For all k, since there are only finitely many possible values for an (n≥k) to take, there exists nmax≥k so that anmax≥an for all n≥k. Now, we can prove by induction that am≥(1−2−nmax−2−nmax−1−⋯−2−(m+1))anmax for all k≤m≤nmax. This is trivial when m=nmax. Now suppose that it holds for m+1, then we have
am+1≤am+2−(m+1)a2(m+1)≤am+2−(m+1)anmax
since 2(m+1)>k. By the induction hypothesis, we get that
am≥am+1−2−(m+1)anmax≥(1−2−nmax−⋯−2−(m+1))anmax,
as desired. In particular, we can set m=k to get that
ak≥(1−2−nmax−⋯−2(k+1))anmax≥1−i=k+1∑∞2−i=1−2−k
where the inequality follows from the fact that anmax≥aN≥1 and the fact that we are extending the finite sum to the infinite sum. □
We first reduce the problem to the case where an=an−1+2−na2n holds for all n≤N. We will show this by showing that we can decrease a1,…,aN−1 to achieve this. Let k∈[N] be the largest such that ak<ak−1+2−ka2k. Let M be the difference between the two sides. Let ak′=ak,ak+1′=ak+1,… and for each i=1,…,k−1 we
let ai′=ai−(1−2−i−1−⋯−2−k+1)M<ai (in particular, ak−1′=ak−1−M). Then
it is clear that ai′≥ai−M, and so
an−1′=an−1−an+an′+2−nM≥an′+2−n(M−a2n)≥an′−2−na2n′.
for all n=1,…,k−1. It is also clear that ak−1′=ak′−2−ka2k′ by definition. Therefore a1′,a2′,… is a sequence not greater than a1,a2,… term-wise that satisfies the inequality and have one additional equality holding. We can therefore iteratively decrease the sequence term-wise so that all the equalities hold for n<N. Then we can simply prove by induction that 1−2−k<ak≤1. In particular, ak>1−2−k holds for the original sequence. □