You know that moment in calculus class when the teacher draws three different rectangles under a curve and says they all give different answers? And then you're just sitting there thinking — wait, which one do I actually trust?
Here's the thing most people miss: the midpoint Riemann sum isn't always the "safe" choice. Plus, everyone acts like it's automatically better than left or right. And yeah, it usually is. But "usually" isn't "always." Sometimes the midpoint Riemann sum is an overestimate. And if you're taking an exam, or building something, or just trying to actually understand what's going on, that matters.
So let's talk about when is midpoint Riemann sum an overestimate — not in a textbook way, but in a way that actually makes sense.
What Is a Midpoint Riemann Sum
A midpoint Riemann sum is a way to estimate the area under a curve. You split the interval into chunks, and instead of using the left edge or the right edge of each chunk to decide the rectangle's height, you use the middle. Now, that's it. Middle of the slice, draw the rectangle up to the curve, add them all up.
The reason people like it is simple: sampling in the middle tends to balance out the errors. If the curve is rising, the left endpoint underestimates and the right overestimates. Midpoint sort of lands between. In practice, it's often shockingly close to the real integral.
But here's what most guides get wrong — they treat "midpoint is more accurate" as "midpoint is always closer and never wrong in direction.Think about it: " That's just not true. The midpoint sum has a bias. It depends entirely on the shape of the function.
Concavity Is the Whole Story
The secret lives in the second derivative. If you remember nothing else: concavity decides the direction of the error.
A function is concave up* when it curves like a bowl — think x². In practice, a function is concave down* when it arches like a frown — think -x². The midpoint rule sits above a concave down curve and below a concave up curve, locally, over each subinterval.
So if the whole function is concave down on the interval you're summing, the midpoint Riemann sum overestimates the true area. Now, every rectangle pokes above the curve a little. Add them up, and you're too high.
Why It Matters
Why does this matter? They memorize "midpoint = good" and move on. Now, because most people skip it. Then they hit a problem where the function is concave down, the midpoint sum is too big, and they mark it as "closest to exact" without checking the sign of the error.
In real applications — numerical integration in engineering, physics simulations, financial modeling — knowing which way your estimate leans is sometimes more useful than knowing how close it is. If you know you're overestimating drag, or overestimating cost, you can plan around it.
And look, on a test, this is a free point. The question "is the midpoint sum an over or underestimate?" shows up constantly. The answer is never "it depends" without context — it depends on concavity, and you can state that cleanly.
What Changes When You Understand This
Once you see the concavity rule, the whole family of Riemann sums clicks. Practically speaking, left and right sums are predictable based on increasing/decreasing. In practice, midpoint and trapezoid are predictable based on concavity. The trapezoid rule is the opposite of midpoint: it overestimates when concave up, underestimates when concave down.
That's why the error formulas for midpoint and trapezoid have opposite signs. But they're like mirror images. Real talk, that relationship is more useful than computing either one by hand a hundred times.
How It Works
Let's break down exactly how the midpoint sum behaves, and why the overestimate happens.
The Basic Setup
Say you've got a function f(x) on [a, b]. You cut it into n equal subintervals, each width Δx. Now, for each slice, you take the midpoint x_i* = (x_{i-1} + x_i)/2. The rectangle height is f(x_i*). Sum them: M_n = Δx · Σ f(x_i*).
That's the estimate. Worth adding: the error is E = M_n − ∫f(x)dx. That said, the true area is the definite integral. When E > 0, you've overestimated.
Why Concave Down Means Overestimate
Picture one slice of a concave down curve. The tangent line at the midpoint sits above the curve. It bends downward. The rectangle using the midpoint height is basically a flat top at that tangent-ish level. Since the curve falls away from the midpoint on both sides, the rectangle covers area that isn't under the curve.
Over one subinterval, the midpoint rectangle area is bigger than the true area under that slice. Do that across all slices on a concave down function, and the total midpoint sum is an overestimate.
For more on this topic, read our article on what is the period in physics or check out what are the 3 parts that make up a nucleotide.
Turns out this is provable with Taylor expansion — the local error term has a factor of f''(ξ), and for midpoint it's negative when f'' is negative. But you don't need the proof. You need the picture: frown curve, midpoint rectangles poke out the top.
When the Function Isn't Uniformly Concave
Here's where it gets messy in real life. They wiggle. Most functions aren't concave down on the whole interval. Part bends up, part bends down.
In that case, the midpoint sum might overestimate on some slices and underestimate on others. The global error is the net of those. So "when is midpoint Riemann sum an overestimate" has a precise answer for clean functions, and a "it depends on the net concavity" answer for messy ones.
If f''(x) < 0 everywhere on [a,b], it's a guaranteed overestimate. If f'' changes sign, you'd need to actually compute or bound the error to know.
A Quick Example
Take f(x) = √(1 − x²) on [0, 1]. The function is concave down the whole time (second derivative is negative). But that's the top-right quarter of a circle. Think about it: midpoint sum with any n? That said, overestimate. You're drawing flat tops under a dome, and they stick out.
Compare to f(x) = x² on [0, 1]. Midpoint underestimates. Concave up. The rectangles tuck under the bowl.
Common Mistakes
Honestly, this is the part most guides get wrong. " It has a bias. They conflate "midpoint is accurate" with "midpoint has no bias.It's just smaller than left/right for smooth functions.
Another mistake: using increasing/decreasing to judge midpoint. Which means midpoint cares about f'' sign. Left and right care about f' sign. No. And if a function is increasing but concave down — say f(x) = ln(x) on [1, 3] — the right sum overestimates, the left underestimates, and the midpoint overestimates (because concave down). People mix those up constantly.
And the big one: assuming more rectangles fixes the direction. On the flip side, more slices makes the overestimate smaller in size, but it's still an overestimate. Halving the error doesn't flip its sign.
Mistaking Trapezoid for Midpoint
The trapezoid rule connects endpoints with a straight line. On concave down, that line sits below the curve, so trapezoid underestimates. Midpoint overestimates. They're backwards. I've seen students "correct" a midpoint overestimate by switching to trapezoid and think they fixed the direction — they did, but only because they flipped to the opposite method, not because they understood why.
Practical Tips
Here's what actually works when you're staring at a problem:
- Check the second derivative first. Before you compute anything, ask: is f'' positive, negative, or mixed? That tells you the direction of midpoint error instantly.
- Sketch the curve. Even a lazy sketch of a frown or a bowl tells you more than a formula. If it arches down, midpoint is too high.
- Don't trust "midpoint = best" on concave down intervals. It's still usually closer than left/right, but it's leaning high.
- Use the error bound. For midpoint, |E| ≤ (b−a)³ / (24n²) · max|f''|. If you know f'' is negative everywhere, drop the absolute value and write E ≈ −(positive)
— meaning the error is negative, so the true integral is below your midpoint estimate.
That last point is worth internalizing: the error bound isn't just a magnitude. Worth adding: the sign of f'' tells you which way the bound leans. If you're only ever reporting "|E| ≤ something," you're throwing away half the information the formula gives you for free.
When None of This Cleanly Applies
Real-world integrands are rarely polite. You might have f(x) = sin(x) + 0.1x² on a wide interval where f'' oscillates. In that case, the midpoint rule produces a sum of local over- and underestimates that partially cancel. Practically speaking, the global error could be tiny and effectively unbiased even though neither "concave up" nor "concave down" describes the whole interval. This is why adaptive quadrature exists: it splits the interval wherever f'' changes character, applies the concave-up/down logic locally, and keeps the bias controlled where it matters.
So the takeaway isn't "midpoint always overestimates" or "midpoint always underestimates.Consider this: " The takeaway is that midpoint error is governed by curvature, not slope, and curvature has a sign. Learn to read that sign, and you'll know not just how wrong your estimate is, but which direction it's wrong in — which is usually the more useful thing to know.
