Your 2.2 µF is 0.68 µF: DC bias and the capacitance you did not buy

A capacitor’s value is measured at essentially zero volts. Then you solder it to a rail and apply the voltage it was sold for, and — if it is a class 2 ceramic — a large fraction of the capacitance goes away.

Not drifts. Goes away, immediately, as a function of the applied DC voltage, and comes back when you remove it. The part is not faulty and the datasheet is not lying. The number on the reel is simply measured under conditions your circuit never reproduces.

Which capacitors do this

Ceramics divide into two classes, and only one of them has the problem.

	Class 1	Class 2
Typical codes	C0G / NP0	X5R, X7R, Y5V
Dielectric	Titanium dioxide	Barium titanate
Permittivity	~10–500	~500–10 000
Voltage dependent	No	Yes
Temperature dependent	No	Yes
Ages	No	Yes

Those properties come from Würth’s summary of the ceramics they use (ANP062). The trade is visible in the permittivity row: class 2 gets its enormous capacitance-per-volume from barium titanate, a ferroelectric whose permittivity is a function of the field across it. Class 1 has none of these behaviours and also, per unit volume, hardly any capacitance. You cannot buy a 10 µF C0G in an 0805.

Aluminium and aluminium-polymer electrolytics show no DC bias effect either.

How much you lose

Würth measured this on real parts. The number that should stick:

An 0805, 2.2 µF, 25 V X7R capacitor, at its full rated voltage, is reduced by 69 % — to 0.68 µF.

That is not an abusive test. It is the part operated exactly at the voltage printed on it. You specified 2.2 µF, you are running it inside its rating, and you have 0.68 µF.

Worse, from the same note: a 47 µF 0805 falls to about 10 µF under bias, an 80 % loss. Their explanation is the mechanism you need in order to predict this yourself.

The rule that lets you predict it

The loss depends on how much ceramic is left between the electrodes.

To get 47 µF into an 0805, the manufacturer stacks a very large number of very thin dielectric layers of very high permittivity material. Thin layers mean a large field for a given applied voltage, and high-permittivity barium titanate saturates hard once the field gets large. The part runs out of dielectric early.

Take the same 0805 case and ask for 2.2 µF at 6.3 V instead, and Würth’s measurement shows the capacitance barely moving at maximum voltage — there is plenty of ceramic between the layers, so it does not need to be highly permeable, and the effect hardly appears.

So, for a fixed case size:

• Higher rated capacitance → worse DC bias loss. You are buying thinner dielectric.
• Higher rated voltage → better. Thicker dielectric for the same value.
• Bigger case → better. More room for both.

Every one of those is the opposite of the direction a BOM cost review pushes you.

What this means for a 100 nF decoupler

Mostly, it does not. A 100 nF X7R in an 0402 rated at 16 V or 50 V, sitting on a 3.3 V rail, is nowhere near the geometry that causes trouble — it is a low-capacitance part in a package with dielectric to spare, running at a fraction of its rating. It is worth knowing about, not worth worrying about.

The place it will hurt you is bulk. The 10 µF and 22 µF parts in small cases that you put next to the regulator, or that the reference design specified, are exactly the high-value-in-a-small-case parts that lose most. A power stage designed around 4 × 22 µF and built with 0805 parts at their rated voltage can easily be running on a third of the capacitance the loop compensation assumed. That shows up as an unstable regulator or a rail that sags on load steps, and it looks like a control-loop problem rather than a component problem.

The same applies to the target-impedance sums in the decoupling calculator: if you enter the printed value for a bulk part, the low-frequency end of the curve is optimistic. Enter the biased value.

What to do

1. Get the curve, not the number. Every serious manufacturer publishes capacitance versus DC bias. Murata’s SimSurfing, Würth’s REDEXPERT and TDK’s equivalents all give it per part number. Look up the actual part, at your actual rail voltage.
2. Derate on purpose. Choose the part so that the biased capacitance meets your requirement, not the nominal one. In practice that often means one voltage rating up, or one case size up.
3. Use class 1 where the value must be exact. Timing, filters, oscillator load capacitors, anything where a 30 % error changes the answer — C0G, and pay for the volume.
4. Consider polymer for bulk. Aluminium-polymer has no DC bias effect and often no worse ESR, and Würth make the point that it can be cheaper than a very high-capacitance ceramic for the same delivered microfarads.

And remember there are two more class 2 behaviours stacked on top of this one: capacitance also moves with temperature (the X7R and X5R codes bound only that variation, over their stated range) and class 2 parts age, losing capacitance logarithmically with time since their last thermal excursion. The capacitor code decoder spells out what each dielectric code actually promises.

None of those three appear in “100 nF ±10 %”.

Sources

• Würth ANP062 — LC Filter Design With MLCCs: Why The Applied Voltage Matters — the class 1 / class 2 property table and the measured 69 % and 80 % losses quoted above.
• Würth ANP109 — Impedance Spectra of Different Capacitor Technologies — how the families compare once you plot impedance rather than capacitance.
• KEMET X7R SMD MLCC datasheet — temperature, DC bias and ageing curves for a specific range.
• IEC 60062:2016 — Marking codes for resistors and capacitors — what the dielectric and tolerance codes are defined to mean.