Why Is My Oscilloscope Frequency Reading Unstable?

How the Freq readout works

An automatic frequency measurement does not "listen" to the signal the way a radio does. In most scopes it takes the captured record, finds the points where the waveform crosses a reference level, usually 50% of the way between its top and base, measures the time between matching crossings to get the period T, and reports f = 1 ÷ T. Everything that makes those crossings unclear, inconsistent or too few makes the reading jump or turn into dashes. Some scopes also have a separate hardware counter that counts edges from the trigger circuit over a fixed gate time, which behaves a little differently and is covered at the end.

1. 01

   Capture

   Trigger fires, record stored

   →

2. 02

   Find top and base

   Sets the 50% threshold

   →

3. 03

   Find crossings

   Rising edges at the threshold

   →

4. 04

   Time them

   Period T between crossings

   →

5. 05

   f = 1 ÷ T

   Shown, then the next capture

Because the period is what is measured, it helps to think in periods. Frequency and period are reciprocals: f = 1 ÷ T and T = 1 ÷ f. A 16 MHz clock has T = 1 ÷ 16,000,000 = 62.5 ns. A small relative error in T becomes the same relative error in f, so a 1 ns misjudged edge on that clock is 1 ÷ 62.5 = 1.6% of the period, and the reading moves from 16.0 MHz to 1 ÷ 61.5 ns ≈ 16.26 MHz. That is why fast signals at modest sample rates show the most jumpy readings.

Causes, checks and fixes

1Get a stable trigger first

A trace that slides or flickers sideways gives every capture a different set of edges, and some captures catch partial or misaligned cycles. Fix the picture before trusting any number. Set the trigger source to the channel you are measuring, put the level about halfway up the signal, choose a rising edge, and use Normal mode once the level is right so the scope only draws captures that actually triggered. If the trace still will not settle, work through how to stop a waveform drifting, and if the scope never triggers at all, see why an oscilloscope trigger is not working.

2Check the measurement is on the right source

Automatic measurements are attached to a source channel that you pick when you add them, and it does not follow the probe. If you move the probe from CH1 to CH2, the Freq readout may still be timing CH1, which is now an unconnected input producing noise and random values. Look at the label beside the reading (often "Freq(1)" or a colored channel tag) and set it to the probed channel. If a channel shows nothing at all, check why there is no signal on one channel. If the measurement is on a math trace, make sure the math output actually crosses its threshold.

3Put the right number of cycles on screen

With less than one complete cycle visible, there are not two matching crossings to time, so the scope shows dashes or a question mark. Go too far the other way and the reading can become erratic too: if the screen holds hundreds of cycles, each cycle is covered by very few points, especially on scopes that compute measurements from a reduced display record rather than the full memory. As an example, if a measurement worked from about 1,000 points and the screen held 500 cycles, each cycle would get only 1,000 ÷ 500 = 2 points, too few to place an edge.

Aim for roughly two to ten cycles across the ten divisions. For a signal of frequency f, time/div ≈ cycles wanted ÷ (10 × f). For four cycles of a 25 kHz signal: 4 ÷ (10 × 25,000) = 16 µs/div, so the nearest standard setting of 10 µs/div (2.5 cycles) or 20 µs/div (5 cycles) both work.

4Stop noise from double-counting edges

On a fast, clean edge the signal passes through the threshold once. On a slow edge with noise riding on it, the trace can cross the threshold several times in quick succession, and a scope without enough hysteresis counts each crossing as a new edge. The measured period then becomes a fraction of the real one and the frequency reads double, triple or randomly high.

You can estimate how wide the danger zone is. The slope of a sine wave at its zero crossing is 2π × f × A, where A is its peak amplitude. A 1 kHz sine with A = 0.5 V rises at 2π × 1,000 × 0.5 ≈ 3,140 V/s, or 3.14 mV per microsecond. With 20 mV of peak-to-peak noise on it, the trace sits within the noise of the threshold for about 20 ÷ 3.14 ≈ 6.4 µs on every edge, plenty of time for several false crossings. Double the amplitude and that window halves; a square wave's fast edges barely have one at all.

Fixes, roughly in order: make the signal fill more of the screen vertically so the scope's threshold logic has more resolution; turn on noise reject or HF reject in the trigger menu, which adds hysteresis or filtering to the trigger path; switch on the bandwidth limit if the signal is slow enough to pass through it; and add averaging. Shortening the probe ground and using a 10x probe also remove a lot of pickup, as explained in why an oscilloscope waveform looks noisy.

5Check the sample rate at slow timebases

The time resolution of a waveform measurement is set by the interval between samples. At 1 GSa/s samples are 1 ns apart; at 250 MSa/s they are 4 ns apart. Without interpolation, one sample of error on a single 1 µs period is 1 ÷ 1,000 = 0.1% at 1 GSa/s but 4 ÷ 1,000 = 0.4% at 250 MSa/s. Most scopes interpolate between samples, so real results are usually better, but the principle holds: fewer samples per period means a coarser, jumpier reading.

Worst-case resolution for a single-period measurement with no interpolation; illustrative arithmetic, not any one scope's specification. Remember that at slow time/div settings a scope reduces its sample rate to fit the window into memory, and turning on more channels can divide the maximum rate on many models. If the reading only goes wrong at slow timebases, check the sample rate shown on screen; if it has fallen below about twice the signal frequency, the scope is measuring an alias. What GSa/s means covers the memory and sample-rate trade-off in more depth.

6Use DC coupling for low-frequency signals

AC coupling puts a capacitor in the input path, forming a high-pass filter whose cutoff is typically a few hertz to around 10 Hz, depending on the scope. Signals near that cutoff are attenuated and phase shifted, so the waveform drifts up and down as its content changes and its crossings of the 50% threshold wander. Pulse trains are affected even at higher frequencies: AC coupling centers the average on zero, so when the duty cycle changes, the whole trace moves and the measured edges move with it. For anything below a few hundred hertz, and for pulses whose duty cycle varies, switch the channel to DC coupling.

7Average the reading and use statistics

Once the setup is right, use the measurement statistics view (mean, minimum, maximum and standard deviation) instead of staring at a single live number. Averaging independent readings shrinks the random part of the spread by the square root of the number averaged. If single readings of a 10 kHz signal scatter with a standard deviation of 40 Hz (0.4%), the mean of 16 readings has a standard deviation of about 40 ÷ √16 = 10 Hz (0.1%), and of 100 readings about 40 ÷ 10 = 4 Hz (0.04%).

Spread (one standard deviation) of a 10 kHz frequency reading as more readings are averaged, starting from an illustrative ±40 Hz. Measuring across several periods on one capture helps the same way: a 1 ns timing error spread over ten 1 µs periods is 1 ÷ 10,000 = 0.01% instead of 0.1%. Averaging only removes random scatter, not a systematic error such as double-counting or a wrong source, so fix those first.

How much variation is normal?

After the setup is fixed, some movement may be the signal itself. Oscillators are only as stable as their timing element, and some signals are deliberately varied. Compare the spread you see with the typical tolerance of the source; the figures below are typical published ranges and vary by part, so check the datasheet for the exact component in your circuit.

Typical frequency tolerance by oscillator type, in parts per million (log scale); upper end of each range shown. Worked examples: a 16 MHz crystal at ±50 ppm may sit anywhere within 16,000,000 × 50 ÷ 1,000,000 = ±800 Hz, so a reading of 16.0006 MHz is perfectly normal, while 16.3 MHz is not a crystal tolerance issue. A 16 MHz ceramic resonator at ±0.5% may be ±80 kHz off. A microcontroller's internal 8 MHz RC oscillator at ±2% may read anywhere from 7.84 to 8.16 MHz and drift as it warms up.

To see real variation edge by edge rather than as a jumpy number, trigger on one edge and watch a later edge spread with persistence on, or use period statistics; the method is in how to compare clock jitter. For crystals specifically, see how to check a crystal oscillator.

Hardware counter vs waveform measurement

A waveform-based Freq measurement times the crossings in one capture, so its resolution depends on the sample interval, the number of periods on screen and any interpolation. On a clean signal with several cycles visible it is commonly good to a fraction of a percent, which is plenty for checking that a PWM, a UART clock or an oscillator is in the right place.

Some scopes also include a hardware frequency counter that counts trigger events over a gate time, often displaying five or six digits. A counter that gates over 100,000 cycles of a 1 MHz signal sees a 0.1 s window, and a one-count uncertainty over that window is 1 ÷ 100,000 = 0.001%, far finer than one capture can resolve. Its absolute accuracy is then limited by the scope's own timebase reference, which on general-purpose scopes is commonly specified in tens of parts per million; check your model's datasheet for the exact figure. Because the counter works from the trigger comparator, it still needs a clean, stable trigger, and noise causes the same double-counting.

For the hand method of reading frequency from the graticule and cross-checking the automatic value, see how to measure frequency with an oscilloscope.