Asymptotics of Recursive Bound

18 Views Asked by At

Suppose that we knew that for $n>N$, we have $$F(n) \le F(n-g(n)) +h(n)$$ for some well behaved functions $g,h$. (for a concrete example, let’s say $N=1,g(n)=n^\alpha, h(n)=n^\beta$ where $\alpha,\beta \in (0,1))$)

I was wondering then how we can get an asymptotic upper bound of $F$. Would this be an easy job for Riemann-Stieltjes integration? (I am not very knowledgeable about this form of integration, so if this is the case, working through my concrete example would be appreciated)

1

There are 1 best solutions below

0
On BEST ANSWER

Here's an approach that works well if you don't care about the leading constant. For simplicity assume $F$ is an increasing function, although one could probably do the technicalities to reduce to this case.

Suppose we have $F(n) \le F(n-n^\alpha) + n^\beta$. Starting from $F(N)$, let's iterate this inequality as many times as it takes so that the argument of $F$ in the upper bound is at most $N/2$. Each time we use the inequality, the argument is decreased by $n^\alpha$ for whatever our current value of $n$ is; since that value (by construction) is always at least $N/2$, we decrease the argument of $F$ by at least $(N/2)^\alpha$. Therefore the argument will go below $N/2$ in at most $(N/2)^{1-\alpha}$ iterations.

At each iteration, the upper bound increases by $n^\beta$ for the current value of $n$; since these values are all at most $N$, the upper bound increases by at most $N^\beta$ each time. The result of this iteration is the new recursive upper bound $$ F(N) \le F(N/2) + (N/2)^{1-\alpha}N^\beta = F(N/2) + 2^{\alpha-1}N^{\beta+1-\alpha}. $$ Now we iterate this inequality for $N$, $N/2$, $N/4$, and so on down to $N=1$ or wherever we want to stop. The resulting upper bound is a geometric series $$ F(N) \le F(1) + 2^{\alpha-1} \big( N^{\beta+1-\alpha} + (N/2)^{\beta+1-\alpha} + (N/4)^{\beta+1-\alpha} + \cdots \big); $$ notice that the exponent $\beta+1-\alpha$ is greater than $1$, so the first term is the dominant term (and we might as well bound the finite geometric series by the infinite geometric series it's the start of, whence the $\cdots$). The result is $$ F(N) \le F(1) + 2^{\alpha-1} \frac{N^{\beta+1-\alpha}}{1-2^{\alpha-1-\beta}}. $$ To get tighter and tighter upper bounds, one can reduce the $2$ in $N/2$ to a smaller and smaller constant; but the order of magnitude of this upper bound is probably correct. The limiting case is perhaps a differential equation? (If one assumes that $F(N)$ is bounded by a power of $N$, this could probably be made rigorous.) I'm not sure that I see a way to use Riemann–Stieltjes integration.