Tchebycheff systems and extremal

problems for generalized moments:

a brief survey
arXiv:1107.3493v1 [math.OC] 18 Jul 2011

Iosif Pinelis∗
Department of Mathematical Sciences
Michigan Technological University
Houghton, Michigan 49931, USA
E-mail: [email protected]

Abstract: A brief presentation of basics of the theory of Tchebycheff and

Markov systems of functions and its applications to extremal problems for
integrals of such functions is given. The results, as well as all the necessary
definitions, are stated in most common terms. This work is motivated by
specific applications in probability and statistics. A few related questions
are also briefly discussed, including the one on the existence of a Tcheby-
cheff system on a given topological space.

AMS 2000 subject classifications: Primary 26D15 49K30, 49K27, 52A40;

secondary 49K35, 52A20, 52A41, 60E15, 41A50, 41A52, 26A48, 26B25,
46N10, 46N30, 90C05, 90C25, 90C26, 90C47, 90C48.
Keywords and phrases: Tchebycheff systems, Markov systems, extremal
problems, generalized moments, exact inequalities, generalized monotonic-
ity, generalized convexity, stochastic orders, best approximation.

The theory of Tchebycheff systems (or, briefly, T -systems) presents powerful

tools that can be used in various extremal problems in analysis, probability,
and statistics; see e.g. monographs [4, 6] and the extensive bibliography there.
However, whereas the Carathéodory Principle (as stated below) and the related
duality principle (see e.g. [6, Theorem 4.1]) have been used extensively in prob-
ability and statistics (see e.g. [4, Chapters XII and XIII] and [12, 10, 11, 9]),
there appear to have been very few (if any) such applications of the theory of
T -systems, even though the latter offers significant advantages. Even in [4], I
have not found applications in probability or statistics of the T -systems theory
per se.
This brief survey was motivated by specific applications in [8]. Here we shall
present basics of the theory of T -systems and its applications to extremal prob-
lems for the corresponding generalized moments. The results, as well as all
the necessary definitions, will be stated in most common terms and thus, it is
hoped, easy to use; the notions of canonical and principal representations will
be avoided here. A few related questions will also be briefly discussed, including
the one on the existence of a T -system on a given topological space.
***
∗ Supported by NSF grant DMS-0805946
1
imsart-generic ver. 2009/05/21 file: arxiv.tex date: July 19, 2011
 Iosif Pinelis/T -systems review 2

For a nonnegative integer n, let g0 , . . . , gn be (real-valued) continuous func-

tions on a compact topological space X. Let M denote the set of all (nonnega-
tive) Borel measures on X. Take any point c = (c0 , . . . , cn ) ∈ Rn+1 such that
n Z o
Mc := µ ∈ M : gi dµ = ci for all i ∈ 0, n 6= ∅; (1)
X

Consider also the condition that the (generalized) polynomial

n
X
λi gi is strictly positive on X, for some (λ0 , . . . , λn ) ∈ Rn+1 . (2)
0

Carathéodory Principle. (See e.g. [3, 5].) If theRtopologcal compact space X

is Hausdorff and (2) holds, then the maximum of X gn+1 dµ over all µ ∈ Mc
is attained at some measure µmax ∈ Mc with card supp µmax 6 n + 1.
Here, as usual, card stands for the cardinality and supp µ denotes the support
set of the measure µ. Note that in [5] it is additionally assumed that g0 = 1,
which is used to provide for the weak compactness of Mc . However, the same
effect is achieved under the more general condition (2).
Remark 1. The condition that X be compact can oftentimes be circumvented
by using, for instance, an appropriate compactification, say X, of X if X is (say)
only locally compact (as, for instance, Rk is). At that, one may be able to find
a function h, which is positive and continuous on X and such that the functions
g0 gn
h , . . . , h can be continuously extended from X to X; sometimes one of the gi ’s
can play the role Pnof h; or, more generally, h can be constructed as (or based on)
a polynomial 0 λi gi . Replacing then g0 , . . . , gn by the continuous extensions
of the functions gh0 , . . . , ghn to X and, correspondingly, replacing the measure µ
by the measure ν (on X and hence on X) defined by the formula dν = h dµ,
one will largely reduce the original optimization problem on X to one on the
compact space X.
The essential fact is that the upper bound n + 1 on the cardinality of the
support of an extremal measure µ given in the Carathéodory Principle can be
approximately halved in the presence of the Tchebycheff or, especially, Markov
property.
Definition 1. The sequence (g0 , . . . , gn ) of functions is a T -system if the re-
strictions of these n + 1 functions to any subset of X of cardinality n + 1 are
linearly independent. If, for each k ∈ 0, n, the initial subsequence (g0 , . . . , gk )
of the sequence (g0 , . . . , gn ) is a T -system, then (g0 , . . . , gn ) is said to be an
M -system (where M refers to Markov).
By Haar’s theorem, linearly independent functions g0 , . . . , gn on X form a
T -system on X if and only the problem of best P uniform approximation of any
given continuous function on X by a polynomial n0 λi gi has a unique solution;
see e.g. [13].
Any T -system satisfies the condition (2); see e.g. [6, Theorem II.1.4].

imsart-generic ver. 2009/05/21 file: arxiv.tex date: July 19, 2011

 Iosif Pinelis/T -systems review 3

For any n > 1 and any topological space X of cardinality > n + 1, if there
exists a T -system (g0 , . . . , gn ) of continuous functions on X, then X is necessarily
Hausdorff. Indeed, take any distinct x0 and x1 in X. Let x2 , . . . , xn be any points
in X such that x0 , . . . , xn are distinct. The restrictions of the functions g0 , . . . , gn
to the set {x0 , . . . , xn } are linearly independent and hence gi (x0 ) 6= gi (x1 ) for
some i ∈ 0, n. Take now any disjoint open sets O0 and O1 in R containing gi (x0 )
and gi (x1 ), respectively. Then the pre-images gi−1 (O0 ) and gi−1 (O1 ) of O0 and
O1 under the mapping gi are disjoint open sets in X containing x0 and x1 ,
respectively. Thus, X is a Hausdorff topological space.
In connection with Remark 1, it should be noted that, clearly, if (g0 , . . . , gk ) is
a T -system or an M -system, then the same is true of the sequence ( gh0 , . . . , ghn ),
for any positive continuous function h. Note also that, if (g0 , . . . , gn ) is a T -
system on a set X ′ containing X and such that card(X ′ \ X) > n, then
(h0 , . . . , hn ) := A(g0 , . . . , gn ) is an M -system on X for some linear (necessarily
nonsingular) transformation A of Rn+1 ; cf. [6, Theorem II.4.1].
A T -system (g0 , . . . , gn ) with n > 1 on (the compact topological space) X
exists only if X is one-dimensional (which will be the case in many applications).
More precisely, if for some n > 1 there exists a T -system of n + 1 functions on
X, then X must be homeomorphic to a subset of a circle; for X ⊆ Rk this was
proved in [7], and for general X in [2] (with an additional restriction) and in
[16]; a further extension of this result to complex T -systems was given in [14],
where one can also find yet another proof of the real-valued version.
In fact, the general case of (real-valued) T -systems can be easily reduced to
the special case with X ⊆ Rk . Indeed, for any natural n consider the mapping
x 7→ r(x) := g(x)/kg(x)k of X into the unit sphere S n in Rn+1 , where g(x) :=
(g0 (x), . . . , gn (x)) and k · k is the Euclidean norm. In view of the T -property
of (g0 , . . . , gn ) and the compactness
of X, this mapping is correctly defined
since g(x) is nonzero for any x ∈ X , one-to-one, and continuous, and hence
a homeomorphism of the compact Hausdorff set X onto the image in S n of X
under the mapping r. In the case n = 1, this also proves the mentioned result
of [7, 2, 16]. Another elementary observation in this regard, presented in [1], is
that a T -system (g0 , . . . , gn ) with n > 1 on X may exist only if X does not
contain a “tripod”, that is a set homeomorphic to the set {(s, 0) ∈ R2 : |s| <
1} ∪ {(0, t) ∈ R2 : 0 < t < 1}.
We shall henceforth consider the case when X = [a, b] for some a and b
such that
−∞ < a < b < ∞. Let (g0 , . . . , gn ) be a T -system on [a, b]. Let
n

det gi (xj ) 0 denote the determinant of the matrix gi (xj ) : i ∈ 0, n, j ∈ 0, n .
This determinant is continuous in (x0 , . . . , xn ) in the (convex) simplex (say Σ)
defined by the inequalities
n a 6 x0 < · · · < xn 6 b and does not vanish anywhere
on Σ. So, det gi (xj ) 0 is constant in sign on Σ.
Definition
2. The sequence (g0 , . . . , gn ) is said to be a T+ -system on [a, b] if
n
det gi (xj ) 0 > 0 for all (x0 , . . . , xn ) ∈ Σ. If (g0 , . . . , gk ) is a T+ -system on [a, b]
for each k ∈ 0, n, then the sequence (g0 , . . . , gn ) is said to be an M+ -system on
[a, b].
Clearly, if (g0 , . . . , gn ) is a T -system on [a, b], then either (g0 , . . . , gn ) or

imsart-generic ver. 2009/05/21 file: arxiv.tex date: July 19, 2011

 Iosif Pinelis/T -systems review 4

(g0 , . . . , gn−1 , −gn ) is a T+ -system on [a, b]. Similarly, if (g0 , . . . , gn ) is an M -

system on [a, b] then, for some sequence (s0 , . . . , sn ) ∈ {−1, 1}n+1 of signs,
(s0 g0 . . . , sn gn ) is an M+ -system on [a, b].
In the case when the functions g0 , . . . , gn are n times differentiable at a point
x ∈ (a, b), consider also the Wronskians
(j)
k
W0k (x) := det gi (x) 0 ,

(j) (0)
where k ∈ 0, n and gi is the jth derivative of gi , with gi := gi ; in particular,
W00 (x) = g0 (x).
Proposition 1. Suppose that the functions g0 , . . . , gn are (still continuous on
[a, b] and) n times differentiable on (a, b). Then, for the sequence (g0 , . . . , gn ) to
be an M+ -system on [a, b], it is necessary that W0k > 0 on (a, b) for all k ∈ 0, n,
and it is sufficient that u0 > 0 on [a, b] and W0k > 0 on (a, b) for all k ∈ 1, n.
Thus, verifying the M+ -property largely reduces to checking the positivity of
several functions of only one variable.
A special case of Proposition 1 (with n = 1 and g0 = 1) is the following well-
known fact: if a function g1 is continuous on [a, b] and has a positive derivative
on (a, b), then g1 is (strictly) increasing on [a, b]; vice versa, if g1 is increasing
on [a, b], then the derivative of g1 (if exists) must be nonnegative on (a, b).
As in this special case, the proof of Proposition 1 in general can be based on
the mean-value theorem; cf. e.g. [4, Theorem 1.1 of Chapter XI], which states
that the requirement for W0k to be strictly positive on the closed interval [a, b]
for all k ∈ 0, n is equivalent to a condition somewhat stronger than being an
M+ -system on [a, b]; in connection with this, one may also want to look at [6,
Theorem IV.5.2]. Note that, in the applications to the proofs of [8, Lemmas 2.2
and 2.3], the relevant Wronskians vanish at the left endpoint of the interval.
The proof of Proposition 1 can be obtained by
n induction on n using the
recursive formulas for the determinants det gi (xj ) 0 and W0n as displayed right
above [4, (5.5) in Chapter VIII] and in [4, (5.6) in Chapter VIII], where we use
gi in place of ψi .
Proposition 2. Suppose that (g0 , . . . , gn+1 ) is an M+ -system on [a, b] or, more
generally, each of the sequences (g0 , . . . , gn ) and (g0 , . . . , gn+1 ) is a T+ -system
on [a, b]. Suppose also that condition (1) holds. Let m := ⌊ n+1 2 ⌋. Then one has
the following.
Rb
(I) The maximum (respectively, the minimum) of a gn+1 dµ over all µ ∈ Mc
is attained at a unique measure µmax (respectively, µmin ) in Mc . More-
over, the measures µmax and µmin do not depend on the choice of gn+1 ,
as long as gn+1 is such that (g0 , . . . , gn+1 ) is a T+ -system on [a, b].
(II) There exist subsets Xmax and Xmin of [a, b] such that Xmax ⊇ supp µmax ,
Xmin ⊇ supp µmin , and
(a) if n = 2m then card Xmax = card Xmin = m + 1, Xmax ∋ b, and
Xmin ∋ a;

imsart-generic ver. 2009/05/21 file: arxiv.tex date: July 19, 2011

 Iosif Pinelis/T -systems review 5

(b) if n = 2m − 1 then card Xmax = m + 1, card Xmin = m, and Xmax ⊇

{a, b}.
Whereas the maximizer µmax and the minimizer µmin are each unique in Mc
for each given c with Mc 6= ∅, in particular applications such as the ones in [8],
one may want to allow the vector c to vary, and then the measures µmax and
µmin will vary with c, and thus the corresponding subsets Xmax and Xmin of [a, b]
may vary. Then the number of real variables needed to describe each of the sets
Xmax and Xmin will be about n+1 2 , that is, half the number of restrictions on the
measure µ and also half the upper bound on card supp µmax in the Carathéodory
Principle; here one should also take into account that, as described in part (II)
of Proposition 2, the sets Xmax and Xmin may have to contain at least one of the
endpoints a and b of the interval [a, b], with the corresponding reduction in the
required number of variables. On the other hand, the Carathéodory Principle
holds for more general systems of functions, defined on a set X of a much more
general class.
To illustrate Proposition 2, one may consider the simplest two special cases
when the conditions of the proposition hold and its conclusion is obvious:
(i) n = 0, g0 (x) ≡ 1, g1 is increasing on [a, b], and c0 > 0; then supp µmax ⊆
{b} and supp µmin ⊆ {a}; in fact, µmax = c0 δb and µmin = c0 δa ; here and
in what follows, δx denotes the Dirac probability measure at point x.
(ii) n = 1, g0 (x) ≡ 1, g1 (x) ≡ x, g2 is strictly convex on [a, b], c0 > 0, and
c1 ∈ [c0 a, c0 b]; then supp µmax ⊆ {a, b} and card supp µmin 6 1; in fact,
µmax = c0b−a
b−c1
δa + c1b−a
−c0 a
δb , and µmin = c0 δc1 /c0 if c0 > 0 and µmin = 0 if
c0 = 0.
These examples also show that the T -property of systems of functions can be
considered as generalized monotonicity/convexity; see e.g. [15] and bibliography
there.
Proof of Proposition 2. Consider two cases, depending on whether c is strictly
or singularly positive; in equivalent geometric terms, this means, respectively,
that c belongs to the interior or the boundary of the smallest closed convex
cone containing the subset {(g0 (x), . . . , gn (x)) : x ∈ [a, b]} of Rn+1 [6, Theo-
rem IV.6.1].
In the first case, when c is strictly positive, both statements of Proposition 2
follow by [6, Theorem IV.1.1]; at that, one should let Xmax = supp µmax and
Xmin = supp µmin . (The condition that c be strictly positive appears to be miss-
ing in the statement of the latter theorem; cf. [4, Theorem 1.1 of Chapter 1.1].)
In the other case, when c is singularly positive, use [6, Theorem III.4.1], which
states that in this case the set Mc consists of a single measure (say µ∗ ), and its
support set X∗ := supp µ∗ is of an index 6 n; that is, ℓ− + 2ℓ + ℓ+ 6 n, where
ℓ− , ℓ, and ℓ+ stand for the cardinalities of the intersections of X∗ with the sets
{a}, (a, b), and {b}. It remains to show that this condition on the index of X∗
implies that there exist subsets Xmax and Xmin of [a, b] satisfying the conditions
(IIa) and (IIb) of Proposition 2 and such that Xmax ∩ Xmin ⊇ X∗ .

imsart-generic ver. 2009/05/21 file: arxiv.tex date: July 19, 2011

 Iosif Pinelis/T -systems review 6

If n = 2m then card(X∗ ∩ (a, b)) = ℓ 6 ⌊ 2m−ℓ2− −ℓ+ ⌋ 6 ⌊ 2m−ℓ

2
−
⌋ = m − ℓ− ; so,
card(X∗ ∪ {b}) 6 ℓ− + (m − ℓ− ) + 1 = m + 1. Adding now to the set X∗ ∪ {b} any
m+ 1 − card(X∗ ∪{b}) points of the complement of X∗ ∪{b} to [a, b], one obtains
a subset Xmax of [a, b] such that Xmax ⊇ X∗ , Xmax ∋ b, and card Xmax = m + 1.
Similarly, there exists a subset Xmin of [a, b] such that Xmin ⊇ X∗ , Xmin ∋ a,
and card Xmin = m + 1.
If n = 2m − 1 then card(X∗ ∩ (a, b)) = ℓ 6 ⌊ 2m−1−ℓ2
− −ℓ+
⌋ 6 m − 1 and hence
card(X∗ ∪ {a, b}) 6 1 + (m − 1) + 1 = m + 1. So, there exists a subset Xmax of
[a, b] such that Xmax ⊇ X∗ , Xmax ⊇ {a, b}, and card Xmax = m + 1. One also
has card X∗ = ℓ− + ℓ + ℓ+ 6 ⌊ 2m−1+ℓ 2
− +ℓ+
⌋ 6 ⌊ 2m+1
2 ⌋ = m. So, there exists a
subset Xmin of [a, b] such that Xmin ⊇ X∗ and card Xmin = m.

References

[1] R. C. Buck. Linear spaces and approximation theory. In On numerical

approximation. Proceedings of a Symposium, Madison, April 21-23, 1958,
Edited by R. E. Langer. Publication no. 1 of the Mathematics Research
Center, U.S. Army, the University of Wisconsin, pages 11–23. The Univer-
sity of Wisconsin Press, Madison, 1959.
[2] P. C. Curtis, Jr. n-parameter families and best approximation. Pacific J.
Math., 9:1013–1027, 1959.
[3] W. Hoeffding. The extrema of the expected value of a function of indepen-
dent random variables. Ann. Math. Statist., 26:268–275, 1955.
[4] S. Karlin and W. J. Studden. Tchebycheff systems: With applications in
analysis and statistics. Pure and Applied Mathematics, Vol. XV. Inter-
science Publishers John Wiley & Sons, New York-London-Sydney, 1966.
[5] A. F. Karr. Extreme points of certain sets of probability measures, with
applications. Math. Oper. Res., 8(1):74–85, 1983.
[6] M. G. Kreı̆n and A. A. Nudel′ man. The Markov moment problem and
extremal problems. American Mathematical Society, Providence, R.I., 1977.
Ideas and problems of P. L. Čebyšev and A. A. Markov and their further
development, Translated from the Russian by D. Louvish, Translations of
Mathematical Monographs, Vol. 50.
[7] J. C. Mairhuber. On Haar’s theorem concerning Chebychev approximation
problems having unique solutions. Proc. Amer. Math. Soc., 7:609–615,
1956.
[8] I. Pinelis. An asymptotically Gaussian bound on the Rademacher tails
(preprint), arXiv:1007.2137 [math.PR].
[9] I. Pinelis. Exact lower bounds on the exponential moments of Winsorized
and truncated random variables (preprint), arXiv:1001.2901 [math.PR].
[10] I. Pinelis. Optimal tail comparison based on comparison of moments.
In High dimensional probability (Oberwolfach, 1996), volume 43 of Progr.
Probab., pages 297–314. Birkhäuser, Basel, 1998.
[11] I. Pinelis. On normal domination of (super)martingales. Electron. J.
Probab., 11:no. 39, 1049–1070, 2006.

imsart-generic ver. 2009/05/21 file: arxiv.tex date: July 19, 2011

 Iosif Pinelis/T -systems review 7

[12] I. S. Pinelis and S. A. Utev. Sharp exponential estimates for sums of

independent random variables. Theory Probab. Appl., 34(2):340–346, 1989.
[13] I. J. Schoenberg. On the question of unicity in the theory of best approxi-
mations. Ann. New York Acad. Sci., 86:682–692 (1960), 1960.
[14] I. J. Schoenberg and C. T. Yang. On the unicity of solutions of problems
of best approximation. Ann. Mat. Pura Appl. (4), 54:1–12, 1961.
[15] M. Shaked and J. G. Shanthikumar. Stochastic orders. Springer Series in
Statistics. Springer, New York, 2007.
[16] K. Sieklucki. Topological properties of sets admitting the Tschebycheff
systems. Bull. Acad. Polon. Sci. Sér. Sci. Math. Astr. Phys., 6:603–606,
1958.

imsart-generic ver. 2009/05/21 file: arxiv.tex date: July 19, 2011