Why a Tangent Line Is the Best Local Linear Model

Most students assume the tangent line is the best straight-line fit over any interval. That is wrong. The tangent line is the best local linear model at exactly one point. It matches the function's value and its instantaneous slope at that base point. No other line can do both. The reason why a tangent line is the best local linear model is baked into its construction: it shares the same output and the same rate of change at the anchor, and that dual match gives the closest possible linear fit right there. Move away from the point and the error grows. The model is local, not global.

At x = a, the tangent line L(x) = f(a) + f'(a)(x - a) satisfies L(a) = f(a) and L'(a) = f'(a). These two conditions lock the line to the curve. Any other line passing through (a, f(a)) would have a different slope, so its value would diverge immediately. The tangent line is the unique linearization that achieves this match. That is what makes it the best local linear model.

Take f(x) = x² with a = 3. f(3) = 9 and f'(3) = 6. The tangent line is L(x) = 9 + 6(x - 3). At x = 3, L(3) = 9 and L'(3) = 6. Both match exactly. Now test a secant line through x = 3 and x = 4, slope (16 - 9)/(4 - 3) = 7, equation S(x) = 9 + 7(x - 3). At x = 3, S(3) = 9 but S'(3) = 7, not 6. The secant line shares the value but not the instantaneous slope. The tangent line is the only line that hits both.

At x = 3.1, L(3.1) = 9 + 6(0.1) = 9.6, while f(3.1) = 9.61. The error is 0.01. The secant line gives S(3.1) = 9 + 7(0.1) = 9.7, error 0.09. The tangent line is nine times closer at 0.1 away from the base point. That gap widens as you move further. At x = 4, L(4) = 15, f(4) = 16, error 1. The secant line through 3 and 4 gives S(4) = 16 exactly, but only because it was built to pass through both endpoints. Over a single interval a secant can beat the tangent. Locally, at the base point, the tangent wins.

To verify the dual match yourself for f(x) = √x at a = 4: f(4) = 2, f'(4) = 1/(2√4) = 0.25. L(x) = 2 + 0.25(x - 4). L(4) = 2, L'(4) = 0.25. At x = 4.04, L(4.04) = 2 + 0.25(0.04) = 2.01, f(4.04) = √4.04 ≈ 2.009975, error ≈ 0.000025. The tangent line calculator will confirm that result in one step. The calculation shows the tangent line delivers a close approximation because it was built from the exact value and the exact slope at the base point. The formula L(x) = f(a) + f'(a)(x - a) is the first-order Taylor polynomial. Taylor's theorem gives the error: E(x) = (1/2)f''(ξ)(x - a)² for some ξ between a and x. For √x at a = 4, f''(x) = -1/(4x^(3/2)), so f''(4) = -1/(4 * 8) = -1/32. For x = 4.04, the error bound is |E(4.04)| ≤ (1/2)(1/32)(0.04)² = (1/64)(0.0016) = 0.000025, which matches the actual error. That tight bound comes from the second derivative being small near a.

The explanation rests on two conditions. Condition one: L(a) = f(a). The tangent line passes through the same point as the curve. Condition two: L'(a) = f'(a). The tangent line has the same instantaneous slope as the curve at that point. No other line can satisfy both simultaneously. A line through (a, f(a)) with a different slope will diverge immediately. A line with the same slope but a different intercept will not pass through the point. The tangent line is the unique intersection of these two constraints.

This dual match makes the tangent line the best linear approximation in the sense of Taylor's theorem. The error grows with the square of the distance from a. The constant in that quadratic term is tied to the second derivative. A smaller second derivative at a means a flatter curve and a tighter error bound. The tangent line is the linearization used in calculus because it captures the first-order behaviour exactly. That is why it is the best local linear model.

The sign of f''(a) tells you whether L(x) is an overestimate or an underestimate near a. If f''(a) > 0, the function is concave up, the tangent line lies below the curve, and L(x) is an underestimate. If f''(a) < 0, concave down, the line lies above, and L(x) is an overestimate. For f(x) = x² at a = 3, f''(3) = 2 > 0, so L(3.1) = 9.6 is below f(3.1) = 9.61. Underestimate. For f(x) = √x at a = 4, f''(4) = -1/32 < 0, so L(4.04) = 2.01 is above f(4.04) ≈ 2.009975. Overestimate.

Students confuse the direction. The rule is simple f'' sign positive means underestimate. Negative means overestimate. Memorise that. The concavity test applies only near the base point. Far from a, the curve may change concavity, flipping the direction.

The most instructive failure case is f(x) = |x| at a = 0. This function is not differentiable at 0. There is no tangent line, no linearization, no local linear model of the form L(x) = f(a) + f'(a)(x - a). The limit of the difference quotient from the left is -1, from the right is +1. They do not agree. The tangent line does not exist. Anyone trying to apply linearisation to |x| at 0 gets no result. The condition for a tangent line is differentiability at the base point.

Even when a tangent line exists, the model is local. The error at distance d is roughly (1/2)|f''(ξ)| d². For f(x) = x² at a = 3, the error at x = 5 is L(5) = 9 + 6(2) = 21, f(5) = 25, error 4. That is 0.16 times f(5). For x = 10, error 9 + 6(7) = 51, f(10) = 100, error 49, nearly half. The quadratic growth makes the linearisation useless far from a. The single thing that most often goes wrong is using the linearisation too far from the base point without checking the error bound.

The tangent line is the best local linear model only at exactly one point. Over any interval longer than zero, a secant line can beat it on total error. The error bound from Taylor's inequality is a guarantee, but the bound itself may be loose if you use a global maximum of f'' instead of the maximum over the subinterval. Always compute the bound for your specific interval, not a generic one.