Source. Laczkovich (1984), Lemma 1, printed pp. 110–112
(PDF pp. 2–4).
Statement
Let n≥1 be an integer, K≥0, and
f:{0,…,n}→R. Suppose ∣f(i)∣≤K and
2f(i)≤f(i+h)+f(i+2h)for all integers i,h with 0≤i<i+h<i+2h≤n.(1)
Then
f(0)≤f(n)+n10K.(2)
Source precision. The source assumes K>0; the case K=0 is
immediate. We state n≥1 because (2) is undefined at zero, and
handle n=1 before the printed induction begins at n=2.
In the last dyadic branch, n=2k+2 already gives the desired bound
at index zero. The further backward step, whose index would be
negative there, is used only when n≥2k+3.
Dependencies. The finite restriction convention in
Definitions
and induction.
Bears on. Problem 1125, through
Theorem 2.
Proof
If K=0, the function vanishes. If n=1, then
f(0)−f(1)≤2K≤10K, even though (1) has no instances.
Hence assume K>0 and n≥2.
For k≥1 and 2k≤n<2k+1, we prove the three estimates
Ak:Bk:Ck:n=2k ⟹ f(0)≤f(n)+2K/2k,n=2k+1 ⟹ f(0)≤f(n)+6K/2k,2k+2≤n<2k+1 ⟹ f(0)≤f(n)+5K/2k.(3)
These imply (2). For Ak this is immediate;
for Bk use 6(2k+1)≤10⋅2k;
and for Ck use n<2k+1.
For k=1, (1) gives
f(0)≤2f(1)+f(2)≤f(2)+K,
which is A1. At n=3 the crude bound
f(0)−f(3)≤2K implies B1.
The range in C1 is empty.
Assume k≥2 and all three assertions at level k−1.
For any n in the current dyadic range, apply Ak−1
to the terminal interval of length 2k−1. It gives
f(n−2k−1)≤f(n)+2k−12K.
The inequality (1) at i=n−2k and h=2k−1 then yields
f(n−2k)≤2f(n−2k−1)+f(n)≤f(n)+2k2K.(4)
At n=2k, this proves Ak.
Next let n=2k+1. Equation (4) bounds f(1). We also have
f(2)≤f(n)+2k−15K.(5)
For k=2, (5) follows from f(2)−f(n)≤2K≤5K/2.
For k≥3, the translated interval from 2 to n has length
2k−1, which satisfies
2k−1+2≤2k−1<2k.
Thus Ck−1 proves (5). Averaging (4) and (5) in
f(0)≤(f(1)+f(2))/2 gives
f(0)≤f(n)+2kK+2k5K=f(n)+2k6K,
proving Bk.
Finally, suppose 2k+2≤n<2k+1 and set j=n−2k≥2.
The just-proved Bk, applied to the terminal interval
from j−1 to n, and (4) give
f(j−1)≤f(n)+2k6K,f(j)≤f(n)+2k2K.
Their average bounds the preceding value:
f(j−2)≤f(n)+2k4K.(6)
If j=2, this already proves Ck. If j≥3, another
application of (1), now to j−3,j−2,j−1, gives
f(j−3)≤f(n)+2k5K.
Both f(j−3) and f(j−2) are therefore at most
f(n)+5K/2k. Repeatedly applying
f(i)≤(f(i+1)+f(i+2))/2 propagates this bound backward to i=0.
For j=3 it is already the bound at zero. This proves Ck,
closes the induction, and proves (2).