π(PI) and the AGM: Evaluating Elliptic Integrals contd.

3 comments

Complete Elliptic Integrals of Second Kind

After getting familiar with the AGM sequences and Landen Transformation, it is time to apply these concepts to evaluate elliptic integrals. Here we are going to focus on elliptic integrals of the second kind. To be more specific we are going to deal with J(a,b) defined by
J(a,b)=π/20a2cos2θ+b2sin2θdθ
Our strategy (well actually Landen's and Legendre's) here will be to analyze the defining integral under the Landen transformation
tan(ϕθ)=batanθ
In fact we will study the effect of a sequence of these transformations applied to the defining integral for J(a,b). Setting the stage with the following initial values
a0=a,b0=b,c1=ab2
and
an+1=an+bn2,bn+1=anbn,cn+1=anbn2
tan(θn+1θn)=bnantanθn
Δn=a2ncos2θn+b2nsin2θn
(note that unlike sequences {an},{bn},{cn}, the sequence {θn} is a sequence of variable amplitudes and not of constant real numbers) we are going to establish the following in order
  1. Δn+1+cn+1cosθn+1=Δn
  2. Δn+1cn+1cosθn+1=anbn/Δn
  3. 2Δn+1=Δn+anbnΔn
  4. dθnΔn=12dθn+1Δn+1
  5. ΔnΔn+1=Δ2n+anbn2
  6. ΔndθnΔn+1dθn+1=cn+1cosθn+1dθn+1anbnΔndθn

Initialization

And now come the laborious manipulations which must be appreciated for the fact they lead to such simple looking results in spite of their inherent complexity. The starting point here is to express θn+1 in terms of θn and calculate the derivative dθn+1/dθn as follows
tan(θn+1θn)=bnantanθn an(tanθn+1tanθn)=bntanθn(1+tanθn+1tanθn) (anbntan2θn)tanθn+1=(an+bn)tanθn tanθn+1=(an+bn)tanθnanbntan2θn cosθn+1=anbntan2θn(anbntan2θn)2+(an+bn)2tan2θn=anbntan2θna2n+b2ntan4θn+(a2n+b2n)tan2θn=anbntan2θn(1+tan2θn)(a2n+b2ntan2θn)=anbntan2θnsec2θna2ncos2θn+b2nsin2θn=ancos2θnbnsin2θnΔn Similarly
sinθn+1=(an+bn)sinθncosθnΔn Also we have (from the defining equation for θn+1)
dθn+1dθn=(an+bn)cos2θn+1(anbntan2θn)sec2θn+2bntan2θnsec2θn(anbntan2θn)2=(an+bn)cos2θn+1an+bntan2θncos2θn(anbntan2θn)2=(an+bn)cos2θn+1ancos2θn+bnsin2θn(ancos2θnbnsin2θn)2 dθn+1dθn=(an+bn)(ancos2θn+bnsin2θn)Δ2n Now that the preliminary setup is done we can prove the desired results
Proof for 1) Δn+1+cn+1cosθn+1=Δn We have
Δn+1=a2n+1cos2θn+1+b2n+1sin2θn+1=(an+bn)2cos2θn+1+4anbnsin2θn+12=(an+bn)(ancos2θnbnsin2θn)2+4anbnsin2θncos2θn2Δn=(an+bn)(ancos2θn+bnsin2θn)22Δn=(an+bn)(ancos2θn+bnsin2θn)2Δn Therefore (from definition of cn+1 and value of cosθn+1)
Δn+1+cn+1cosθn+1=(an+bn)(ancos2θn+bnsin2θn)2Δn+(anbn)(ancos2θnbnsin2θn)2Δn=2(a2ncos2θn+b2nsin2θn)2Δn=Δ2nΔn=Δn Proof for 2) Δn+1cn+1cosθn+1=anbn/Δn Similarly
Δn+1cn+1cosθn+1=(an+bn)(ancos2θn+bnsin2θn)2Δn(anbn)(ancos2θnbnsin2θn)2Δn=2anbn(cos2θn+sin2θn)2Δn=anbnΔn Proof for 3) is obtained from adding 1) and 2).
Proof for 4) dθnΔn=12dθn+1Δn+1 First we note that
dΔndθn=(b2na2n)sinθncosθnΔn=c2nsinθncosθnΔn Differentiating 3) with respect to θn we get
2dΔn+1dθn+1dθn+1dθn=dΔndθnanbnΔ2ndΔndθn 2dΔn+1dθn+1dθn+1=(dΔndθnanbnΔ2ndΔndθn)dθn 2c2n+1sinθn+1cosθn+1Δn+1dθn+1=c2nsinθncosθnΔn(Δ2nanbnΔ2n)dθn Now putting the values of cosθn+1 and sinθn+1 and noting that
cn+1=anbn2,c2n=a2nb2n Δ2nanbn=Δ2nanbn(cos2θn+sin2θn)=(anbn)(ancos2θnbnsin2θn) we get
(anbn)2(an+bn)(ancos2θnbnsin2θn)sinθncosθn2Δ2ndθn+1Δn+1 =(a2nb2n)sinθncosθn(anbn)(ancos2θnbnsin2θn)Δ2ndθnΔn 12dθn+1Δn+1=dθnΔn Proof for 5) ΔnΔn+1=Δ2n+anbn2 We have
Δ2n+anbn=Δ2n+anbn(cos2θn+sin2θn)=(an+bn)(ancos2θn+bnsin2θn)=Δ2ndθn+1dθn=Δ2n2Δn+1Δn=2ΔnΔn+1 (from (4) above)
Proof for 6) ΔndθnΔn+1dθn+1=cn+1cosθn+1dθn+1anbnΔndθn The result in 5) can be rewritten as
Δn=Δ2n+anbn2Δn+1 Replacing Δn+1 using 4) we get
Δn=Δ2n+anbnΔndθndθn+1 Δndθn+1=(Δn+anbnΔn)dθn Replacing Δn on L.H.S. using 1) we get
Δn+1dθn+1+cn+1cosθn+1dθn+1=Δndθn+anbnΔndθn ΔndθnΔn+1dθn+1=cn+1cosθn+1dθn+1anbnΔndθn
Having established these results it may still not be clear to the reader as to the overall goal of proving these results. The last result gives some indication about the path to reach the goal. It tries to estimate the change in the value of the quantity Δdθ as we apply the Landen transformation. Upon integrating these differentials we can find the change in the integral defining J(a,b) effected due to the Landen transformation. And we hope to add these changes involved in the series of Landen transformations to finally get the value of J(a,b).

It turns out that the series so obtained does not converge (because of the simple fact that Δn does not tend to zero as n). So we modify our strategy a little bit and instead study the changes in the quantity (Δna2n/Δn) (this does tend to zero as n).

Armed with this insight we can see that
a2ndθnΔna2n+1dθn+1Δn+1=(a2n2a2n+1)dθnΔn=(a2nb2n2anbn)dθnΔn Thus
a2ndθnΔna2n+1dθn+1Δn+1=12c2ndθnΔnanbnΔndθn Using this result with 6) we get
(Δna2nΔn)dθn(Δn+1a2n+1Δn+1)dθn+1=cn+1cosθn+1dθn+112c2ndθnΔn=cn+1cosθn+1dθn+1122nc2ndθoΔo (using (4) to get from dθn to dθo)
Putting n=0,1,2, and on adding and integrating the result we get (here we use the fact that (Δna2n/Δn) tends to zero as n)
θ00Δ0dθa20θ00dθΔ0=c1sinθ1+c2sinθ2+c3sinθ3+ 12(c20+2c21+4c22+8c23+)θ00dθΔ0 Putting θ0=π/2 and noting that θ1=π,θ2=2π (and in general θn=2n1π) we get
J(a,b)=(a212n=02nc2n)I(a,b)
and finally the elliptic integral of second kinds gets expressed in terms of AGM and the associated sequence.

Formula for J(a, b): An Alternative Approach

The above proof of the formula for J(a,b) is somewhat cumbersome and indirect. By studying a closely related integral L(a,b) defined by
L(a,b)=(a2b2)π/20sin2θdθa2cos2θ+b2sin2θ
one can arrive at the formula for J(a,b) established earlier.

As can be easily seen, this integral is related to J(a,b) and I(a,b) as follows
L(a,b)=a2I(a,b)J(a,b)
The idea here is to find a relation between L(an,bn) and L(an+1,bn+1). To simplify writing it is sufficient to consider the relationship between L(a,b) and L(a1,b1). We start with the substitution
t2=a2cos2θ+b2sin2θ
so that
tdt/dθ=(b2a2)sinθcosθ=(a2t2)(t2b2)
and therefore the integral gets transformed into
L(a,b)=aba2t2t2b2dt
Now the substitution x=(1/2)(t+ab/t) leads us to
a21x2=(a2t2)(t2b2)/4t2
x2b21=(t2ab)2/4t2
dx/dt=(t2ab)/2t2
so that
dt(a2t2)(t2b2)=±12dx(a21x2)(x2b21)
The ambiguous sign on the right can be resolved if we observe that when t increases from b to b1, x decreases from a1 to b1 (so that we select the - sign here) and when t increases from b1 to a, x increases from b1 to a1 (so that we select the + sign here).

Also note that we have
t=x±x2b21
which leads us to
L(a,b)=ab(a2t2)(a2t2)(t2b2)dt=abϕ(t)(a2t2)(t2b2)dt=12a1b1ψ(x)(a21x2)(x2b21)dx where \phi(t) = a^{2} - t^{2} and
\begin{aligned}\psi(x) &= \phi(x + \sqrt{x^{2} - b_{1}^{2}}) + \phi(x - \sqrt{x^{2} - b_{1}^{2}})\\ &= a^{2} - (x + \sqrt{x^{2} - ab})^{2} + a^{2} - (x - \sqrt{x^{2} - ab})^{2}\\ &= 2a^{2} + 2ab - 4x^{2} = a^{2} - b^{2} + (a^{2} + 2ab + b^{2}) - 4x^{2}\\ &= (a^{2} - b^{2}) + 4(a_{1}^{2} - x^{2})\end{aligned} We thus get
L(a, b) = \frac{a^{2} - b^{2}}{2}\int_{b_{1}}^{a_{1}}\frac{dx}{\sqrt{(a_{1}^{2} - x^{2})(x^{2} - b_{1}^{2})}} + 2\int_{b_{1}}^{a_{1}}\sqrt{\frac{a_{1}^{2} - x^{2}}{x^{2} - b_{1}^{2}}}\,\,dx
The first integral on the right hand side is I(a_{1}, b_{1}) which is the same as I(a, b) (use the substitution x^{2} = a_{1}^{2}\cos^{2}\theta + b_{1}^{2}\sin^{2}\theta) so we have the recursive relation for L(a,b) as follows
L(a, b) = (1 / 2)(a^{2} - b^{2})I(a, b) + 2L(a_{1}, b_{1})
Dividing both sides by I(a, b) we get
\frac{L(a_{0}, b_{0})}{I(a, b)} = \frac{1}{2}c_{0}^{2} + 2\frac{L(a_{1}, b_{1})}{I(a, b)}
and a repeated application of this identity gives us
\frac{L(a, b)}{I(a, b)} = \frac{1}{2}(c_{0}^{2} + 2c_{1}^{2} + \cdots + 2^{n - 1}c_{n - 1}^{2}) + 2^{n}\frac{L(a_{n}, b_{n})}{I(a, b)}
Now 2^{n}L(a_{n}, b_{n}) \rightarrow 0 as n \rightarrow \infty (this is clear from the fact that 2^{n}(a_{n}^{2} - b_{n}^{2}) \rightarrow 0 as n \rightarrow \infty).
Therefore on letting n \rightarrow \infty we get
L(a, b) = \left(\frac{1}{2}\sum \limits_{n = 0}^{\infty}2^{n}c_{n}^{2}\right)I(a, b)
Using the relation between L(a, b) and J(a, b) we get
J(a, b) = \left(a^{2} - \frac{1}{2}\sum\limits_{n = 0}^{\infty}2^{n}c_{n}^{2}\right)I(a, b)
We are now in a position to state and prove the Brent-Salamin AGM formula for \pi and this will be done in the next post.

Print/PDF Version

3 comments :: π(PI) and the AGM: Evaluating Elliptic Integrals contd.

Post a Comment

  1. Hi there,

    I really like this blog, I’ve found it fascinating. I was wondering if you could tell me where you found the alternative proof above?

    Thanks.

  2. Hi Bob,
    I read the alternate proof in some paper online. It’s an old entry so I don’t remember the source now. Sorry.

  3. Hi, Paramanand,
    I found the limits in this post interesting.

    From the Landen transformation section:

    \theta_{n+1} = \arctan(\frac{b_n}{a_n} tan(\theta_n)) + \theta_n

    Near \pi / 2 , the \tan() approaches \infty . Therefore, it doesn't matter much what b_n and a_n are, the \arctan() is still \pi / 2 , and thus \theta_{n+1} = 2 \theta_n .

    Later on, it took me a while to see the double sign reversal with \psi(x) .

    As noted above, corresponding to t going up from b \to a , x goes down a_1 \to b_1 and then up again from b_1 \to a_1 . This is reflected in the 2 versions of \psi(x) .
    Since the overall integral is stated as \int_{b_1}^{a_1} , the result of the first part of the integral range above must be sign reversed. This is the 2nd \psi(x) in the expansion above. This corresponds to the part of the integral where the bottom \sqrt{..} \space is negative, resulting in a positive \psi(x) component. The other \psi(x) has no sign reversals.

    Thanks for the http://math.stackexchange.com/questions/ask hint for previewing. It was very helpful for checking the formatting. I hope it worked ok.

    Thanks, a very informative section. --gjk