DP Pattern

Imagine you're assembling a huge jigsaw puzzle and you keep solving the same corner section over and over — every time you come back to it. That's what naive recursion does. Dynamic Programming (DP) says: solve each piece once, write the answer on a sticky note, and look it up if you need it again.

That simple idea — cache results, never recompute — turns exponential algorithms into polynomial ones.

Two Properties You Need

Overlapping Subproblems: The recursive solution calls the same sub-problems many times. Classic example: Fibonacci numbers. Computing F(5) calls F(4) and F(3). F(4) calls F(3) and F(2). F(3) is computed twice, F(2) is computed three times... Without caching, computing F(50) makes over a trillion function calls. With caching: 50 unique calls, done.

Optimal Substructure: The optimal answer to the big problem is built from optimal answers to smaller sub-problems. Shortest path from A to C through B: the A→B segment must be the shortest A→B path, and the B→C segment must be the shortest B→C path. If either sub-path were suboptimal, you could improve it and get a better overall path — contradiction. So optimal substructure holds.

Both properties together mean: compute each sub-problem once in optimal fashion, combine them, and you're done.

The DP Design Framework

1. Define the state: What does dp[i] (or dp[i][j]) represent? Be precise and write it out. Wrong state definition → everything downstream is wrong.

2. Write the recurrence (transition): Express dp[i] in terms of smaller indices. This is the hardest and most creative step.
• Fibonacci: dp[i] = dp[i-1] + dp[i-2]
• 0/1 Knapsack: dp[i][w] = max(dp[i-1][w], dp[i-1][w-wt[i]] + val[i])
• LCS: dp[i][j] = dp[i-1][j-1]+1 if s1[i]==s2[j] else max(dp[i-1][j], dp[i][j-1])

3. Define base cases: The smallest sub-problems you can answer directly.
• Fibonacci: dp[0]=0, dp[1]=1
• Knapsack: dp[0][w]=0 for all w (no items → no value)

4. Fill order or memoize: Either fill the table bottom-up (tabulation) ensuring all dependencies are computed before they're needed, or add a cache to the top-down recursion (memoization).

DP vs Greedy vs Divide & Conquer

All three break problems into smaller pieces, but differently:
• Divide & Conquer: subproblems are independent — solve each once, combine. (Merge sort, closest pair.)
• Greedy: makes one locally-optimal choice, never revisits. Works only when the greedy-choice property holds.
• DP: subproblems overlap, caches results. Considers all choices at each step. More powerful than greedy but typically slower.

Rule of thumb: try greedy first. If you can find a counterexample to the greedy choice, switch to DP.

DP Framework — Fibonacci (Naive → Memo → Tab → Space-Optimised)

import sys
from functools import lru_cache

# 1. Naive — exponential O(2^n)
def fib_naive(n):
    if n <= 1: return n
    return fib_naive(n-1) + fib_naive(n-2)

# 2. Memoization (top-down) — O(n)
@lru_cache(maxsize=None)
def fib_memo(n):
    if n <= 1: return n
    return fib_memo(n-1) + fib_memo(n-2)

# 3. Tabulation (bottom-up) — O(n) time, O(n) space
def fib_tab(n):
    if n <= 1: return n
    dp = [0] * (n+1)
    dp[1] = 1
    for i in range(2, n+1):
        dp[i] = dp[i-1] + dp[i-2]
    return dp[n]

# 4. Space-optimised — O(n) time, O(1) space
def fib_opt(n):
    a, b = 0, 1
    for _ in range(n):
        a, b = b, a+b
    return a

for fn, name in [(fib_memo,'memo'),(fib_tab,'tab'),(fib_opt,'opt')]:
    print(f"fib(10) [{name}] = {fn(10)}")

Output

fib(10) [memo] = 55
fib(10) [tab] = 55
fib(10) [opt] = 55

Complexity

🐢 Naive Fibonacci (no cache)Every call spawns two more; the same sub-problems are recomputed exponentially many times

Exponential blowup\(O(2^n)\)

⚡ Fibonacci with memoizationn unique states (F(0)..F(n)); each computed once and cached

Each state computed once\(O(n)\)

📏 General 1D DPn states, \(O(1)\) transition each = \(O(n)\) total

Linear in state count\(O(n)\)

📐 General 2D DPn × m state grid, \(O(1)\) transition each — LCS, edit distance, etc.

Quadratic\(O(n · m)\)

📦 Space (rolling array trick)When recurrence only looks back a fixed number of rows, discard older rows

Reduced table\(O(1)\) to \(O(row size)\)

Two Properties You Need

The DP Design Framework

DP vs Greedy vs Divide & Conquer

DP Framework — Fibonacci (Naive → Memo → Tab → Space-Optimised)

Complexity

Quick check