On this concept: Formula

Not sure which lesson to do first? Find your level and get a plan →

Math · Sets & Logic · Grade 9-12

Sensitivity (Meta)

Sensitivity (meta) measures how strongly a result responds to small changes in inputs, parameters, or assumptions — high sensitivity means a tiny input change causes a big output change.

sensitivity≈ΔoutputΔinput (how much the output changes per unit change in input)

Quick Answer

Sensitivity (meta) measures how strongly a result responds to small changes in inputs, parameters, or assumptions — high sensitivity means a tiny input change causes a big output change. Compute it as the change in output divided by the change in input that caused it. Compare that ratio across inputs to find the one worth measuring carefully, and watch for places where it suddenly becomes large, such as near a double root, because results there are fragile.

Why This Matters

Knowing which input an answer is most sensitive to tells you where to spend effort: measure that one carefully and don't fuss over the rest. It also tells you how far to trust a result. Without it, people polish an input that barely matters, or report a precise-looking answer that a tiny measurement error could overturn.

Intuitive Explanation

A square garden of side s metres has area s2. Nudge the side from 10 m to 10.1 m and the area goes from 100 to 102.01 m²: an output change of 2.01 for an input change of 0.1, or about 20 m² per metre. Do the same starting at 2 m: the area goes from 4 to 4.41, about 4.1 m² per metre. Same formula, different sensitivity. Sensitivity belongs to a point, which is why you compute it where you actually are; for a smooth formula the exact value is the derivative, here 2s.

Sometimes sensitivity explodes. The equation x2−5x+6=0 has roots 2 and 3; change the 6 to 5.99 and the roots move to about 1.990 and 3.010, a shift of 0.01 for a change of 0.01. Now take x2−2x+1=0, with a double root at 1. Change the 1 to 0.99 and the roots become 1±0.01, that is 0.9 and 1.1: a change of 0.01 moved them by 0.1, ten times as much. Near a double root, tiny errors in the coefficients become big errors in the answer.

A raw change is not yet a sensitivity. "The area went up by 2.01" means little until you divide by the 0.1 that caused it; otherwise a big push looks like a sensitive input. Sensitivity is also not the same as robustness: sensitivity is the number, and robustness is the verdict you reach when that number is small for the inputs that might go wrong.

Think of it like this. Sensitivity is the gearing between a dial and a needle: on a low-sensitivity dial a full turn barely moves the needle, while on a high one a hair's turn swings it across the scale. Unlike a fixed gear, the ratio can change with where the dial is set, as it does near a double root.

Core idea. Sensitivity is the ratio of how much the output changes to how much the input changed.


When to Use

Use sensitivity when a problem asks which input matters most, how much an answer changes if a measurement is slightly off, or whether a result is fragile. Nudge one input by a small amount, recompute, and divide the output change by the input change; for a smooth formula, the derivative gives the exact rate. If the question is whether a result survives bad data or broken assumptions, that is robustness; if errors from several measured inputs combine through a calculation, use error analysis.

Tip. Ask: Am I measuring how much the output moves PER unit change in the input?

How to Recognize It

Before measuring sensitivity, pick one output to watch and one input to nudge, and keep every other input fixed while you do it.

  1. Does the problem give a small change in one input, or let you choose one, with the rest held fixed?

    If yes, recompute the output and divide its change by the input change. If several inputs are uncertain at once and you need the total uncertainty of the answer, that is error analysis or error propagation.

  2. Are the inputs measured in different units, such as dollars and percentage points?

    If yes, raw ratios cannot be compared directly; compare percentage changes instead, or the effect of a realistic change in each input. With matching units, the plain ratio of output change to input change is enough.

  3. Is the relationship a smooth formula that you can differentiate?

    If yes, the derivative gives the exact rate: for A=s2, dAds=2s, so at s=10 the area grows about 20 square units per unit of side. For a table or a simulation, use a small finite nudge instead.

  4. Does the question ask whether a conclusion survives bad data, rather than how much it moves?

    If yes, the issue is robustness, a verdict rather than a rate. Sensitivity supplies the evidence: a conclusion with low sensitivity to a doubtful input is robust to it.

Sensitivity (Meta) compared with Robustness, Slope / derivative, and Error propagation

Each row describes how an output answers to changes in its inputs, so the ideas overlap. Ask which you need: a rate of response to one input, a verdict on whether a result holds up, the exact rate of a smooth function, or the total uncertainty from several inputs.

Sensitivity (Meta)

Meaning
Use this when you need to know how much a result responds per unit change in an input, to find which inputs dominate or whether a result is fragile.
Key test
Am I measuring how much the output moves PER unit change in the input?
Formula
sensitivity≈ΔoutputΔinput (how much the output changes per unit change in input)
Example
A profit formula gives profit =5q−200. How sensitive is profit to a change in quantity q?

Robustness

Meaning
The desirable property of surviving violated assumptions; sensitivity is the underlying measure of response.
Key test
Use when judging whether a result holds up, not quantifying its response rate.
Example
Median surviving an outlier

Slope / derivative

Meaning
The exact rate of change of a function, the calculus version of sensitivity for a smooth relationship.
Key test
Use when the relationship is a known function and you want its instantaneous rate.
Formula
dfdx
Example
Rate of f at a point

Error propagation

Meaning
How input errors COMBINE through a multi-step calculation, using sensitivities as ingredients.
Key test
Use when tracking total uncertainty through several operations.
Example
Combining measurement errors in A=lw

Formula & Notation

sensitivity≈ΔoutputΔinput (how much the output changes per unit change in input)
The condition number κ=∣x⋅f′(x)f(x)∣ measures relative sensitivity; κ≫1 means the problem is ill-conditioned

How to read it: Δ denotes a small change; high ΔoutputΔinput means high sensitivity

Worked Examples

Example 1 — Which input matters

Easy

A profit formula gives profit =5q−200. How sensitive is profit to a change in quantity q?
Show the steps and the answer
  1. Pick a starting point, say q=100: profit =5(100)−200=300 dollars.

    Sensitivity compares two nearby outputs, so we need a first one.

  2. Nudge the input: q=101 gives profit =5(101)−200=305 dollars.

    A change of one unit makes the ratio easy to read.

  3. Divide: ΔprofitΔq=305−300101−100=5 dollars per unit.

    The raw change of $5 becomes a sensitivity only after dividing by the input change that caused it.

  4. Try another point: from q=10 to q=12, profit goes from −150 to −140, and 102=5 again.

    The formula is linear, so its sensitivity is the coefficient 5 everywhere; the −200 fixed cost shifts profit but doesn't change the rate.

Answer: Sensitivity =5 dollars per unit

Takeaway: For a linear formula the sensitivity is the coefficient of the input; the constant term moves the output but not how strongly it responds.

Example 2 — Judging survival instead

Standard

You ask whether the profit estimate stays usable if one cost figure is slightly wrong. Is that sensitivity?
Show the steps and the answer
  1. Read what is asked: whether the estimate stays usable, a yes-or-no judgement, not a rate.

    Asking whether a result survives a wrong input is robustness; sensitivity would ask how many dollars profit moves per dollar of cost error.

  2. Compute the sensitivity that feeds the judgement: if the $200 fixed cost is really $210, profit at q=100 drops from $300 to $290.

    Profit falls by one dollar for every dollar of extra cost, a sensitivity of −1.

  3. Judge survival: a $10 error changes profit by about 3%, so the estimate stays usable for planning.

    Robustness is this verdict; it uses the sensitivity but answers a different question.

Answer: It's a robustness question

Takeaway: Sensitivity gives the number (−1 dollar of profit per dollar of cost); robustness is the judgement that a small cost error leaves the estimate usable.

Example 3 — Find the mistake: comparing raw changes

Application

A rectangular field is 50 m long and 10 m wide, so its area is 500 m². Mia adds 5 m to the length and gets 550 m²; separately she adds 0.5 m to the width and gets 525 m². She writes: "Length changed the area by 50, width by only 25, so the area is twice as sensitive to the length." Find the mistake and correct it.
Show the steps and the answer
  1. Check Mia's areas: 55×10=550 and 50×10.5=525.

    The areas are right; the error is in how she compares them.

  2. Divide each change by its own input change: length gives 505=10 m² per metre, width gives 250.5=50 m² per metre.

    Mia pushed the length ten times as far as the width, so her raw changes compare pushes of different sizes.

  3. Explain the numbers: one extra metre of length adds a 1×10 strip, one extra metre of width adds a 50×1 strip.

    Each sensitivity equals the length of the other side, which is why the width, running along the long side, matters more.

  4. Compare: 5010=5, so the area is five times as sensitive to the width.

    A 1 m measuring error in the width would therefore cost five times as much area as the same error in the length.

Answer: Area changes by 50 m² per metre of width but only 10 m² per metre of length: five times more sensitive to the width.

Takeaway: Compare sensitivities only after dividing each output change by the input change that caused it.

Where You'll See It

  • Mortgage rate versus house price

    On a $300,000, 30-year loan at 6% the payment is about $1,799 a month. Raising the rate to 7% adds about $197; adding $10,000 to the price at 6% adds only about $60. So one percentage point of interest costs as much each month as roughly $33,000 more on the price.

  • Weather forecasts and tiny rounding

    In 1961 Edward Lorenz restarted a weather simulation with 0.506 typed in place of 0.506127, and the simulated weather soon looked nothing like the first run. Forecast models are that sensitive to their starting data, so forecasters run many versions with slightly different starts; when the versions disagree, the forecast is uncertain.

Common Mistakes

  1. The slip: Reporting the raw output change as sensitivity

    Instead: divide by the input change to get the ratio.

  2. The slip: Ignoring low-sensitivity inputs entirely

    Instead: they're cheap to estimate, but identifying them is the point of the analysis.

  3. The slip: Confusing sensitivity (the measure) with robustness (the property)

    Instead: high sensitivity is what makes a result NOT robust.


Mini Practice

Try these on your own. Tap Reveal when you want to check.

  1. A taxi fare is F=3+2.5d dollars for a trip of d km. How sensitive is the fare to distance? How much more does a trip 2 km longer cost?

    Hint: Which number tells you the change per extra km?

  2. For a square, A=s2. When s goes from 10 cm to 10.1 cm, find the change in area and the sensitivity ΔAΔs.

    Hint: Work out 10.12 first.

  3. A box has length 10 cm, width 5 cm and height 2 cm, so V=lwh. Increase each dimension by 1 cm on its own. Which one is the volume most sensitive to?

    Hint: Recompute the volume three times, changing one dimension each time.

  4. Revenue is R=p(500−10p) dollars at a ticket price of p dollars. Compare the change in revenue when p rises from 20 to 21 and from 24 to 25.

    Hint: Compute R at all four prices.

  5. Analyst 1 reports: "A 1% rise in fuel price lowers profit by $3,000." Analyst 2 reports: "The plan stays profitable for any fuel price up to 20% higher." Which statement measures sensitivity?

    Hint: Which one gives an amount of change per unit of input?

  6. Raising a price from $8 to $10 cut weekly sales from 500 to 400. Kim reports, "The sensitivity of sales to price is 100." Correct her.

    Hint: Sensitivity is a change per unit of input. By how many dollars did the price change?

Want the full set?

50 practice questions for this concept — free to try, every one with a complete worked solution showing the why, not just the answer.

Frequently Asked Questions

Why do people say a butterfly can cause a hurricane?

It is a picture of extreme sensitivity. In 1961 Edward Lorenz reran a weather simulation after typing a starting value as 0.5060.506 instead of 0.5061270.506127. The tiny difference grew until the two forecasts had nothing in common. Systems like the weather, where small changes in the input are amplified over time, are called chaotic, and they are why detailed forecasts become unreliable after about two weeks.

What is elasticity, and is it the same as sensitivity?

Elasticity is sensitivity measured in percentages, and economists use it constantly. If a 10% price rise cuts sales by 20%, the elasticity is −20%10%=−2\frac{-20\%}{10\%} = -2. Working in percentages removes the units, so you can compare how strongly sales respond to price for bread and for cars, even though their prices differ by thousands of dollars.

Why does a tiny change near a double root move the answer so much?

Take x2=cx^2 = c. At c=0c = 0 there is one root, 0. Change cc to 0.00010.0001 and the roots jump to ±0.01\pm 0.01, a move 100 times the size of the change. Near a double root the graph is flat where it touches the axis, so lifting or lowering it slightly shifts the crossing points a long way sideways. Results computed near such points need extra care.

Should I nudge an input up or down when estimating sensitivity?

Either works for a quick estimate, but trying both is better. For y=x2y = x^2 at x=3x = 3, nudging up by 0.1 gives 9.61−90.1=6.1\frac{9.61 - 9}{0.1} = 6.1, and nudging down gives 9−8.410.1=5.9\frac{9 - 8.41}{0.1} = 5.9. Their average, 6, is exactly the slope 2x2x at that point. If the up and down answers differ a lot, the output responds differently in each direction, and you should report both.

Learning Path

Sensitivity (Meta)

You are here

Next →

Robustness
You need Local vs Global Behavior first: a function can be steep in one region and flat in another, so sensitivity belongs to a point, not to a whole formula. Sensitivity adds a number for that steepness, ΔoutputΔinput, for any input you can nudge. Next, Robustness uses those numbers to judge whether a method or result survives messy data and loosened assumptions.

See Also