Ann.

Henri Poincaré 9 (2008), 625–638

c 2008 Birkhäuser Verlag Basel/Switzerland
1424-0637/040625-14, published online May 30, 2008
DOI 10.1007/s00023-008-0367-7 Annales Henri Poincaré

Reflections on Refraction (Geometrical Optics)

Michel Mendès France and Ahmed Sebbar

à Jacques Peyrière

Abstract. Consider an optical system made of an unknown number N of layers

of homogeneous transparent plates with different unknown refraction indices.
Observing beams of monochromatic light through the system, find the num-
ber N of plates together with their respective indices and their thicknesses.
The mathematical analysis of the problem involves the so-called Hadamard
quotient of two power series.
We shall also discuss fractal optical systems consisting of infinitely many
infinitely thin plates. If the index of refraction varies in an erratic way there
may be multiple refraction. These systems could be called “refractals”.
We conclude the paper with independent considerations on a general
system consisting of one plate with continuous varying index n(x, y) ≥ 1. To
determine the function (x, y) → n(x, y) seems to be a difficult problem. Our
contribution to solving it is thus very modest.

1. Determining the index of refraction

Fascinated by the simplicity of geometrical optics and by the Snell–Descartes law
“sin i = n sin r”, we ask the natural question which may be well known to special-
ists. Consider a non-crystalline transparent plate whose unknown index of refrac-
tion n(x, y) ≥ 1 depends on the point (x, y) within the plate. We ignore partial
reflection and absorption. Analyzing the outcoming beams in terms of monochro-
matic incident beams should enable one to recover the function (x, y) → n(x, y).
We shall discuss briefly the general problem using geometry at the very end
of our work. The main part of the paper is concerned with the much simpler
case where the optical system consists of N transparent layers of respective thick-
nesses a1 , a2 , . . . , aN and with respective indices of refraction n1 , n2 , . . . , nN . Our
approach is analytical.
626 M. Mendès France and A. Sebbar Ann. Henri Poincaré

θ
M

O
θ Α x

Figure 1. Ray through plates.

The system is represented in the (x, y) plane so that the front face of the
plate is the y-axis and the opposite face is the vertical line
N

x= aℓ = d .
ℓ=1

An incident ray hits the plate at the origin O at angle θ ∈ [0, π/2[, penetrates
within the system and meets the opposite side at M whose coordinates are x = d,
y = φ(θ). Does the knowledge of θ → φ(θ) determine the number of layers, their
indices and their thicknesses?
In Sections 6 and 7 we study a somewhat limit case. There we are given
the refraction indices and we ask what can be said about φ(θ). We shall discover
that for certain distribution of indices the outcoming ray may well be multiple.
Of course this multiple refraction has nothing to do with the well known double
refraction of calcite.

2. The mathematical formulation

Let θk be the angle of the ray as it penetrates the k th plate. The Snell–Descartes
law asserts that
sin θ = n1 sin θ1 = n2 sin θ2 = . . .
Let M0 = O, M1 , M2 , . . . be the points where the ray hits the plates 1, 2, 3, . . .
The ordinate of these points are
y0 = 0 , y1 = a1 tan θ1 , . . . , yk = yk−1 + ak tan θk , . . .
Vol. 9 (2008) Reflections on Refraction (Geometrical Optics) 627

Mk
θk+1
θk

Mk−1 θk
θk−1

Figure 2. Indices.

Since
sin θk sin θ
tan θk =  =
2
1 − sin θk n2k − sin2 θ
we have
N
 aℓ sin θ
φ(θ) = yN =  . (1)
ℓ=1 n2ℓ − sin2 θ
More generally, should the index vary continuously then
 ∞
sin θ dσ(x)
φ(θ) =  (2)
0 n2 (x) − sin2 θ
where σ is a bounded non-decreasing function. We shall come back to this aspect in
Sections 6 and 7. The above representation is actually general since it encompasses
also the discrete case where σ is then a step function and

dσ = a ℓ δ x ℓ , aℓ > 0
ℓ

where δxℓ is the Dirac measure at xℓ .

The question raised in Section 1 reduces to the problem of expressing
N, n1 , n2 , . . . , nN , a1 , a2 , . . . , aN in terms of φ(θ). For a given φ, if these values
do exist they are not unique since permuting the layers does not change φ (pro-
vided all the indices are finite). Therefore in this case, with no loss of generality
we can always assume 1 ≤ n1 < n2 < · · · < nN and agree that layers of same
index are put together to form a single plate.
628 M. Mendès France and A. Sebbar Ann. Henri Poincaré

3. A functional equation
Put ξ = sin θ so that 0 ≤ ξ < 1. Define Ψ(sin θ) = φ(θ). Equation (1) reads
N − 12
ξ2

 ξ
Ψ(ξ) = aℓ 1− 2 .
nℓ nℓ
ℓ=1

The problem now is to find N, aℓ > 0 and the increasing sequence 1 ≤ n1 < n2 <
· · · < nN in terms of Ψ. This is of course quite trivial. Indeed n1 is the smallest
singularity of the odd analytic function
2k+1
Ψ(ξ) = Σ∞
k=0 f2k+1 ξ

so
1 1
= lim |f2k+1 | 2k+1 .
n1 k→∞
Then
Ψ(ξ)

a1 = lim n21 − ξ 2 .
ξ→n1 ξ
The couple (n2 , a2 ) is obtained in a similar way by considering the function Ψ1
a1 ξ
Ψ1 (ξ) = Ψ(ξ) −  2 .
n1 − ξ 2
At step l
l−1
 aj ξ
Ψl−1 = Ψ(ξ) − 
j=1 n2j − ξ 2
determines nl and al . At some point, ΨN ≡ 0 and the algorithm ends.
It may be interesting to observe that the above problem falls in a general
scheme, namely to discuss the equation
N  
 ξ
F (ξ) = al H (3)
nl
l=1

where F and H are given functions and where N, al , nl are unknown. In our context
we assume that H and therefore F are odd functions holomorphic in a neighbour-
hood of the origin:
 
F (ξ) = f2k+1 ξ 2k+1 , H(ξ) = h2k+1 ξ 2k+1 .
k≥o k≥o

We shall suppose that for all k, h2k+1 = 0. The previous case corresponds to
 1
2 − 21 k −2

H(ξ) = ξ(1 − ξ ) = (−1) ξ 2k+1
k
k≥0

and F = Ψ.
Vol. 9 (2008) Reflections on Refraction (Geometrical Optics) 629

Define the Hadamard quotient of the power series F and H by

 f2k+1
F/H(z) = z 2k+1 .
h2k+1
k≥0

Theorem 1. Under the above assumptions, the following two equalities are equiv-
alent
N  
 ξ
F (ξ) = al H
nl
l=1
and
N
 z
F/H(z) = al n l .
n2l − z 2
l=1

Proof. The first equality reads

N  2k+1

2k+1
  ξ
f2k+1 ξ = al h2k+1 .
nl
k≥o l=1 k≥o

Therefore
N
al
f2k+1 = h2k+1 2k+1
(4)
n
l=1 l
 f2k+1 N
   z 2k+1  N
z
z 2k+1 = al = al n l 2 . 
h2k+1 nl nl − z 2
k≥0 l=1 k≥0 l=1

Corollary 1. Given Ψ, the number N of plates of the optical system is the number
of positive poles of the rational function
N
1
 z
Ψ/z(1 − z 2 )− 2 = al n l .
n2l − z2
l=1
The positive poles n1 , n2 , . . . nN are the respective indices of refraction and the
respective thicknesses of the plates are given by
1
al = − residue at nl .
2nl
 1
Corollary 2. The only optical system such that for all k, f2k+1 ∈ −k2 Z is the
trivial one, i.e., its index of refraction is uniformly equal to 1. Then necessarily
 1
f2k+1 = (−1)k −k2 a where a is a positive integer.
Proof. According to the hypothesis
1

Ψ/z(1 − z 2 )− 2 = e2k+1 z 2k+1
k

where e2k+1 ∈ Z. The radius of convergence is then surely ≤ 1 (it cannot vanish).
Since it is R = n1 ≥ 1, the radius equals 1. A classical theorem of Polya’s [5]
630 M. Mendès France and A. Sebbar Ann. Henri Poincaré

asserts that a power series with integral coefficients whose radius of convergence
is 1, is either a rational function or a transcendental function where |z| = 1 is a
natural boundary. Since
1
 aℓ n ℓ z
Ψ/z(1 − z 2 )− 2 =
n2ℓ − z 2
ℓ
∞ n(x)zdσ(x)
or maybe 0 n2 (x)−z 2 , then necessarily
1 a1 z
Ψ/z(1 − z 2 )− 2 =
1 − z2
so that
a1 ξ
Ψ(ξ) =  . 
1 − ξ2

4. An algebraic statement
Given the function φ(θ), we know that the optical system is determined given that
the index of refraction increases. What if we only know some derivatives of φ at
the origin?
Theorem 2. Suppose there exists no nontrivial polynomial R ∈ Z[X0 , X1 , . . . , X2N ]
of degree at most (2N + 1)! such that
′′′
R φ′ (0), φ (0), . . . , φ(4N +1) (0) = 0 .
Then the optical system consists of at least N + 1 plates.
Proof. The derivatives of φ are rational linear combinations of those of Ψ so that
the theorem can be stated with the coefficients f2k+1 of the Taylor expansion of Ψ
2N

Ψ(ξ) = f2k+1 ξ 2k+1 + O(ξ 4N +3 ) .
k=0

Equation (4) reads

N
 al f2k+1
2k+1
=  1
n
l=1 l (−1)k −k2
for k = 0, 1, . . . . In order to simplify notations in the following discussions put
al 1 f2k+1
= αl , = νl ,  1 = gk .
nl n2l (−1)k −k2
Consider the system of 2N + 1 equations where the αl and νl could be thought of
as 2N unknowns and where the gk , k = 0, 1, . . . , 2N are given
N

αl νlk = gk , 0 ≤ k ≤ 2N . (5)
l=1
Vol. 9 (2008) Reflections on Refraction (Geometrical Optics) 631

The number 2N +1 of equations exceeds the number 2N of unknown and therefore

unless the parameters gk are linked the system should have no solutions. And
indeed a classical result as found for example in O. Perron’s book [4] (Algebra I,
p. 129, satz 57) states:
Given M + 1 polynomials in M variables of respective degrees
m1 , m2 , . . . , mM+1 with a common zero, these polynomials are neces-
M+1
sarily linked by an algebraic polynomial of degree at most j=1 mj .
Therefore if the system (5) is solvable then necessarily there exist a nontrivial
polynomial R of degree (2N + 1)! such that
R (g0 , g1 , . . . , g2N ) = 0 .
If no such polynomial exists then (5) has no solution and the optical system must
contain at least N + 1 plates.
Theorem 2 is a partial answer to a question of Ken Falconner’s who asked
what can be said about an optical system where φ(θ) is not precisely known. Our
result shows that if we know that the M first nonzero derivatives of Ψ are not
linked by an algebraic relation of degree (2N + 1)! then necessarily the number of
plates N is larger than M2 . 

5. A remarkable example with infinitely many layers

Suppose
 1
 −2
Ψ(ξ) = (−1)k ζ(2k + 1 + α)ξ 2k+1
k
k≥0
where α > 0 is a given parameter and where ζ is the Riemann function. We then
know that
1
 aℓ n ℓ z 
Ψ/z(1 − z 2 )− 2 = 2 2
= ζ(2k + 1 + α)z 2k+1
nℓ − z
1≤ℓ 0≤k
  1
= z 2k+1
ℓ2k+1+α
0≤k 1≤ℓ
 1 ℓz
= .
ℓα ℓ2 − z 2
1≤ℓ

Equating both sides, we get nℓ = ℓ and aℓ = ℓ−α . If α > 1 the total thickness
of the optical system is finite. As ℓ increases to infinity, nℓ tends to infinity. At
the limit, the luminous rays are parallel to the x-axis within the system. This is
somehow a paradoxical situation since the outcoming rays keep the same horizontal
direction contrary to the finite case where the outcoming ray is parallel to the
incoming ray. Paradoxical as this may seem, we should not be too surprised. Infinity
often creates strange situations. If for example the horizontal ray hits the system
from the right hand side, the outcoming ray on the left of the system is completely
632 M. Mendès France and A. Sebbar Ann. Henri Poincaré

undetermined and we could agree to say that it diffuses. Of course we are here far
from the real world where the highest known indices seem not to exceed 4.
Suppose now α = 1. The optical system is infinitely thick. For all θ <
π/2, φ(θ) stays finite since
1 sin θ
 <∞
ℓ ℓ2 − sin2 θ
ℓ≥1

and the rays become asymptotically parallel to Ox.

In this last case α = 1,
∞  
1
 z 1 1 π
Ψ/z(1 − z 2 )− 2 = = −
ℓ2 − z 2 2 z tan πz
ℓ=1

hence the identity

∞   
 1 z z 1 1 π
√ √ = − .
ℓ ℓ − z2
2 1−z 2 2 z tan πz
ℓ=1

Similarly we have
∞
(−1)ℓ−1
  
 z z 1 π 1
√ √ = − .
ℓ ℓ − z2
2 1−z 2 2 sin πz z
ℓ=1

But this last identity does not correspond to any optical system since
aℓ = (−1)ℓ−1 /ℓ cannot represent the thickness of a layer! By adding and sub-
stracting the two identities we obtain
∞ 
 1 z z 1 π π
√ = −
2ℓ + 1 (2ℓ + 1)2 − z 2

1−z 2 4 sin πz tan πz
ℓ=0
∞   
 1 z z 1 2 π
√ = −
4 z tan π z2

2ℓ (2ℓ) − z
2 2 1−z 2
ℓ=1

which both corresponds to optical systems. The second one is clearly seen to be
equivalent to the very first one at the beginning of this remark.

6. The continuous case

Upon to now we considered optical systems with finitely many homogenous plates
except in the last section. Here we discuss the general case
 ∞
dσ(x)
Ψ(ξ) = ξ
0 n2 (x) − ξ 2
which is easily seen to be equivalent to
∞
n(x)dσ(x)

2 − 12
Ψ/z(1 − z ) = z .
0 n2 (x) − z 2
Vol. 9 (2008) Reflections on Refraction (Geometrical Optics) 633

Even if we assume n(x) increasing, it does not seem to be easy to determine n(x)
and dσ given Ψ. The moment problem we are led to
 ∞
dσ f2k+1
 2k+1
= (−1)k  −1
0 n(x) k
2

is not simple.
We limit ourselves to illustrate the continuous case with n(x) = eαx , α > 0
and dσ = xβ dx, β < −1, and compute Ψ or its coefficients f2k+1 . As we just noted
 −1   ∞
f2k+1 = (−1)k 2 e−α(2k+1)x xβ dx
k 0
 −1   ∞
du
= (−1)k 2 e−u uβ  β+1
k 0 α(2k + 1)
 −1  
k 2
Γ β+1
= (−1) β+1
k

α(2k + 1)
Γ β + 1 Γ k + 21
 
1
= √ β+1 .
πα k! (2k + 1)β+1
And therefore
1
 
Γ β+1  Γ k+ 2 ξ 2k+1
Ψ(ξ) = √ β+1 .
πα k! (2k + 1)β+1
k≥0

In particular, for β = 0
∞
zeαx dx 1 1+z

2 − 12
Ψ/z(1 − z ) = = log
0 e2αx − z 2 2α 1−z
so that
1

1 Γ k+ 2 ξ 2k+1
Ψ(ξ) = √
πα k! 2k + 1
k≥0
1 1+ξ 1
= log ⋆ ξ(1 − ξ 2 )− 2
2α 1−ξ
where ⋆ represents the Hadamard product.

7. Refractals
Consider once again the case of a unique plate with varying index x → n(x). As
already mentioned
 ∞
sin θ dσ(x)
φ(θ) = 
0 n2 (x) − sin2 θ
provided the function n(.) is σ-integrable.
634 M. Mendès France and A. Sebbar Ann. Henri Poincaré

If however it is not, it is bounded from below and from above by two mea-
surable functions. We can then consider both the lower and the upper integrals
 ∞ ⋆ ∞
sin θ dσ(x) sin θ dσ(x)
φ∗ (θ) =  and φ∗ (θ) =  .
⋆ 0 n2 (x) − sin2 θ 0 n2 (x) − sin2 θ

If φ∗ (θ) < φ∗ (θ) we interpret this as multirefraction: the incoming θ-ray emerges
as a beam of width (φ∗ (θ) − φ∗ (θ)) cos θ. We propose to call such an optical system
a refractal.
Refractals do not diffract light nor they diffuse, disperse nor scater light. The
phenomenon is entirely different but unfortunately no experiment will ever demon-
strate it! The real world ignores non-measurable functions which as we mathemati-
cians know, depend on the Axiom of choice!
Here we are far away from the trends of today’s philosophy according to
which there is no room for metaphysics within physics. Refractals are an instance
of what could be called “abstract physics” which describes a non-existing world.

8. The average of φ(θ)

The average of
∞
sin θ dσ(x)

φ(θ) = 
0 n2 (x) − sin2 θ
with respect to the incident angle θ ∈ [0, π/2[ and with respect to the thickness of
the system is
  π
2 ∞ 2 sin θ dθdσ(x)
 ∞
M(φ) = dσ(x)
π 0

0 n2 (x) − sin2 θ 0

which is easily seen to be

∞  ∞
1 n(x) + 1

M(φ) = log dσ(x) dσ(x) .
π 0 n(x) − 1 0

If the index of refraction n(x) is constant from one layer to the other, then
 2
1 n+1 1 n+1
M(φ) = log = log .
π n−1 2π n−1
The ratio ( n−1 2
n+1 ) is a well known quantity called the reflectivity R of the surface
of the plate (see for example M. Born, E. Wolf; Principles of Optics, page 42 [1]).
Then

R = exp − 2πM(φ) .
The right hand side of the formula makes sense even for non-constant indices. Our
formula could then be considered as an average measure of the reflectivity.
Vol. 9 (2008) Reflections on Refraction (Geometrical Optics) 635

Here is a simple application. Given the measure dσ, minimize the harmonic
average n̂
 ∞  ∞ 
1 dσ(x)
= , dσ(x) = 1
n̂ 0 n(x) 0
knowing that the mean M(φ) is equal to a known value M . The problem is to find
the extrema of the above integral given the constraint
 ∞
n(x) + 1
log dσ(x) = πM .
0 n(x) −1
Following Lagrange we introduce the parameter λ:
 
d 1 n+1
+ λ log =0
dn n n−1
 
1 1 1
− 2 +λ − =0
n n+1 n−1
and therefore
1
n2 (x) =
1 − 2λ
which shows that the index is constant. Then
n+1
log = πM
n−1
−1
and finally n̂ = n = (tanh(πM/2)) . The system consists of one plate with
constant index, and this indeed minimizes n̂.

9. Blind spots
Consider a plate P of thickness d where now the index of refraction n(x, y) ≥ 1
varies continuously with the point p = (x, y) ∈ P . We actually assume that the
map p → n(p) is continuously differentiable on P . As before, the front face ∂P0 is
identified with the axis x = 0 and the back face ∂Pd with x = d.
To each A ∈ ∂P0 we associate the gradient of n at A denoted grad n(A). If
none of the incident rays at A reach ∂Pd , A is said to be a blind point or blind
spot. We shall show that if there exists such a point then the plate P cannot be
too thin. Before stating our theorem we define
|||grad log n||| = sup ||grad log n(p)|| .
p∈P

Theorem 3. If A is a blind point then the thickness d of the plate satisfies the
inequality
  ∂n(A)  
1 
∂y

d> 1− .
|||grad log n||| ||grad n(A)||
636 M. Mendès France and A. Sebbar Ann. Henri Poincaré

If all the spots of ∂P0 are blind then P is opaque even though sup n(p) may
be finite. No rays from either side of P manage to reach the opposite side , and
then
  ∂n(A)  
1 
∂y

d> 1 − inf .
|||grad log n||| A∈∂P ||grad n(A)||

In this formula, ∂P = ∂P0 ∪ ∂Pd .

The second hand side of the inequality in Theorem 3 is invariant if the index n
is multiplied by a positive scalar. In other words, if the plate P is plunged into a
medium of constant index ν and if for that new optical system A is blind , then
the inequality holds. It can be shown that the inequality is sharp for ν = 0, since
then all incident rays at A are reflected following the tangent grad n(A). In terms
of physics this is of course nonsense since ν = 0 corresponds to a medium where
the speed of light is infinitely larger than in vacuum.

10. Proof of Theorem 3

The proof requires two steps:
Step 1: It is well known that if the index of refraction n varies continuously from one
point to another then the luminous ray Γ satisfy the partial differential equation
 
d dp
grad n(p) = n(p) , p∈Γ
ds ds
where s is the arclength of Γ (see for example W. Pauli [3], p. 5 and p. 6). A simple
and classical consequence is that if R(p) is the radius of curvature of Γ at p, then
R(p) = ||grad log n(p)||−1
and therefore
R(p) ≥ |||grad log n|||−1 .

Step 2: Consider the incident ray at A which is colinear with grad n(A). As it
penetrates P it is not immediately deflected since it coincides with the normal to
the curve n(x, y) = n(A). Both the angles of incidence and refraction at A are
equal to the acute angle ω between grad n(A) and the x−axis . Therefore
  ∂n(A)  
 
∂x
ω = arccos .
||grad n(A)||

Consider two circles tangent at A to grad n(A), of radius R = |||grad log n|||−1 .
According to Step 1, the refracted luminous curve Γ cannot penetrate neither
of the two discs if d ≤ d0 where d0 is defined on the figure. In other terms, if d ≤ d0
the luminous ray Γ necessarily reaches ∂Pd . Therefore, if A is a blind spot, d > d0 .
Vol. 9 (2008) Reflections on Refraction (Geometrical Optics) 637

d0
• grad n(A)

A ω x
•

Figure 3. Definition of d0 .

To conclude the proof of the theorem we are left to compute d0 . Easily seen,
d0 = R(1 − sin ω). Now
  ∂n(A)  
 
∂x
sin ω = sin arccos
||grad n(A)||
  2  12
∂n(A)
= 1− ||grad n(A)||−1
∂x
 
 ∂n(A) 
=  ||grad n(A)||−1 .
∂y 

11. Appendix
The functional Equation (3) met in Section 3 is quite general. Here is an example in
Number Theory where the motivations are obviously entirely different and where F
and H are no longer continuous. Put H(ξ) = ξ1 [ξ] ([.] represent the integer part).
Define
∞
F (ξ) = aℓ H(ξ/ℓ)
ℓ=1
where the unknown real sequence aℓ verifies lim inf ℓaℓ > −∞. A. E. Ingham es-
tablishes the subtle Tauberian theorem according to which
∞
lim F (ξ) = λ ⇒ aℓ = λ
ξ→∞
ℓ=1
(see [2]).
638 M. Mendès France and A. Sebbar Ann. Henri Poincaré

Acknowledgements
Many thanks to Jean Creignou and particularly to Alain Yger.

References
[1] M. Born, E. Wolf, Principles of Optics, Pergamon Press, sixth edition reprinted 1989.
[2] A. E. Ingham, Some Tauberian theorems connected with prime numbers, Jour. Lon-
don Math. Soc. 20 (1945), 171–180.
[3] W. Pauli, Optics and the Theory of Electrons, Pauli Lectures on Physics 2, Dover
Publ. 2000.
[4] O. Perron, Algebra I, Die Grundlagen , Walter de Gruyter, Berlin 1951.
[5] G. Pólya, Über Potenzreihen mit ganzzahligen Koeffizienten, Math. Annalen 77
(1916), 497–518.

Michel Mendès France and Ahmed Sebbar

Institut de Mathématiques de Bordeaux
UMR 5251
Université Bordeaux I
F-33405 Talence Cedex
France
e-mail: michel.mendes-france@math.u-bordeaux1.fr
sebbar@math.u-bordeaux1.fr

Communicated by Vincent Rivasseau.

Submitted: December 20, 2007.
Accepted: February 20, 2008.