore, the current stack of codes is mostly built for the tracking of water parcels or passive particulates. While the
CMS and TRACMASS do support the addition of diffusive processes through Markovian stochastic models (e.g. Griffa, 1996),
it is non-trivial to incorporate ‘behaviour’ of particulates to these codes. Effortless incorporation of behaviour such as sinking,
fragmentation, or even swimming to particulates would simplify exploration of the dynamics of particulates such as fish,
icebergs and marine debris.
Here, we describe a novel framework for computing Lagrangian particle trajectories, named Parcels (‘Probably A Really
Computationally Efficient Lagrangian Simulator’). Being developed from the ground up with scalability and performance
in mind, we hope that this Parcels framework will be able to keep up with OGCM development for the coming decades,
particularly by being scalable and efficient at reading in hydrodynamic data. We have furthermore focussed on flexibility and
customisability of the particle dispersion schemes, so that it is relatively straightforward to add new functionality such as active
particle behaviours.
We have decided to brand this version of Parcels as v0.9, signalling that while in principle it is feature-complete, the code
is not nearly as fast and efficient as we envision it to be in the future. Improving performance will be the main priority as we
work towards v1.0. We invite all interested researchers to contribute to the development by starting to use the code.
While development efforts of Parcels focus on oceanographic applications, the Parcels framework should in principle also be
adaptable to atmospheric particle tracking simulations. Models such as FLEXPART (Stohl and James, 2005) and the MetOffice
NAME model (Jones et al., 2007) are state-of-the-art and have an excellent track-record in the field of atmospheric dispersion
modelling, but perhaps some of the ideas presented here could be incorporated or used in these models too.
This paper is structured as follows: in the next section, we will describe the philosophy behind the Parcels code. We then
present a worked-out example of an application of Parcels for an actual scientific experiment in Sect. 3. Following that, we
evaluate the accuracy of the code in Sect. 4, by comparison to analytical solutions in idealised test. We provide a future outlook
in Sect. 5, before concluding in Sect. 6.
2
2
Prototype design and philosophy
A key contribution of the new Parcels v0.9 framework is to define a set of interfaces and composable abstractions that encapsulate the various processes required to create dynamic and extensible Lagrangian models that feature direct interactions between
particles and an associated OGCM grid. The design follows modern scientific software engineering practices, providing high
levels of modularity and flexibility with a clear intent to further specialize various sub-components at a later stage. The interfaces provided in Parcels are therefore intended to capture the general domain-specific challenges posed by particle tracking
for Lagrangian Ocean Analysis. The overall design philosophy, as well as the structure of the code, are driven by three major
design considerations:
– Extensibility – While the core algorithm of Lagrangian particle models is concerned with the advection and dispersion
of passive particles that constitute infinitely small point parcels, practical oceanographic applications often require more
complex behaviour of the particles. Potential extensions towards individual-based modelling of particulates to simulate
biological species or marine debris will require extensions to particle data definitions and programmable behavioural
customisation at a per-particle level.
– Compatibility – Particle tracking in oceanography requires the close coupling of computational particles to velocity
data that defines the hydrodynamic flow field. Parcels aims to make as little assumptions about the nature and structure
of the hydrodynamic fields as possible, so as to be compatible with various types of OGCMs and data formats. While
the focus in this v0.9 is on utilising offline data, this includes considerations for at-runtime coupling with OGCMs in the
future.
– Dynamic data – Particle data is sparse in nature and can, depending on application context, exhibit very dynamic data
access patterns where new particles are inserted and deleted from the active set at runtime. For this reason, structured
compile-time performance optimisations and parallelisation strategies are insufficient, and Just-In-Time scheduling is
required to handle the amorphous data parallelism inherent in dynamic particle applications (Pingali et al., 2011).
The above list of requirements suggests that a static compile-time approach is likely to provide insufficient flexibility to
adjust to the various scientific contexts in which oceanographic particle tracking might be utilised. For this reason Parcels is
based on the domain-specific languages paradigm, which aims to decouple the problem definition as defined by the scientific
modeller from the implementation that is ultimately executed on a particular hardware architecture. This approach is based
on automated code generation at runtime and creates a separation of concerns between domain scientists and computational
experts that allows hardware-specific performance optimisation and thus greater flexibility with respect to advances in highperformance computing resources.
Since the prototype of the Parcels framework presented here provides a conceptual blueprint for future versions, we define a
clear set of abstractions for the following three software layers:
– User-facing API – The primary objective of Parcels is to provide a user-friendly, clear and concise API for scientists to
perform oceanographic particle tracking experiments with very little effort, while leaving room for customisations that
3
go beyond traditional configuration files. For this reason Parcels provides a high-level Python API that enables users to
define a complete model in a small number of lines of code (see examples in Sect. 3). For more advanced models, the
API also provides enough scope to fully control the variable layout of particles in memory, as well as to define custom
behaviour via individual kernel operations.
– Execution layer – The transient nature of Lagrangian particles implies that many practical oceanographic applications
rely on particle sets that may grow and shrink dynamically, while also relying on external hydrodynamic field data
that might be sampled at a timestep much different from the primary particle loop. This complex parameter variability
entails that the core loop that updates individual particle states needs to be highly dynamic and flexible, as well as highly
optimised for large-scale applications. Parcels aims to encapsulate the core parameters of the particle update loop so
as to establish an interface for integration with a variety of external host OGCMs, and leaves enough scope for more
advanced performance optimisations in the future.
– Data layout – The two fundamental types of data involved in Lagrangian particle tracking algorithms constitute field data
provided by the external OGCM, as well as data on the particle state. Since the data layout for particle data might change
with future performance optimisations, and the memory layout of field data depends on the OGCM implementation,
Parcels provides high-level abstractions for both types of data, allowing the actual data layout in memory to change.
The abstractions shown in Fig. 1 comprise the core functionalities provided by the framework. The primary input in the user
layer consists of generic definitions of the particle variables for individual types of particles, alongside an interface to define
the computation kernels. Parcels’ core execution loop uses this information to update particle data given external parameters,
such as timestepping constraints, and interpolated hydrodynamic field data. Thus, given a stable user-level API and a highly
modular code structure, it is possible to implement various applications and experiments without commiting to a particular
implementation, while leaving enough scope for further development and future performance optimisation ‘under the hood’.
2.1
Programmable user interface
The prototype presented in this paper provides a highly flexible user API that allows users to define complete models via the
Python programming language. The user hereby manages creation, execution and customisation of individual sets of particles,
as well as combinations of computational kernels to update the particle state. In contrast to traditional configuration files, this
approach provides the user with native compatability with the open-source libraries and tools available in the scientific Python
ecosystem.
The key components of Parcels’ overall class structure are depicted in Fig. 2. The definition of the variables that constitute
a single particle is hereby encapsulated in the Particle class, while container objects of type ParticleSet provide the
runtime handling and management of particle data. Python descriptor objects are used to generically define the compound data
type underlying each type of particle, leaving allocation and memory layout choices to the particular implementation of the
data container structure.
4
The computational behaviour of particles is encapsulated through the Kernel. Parcels provides a set of pre-defined advection methods, as well as allowing users to define custom behaviour programmatically. Multiple kernels can be concatenated,
allowing users to incrementally build complex behaviour from individual components.
2.1.1
Advection algorithm
At its core, computing Lagrangian particle trajectories is equivalent to solving the following equation:
t+∆t
Z
v(x, τ ) dτ + ∆X b (t),
X(t + ∆t) = X(t) +
(1)
t
where X is the three-dimensional position of a particle, v(x, t) is the three-dimensional velocity field at that location from an
OGCM, and ∆X b (t) is a change in position due to ‘behaviour’. The latter can itself be an integration of a (three-dimensional)
velocity field, for example when a particle sinks downward because of a negative buoyancy force.
In Parcels, the trajectory equation (1) is by default time-stepped using a 4th order Runge-Kutta scheme, although schemes for
Euler-Forward and adaptive Runge-Kutta-Fehlberg integration (RKF45, e.g. Alexander, 1990) are also provided. In principle,
the Parcels framework should be flexible enough to also implement integration using the discrete analytical streamtube method
(Blanke and Raynaud, 1997; Döös et al., 2017).
2.1.2
Custom kernels
Lagrangian particle tracking in the ocean often involves more complex displacement schemes than simple velocity-driven advection. For example, in the presence of turbulence, a Random Walk kernel or Brownian motion is required, while ocean ecology models often include active locomotion. Parcels therefore allows users to create generic kernel functions by providing native Python functions that adhere to the function signature KernelName(particle, fieldset, time, dt). Within
these kernel functions, users can access built-in particle state variables, such as particle.lat and particle.lon, or
user-defined ones. Access to field data from within kernels is provided through the fieldset object, which provides fields as
named properties, for example fieldset.U for the zonal velocity. Interpolation of field data is implemented via overloaded
member access on the field object (square bracket notation), allowing user to express field sampling as fieldset.fieldname[time,
lon, lat, depth].
In addition to kernels that update the internal state of particles, Parcels’ execution engine also enables users to customize
the behaviour of particles under various error conditions. For this, a similar type of kernel function can be created and passed
to the execution call, mapped to a particular error type that might be triggered during the main particle update, for example
OutOfBoundsError.
5
2.2
Execution and JIT compilation
The update of the internal state of particles is facilitated by a dynamic loop, which applies a user-defined combination of kernels
to each particle in a ParticleSet. The primary particle update loop can either be run with a forward timestepping, or in a
time-backward mode to enable inverse modelling. For this central update loop, Parcels provides two modes of execution:
– Scipy mode: A pure Python mode that utilises interpolator objects provided by the Scientific Python package
(SciPy) to perform interpolation of field data. This mode is primarily intended as a debug option due to the performance
penalty of running kernels in the Python interpreter itself.
– JIT mode: Runtime code generation and Just-In-Time compilation (JIT) are utilised to generate low-level C code that
performs the particle state update and field data interpolation. The code generation engine hereby primarily translates
a restricted subset of the Python language into equivalent C code, while a set of utility modules provides auxiliary
functionality such as random number generation or mathematical utilities (math.h).
The execution mode of the particle update loop is determined by the type of the particle (ScipyParticle or JITParticle)
used to create the ParticleSet. Development of new features in the current Parcels prototype is strongly driven by the fact
that both modes are intended to be semantically equivalent. This means that new features can rapidly be developed using the
full flexibility of the Python interpreter, providing a template implementation and test case for implementation in the computationally more efficient JIT mode.
Parcels’ dynamic update loop also provides an interval keyword to impose a secondary sub-timestepping that allows
for direct coupling with a host OGCM in the future. The dynamic composition of multiple timestepping intervals might also
be used for future data and performance optimisation strategies, for example directed prefetching of regional field data. Such
strategies, as well as a potentially more intricate execution engine, have to be explored carefully to successfully tackle the
big-data challenges facing Lagrangian tracking codes in the petascale age.
2.3
Interpolation
The interaction of particles with their enclosing fields is currently limited to interpolating field data onto the current particle
position. In the SciPy debug mode this is facilitated by scipy.interpolate.RegularGridInterpolator objects
and supports linear and nearest-neighbour interpolation. Equivalent low-level C routines are also included in the Parcels source
code as macros that can be inlined into the generated C kernel code by the code generation engine. More advanced interpolation
methods, such as quadratic, cubic or spline interpolation, may easily be added in future releases if a fast C implementation can
be provided with Parcels’ internal header files.
One of key performance advantages of using runtime code generation is the ability to inline bespoke grid interpolation
methods with the user-defined kernels in Parcels to avoid the Python interpreter overhead of repeatedly calling native Python
interpolation functions. This overhead can be quite significant due to the high frequency at which the associated field data needs
to be sampled. This can be illustrated using the “Steady-state flow around a peninsula” test case discussed in Sect. 4.2.4, where
6
100 particles are advected for 20 hours with a timestep size of 30 seconds. While the sequential execution time of the pure
Python implementation runs in 305.92 seconds, the auto-generated JIT kernels can run the same experiment in 1.74 seconds, a
speedup of over 150×.
2.4
External field data
Parcels v0.9 supports external field data from NetCDF files, with a configurable interface to describe the input data and variable
structure. The data is encapsulated in individual Field objects, which are accessible from within particle kernels via provided
interpolation routines. Individual fields are stored in a FieldSet container class, which may also provide global meta-data
to the kernel execution engine at runtime.
Currently, only linear interpolation schemes are implemented in Parcels, both in space and in time. In space, Parcels can
currently only work on regular grids (i.e. where the grid dimensions are functions of only longitude, only latitude or only
depth). However, support for unstructured grids is a priority for the next release of the code, Parcels v1.0.
3
A worked-out example: tracking virtual foraminifera in the Agulhas region
To highlight some of the prototype design and philosophies of the Parcels API, we here present a worked-out example code of
a previously-published scientific experiment. This example follows the experimental design of Van Sebille et al. (2015), where
the goal was to investigate the temperatures that planktic foraminifera experience during their lifespan as they drift with the
currents in the upper ocean. In particular, that study looked at the variability of lifespan-averaged temperatures of foraminifera
that all end up on one single location on the ocean floor (e.g. Peeters et al., 2004; Katz et al., 2010).
Figure 1b of Van Sebille et al. (2015) depicted the origin of virtual planktic foraminifera that end up on a site just off the
coast of Cape Town (17.3◦ E, 34.7◦ S), at 2,440 meter water depth. The virtual particles were released at that site and then
tracked in time-backward mode. There were two phases to the experiment: in the sinking phase, the foraminifera were tracked
back as they sunk at 200 meter per day to the ocean floor, while being advected by the (deep) ocean circulation. In the lifespan
phase, the particles were then tracked further backward in time as they were advected by the horizontal circulation at their 50m
dwelling depth. During this last phase, temperature along their trajectory was recorded at daily interval.
While the original experiment was computed with the Connectivity Modelling System (Paris et al., 2013), here we have
re-coded it using the Parcels API. This experiment setup is a fitting one, as it combines a number of the API highlights of
Parcels: custom kernels, NetCDF I/O, and field sampling. The full Python code for this experiment in Parcels is available at
https://doi.org/10.5281/zenodo.823994. Below, we emphasise some of the key statements in the Python script.
3.1
Reading the FieldSet
The hydrodynamic fields that carry the foraminifera come from the OFES model (Masumoto et al., 2004) and can be accessed from http://apdrc.soest.hawaii.edu/datadoc/ofes/ncep_0.1_global_3day.php. Three-dimensional velocities and temperature are available on 1/10◦ horizontal resolution, on 54 vertical levels, and are stored as three-day averages. The bash script
7
get_ofesdata_agulhas.sh provided at https://doi.org/10.5281/zenodo.823994 was used to download snapshot numbers 3165 to 3289, covering the year 2006, in a subdomain around the core site off Cape Town (note, the total file size is
6GB).
While the 6GB file size for this example is not excessively large and could in principle be loaded into memory all at once,
this will not be possible for FieldSets with larger regional domains or longer time series. Hence, Parcels provides a system
to read in hydrodynamic fields during particle integration, at any time storing only three consecutive timeslices (e.g. Paris et al.,
2013). See also Section 3.4.
After the first three days of hydrodynamic fields are read in through a call to the user-defined set_ofes_fieldset
function (see the example_corefootprintparticles.py script for the exact formulation of this function, which
requires as input a set with filenames, provided as a list of arbitrary length), three global constants are added to the FieldSet
fieldset . add_constant ( ’ dwellingdepth ’ , 50.)
fieldset . add_constant ( ’ sinkspeed ’ , 200./86400)
fieldset . add_constant ( ’ maxage ’ , 30.*86400)
These constants will be used later in the custom kernels controlling the movement of the particles.
3.2
Defining the ParticleSet
Apart from information on their location and time, the virtual foraminifera particles will need two extra Variables: the sea
water temperature at their present location, and their age. Therefore, we define a new particle class, which inherits from the
standard JITParticle:
class ForamParticle ( JITParticle ):
temp = Variable ( ’ temp ’ , dtype = np . float32 , initial = np . nan )
age = Variable ( ’ age ’ , dtype = np . float32 , initial =0.)
And we then define a ParticleSet containing a single particle as
pset = ParticleSet ( fieldset = fieldset , pclass = ForamParticle , lon =[17.3] , lat =[ -34.7] ,
depth =[2440] , time = fieldset . U . time [ -1])
3.3
Defining the custom kernels
We need to define four custom kernels: one that causes the particle to sink after it dies, one that keeps track of its age and deletes
it once it reaches its maximum age, one that samples the temperature at its location, and one that deletes the particle when it
reaches a boundary of the domain (since we only have hydrodynamic data in a subset of the global OFES domain). Note that
while in principle the first three could be written in one Kernel, here we write three separate kernels and then concatenate these
with the built-in AdvectionRK4_3D kernel.
The first kernel, controlling the sinking of the particle after it died (i.e. the first twelve days in our reverse-time experiment),
can be written as
def Sink ( particle , fieldset , time , dt ):
if particle . depth > fieldset . dwellingdepth :
particle . depth = particle . depth + fieldset . sinkspeed * dt
else :
particle . depth = fieldset . dwellingdepth
8
The second kernel, which keeps track of the age and deletes the particle when it reaches maxage, can be written as
def Age ( particle , fieldset , time , dt ):
if particle . depth <= fieldset . dwellingdepth :
particle . age = particle . age + math . fabs ( dt )
if particle . age > fieldset . maxage :
particle . delete ()
The third kernel, which samples the temperature, can be written as
def SampleTemp ( particle , fieldset , time , dt ):
particle . temp = fieldset . temp [ time , particle . lon , particle . lat , particle . depth ]
These three kernels are then concatenated with the AdvectionRK4_3D kernel as
kernels = pset . Kernel ( AdvectionRK4_3D ) + Sink + SampleTemp + Age
Where at least one of the kernels needs to be cast into a Kernel object for the overloading of the + operator as a kernel
concatenator to work.
Finally, the kernel that deletes a particle if it reaches one of the lateral boundaries and which will be invoked through the
error recovery execution is
def DeleteParticle ( particle , fieldset , time , dt ):
particle . delete ()
3.4
Executing the particle set
The ParticleSet can now be integrated with a call to pset.execute(). This method requires as input the list of kernels,
the starttime of the execution loop, the runtime of the execution loop, the Runge Kutta integration timestep (here taken to be
5 minutes), the interval at which output is written (here once per day), and the recovery kernel that gets called when a Particle
crosses the boundary of the regional domain.
As mentioned in section 3.1, only three timeslices are held in memory at any one time. The loading of new fields is controlled
by the fieldset.advancetime() method, which replaces the oldest timeslice with a new one (held in this case in
[snapshots[s]]). This also means that the executing of the ParticleSet has to be done within a loop:
for s in range ( len ( snapshots ) -5 , -1 , -1):
pset . execute ( kernels , starttime = pset [0]. time , runtime = delta ( days =3) ,
dt = delta ( minutes = -5) , interval = delta ( days = -1) ,
recovery ={ ErrorCode . ErrorOutOfBounds : DeleteParticle })
fieldset . advancetime ( set_ofes_fieldset ([ snapshots [ s ]]))
There is another reason to call the pset.execute method within a loop: it allows for a new particle to be released every
three days (the frequency with which hydrodynamic data is available). This happens within the for-loop through a call to
pset . add ( ForamParticle ( lon =[17.3] , lat =[ -34.7] , depth =[2440] , fieldset = fieldset ))
3.5
Saving and plotting the output
The Parcels framework allows for storing of the locations of the particle to disk on-the-fly in NetCDF files, following the Discrete Sampling Geometries section of http://cfconventions.org/cf-conventions/v1.6.0/cf-conventions.html\#discrete-sampling-geometries,
9
and is hence CF-1.6-compliant. Storing of the particle trajectories and properties such as age and along-track temperature happens in the for-loop through calls to
pfile . write ( pset , pset [0]. time )
Since particles are continually added to and deleted from the ParticleSet, the ParticleFile needs to be stored in ‘indexed’ format, where for each variable all particle states are written in one long vector.
pfile = ParticleFile ( outfile , pset , type = " indexed " )
These long vectors in Indexed format, however, are not very easy to work with, so Parcels provides the utility script
convert_IndexedOutputToArray to convert an Indexed NetCDF file to array format.
The particle trajectories can then be plotted using the matplotlib and Basemap libraries, see Fig. 3. This figure shows
the temperature recorded on each day during the lifespan of all virtual particles. It highlights that foraminifera that end up on
the ocean floor off Cape Town travel hundreds to thousands of kilometers during their lifespan, and that while some originate
from the Agulhas Current as far north as 27◦ S, others originate from the much colder Southern Ocean south of 40◦ S.
4
Model evaluation
Evaluation of a code-base’s accuracy and performance is a key component of its validation and roll-out. For this Parcels v0.9,
performance and speed are not a priority; these will be the focus for the v1.0 release (see also Sect. 5). Instead, while developing
Parcels v0.9 we have concentrated on accuracy.
4.1
Unit tests and continuous integration
Following best practices in software engineering, we have incorporated Unit Testing and Continuous Integration into the
development cycle of Parcels. Every push of code changes to github automatically triggers a validation of the entire code base
(an important component of the Continuous Integration paradigm), through the travis-ci.org web service.
The validation of the code base is done through so-called unit tests; small snippets of code that test individual components of
the codebase. Parcels v0.9 has over 150 of these unit tests, which check the integrity and consistency of the codebase. Where
relevant, these unit tests are run in both Scipy and JIT mode, to test both modes of executing the kernels.
The following Python snippet shows a typical example of a unit test for Parcels (as included in the test_particle_sets.py
file). It performs the test that Particles in a ParticleSet indeed get their assigned longitudes and latitudes. While this
may seem a trivial test, these kinds of unit tests can help prevent bugs.
@pytest . mark . parametrize ( ’ mode ’ , [ ’ scipy ’ , ’ jit ’ ])
def t e st _ p s e t _ c r e a t e _ l o n _ l at ( fieldset , mode , npart =100):
lon = np . linspace (0 , 1 , npart , dtype = np . float32 )
lat = np . linspace (1 , 0 , npart , dtype = np . float32 )
pset = ParticleSet ( fieldset , lon = lon , lat = lat , pclass = ptype [ mode ])
assert np . allclose ([ p . lon for p in pset ] , lon , rtol =1 e -12)
assert np . allclose ([ p . lat for p in pset ] , lat , rtol =1 e -12)
10
Ideally, the full set of unit tests means that no change of the code can ever break another part of the code, since some of the
unit tests would then fail. Of course, in reality the completeness of the unit tests can never be guaranteed, but during Parcels
development we have attempted to provide unit tests for a broad spectrum of the Parcels functionality and code.
4.2
Idealised and analytic test cases
Following the list of standard tests of particle tools, as described in Sec. 6 of Van Sebille et al. (submitted), we have validated
the accuracy of Parcels v0.9 against seven idealised and analytical test cases. In this section we will describe the results in
detail. All test cases are run with Runge-Kutta4 integration and in JIT mode. In each case, the hydrodynamic velocities are
generated within the Python scripts and converted directly to a FieldSet (i.e. without first storing these fields in NetCDF
format).The Python code for all testcases is available at https://doi.org/10.5281/zenodo.823994.
4.2.1
Radial rotation with known period
The first test case is that of a simple counter-clockwise solid-body rotation with a period of 24 hours. Velocities are defined on
a (20 × 20) km Arakawa A-grid centered at the origin with a 100 m horizontal resolution. Solid-body radial velocities (u, v) =
(−ωr sin(φ), ωr cos(φ)), with r and φ the radius and angle from the origin and ω = 2π/86, 400 s the angular frequency, are
then computed on that grid.
Four particles are started at x = 0 km and y = (1000, 2000, 3000, 4000) km and then advected for 24 hours, using an RK4
timestep of 5 minutes, and with particle positions stored every hour (Fig. 4a). All four particles indeed follow the flow for the
full circle. The maximum distance error after this 24 hours advection is less than 3mm, on path lengths of more than 5km.
4.2.2
Longitudinal shear flow
The second test case tests the ability of the Parcels code to convert between spherical longitude/latitude space and local flat
Euclidian space. When defining a FieldSet on a spherical mesh, Parcels automatically performs this conversion under the
hood. To test its accuracy, an idealised flow on a sphere at 1◦ horizontal resolution is created, with a uniform zonal velocity of
1 m/s and no meridional velocity. A total of 31 particles are then released on a north-south line, with a meridional spacing of
3◦ . These particles are advected for 57 days, using an RK4 timestep of 5 minutes and output saved every day (Fig. 4b). The
main panel shows trajectories in planar projection, with the inset showing the same trajectories in orthographic projection.
At a speed of 1 m/s, the particles travel 4.9·106 m in the 57 days. At the equator, this amounts to almost 45◦ of longitude, but
because of the cosine-dependence of zonal distance with latitude, particles closer to the poles travel farther in degrees (main
panel in Fig. 4b). The inset of Fig. 4b, nevertheless, shows that in an orthographic projection, all particles travel the same
distance.
11
4.2.3
Advection due to a time-oscillating zonal flow
The third test case tests the ability of Parcels to cope with simple time-varying flow. The flow in this case is a uniform meridional
flow of v = A = 0.1 m/s, and an oscillating zonal flow with u(t) = A cos(ωt) where ω = 2π/T and the period is T = 1 day.
The time resolution of the FieldSet is 5 minutes, and since the flow is constant in space there are only two grid cells in each
of the horizontal directions. A total of 20 particles are then released on a zonal line at y = 0 km and advected for 4 days, using
an RK4 timestep of 5 minutes and storing output every 3 hours (Fig. 4c).
The analytical flow for the paths of these particles is y(t) = At and x(t) = x0 + A/ω sin(ωt) where ω = 2π/T and x0 is
the zonal start location of the particle. Indeed, all particles follow these analytical pathways very closely (Fig. 4c), with largest
positional errors after 4 days being 6 cm in the zonal direction and 4 mm in the meridional direction.
4.2.4
Steady-state flow around a peninsula
The test case of steady-state flow around a peninsula follows a description by Ådlandsvik et al. (2009) and was also used as a
validation test case in the article describing the Connectivity Modeling System (Paris et al., 2013). Starting from the analytical
expression for a streamfunction Ψ of a steady-state flow around a peninsula, analytical expressions of the zonal and meridional
component of velocity are solved on a (1◦ × 0.5◦ ) Arakawa A-grid at 1/100◦ horizontal resolution. A set of 20 particles is
seeded just off the western edge of the domain, and then advected with the flow for 24 hours using an RK4 timestep of 5
minutes and particle positions stored every hour (Fig. 4d, where the brown semi-circle is the peninsula).
Since the particles should follow streamlines, a comparison of the interpolated streamfunction value at t = 24 hours to that
at t = 0 hours gives an estimate of the error. The largest error is 0.008 m2 /s, which corresponds to a positional error of 10−5
degrees, or 1 meter. Indeed, Fig. 4d shows that the particle trajectories closely follow the dashed streamlines.
4.2.5
Steady-state flow in a Stommel gyre and western boundary current
The test case of the Stommel gyre follows a description in Fabbroni (2009), and provides an analytical solution to the streamfunction field of a Stommel gyre and western boundary current. Here, we compute the meridional and zonal central derivatives
of this streamfunction field to generate zonal and meridional velocities, respectively, on a (10, 000 × 10, 000) km Arakawa
A-grid at 50 km horizontal resolution. A set of four particles is seeded on a line crossing the western boundary, at y = 5,000
km, and then advected for 50 days with an RK4 timestep of 5 minutes and the particle positions stored every 24 hours (Fig. 4e).
Since the particles should follow streamlines, the deviation of particles from the streamlines is a measure of the accuracy
of the method. Fig. 4e shows that all three particles stay close to their streamline throughout the 50 day advection period. The
largest error is 0.05 m2 /s, which corresponds to a positional error of less than 5 km.
4.2.6
Damped inertial oscillation on a geostrophic flow
The test case of a damped inertial oscillation on a geostrophic flow follows Fabbroni (2009) and Döös et al. (2013). In this
test case, the velocity varies over the entire domain, following an analytical time-dependent equation. Here, we use a time
12
resolution of 5 minutes for the velocity field. A particle is then seeded at the origin and advected for four days, with a RK4
timestep of 5 minutes and output stored every hour (Fig. 4f). After four days of advection, the positional error of the particle,
as compared to the analytical solution, is less than 5 cm.
4.2.7
Brownian motion with uniform Kh
The test case of Brownian motion with uniform Kh tests for the accuracy and implementation of the random number generator.
Here, a total of 100,000 particles are seeded at the origin of a (60×60) km grid centered around the origin with zero velocities,
and then diffused using a normal variate random number distribution with Kh = 100 m2 /s. The particles are diffused for 1 day
with a timestep of 5 minutes (Fig. 4g). The two-dimensional normalised histogram agrees very well with the analytical solution
√
of this Brownian motion: a two-dimensional Gaussian with a mean at the origin and standard deviation of σ = 2Kh ∆t = 4.16
km.
5
Future outlook
As mentioned before, Parcels v0.9 is a prototype. The core contributions of this paper are both the API, as well as the design
philosophy which enables a wide range of valuable future improvements of the framework. Below, we discuss some of the
conceptual ideas for these planned improvements.
5.1
Performance optimisation
The primary performance optimisation in version 0.9 of Parcels is the automated generation of C kernel code to allow inlining
of field evaluation routines. However, several future optimisations have been planned during the design of the code, based
around considerations for irregular data processing. Since dynamic addition and deletion of particles is a common feature
of many oceanographic use cases, no assumptions about data layout or iteration protocol have been made in the high-level
API of particle sets, allowing more optimised implementations in the future. The use of dynamic code generation at runtime
also enables further automated specialisation of kernel code, while allowing us to define a clear initial interface for kernel
customisation.
In addition to optimising the execution of particle kernels, the extensive interaction with hydrodynamic field data constitutes
a considerable cost of the overall computation - a cost that is likely to dominate overall execution if large sets of hydrodynamic
field data are to be read from files. Multiple potential approaches can be considered in future versions of Lagrangian particle
tracking codes:
– Directly coupled (online) runs within the host OGCM can completely avoid the bandwidth bottlenecks imposed by
reading dense field data from disk, at the expense of additional computation. For simulations at local scales with a high
particle density, this trade-off might prove beneficial, for example for regional studies on marine ecology.
13
– For global-scale models that require offline hydrodynamic field data but feature a low particle density with high localization, the total volume of data read from disk might be drastically reduced by explicitly prefetching local subsets of field
data based on particle locations. Such a mechanism would require the use of additional geospatial indexing methods,
for example via octrees or r-trees (Isaac et al., 2015; Schubert et al., 2013), that decompose the grid into individual
sub-regions and provide fast indexing methods. Using explicit prefetch directives in the dynamic execution loop might
also enable overlapping of asynchronous file reads with effective computation to further amortize file I/O overheads.
The modularity of Parcels’ internal abstractions, as well as the composability of kernels and the flexibility provided by the
dynamic execution loop should facilitate extensive experimentation and exploration with such advanced optimization techniques, without the need for users to change any high-level algorithmic definitions. The use of advanced data handling and
task-scheduling libraries, such as Dask (Rocklin, 2015) or Xarray (Hoyer and Hamman, 2017), might also be utilised to
quickly achieve efficient out-of-core data management in Parcels.
5.1.1
Towards parallellisation
The current version of Parcels is not in itself parallel due to two restrictions:
– The primary input format of field data in the v0.9 prototype is NetCDF-based field data, so that parallellisation requires
an explicit domain decomposition and a parallel file reader. The current version of the netcdf Python package does
not provide these features. Alternative implementations of the NetCDF file format, such as Xarray, might be leveraged
in future versions of Parcels to provide parallel data management.
– Exchanging particle information between parallel processors is currently not supported, although it is deemed a critical
feature for the next release (v1.0).
5.2
Community building and kernel sharing
One of the key ideas between the development of Parcels is for it to be a flexible and extendable codebase, where particle
behaviour can easily be customised. The worked out example in Sect. 3 shows that many types of behaviour (sinking, aging,
etc) can be coded in a few lines of Python code.
The customisability of Parcels enables a multitude of oceanographic modelling, from water parcels to plankton to plastic
litter to fish. We therefore envision an active community of Parcels users who share and discuss kernel development. We
encourage anyone who wishes to share their custom kernels to upload them onto github, and we will provide a properly
referenced library of user-contributed kernels for others to reuse on oceanparcels.org.
5.3
Towards runtime integration with OGCMs
Although the current version of Parcels primarily uses off-line field data, the overall design of the particle exectuion engine
is designed to be compatible with a variety of OGCMs for directly coupled (at-runtime) simulations. In particular, the current
14
Field interface can easily be extended to provide interpolation routines for various types of field data, for example based on
unstrctured meshes, while the primary particle update loop provides a mechanism for host models to dictate a model timestep
size that varies from that of the particle update. Moreover, the explicit generation of C code allows Parcels kernel code to
be easily injected into existing ocean modelling frameworks, while the provision of error-recovery kernels can guarantee
progression of the coupled model.
6
Conclusions
Here, we have introduced a new framework for Lagrangian ocean analysis that focusses on customisability, flexibility and
ease-of-use. This v0.9 of Parcels is very much a prototype, providing a proof-of-concept of the API and showcasing how
it can be used to create high-level Python code for full-fledged scientific experiments. We also assess the accuracy of the
current implementation, with the idea to provide a benchmark for future versions. Future development will focus on increasing
efficiency of the framework, and also towards providing easy tools to port the generated C-code of Parcels experiments to
at-runtime integration within OGCMs.
7
Code availability
The code for Parcels is licensed under the MIT license and is available through github at github.com/OceanParcels/parcels.
The version 0.9 described here is archived at Zenodo at https://doi.org/10.5281/zenodo.823562. More information is available
on the project webpage at oceanparcels.org.
Author contributions. ML and EvS developed the code and wrote the manuscript jointly.
Competing interests. The authors declare no competing interests
Acknowledgements. The initial ideas for the Parcels framework were the result of very fruitful discussions with the attendees of the “Future of
Lagrangian Ocean Modelling” workshop, held at Imperial College London, UK, in September 2015. Funding for this workshop was provided
through an EPSRC Institutional Sponsorship grant to EvS under reference number EP/N50869X/1. EvS is supported through funding from
the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No
715386). The OFES simulation was conducted on the Earth Simulator under the support of JAMSTEC. We thank Joe Scutt-Phillips, Ronan
McAdam, Joel Kronberg, Thomas Stokes, Nathaniel Tarshish, Michael Hart-Davis, Birgit Sutzl, Ben Snowball, Samuel Wetherell and David
Ham for their support in testing and developing aspects of the Parcels code.
15
References
Ådlandsvik, B., Bartsch, J., Brickman, D., Browman, H. I., Edwards, K., Fiksen, Ø., Gallego, A., Hermann, A. J., Hinckley, S., Houde, E.,
Huret, M., Irisson, J.-O., Lacroix, G., Leis, J. M., McCloghrie, P., Megrey, B. A., Miller, T., Van der Molen, J., Mullon, C., North, E. W.,
Parada, C., Paris, C. B., Pepin, P., Petitgas, P., Rose, K., Thygesen, U. H., and Werner, C.: Manual of recommended practices for modelling
physical – biological interactions during fish early life, Tech. rep., 2009.
Alexander, R.: Solving ordinary differential equations I: Nonstiff problems (E. Hairer, SP Norsett, and G. Wanner), SIAM Review, 32, 485,
1990.
Beron-Vera, F. J. and LaCasce, J. H.: Statistics of Simulated and Observed Pair Separations in the Gulf of Mexico, Journal of Physical
Oceanography, 46, 2183–2199, 2016.
Blanke, B. and Raynaud, S.: Kinematics of the Pacific Equatorial Undercurrent: An Eulerian and Lagrangian approach from GCM results,
Journal of Physical Oceanography, 27, 1038–1053, 1997.
Cetina-Heredia, P., Roughan, M., van Sebille, E., Feng, M., and Coleman, M. A.: Strengthened currents override the effect of warming on
lobster larval dispersal & survival, Global Change Biology, 21, 4377–4386, 2015.
Cowen, R. K., Paris, C. B., and Srinivasan, A.: Scaling of connectivity in marine populations, Science, 311, 522–527, 2006.
Doblin, M. A. and van Sebille, E.: Drift in ocean currents impacts intergenerational microbial exposure to temperature, Proceedings of the
National Academy of Sciences, pp. 201521 093–11, 2016.
Döös, K.: Interocean Exchange of Water Masses, Journal of Geophysical Research: Oceans, 100, 13 499–13 514, 1995.
Döös, K., Nycander, J., and Coward, A. C.: Lagrangian decomposition of the Deacon Cell, J. Geophys. Res. Ocean., 113, C07 028,
doi:10.1029/2007JC004351, http://dx.doi.org/10.1029/2007JC004351, 2008.
Döös, K., Kjellsson, J., and Jonsson, B. F.: TRACMASS—A Lagrangian Trajectory Model, in: Preventive Methods for Coastal Protection,
pp. 225–249, Springer International Publishing, Heidelberg, 2013.
Döös, K., Jönsson, B., and Kjellsson, J.: Evaluation of oceanic and atmospheric trajectory schemes in the TRACMASS trajectory model v6.0,
Geoscientific Model Development, 10, 1733–1749, doi:10.5194/gmd-10-1733-2017, http://www.geosci-model-dev.net/10/1733/2017/,
2017.
Fabbroni, N.: Numerical simulations of passive tracers dispersion in the sea, Ph.D. thesis, Universita di Bologna, 2009.
Griffa, A.: Applications of stochastic particle models to oceanographic problems, in: Stochastic modelling in physical oceanography, pp.
113–140, Springer, 1996.
Haller, G.: Lagrangian Coherent Structures, Annual Review of Fluid Mechanics, 47, 137–162, doi:10.1146/annurev-fluid-010313-141322,
2015.
Hellweger, F. L., van Sebille, E., and Fredrick, N. D.: Biogeographic patterns in ocean microbes emerge in a neutral agent-based model,
Science, 345, 1346–1349, 2014.
Hoyer, S. and Hamman, J.: xarray: N-D labeled Arrays and Datasets in Python, Journal of Open Research Software, 5(1), 10,
doi:10.5334/jors.148, http://doi.org/10.5334/jors.148, 2017.
Isaac, T., Burstedde, C., Wilcox, L. C., and Ghattas, O.: Recursive Algorithms for Distributed Forests of Octrees, SIAM Journal on Scientific
Computing, 37, C497–C531, doi:10.1137/140970963, https://doi.org/10.1137/140970963, 2015.
16
Jones, A., Thomson, D., Hort, M., and Devenish, B.: The U.K. Met Office’s Next-Generation Atmospheric Dispersion Model, NAME III,
in: Air Pollution Modeling and Its Application XVII, edited by Borrego, C. and Norman, A.-L., pp. 580–589, Springer US, Boston, MA,
2007.
Jönsson, B. F., Salisbury, J. E., and Mahadevan, A.: Large variability in continental shelf production of phytoplankton carbon revealed by
satellite, Biogeosciences, 8, 1213–1223, doi:10.5194/bg-8-1213-2011, http://www.biogeosciences.net/8/1213/2011/, 2011.
Jutzeler, M., Marsh, R., Carey, R. J., White, J. D. L., Talling, P. J., and Karlstrom, L.: On the fate of pumice rafts formed during the 2012
Havre submarine eruption, Nature Communications, 5, 3660, 2014.
Katz, M. E., Cramer, B. S., Franzese, A. M., Hoenisch, B., Miller, K. G., Rosenthal, Y., and Wright, J. D.: Traditional and emerging
geochemical proxies in foraminifera, Journal of Foraminiferal Research, 40, 165–192, 2010.
Kool, J. T., Moilanen, A., and Treml, E. A.: Population connectivity: recent advances and new perspectives, Landscape Ecology, 28, 165–185,
doi:10.1007/s10980-012-9819-z, http://dx.doi.org/10.1007/s10980-012-9819-z, 2013.
Lebreton, L. C. M., Greer, S. D., and Borerro, J. C.: Numerical modelling of floating debris in the world’s oceans, Marine Pollution Bulletin,
64, 653–661, 2012.
Marsh, R., Ivchenko, V. O., Skliris, N., Alderson, S., Bigg, G. R., Madec, G., Blaker, A. T., Aksenov, Y., Sinha, B., Coward, A. C., Le Sommer,
J., Merino, N., and Zalesny, V. B.: NEMO–ICB (v1.0): interactive icebergs in the NEMO ocean model globally configured at eddypermitting resolution, Geoscientific Model Development, 8, 1547–1562, 2015.
Masumoto, Y., Sasaki, H., Kagimoto, T., Komori, N., Ishida, A., Sasai, Y., Miyama, T., Motoi, T., Mitsudera, H., Takahashi, K., Sakuma,
H., and Yamagata, T.: A fifty-year eddy-resolving simulation of the world ocean - Preliminary outcomes of OFES (OGCM for the Earth
Simulator), Journal of the Earth Simulator, 1, 2004.
Paris, C. B., Cowen, R. K., Claro, R., and Lindeman, K. C.: Larval transport pathways from Cuban snapper (Lutjanidae) spawning aggregations based on biophysical modeling, Marine Ecology-Progress Series, 296, 93–106, 2005.
Paris, C. B., Helgers, J., van Sebille, E., and Srinivasan, A.: Connectivity Modeling System: A probabilistic modeling tool for the multi-scale
tracking of biotic and abiotic variability in the ocean, Environmental Modelling & Software, 42, 47–54, 2013.
Peeters, F. J. C., Acheson, R., Brummer, G.-J. A., de Ruijter, W. P. M., Schneider, R. R., Ganssen, G. M., Ufkes, E., and Kroon, D.: Vigorous
exchange between the Indian and Atlantic oceans at the end of the past five glacial periods, Nature, 430, 661–665, 2004.
Pingali, K., Nguyen, D., Kulkarni, M., Burtscher, M., Hassaan, M. A., Kaleem, R., Lee, T.-H., Lenharth, A., Manevich, R., Méndez-Lojo,
M., Prountzos, D., and Sui, X.: The Tao of Parallelism in Algorithms, SIGPLAN Not., 46, 12–25, doi:10.1145/1993316.1993501, http:
//doi.acm.org/10.1145/1993316.1993501, 2011.
Qin, X., Menviel, L., Sen Gupta, A., and van Sebille, E.: Iron sources and pathways into the Pacific Equatorial Undercurrent, Geophysical
Research Letters, pp. n/a–n/a, doi:10.1002/2016GL070501, 2016GL070501, 2016.
Rocklin, M.: Dask: Parallel Computation with Blocked algorithms and Task Scheduling, in: Proceedings of the 14th Python in Science
Conference, edited by Huff, K. and Bergstra, J., pp. 130 – 136, 2015.
Rühs, S., Durgadoo, J. V., Behrens, E., and Biastoch, A.: Advective timescales and pathways of Agulhas leakage, Geophysical Research
Letters, 40, 3997–4000, doi:10.1002/grl.50782, http://doi.wiley.com/10.1002/grl.50782, 2013.
Schubert, E., Zimek, A., and Kriegel, H.-P.: Geodetic Distance Queries on R-Trees for Indexing Geographic Data, pp. 146–164, Springer
Berlin Heidelberg, Berlin, Heidelberg, doi:10.1007/978-3-642-40235-7_9, https://doi.org/10.1007/978-3-642-40235-7_9, 2013.
Stohl, A. and James, P.: A Lagrangian analysis of the atmospheric branch of the global water cycle. Part II: Moisture transports between
earth’s ocean basins and river catchments, Journal of Hydrometeorology, 6, 961–984, 2005.
17
Teske, P. R., Sandoval-Castillo, J., van Sebille, E., Waters, J., and Beheregaray, L. B.: On-shelf larval retention limits population connectivity
in a coastal broadcast spawner, Marine Ecology-Progress Series, 532, 1–12, 2015.
Van Sebille, E., Scussolini, P., Durgadoo, J. V., Peeters, F. J. C., Biastoch, A., Weijer, W., Turney, C. S. M., Paris, C. B., and Zahn, R.: Ocean
currents generate large footprints in marine palaeoclimate proxies, Nature Communications, 6, 6521, 2015.
Van Sebille, E., Griffies, S. M., Abernathey, R., Adams, T. P., Berloff, P., Biastoch, A., Blanke, B., Chassignet, E. P., Cheng, Y., Cotter, C. J.,
Deleersnijder, E., Doöös, K., Drake, H., Drijfhout, S., Gary, S. F., Heemink, A. W., Kjellsson, J., Koszalka, I. M., Lange, M., Lique, C.,
MacGilchrist, G. A., Marsh, R., Mayorga Adame, C. G., McAdam, R., Nencioli, F., Paris, C. B., Piggott, M. D., Polton, J. A., Rühs, S.,
Shah, S. H. A. M., Thomas, M. D., Wang, J., Wolfram, P. J., Zanna, L., and Zika, J. D.: Lagrangian ocean analysis: fundamentals and
practices, Ocean Modeling, submitted.
18
User
Particle
Kernel
Define variables
Define computation
Visualization
Particle Loop
Execution
Data
External OGCM
Update particles
Particle Data
Field Data
File I/O
Interpolation
External data
Figure 1. Conceptual abstractions (dark) and functionalities encapsulated in the Parcels prototype in relation to external components (light).
19
.from_list()
ParticleSet.from_list
Field
is one of the methods
used to define the starting
positions of Particles.
ParticleSet objects are
Field objects hold the
data of the hydrodynamic
fields, stored as 4dimensional (lon, lat,
depth, time) numpy
arrays.
.from_netcdf()
FieldSet.from_netcdf
is the method used to
read hydrodynamic fields
in NetCDF data using info
of Fields in dictionaries.
ParticleSet
Particle
Particle objects contain
the position and other
variables of each particle
in the ParticleSet.
FieldSet
the main objects in
Parcels. They contain a
FieldSet and a list of
Particles.
The .from_list,
.execute and .show are
the most important
methods defined on
ParticleSets.
FieldSet objects are
collections of Fields. At
least a U and V Field are
.execute()
ParticleSet.execute
is the method used to
actually compute the
evolution of particles by
executing Kernel objects.
.show()
ParticleSet.show
is the method used to plot
particle positions,
optionally overlayed on a
Field.
required for Parcels to
work.
Kernel
Kernels are little
snippets of code that get
run when a ParticleSet
is executed. Parcels comes
with some build-in kernels
like 4th order Runge-Kutta
advection, but it is very
easy to create custom
kernels. Multiple kernels
can be concatenated with
the + operator.
Figure 2. Class diagram of the Parcels v0.9 prototype implementation. Classes are depicted in blue, methods in green. Note that not all
methods and classes are shown in this diagram.
20
[ C]
20
30°S
18
16
14
40°S
10°E
20°E
30°E
12
Figure 3. Footprints of virtual foraminifera ending up on the ocean floor just off Cape Town in the Agulhas region. This experiment is a
Parcels implementation of the study described in Van Sebille et al. (2015), and this figure can be compared to Fig. 1b in that paper. The
magenta dot is the location of the sediment core, from which virtual particles are first tracked back until they reach their 50m dwelling depth
(black dots), and then further tracked back for their 30-day lifespan. Temperatures (in degrees Celcius) are recorded each day throughout
their lifespan and shown as colours. The code for this experiment and plotting is available at https://doi.org/10.5281/zenodo.823994.
21
(a) Radial rotation with known period
(b) Longitudinal shear flow
0
2
Meridional distance [km]
Latitude [degrees]
Meridional distance [km]
30
40
2
20
0
25
20
15
10
5
20
4
0
4
2
0
Zonal distance [km]
2
4
0
(d) Steady-state flow around a peninsula
20
40
60
Longitude [degrees]
80
10
(e) Steady-state flow in a Stommel gyre and western boundary current
0.25
0.20
0.15
0.10
5
10
1
Meridional distance [km]
Meridional distance [km]
0.30
0
Zonal distance [km]
0
8000
0.35
5
(f) Damped inertial oscillation on a geostrophic flow
10000
0.40
Latitude [degrees]
(c) Advection due to a time-oscillating zonal flow
35
60
4
6000
4000
2
3
4
2000
0.05
0.1
0.2
0.3
0.4 0.5 0.6
Longitude [degrees]
0.7
0
0.8
0
2000
4000
6000
Zonal distance [km]
8000
(g) Brownian motion with a uniform Kh
Meridional distance [km]
10
10000
5
0
2
4
6
8
Zonal distance [km]
10
12
0.010
0.008
5
0.006
0
0.004
5
0.002
10
15
15
10
5
0
5
Zonal distance [km]
10
0.000
Figure 4. Evaluation of trajectory accuracy in Parcels v0.9, following the seven idealised and analytical test case described in in Sect. 6 of
Van Sebille et al. (submitted): (a) radial rotation with known period; (b) longitudinal shear flow; (c) advection due to a time-oscillating zonal
flow; (d) steady-state flow around a peninsula; (e) steady-state flow in a Stommel gyre and western boundary current; (f) damped inertial
oscillation on a geostrophic flow; and (g) Brownian motion with a uniform Kh . In the upper six panels, the coloured lines are the particle
trajectories and the black dashed lines are the analytical solutions. In panel (g), the colouring shows the density of particles, and the contours
show the probability density function of the equivalent analytical solution (a two-dimensional Gaussian).
22
| 5cs.CE
|
arXiv:1711.06400v2 [math.RT] 23 Jan 2018
GENERICALLY FREE REPRESENTATIONS II: IRREDUCIBLE
REPRESENTATIONS
SKIP GARIBALDI AND ROBERT M. GURALNICK
Abstract. We determine which faithful irreducible representations V of a
simple linear algebraic group G are generically free for Lie(G), i.e., which V
have an open subset consisting of vectors whose stabilizer in Lie(G) is zero.
This relies on bounds on dim V obtained in prior work (part I), which reduce
the problem to a finite number of possibilities for G and highest weights for
V , but still infinitely many characteristics. The remaining cases are handled
individually, some by computer calculation. These results were previously
known for fields of characteristic zero; we provide a shorter proof that gives the
result with very mild hypotheses on the characteristic. (The few characteristics
not treated here are settled in part III.) Combining these results with those
of Guralnick–Lawther–Liebeck shows that for any irreducible module for a
simple algebraic group, there is a generic stabilizer (as a group scheme) and
gives a classification of the generic stabilizers in all cases. These results are
also related to questions about invariants.
Let G be a simple linear algebraic group over a field k acting faithfully on a
vector space V . In the special case k = C, there is a striking dichotomy between the
properties of irreducible representations V whose dimension is small (say, ≤ dim G)
versus those whose dimension is large, see for example [AVE68], [Èla72], [Pop88],
and [PV94, §8.7]. For example, if dim V < dim G, then (trivially) the stabilizer
Gv of a vector v ∈ V is nontrivial. Conversely (and nontrivially) for V hardly
bigger than dim G, the stabilizer Gv for generic v ∈ V is trivial, i.e., 1; in this
case one says that V is generically free or G acts generically freely on V . This
property has taken on increased importance recently due to applications in Galois
cohomology and essential dimension, see [Rei10] and [Mer13] for the theory and
[BRV10], [GG17b], [Kar10], [LMMR13], [Löt13], etc. for specific applications.
With applications in mind, it is desirable to extend the results on generically
free representations to all fields. In that setting, [GLL18] has shown that, if V is
irreducible and dim V is large enough, then G(k), the group of k-points of G, acts
generically freely. Equivalently (when k is algebraically closed), the stabilizer Gv of
a generic v ∈ V is an infinitesimal group scheme. For applications, one would like
to say that Gv is not just infinitesimal but is the trivial group scheme, for which
one needs to know that the Lie algebra g of G acts generically freely on V , i.e.,
gv = 0. The two conditions are related in that dim Gv ≤ dim gv , so in particular if
gv = 0, then Gv is finite. Conversely, if Gv (k) = 1, gv can be nontrivial (i.e., Gv
may be a nontrivial infinitesimal group scheme), see Example 3.2.
In a previous paper, [GG17a], we proved that, roughly speaking, if dim V is large
enough (where the bound is Θ((rank G)2 )), then V is a generically free g-module.
2010 Mathematics Subject Classification. 20G05 (primary); 17B10 (secondary).
1
2
S. GARIBALDI AND R.M. GURALNICK
In this paper, we restrict our focus to irreducible modules and settle the question
of whether or not g acts generically freely.
Theorem A. Let G be a simple linear algebraic group over a field k such that char k
is not special for G and let ρ : G → GL(V ) be a restricted irreducible representation
of G. If dim V ≤ dim G or (G, ρ, k) is one of the five representations in Table 1,
then g does not act virtually freely on V . Otherwise, g acts virtually freely on V .
G
A7
A8
D8
char k
weight dim V dim gv G char k weight dim V dim gv
2
∧4
70
4
C4
3
0100
40
2
3
3
∧ or ∧6
84
3
B2
5
11
12
1
2
half-spin
128
4
Table 1. Irreducible and restricted representations V of simple G
with dim V > dim G that are not virtually free for g. For each, the
stabilizer gv of a generic v ∈ V is a toral subalgebra, and dim gv is
given for the case where G is simply connected.
We say that g acts virtually freely on V if the stabilizer gv of a generic vector
v ∈ V satisfies gv = ker dρ, i.e., gv is as a small as possible. This notion is the
natural generalization of “generically freely” to allow for the possibility that G does
not act faithfully. (For V as in Theorem A and not the trivial representation, ker dρ
is a proper ideal in g, so is contained in z(g) by the hypothesis that char k is not
special for G. Thus ker dρ = ker λ|z(g) for λ the highest weight of V , which depends
only on the equivalence class of λ modulo the root lattice.)
We remark that the exceptions in Theorem A, listed in Table 1, can be divided
into two types. In the left column are three “θ-group” representations, which arise
from embedding g in some larger Lie algebra with a finite grading, and the generic
stabilizer Gv is a non-étale, non-infinitesimal finite group scheme. (Premet’s appendix in [GG17b] gives a detailed study of the half-spin representation of D8 in
characteristic 2. For the other two representations in the left column, see Remark
6.3, [Aul01, 4.8.2, 4.9.2], or [GLL18, 4.1].) In the right column are two representations where the generic stabilizer is a nonzero infinitesimal group scheme, see
Examples 3.2 and 4.2.
Recall that every irreducible representation V of G has a highest weight λ, and
we consider separately three cases. Specifically, we may write λ = λ0 + pλP
1 where
p := char k 6= 0 and λ0 is restricted, i.e., when λ0 is expressed as a sum ω cω ω
where the sum runs over the fundamental dominant weights ω, we have 0 ≤ cω < p
for all ω. Theorem A treats the case where λ is itself restricted, i.e., when λ1 = 0.
When λ0 = 0, V is the Frobenius twist of the irreducible module of G with highest
weight λ1 , so g acts trivially on V . The next theorem treats the remaining case,
where λ0 and λ1 are both nonzero; in that case L(λ) ∼
= L(λ0 ) ⊗ L(λ1 )[p] , where
both L(λ0 ), L(λ1 ) 6∼
k,
i.e.,
L(λ)
is
irreducible
and
tensor
decomposable. (The G=
representation V is tensor decomposable if there exist G-representations V ′ , V ′′ 6∼
=k
′
′′
such that V ∼
V
⊗
V
.)
=
Theorem B. Let G be a simple linear algebraic group over a field k such that
char k is not special for G. If ρ : G → GL(V ) is an irreducible representation that
is tensor decomposable, then for generic v ∈ V , Lie(G)v = ker dρ.
GENERICALLY FREE IRREDUCIBLE REPRESENTATIONS
3
The main results here and in [GG17a] do not address the case when char k is
special for G. Results for that case will appear in [GG18].
Notation. For convenience of exposition, we will assume in most of the rest of
the paper that k is algebraically closed of characteristic p 6= 0. This is only for
convenience, as our results for p prime immediately imply the corresponding results
for characteristic zero: simply lift the representation from characteristic 0 to Z and
reduce modulo a sufficiently large prime.
We say that char k is special for G if char k = p 6= 0 and the Dynkin diagram of
G has a p-valent bond, i.e., if char k = 2 and G has type Bn or Cn for n ≥ 2 or type
F4 , or if char k = 3 and G has type G2 . (Equivalently, these are the cases where G
has a very special isogeny.) This definition is as in [Ste63, §10]; in an alternative
history, these primes might have been called “extremely bad” because they are a
subset of the very bad primes — the lone difference is that for G of type G2 , the
prime 2 is very bad but not special.
Let G be an affine group scheme of finite type over k, which we assume is
algebraically closed. If G is additionally smooth, then we say that G is an algebraic
group. An algebraic group G is simple if it is not abelian, is connected, and has no
connected normal subgroups 6= 1, G; for example SLn is simple for every n ≥ 2.
If G acts on a variety X, the stabilizer Gx of an element x ∈ X(k) is a subgroup-scheme of G with points
Gx (R) = {g ∈ G(R) | gx = x}
for every k-algebra R. Statements “for generic x” means that there is a dense open
subset U of X such that the property holds for all x ∈ U .
If Lie(G) = 0 then G is finite and étale. If additionally G(k) = 1, then G is the
trivial group scheme Spec k.
We write g for Lie(G) and similarly spinn for Lie(Spinn ), etc. We put z(g) for
the center of g; it is the Lie algebra of the (scheme-theoretic) center of G. As
char k = p, the Frobenius automorphism of k induces a “p-mapping” x 7→ x[p] on
n
g. An element x ∈ g is nilpotent if x[p] = 0 for some n > 0, toral if x[p] = x, and
semisimple if x is contained in the Lie p-subalgebra of g generated by x[p] .
In this paper, highest weights are numbered as in Table 2, imitating [Lüb01]
to make it convenient to refer
P to that paper. We write c1 c2 c3 · · · cℓ as shorthand
for the dominant weight
ci ωi , where ωi is the fundamental dominant weight
corresponding to the vertex i in Table 2.
(Aℓ ) r
r
1
2
r· · · r
3
ℓ−2
r
ℓ−1
r
ℓ
(Cℓ ) r
1
2
(Bℓ ) r
1
>
r
2
r· · · r
3
ℓ−2
r
ℓ−1
r
ℓ
(Dℓ )
1
<
r
2
r
❩❩r
r✚✚3
r· · · r
3
ℓ−2
r· · · r
4
ℓ−2
r
ℓ−1
r
ℓ−1
r
ℓ
r
ℓ
Table 2. Dynkin diagrams of simple root systems of classical
type, with simple roots numbered as in [Lüb01].
Acknowledgements. Guralnick was partially supported by NSF grants DMS1265297, DMS-1302886, and DMS-1600056.
4
S. GARIBALDI AND R.M. GURALNICK
1. Results from part I
For a representation V of simple G with V g = 0 (such as in Theorem A in this
paper), we showed in part I ([GG17a, Th. A]): if dim V > b(G) for b(G) as in Table
3, then g acts virtually freely on V . Furthermore, this was used to verify Theorem
A except in the following cases:
(1) G has type Aℓ for 2 ≤ ℓ ≤ 19;
(2) G has type Bℓ or Cℓ with 2 ≤ ℓ ≤ 11;
(3) G has type Dℓ for 4 ≤ ℓ ≤ 11.
type of G
Aℓ
Aℓ
Bℓ
Cℓ
Dℓ
Dℓ
char k
b(G)
type of G char k
6= 2
2.25(ℓ + 1)2
G2
6= 3
=2
2(ℓ + 1)2 − 2
F4
6= 2
6= 2
8ℓ2
E6
any
6= 2
6ℓ2
E7
any
2(2ℓ − 1)2
E8
any
6= 2
=2
4ℓ2
Table 3. Bound b(G) from part I
b(G)
48
240
360
630
1200
Put g := Lie(G). For x ∈ g, put
V x := {v ∈ V | dρ(x)v = 0}.
We are going to verify the inequality
(1.1)
dim(Ad(G)x) + dim V x < dim V
for various x ∈ g. The following lemma is Lemma 2.6 in [GG17b]; it resembles
[AP71, Lemma 4] and [Gue97, §3.3].
Lemma 1.2. Suppose k is an algebraically closed field of characteristic p > 0 and
h is a G-invariant subspace of g.
n
(1) If, for every x ∈ g \ h such that x[p] = x or x[p] = 0 for some n, (1.1)
holds, then for generic v ∈ V , gv ⊆ h.
(2) If h consists of semisimple elements and equation (1.1) holds for every
x ∈ g \ h with x[p] ∈ {0, x}, then for generic v in V , gv ⊆ h.
Taking h = z(g) in Lemma 1.2, we see that verifying (1.1) for noncentral nilpotent
and toral x ∈ g implies that gv ⊆ z(g) for generic v ∈ V .
Theorem 12.2 in [GG17a] proved a somewhat stronger result than the one stated
at the start of this section: If V is a representation of a simple group G such
that V g = 0 and dim V > b(G), then (1.1) holds for all noncentral x ∈ g with
x[p] ∈ {0, x}.
2. Constructing representations in Magma
In order to prove Theorem A, the results of [GG17a] reduce us to considering
a finite list of irreducible representations, each of which we will consider. Some of
these will be dealt with by invoking calculations done on a desktop computer using
Magma [BCP90], which we now explain. (Code and output are attached to the
arxiv version of this article.)
GENERICALLY FREE IRREDUCIBLE REPRESENTATIONS
5
The built-in Magma instructions IrreducibleRootDatum and LieAlgebra produce the Lie algebra g of a simple algebraic group over Fq for q a power of a prime
p; the main cases of concern in this paper are for p ≤ 7. For a given highest weight
λ, HighestWeightModule gives a homomorphism ρ from g to matrices, and one can
identify the standard module with highest weight λ with the space of row vectors
v where the action by g is v 7→ vρ(x) for x ∈ g. The vector (1, 0, . . . , 0) is a highest weight vector and it generates a submodule V that is irreducible with highest
weight λ.
For any row vector v, it is then a matter of linear algebra to compute the stabilizer gv , i.e., the subspace of x ∈ g such that vρ(x) = 0. It is determined by
Kernel(VerticalJoin([vρ(y):y in Basis(g)])).
To verify that a particular V is virtually free, we use Random(V) to generate
random vectors v ∈ V . For each, we compute dim gv . By upper semicontinuity of
dimension, dim gv is at least as big as dim gw for w generic in V . Therefore, if we
find any v ∈ V with dim gv = dim ker dρ, we have verified that the representation
is virtually free.
e → G is a central isogeny; note that the differential dq : g̃ → g
Remark. Suppose q : G
need not be surjective, i.e., q need not be étale. Nonetheless, if g acts virtually freely
on V , then so does g̃. Therefore, in the computer calculations described above we
e such that ρ is defined.
construct G to be the largest central quotient of G
(If we instead assume that g̃ acts virtually freely on V , it need not follow that g
e = Sp8 , G = PSp8 , char k = 2, and V is the
does. For example, this occurs when G
16-dimensional irreducible “spin” representation, see [GG18, Example 6.8].)
3. Examples where g does not act virtually freely
Lemma 3.1. Let V be a representation of a simple algebraic group G. If there is
a v ∈ V such that
(1) L := gv is a toral subalgebra;
(2) dim zg (L) = rank G; and
(3) dim G − rank G = dim V − dim V L ,
then there is an open subset U of V containing v such that gu is a G-conjugate of
L for every u ∈ U .
Proof. The torus L is contained in Lie(T ) for some maximal torus T in G [Hum67,
Th. 13.3]. Since zg (L) contains Lie(T ), the two are equal.
Put Û for the set of v ′ ∈ V L such that dim gv′ is minimal; it is open in V L . On
the one hand, L ⊆ gv′ , and on the other hand, v ∈ V L , so dim gv′ ≤ dim L, whence
gv′ = L and v ∈ Û .
If g ∈ G(k) normalizes L, then it also normalizes zg (L) = Lie(T ), hence T , so
dim NG (L) = rank G. In particular, for v ′ ∈ Û , the space of g ∈ G(k) with gv ′ = v
has dimension at most rank G and the map G × V L → V defined by (g, w) 7→ gw,
then, has image of dimension at least dim G − rank G + dim V L = dim V . That is,
the map is dominant, so an open subset of V consists of elements whose stabilizer
in g is conjugate to L.
Example 3.2 (C4 , 0100, p = 3). Consider now G = Sp8 over a field k of characteristic 3. (See Prop. 7.4(2) for the case char k 6= 2, 3.) It has a unique irreducible
representation V with dim V = 40 [Lüb01], which occurs as a quotient of the Weyl
6
S. GARIBALDI AND R.M. GURALNICK
module of dimension 48 contained in ∧3 (k 8 ) (with k 8 as the other composition factor), compare [PS83] or Proposition 7.4. Using Magma, one can construct V (say,
with k = F3 ) as in the preceding section and verify that for a random v ∈ V , in
the notation of Lemma 3.1, dim L = 2 and dim V L = 8. It follows that g does
not act virtually freely on V . On the other hand, Gv (k) = 1 for generic v ∈ V
by [GLL18], so this is an example of a representation where the scheme-theoretic
generic stabilizer Gv is a nontrivial and infinitesimal group scheme.
Lemma 3.1 shows also that the other representation in the right column of Table
1 is not virtually free, see Example 4.2.
4. Representations defined over a localization of the integers
Recall that G is defined over an algebraically closed field k of characteristic p,
and in particular is split. Let now R be a subring of Q with homomophisms to Fp
and to a field K containing a primitive p-th root of unity ζ (e.g., take R = Z and
K = C). There exists a smooth affine group scheme GR over R such that GR × k
is isomorphic to G.
Lemma 4.1. Let ρ : GR → GL(V ) be a homomorphism of group schemes over R
for some free R-module V . Then the following are equivalent:
(1) dim xG + dim(Vk )x < dim V for all noncentral x ∈ g such that x[p] = x.
(2) dim g GK + dim(VK )g < dim V for all noncentral g ∈ GR (K) such that
g p = 1.
Here and below we use the shorthand XF for XR × F , where XR is an R-scheme
and there is an implicit homomorphism R → F .
Proof. This is essentially §3.4 in [Aul01], which we reproduce here for the convenience of the reader. Pick a split maximal torus TR in GR and a basis τ1 , . . . , τℓ
of the lattice of cocharacters Gm → TR . Identifying the Lie algebra of Gm with k,
the elements hj := dτj (1) make up a basis of the Lie algebra t of TR × Fp such that
[p]
hj = hj . This gives a bijection of toral elements in t with elements of order p in
TK via
X
Y
ψ:
cj hj 7→
τj (ζ cj ) for cj ∈ Fp .
The character lattices of TK and t are naturally identified; indeed there is a basis
χ1 , . . . , χℓ such that χi (hj ) = δij and
P χi ◦ τj : Gm → Gm is the
P identity for
Q i=j
and trivial for i 6= j. Writing χ = P di χi , we P
find that χ(ψ( cj hj )) = i ζ ci di ,
which is 1 in K if and only if χ( cj hj ) =
ci di is 0 in k. Decomposing V
as a sum of weight spaces relative to TR , we find that for toral x ∈ t, we have
dim(Vk )x = dim(VK )ψ(x) .
The centralizer in g of x and the centralizer in GK of ψ(x) contain Lie(Tk ) and
TK , so their identity components are generated by that and the root subalgebras
or subgroups corresponding to roots vanishing on x or ψ(x) respectively. As in the
preceding paragraph, we find that the centralizers of x and ψ(x) have the same
dimension, hence (a) x is central in g if and only if ψ(x) is central in GK and (b)
dim xG = dim ψ(x)GK . The equivalence of (1) and (2) follows.
Example 4.2 (B2 , 11). Let G = Spin5 ∼
= Sp4 and take V to be the irreducible
representation of dimension 12 (if char k = 5) or 16 (if char k 6= 5). It occurs as
GENERICALLY FREE IRREDUCIBLE REPRESENTATIONS
7
a composition factor of the tensor product X of the two fundamental irreducible
representations.
In case char k = 5, we apply Lemma 3.1. One finds dim L = 1 and dim V L = 4,
so V is not virtually free. We remark that in this case again Gv (k) = 1, so Gv is a
nonzero infinitesimal group scheme.
In case char k = 2, we verify that V is generically free for g using Magma as in
§2.
So assume char k 6= 2, 5. As X is self-dual, it is a direct sum of V and X/V ,
the natural representation of Sp4 . In this case we argue that V is virtually free by
verifying (1.1). We get it for toral elements via Lemma 4.1, where we refer to the
literature (in this case [GLL18]) for the corresponding result for Sp4 (C).
A long root element x has a single Jordan block of size 2 on the natural module
and 2 Jordan blocks of size 2 on the 5-dimensional module. Since char k 6= 2, x
has partition (32 , 25 , 14 ) on X, so dim X x = 11. Since V /X is the 4-dimensional
symplectic module, dim(V /X)x = 3, so dim V x = 8. As dim xG = 4, (1.1) is
verified.
For any other nilpotent, the closure of xG in sp4 contains a nilpotent with partition (2, 2), so dim V x ≤ 6; as dim xG ≤ 8, the inequality is verified.
Example 4.3 (B3 , 101). Let G = Spin7 and take V to be the irreducible representation of dimension 40 (if char k = 7) or 48 (if char k 6= 7). It occurs as a
composition factor of the tensor product X of the natural and spin representations.
In case char k = 2 or 7, we construct the representation in Magma as in §2 and
observe that it is generically free.
So suppose char k 6= 2, 7. Then X is self-dual so it is a direct sum of V and X/V ,
the spin representation. As in the preceding example, we argue that V is virtually
free by verifying (1.1) and we similarly are reduced to considering nonzero nilpotent
x. For x with partition (32 , 1) on the natural representation, dim V x ≤ 22 and
dim xG = 14. If x has partition (7) or (5, 12 ), then dim xG ≤ 18 and dim V x ≤ 22
(by specialization). A long root element x (partition (22 , 13 )), has dim xG = 8 and
dim V x = 34. The remaining possibilities for x have partition (3, 22 ) or (3, 14 ),
which have dim xG = 12 or 10 and by specialization dim V x ≤ 34.
Example 4.4 (D4 , 1001). Consider the representation V of G = Spin8 with highest
weight 1001. In case char k = 2, dim V = 48 and g acts with 1-dimensional kernel;
we verify with Magma that V is virtually free for g.
So suppose char k 6= 2, in which case dim V = 56. Writing Vi with i = 1, 2, 3
for the three inequivalent irreducible 8-dimensional representations, we find X :=
V1 ⊗ V2 ∼
= V ⊕ V3 . Lemma 4.1 verifies the inequality (1.1) for x toral.
Suppose that x is nonzero nilpotent with dim xG < 23. Certainly dim V x ≤
dim V y for a root element y. Such a y has two Jordan blocks of size 2 on the
Vi ’s, and so y acts on X with partition (34 , 217 , 118 ). Thus dim X y = 39 and
dim V y = dim X y − dim V3y = 33, and the inequality is verified for x.
We now divide into cases based on the partition of x on one of the Vi ’s. If x
only has Jordan blocks of size at most 3, then dim xG < 21 and we are done by the
previous paragraph.
If x has two Jordan blocks of size 4, then dim V x < 16. If x has a Jordan block
of size ≥ 5, then dim V x < 20. In either case, as dim xG ≤ 24, the inequality is
verified.
8
S. GARIBALDI AND R.M. GURALNICK
Example 4.5 (D5 , 20000). Consider the representation V of G = Spin10 with
highest weight 20000 of dimension 126. If char k = 2, then 126 > b(G) and the
inequality (1.1) holds for noncentral x ∈ g with x[2] ∈ {0, x} by [GG17a]. So assume
char k 6= 2. For one of the half-spin representations X, the second symmetric power
S 2 X is a direct sum of V and the natural 10-dimensional module.
A root element x ∈ g has a 12-dimensional fixed space on X and so has 4
nontrivial Jordan blocks. On S 2 X, it has a fixed space of dimension 84 hence
dim V x = 72. Therefore, for every nonzero nilpotent x ∈ g, we have dim V x ≤
72 and of course dim xG ≤ dim G − rank G = 40, verifying the inequality. For
noncentral toral x, the inequality is verified as in the preceding examples.
Example 4.6 (C5 , 10000). Let now V be the irreducible representation of G = Sp10
with highest weight 10000. In case char k = 3, dim V = 122 and one checks using
Magma that a generic vector has trivial stabilizer. So assume char k > 3, in which
case dim V = 132.
The inequality holds for noncentral toral elements by Lemma 4.1, so we verify
the inequality for nilpotent elements of sp10 . Restricting to the Levi subgroup
Sp8 , the representation decomposes as a direct sum of irreducibles X ⊕ Y ⊕ Y
where dim X = 48 and dim Y = 42. Since char k > 3, X is a submodule of ∧3 k 8
with quotient k 8 . The restriction of Y to the Levi Sp6 in Sp8 is a direct sum of
irreducibles Y ′ ⊕ Y ′ ⊕ Y ′′ where dim Y ′ = dim Y ′′ = 14, Y ′ is a submodule of
∧3 k 6 with quotient k 6 and Y ′′ is a submodule of ∧2 k 6 with quotient k. Using
these decompositions, we find that a long root x ∈ sp6 ⊂ sp10 has dim V x = 90
and nilpotent y ∈ sp6 ⊂ sp10 with partition (4, 16 ) has dim V y = 19. In view
of the fomer, it suffices to consider nilpotent z ∈ g such that dim CSp10 (z) ≤ 13
Such a z has a Jordan block of size at least 4 and so specializes to y. Then
dim z Sp10 + dim V z ≤ 50 + 19, verifiying the inequality.
5. Example: symmetric squares and wedge squares
Recall that k is assumed algebraically closed of characteristic p ≥ 0. Put gln (k)
for the Lie algebra of n-by-n matrices with entries in k. We first note that, for
x ∈ gln (k), CGLn (x) is the group of units in the associative k-algebra gln (k)x .
Therefore, dim xGLn = dim [gln (k), x] and we have:
Lemma 5.1. Let x ∈ gln (k). Then dim xGLn + dim gln (k)x = n2 .
Lemma 5.2. Let x ∈ gln = gl(V ) with x a regular nilpotent element.
(1) The number of Jordan blocks of x on gl(V ) and V ⊗ V is n.
If furthermore char k 6= 2, then
(2) the number of Jordan blocks of x on Sym2 (V ) is n/2 for n even and (n+1)/2
for n odd; and
(3) the number of Jordan blocks of x on ∧2 (V ) is n/2 if n is even and (n − 1)/2
if n is odd.
Proof. As x is nilpotent, V and V ∗ are equivalent k[x]-modules, hence the number
of Jordan blocks on V ⊗ V and gl(V ) is the same and is also independent of the
characteristic by Lemma 5.1. If the characteristic of k is not 2, then V ⊗ V =
Sym2 (V ) ⊕ ∧2 (V ). if we take x in characteristic 0 (over the integers) and reduce
modulo the characteristic, we see that the number of Jordan blocks on each of
GENERICALLY FREE IRREDUCIBLE REPRESENTATIONS
9
Sym2 (V ) and ∧2 (V ) can only increase in positive characteristic and so we have
equality. Thus, we may assume that k has characteristic 0.
In characteristic 0, we view V as a module under a principal SL2 and see that
V ⊗V ∼
= L(n − 1) ⊗ L(n − 1) ∼
= L(2n − 2) ⊕ L(2n − 4) ⊕ · · · ⊕ L(0), proving (1).
Examining the weights shows that ∧2 V ∼
= L(2n − 4) ⊕ L(2n − 8) ⊕ · · · , proving (3),
from which (2) follows.
Lemma 5.3. Let x ∈ gln (k) = gl(V ) with k of characteristic p 6= 2. Assume that
x has r1 Jordan blocks of even size and r2 Jordan blocks of odd size. Let s be the
number of Jordan blocks of x on Sym2 (V ) and a the number of Jordan blocks on
∧2 (V ). Then s − a = r2 .
Proof. Let J1 , . . . , Jm be the Jordan blocks for x. Then Ji ⊗ Jj for i < j occurs in
both Sym2 (V ) and ∧2 (V ). The other terms are Sym2 (Ji ) on Sym2 (V ) and ∧2 (Ji ).
Thus, the difference in the number of Jordan blocks on Sym2 (V ) and ∧2 (V ) is just
the sum of the differences on Sym2 (Ji ) and ∧2 (Ji ) and the result follows by the
previous lemma.
Put λ for the highest weight of the natural module of son , i.e., λ = ω⌊n/2⌋ . We
can now show that son acts freely on L(2λ) in characteristic not 2 by proving that
our standard inequality (1.1) holds. See [GG15, Example 10.7] or [GLL18, §4.1]
for another proof that the generic stabilizer is an elementary abelian 2-group (as a
group scheme).
Lemma 5.4. Let g = son (k) = so(V ) with n ≥ 5 and k of characteristic p =
6 2.
Set W = L(2λ). If x ∈ g is nonzero nilpotent or noncentral semisimple, then
dim xG + dim W x < dim W .
Proof. If x is semisimple, by considering weights on V , Sym2 (V ) and ∧2 (V ), we
see that dim Sym2 (V )x − dim ∧2 (V )x = dim V x , using that p 6= 2. Since dim xG +
dim ∧2 (V )x = dim G, we see that
dim xG + dim Sym2 (V )x = dim G + dim V x = dim Sym2 (V ) − (dim V − dim V x ),
which is at most dim Sym2 (V ) − 2, because the fixed space of x has codimension
at most 2. Since L(2λ) is a summand of Sym2 (V ) as an x-module and x is trivial
on a complement, the result follows.
If x is nilpotent, we argue similarly using the previous lemma. We obtain that
dim xG + dim Sym2 (V )x = dim Sym2 (x) − (n − r2 ). The result follows immediately
if r2 < n − 2. If r2 = n − 2, then x has a single Jordan block of size 3 and all other
blocks of 1. In particular, this implies that dim W x < dim Sym2 (V )x and the result
follows.
6. Example: Vinberg representations
Let G be an algebraic group over a field k and suppose θ ∈ Aut(G)(k) has finite
order m not divisible by char k. Then θ gives a Z/m-grading g = ⊕i∈Z/m gi on
the Lie algebra g of G (clear) and the sub-scheme G0 of fixed points is smooth
by [Con14, Exercise 2.4.10]. In this section we will assume furthermore that G is
semisimple simply connected, in which case G0 is connected reductive and can be
described explicitly using the recipe in [Ste68, §8]. Representations (G0 , g1 ) arising
in this way are sometimes called Vinberg representations or θ-groups.
10
S. GARIBALDI AND R.M. GURALNICK
Lemma 6.1. Let T be a maximal torus in a simple algebraic group G over a field
k. Then, (1) G has type Cn for some n ≥ 1 and char k = 2 or (2) for a generic
t ∈ Lie(T ), the transporter {x ∈ Lie(G) | [x, t] ∈ Lie(T )} equals Lie(T ).
Proof. Write x as a sum of an element x0 ∈ Lie(T ) and elements xα in the root
subalgebra for each root α. Choose t ∈ Lie(T ) generic and suppose [x, t] ∈ Lie(T ),
i.e., dα(t)xα = [xα , t] = 0 for all α. If (1) fails, then an exercise with roots as
in [CR10, Lemma 2.13] shows that dα(t) 6= 0, proving that xα = 0, whence the
claim.
Example 6.2 (m = 2). Suppose θ ∈ Aut(G)(k) has order 2 and acts on a maximal
torus T via θ(t) = t−1 for t ∈ T , so Lie(T ) is contained in g1 . As char k 6= 2, the
centralizer of a generic element in Lie(T ) in Lie(G) is just Lie(T ) which misses
g0 , whence g0 acts virtually freely on g1 . More precisely, as a group scheme, the
stabilizer of a generic element of Lie(T ) in G0 is the 2-torsion subgroup of T . In
this way, if we pick a subgroup H of G0 , we conclude that h acts generically freely
on g1 . We now consider examples where this applies; in each case a generic element
of g1 is a regular semisimple element of g, see [RLYG12, §7] or [GLL18, §4.1].
(1): Take G to have type E6 and θ to be an outer automorphism so that G0 is
the adjoint group PSp8 of type C4 , compare, for example, [GPT09, §5]. In that
case, g0 = sp8 and g1 is the Weyl module with highest weight 1000 (the “spin”
representation). If char k 6= 3 (and 6= 2), then the representation g1 is irreducible
of dimension 42.
If char k = 3, g1 has head the irreducible representation of dimension 41 and
radical k = z(g). Let v be a regular semisimple element of Lie(T ) ⊂ g1 . The
stabilizer in g0 = sp8 of the image of v in g1 /k transports v into z(g), and therefore
belongs to Lie(T ) ∩ g0 = 0 by Lemma 6.1. In particular, sp8 acts generically freely
on the irreducible representation g1 /k.
(2): Take G to be E8 and θ to be such that G0 has type D8 . In this case, G0 is
a half-spin group Spin16 /µ2 and g1 is the 128-dimensional half-spin representation.
We conclude that it acts generically freely when char k 6= 2. (Regardless of char k,
the generic stabilizer in G0 is (Z/2)4 × (µ2 )4 as a group scheme, see [GG17b,
Th. 1.2].)
(3): Take G to be E7 and θ to be such that G0 = SL8 /µ4 . In this case, g1 is
the representation ∧4 k 8 , which is generically free for char k 6= 2. (We provide a
stronger result in Prop. 7.1(1).)
(4): Take G to be SLn with θ(g) = g −⊤ , so G0 = SOn and g1 is the Weyl module
with head L(2λ) as in Lemma 5.4.
The representation ∧3 k 9 of G0 = SL9 /µ3 arises also in this way when G = E8
and m = 3, see [VE78] for a detailed analysis of the orbits in the case char k = 0. A
generic element of g1 is regular semisimple as an element of g as in the references in
Example 6.2 ([GLL18] produces an explicit regular nilpotent element), and we find
that sl9 acts generically freely on ∧3 k 9 . We will provide a stronger result below in
Prop. 7.1(1).
Remark 6.3. The setup above can be generalized to accomodate the case where
char k divides m. Instead of an element θ ∈ Aut(G)(k), one picks a homomorphism
of group schemes µm → Aut(G) defined over k. Again one obtains a Z/m-grading
on g and an action of µm on G such that G0 is smooth. Some statements about
the representation g1 of G0 from [Vin76] or [Lev09] do not hold in this generality.
GENERICALLY FREE IRREDUCIBLE REPRESENTATIONS
11
For example, the representations from Example 6.2(2) and (3) with char k = 2 and
∧3 SL9 with char k = 3, are not virtually free. This can be seen by computationally
verifying that Lemma 3.1 applies; in each of these three cases the stabilizer of a
generic vector is a toral subalgebra whose dimension we list in Table 1. Alternatively, for x ∈ g1 , x[p] is in zg0 (x), so finding any x with x[p] not in the kernel of the
representation (as is done in [Aul01, Prop. 4.8.2, 4.9.2]) suffices to show that the
representation is not virtually free.
For the spin representation of Sp8 , 2 is a special prime so is treated in [GG18].
7. Example: 3rd and 4th exterior powers
We now consider the representation ∧e (k n ) of SLn and its analogues for SOn and
Spn . Whether or not such representations are virtually free has previously been
considered in [Aul01] and elsewhere. We will check here the stronger condition of
whether or not inequality (1.1) holds for x ∈ sln .
Proposition 7.1. For the representation V := ∧e (k n ) of SLn and noncentral x ∈
sln with x[p] ∈ {0, x}, we have:
(1) If (a) e = 3 and n ≥ 10; (b) e = 3, n = 9, and char k 6= 2, 3; (c) e = 4 and
n ≥ 9; or (d) e = 4, n = 8, and char k 6= 2, then dim xSLn +dim V x < dim V
and Lie(SLn /µgcd(e,n) ) acts generically freely on V .
(2) If (a) e = 3, n = 9, and char k = 2, 3 or (b) e = 4, n = 8, and char k = 2,
then dim xSLn + dim V x ≤ dim V .
Proof. If x is noncentral toral, then we may find a nilpotent y in the closure of the
Gm SLn orbit of x such that dim y SLn = dim xSLn and (automatically) dim V x ≤
dim V y [GG17a, Lemma 3.2]. Therefore, we may assume that x[p] = 0.
If n > 16, then dim V = n3 > 2.25n2 ≥ b(SLn ), and (1.1) holds by [GG17a].
So suppose n ≤ 16. We calculate dim xSLn , which does not depend on char k,
using the well known formulas from, for example, [LS12, p. 39]. For the other term
in (1.1), dim V x , we view V as a representation of SL2 where a nilpotent element
acts as x on V . Arguing as in [McN00, §3.4], we find that if char k > en, then the
Jordan form of x acting on V is the same as in characteristic zero. Therefore, it
suffices to check the inequality over Q and over Fp for 2 ≤ p < en. This is quickly
done via computer. For the convenience of the reader, Table 4 lists the partitions
corresponding to nilpotent x for which we have equality in (2). In case gcd(e, n) = 1,
this shows that sln acts generically freely on ∧e k n ; when gcd(e, n) > 1, we verify
that Lie(SLn /µgcd(e,n) ) acts generically freely using Magma.
partition of x dim xG dim V x
(24 , 1)
40
44
(9)
72
12
(33 )
54
30
4
∧ sl8
2
(8)
56
14
(42 )
48
22
(24 )
32
38
Table 4. Complete list of nilpotent elements x from Proposition
7.1(2) where equality holds.
representation char k
∧3 sl9
2
∧3 sl9
3
12
S. GARIBALDI AND R.M. GURALNICK
Trivectors and SOn . Consider now SOn with n ≥ 9. The representation ∧3 (k n )
is a fundamental Weyl module and it is irreducible representation if char k 6= 2, see
for example [Jan03, II.8.21] and [McN00, Remark 3.4].
Proposition 7.2. For SOn with n ≥ 9 (over any field k) and V := ∧3 (k n ),
the inequality (1.1) holds for all nonzero x ∈ son with x[p] ∈ {0, x}, and g acts
generically freely on V .
Proof. Under the tautological inclusion SOn ֒→ SLn , suppose the inequality holds
for x viewed as an element of sln . Then as dim xSOn ≤ dim xSLn , the inequality
holds also for x as an element of son , completing the proof in case n ≥ 10, or n = 9
and char k 6= 2, 3 (Prop. 7.1).
So suppose n = 9 and char k = 2 or 3. Write y for the image of x in sln
if x is nilpotent, and for the image of the nilpotent specialization of x as in the
proof of Prop. 7.1 if x is toral. As in the previous paragraph, we are done if the
inequality holds for y, and therefore we may assume that y has partition (24 , 1),
(9), or (33 ) as in the first three lines of Table 4. In any of these cases, we have
dim xSO9 + dim V x ≤ 32 + dim V y < 76, completing the proof.
When char k = 2, the representation ∧3 k 2ℓ of SO2ℓ is no longer irreducible, as
the invariant alternating bilinear form generates a submodule. We now study the
irreducible module L(ωℓ−2 ) with the same highest weight.
Proposition 7.3. Suppose char k = 2 and ℓ ∈ {5, 6, 7}. For noncentral x ∈ so2ℓ
with x[2] ∈ {0, x} and V := L(ωℓ−2 ), inequality (1.1) holds. Moreover, so2ℓ acts
generically freely on V .
Proof. Recall that dim L(ω3 ) = 100, 208, 336 for ℓ = 5, 6, 7 respectively [Lüb01].
Suppose first that x[2] = 0. As every Jordan block of x has size at most 2 and
even blocks occur with even multiplicity, the element x has partition (22r , 12ℓ−4r )
for some 1 ≤ r ≤ ⌊r/2⌋ and dim xG ≤ 4r(ℓ − r) as in [GG17a, Example 9.5], so
dim xG ≤ 24, 36, 48.
For x a root element (i.e., r = 1), dim(∧3 k 2ℓ )y = 76, 146, 252 respectively, so
the inequality holds for nilpotent x if dim xG < 24, 62, 84. We are reduced to
considering the class x ∈ so10 with partition (24 , 12 ). Such an x has dim xG ≤ 24
and dim V x ≤ 64, verifying the inequality.
Suppose now that x[2] = x, so as in the proof of [GG17a, Cor. 9.6] we have
dim xG ≤ 24, 36, 48 respectively. We find a nilpotent y ∈ Gm Ad(SOn )x with
dim(k n )y ≥ 2. Now ∧3 k n is self-dual, so we find that ∧3 k n ∼
= V ⊕ k n if n = 12
n
n
and V ⊕ k ⊕ k if n = 10, 14 as SOn -modules. (That the other component is k n
or has composition factors k n , k n is clear from inspecting weight multiplicities, and
that it is a direct sum in the latter case follows because Ext1SOn (k n , k n ) = 0 [Jan03,
II.2.12(1)].) Therefore, as dim(∧3 k n )y ≤ 76, 146, 252 respectively, we have
dim V x ≤ dim V y ≤ 72, 144, 248.
Putting this together, we find dim xG + dim V x ≤ 96, 180, 296.
Trivectors and Sp2ℓ . The natural representation of Sp2ℓ has an invariant alternating bilinear form b. The subspace V (ωℓ−2 ) of ∧3 k 2ℓ spanned
by those v1 ∧v2 ∧v3
with b(vi , vj ) = 0 for all i, j is a submodule of dimension 2ℓ
3 − 2ℓ; it is the Weyl
module with highest weight ωℓ−2 . It is irreducible, i.e., V (ωℓ−2 ) = L(ωℓ−2 ), if and
GENERICALLY FREE IRREDUCIBLE REPRESENTATIONS
13
only if ℓ − 1 is nonzero in k; otherwise V (ωℓ−2 ) has socle the natural module k 2ℓ
and head L(ωℓ−2 ) [PS83, Th. 2(i)].
Proposition 7.4. In the notation of the preceding paragraph, if
(1) ℓ ≥ 5 and char k > 2; or
(2) ℓ = 4 and char k > 3,
then for V := V (ωℓ−2 ) or L(ωℓ−2 ), inequality (1.1) holds for all nonzero x ∈ sp2ℓ
with x[p] ∈ {0, x}, and therefore sp2ℓ acts generically freely on V .
Proof. If ℓ > 6, then dim V (ωℓ−2 ) ≥ dim L(ωℓ−2 ) ≥ 2ℓ
3 − 4ℓ > b(Sp2ℓ ), and the
conclusion holds by the results of [GG17a]. So suppose ℓ = 4, 5, or 6.
First suppose that x ∈ sp2ℓ has x[p] = 0. If ℓ = 5 or 6, we have dim L(ωℓ−2 )x ≤
dim V (ωℓ−2 )x ≤ dim(∧3 k 2ℓ )x and one checks that dim xSp2ℓ + dim(∧3 k 2ℓ )x <
dim L(ωℓ−2 )x , which need only be done for small characteristics as in the proof
of Proposition 7.1 and therefore amounts to a computer calculation. If ℓ = 4 (and
char k > 3), then ∧3 k 8 is a direct sum of V and k 8 , so dim V x = dim(∧3 k 8 )x −
dim(k 8 )x and the same computer calculations verify (1.1).
For V (ωℓ−2 ) and x toral, we appeal to Lemma 4.1.
Thus we are reduced to the case ℓ = 6, k = F5 , V = L(ω4 ), and x[p] = x. One
can construct the Lie algebra g := sp12 and the Weyl module V (ωℓ−2 ) in Magma.
Certainly Lie(CG (x)) = zg (x) = ker(ad(x)), and as CG (x) is smooth we find that
dim xG = dim(im ad(x)).
For the other term in (1.1), x acts semisimply on V (ω3 ), so
dim L(ωℓ−2 )x = dim(∧3 k 12 )x − 2 dim(k 12 )x .
Using these expressions, the computer can immediately verify the inequality for all
toral x in a maximal toral subalgebra of g.
8. Proof of Theorem A
In this section, we adopt the hypotheses of Theorem A and deduce it from what
has gone before. Recall that V is assumed to be restricted and irreducible for G,
hence irreducible for g [Ste63]. The main result of [GG17a] allows us to assume that
G has type Aℓ , Bℓ , Cℓ or Dℓ of bounded ℓ as in §1 and that dim V ≤ b(G) for b(G)
as in Table 3. Therefore, V appears in tables A.6–A.48 in [Lüb01].1 (Note that the
search space remains infinite: while there are only finitely many possibilities for G
and for the highest weight of V , we have not exhibited any upper bound on char k.)
We aim to show that V is virtually free as a representation of g if and only if
dim V > dim G and (G, char k, V ) does not appear in Table 1.
If dim V < dim G − dim z(g), then certainly g cannot act virtually freely on V .
If dim G ≥ dim V ≥ dim G − dim z(g), then examining the tables shows that V is
the irreducible representation with highest weight the highest root, which is not
virtually free as in [GG17a, Example 13.2]. Therefore we assume for the rest of this
section that dim V > dim G.
1For A
18 and A19 , we refer to the extended table available on his webpage at http://www.
math.rwth-aachen.de/~Frank.Luebeck/chev/WMSmall/A18-130321.html
14
S. GARIBALDI AND R.M. GURALNICK
Type A. For Aℓ with ℓ = 16, 17, 18, 19, there are no restricted irreducible representations of dimension between dim G and b(G), according to Lübeck’s tables.
Thus we are reduced to considering 3 ≤ ℓ ≤ 15.
Proposition 7.1 treats the representations ∧3 SLℓ+1 and ∧4 SLℓ+1 , apart from a
few cases. For ∧3 SL9 when char k = 2, sl9 acts generically freely on V by [Aul01,
Prop. 4.8.3], by reasoning as in §6, or as can be checked in Magma, despite the
failure of inequality (1.1). The representations ∧3 SL9 when char k = 3 or ∧4 SL8
when char k = 2 are not virtually free, see Remark 6.3.
We refer to [Gue97, Th. 4.3.2] for the representation of Aℓ (3 ≤ ℓ ≤ 9) with
highest weight 0 · · · 011 with char k = 3 (Claim 12 on p. 97 of ibid., or check with
Magma); of Aℓ (2 ≤ ℓ ≤ 9) with highest weight 0 · · · 03 (and char k > 3 so it is
restricted); of A3 with highest weight 004 and dimension 35; of A3 with highest
weight 102 and char k 6= 5; of A2 with highest weight 04 and dimension 15; and of
A2 with highest weight 13 or 22 with char k = 5.
We refer to [Aul01] to see that the following are virtually free: the representation of Aℓ (ℓ = 3, 4, 5) and char k 6= 3 with highest weight 0 · · · 011 (§4.5); the
representation of A2 with highest weight 12 and dimension 15 when char k 6= 2
(§4.1); the representation of A3 with highest weight 102 and dimension 32 when
char k = 5 (§4.2); the representation of A4 with highest weight 0101 and dimension
40 or 45 (§4.6); the representation of A4 with highest weight 0200 and dimension
45 or 50 when char k 6= 2 (§4.7); the representation of A4 with highest weight 0110
and dimension 51 when char k = 3 (§4.4); and the representation of A5 with highest
weight 01001 and dimension 78 when char k = 5 (§4.6).
The representation of A3 with highest weight 020 and char k 6= 2 is the representation S 2 (k 6 ) of so6 , and the inequality holds by Lemma 5.4.
Types B and D. For G of type Dℓ with 4 ≤ ℓ ≤ 11 or Bℓ with 2 ≤ ℓ ≤ 11 and
char k 6= 2, the representation with highest weight 0 · · · 02 is handled by Lemma
5.4. The half-spin representations of Bℓ for 7 ≤ ℓ ≤ 11 and Dℓ for ℓ = 9, 10, 11
(corresponding to Spinn for n = 15 and 17 ≤ n ≤ 23) are virtually free but for
D8 it has generic stabilizer (Z/2)4 × µ42 as a group scheme, so it is virtually free in
when char k 6= 2 and is not virtually free when char k = 2 [GG17b].
The representation ∧3 (k n ) of SOn with n = 9, . . . , 13 (i.e., B4 , B5 , B6 , D5 , D6 )
with char k 6= 2 is virtually free by Proposition 7.2.
The representations of B2 with highest weight 11 and B3 with highest weight
101 are handled in Examples 4.2 and 4.3.
The representation of B3 with highest weight 200 and dimension 35 appears
as a summand in S 2 X for X the (8-dimensional spin representation); we have
S2X ∼
= V ⊕ k. The action on V factors through the action of D4 as in Lemma
5.4, whence we have the inequality for V . Similarly the representation of B4 with
highest weight 2000 and dimension 126 is virtually free because it factors through
the virtually free representation of D5 as in Example 4.5.
We refer to [Gue97, Th. 4.3.3] for the representations of B4 with highest weight
1001 and dimension 112 or 128; of B3 with highest weight 011 and dimension 63
and char k = 3; of B3 with highest weight 110 and dimension 64 and char k = 5; of
B2 with highest weights 30, 12, 03, or 21.
The representation of D4 with highest weight 1001 has dimension greater than
dim D4 and is virtually free as in Example 4.4.
GENERICALLY FREE IRREDUCIBLE REPRESENTATIONS
15
For the representation of D5 with highest weight 10001 of dimension 144 with
char k 6= 2, 5, [Gue97, Th. 4.3.5] proves it is virtually free. If char k = 2, that
representation has dimension 144 > b(G) and the inequality holds. If char k = 5,
one checks with Magma that a random vector has zero-dimensional stabilizer.
The representation with highest weight 0 · · · 0100 of Dℓ for ℓ = 5, 6, 7 and
char k = 2 is virtually free by Prop. 7.3.
The representation of D5 with highest weight 20000 of dimension 126 is virtually
free by Example 4.5.
Type C. Type C is similar to types B and D. The only case for which we refer
to [Gue97] is type C3 with highest weight 011 of dimension 50 with char k = 3
(Th. 4.3.4), which can also be checked using Magma. The representations ∧30 (Sp2ℓ )
with ℓ = 4, 5, 6 and char k not dividing ℓ − 1 or C6 with highest weight 000100 are
generically free by Proposition 7.4. For C4 with highest weight 0100 and dimension
40 in characteristic 3, see Example 3.2. The representation of C5 with highest
weight 10000 is generically free by Example 4.6.
We use Magma to verify that a random vector has trivial stabilizer when char k =
3 for C5 with highest weight 01000 and dimension 121.
The remaining representation V of C4 with highest weight 1000, which was
treated in Example 6.2. It has dim V > dim C4 and V is generically free.
9. Proof of Theorem B
In this section we will prove Theorem B.
Lemma 9.1. Let G be a simple algebraic group. For every dominant weight λ 6= 0,
i
i
g acts virtually freely on L(λ) ⊗ L(λ)[p] and L(λ) ⊗ (L(λ)∗ )[p] for all i > 0.
Proof. Suppose first that G = SLn and L(λ) is the natural module. The representation dρ of sln on V is equivalent to a direct sum of n copies of the natural
module, i.e., is equivalent to sln acting on n-by-n matrices by left multiplication.
A generic matrix v is invertible, so the generic stabilizer (sln )v is zero. (We remark
that the group SLn has finitely many orbits on P(V ) [GLMS97, Lemma 2.6].)
Otherwise, the representation factors through SL(L(λ)), and the previous paragraph shows that sl(L(λ)) acts generically freely.
Note that, in the lemma, the inequality (1.1) need not hold. Specifically, a root
element x ∈ sln has dim xSLn = 2(n − 1) and kernel of dimension n − 1 on the
natural module, so we find dim xSLn + dim V x = dim V + n − 2 for V a sum of n
copies of the natural module.
Example 9.2. Consider now SOn for n ≥ 3 and suppose that char k 6= 2 or n is
even. Take V to be a direct sum of n − 1 copies of the natural module V0 . Let
v ∈ V be a vector each of whose n − 1 components is a generic element of V0 ; in
particular, the SOn -invariant quadratic form q is nonzero on each component of v.
The n − 1 components generate a codimension 1 subspace U of V0 on which the
bilinearization of q is nondegenerate if char k 6= 2 or has a 1-dimensional radical
on which q does not vanish if char k = 2. In either case, an element of son that
annihilates U is zero on V0 , hence son acts generically freely on V . (See [BGS17]
for more general arguments in a similar vein.)
16
S. GARIBALDI AND R.M. GURALNICK
Proof of Theorem B. Adopt now the hypotheses of Theorem B. Note that p :=
char k 6= 0 because when char k = 0 irreducible representations P
are tensor indecomposable. We may write the highest weight of V as λ = λ0 + i>0 pi µi where λ0
and each µi is a restricted dominant weight. Since char k is not special for G, the
statement that V is tensor decomposable means that at least two of these weights
are nonzero [Sei87, 1.6]. If λ0 = 0, then V is a Frobenius twist of an irreducible
representation and g acts P
trivially, so V is virtually free. So assume λ0 6= 0. Then
λ = λ0 + pλ1 for λ1 := i>0 pi−1 µi 6= 0 and V ∼
= L(λ0 ) ⊗ L(λ1 )[p] . As g acts
[p]
trivially on L(λ1 ) , the representation V of g is the same as a sum of dim L(λ1 )
copies of L(λ0 ).
Let m(G) be the dimension of the smallest nonzero restricted irreducible representation of G. If dim L(λ0 ) > b(G)/m(G) for b(G) as in Table 3, then dim V >
b(G) and g acts virtually freely on V by part I. If dim L(λ0 ) = m(G), then V contains L(λ0 ) ⊗ L(λ0 )[p] as a summand, and we are done by Lemma 9.1. Therefore,
it remains to inspect ⊕m(G) L(λ0 ) for those nonzero restricted dominant weights λ0
with m(G) < dim L(λ0 ) ≤ b(G)/m(G), the list of which may be extracted from the
tables in [Lüb01].
For G2 , the tables show there are no λ0 to be inspected. For the other exceptional
groups, m(G) > b(G)/m(G). Thus the theorem is proved for all exceptional types.
Types B, C, and D. With type Cℓ for ℓ ≥ 2, m(G) = 2ℓ, so b(G)/m(G) = 3ℓ and
the unique reduced irreducible representation of dimension in the target range is
C2 where L(λ0 ) is the 5-dimensional fundamental irreducible representation, i.e., g
is sp4 acting on four copies of the 5-dimensional module and so a generic stabilizer
in sp4 is the annihilator of a 4-dimensional space. We are not in characteristic 2, so
generically this is non-degenerate but if something kills a non-degenerate 4-space
it preserves the orthogonal complement and so also the entire 5-dimensional space
(since the trace is 0) and so a generic stabilizer is central.
For type Dℓ with ℓ ≥ 4, m(G) = 2ℓ. The unique dominant weight λ0 that must
be considered is for type D5 with char k 6= 2 and L(λ0 ) a half-spin representation.
One checks that for regular x ∈ spin10 , dim L(λ0 )x < 12, so (1.1) holds. On the
other hand, for non-regular and noncentral x ∈ spin10 , dim L(λ0 )x ≤ 12 by [GG17b,
Prop. 2.1(i)] and again (1.1) holds, completing the proof of the theorem for type
D.
For type Bℓ with ℓ ≥ 2, m(G) = 2ℓ+1. The only λ0 to be considered is B3 where
L(λ0 ) is the 8-dimensional spin representation. In this case, the representation
factors through Spin8 as a vector representation, and we apply Example 9.2 to see
that the generic stabilizer in so8 is trivial and therefore the same is true for spin7 .
Type A. For type Aℓ with ℓ ≥ 1, we will verify that (1.1) holds for x ∈ slℓ+1 with
x[p] = 0. This will imply that the inequality also holds for noncentral toral x as in
the proof of Proposition 7.1, whence the claim.
For type A1 , L(λ0 ) is virtually free unless λ0 = 1 or 2 [GG17a, Example 1.4]
so we need only consider V = L(2) ⊕ L(2) when char k ≥ 3, i.e., where V is a
sum of two copies of the adjoint representation. For x nonzero nilpotent we have
dim xSLn = 2 and dim L(2)x = 1, so (1.1) holds.
For ℓ ≥ 2, we again refer to [Lüb01] to enumerate the weights λ0 such that
ℓ + 1 < dim L(λ0 ) ≤ b(G)/(ℓ + 1), yielding three cases to be checked:
(i) A2 , where char k 6= 2 and L(λ0 ) = S 2 (k 3 );
GENERICALLY FREE IRREDUCIBLE REPRESENTATIONS
17
(ii) A3 , where L(λ0 ) = ∧2 (k 4 ); and
(iii) A4 , where char k 6= 2 and L(λ0 ) = ∧2 (k 5 ).
We write the details of the verification in the most complicated case, (iii). For
x ∈ sl5 a root element, i.e., nilpotent with partition (2, 13 ), we have dim xG = 8
and dim L(λ0 )x = 7, and 8 + 5 · 7 = 43 < 50. For x nilpotent with partition (2, 2, 1),
we have dim xG = 12 and dim L(λ0 )x = 6, and 12 + 5 · 6 = 42 < 50.
For x nilpotent with partition (3, 12 ), we have dim L(λ0 )x = 4 and dim xG ≤
dim G − rank G = 20. Consequently, for every nilpotent y ∈ sl5 such that x ∈
Ad(G)y, we have:
dim y G + dim V y ≤ (dim G − rank G) + m(G) · dim L(λ0 )x < dim V.
Thus we have verified the inequality (1.1) for every nonzero nilpotent in sl5 .
For cases (i) and (ii) respectively, the role of (3, 12 ) is instead played by the
smallest class xG such that dim L(λ0 )x < dim L(λ0 ) − (dim G − rank G)/m(G),
where the right side is 4 or 3 respectively.
References
[AP71]
E.M. Andreev and V.L. Popov, Stationary subgroups of points of general position in
the representation space of a semisimple Lie group, Functional Anal. Appl. 5 (1971),
no. 4, 265–271.
[Aul01]
A. Auld, Exceptional modular representations of special linear Lie algebras, Ph.D.
thesis, University of Manchester, February 2001.
[AVE68]
E.M. Andreev, E.B. Vinberg, and A.G. Elashvili, Orbits of greatest dimension in
semi-simple linear Lie groups, Funct. Anal. Appl. 1 (1968), 257–261.
[BCP90]
W. Bosma, J. Cannon, and C. Playoust, The Magma algebra system I: the user
language, J. Symbolic Computation 9 (1990), 677–698.
[BGS17]
T. Burness, R. Guralnick, and J. Saxl, On base sizes for algebraic groups, J. Eur.
Math. Soc. (JEMS) 19 (2017), no. 8, 2269–2341.
[BRV10]
P. Brosnan, Z. Reichstein, and A. Vistoli, Essential dimension, spinor groups, and
quadratic forms, Ann. of Math. (2) 171 (2010), no. 1, 533–544.
[Con14]
B. Conrad, Reductive group schemes, Autour des Schémas en Groupes I, Panoramas
et synthèses, vol. 42–43, Société Mathématique de France, 2014.
[CR10]
P.-E. Chaput and M. Romagny, On the adjoint quotient of Chevalley groups over
arbitrary base schemes, J. Inst. Math. Jussieu 9 (2010), no. 4, 673–704.
[Èla72]
A. G. Èlašvili, Canonical form and stationary subalgebras of points in general position
for simple linear Lie groups, Funkcional. Anal. i Priložen. 6 (1972), no. 1, 51–62.
[GG15]
S. Garibaldi and R.M. Guralnick, Simple groups stabilizing polynomials, Forum of
Mathematics: Pi 3 (2015), e3 (41 pages).
, Generically free representations I: Large representations, November 2017,
[GG17a]
arxiv:1711.05502.
[GG17b]
, Spinors and essential dimension, Compositio Math. 153 (2017), 535–556,
with an appendix by A. Premet.
[GG18]
, Generically free representations III: Exceptionally bad characteristic, January 2018, arxiv:1801.06915.
[GLL18]
R.M. Guralnick, R. Lawther, and M. Liebeck, Generic stabilizers for actions of simple
algebraic groups, in preparation, 2018.
[GLMS97] R.M. Guralnick, M.W. Liebeck, D. Macpherson, and G.M. Seitz, Modules for algebraic
groups with finitely many orbits on subspaces, J. Algebra 196 (1997), 211–250.
[GPT09]
S. Garibaldi, R. Parimala, and J.-P. Tignol, Discriminant of symplectic involutions,
Pure Applied Math. Quarterly 5 (2009), no. 1, 349–374.
[Gue97]
M. Guerreiro, Exceptional representations of simple algebraic groups in prime characteristic, Ph.D. thesis, University of Manchester, 1997, arxiv:1210.6919.
18
[Hum67]
S. GARIBALDI AND R.M. GURALNICK
J.E. Humphreys, Algebraic groups and modular Lie algebras, Memoirs of the American Mathematical Society, No. 71, American Mathematical Society, Providence, R.I.,
1967.
[Jan03]
J.C. Jantzen, Representations of algebraic groups, second ed., Math. Surveys and
Monographs, vol. 107, Amer. Math. Soc., 2003.
[Kar10]
N. Karpenko, Canonical dimension, Proceedings of the International Congress of
Mathematicians 2010, World Scientific, 2010.
[Lev09]
P. Levy, Vinberg’s θ-groups in positive characteristic and Kostant-Weierstrass slices,
Transform. Groups 14 (2009), no. 2, 417–461.
[LMMR13] R. Lötscher, M. MacDonald, A. Meyer, and Z. Reichstein, Essential dimension of
algebraic tori, J. Reine Angew. Math. 677 (2013), 1–13.
[Löt13]
R. Lötscher, A fiber dimension theorem for essential and canonical dimension, Compositio Math. 149 (2013), no. 1, 148–174.
[LS12]
M. Liebeck and G. Seitz, Unipotent and nilpotent classes in simple algebraic groups
and Lie algebras, Math. Surveys and Monographs, no. 180, Amer. Math. Soc., 2012.
[Lüb01]
F. Lübeck, Small degree representations of finite Chevalley groups in defining characteristic, LMS J. Comput. Math. 4 (2001), 135–169.
[McN00]
G.J. McNinch, Semisimplicity of exterior powers of semisimple representations of
groups, J. Algebra 225 (2000), 646–666.
[Mer13]
A. Merkurjev, Essential dimension: a survey, Transform. Groups 18 (2013), no. 2,
415–481.
[Pop88]
A.M. Popov, Finite isotropy subgroups in general position of simple linear Lie groups,
Trans. Moscow Math. Soc. (1988), 205–249, [Russian original: Trudy Moskov. Mat.
Obschch. 50 (1987), 209–248, 262].
[PS83]
A.A. Premet and I.D. Suprunenko, The Weyl modules and the irreducible representations of the symplectic group with the fundamental highest weights, Comm. Algebra
11 (1983), no. 12, 1309–1342.
[PV94]
V.L. Popov and E.B. Vinberg, Invariant theory, Encyclopedia of Mathematical Sciences, vol. 55, pp. 123–284, Springer-Verlag, 1994.
[Rei10]
Z. Reichstein, Essential dimension, Proceedings of the International Congress of
Mathematicians 2010, World Scientific, 2010.
[RLYG12] M. Reeder, P. Levy, J.-K. Yu, and B.H. Gross, Gradings of positive rank on simple
Lie algebras, Transform. Groups 17 (2012), no. 4, 1123–1190.
[Sei87]
G.M. Seitz, The maximal subgroups of classical algebraic groups, Mem. Amer. Math.
Soc. 67 (1987), no. 365.
[Ste63]
R. Steinberg, Representations of algebraic groups, Nagoya Math. J. 22 (1963), 33–56,
[= Collected Papers, pp. 149–172].
, Endomorphisms of linear algebraic groups, Mem. Amer. Math. Soc. 80
[Ste68]
(1968), 1–108, (= Collected Papers #23).
[VE78]
E.B. Vinberg and A.G. Elashvili, A classification of the three-vectors of ninedimensional space, Trudy Sem. Vektor. Tenzor. Anal. 18 (1978), 197–233.
[Vin76]
E.B. Vinberg, The Weyl group of a graded Lie algebra, Math. USSR Izv. 10 (1976),
463–495.
| 4math.GR
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.