The pi in matlab is not a real 'pi',
but it is only a floating-point number close to 'pi'.
Trigonometric functions around pi may have errors close to machine precision.
See:
cos(pi/2) = 6.12323399573677e-17
sin(pi) = 1.22464679914735e-16
exp(i*pi/2) = 6.12323399573677e-17 + 1i
exp(i*pi) = -1 + 1.22464679914735e-16i
Note that sin(pi) returns '1.22464679914735e-16' and it is true somehow.
These unwanted results do not come from function evaluations. It is because pi value cannot be represented in floating-point number.
One approach to go around it is to use symbolic toolbox: sin(sym(pi)) returns 0. Note that sym(pi) is the exact pi.
See http://www.mathworks.co.kr/kr/help/symbolic/sin.html
There is another way without using symbolic toolbox.
You had better use degrees rather than radians
when your application needs to use correct values around such angles.
Note that sind(180) returns '0'.
The functions sind and cosd use degrees:
cosd(90)
sind(180)
Or, we can define functions with radian input:
cos_rad = @(theta) cosd(theta/pi*180)
sin_rad = @(theta) sind(theta/pi*180)
cos_rad(pi/2) = 0
sin_rad(pi) = 0
The exp function is not easy to handle with degree.
You can try the following definition:
exp_rad = @(theta) exp(real(theta)).*(cos_rad(imag(theta)) + i* sin_rad(imag(theta)))
exp_rad(i*pi/2)
exp_rad(i*pi)
The function exp_rad slightly abuses the definition pi.
This would be a trick for handling 'pi' as if it is a real 'pi'.
pi is not a rational number, but the 180 degrees are rational number.
Converting radians to degrees makes the exact result around pi.
The exp_rad also responses with matrix input:
A = [log(2) 0; i*pi/2 i*pi];
exp(A)
expm(A)
exp_rad(A)
[V,D]=eig(A);
V*diag(exp_rad(diag(D)))/V % expm_rad(A) example

Now we can provide 4 functions with abused pi definition:
cos_rad.m
sin_rad.m
exp_rad.m
expm_rad.m

Those show correct results around pi/2 and pi points if we want to take 'pi' as the real 'pi'.
These files are very short sample codes for understanding.
You can use them when your calculation needs to be exact zero with sin(N*pi) or cos((N+1/2)*pi).

For large angle > 6*pi radians, the *_rad functions happen to show non-zero at the intended points due to further loss of precision of large multiple of pi. See the comments on sin(pi) at http://www.mathworks.co.kr/company/newsletters/articles/the-tetragamma-function-and-numerical-craftsmanship.html

We can count the number of points where abs(sin(k*pi))<eps or abs(cos(k*pi+pi*0.5))<eps in order to see how accurate the trigonometric functions calculate for large k's.
N=5000;
tt=linspace(-N,N,8*N+1)*pi;
% count the points where the output is less than epsilon.
sum(abs(sin(tt))<eps(tt))
ans = 10001
sum(abs(cos(tt))<eps(tt))
ans = 10000
These points can be assumed to be 0 without loss of sin or cos natures.
Note that sum(abs(sin(tt))==0) and sum(abs(cos(tt))==0) return 1 and 0 because of floating-point errors.
Find more examples at test_pi_functions.m

Again we define *_eps functions that return zeros when their outputs are less than eps:
More functions are added:
cos_eps.m
sin_eps.m
exp_eps.m
expm_eps.m

Due to output comparison, _eps functions are slow a bit. If we know the phase angles are less 6*pi, we can still use _rad functions.

--jo