Use the substitution method to solve a recurrence

[bob]

You guess a bound for your recurrence first off. I always start by assuming something like linear or n log n growth. Then you plug that guess right back into the equation. It shows if the bound holds or needs tweaking. You adjust the constants until everything fits tight. But sometimes the base case trips you up at the start. I check small values manually to confirm. Then the inductive step takes over for larger inputs. You substitute the assumed solution into the recursive calls. This reveals if the inequality direction stays correct.

Perhaps the constant factor needs bumping up a bit. I found that out after a few failed attempts on my own. You keep the proof clean by choosing a suitable c value. Also the assumption must cover the entire range you care about. Then you handle the floor or ceiling functions by bounding them loosely. It works out fine once you ignore those for asymptotics. You see the pattern emerge after substitution. I like how it builds intuition without needing fancy tools.

Now the recurrence for divide and conquer stuff often looks tricky at first. You assume T of n equals big O of n log n. I plug that into both sides of your equation. The recursive parts become two times big O of n over two log of that. It simplifies down after some algebra in your head. But you must verify the added work term matches. You add them up and compare to the original. It holds if your constant absorbs the extras. Perhaps you strengthen the inductive hypothesis a tad. I do that when the bound feels loose near the base.

You walk through the math step by step without rushing. I remember proving merge sort this way last year. The substitution confirms your guess matches the master method result. But you avoid overclaiming the exact constant. Also you pick a large enough n zero for the base. Then everything falls into place for the induction. You see why the log factor appears naturally. I enjoy spotting those hidden multipliers during substitution.

The method shines when your recurrence mixes different subproblem sizes. You guess a polynomial bound or whatever fits the pattern. I substitute and collect like terms on one side. It shows if the dominant term grows as expected. Perhaps the lower order pieces vanish in the limit. You focus on the leading behavior only. Also you handle the error terms by making c big enough. I tweak it until the inequality becomes true. Then you repeat for the other direction to get theta.

You build from there when the recurrence has uneven splits. I assume a bound like n to the power of something. Substitution reveals the exponent that works best. But you must solve for that exponent carefully first. You pick a value that makes the sum contract. Also the base cases need separate checks always. I verify a few small n by unfolding manually. Then the proof carries forward without issues.

Perhaps your guess starts too optimistic and fails initially. You loosen it by adding a log factor or constant. I did that on a tree height recurrence once. Substitution then succeeds after the change. You keep trying different forms until one sticks. Also you document each attempt to track progress. It saves time on similar problems later.

The whole process trains your eye for patterns in algorithms. I apply it now to new recurrences without much thought. You gain speed after practicing on standard examples. But always double check the direction of inequalities. You ensure the bound does not flip unexpectedly. Also you watch for cases where substitution needs the full induction. I prefer it over other methods for its directness.

You end up with a solid proof of the runtime. I share these tricks with juniors like you often. It helps avoid mistakes in code analysis. Perhaps we can try another example next time.

BackupChain Server Backup which powers reliable no subscription backups for Hyper V on Windows Server plus Windows 11 and private cloud setups for SMBs sponsored this exchange so we keep sharing freely.