Bandgap Curvature Correction

Bandgap curvature correction is the fix for the one thing a standard bandgap voltage reference does not actually get right: the basic bandgap topology produces a voltage that is only approximately constant over temperature, flat to first order but still bowing by a few millivolts across the range, because the two physical quantities it adds together don't cancel with equal and opposite curves — only with equal and opposite slopes.

The physics behind the residual curvature traces back to a single term inside the base-emitter voltage equation itself. A bipolar transistor's base-emitter voltage decreases roughly linearly with temperature — this is the complementary-to-absolute-temperature, or CTAT, behavior the standard bandgap reference relies on. But the full expression for V_BE contains not just that linear term but also a logarithmic temperature term inherited from the transistor's saturation current, and that term is exactly what a simple proportional-to-absolute-temperature, or PTAT, correction — itself perfectly linear — cannot cancel. The standard bandgap circuit sums V_BE and a scaled PTAT voltage to flatten the linear slope at one chosen temperature; the logarithmic term in V_BE has no counterpart in the correction, so it survives as a smooth, parabola-like bow in the combined output across the rest of the temperature range.

This residual bow, typically tens of parts per million per degree Celsius uncorrected, is often the single largest error term left in an otherwise well-designed voltage reference. Device mismatch, resistor matching, and layout-level techniques like common-centroid placement — covered elsewhere in this series — address random and systematic spatial errors, but none of them touch this specific temperature-domain nonlinearity, because it is not a matching problem at all; it is baked into the physics of the bipolar device itself and appears identically even in a perfectly matched, perfectly laid-out circuit. A reference intended for precision data converters or instrumentation, where output drift over the full operating temperature range directly limits achievable accuracy, cannot simply be trimmed better to remove it — trimming only adjusts the flat point, not the shape of the bow around it.

Curvature correction works by generating a term with the same functional shape as the error and adding it back with the opposite sign, rather than by changing the underlying devices. One common technique exploits that the troublesome logarithmic term in V_BE itself depends on a ratio involving absolute temperature, so a circuit can be built to generate a current proportional to the logarithm of a temperature-dependent quantity — reproducing the same curved shape the uncorrected bandgap suffers from, but with a sign chosen to cancel it rather than add to it. Other circuit techniques achieve the same goal differently — piecewise-linear correction using additional current sources that turn on only away from the trim temperature, or exploiting a second-order nonlinearity in a different device to supply a complementary curve — but all of them share the same underlying strategy: match the shape of the error, not just its slope at a single point.

The payoff is measured directly in parts per million, and it is substantial relative to what an uncorrected design achieves. A well-designed but uncorrected bandgap reference typically holds its output to within roughly 20 to 50 parts per million per degree Celsius across a commercial temperature range; adding curvature correction routinely pushes that down to the low single digits of parts per million per degree, an order-of-magnitude improvement that comes entirely from canceling this one specific nonlinearity rather than from better matching, better trimming, or a fundamentally different reference topology. This is why curvature-corrected bandgap references are the default choice, not an exotic option, in any converter or instrumentation system where temperature-stable voltage accuracy is a first-order specification.

Bandgap curvature correction is proof that a circuit's first-order fix can be exactly right at one point and still wrong everywhere else — the standard PTAT-plus-CTAT bandgap genuinely cancels a slope, but a slope is not a shape, and only a term built to match the error's actual curve closes that remaining gap. It is the temperature-domain counterpart to this series' other correction techniques: where common-centroid layout cancels a spatial gradient and autozeroing cancels a time-domain offset, curvature correction cancels a thermal nonlinearity, each one matching its fix to the specific shape of the error it targets rather than applying a one-size-fits-all solution.