Search Variants Beyond Binary

Binary search isn’t the only game in town. A handful of specialized variants beat plain binary search in specific scenarios — sometimes by enough to be the right answer for a real-world system. This page catalogs them.

You won’t see most of these in interviews. You will see them in production code, library internals, and the occasional research paper. Knowing they exist is the difference between “I always reach for binary search” and “I pick the right tool.”

1. Jump Search — for when comparisons are cheap but seeks are expensive

Idea. Instead of halving the search space each step, jump forward by a fixed step size until you overshoot the target, then linearly walk back through the last block.

• Optimal step size: √n. With this step, total comparisons ≈ √n + √n = 2√n.
• Time: O(√n).
• Space: O(1).

import math
 
def jump_search(arr, target):
    n = len(arr)
    step = int(math.sqrt(n))
    prev = 0
 
    # Jump forward until we overshoot
    while prev < n and arr[min(step, n) - 1] < target:
        prev = step
        step += int(math.sqrt(n))
        if prev >= n:
            return -1
 
    # Linear walk back in the last block
    for i in range(prev, min(step, n)):
        if arr[i] == target:
            return i
    return -1

int jumpSearch(const vector<int>& arr, int target) {
    int n = arr.size();
    int step = (int)sqrt(n);
    int prev = 0;
 
    while (prev < n && arr[min(step, n) - 1] < target) {
        prev = step;
        step += (int)sqrt(n);
        if (prev >= n) return -1;
    }
    for (int i = prev; i < min(step, n); i++) {
        if (arr[i] == target) return i;
    }
    return -1;
}

public int jumpSearch(int[] arr, int target) {
    int n = arr.length;
    int step = (int) Math.sqrt(n);
    int prev = 0;
 
    while (prev < n && arr[Math.min(step, n) - 1] < target) {
        prev = step;
        step += (int) Math.sqrt(n);
        if (prev >= n) return -1;
    }
    for (int i = prev; i < Math.min(step, n); i++) {
        if (arr[i] == target) return i;
    }
    return -1;
}

Where it wins: when seeking to a random position is expensive but reading sequentially is cheap. Think: a sorted array on disk where each seek is a head movement. Jump search minimizes seeks; binary search needs more seeks but does fewer overall reads. Mostly a curiosity in modern systems — RAM/SSD have made the trade-off moot.

2. Interpolation Search — for uniformly distributed data

Idea. Binary search always picks the middle. Interpolation search picks where it guesses the target is, based on the value distribution.

If you’re looking for 25 in [10, 20, 30, 40, ..., 100], the middle (50) is way off. But linear interpolation says 25 should be ~1/6 of the way through — pick that index instead.

pos = lo + ((target - arr[lo]) * (hi - lo)) / (arr[hi] - arr[lo])

• Time: O(log log n) on uniformly distributed data — much faster than binary search.
• Time: O(n) worst case on skewed data.
• Space: O(1).

def interpolation_search(arr, target):
    lo, hi = 0, len(arr) - 1
    while lo <= hi and arr[lo] <= target <= arr[hi]:
        if lo == hi:
            return lo if arr[lo] == target else -1
        # Estimated position
        pos = lo + ((target - arr[lo]) * (hi - lo)) // (arr[hi] - arr[lo])
        if arr[pos] == target:
            return pos
        if arr[pos] < target:
            lo = pos + 1
        else:
            hi = pos - 1
    return -1

int interpolationSearch(const vector<int>& arr, int target) {
    int lo = 0, hi = arr.size() - 1;
    while (lo <= hi && target >= arr[lo] && target <= arr[hi]) {
        if (lo == hi) return arr[lo] == target ? lo : -1;
        int pos = lo + ((long long)(target - arr[lo]) * (hi - lo)) / (arr[hi] - arr[lo]);
        if (arr[pos] == target) return pos;
        if (arr[pos] < target) lo = pos + 1;
        else                   hi = pos - 1;
    }
    return -1;
}

Where it wins: phone-book-like data — keys roughly uniformly distributed. Database systems with statistics on key distributions sometimes use this for index lookups.

Where it loses: non-uniform data (most real-world keys), or duplicate values that cause arr[hi] - arr[lo] = 0 (must guard against division by zero).

3. Exponential Search — for unbounded or unknown-size sorted data

Idea. First, find the range where the target could be by doubling: check arr[1], arr[2], arr[4], arr[8], arr[16], ... until you overshoot. Then run binary search inside the bracketed range.

• Time: O(log n) — same as binary, but doesn’t need to know n upfront.
• Space: O(1).

def exponential_search(arr, target):
    n = len(arr)
    if arr[0] == target: return 0
 
    # Phase 1: find range by doubling
    i = 1
    while i < n and arr[i] < target:
        i *= 2
 
    # Phase 2: binary search in [i // 2, min(i, n) - 1]
    lo, hi = i // 2, min(i, n) - 1
    while lo <= hi:
        mid = lo + (hi - lo) // 2
        if arr[mid] == target: return mid
        if arr[mid] < target: lo = mid + 1
        else:                 hi = mid - 1
    return -1

Where it wins:

• Infinite or unbounded data — searching in an infinite sorted stream where you don’t know the size.
• Target near the beginning — phase 1 finds the bracketing range fast.
• Sorted-but-no-length API — like searching for an element in a stream where you can only ask “what’s the i-th element?”

This is what Python’s bisect module and Go’s sort.Search lean on under the hood for sorted slices of unknown size.

4. Fibonacci Search — when division is expensive

Idea. Instead of dividing the range in half (requires / 2), divide it using Fibonacci numbers (only requires subtraction).

• Time: O(log n).
• Space: O(1).
• Comparisons: very similar to binary search.

def fibonacci_search(arr, target):
    n = len(arr)
    fib2, fib1 = 0, 1
    fib = fib2 + fib1
    while fib < n:
        fib2, fib1 = fib1, fib
        fib = fib2 + fib1
 
    offset = -1
    while fib > 1:
        i = min(offset + fib2, n - 1)
        if arr[i] < target:
            fib, fib1 = fib1, fib2
            fib2 = fib - fib1
            offset = i
        elif arr[i] > target:
            fib, fib1 = fib2, fib1 - fib2
            fib2 = fib - fib1
        else:
            return i
    if fib1 == 1 and offset + 1 < n and arr[offset + 1] == target:
        return offset + 1
    return -1

Try it as an exercise. The main reason to include it here: it’s the bridge between the recurrence-relation thinking from Day 5 (Fibonacci recurrence) and concrete divide-and-conquer.

5. Ternary Search — for unimodal functions

Idea. Binary search assumes the data is sorted (monotonic). Ternary search assumes it’s unimodal — strictly increasing then strictly decreasing (or vice versa). Splits into three parts and discards either the leftmost or rightmost third.

def ternary_search(lo, hi, f, eps=1e-9):
    """Find x in [lo, hi] that maximizes the unimodal function f."""
    while hi - lo > eps:
        m1 = lo + (hi - lo) / 3
        m2 = hi - (hi - lo) / 3
        if f(m1) < f(m2):
            lo = m1
        else:
            hi = m2
    return (lo + hi) / 2

• Time: O(log_1.5 n) ≈ O(log n) — slightly more comparisons than binary search.
• Space: O(1).

Where it wins: continuous optimization on unimodal functions — finding the peak of a parabola, the minimum of a convex function. Used in numerical optimization libraries.

Interview relevance: Find Peak Element (practice problem) is a discrete cousin — though you usually solve it with plain binary search.

A side-by-side comparison

What about hash tables?

If you can give up the “sorted” requirement entirely — hash tables get you O(1) average lookup, with no comparisons at all. That’s the topic of Day 8. The trade-off: hash tables don’t support range queries, ordered iteration, or finding next-greater / next-smaller. Pick based on what queries you need.

The mental ranking for interviews

For 99% of interview problems:

1. Binary search (or the unified template) is the answer.
2. Linear search if the data isn’t sorted and there’s no time to sort.
3. Hash table if you don’t need order.

Reach for the variants only when a specific constraint demands it. Otherwise, master the template on the previous page and you’ve covered the entire chapter’s interview surface.

Next: applications — where searching is the silent infrastructure of real systems.