Integral over the hypersphere
up vote
4
down vote
favorite
Assume I have a diagonal matrix $L$ of size $n$. I want to compute the following integral:
$$I_n(L) equiv int_{(mathbb{S}^{n-1})^2} mathrm{d}sigma(x) mathrm{d}sigma(x') exp[n x^top L , x']$$
In which $mathbb{S}^{n-1}$ is the unit sphere in $mathbb{R}^n$, and $ mathrm{d}sigma$ is the usual measure on $mathbb{S}^{n-1}$, with $int_{mathbb{S}^{n-1}} mathrm{d}sigma(x) = 2 pi^{n/2}/Gamma(n/2)$.
I am looking for either a general form of $I_n(L)$ or simply an asymptotic of $frac{1}{n}log I_n(L)$ as $n to infty$.
I tried but could not find such results. Any help would be appreciated. Thanks !
integration gaussian-integral
add a comment |
up vote
4
down vote
favorite
Assume I have a diagonal matrix $L$ of size $n$. I want to compute the following integral:
$$I_n(L) equiv int_{(mathbb{S}^{n-1})^2} mathrm{d}sigma(x) mathrm{d}sigma(x') exp[n x^top L , x']$$
In which $mathbb{S}^{n-1}$ is the unit sphere in $mathbb{R}^n$, and $ mathrm{d}sigma$ is the usual measure on $mathbb{S}^{n-1}$, with $int_{mathbb{S}^{n-1}} mathrm{d}sigma(x) = 2 pi^{n/2}/Gamma(n/2)$.
I am looking for either a general form of $I_n(L)$ or simply an asymptotic of $frac{1}{n}log I_n(L)$ as $n to infty$.
I tried but could not find such results. Any help would be appreciated. Thanks !
integration gaussian-integral
Looks a bit hairy. The only case I can do is that of $n=2$ when the modified Bessel function $I_0$ pops out. Then again, I suck at integrals.
– Jyrki Lahtonen
Nov 12 at 10:09
Maybe brute force using en.wikipedia.org/wiki/N-sphere#Spherical_coordinates :D
– Stockfish
Nov 12 at 10:09
I didn't manage brute force with hyper-spherical coordinates...
– seamp
Nov 12 at 10:20
The integral is invariant to rotation, and so you may have some luck changing variables. If you take a look in the book "Multivariate Statistics: A vector space approach" by Morris L Eaton. It gives some really nice invariance theorems, and proofs about the Wishart distribution. It might take a little bit of work.
– Daniel Beale
Nov 12 at 11:00
Does the dimension $n$ of the matrix $L$ is the same as the parameter $n$ in the exponential ? In which case, an asymptotic of $frac1{n}log I_n(L)$ would depend on a sequence of diagonal matrices $L_n$ ?
– user120527
Nov 13 at 11:21
add a comment |
up vote
4
down vote
favorite
up vote
4
down vote
favorite
Assume I have a diagonal matrix $L$ of size $n$. I want to compute the following integral:
$$I_n(L) equiv int_{(mathbb{S}^{n-1})^2} mathrm{d}sigma(x) mathrm{d}sigma(x') exp[n x^top L , x']$$
In which $mathbb{S}^{n-1}$ is the unit sphere in $mathbb{R}^n$, and $ mathrm{d}sigma$ is the usual measure on $mathbb{S}^{n-1}$, with $int_{mathbb{S}^{n-1}} mathrm{d}sigma(x) = 2 pi^{n/2}/Gamma(n/2)$.
I am looking for either a general form of $I_n(L)$ or simply an asymptotic of $frac{1}{n}log I_n(L)$ as $n to infty$.
I tried but could not find such results. Any help would be appreciated. Thanks !
integration gaussian-integral
Assume I have a diagonal matrix $L$ of size $n$. I want to compute the following integral:
$$I_n(L) equiv int_{(mathbb{S}^{n-1})^2} mathrm{d}sigma(x) mathrm{d}sigma(x') exp[n x^top L , x']$$
In which $mathbb{S}^{n-1}$ is the unit sphere in $mathbb{R}^n$, and $ mathrm{d}sigma$ is the usual measure on $mathbb{S}^{n-1}$, with $int_{mathbb{S}^{n-1}} mathrm{d}sigma(x) = 2 pi^{n/2}/Gamma(n/2)$.
I am looking for either a general form of $I_n(L)$ or simply an asymptotic of $frac{1}{n}log I_n(L)$ as $n to infty$.
I tried but could not find such results. Any help would be appreciated. Thanks !
integration gaussian-integral
integration gaussian-integral
edited Nov 12 at 10:52
asked Nov 9 at 15:20
seamp
32812
32812
Looks a bit hairy. The only case I can do is that of $n=2$ when the modified Bessel function $I_0$ pops out. Then again, I suck at integrals.
– Jyrki Lahtonen
Nov 12 at 10:09
Maybe brute force using en.wikipedia.org/wiki/N-sphere#Spherical_coordinates :D
– Stockfish
Nov 12 at 10:09
I didn't manage brute force with hyper-spherical coordinates...
– seamp
Nov 12 at 10:20
The integral is invariant to rotation, and so you may have some luck changing variables. If you take a look in the book "Multivariate Statistics: A vector space approach" by Morris L Eaton. It gives some really nice invariance theorems, and proofs about the Wishart distribution. It might take a little bit of work.
– Daniel Beale
Nov 12 at 11:00
Does the dimension $n$ of the matrix $L$ is the same as the parameter $n$ in the exponential ? In which case, an asymptotic of $frac1{n}log I_n(L)$ would depend on a sequence of diagonal matrices $L_n$ ?
– user120527
Nov 13 at 11:21
add a comment |
Looks a bit hairy. The only case I can do is that of $n=2$ when the modified Bessel function $I_0$ pops out. Then again, I suck at integrals.
– Jyrki Lahtonen
Nov 12 at 10:09
Maybe brute force using en.wikipedia.org/wiki/N-sphere#Spherical_coordinates :D
– Stockfish
Nov 12 at 10:09
I didn't manage brute force with hyper-spherical coordinates...
– seamp
Nov 12 at 10:20
The integral is invariant to rotation, and so you may have some luck changing variables. If you take a look in the book "Multivariate Statistics: A vector space approach" by Morris L Eaton. It gives some really nice invariance theorems, and proofs about the Wishart distribution. It might take a little bit of work.
– Daniel Beale
Nov 12 at 11:00
Does the dimension $n$ of the matrix $L$ is the same as the parameter $n$ in the exponential ? In which case, an asymptotic of $frac1{n}log I_n(L)$ would depend on a sequence of diagonal matrices $L_n$ ?
– user120527
Nov 13 at 11:21
Looks a bit hairy. The only case I can do is that of $n=2$ when the modified Bessel function $I_0$ pops out. Then again, I suck at integrals.
– Jyrki Lahtonen
Nov 12 at 10:09
Looks a bit hairy. The only case I can do is that of $n=2$ when the modified Bessel function $I_0$ pops out. Then again, I suck at integrals.
– Jyrki Lahtonen
Nov 12 at 10:09
Maybe brute force using en.wikipedia.org/wiki/N-sphere#Spherical_coordinates :D
– Stockfish
Nov 12 at 10:09
Maybe brute force using en.wikipedia.org/wiki/N-sphere#Spherical_coordinates :D
– Stockfish
Nov 12 at 10:09
I didn't manage brute force with hyper-spherical coordinates...
– seamp
Nov 12 at 10:20
I didn't manage brute force with hyper-spherical coordinates...
– seamp
Nov 12 at 10:20
The integral is invariant to rotation, and so you may have some luck changing variables. If you take a look in the book "Multivariate Statistics: A vector space approach" by Morris L Eaton. It gives some really nice invariance theorems, and proofs about the Wishart distribution. It might take a little bit of work.
– Daniel Beale
Nov 12 at 11:00
The integral is invariant to rotation, and so you may have some luck changing variables. If you take a look in the book "Multivariate Statistics: A vector space approach" by Morris L Eaton. It gives some really nice invariance theorems, and proofs about the Wishart distribution. It might take a little bit of work.
– Daniel Beale
Nov 12 at 11:00
Does the dimension $n$ of the matrix $L$ is the same as the parameter $n$ in the exponential ? In which case, an asymptotic of $frac1{n}log I_n(L)$ would depend on a sequence of diagonal matrices $L_n$ ?
– user120527
Nov 13 at 11:21
Does the dimension $n$ of the matrix $L$ is the same as the parameter $n$ in the exponential ? In which case, an asymptotic of $frac1{n}log I_n(L)$ would depend on a sequence of diagonal matrices $L_n$ ?
– user120527
Nov 13 at 11:21
add a comment |
1 Answer
1
active
oldest
votes
up vote
1
down vote
accepted
There is a way to get an infinite series using the expansion of $exp$, however, it may be more efficient to compute the result using quadrature. The steps below first expand $exp$, and use the multinomial theorem to get a polynomial expansion for the integrand. Taking the integral inside the sum to the monomials we can evaluate the integral directly to obtain an analytic (closed form) expression for the integral.
For clarity, first denote the integral by $I$, i.e.
begin{align}
I := int_{S_n}int_{S_n}exp(n textbf{x}^T L textbf{y}) dsigma_n dsigma_n
end{align}
Power series for exp
If we first consider the expansion of the function argument in the integral,
begin{align}
exp(n textbf{x}^T L textbf{y}) =& sum_{k=0}^infty frac{(ntextbf{x}^T L textbf{y})^k}{k!} \
=& sum_{k=0}^infty frac{left(n sum_{r=1}^M x_r L_{rr}y_r right)^k}{k!} tag{1}
end{align}
It is well known that the series converges uniformly, and the spherical integral is linear (see [1]), which means we can take the integrals inside the summation,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!} int_{S_n}int_{S_n}left( sum_{r=1}^M x_r L_{rr}y_r right)^kdsigma_n dsigma_n tag{2}
end{align}
Use of the multinomial theorem
Using the multinomial theorem on the inner bracket gives,
begin{align}
left( sum_{r=1}^M x_r L_{rr}y_r right)^k =& sum_{j_{1} + ldots + j_{M} = k} frac{k!}{j_{1}!ldots j_{M}!} prod_{t=1}^M (L_{tt} y_t x_t)^{j_t} tag{3}
end{align}
We can then put (3) into (2) and distribute the integrals further over the addition to give,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k} rho_k int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n tag{4}
end{align}
where $rho_k = frac{k!}{j_{1}!ldots j_{M}!} $.
Integrating a monomial
The spherical integral of a monomial is discussed in detail in [2], but the main result gives,
begin{align}
int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n = left{ begin{array}{rl} left( 2frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_t)} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} & : j_{t} mbox{ all even} \ 0& : mbox{otherwise} end{array} right .tag{5}
end{align}
Where $q_{t} = frac{1}{2} (j_{t} + 1)$. Putting (5) into (4) then gives,
begin{align}
I =& 4 sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k \ forall t mod(j_{t},2) = 0 } rho_k left(frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_{t})} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} tag{6}
end{align}
Simplification
Since the terms are non-zero only when all of the the $j_t$, and hence $k$, are even; after some cancellations, the sum can be written,
begin{align}
I =& 4 sum_{k=0}^infty sum_{j_{1} + ldots + j_{M} = k} frac{n^{2k}}{prod_{t=1}^M (2j_{t})!}left(frac{prod_{t=1}^MGamma(j_{t} + frac{1}{2})}{Gamma(frac{M}{2} + sum_{t=1}^M j_{t})} right)^2 prod_{t=1}^M L_{tt}^{2j_{t}} \
=& 4 sum_{k=0}^infty frac{n^{2k}}{Gamma(frac{M}{2} + k)^2} sum_{j_{1} + ldots + j_{M} = k}prod_{t=1}^M frac{left(Gamma(j_{t} + frac{1}{2}) L_{tt}^{j_{t}}right)^2}{ (2j_{t})!}tag{7}
end{align}
[1] Baker, John A., Integration over spheres and the divergence theorem for balls, Am. Math. Mon. 104, No. 1, 36-47 (1997). ZBL0877.26008.
[2] Folland, Gerald B., How to integrate a polynomial over a sphere, Am. Math. Mon. 108, No. 5, 446-448 (2001). ZBL1046.26503.
Thanks a lot for the very clear and detailed answer !
– seamp
Nov 13 at 14:18
add a comment |
1 Answer
1
active
oldest
votes
1 Answer
1
active
oldest
votes
active
oldest
votes
active
oldest
votes
up vote
1
down vote
accepted
There is a way to get an infinite series using the expansion of $exp$, however, it may be more efficient to compute the result using quadrature. The steps below first expand $exp$, and use the multinomial theorem to get a polynomial expansion for the integrand. Taking the integral inside the sum to the monomials we can evaluate the integral directly to obtain an analytic (closed form) expression for the integral.
For clarity, first denote the integral by $I$, i.e.
begin{align}
I := int_{S_n}int_{S_n}exp(n textbf{x}^T L textbf{y}) dsigma_n dsigma_n
end{align}
Power series for exp
If we first consider the expansion of the function argument in the integral,
begin{align}
exp(n textbf{x}^T L textbf{y}) =& sum_{k=0}^infty frac{(ntextbf{x}^T L textbf{y})^k}{k!} \
=& sum_{k=0}^infty frac{left(n sum_{r=1}^M x_r L_{rr}y_r right)^k}{k!} tag{1}
end{align}
It is well known that the series converges uniformly, and the spherical integral is linear (see [1]), which means we can take the integrals inside the summation,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!} int_{S_n}int_{S_n}left( sum_{r=1}^M x_r L_{rr}y_r right)^kdsigma_n dsigma_n tag{2}
end{align}
Use of the multinomial theorem
Using the multinomial theorem on the inner bracket gives,
begin{align}
left( sum_{r=1}^M x_r L_{rr}y_r right)^k =& sum_{j_{1} + ldots + j_{M} = k} frac{k!}{j_{1}!ldots j_{M}!} prod_{t=1}^M (L_{tt} y_t x_t)^{j_t} tag{3}
end{align}
We can then put (3) into (2) and distribute the integrals further over the addition to give,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k} rho_k int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n tag{4}
end{align}
where $rho_k = frac{k!}{j_{1}!ldots j_{M}!} $.
Integrating a monomial
The spherical integral of a monomial is discussed in detail in [2], but the main result gives,
begin{align}
int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n = left{ begin{array}{rl} left( 2frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_t)} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} & : j_{t} mbox{ all even} \ 0& : mbox{otherwise} end{array} right .tag{5}
end{align}
Where $q_{t} = frac{1}{2} (j_{t} + 1)$. Putting (5) into (4) then gives,
begin{align}
I =& 4 sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k \ forall t mod(j_{t},2) = 0 } rho_k left(frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_{t})} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} tag{6}
end{align}
Simplification
Since the terms are non-zero only when all of the the $j_t$, and hence $k$, are even; after some cancellations, the sum can be written,
begin{align}
I =& 4 sum_{k=0}^infty sum_{j_{1} + ldots + j_{M} = k} frac{n^{2k}}{prod_{t=1}^M (2j_{t})!}left(frac{prod_{t=1}^MGamma(j_{t} + frac{1}{2})}{Gamma(frac{M}{2} + sum_{t=1}^M j_{t})} right)^2 prod_{t=1}^M L_{tt}^{2j_{t}} \
=& 4 sum_{k=0}^infty frac{n^{2k}}{Gamma(frac{M}{2} + k)^2} sum_{j_{1} + ldots + j_{M} = k}prod_{t=1}^M frac{left(Gamma(j_{t} + frac{1}{2}) L_{tt}^{j_{t}}right)^2}{ (2j_{t})!}tag{7}
end{align}
[1] Baker, John A., Integration over spheres and the divergence theorem for balls, Am. Math. Mon. 104, No. 1, 36-47 (1997). ZBL0877.26008.
[2] Folland, Gerald B., How to integrate a polynomial over a sphere, Am. Math. Mon. 108, No. 5, 446-448 (2001). ZBL1046.26503.
Thanks a lot for the very clear and detailed answer !
– seamp
Nov 13 at 14:18
add a comment |
up vote
1
down vote
accepted
There is a way to get an infinite series using the expansion of $exp$, however, it may be more efficient to compute the result using quadrature. The steps below first expand $exp$, and use the multinomial theorem to get a polynomial expansion for the integrand. Taking the integral inside the sum to the monomials we can evaluate the integral directly to obtain an analytic (closed form) expression for the integral.
For clarity, first denote the integral by $I$, i.e.
begin{align}
I := int_{S_n}int_{S_n}exp(n textbf{x}^T L textbf{y}) dsigma_n dsigma_n
end{align}
Power series for exp
If we first consider the expansion of the function argument in the integral,
begin{align}
exp(n textbf{x}^T L textbf{y}) =& sum_{k=0}^infty frac{(ntextbf{x}^T L textbf{y})^k}{k!} \
=& sum_{k=0}^infty frac{left(n sum_{r=1}^M x_r L_{rr}y_r right)^k}{k!} tag{1}
end{align}
It is well known that the series converges uniformly, and the spherical integral is linear (see [1]), which means we can take the integrals inside the summation,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!} int_{S_n}int_{S_n}left( sum_{r=1}^M x_r L_{rr}y_r right)^kdsigma_n dsigma_n tag{2}
end{align}
Use of the multinomial theorem
Using the multinomial theorem on the inner bracket gives,
begin{align}
left( sum_{r=1}^M x_r L_{rr}y_r right)^k =& sum_{j_{1} + ldots + j_{M} = k} frac{k!}{j_{1}!ldots j_{M}!} prod_{t=1}^M (L_{tt} y_t x_t)^{j_t} tag{3}
end{align}
We can then put (3) into (2) and distribute the integrals further over the addition to give,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k} rho_k int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n tag{4}
end{align}
where $rho_k = frac{k!}{j_{1}!ldots j_{M}!} $.
Integrating a monomial
The spherical integral of a monomial is discussed in detail in [2], but the main result gives,
begin{align}
int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n = left{ begin{array}{rl} left( 2frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_t)} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} & : j_{t} mbox{ all even} \ 0& : mbox{otherwise} end{array} right .tag{5}
end{align}
Where $q_{t} = frac{1}{2} (j_{t} + 1)$. Putting (5) into (4) then gives,
begin{align}
I =& 4 sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k \ forall t mod(j_{t},2) = 0 } rho_k left(frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_{t})} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} tag{6}
end{align}
Simplification
Since the terms are non-zero only when all of the the $j_t$, and hence $k$, are even; after some cancellations, the sum can be written,
begin{align}
I =& 4 sum_{k=0}^infty sum_{j_{1} + ldots + j_{M} = k} frac{n^{2k}}{prod_{t=1}^M (2j_{t})!}left(frac{prod_{t=1}^MGamma(j_{t} + frac{1}{2})}{Gamma(frac{M}{2} + sum_{t=1}^M j_{t})} right)^2 prod_{t=1}^M L_{tt}^{2j_{t}} \
=& 4 sum_{k=0}^infty frac{n^{2k}}{Gamma(frac{M}{2} + k)^2} sum_{j_{1} + ldots + j_{M} = k}prod_{t=1}^M frac{left(Gamma(j_{t} + frac{1}{2}) L_{tt}^{j_{t}}right)^2}{ (2j_{t})!}tag{7}
end{align}
[1] Baker, John A., Integration over spheres and the divergence theorem for balls, Am. Math. Mon. 104, No. 1, 36-47 (1997). ZBL0877.26008.
[2] Folland, Gerald B., How to integrate a polynomial over a sphere, Am. Math. Mon. 108, No. 5, 446-448 (2001). ZBL1046.26503.
Thanks a lot for the very clear and detailed answer !
– seamp
Nov 13 at 14:18
add a comment |
up vote
1
down vote
accepted
up vote
1
down vote
accepted
There is a way to get an infinite series using the expansion of $exp$, however, it may be more efficient to compute the result using quadrature. The steps below first expand $exp$, and use the multinomial theorem to get a polynomial expansion for the integrand. Taking the integral inside the sum to the monomials we can evaluate the integral directly to obtain an analytic (closed form) expression for the integral.
For clarity, first denote the integral by $I$, i.e.
begin{align}
I := int_{S_n}int_{S_n}exp(n textbf{x}^T L textbf{y}) dsigma_n dsigma_n
end{align}
Power series for exp
If we first consider the expansion of the function argument in the integral,
begin{align}
exp(n textbf{x}^T L textbf{y}) =& sum_{k=0}^infty frac{(ntextbf{x}^T L textbf{y})^k}{k!} \
=& sum_{k=0}^infty frac{left(n sum_{r=1}^M x_r L_{rr}y_r right)^k}{k!} tag{1}
end{align}
It is well known that the series converges uniformly, and the spherical integral is linear (see [1]), which means we can take the integrals inside the summation,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!} int_{S_n}int_{S_n}left( sum_{r=1}^M x_r L_{rr}y_r right)^kdsigma_n dsigma_n tag{2}
end{align}
Use of the multinomial theorem
Using the multinomial theorem on the inner bracket gives,
begin{align}
left( sum_{r=1}^M x_r L_{rr}y_r right)^k =& sum_{j_{1} + ldots + j_{M} = k} frac{k!}{j_{1}!ldots j_{M}!} prod_{t=1}^M (L_{tt} y_t x_t)^{j_t} tag{3}
end{align}
We can then put (3) into (2) and distribute the integrals further over the addition to give,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k} rho_k int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n tag{4}
end{align}
where $rho_k = frac{k!}{j_{1}!ldots j_{M}!} $.
Integrating a monomial
The spherical integral of a monomial is discussed in detail in [2], but the main result gives,
begin{align}
int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n = left{ begin{array}{rl} left( 2frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_t)} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} & : j_{t} mbox{ all even} \ 0& : mbox{otherwise} end{array} right .tag{5}
end{align}
Where $q_{t} = frac{1}{2} (j_{t} + 1)$. Putting (5) into (4) then gives,
begin{align}
I =& 4 sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k \ forall t mod(j_{t},2) = 0 } rho_k left(frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_{t})} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} tag{6}
end{align}
Simplification
Since the terms are non-zero only when all of the the $j_t$, and hence $k$, are even; after some cancellations, the sum can be written,
begin{align}
I =& 4 sum_{k=0}^infty sum_{j_{1} + ldots + j_{M} = k} frac{n^{2k}}{prod_{t=1}^M (2j_{t})!}left(frac{prod_{t=1}^MGamma(j_{t} + frac{1}{2})}{Gamma(frac{M}{2} + sum_{t=1}^M j_{t})} right)^2 prod_{t=1}^M L_{tt}^{2j_{t}} \
=& 4 sum_{k=0}^infty frac{n^{2k}}{Gamma(frac{M}{2} + k)^2} sum_{j_{1} + ldots + j_{M} = k}prod_{t=1}^M frac{left(Gamma(j_{t} + frac{1}{2}) L_{tt}^{j_{t}}right)^2}{ (2j_{t})!}tag{7}
end{align}
[1] Baker, John A., Integration over spheres and the divergence theorem for balls, Am. Math. Mon. 104, No. 1, 36-47 (1997). ZBL0877.26008.
[2] Folland, Gerald B., How to integrate a polynomial over a sphere, Am. Math. Mon. 108, No. 5, 446-448 (2001). ZBL1046.26503.
There is a way to get an infinite series using the expansion of $exp$, however, it may be more efficient to compute the result using quadrature. The steps below first expand $exp$, and use the multinomial theorem to get a polynomial expansion for the integrand. Taking the integral inside the sum to the monomials we can evaluate the integral directly to obtain an analytic (closed form) expression for the integral.
For clarity, first denote the integral by $I$, i.e.
begin{align}
I := int_{S_n}int_{S_n}exp(n textbf{x}^T L textbf{y}) dsigma_n dsigma_n
end{align}
Power series for exp
If we first consider the expansion of the function argument in the integral,
begin{align}
exp(n textbf{x}^T L textbf{y}) =& sum_{k=0}^infty frac{(ntextbf{x}^T L textbf{y})^k}{k!} \
=& sum_{k=0}^infty frac{left(n sum_{r=1}^M x_r L_{rr}y_r right)^k}{k!} tag{1}
end{align}
It is well known that the series converges uniformly, and the spherical integral is linear (see [1]), which means we can take the integrals inside the summation,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!} int_{S_n}int_{S_n}left( sum_{r=1}^M x_r L_{rr}y_r right)^kdsigma_n dsigma_n tag{2}
end{align}
Use of the multinomial theorem
Using the multinomial theorem on the inner bracket gives,
begin{align}
left( sum_{r=1}^M x_r L_{rr}y_r right)^k =& sum_{j_{1} + ldots + j_{M} = k} frac{k!}{j_{1}!ldots j_{M}!} prod_{t=1}^M (L_{tt} y_t x_t)^{j_t} tag{3}
end{align}
We can then put (3) into (2) and distribute the integrals further over the addition to give,
begin{align}
I = sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k} rho_k int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n tag{4}
end{align}
where $rho_k = frac{k!}{j_{1}!ldots j_{M}!} $.
Integrating a monomial
The spherical integral of a monomial is discussed in detail in [2], but the main result gives,
begin{align}
int_{S_n}int_{S_n} prod_{t=1}^M (L_{tt} y_t x_t)^{j_{t}}dsigma_n dsigma_n = left{ begin{array}{rl} left( 2frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_t)} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} & : j_{t} mbox{ all even} \ 0& : mbox{otherwise} end{array} right .tag{5}
end{align}
Where $q_{t} = frac{1}{2} (j_{t} + 1)$. Putting (5) into (4) then gives,
begin{align}
I =& 4 sum_{k=0}^infty frac{n^k}{k!}sum_{j_{1} + ldots + j_{M} = k \ forall t mod(j_{t},2) = 0 } rho_k left(frac{prod_{t=1}^MGamma(q_{t})}{Gamma(sum_{t=1}^M q_{t})} right)^2 prod_{t=1}^M L_{tt}^{j_{t}} tag{6}
end{align}
Simplification
Since the terms are non-zero only when all of the the $j_t$, and hence $k$, are even; after some cancellations, the sum can be written,
begin{align}
I =& 4 sum_{k=0}^infty sum_{j_{1} + ldots + j_{M} = k} frac{n^{2k}}{prod_{t=1}^M (2j_{t})!}left(frac{prod_{t=1}^MGamma(j_{t} + frac{1}{2})}{Gamma(frac{M}{2} + sum_{t=1}^M j_{t})} right)^2 prod_{t=1}^M L_{tt}^{2j_{t}} \
=& 4 sum_{k=0}^infty frac{n^{2k}}{Gamma(frac{M}{2} + k)^2} sum_{j_{1} + ldots + j_{M} = k}prod_{t=1}^M frac{left(Gamma(j_{t} + frac{1}{2}) L_{tt}^{j_{t}}right)^2}{ (2j_{t})!}tag{7}
end{align}
[1] Baker, John A., Integration over spheres and the divergence theorem for balls, Am. Math. Mon. 104, No. 1, 36-47 (1997). ZBL0877.26008.
[2] Folland, Gerald B., How to integrate a polynomial over a sphere, Am. Math. Mon. 108, No. 5, 446-448 (2001). ZBL1046.26503.
edited Nov 13 at 11:12
answered Nov 13 at 9:29
Daniel Beale
66629
66629
Thanks a lot for the very clear and detailed answer !
– seamp
Nov 13 at 14:18
add a comment |
Thanks a lot for the very clear and detailed answer !
– seamp
Nov 13 at 14:18
Thanks a lot for the very clear and detailed answer !
– seamp
Nov 13 at 14:18
Thanks a lot for the very clear and detailed answer !
– seamp
Nov 13 at 14:18
add a comment |
Sign up or log in
StackExchange.ready(function () {
StackExchange.helpers.onClickDraftSave('#login-link');
});
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
StackExchange.ready(
function () {
StackExchange.openid.initPostLogin('.new-post-login', 'https%3a%2f%2fmath.stackexchange.com%2fquestions%2f2991496%2fintegral-over-the-hypersphere%23new-answer', 'question_page');
}
);
Post as a guest
Required, but never shown
Sign up or log in
StackExchange.ready(function () {
StackExchange.helpers.onClickDraftSave('#login-link');
});
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
Sign up or log in
StackExchange.ready(function () {
StackExchange.helpers.onClickDraftSave('#login-link');
});
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
Sign up or log in
StackExchange.ready(function () {
StackExchange.helpers.onClickDraftSave('#login-link');
});
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Sign up using Google
Sign up using Facebook
Sign up using Email and Password
Post as a guest
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Required, but never shown
Looks a bit hairy. The only case I can do is that of $n=2$ when the modified Bessel function $I_0$ pops out. Then again, I suck at integrals.
– Jyrki Lahtonen
Nov 12 at 10:09
Maybe brute force using en.wikipedia.org/wiki/N-sphere#Spherical_coordinates :D
– Stockfish
Nov 12 at 10:09
I didn't manage brute force with hyper-spherical coordinates...
– seamp
Nov 12 at 10:20
The integral is invariant to rotation, and so you may have some luck changing variables. If you take a look in the book "Multivariate Statistics: A vector space approach" by Morris L Eaton. It gives some really nice invariance theorems, and proofs about the Wishart distribution. It might take a little bit of work.
– Daniel Beale
Nov 12 at 11:00
Does the dimension $n$ of the matrix $L$ is the same as the parameter $n$ in the exponential ? In which case, an asymptotic of $frac1{n}log I_n(L)$ would depend on a sequence of diagonal matrices $L_n$ ?
– user120527
Nov 13 at 11:21