A curved graph looks almost straight when you zoom in near a differentiable point. Linear approximation turns that observation into a calculation: use the tangent line as a nearby substitute for the function.
Build the line from a value and a slope
At the point is and the tangent slope is The point-slope equation gives
The quantity is the change in input. Multiplying it by the local rate estimates the change in output. Adding restores the starting value. This explains why alone is usually not the tangent line.
Differentiability at a is essential. At a corner, such as at zero, there is no single finite tangent slope to use. Even for a smooth function, the approximation is local: moving farther from a can make it inaccurate.
Example: estimate a square root
To estimate choose and This anchor is close to 4.1 and has an easy exact value. We have and so
The tangent model is Therefore
The exact value is about 2.02485. The approximation is slightly high. Since for positive , the graph is concave down and lies below its tangent line near 4. Concavity tells you the direction of the error without requiring the exact square root.
- (dashed)
tangent
The curve y = sqrt(x) and its tangent at (4, 2) are shown. At x = 4.1 the tangent estimate is slightly above the curve.
Near , the tangent closely tracks the square-root curve and lies slightly above it.
Describe this figure
The curve and its tangent at (4, 2) are shown. At the tangent estimate is slightly above the curve.
Curves: ; tangent.
Marked points: anchor: filled marker at (4, 2).
Example: interpret a tangent model with units
Suppose a tank contains liters after 3 minutes and liters per minute. Near that time,
At 3.2 minutes, this predicts liters. The negative derivative means the amount is decreasing; it does not mean the volume itself is negative. The product of liters per minute and minutes has units of liters, as the model requires.
You are not claiming that the tank drains at exactly 8 liters per minute forever. The derivative describes the instantaneous rate at the anchor time. A long-term prediction would need further information about the rate.
Differentials describe the same estimate
Writing the estimated output change is The actual change is In general, and are close for small input changes, but they are not equal.
For the square-root example, and That is the estimated increase, not the estimated final square root. Confusing a change with a final value is a common error.
If the function is concave up throughout the relevant interval, its tangent approximation is an underestimate; if concave down, it is an overestimate. If concavity changes, inspect the interval before making either claim.
Use an error estimate when more information is known
Concavity gives the direction of error, but not its size. If between the anchor a and the target , the linear approximation error is at most This is the degree-one Taylor remainder bound. Halving the distance from the anchor makes this bound one quarter as large when the same M applies.
For the square-root example on The error is therefore at most The actual error, about 0.000154, fits inside that guarantee. This distinguishes a justified accuracy statement from simply reporting many calculator digits.
Try it yourself
Use a tangent line at to estimate Pause and identify the anchor value, derivative, and input change.
Answer: For and Thus The cube-root graph is concave down for positive , so this is an overestimate.
What to remember
A tangent approximation uses both the function value and its derivative at an easy nearby anchor. Keep the input change, output change, and final value distinct, and use concavity to assess the error.