The Problem of Moments: A Bunch of Classical Results With Some Novelties

We summarize significant classical results on (in)determinacy of measures in terms of their finite positive integer order moments. Well-known is the role of the smallest eigenvalues of Hankel matrices, starting from Hamburger's results a century ago and ending with the great progress made only in recent times by C. Berg and collaborators. We describe here known results containing necessary and sufficient conditions for moment (in)determinacy in both Hamburger and Stieltjes moment problems. In our exposition we follow an approach different from that commonly used. There are novelties well complementing the existing theory. Among them are: (a) to emphasize on the geometric interpretation of the indeterminacy conditions; (b) exploit fine properties of the eigenvalues of perturbed symmetric matrices allowing to derive new lower bounds for the smallest eigenvalues of Hankel matrices; these bounds are used for concluding indeterminacy; (c) provide new arguments to confirm classical results; (d) give new numerical illustrations involving commonly used probability distributions.

Question 2. (uniqueness).Is µ the only measure with the moments {m k } ∞ k=0 ?If the answer to Question 2 is 'yes', we say that µ is moment determinate, or, that µ is uniquely determined by its moments.Otherwise, if the answer is 'no', µ is moment indeterminate, or, µ is not determined uniquely by its moments.In such a case, available is the following deep and non-trivial result, see Berg-Christensen (1981): General result.Suppose that µ is a measure with finite all moments of positive integer order.If µ is non-unique, then there are infinitely many measures of any kind, discrete, absolutely continuous or singular, all with the same moments as µ.
There is a long, rich and amazingly interesting history which originates in works by P.L. Chebyshev (1832-1894) and A. A. Markov (1856A. Markov ( -1922)).The systematic development of the moment problem is due to T.J. Stieltjes (1856Stieltjes ( -1894)), see his memoir Stieltjes (1894).In this remarkable work he was the first to show that the answer to Question 2 can be 'no' by describing explicitly different measures on R + sharing the same moments.
The answers to the above questions depend on both, the support U of µ and the moment sequence {m k } ∞ k=0 .Adopted in the literature are the following names: Hausdorff moment problem, if U is bounded; Stieltjes moment problem, if U is unbounded and U ⊂ R + , and Hamburger moment problem, if U is unbounded and U ⊂ R.
If U = R, we say that {m k } ∞ k=0 is a Hamburger moment sequence, while for U = R + , {m k } ∞ k=0 is a Stieltjes moment sequence.We do not deal with the Hausdorff moment problem since any such a measure, when exists, is uniquely determined by its moments.
In this paper we use standard terminology and notations generally accepted in analysis in works on moment problems.Telling that {m k } ∞ k=0 is a moment sequence always means that there is a measure µ 'behind', i.e., there exists µ which produces these moments.
We need a few words for terminology clarity: If we are given a measure µ with finite moments, then µ produces its only one moment sequence {m k } ∞ k=0 .It is the measure µ which is either determinate or indeterminate, hence no reasons to stick 'determinate' or 'indeterminate' to the moment sequence.
We are interested in the (in)determinacy property of a measure µ with unbounded support and finite all moments {m k } ∞ k=0 .Notice, in general, a Stieltjes moment sequence {m k } ∞ k=0 can also be considered as a Hamburger moment sequence.One important, not intuitive and nontrivial fact is the possibility for a measure µ with finite moments {m k } to be determinate in Stieltjes sense and indeterminate in Hamburger sense.We refer, e.g., to Shohat-Tamarkin (1943), p. 75, Berg-Valent (1994), p. 165, Schmüdgen (2017), p. 183; see also Theorem 3.2 given below in Section 3.
The basic Hankel matrix H n is based on all moments m 0 , m 1 , . . ., while H n,p , for p = 1, 2, 3, 4, the 'shifted' Hankel matrices are formed as follows: H n,1 = (m i+j+1 ) n i,j=0 is based on the 'shifted' moment sequence {m 1 , m 2 , . ..} which is generated by the measure µ 1 with dµ 1 = xdµ; H n,2 = (m i+j+2 ) n i,j=0 is based on the 'shifted' moment sequence {m 2 , m 3 , . ..} generated by the measure µ 2 with dµ 2 = x 2 dµ; similarly for H n,3 and H n,4 .For the determinants of Hankel matrices we use the following notations: In this paper we use c 0 , c 1 , c 2 to denote positive constants which depend on some fixed moments, but we omit them as explicit arguments.For simplicity, if the moments m p+1 , . . ., m p+2n are fixed and we allow m p , the moment preceding m p+1 , to 'vary', then instead of the full notations H n,p (m p , m p+1 , . . ., m p+2n ) and D n,p (m p , m p+1 , . . ., m p+2n ), we write H n,p (m p ) and D n,p (m p ).
Recall a fact from Shohat-Tamarkin (1943), Theorems 1.2 -1.3, which is related to our Question 1: A sequence of real numbers {m k } ∞ k=0 is the moment sequence of a measure µ with support on R (Hamburger case) if and only if all Hankel matrices {H n } are nonnegative definite, which is equivalent to the non-negativity of all determinants {D n }.If the support of µ is R + (Stieltjes case), we have a similar statement, however now we need the two sequences of Hankel matrices, {H n } and {H n,1 }, to be non-negative definite, or, in terms of their determinants, that D n ≥ 0 and D n,1 ≥ 0, n = 0, 1, 2, . . .
In both cases, Hamburger and Stieltjes, if for some n 0 we have D n 0 = 0, then D n = 0 for all n > n 0 and µ is a discrete measure concentrated on a finite number of points.We also say that µ has a finite spectrum, i.e the support is bounded, which implies that the measure µ is unique.
We are interested in the phenomenon 'indeterminacy' ('nonuniqueness') of the measure µ.This means that the support of µ has to be unbounded and its support not to be reducible to a finite set of points.This is why we deal with Hankel matrices which are strictly positive definite, equivalently, their determinants are strictly positive, D n > 0, D n,1 > 0, n = 0, 1, 2, . . . .Since D n , as well D n,1 , are positive, then all eigenvalues of the Hankel matrices H n and H n,p are positive.We use the notation λ k (H n,p ), k = 1, 2, . . ., n + 1, for the kth largest eigenvalue of H n,p .Thus It is well-known, see, e.g., Berg-Chen-Ismail (2002), that the smallest eigenvalue of the matrix H n,p is given by the Rayleigh relation and that the positive numerical sequence {λ 1 (H n,p )} ∞ n=1 is decreasing in n.The smallest eigenvalues of Hankel matrices are fundamentally involved when studying (in)determinacy of measures; see the historical paper by Hamburger (1920) and the more recent work by Chen-Lawrence (1999), Berg-Chen-Ismail (2002) and Berg-Szwarc (2011).
To mention, in the case of indeterminacy the inverse Hankel matrices are involved.Indeed, Berg-Chen-Ismail (2002) obtained the lower bound for the smallest eigenvalue in terms of the trace of the inverse matrix (see their eqns.1.10, 1.11, 1.12), starting from the fact that 1/λ min of the Hankel matrix is equal to λ max of the inverse of that Hankel matrix.

Two classical results
For a very long time available in the literature were classical results expressed in terms of the smallest eigenvalues of Hankel matrices; see, e.g., Hamburger (1920).A remarkable progress was made only in more recent times by Berg-Thill (1991) and Berg-Chen-Ismail (2002).These authors proved fundamental results which can be summarized as follows (the letter 'H' stands for Hamburger, 'S' stands for Stieltjes): Classical Result H.In the Hamburger moment problem, the measure µ is uniquely determined by its moments {m k } ∞ k=0 if and only if the sequence of the smallest eigenvalues of the basic Hankel matrices {H n } ∞ n=1 converges to zero as n → ∞ : lim n→∞ λ 1 (H n ) = 0. Equivalently, µ is indeterminate by its moments {m k } ∞ k=0 if and only if the sequence of the smallest eigenvalues of the basic Hankel matrices {H n } ∞ n=1 converges to a strictly positive number as n → ∞ : lim n→∞ λ 1 (H n ) = c 0 , c 0 > 0.
Classical Result S. In the Stieltjes moment problem, the measure µ is non-uniquely determined by its moments {m k } ∞ k=0 if and only if the sequences of the smallest eigenvalues of the basic Hankel matrices {H n } ∞ n=1 and of the shifted Hankel matrices {H n,1 } ∞ n=1 both converge to strictly positive numbers: Equivalently, the measure µ is determinate by its moments {m k } ∞ k=0 if and only if at least one of the sequences of the smallest eigenvalues of the basic Hankel matrices {H n } ∞ n=1 and of the shifted Hankel matrices {H n,1 } ∞ n=1 converges to zero as n → ∞.Available in the literature are equivalent variations of the formulations of the above results.The proofs, however, may rely on different ideas and techniques.

About the novelties in our approach
Crucial in our approach is to exploit the following : • The geometric interpretation of the indeterminacy conditions as developed by Merkes-Wetzel (1976).
• Properties of the eigenvalues of perturbed symmetric matrices in the spirit of Golub-Van Loan (1996) and Wilkinson (1995).
Both these are among the novelties in our exposition.They are properly used and combined with results from Shohat-Tamarkin (1943), Akhiezer (1965) and Schmüdgen (2017) and a frequent referring to Berg-Chen-Ismail (2002) or Berg-Thill (1991).Going this way we arrive at a unified presentation of classical results in both Hamburger and Stieltjes cases.
As far as we are aware, there is no work, until now, giving such a presentation of most significant classical results on moment problems based on ideas and techniques similar to those used in this paper.We found a little strange that the paper by Merkes-Wetzel (1976) was somehow neglected for a long time.It is not in the list of references in papers and books written by leading specialists on the moment problem.The only proper citation and comments are given by Wulfsohn (2006).In our opinion the geometric interpretation of the indeterminacy conditions has a value by its own, it is fresh and convincing, and deserves attention.The idea is quite simple.Based on the complete moment sequence {m k } ∞ k=0 we build up the so-called parabolic limit region in the plane, and then we look at the position of the point (m 0 , m 1 ).All depends on where this point is located: inside or outside of the region, or on its boundary.Later on we give details and clear graphical illustrations.
We exploit intensively several properties of perturbed symmetric matrices, which allows to derive new lower bound used to conclude the indeterminacy property.Our bound is comparable with the lower bound derived in Berg-Chen-Ismail (2002) by using orthogonal polynomials.
We provide a little different arguments, based on Krein-Nudelman (1977) for concluding the determinacy property.

Moment determinacy in Probability theory
It is worth mentioning that there are results which are of the sort "if and only if ". Usually they are compactly formulated, fundamental in their content, and mathematically beautiful.However, such results are difficult to prove and the conditions involved are practically impossible to check, hence the name 'uncheckable conditions'.
If one assumes that m 0 = 1, i.e., that the total mass is µ(U) = 1, then µ is a probability measure.Well-known is the important role played by the moments in Probability and Statistics, and especially in their applications.This is why a special attention has been paid over a century on finding another sort of 'relatively easier' conditions which are only sufficient or only necessary for either determinacy or indeterminacy of a probability distribution.Nowadays, a variety of 'checkable conditions' (Cramér, Hardy, Carleman, Krein) are available in the literature.The checkable conditions have their analytical value and are more than useful in several applied areas, see, e.g., Janssen-Mirbabayi-Zograf (2021).
The paper by Lin ( 2017) is a rich and valuable source of information on classical and recent results on moment determinacy of probability distributions; see also Stoyanov-Lin-Kopanov (2020).The present paper is intrinsically related to another subsequent paper which is in preparation, see Lin-Stoyanov (2023).

Structure of this paper
The rest of the paper is organized as follows.In Section 2 we treat the Hamburger case and discuss conditions for (in)determinacy.Based on the geometric interpretation of the indeterminacy conditions, we re-derive in a different way already known results by Berg-Chen-Ismail (2002).In Section 3 we follow the same line of reasoning and establish results in the Stieltjes case announced in Berg-Thill (1991).In both cases we provide necessary and sufficient determinacy conditions in terms of the asymptotic behavior, as n → ∞, of two sequences of smallest eigenvalues, namely λ 1 (H n ) and λ 1 (H n,1 ).We also provide a new lower bound for λ 1 (H n,1 ) which is related to the indeterminacy of the measure involved.Section 4 presents details on the smallest eigenvalues and their lower bounds calculated in different ways.The numerical illustrations involve commonly used probability distributions.Brief concluding comments are given in Section 5.

Hamburger moment problem
Before moving further, we discuss some known tools and results which will be used in the sequel.

Limit parabolic region
From Shohat-Tamarkin (1943), p. 5 or Akhiezer (1965), p. 30, we know what {m k } ∞ k=0 is the moment sequence of a measure µ with support R if and only if D n ≥ 0 for all n = 0, 1, 2, . ... In such a case an H-sequence is called positive non-negative.We are going to deal with the interesting case of strictly positive H-sequences, in which case all D n > 0.
Suppose now that {m k } ∞ k=0 is a moment sequence for which we 'keep fixed' the moments {m 2 , m 3 , . ..}, while we treat as 'varying continuously' the moments m 0 and m 1 .If letting m 0 = x, m 1 = y, the whole moment sequence can be written as {x, y, m 2 , m 3 , . ..}.The numbers x and y, i.e., the moments preceding m 2 , can not be arbitrary.Moreover, if for p = 1, 2, 3, 4, the moments m p+1 , m p+2 , . . .are fixed, there is always a range for the possible values of the 'previous' moment m p , the moment just before m p+1 .E.g., the following two-sided bound for m p holds: p+1) .Here the second relation is the Lyapunov's inequality.The lower bound a − p,n , i.e., the smallest possible value of m p , is the unique number such that D n,p (a − p,n , m p+1 , . . ., m p+2n ) = 0. We now turn to an H-moment sequence and the geometric interpretation of the indeterminacy conditions for the corresponding measure.Following Merkes-Wetzel (1976), we fix n and consider the moments (m 0 , m 1 , m 2 , . . ., m 2n ) .We keep 'untouched' m 2 , . . ., m 2n , and assume that m 0 = x and m 1 = y are 'varying continuously'.For each n the relation D n (x, y) ≥ 0 defines a closed convex region, P n := {(x, y) ∈ R 2 : D n (x, y) ≥ 0}, which is bounded by a proper parabola with horizontal axis and vertex in the right-half plane.Since for n = 1, 2, . . ., the Hankel matrices H n (x, y) are positive semidefinite, the regions P n are nested, P n ⊂ P n−1 .Of interest is their intersection P := ∞ n=1 P n , called a limit parabolic region.One possibility is that P is a 'proper' closed region in the right-half plane such that P is bounded by a proper parabola and containing the initially given moments (m 0 , m 1 ).The other possibility is P to degenerate to just a ray, as explained below.

Determinacy criteria and their geometric meaning
In this subsection we provide a geometric interpretation of the determinacy criterion in the Hamburger case.The next two results, Theorems 1 and 2, play a fundamental role.
Theorem 1. (Shohat-Tamarkin (1943), Theorem 2.18).Let µ be a measure associated with the strictly positive definite Hamburger moment sequence {m k } ∞ k=0 .Then µ is determinate (unique) if and only if at least one of the two sequences of ratios has a limit zero as n → ∞, i.e., either lim n→∞ (Merkes-Wetzel (1976), Theorem 1).Let {x, y, m k } ∞ k=2 be a Hamburger moment sequence for the measure µ.Then µ is indeterminate if and only if the point (x, y) is an interior point of the limit parabolic region P.
Theorem 1 has an interesting and enlightening geometric meaning.In fact, the following relations hold: Here a − 0,n comes from the equation 2,n−1 := 1/L n is equal to the length of the 'latus rectum' of the bounding parabola; see Merkes-Wetzel (1976), Lemma 1.The sequence of positive numbers {L n } ∞ n=1 is nondecreasing in n.Then, as n → ∞ and L n → ∞, the length of the latus rectum 1/L n tends to zero and the limit parabolic region P becomes a ray; see Figure 1 (the red bold line).
There are two possibilities for the measure µ with the moments {m k }.Case 1. (The measure µ is H-indeterminate).From Theorem 2 we have that It follows that the limit parabolic region is bounded by a non-degenerate parabola and, notice, the point (m 0 , m 1 ) is interior for the limit parabolic region P, Figure 1.Now, from Theorem 2 the following result promptly comes: Since the point (m 0 , m 1 ) is on the boundary of the limit parabolic region, we have lim n→∞ D n = 0 and hence lim n→∞ λ 1 (H n ) = 0. Consequently we arrive at the following statement: Corollary 1.If the Hamburger moment sequence {m k } ∞ k=0 generates a measure µ which is determinate, then the monotonic decreasing sequence {λ 1 (H n )} ∞ n=1 has limit zero as n → ∞, i.e., lim n→∞ λ 1 (H n ) = 0.

Eigenvalues of perturbed symmetric matrices
Recall first that Berg-Chen-Ismail (2002) use orthonormal polynomials to find a lower bound for the smallest eigenvalue λ 1 (H n ), see their eq.(1.14).If taking large n and the fact that the limit lim n→∞ λ 1 (H n ) is strictly positive, they conclude the indeterminacy of the measure associated with a Hamburger moment sequence.
We are going now to derive properties of the smallest eigenvalues of a family of 'perturbed' symmetric matrices.This allows to derive a lower bound for lim n→∞ λ 1 (H n ), which is a little different from the quantity found in the above cited paper.This, together with Theorems 1 and 2, leads to an alternative proof of Theorem 1.1 in Berg-Chen-Ismail (2002).Moreover, this approach can and will be followed to study the Stieltjes case; see Section 3.
So, suppose that we have the H-moment sequence {m k } ∞ k=0 and let m 0 > a − 0,n for some fixed n.For a 'small' number ε ≥ 0 and e 1 = (1, 0, . . ., 0) being the unit fundamental row-vector in R n+1 we define the following (n + 1) × (n + 1) matrix: which is a symmetric, non-negative matrix of rank 1.This E will play the role of a 'perturbation matrix' of a Hankel matrix.Consider now the (n + 1) × (n + 1) matrices Here H n (m 0 ) is the Hankel matrix based on the moments m 0 , m 1 , . . ., m 2n , while the matrix As a consequence of the simple form of E, the eigenvalue λ k (H n (m 0 )) of the matrix H n (m 0 ) and the eigenvalue λ k (H n (m 0 + ε)) of the perturbed matrix H n (m 0 + ε) are closely related to one other as follows (see, e.g., Wilkinson (1965), formula (41.8), p. 94-98 and Golub-Van Loan (1996), Theorem 8.1.8,p. 397): where Now we suitably specify (2.1) by using the assumption that m 0 > a − 0,n and by setting ε := m 0 − a − 0,n .Hence (2.1) and (2.2) become (2.4) Applying Lagrange's theorem to the right-hand side of (2.3) on the interval [a − 0,n , m 0 ] we obtain that for some number θ n (m 0 ) ∈ (a − 0,n , m 0 ) one holds (2.5) Eq. (2.5) compared with (2.3) yields Let us list some consequences from (2.5).First, all eigenvalues λ k (H n (m 0 )) are monotonically increasing with respect to m 0 ; since 0 ≤ b k (θ n (m 0 ); n) ≤ 1, (2.1) implies that λ k (H n (m 0 )) < m 0 − a − 0,n , which means that each λ k (H n (m 0 )), k = 1, ..., n, is shifted by an amount which lies between zero and the positive number ε = m 0 − a − 0,n .Second, combining (2.5) and (2.4) and taking m 0 → ∞, for k = 1, . . ., n, we see that Third, from the relations , coming from Theorem 1 and their geometric meaning, one has Differentiating both sides with respect to m 0 , we find (2.6) Combining (2.6) with (2.5) shows that one holds b 1 (θ n (m 0 ); n) > b 1 (m 0 ; n).Hence we conclude that the function b 1 (m 0 ; n) is monotonic decreasing as m 0 increases, which means that λ 1 (H n (m 0 )) is a concave function.Summarizing our findings above show that in this H-indeterminate case we deal with two strictly positive quantities, namely m 0 − a − 0,∞ > 0 and λ 1 (H n ) > 0. We use these facts to prove the following statement.
Lemma 1. Assume that m 0 > a − 0,∞ .Then, as n → ∞, the sequence {b 1 (m 0 ; n)} ∞ n=1 admits a positive limit, denoted by b 1 (m 0 ; ∞), in the sense that for a suitable number m0 > m 0 , we will have Proof.We observe first that b 1 (m 0 ; ∞) cannot be directly drawn by taking the limit in (2.6), although each of the quantities (m 0 − a − 0,n ) and λ 1 (H n ) has a limit.Our arguments are given in the next three steps.
Step 3. Combining Step 1 and Step 2 we see that there exists a number m0 > m 0 such that θ ∞ ( m0 ) = m 0 .
From (2.5) with m 0 = m0 then it follows that This means that the number b 1 (m 0 ; ∞) indeed exists.□ From (2.6) we find the required lower bound for the eigenvalue λ 1 (H n ): which was one of our goals.With (2.7) in hands we are in a position to formulate and confirm the validity of the following well-known result, Theorem 1.1 in Berg-Chen-Ismail (2002).Notice, we are arriving at this result in a different way.Hamburger (1920) involves the basic matrices {H n } ∞ n=1 and the shifted matrices {H n,2 } ∞ n=1 .The proof in Berg-Chen-Ismail (2002), as well our proof above, do not rely on results in Hamburger (1920).
We can say a little more.Recall the following known chain of relations from Akhiezer (1965), pp.84-85: , where p n (x) is the nth orthonormal polynomial for the measure µ.Then, if the function ρ(x) = 1/ ∞ n=0 p 2 n (x) ̸ = 0 for all x ∈ R, the measure µ is indeterminate.It remains to see that the lower bound in (2.7) is then replaced by .
In a sense, this lower bound is similar to the bound found by Berg-Chen-Ismail (2002), eqs.
(1.14)-(1.15).In both cases the conclusion drawn is of course the same.

Stieltjes moment problem
In this case we develop a procedure which is similar to that followed in the Hamburger case, with some specifics.We rely essentially on two known results, Theorems 4 and 5.
Theorem 4. (Merkes-Wetzel (1976) D n−1,3 = 0.It is useful to mention that Theorem 4 has the following geometric meaning: , where the number a − 1,n is the unique solution of the equation D n,1 (a − 1,n ) = 0.The numerical sequence {a − 1,n } ∞ n=1 is monotonic nondecreasing and, as n → ∞, convergent to a limit, say a − 1,∞ , where a − 1,∞ ≤ m 1 .Theorem 5. (Merkes-Wetzel (1976), Theorem 2) The strictly positive definite Stieltjes moment sequence {m k } ∞ k=0 generates an indeterminate measure µ if and only if two conditions are satisfied: (i) the point (m 0 , m 1 ) is interior for the limit parabolic region P; (ii) a − 1,∞ < m 1 .We recall that every Stieltjes sequence can also be considered as a Hamburger sequence; see, e.g., Chihara (1968).Hence, the existence of the limit parabolic region P is assured and defined by the two relations, D n (x, y) ≥ 0 and D n,1 (y) ≥ 0. Her we use the notations x = m 0 and y = m 1 .Either P = {(x, y) : x ≥ a − 0,∞ , y = a − 1,∞ } is a ray, or P is the intersection of proper limit parabolic regions in the half plain {y ≥ a − 1,∞ }.Consider the shifted Hankel matrix H n,1 (m 1 ), its smallest eigenvalue λ 1 (H n,1 (m 1 )) and the perturbation matrix E, the same as previously defined in the Hamburger case.First, we want to show that in the S-indeterminate case the estimate of λ 1 (H n (m 0 )) as m 0 varies and the estimate of λ 1 (H n,1 (m 1 )) as m 1 varies are equivalent procedures, just replace D n and m 0 with D n,1 and m 1 .Now we need a result, which is a criterion for S-indeterminacy: Suppose that µ and µ 1 are measures associated with the moment sequence {m k } ∞ k=0 and the shifted moment sequence {m k+1 } ∞ k=0 , respectively.Then µ is S-indeterminate if and only if both µ and µ 1 are H-indeterminate.
This statement is from Krein-Nudelman (1977), p. 199, P.6.8., where it is left as an exercise to the readers.For the sake of completeness we include here the proof.
Indeed, assume that µ is S-indeterminate.This implies the H-indeterminacy of µ and S-indeterminacy of µ 1 .The latter yields H-indeterminacy of µ 1 .If assuming that both µ and µ 1 are H-indeterminate, we use Theorem 1 (formulated for determinate measures).Thus we have two limiting relations, Dn D n−1,2 → c 0 and D n,1 D n−1,3 → c 1 , where c 0 > 0 and c 1 > 0. By Theorem 4 we conclude that µ is S-indeterminate.
We can use Theorem 2 and describe alternatively the S-indeterminacy and also the H-indeterminacy in geometric terms.For this purpose we introduce two limit parabolic regions, P H in the Hamburger case, and P S in the Stieltjes case.With the convention x = m 0 , y = m 1 , z = m 2 , we define: As before, µ is a measure corresponding to the Stieltjes moment sequences {m k } ∞ k=0 .We have the following transparent interpretation: The measure µ is S-indeterminate, hence also H-indeterminate, if and only if two conditions are satisfied: (i) the point (m 0 , m 1 ) is interior for the region P H ; (ii) the point (m 1 , m 2 ) is interior for the region P S .

Since a −
1,∞ < m 1 , we refer to (2.7) and write down the following lower bound of the smallest eigenvalue λ 1 (H n,1 ) of the shifted Hankel matrix H n,1 : Notice, this bound is related to the determinacy of the measure µ 1 with dµ 1 = x dµ.
Let us summarize the above findings: If a measure is S-indeterminate, it is also Hindeterminate.However, an S-determinate measure can be either H-determinate or Hindeterminate.Thus we have the cases, briefly discussed bellow.
One possibility to construct a measure which is S-determinate and H-indeterminate is to start with a moment sequence {m k } ∞ k=0 associated with S-indeterminate measure, hence this moment sequence corresponds also to H-indeterminate measure.Then the idea is to modify this sequence and get another one, { mk } ∞ k=0 , associated with an S-determinate measure which is H-indeterminate.Such a specific construction is given in Schümdgen ( 2017 Case 3. (µ is S-determinate and H-determinate).For a given moment sequence, Sdeterminacy means that there is only one measure with support [0, ∞).Regarding Hdeterminacy, Corollary 1 provides an exhaustive answer.
Clearly, important is the value of the limit lim n→∞ λ 1 (H n,1 ).Combining Theorems 1, 2, 3 and 4, we find that there are four possible limit parabolic regions, (a) -(d), and they are all feasible.Below are the details.
(a) From Case 1 (S-indeterminate and H-indeterminate), we deal with a moment sequence {m k } ∞ k=0 whose associated discrete measure, say ν, has a mass at 0 (this comes from the H-indeterminacy condition a − 0,∞ < m 0 ).Next, consider the measure ν related to ν via the relation ν = ν − ν({0})δ 0 .The moments sequence { mk } ∞ k=0 of the measure ν differs from {m k } ∞ k=0 only at the zero-th entry (see Berg-Christensen, (1981), Theorem 7, p. 111).Hence ν has mass zero at 0, so that ã− 0,∞ = m0 .From Theorems 1 and 3 and their geometric meaning we conclude for ν both properties, H-determinacy and S-determinacy.The limit parabolic region is nondegenerate with ( m0 , m1 ) on its boundary and ã− 1,∞ < m1 , so that one holds lim n→∞ λ 1 (H n,1 ) = c 1 > 0. See Figure 4. (b) Here we have a nondegenerate limit parabolic region such that the point (m 0 , m 1 ) is on its boundary (which implies H-determinacy) and that a − 1,∞ = m 1 .The latter implies S-determinacy and then lim n→∞ λ 1 (H n,1 ) = 0.In this case the unique measure, say μ, is related to the measure µ, involved in the above Case 2 (S-determinate and H-indeterminate) by the relation μ = µ − µ({0})δ 0 .By analogy with the previous item (a), the measure μ has a moment sequence { mk } ∞ k=0 which differs from {m k } ∞ k=0 only by the very first entry indexed by 0 (zero).(c) and (d) Here the limit parabolic regions are rays so that from the relation a − 1,∞ = m 1 it follows that lim n→∞ λ 1 (H n,1 ) = 0. Hence in both, (c) and (d), we have that a − 0,∞ < m 0 .Graphically, see the red lines in Figure 3 and in Figure 5, respectively.It is interesting to mention that, if having S-determinate and H-determinate, each of the relations lim n→∞ λ 1 (H n,1 ) = 0 and lim n→∞ λ 1 (H n,1 ) = c 1 may occur.
The findings in Cases 1 -3 above can be summarized as follows: then the condition lim n→∞ λ 1 (H n,1 ) = 0 is necessary and sufficient for µ to be S-determinate.• If µ is H-determinate, then lim n→∞ λ 1 (H n ) = 0 is a necessary and sufficient condition for µ to be S-determinate.Notice that each of the relations lim n→∞ λ 1 (H n,1 ) = 0 may lim n→∞ λ 1 (H n,1 ) = c 1 > 0 may occur.
The arguments used in Cases 1, 2, 3 above can be alternatively expressed in terms of the smallest eigenvalues of Hankel matrices thus arriving at a result which is equivalent to the known result in Berg-Thill (1991), Proposition 2.3.Theorem 6.A Stieltjes moment sequence {m k } ∞ k=0 corresponds to exactly one measure on the positive real axis if and only if the smallest eigenvalues of either H n or of H n,1 tend to 0, as n → ∞, i.e.
Remark 2. If we do not involve the lower bounds for the smallest eigenvalues of Hankel matrices, Theorem 6 can be easily proved by combining the Krein-Nudelman's statement used above, with the main result in Berg-Chen-Ismail (2002).Indeed, from Krein-Nudelman's result, we have that the measure µ with moments {m k } ∞ k=0 is S-indeterminate if and only if H-indeterminate are both the measure µ with moments {m k } ∞ k=0 and the measure µ 1 with the shifted moments {m k+1 } ∞ k=0 .The opposite statement is: In terms of Theorem 1.1 in Berg-Chen-Ismail (2002), the last statement sounds as follows: We turn now to an important result relating S-determinacy and H-determinacy.Such a result is proved by Schmüdgen (2017), Corollary 8.9, p. 183, in the framework of the operator-theoretic approach and by Heyde (1963), Theorem 1, p. 91, using continued fractions.We give a different short proof involving limit parabolic regions.
Theorem 7. Suppose {m k } ∞ k=0 is a Stieltjes moment sequence associated with the measure µ.If µ is S-determinate with zero mass at zero, µ({0}) = 0, then µ considered on R, is also H-determinate.Proof.Since µ is an S-determinate measure on [0, ∞), there are two options.One is that µ is H-indeterminate.Then, referring to Remark 1 after Case 2 above, µ must have a mass at 0, which is not the case.Thus it remains the second option for µ, namely, that µ is H-determinate.Indeed, if turning to Theorems 1, 2, 3 and 4, we see that the limit parabolic regions in items (a), (b) and (d), see Case 3 above, are compatible with the statement of Theorem 7. Notice, however, the parabolic region (no figure) in item (c) has to be excluded, because of the appearance of a mass µ({0}) > 0, which contradicts the assumption.□ All measures/distributions satisfying the conditions of Theorem 7 are related to shifted Hankel matrices whose smallest eigenvalues may have different limits, as n → ∞, e.g., λ 1 (H n,1 ) → c 1 > 0 or λ 1 (H n,1 ) → 0 Clearly, Theorem 7 can be formulated in other equivalent forms.
It is useful to provide here a result of Heyde (1963), his 'Theorem B', which we paraphrase as follows: Theorem 8. We are given a Stieltjes moment sequence {m 0 = 1, m k } ∞ k=1 and let the associated probability measure µ be S-determinate with no mass at zero: µ({0}) = 0.
Then it is possible that the new moment sequence generates a new distribution, say µ * , which is S-indeterminate.
Proof.Indeed, from Theorem 7, the measure µ with moments {m 0 = 1, m k } ∞ k=1 is Sdeterminate and H-determinate and these properties are in agreement with the conclusions from the limit parabolic regions in items (a), (b) and (d).The assumption µ({0}) > 0 changes the picture.The normalized measure µ * for the new sequence m is compatible with the following subcases: (i)

More on the lower bounds
Let us start with the Hamburger case.It is well-known and we have seen in the previous sections that there are some quantities which are important for deciding whether or not a measure is determinate or indeterminate.The modern technology allows 'easily' to perform computations with good accuracy.Thus, in principle, having computed some quantities, would allow to make either preliminary or definite conclusions.
Thus, the relationship (4.2) between the two lower bounds is established.□ Having (4.2), we can take a limit, as n → ∞, by preserving all relations.We use the fact that the numbers a − 0,n have a limit, as n → ∞, and we used for it the notation a − 0,∞ .The conclusion, after passing to the limit, is as follows: Relations (4.2) and (4.3) can eventually be validated through numerical examples.However first the assumptions (4.1) have to be verified and guaranteed.For practical purposes, we can take n large enough, and follow three steps: Step 1.The quantity ρ 0 =: 1 2π 2π 0 ∞ k=0 |p k (e iθ )| 2 dθ to be replaced by Tr (H −1 n ).
Step 2. The difference m 0 − a − 0,∞ to be replaced by 1/H −1 n (1, 1).This is justified by the fact that m Step 3. The number b 1 (m 0 ; ∞) will also be replaced appropriately.The matrix H n is symmetric and diagonalizable and let us assume that λ 1 (H n ) is a simple eigenvalue with its associated eigenvector v = v(m 0 ; n).We use the notation H n (x) to indicate that the moment m 0 = x can 'vary', so we find that b 1 (m 0 Golub-Van Loan (1996), p. 323).Then for large n, b 1 (m 0 ; ∞) can be replaced by v T Ev v T v .The eigenvector v can be calculated efficiently by using the Inverse Power method.
Remark 3. As we will see below, all numerically computed lower bounds in both Hamburger and Stieltjes cases look tight.In the H-indeterminate case, we have for λ 1 (H n ) a lower bound, a positive constant.This agrees with a result in Berg-Szwarc (2011), Theorem 4.4.These authors proved that, as n → ∞, the kth smallest eigenvalue of H n , λ k (H n (m 0 )), tends rapidly to infinity with k.This means that the two neglected terms n+1

Numerical illustrations
While up to this place of our paper, we have used traditional terminology, notations and arguments from Analysis, now we turn to standard probabilistic terminology and arguments.The main is the same, what is different is clear.
We have chosen two popular and frequently used probability distributions, namely, the Weibull distribution, which includes the exponential distribution, and the Lognormal distribution.Their (in)determinacy properties are well described and available in the literature.The reader can consult, e.g., Lin (2017) or Stoyanov-Lin-Kopanov (2020).Our goal now is to use the lower bounds for the smallest eigenvalues of Hankel matrices and, in a sense, confirm these (in)determinacy properties.We give two examples.Example 1 is related to the content of the paper by Chen-Lawrence (1999) dealing with the weight function w(x) = exp(−x β ), x > 0, β > 0. Notice, after normalizing w, it becomes the density function of the Weibull distribution (also called 'generalized gamma distribution').Example 2 is similar to Example 3.1 in Berg-Chen-Ismail (2002), in which the authors start with the weight function w(x) = x f (x), x > 0, where f is the standard lognormal density.The treatment in these two papers is entirely analytic, no probabilistic notions involved.
Moreover, for any power Y r , r > 0, we easily find the density function, hence the distribution function (the measure), and see that all moments m k (Y r ), k = 1, 2, . . ., are finite; they are expressed via the Euler gamma function.The interesting property is that Y r is determinate for r ∈ [0, 2], and indeterminate for r > 2.
These conclusions can be derived from computed lower bounds of the smallest eigenvalues of Hankel matrices.We can write the matrices H n , H n,1 and compute that if r ∈ (0, 2], then λ 1 (H n ) → 0 and λ 1 (H n,1 ) → 0 for large n.This confirms that indeed Y r is Sdeterminate and also H-determinate.It is not surprising to observe that if r is 'close' to the boundary r = 2, the convergence to zero is quite slow.
It is instructive to make one step more by considering the random variable Z, where Z = Y 3 , for Y ∼ Exp(1); L(Z) = Law(Z).

Brief concluding comments
We exploit the geometric interpretation of indeterminacy conditions and fine properties of the eigenvalues of perturbed symmetric matrices and give a unified exposition of a bunch, or a series of classical results in the problem of moments.Our paper throws an additional light to the phenomena 'uniqueness' and 'nonuniqueness' (determinacy and indeterminacy) of measures in terms of their moments.In general, it is always useful to have in our disposal different approaches, ideas and techniques which lead either to the same final conclusions or allow to establish new results.This enhances the theory and makes it more applicable by providing the freedom to choose and use properly the most appropriate tools.Despite the practical difficulty, in fact the impossibility, to check the determinacy or indeterminacy conditions involving the smallest eigenvalues of Hankel matrices, the results discussed in this paper have been, are and will remain fundamental in mathematics.The numerical illustrations given above indicate that at least some of these results can be adopted and used in applications when dealing with specific measures/probability distributions.

Figure 3 :
Figure 3: Parabolic region for S-det -H-indet case ), Example 8.11, p. 183.It is shown (we do not give details here) how to calculate a proper constant u > 0 and define the new moments by mk = k j=0 k j (−u) j m k−j for k = 0, 1, 2, . . . .Related relevant details can be found in Simon (1998), p. 96, Theorem 3.3.
Theorem 3. Suppose that µ is a measure associated with the Hamburger moment sequence {m k } ∞ k=0 and let λ 1 (H n ) be the smallest eigenvalue of the Hankel matrix H n .Then µ is determinate if and only if lim Proof.In one direction, if we assume that µ is H-determinate, from Corollary 1 we have lim n→∞ λ 1 (H n ) = 0.In the other direction, if µ is H-indeterminate, in view of Lemma 1, λ 1 (H n ) would have a positive lower bound (2.7).□ Most important, this is a criterion for determinacy in the Hamburger moment problem.Remarkable is the fact that involved are only the basic Hankel matrices {H n } ∞ n=1 .Recall that the classical result of µ * is H-indeterminate and S-indeterminate; (ii) µ * is H-indeterminate and S-determinate; (iii) µ * is H-determinate and S-determinate.In fact, subcase (i) proves Theorem B in Heyde (1963).□ It is useful to add a few words.The statement in Heyde's Theorem B means that the shifted Hankel matrix H * n,1 based on the new moment sequence {m * 1, 2, . . ., are finite; c β and m k can be expressed via the Euler gamma function.The (in)determinacy property of X depends on the value of β.It turns out, β = 1 2 is the boundary point: if β ≥ 1 2 , X and its distribution Wei(β) are determinate, while they are indeterminate for any β ∈ (0, 1 2 ).Consider now {m k } ∞ k=0 as a Stieltjes moment sequence and as a Hamburger moment sequence.We want to make the above conclusions by computing the lower bounds of the smallest eigenvalues of the Hankel matrices H n and H n,1 .