Abstract
The defect of a function is defined as the difference between the measure of the positive and negative regions. In this paper, we begin the analysis of the distribution of defect of random Gaussian spherical harmonics. By an easy argument, the defect is non-trivial only for even degree and the expected value always vanishes. Our principal result is evaluating the defect variance, asymptotically in the high-frequency limit. As other geometric functionals of random eigenfunctions, the defect may be used as a tool to probe the statistical properties of spherical random fields, a topic of great interest for modern cosmological data analysis.
Export citation and abstract BibTeX RIS
1. Introduction
In recent years, a lot of interest has been drawn by the analysis of the geometric features of random eigenfunctions for the spherical Laplacian. More precisely, let us denote as usual by {Ylm( · )}m = −l, ..., l, for l = 1, 2, ..., the set of (real-valued) spherical harmonics, i.e. the array of real-valued functions satisfying
where denotes the spherical Laplacian (see for instance [16, 18]). The random model we shall focus on is
where the coefficients {alm} are independent standard Gaussian with zero-mean and unit variance.
The random field fl is isotropic, meaning that for any and , the distribution of the random vector {fl(x1), ..., fl(xl)} is invariant under rotations, i.e. equal to the distribution of {fl(g · x1), ..., fl(g · xl)} for any rotation g ∈ SO(3) on the sphere. Also, fl is centred Gaussian, with the covariance function
where Pl are the usual Legendre polynomials defined by Rodrigues' formula
and d(x, y) is the (spherical) geodesic distance between x and y. As is well known, Legendre polynomials are orthogonal w.r.t. the constant weight ω(t) ≡ 1 on [ − 1, 1], see section 5 for more discussion and details.
1.1. Background
Random eigenfunctions for the spherical Laplacian naturally emerge in a number of different physical contexts. A particularly active area is related to the analysis of isotropic spherical random fields on the sphere, as motivated for instance by the analysis of cosmic microwave background radiation (CMB), see for instance [4, 5] and the references therein. Under these circumstances, eigenfunctions like fl represent the (normalized) Fourier components of the field, i.e. the following orthogonal expansion holds, in the L2 sense:
where cl ⩾ 0 is a deterministic sequence (depending on the angular power spectrum) which encodes the full correlation structure of f.
In the CMB literature, f(x), and hence the components fl, are actually observed by highly sophisticated satellite experiments such as WMAP by NASA and Planck by ESA. It is then common practice to analyse geometric functionals of the observed CMB radiation to constraint the statistical properties of the underlying fields, e.g. to test for isotropy and/or Gaussianity. For instance, the three Minkowski functionals, providing the area, the boundary length and the topological genus of excursion sets over a given level have been applied on CMB data by a huge number of authors, including [7, 8, 12]. Much effort has been spent in analysing unexpected features, for instance, the so-called cold spot (see [2, 3] and the references therein), whose statistical significance can also be evaluated by means of local curvature properties of isotropic Gaussian fields [6].
In this work, we shall focus on one of the most important geometric functionals, namely the defect. The defect (or 'signed area', see [1]) of a function is defined as
Here is the Heaviside function,
where is the usual indicator function of the set A and dx is the Lebesgue measure. The defect is hence the difference between the areas of positive and negative inverse image of ψ, respectively.
The primary focus of this paper is the distribution of the defect of the random Gaussian fields fl. Note that for odd l, fl is always odd, so that in this case the defect vanishes identically (), and therefore has the non-trivial distribution for even l only. A possible alternative to avoid trivialities is to restrict ourselves to a subset of the sphere, the most natural choice being a hemisphere, i.e. we may choose any hemisphere and define
in this paper we study for even l only.
1.2. Statement of the main result
Evaluating the expectation of is trivial. Indeed, note that integration over is exchangeable with expectation, so that
and vanishes for every , by the symmetry of the Gaussian distribution. We just established the following lemma.
Lemma 1.1. For every l = 1, 2, ..., we have
The main result of this paper concerns the asymptotic behaviour of the defect variance:
Theorem 1.2. As l → ∞ along even integers, the defect variance is asymptotic to
where C > 0 is a positive constant.
The constant C in (5) may be expressed in terms of the infinite (conditionally convergent) integral
See section 2 for some details on the integral on the RHS of (6). We do not know whether one can evaluate C explicitly; however, we shall be able to show that
(see lemma 5.5).
One should compare the statement of theorem 1.2 to empirical results of the study conducted by Blum, Gnutzmann and Smilansky [1]. The authors of that work studied the defect (or, as they refer to it, the 'signed area') of random monochromatic waves on various planar domains. For the particular case of a unit circle, they found that the order of magnitude of defect variance is consistent to (5) per unit area, with leading constant evaluated numerically as ≈0.0386. The corresponding defect variance per unit area in our situation is
with the leading constant3
1.3. Previous work
To put our results in a proper perspective and explain the technical difficulties to be handled, we need to briefly recall some results from [11]. In that work we studied the asymptotic behaviour of excursion sets
of fl, as l → ∞. As we will recall below, a rather peculiar phenomenon can be shown to hold for z ≠ 0; namely, the asymptotic distribution of the area of these excursion sets is fully degenerate over z, i.e. it corresponds to a Gaussian random variable times a deterministic function of z. This result is intuitively due to an asymptotic degeneracy in the excursion set functionals, which turns out to be dominated by a single polynomial (quadratic) term.
This degeneracy, however, does not hold for the special case z = 0. More precisely, in [11] we focused on the asymptotic behaviour of the empirical measure of random spherical harmonics, defined as
. A key step in that paper is the asymptotic expansion
where Hq(.) are standard Hermite polynomials, is the cumulative distribution function of a standard Gaussian variable and the deterministic functions Vq(z) can be explicitly provided in terms of higher order derivatives of Φ(z), Vq(z) = ( − 1)qΦ(q)(z) (see [11] for more discussion and details).
It turns out that, as l → ∞,
converges by distribution to the standard Gaussian, whereas for q ⩾ 3,
moreover, V2(z) = −zϕ(z), where ϕ(z) is the standard Gaussian probability density, clearly does not vanish for all z ≠ 0, whence the asymptotic behaviour of
is easily seen to be Gaussian, uniformly over z. Furthermore, the limiting process is completely degenerate with respect to z, a feature to which we shall come back later.
For z ≠ 0, the asymptotic behaviour of the area functional for the excursion sets is hence fully understood. The previous argument, however, fails for z = 0, as in this case the leading term is null and each summand in the asymptotic expansion (7) becomes relevant. Up to a linear transformation, this is clearly equivalent to the defect functional, indeed
Thus, the case z = 0 is the most challenging from the mathematical point of view, and, at the same time, the most interesting from the point of view of geometric interpretation.
1.4. Overview of the paper
The plan of the paper is as follows. In section 2, we provide the main ideas behind our principal arguments, to help the reader understand the material to follow; in section 3, we discuss the relation of our results to recent works on the distribution of nodal lengths and level curves for random eigenfunctions and related conjectures; section 4 provides the proof of the main results; section 5 contains auxiliary lemmas, we believe of independent interest, on the asymptotic behaviour of moments of Legendre polynomials.
2. On the proof of theorem 1.2
To establish our results, we shall need a detailed analysis of the odd moments of Legendre polynomials
The rationale for this can be explained as follows. It is relatively easy to express (up to a constant) the defect variance as
where sin θdθ is (up to a constant) the uniform measure on the sphere in the spherical coordinates (see lemma 4.1). It then remains to understand the asymptotic behaviour of Il, a task which is put forward in proposition 4.2; here we provide the main ideas underlying its proof. We know from Hilb's asymptotics (lemma 5.1) that Pl(cos θ) has a scaling limit: as l → ∞, for any fixed
we have
In fact, the latter estimate holds uniformly for ψ = o(l), and it may be shown that, as a tail of a conditionally convergent integral, the contribution of the other regime ψ ≳ l is negligible (see the proof of proposition 4.2).
In the following, we neglect the difference between l and . It is then natural to try to replace Pl(cos θ) in (8) with its scaling limit; a formal substitution yields heuristically
where we replaced with , which is justified for ψ = o(l). This is inconsistent with the statement of theorem 1.2; the integral
diverges, so some more care is needed to transform (10). Moreover, the last integrand in (10) is somewhat different from the integrand in integral (6) defining C (neglecting the constant in front of the integral in (6)) in that we subtract J0(ψ) from the arcsine. Note that the factor ψ in both (10) and (6) is reminiscent of the uniform measure sin θ dθ on the sphere.
There is a subtlety that explains the latter discrepancy between the integrand in definition (6) of C and the integrand in (10). One way to justify the substitution of the scaling limit of Pl(cos θ) (i.e. the second step of (10)) is expanding the arcsine in the integral of (8) into the Taylor series around the origin
for some explicitly given coefficients ak (see (18) and the formula immediately after). We then need to evaluate all the odd moments of Pl(cos θ) w.r.t. the measure sin θ dθ on . It turns out that the moments (appropriately scaled) are asymptotically equivalent to the corresponding moments of the Bessel function on w.r.t. the measure ψ dψ, with the exception for the first moment (lemma 5.2). Indeed, Pl(cos θ) integrates to zero, whereas diverges; this phenomenon also accounts for the divergence of (11). To account for this difference, we need to subtract J0(ψ) from the arcsine in integrand (10), thus obtaining a conditionally convergent integral. This is indeed the heuristic explanation for the discrepancy between (10) and (6).
To summarize this discussion, we note the following. The variance is given, up to an explicit constant, by integral (8). The constant C in (6) is reminiscent of Il up to scaling (9): the expression Pl(cos θ) is replaced by its scaling limit J0(ψ), and the measure sin θdθ becomes ψdψ on after scaling. Subtracting J0(ψ) from the integrand in (6) accounts for the fact that for even l, the integral of Pl(cos θ) against sin θdθ vanishes, whereas this is not the case of integral of the scaling limit J0(ψ) against ψdψ.
3. Discussion
As reported in section 1.3, the asymptotics for the variance of the excursion sets is given for z ≠ 0 by
(here ϕ is the Gaussian probability density); for the defect (z = 0), this gives only an upper bound . The principal result of this paper (theorem 1.2) states that the defect variance (5) is of the order of magnitude l−2. Our explanation for this discrepancy is the disappearance, for z = 0, of the quadratic term in the Hermite expansion of the function see [11].
We should compare this situation to the length distribution of the level curves. For , let
be the (random) length of the level curve f−1l(t), the most important case being that of t = 0; the corresponding level curve is called the nodal line. It is known [19] that for , the expected length is
The variance is asymptotic to
for t ≠ 0, whereas for the nodal length,
It is therefore natural to conjecture that the discrepancy of the defect and the empirical measure is somehow related to the discrepancy in the variance of the nodal length. However, no simple explanation, as in the case of the defect, is known to explain the discrepancy between the nodal length and the non-vanishing level curve length.
Recall that for the two random variables X, Y, the correlation is defined as
It is known [19] that for every , become asymptotically fully dependent for large l, in the sense that4
In fact, a much stronger statement regarding the rate of convergence was proven. It then seems that the asymptotic degeneracy of Φl(z) is related to the asymptotic full dependence of the level lengths (12).
A possible explanation for phenomenon (12) is the following conjecture due to Mikhail Sodin. Let and for , let (the 'local length') be the (random) length of the unique component of f−1l(t) that contains x inside (or 0, if fl does not cross the level t).
Conjecture 3.1 (Mikhail Sodin). The local lengths are asymptotically fully dependent in the sense that for every and ,
Heuristically, conjecture 3.1 is stronger than (12), since can be viewed as a summation of over a set of points . More rigorously, we should assume that fl(x) ≠ t; it is immediate to see that this is satisfied almost surely.
4. Proof of theorem 1.2
The first lemma is probably already known, but we failed to locate a direct reference and thus we report it for completeness. The proof is based on standard properties of Gaussian random variables.
Lemma 4.1. For even l, we have
We postpone the proof of lemma 4.1 until the end of this section. Let us denote the integral in (13)
so that evaluating the defect variance is equivalent to evaluating Il, which is done in the following proposition (for even l).
Proposition 4.2. As l → ∞ along even numbers, we have
where
Moreover, the constant C1 is strictly positive.
Proof of theorem 1.2 (assuming proposition 4.2). Formula (13) together with proposition 4.2 yields
The positivity of the constant C = 32πC1 follows directly from the positivity of C1, which is claimed in proposition 4.2.
Proof of proposition 4.2. As was explained in section 2, to extract the asymptotics of Il, we will expand the arcsine on the RHS of (14) into the Taylor series around the origin; we will encounter only the odd moments of Pl(cos θ), due to the arcsine being an odd function. As was pointed out in section 2, the function Pl(cos θ) differs from its scaling limit in that the integral of the former vanishes, whereas the integral of the latter diverges; all the other odd moments of Pl(cos θ) are asymptotic to the (properly scaled) corresponding moments of the Bessel function (see lemma 5.2). To account for this discrepancy, we subtract Pl(cos θ) from the integrand and add it back separately. To this end, we write
so that the vanishing of the integral of the latter
for even l, implies
The advantage of the latter representation (17) over (14) is that the only powers that will appear in the Taylor expansion of the arcsine on the right-hand side of (17) are of order ⩾3, so that the moments of Pl(cos θ) are all identical to the corresponding moments of the scaling limit. Intuitively, this means that we may replace the appearances of Pl(cos θ) in (17) by its scaling limit. The rest of the present proof is a rigorous argument that will establish the latter statement.
Let
where
are the Taylor coefficients of the arcsine. Note that all the terms in expansion (18) are positive, and by the Stirling formula, the coefficients are asymptotic to
for some c > 0, so that, in particular, the Taylor series (18) is uniformly absolutely convergent. Therefore, we may write
We know from lemma 5.2, that for every k ⩾ 1,
with c2k + 1 given by (32); comparing this result to (21), it is natural to expect that
where
We are going to formally prove (23) immediately; however, first we evaluate the constant C2 by summing up the series in (24) and validate that indeed C2 = C1 in (16), as claimed. We plug (32) into (24) to formally compute
where to obtain the third equality we used (18) again. To justify the exchange of the summation and integration order, we consider the finite summation
using (20) and (30) to bound the contribution of tails, and take the limit m → ∞.
We now turn to proving (23). To this end we expand (18) into a finite-degree Taylor polynomial while controlling the tail using lemma 5.7. Indeed, using (20) and (45), we easily obtain
since |Pl(t)| ⩽ 1 for every l and t ∈ [ − 1, 1], so that |Pl(t)|k is monotone decreasing with k; the notation A '≲' B here and everywhere in this paper means that there exists a constant C > 0 (which may vary) so that A ⩽ C · B.
We then have for every m (m = m(l) to be chosen)
and plugging (22) (a direct consequence of lemma 5.2) into (25) finally yields
with
It is clear that (26) implies (23) (recall definition (24) of C2 and note that as m → ∞, C2, m → C2), which concludes the proof of statement (15) of the present proposition.
It then remains to prove the positivity of the constant C1. While its nonnegativity C1 ⩾ 0 is clear since it is the leading constant for the integral I1, which, up to an explicit positive constant, equals the variance of a random variable via (13), the strict positivity is less obvious. To this end, we recall that C1 = C2, the latter being given by (24). Note that all the Taylor coefficients ak of arcsine are positive (see (19)). Hence, the positivity of C2 (and thus also of C1) follows from the second statement of lemma 5.2, which claims that ck ⩾ 0 are nonnegative, and c3 > 0 is explicitly given.
Proof of lemma 4.1. The result of the present lemma is an artefact of the following general fact (see e.g. [13, 14]). Let (X1, X2) be a two-variate centred Gaussian random variable with the covariance matrix
|r| ⩽ 1, and for i = 1, 2, define the random variables
where is the Heaviside function (4). Then Bi are both mean zero with the covariance given by
To justify the latter statement, one may identify the explicit expression for Cov(B1, B2) as a special integral whose value is known. As an alternative, one may expand Cov(B1, B2) into a Taylor series as a function of r around r = 0 via the Hermite expansion for as in [11], and note that the coefficients in the expansion are identical to the Taylor coefficients of the arcsine.
Using definition (3) of the defect, we may exchange the order of taking the expectation and integrating to write
by the isotropic property of the random field fl, where N is the northern pole. Note that for every , fl(N) and fl(x) are jointly Gaussian, centred, with unit variance and covariance equal to
where (θ, ϕ) are the spherical coordinates of x; the latter follows from definition (2) of the covariance function. Therefore, as was explained earlier, for every ,
We then evaluate the latter integral in (27) in the spherical coordinates as
which, taking into account l being even (and thus Pl(t) is also even), is statement (13) of the present lemma.
5. Moments of Legendre polynomials
We start by recalling a basic fact on the asymptotic behaviour of Legendre polynomials (see for instance [15]); as usual, we shall denote by Jν the Bessel function of the first kind.
Lemma 5.1 (Hilb's asymptotics). For any > 0, C > 0, we have
where
uniformly w.r.t. l ⩾ 1, θ ∈ [0, π − ].
Note that we will use lemma 5.1 only for the range , so that we may forget about the altogether. We also recall that (see again [15])
Lemma 5.2. Let j ⩾ 5 or j = 3. Then
where the constants cj are given by
the RHS of (32) being absolutely convergent for j ⩾ 5 and conditionally convergent for j = 3. Moreover, for every j as above, the constants cj ⩾ 0 are nonnegative, and c3 > 0 is positive, given explicitly by
Proof. By Hilb's asymptotics, we have
The contribution of the error term to (34) is (exploiting being bounded)
Now
and using (30), we may bound the second integral in (35) as
Plugging the last couple of estimates into (35) , and finally into (34), we obtain for j ⩾ 3
Therefore, we are to evaluate
where we denote for brevity. The statement of the present lemma is then equivalent to
where cj is defined by (32).
The main idea is that the main contribution comes from the intermediate range 1 ≲ ψ ⩽ L for any > 0; here we may replace the factor
with 1. To make this argument precise, we write for
so that also
Therefore, the integral in (37) is
Note that as l → ∞ (equivalently L → ∞), the main term on the RHS of converges to cj; indeed
so that it remains to bound the error term. Now
and we use (30) to bound the latter as
so that upon plugging the latter into (41) yields
plugging the latter into (39) yields
so that (40) implies (37) for j ⩾ 5, which was equivalent to the statement of the present lemma in this case.
It then remains to prove the result for j = 3, for which we have to work a little harder due to the conditional convergence of integral (32); to treat this technicality we will have to exploit the oscillatory behaviour of the Bessel function, and not only its decay (30). It is well known that
so that
and hence the integral on the RHS of (32) is indeed convergent, by integration by parts.
Now we choose a (large) parameter K ≫ 1 and divide the integration range into [0, K] and ; the main contribution comes from the first term, whence we need to prove that the latter vanishes. Indeed, we use (42) to bound
where we use integration by parts with the bounded function
to bound the contribution leading term in (42); it is easy to bound the contribution of the error term in (42) using the crude estimate (30).
On [0, K], we use (38) to write
the former clearly being convergent to c3. Combining (43) with (44), we obtain
this implies (37) for j = 3 (which is equivalent to the statement of the present lemma for j = 3) upon choosing the parameter K growing to infinity sufficiently slowly (e.g. ). Since j = 3 was the only case that was not covered earlier, this concludes the proof of the moment part of lemma 5.2.
It then remains to prove the nonnegativity statement of cj, and the explicit expression (33) for c3. In fact, both of those statements follow from the computation we performed in our previous paper [11], p 18. We report the relevant results in lemma 5.3 below.
We recall an alternative characterization for moments of Legendre polynomials from [11]. For given positive integers l1, l2, l3, we introduce the so-called Clebsch–Gordan coefficients , which are different from zero if and only if l1, l2, l3 are such that l1 + l2 + l3 is even and li + lj ⩽ lk for all permutations i, j, k = 1, 2, 3. The Clebsch–Gordan coefficients are well known in group representation theory (they intertwine alternative representations for SO(3)) and in the quantum theory of angular momentum; we do not provide more details here, but refer instead to [11] or to standard references such as [16, 17] (see also [9]). The results provided in [11] are as follows.
Lemma 5.3 ([12], appendix A). For all even l, we have
and
Also, for j ⩾ 5,
Remark 5.4. The previous discussion yields the following interesting corollary: as l → ∞,
and for j ⩾ 5,
The following lemma establishes the lower bound for the constant C in theorem 1.2 claimed in section 1.2. Its proof is straightforward.
Lemma 5.5. For C as in (5), we have
Proof. Recall that , where C1 = C2 is given by (24). Since all the aj and ck are nonnegative, lemma 5.5 follows from bounding the first term in the series
where we used a1 = 1/6 and (33).
Remark 5.6. As a final remark, we note that the moments of Pl(cos θ) are themselves of some physical interest. In particular, while analysing the relationship between asymptotic Gaussianity and ergodicity of isotropic spherical random fields, reference [10] considered the random field
where denotes standard spherical harmonics (for m = 0) and g ∈ SO(3) is a uniformly distributed random rotation in It is readily seen that the resulting field is isotropic, and each of its Fourier components at frequency l is marginally distributed as (for any
Up to normalization constants, the moments of the marginal law are exactly those of the Legendre polynomials we established earlier in the paper. More precisely, if we focus (as in [10]) on we obtain immediately, for any
and
whereas [18]
and hence all moments of order q ⩾ 4 diverge.
Note that in lemma 5.2, we worked relatively hard to establish the precise asymptotics for the moments of Legendre polynomials. A much cruder version of the same argument gives a uniform upper bound for the fifth moment of the absolute value of the Legendre polynomials.
Lemma 5.7. We have the following uniform upper bound for the fifth moment of the absolute value of Legendre polynomials
where the constant involved in the `O'-notation is universal.
Proof. The proof follows along the same lines as the beginning of the proof of lemma 5.2 for k ⩾ 5, except that we use the crude upper bound (30) whenever we reach (36), and the boundedness of on .
Acknowledgments
We wish to thank Zeév Rudnick for suggesting the problem, many stimulating and fruitful discussions and useful comments on an earlier version of this paper, and Mikhail Sodin for many stimulating and fruitful discussions. We are grateful to the anonymous referees for many useful comments on how to improve the presentation of the present work. A substantial part of this research was done during the second author's visit to the University of Rome 'Tor Vergata', and he would like to acknowledge the extremely friendly and stimulating environment in the institution and the generous financial support. He was supported by the Knut and Alice Wallenberg Foundation, grant KAW.2005.0098.
Footnotes
- 3
Numerical computation of the constant (6) reveals that it is consistent with Blum–Gnutzmann–Smilansky, taking into account the normalization and natural symmetry of the sphere.
- 4
The asymptotic full dependence (12) was proven under the technical assumption t1, t2 ≠ 0; it is natural to conjecture, though, that a slight modification of the same argument will work if either of ti vanishes.