Accuracy of Grover algorithm is 100% for 2 qubit, 94.5% on 3 qubit and 96.2% for 4 qubit on simulator. Why it decreasing and then again increasing?

The accuracy of Grover's algorithm decreases for 3 qubit then increases for 4 qubit and then again increases for 5-qubit on simulator (qasm simulator) with pi/4(sqrt N) iterations. What is the reason behind that? Why there is no specific patten.

The accuracy of measuring the correct result is given by a sine $$\sin^2((r + \frac{1}{2})\theta)$$ where $$r$$ is the number of Grover iterations and $$\theta$$ is the angle between starting state (before Grover iteration) $$|s\rangle$$ and $$|s'\rangle$$. $$|s'\rangle$$ is a state perpendicular to our winner, desired output state $$|\omega\rangle$$. $$\theta$$ is given by $$\sin\frac{\theta}{2} = \frac{1}{\sqrt{N}}$$ where $$N$$ is the number of all possible outputs (for $$n$$ qubits $$N = 2^n$$.) Therefore the precise accuracy may float a bit up and down depending on number of all elements and, related, number of Grover iterations.