Why are you taking just half of the spectrum?
You need to apply the modification to the entire frequency range (i.e., both positive and negative frequencies). The "fft" needs the amplitudes from both sides of the frequency spectrum to correctly construct the signal in the time domain. The fact that the real parts of the amplitudes in the frequency domain are symmetric about 0 frequency and the imaginary parts of the amplitudes in the frequency domain are anti-symmetric about 0 frequency is what "forces" the data to be 100% real in the time domain. If you are missing half the spectrum, which makes for no symmetry, then your time domain product will be complex (i.e., have both real and imaginary parts).
Just so I understand what is going on, I usually construct and modify things in the frequency domain by taking the following into account:
g = rand(1,512); % define a random time series "g"
dt = 0.1; % define a time increment (seconds per sample)
N = length(g);
Nyquist = 1/(2*dt); % Define Nyquist frequency
df = 1 / (N*dt); % Define the frequency increment
G = fftshift(fft(g)); % Convert "g" into frequency domain and shift to frequency range below
f = -Nyquist : df : Nyquist-df; % define frequency range for "G"
You can check and confirm for yourself that the number of amplitudes in "g" are the same number of amplitudes in "G" - this is a property of the Fourier transform! Now you can do things to the amplitudes according to what frequency they represent. For example, you can take the derivative of the time signal by doing this in the frequency domain:
G_new = G.*(1j*2*pi*f);
Then convert back to the time domain by:
g_new = ifft(ifftshift(G_new));
