Digital resolution filters

The digital resolution filters used in Keysight spectrum analyzers have an effect on
sweep time that is different from the effects we’ve just discussed for analog filters. For More information:
swept analysis, the speed of digitally implemented filters, with no further processing, A more detailed
can show a two to four times improvement. discussion about fast
sweep measurements
However, the X-Series signal analyzers with Option FS1 are programmed to correct can be found in Using
for the effect of sweeping too fast for resolution bandwidths between about 3 kHz and Fast-Sweep Techniques
to Accelerate Spur
100 kHz. As a result, sweep times that would otherwise be many seconds may be
Searches – Application
reduced to milliseconds, depending upon the particular settings. See Figure 2-15. The Note, literature number
sweep time without the correction would be 79.8 seconds (2-15a). Installation of Option 5991-3739EN
FS1 would show a sweep time of 1.506 s (2-15b). For the widest resolution bandwidths,
sweep times are already very short. For example, using the formula with k = 2 on a span
of 1 GHz and a RBW of 1 MHz, the sweep time calculates to just 2 msec.

For narrower resolution bandwidths, analyzers such as the Keysight X-Series use fast
Fourier transforms (FFTs) to process the data, also producing shorter sweep times
than the formula predicts. The difference occurs because the signal being analyzed is
processed in frequency blocks, depending upon the particular analyzer. For example, if
the frequency block was 1 kHz, then when we select a 10-Hz resolution bandwidth, the
analyzer is in effect simultaneously processing the data in each 1-kHz block through 100
contiguous 10-Hz filters. If the digital processing were instantaneous, we would expect
sweep time to be reduced by a factor of 100. In practice, the reduction factor is less,
but is still significant. For more information on the advantages of digital processing, refer
to Chapter 3.

Figure 2-15a. Full span sweep speed, RBW of 20 kHz, Figure 2-15b. Full span sweep speed, RBW of 20 kHz, with
without Option FS1 Option FS1

Find us at www.keysight.com Page 27

Envelope detector 6
Older analyzers typically converted the IF signal to video with an envelope detector 7. In
its simplest form, an envelope detector consists of a diode, resistive load and low-pass More information:
filter, as shown in Figure 2-16a. The output of the IF chain in this example, an amplitude Additional information on
modulated sine wave, is applied to the detector. The response of the detector follows envelope detectors can
be found in Spectrum
the changes in the envelope of the IF signal, but not the instantaneous value of the IF
and Signal Analyzer
sine wave itself. Measurements and Noise–
Application Note, literature
For most measurements, we choose a resolution bandwidth narrow enough to resolve
number 5966-4008E.
the individual spectral components of the input signal. If we fix the frequency of the LO
so that our analyzer is tuned to one of the spectral components of the signal, the output
of the IF is a steady sine wave with a constant peak value. The output of the envelope
detector will then be a constant (DC) voltage, and there is no variation for the detector
to follow.

However, there are times when we deliberately choose a resolution bandwidth wide
enough to include two or more spectral components. At other times, we have no
choice. The spectral components are closer in frequency than our narrowest bandwidth.
Assuming only two spectral components within the pass band, we have two sine waves
interacting to create a beat note, and the envelope of the IF signal varies, as shown in
Figure 2-16b, as the phase between the two sine waves varies.

t t

IF signal

Figure 2-16a. Envelope detector

Figure 2-16b. Output of the envelope detector follows the peaks of the IF signal

6. The envelope detector should not be confused with the display detectors. See “Detector types” later in this chapter.
7. A signal whose frequency range extends from zero (DC) to some upper frequency determined by the circuit elements.
Historically, spectrum analyzers with analog displays used this signal to drive the vertical deflection plates of the CRT
directly. Hence it was known as the video signal.

Find us at www.keysight.com Page 28

The width of the resolution (IF) filter determines the maximum rate at which the envelope
of the IF signal can change. This bandwidth determines how far apart two input
sinusoids can be so that after the mixing process they will both be within the filter at
the same time. Let’s assume a 22.5-MHz final IF and a 100-kHz bandwidth. Two input
signals separated by 100 kHz would produce mixing products of 22.45 and 22.55 MHz
and would meet the criterion. See Figure 2-16b. The detector must be able to follow
the changes in the envelope created by these two signals but not the 22.5-MHz IF
signal itself.

The envelope detector is what makes the spectrum analyzer a voltmeter. Let’s duplicate
the situation above and have two equal-amplitude signals in the pass band of the IF at
the same time. A power meter would indicate a power level 3 dB above either signal,
that is, the total power of the two. Assume that the two signals are close enough so
that, with the analyzer tuned half-way between them, there is negligible attenuation due
to the roll-off of the filter  8.

The analyzer display will vary between a value that is twice the voltage of either (6 dB
greater) and zero (minus infinity on the log scale). We must remember that the two
signals are sine waves (vectors) at different frequencies, and so they continually change
in phase with respect to each other. At some time they add exactly in phase; at another,
exactly out of phase.

So the envelope detector follows the changing amplitude values of the peaks of the
signal from the IF chain but not the instantaneous values, resulting in the loss of phase
information. This gives the analyzer its voltmeter characteristics.

Digitally implemented resolution bandwidths do not have an analog envelope detector.

Instead, the digital processing computes the root sum of the squares of the I and Q
data, which is mathematically equivalent to an envelope detector. For more information
on digital architecture, refer to Chapter 3.

Displays
Up until the mid-1970s, spectrum analyzers were purely analog. The displayed trace
presented a continuous indication of the signal envelope, and no information was lost.
However, analog displays had drawbacks. The major problem was in handling the long
sweep times required for narrow resolution bandwidths. In the extreme case, the display
became a spot that moved slowly across the cathode ray tube (CRT), with no real trace
on the display. So a meaningful display was not possible with the longer sweep times.

8. For this discussion, we assume the filter is perfectly rectangular.

Find us at www.keysight.com Page 29

Keysight (part of Hewlett-Packard at the time) pioneered a variable-persistence storage
CRT in which we could adjust the fade rate of the display. When properly adjusted,
the old trace would just fade out at the point where the new trace was updating the
display. This display was continuous, had no flicker and avoided confusing overwrites. It
worked quite well, but the intensity and the fade rate had to be readjusted for each new
measurement situation.

When digital circuitry became affordable in the mid-1970s, it was quickly put to use
in spectrum analyzers. Once a trace had been digitized and put into memory, it was
permanently available for display. It became an easy matter to update the display at a
flicker-free rate without blooming or fading. The data in memory was updated at the
sweep rate, and since the contents of memory were written to the display at a flicker-
free rate, we could follow the updating as the analyzer swept through its selected
frequency span just as we could with analog systems.

Detector types
With digital displays, we had to decide what value should be displayed for each display
data point. No matter how many data points we use across the display, each point must
represent what has occurred over some frequency range and time interval (although we
usually do not think in terms of time when dealing with a spectrum analyzer).

It is as if the data for each interval is thrown into a bucket and we apply whatever math
is necessary to extract the desired bit of information from our input signal. This datum is
put into memory and written to the display. This process provides great flexibility.

Here we will discuss six different detector types.

In Figure 2-17, each bucket contains data from a span and timeframe that is determined by
these equations:

Frequency: bucket width = span/(trace points – 1)

Time: bucket width = sweep time/(trace points – 1)

The sampling rates are different for various instruments, but greater accuracy is
obtained from decreasing the span or increasing the sweep time because the number
of samples per bucket will increase in either case. Even in analyzers with digital
IFs, sample rates and interpolation behaviors are designed to be the equivalent of
continuous-time processing.

Find us at www.keysight.com Page 30

Figure 2-17. Each of the 1001 trace points (buckets) covers a 100-kHz frequency span
and a 0.01-millisecond time span

The “bucket” concept is important, as it will help us differentiate the six detector types:

• Sample
• Positive peak (also simply called peak)
• Negative peak
• Normal
• Average
• Quasipeak

Find us at www.keysight.com Page 31

The first three detectors, sample, peak, and negative peak are easy to understand
and are visually represented in Figure 2-18. Normal, average, and quasipeak are more
complex and will be discussed later.

One bucket

Positive peak

Sample

Negative peak

Figure 2-18. The trace point saved in memory is based on the detector type algorithm

Let’s return to the question of how to display an analog system as faithfully as possible
using digital techniques. Let’s imagine the situation illustrated in Figure 2-19. We have a
display that contains only noise and a single CW signal.

Figure 2-19. When digitizing an analog signal, what value should be displayed at each point?

Find us at www.keysight.com Page 32

Sample detection
As a first method, let us simply select the data point as the instantaneous level at the
center of each bucket (see Figure 2-18). This is the sample detection mode. To give the
trace a continuous look, we design a system that draws vectors between the points.
Comparing Figure 2-19 with 2-20, it appears that we get a fairly reasonable display. Of
course, the more points there are in the trace, the better the replication of the analog
signal will be. The number of available display points can vary for different analyzers. On
X-Series signal analyzers, the number of display points for frequency domain traces can
be set from a minimum of 1 point to a maximum of 40,001 points. As shown in Figure
2-21, more points do indeed get us closer to the analog signal.

Figure 2-20. Sample display mode using Figure 2-21. More points produce a display
10 points to display the signal shown in closer to an analog display
Figure 2-17

While the sample detection mode does a good job of indicating the randomness
of noise, it is not a good mode for analyzing sinusoidal signals. If we were to look
at a 100-MHz comb on a high-performance X-Series signal analyzer, we might
set it to span from 0 to 26.5 GHz. Even with 1,001 display points, each display point
represents a span (bucket) of 26.5 MHz. This is far wider than the maximum 8-MHz
resolution bandwidth.

Find us at www.keysight.com Page 33

As a result, the true amplitude of a comb tooth is shown only if its mixing product
happens to fall at the center of the IF when the sample is taken. Figure 2-22a shows
a 10-MHz span with a 750-Hz bandwidth using sample detection. The comb teeth
should be relatively equal in amplitude, as shown in Figure 2-22b (using peak detection).
Therefore, sample detection does not catch all the signals, nor does it necessarily reflect
the true peak values of the displayed signals. When resolution bandwidth is more narrow
than the sample interval (the bucket width), sample mode can give erroneous results.

Figure 2-22a. A 10-MHz span of a 250-kHz comb in the Figure 2-22b. The actual comb over a 10-MHz span
sample display mode using peak (positive) detection

Peak (positive) detection

One way to insure that all sinusoids are reported at their true amplitudes is to display
the maximum value encountered in each bucket. This is the positive peak detection
mode, or peak. This mode is illustrated in Figure 2-22b. Peak is the default mode
offered on many spectrum analyzers because it ensures that no sinusoid is missed,
regardless of the ratio between resolution bandwidth and bucket width. However, unlike
sample mode, peak does not give a good representation of random noise because it
only displays the maximum value in each bucket and ignores the true randomness of the
noise. So spectrum analyzers that use peak detection as their primary mode generally
also offer sample mode as an alternative.

Find us at www.keysight.com Page 34

Negative peak detection
Negative peak detection displays the minimum value encountered in each bucket. It is
generally available in most spectrum analyzers, though it is not used as often as other
types of detection. Differentiating CW from impulsive signals in EMC testing is one
application where negative peak detection is valuable. Later in this application note, we
will see how negative peak detection is also used in signal identification routines when
you use external mixers for high-frequency measurements.

Normal detection
To provide a better visual display of random noise than offered by peak mode and
yet avoid the missed-signal problem of the sample mode, the normal detection mode
(informally known as Rosenfell9 mode) is offered on many spectrum analyzers. Should
the signal both rise and fall, as determined by the positive peak and negative peak
detectors, the algorithm classifies the signal as noise.

In that case, an odd-numbered data point displays the maximum value encountered
during its bucket. And an even-numbered data point displays the minimum value
encountered during its bucket. See Figure 2-25. Normal and sample modes are
compared in Figures 2-23a and 2-23b.10

Figure 2-23a. Normal mode Figure 2-23b. Sample mode

9. Rosenfell is not a person’s name but rather a description of the algorithm that tests to see if the
signal rose and fell within the bucket represented by a given data point. It is also sometimes
written as “rose’n’fell.”
10. Because of its usefulness in measuring noise, the sample detector is usually used in “noise
marker” applications. Similarly, the measurement of channel power and adjacent-channel power
requires a detector type that gives results unbiased by peak detection. For analyzers without
averaging detectors, sample detection is the best choice.

Find us at www.keysight.com Page 35

What happens when a sinusoidal signal is encountered? We know that as a mixing
product is swept past the IF filter, an analyzer traces out the shape of the filter on
the display. If the filter shape is spread over many display points, we encounter a
situation in which the displayed signal only rises as the mixing product approaches
the center frequency of the filter and only falls as the mixing product moves away from
the filter center frequency. In either of these cases, the positive-peak and negative-
peak detectors sense an amplitude change in only one direction, and, according to
the normal detection algorithm, the maximum value in each bucket is displayed. See
Figure 2-24.

Figure 2-24. Normal detection displays maximum values in buckets where the
signal only rises or only falls

Find us at www.keysight.com Page 36

Figure 2-25. Trace points selected by the normal detection algorithm

What happens when the resolution bandwidth is narrow, relative to a bucket? The signal
will both rise and fall during the bucket. If the bucket happens to be an odd-numbered
one, all is well. The maximum value encountered in the bucket is simply plotted as the
next data point. However, if the bucket is even-numbered, then the minimum value in the
bucket is plotted. Depending on the ratio of resolution bandwidth to bucket width, the
minimum value can differ from the true peak value (the one we want displayed) by a little
or a lot. In the extreme, when the bucket is much wider than the resolution bandwidth,
the difference between the maximum and minimum values encountered in the bucket
is the full difference between the peak signal value and the noise. This is true for the
example in Figure 2-25. See bucket 6. The peak value of the previous bucket is always
compared to that of the current bucket. The greater of the two values is displayed if
the bucket number is odd, as depicted in bucket 7. The signal peak actually occurs in
bucket 6 but is not displayed until bucket 7.

Find us at www.keysight.com Page 37

The normal detection algorithm:
If the signal rises and falls within a bucket, then even-numbered buckets display the
minimum (negative peak) value in the bucket. The maximum is remembered. Odd-
numbered buckets display the maxi(positive peak) value determined by comparing the
current bucket peak with the previous (remembered) bucket peak. If the signal only
rises or only falls within a bucket, the peak is displayed. See Figure 2-25.

This process may cause a maximum value to be displayed one data point too far to
the right, but the offset is usually only a small percentage of the span. Some spectrum
analyzers, such as high-performance X-Series signal analyzers, compensate for this
potential effect by moving the LO start and stop frequencies.

Another type of error occurs when two peaks are displayed when only one actually
exists. Figure 2-26 shows this error. The outline of the two peaks is displayed using
peak detection with a wider RBW.

So peak detection is best for locating CW signals well out of the noise. Sample is best for
looking at noise, and normal is best for viewing signals and noise.

Figure 2-26. Normal detection can show two peaks when only one peak actually exists

Find us at www.keysight.com Page 38

Average detection
Although modern digital modulation schemes have noise-like characteristics, Sample
detection does not always provide us with the information we need. For instance,
when taking a channel power measurement on a W-CDMA signal, integration of the
rms values is required. This measurement involves summing power across a range of
analyzer frequency buckets. Sample detection does not provide this capability.

While spectrum analyzers typically collect amplitude data many times in each bucket, sample
detection keeps only one of those values and throws away the rest. On the other hand, an
averaging detector uses all the data values collected within the time (and frequency) interval of a
bucket. Once we have digitized the data, and knowing the circumstances under which they were
digitized, we can manipulate the data in a variety of ways to achieve the desired results.

Some spectrum analyzers refer to the averaging detector as an rms detector when it
averages power (based on the root mean square of voltage). Keysight X-Series signal
analyzers have an average detector that can average the power, voltage or log of the
signal by including a separate control to select the averaging type:

Power (rms) averaging computes rms levels, by taking the square root of the average
of the squares of the voltage data measured during the bucket interval. This computed
voltage is squared and divided by the characteristic input impedance of the spectrum
analyzer, normally 50 ohms. Power averaging calculates the true average power, and is
best for measuring the power of complex signals.

Voltage averaging averages the linear voltage data of the envelope signal measured
during the bucket interval. It is often used in EMI testing for measuring narrowband
signals (this topic will be discussed further in the next section). Voltage averaging is also
useful for observing rise and fall behavior of AM or pulse-modulated signals such as
radar and TDMA transmitters.

Log-power (video) averaging averages the logarithmic amplitude values (dB) of the
envelope signal measured during the bucket interval. Log power averaging is best for
observing sinusoidal signals, especially those near noise.11

Thus, using the average detector with the averaging type set to power provides true
average power based upon rms voltage, while the average detector with the averaging
type set to voltage acts as a general-purpose average detector. The average detector
with the averaging type set to log has no other equivalent.

11. See Chapter 5, “Sensitivity and Noise.”

Find us at www.keysight.com Page 39

Average detection is an improvement over using sample detection for the determination
of power. Sample detection requires multiple sweeps to collect enough data points to
give us accurate average power information. Average detection changes channel power
measurements from being a summation over a range of buckets into integration over
the time interval representing a range of frequencies in a swept analyzer. In a fast Fourier
transfer (FFT) analyzer12, the summation used for channel power measurements changes
from being a summation over display buckets to being a summation over FFT bins.

In both swept and FFT cases, the integration captures all the power information available,
rather than just that which is sampled by the sample detector. As a result, the average
detector has a lower variance result for the same measurement time. In swept analysis,
it also allows the convenience of reducing variance simply by extending the sweep time.

EMI detectors: average and quasipeak detection

An important application of average detection is for characterizing devices for
electromagnetic interference (EMI). In this case, voltage averaging, as described in
the previous section, is used for measuring narrowband signals that might be masked
by the presence of broadband impulsive noise. The average detection used in EMI
instruments takes an envelope-detected signal and passes it through a low-pass filter
with a bandwidth much less than the RBW. The filter integrates (averages) the higher-
frequency components such as noise. To perform this type of detection in an older
spectrum analyzer that doesn’t have a built-in voltage averaging detector function, set
the analyzer in linear mode and select a video filter with a cut-off frequency below the
lowest PRF of the measured signal.

Quasipeak detectors (QPD) are also used in EMI testing. QPD is a weighted form of
peak detection. The measured value of the QPD drops as the repetition rate of the
measured signal decreases. Thus, an impulsive signal with a given peak amplitude
and a 10-Hz pulse repetition rate will have a lower quasipeak value than a signal with
the same peak amplitude but having a 1-kHz repetition rate. This signal weighting is
accomplished by circuitry with specific charge, discharge and display time constants
defined by CISPR13.

12. Refer to Chapter 3 for more information on the FFT analyzers. They perform math computations on
many buckets simultaneously, which improves measurement speed.
13. CISPR, the Inter national Special Committee on Radio Interference, was established in 1934 by a
group of international organizations to address radio interference. CISPR is a non-governmental group
composed of National Committees of the International Electrotechnical Commission (IEC), as well as
numerous international organizations. CISPR’s recommended standards generally form the basis for
statutory EMC requirements adopted by governmental regulatory agencies around the world.

Find us at www.keysight.com Page 40

QPD is a way of measuring and quantifying the “annoyance factor” of a signal. Imagine
listening to a radio station suffering from interference. If you hear an occasional “pop”
caused by noise once every few seconds, you can still listen to the program without
More information:
too much trouble. However, if that same amplitude pop occurs 60 times per second,
A more detailed dis-
it becomes extremely annoying, making the radio program intolerable to listen to.
cussion about noise
markers can be found
Averaging processes in Spectrum and Signal
There are several processes in a spectrum analyzer that smooth the variations in Analyzer Measurements
and Noise – Application
envelope-detected amplitude. The first method, average detection, was discussed
Note, literature number
previously. Two other methods, video filtering and trace averaging, are discussed next.14 5966-4008E

Video filtering
Discerning signals close to the noise is not just a problem when performing EMC
tests. Spectrum analyzers display signals plus their own internal noise, as shown in
Figure 2-27. To reduce the effect of noise on the displayed signal amplitude, we often
smooth or average the display, as shown in Figure 2-28. Spectrum analyzers include a
variable video filter for this purpose. The video filter is a low-pass filter that comes after
the envelope detector and determines the bandwidth of the video signal that will later be
digitized to yield amplitude data. The cutoff frequency of the video filter can be reduced
to the point where it becomes smaller than the bandwidth of the selected resolution
bandwidth (IF) filter. When this occurs, the video system can no longer follow the more
rapid variations of the envelope of the signal(s) passing through the IF chain.

Figure 2-27. Spectrum analyzers display signal Figure 2-28. Display of Figure 2-27 after full smoothing
plus noise

14. A fourth method, called a noise marker, is discussed in Chapter 5, “Sensitivity and Noise.”

Find us at www.keysight.com Page 41

Figure 2-29. Smoothing effect of VBW-to-RBW ratios of 3:1, 1:10, and 1:100

The result is an averaging or smoothing of the displayed signal.

The effect is most noticeable in measuring noise, particularly when you use a wide-
resolution bandwidth. As we reduce the video bandwidth, the peak-to-peak variations
of the noise are reduced. As Figure 2-29 shows, the degree of reduction (degree of
averaging or smoothing) is a function of the ratio of the video to resolution bandwidths.
At ratios of 0.01 or less, the smoothing is very good. At higher ratios, the smoothing is
not as good. The video filter does not affect any part of the trace that is already smooth
(for example, a sinusoid displayed well out of the noise).

If we set the analyzer to positive peak detection mode, we notice two things: First, if
VBW > RBW, then changing the resolution bandwidth does not make much difference
in the peak-to-peak fluctuations of the noise. Second, if VBW < RBW, changing the
video bandwidth seems to affect the noise level. The fluctuations do not change much
because the analyzer is displaying only the peak values of the noise. However, the
noise level appears to change with video bandwidth because the averaging (smoothing)
changes, thereby changing the peak values of the smoothed noise envelope. See Figure
2-30a. When we select average detection, we see the average noise level remains
constant. See Figure 2-30b.

Find us at www.keysight.com Page 42

Because the video filter has its own response time, the sweep time increases approximately
inversely with video bandwidth when the VBW is less than the resolution bandwidth. The
sweep time (ST) can therefore be described by this equation:

k(Span)
ST ≈
(RBW)(VBW)

The analyzer sets the sweep time automatically to account for video bandwidth as well
as span and resolution bandwidth.

Figure 2-30a. Positive peak detection mode: reducing video Figure 2-30b. Average detection mode: noise level
bandwidth lowers peak noise but not average noise. remains constant, regardless of VBW-to-RBW ratios
(3:1, 1:10 and 1:100).

Trace averaging
Digital displays offer another choice for smoothing the display: trace averaging. Trace
averaging uses a completely different process from the smoothing performed using the
average detector. In this case, averaging is accomplished over two or more sweeps
on a point-by-point basis. At each display point, the new value is averaged in with the
previously averaged data:

( )
Aavg = n– 1 A prior avg +
n ( 1n ) A n where

A avg = new average value

A prior avg = average from prior sweep

An= measured value on current sweep

n = number of current sweep

Find us at www.keysight.com Page 43

Thus, the display gradually converges to an average over a number of sweeps. As
with video filtering, we can select the degree of averaging or smoothing. We do this
by setting the number of sweeps over which the averaging occurs. Figure 2-31 shows
trace averaging for different numbers of sweeps. While trace averaging has no effect on
sweep time, the time to reach a given degree of averaging is about the same as with
video filtering because of the number of sweeps required.

Figure 2-31. Trace averaging for 1, 5, 20 and 100 sweeps, top to bottom (trace position
offset for each set of sweeps

In many cases, it does not matter which form of display smoothing we pick. If the signal is
noise or a low-level sinusoid very close to the noise, we get the same results with either
video filtering or trace averaging.

However, there is a distinct difference between the two. Video filtering performs
averaging in real time. That is, we see the full effect of the averaging or smoothing at
each point on the display as the sweep progresses. Each point is averaged only once,
for a time of about 1/VBW on each sweep. Trace averaging, on the other hand, requires
multiple sweeps to achieve the full degree of averaging, and the averaging at each point
takes place over the full time period needed to complete the multiple sweeps.

As a result, we can get significantly different results from the two averaging methods on
certain signals. For example, a signal with a spectrum that changes with time can yield
a different average on each sweep when we use video filtering. However, if we choose
trace averaging over many sweeps, we will get a value much closer to the true average.
See Figures 2-32a and 2-32b.

Find us at www.keysight.com Page 44

Figures 2-32a and 2-32b show how video filtering and trace averaging yield different
results on an FM broadcast signal.

Figure 2-32a. Video filtering Figure 2-32b. Trace averaging

Time gating
Time-gated spectrum analysis allows you to obtain spectral information about signals
occupying the same part of the frequency spectrum that are separated in the time
domain. Using an external trigger signal to coordinate the separation of these signals,
you can perform the following operations:

• Measure any one of several signals separated in time (For example, you can
separate the spectra of two radios time-sharing a single frequency.)
• Measure the spectrum of a signal in one time slot of a TDMA system
• Exclude the spectrum of interfering signals, such as periodic pulse edge transients
that exist for only a limited time

Why time gating is needed

Traditional frequency-domain spectrum analysis provides only limited information for
certain difficult-to-analyze signals. Examples include the following signal types:

• Pulsed RF
• Time multiplexed
• Time domain multiple access (TDMA)
• Interleaved or intermittent
• Burst modulated

In some cases, time-gating capability enables you to perform measurements that would
otherwise be very difficult, if not impossible to make.

Find us at www.keysight.com Page 45

Measuring time division duplex signals
To illustrate using time-gating capability to perform difficult measurements, consider
Figure 2-33a, which shows a simplified digital mobile-radio signal in which two radios,
#1 and #2, are time-sharing a single frequency channel. Each radio transmits a single
1-ms burst, then shuts off while the other radio transmits for 1 ms. The challenge is to
measure the unique frequency spectrum of each transmitter.

Unfortunately, a traditional spectrum analyzer cannot do that. It simply shows the Figure 2-33b. Frequency
combined spectrum, as seen in Figure 2-33b. Using the time-gating capability and an spectrum of combined
external trigger signal, you can see the spectrum of just radio #1 (or radio #2 if you signals. Which radio
produces the spurious
wish) and identify it as the source of the spurious signal shown, as in Figure 2-33c. emissions?

Time gating can be achieved using three different methods we will discuss below.
However, there are certain basic concepts of time gating that apply to any implementation.
In particular, you must have, or be able to set, the following four items:

• An externally supplied gate trigger signal

• The gate control or trigger mode (edge or level) (The X-Series signal analyzers can
be set to gate-trigger holdoff to ignore potential false triggers.)
• The gate delay setting, which determines how long after the trigger signal the gate
actually becomes active and the signal is observed
Figure 2-33c. The time-
• The gate length setting, which determines how long the gate is on and the signal
gated spectrum of signal #1
is observed identifies it as the source
of spurious emission

Figure 2-33d. The time-

gated spectrum of signal
#2 shows it is free of
spurious emissions

Figure 2-33a. Simplified digital mobile-radio signal in the time domain

Find us at www.keysight.com Page 46

Controlling these parameters will allow us to look at the spectrum of the signal during
a desired portion of the time. If you are fortunate enough to have a gating signal that is
only true during the period of interest, you can use level gating, as shown in Figure 2-34.
However, in many cases the gating signal will not perfectly coincide with the time we
want to measure the spectrum. Therefore, a more flexible approach is to use edge
triggering in conjunction with a specified gate delay and gate length to precisely define
the time period in which to measure the signal.

Consider the GSM signal with eight time slots in Figure 2-35. Each burst is 0.577 ms
and the full frame is 4.615 ms. We may be interested in the spectrum of the signal
during a specific time slot. For the purposes of this example, let’s assume we are using
only two of the eight available time slots (time slots 1 and 3), as shown in Figure 2-36.
When we look at this signal in the frequency domain in Figure 2-37, we observe an
unwanted spurious signal present in the spectrum. In order to troubleshoot the problem
and find the source of this interfering signal, we need to determine the time slot in
which it is occurring.

RF signal

Gate signal

Figure 2-34. Level triggering: the spectrum analyzer Figure 2-35. A TDMA format signal (in this case, GSM)
only measures the frequency spectrum when the with 8 time slots, time slot zero is “off”.
gate trigger signal is above a certain level

Find us at www.keysight.com Page 47

If we wish to look at time slot 3, we set up the gate to trigger on the rising edge of the
burst in time slot 3, and, then specify a gate delay of 1.4577 ms and a gate length of
461.60 µs, as shown in Figure 2-38. The gate delay assures that we only measure the
spectrum of time slot 3 while the burst is fully on. Note that the gate start and stop
value is carefully selected to avoid the rising and falling edge of the burst, as we want
to allow time for the RBW filtered signal to settle out before we make a measurement.
Figure 2-39 shows the spectrum of time slot 3, which reveals that the spurious signal is
not caused by this burst.

Three methods are commonly used to perform time gating:

• Gated FFT
• Gated LO
• Gated video

Figure 2-36. A zero span (time domain) view of Figure 2-37. Frequency domain view of the GSM signal
the GSM signal with only time slots 1 and 3 “on”. with 2 time slots “on” showing an unwanted spurious
signal present in the spectrum.

Figure 2-38. Time gating is used to look at the Figure 2-39. Spectrum of time slot 3 reveals that
spectrum of the GSM time slot 3. the spurious signal is not caused by this burst.

Find us at www.keysight.com Page 48

Gated FFT
The Keysight X-Series signal analyzers have built-in FFT capabilities. In this mode, the
data is acquired for an FFT starting at a chosen delay following a trigger. The IF signal
is digitized and captured for a time period of 1.83 divided by resolution bandwidth.
An FFT is computed based on this data acquisition and the results are displayed as
the spectrum. Thus, the spectrum is that which existed at a particular time of known
duration. This is the fastest gating technique when the span is not wider than the FFT
maximum width.

To get the maximum possible frequency resolution, choose the narrowest available
RBW with a capture time that fits within the time period of interest. You may not
always need that much resolution, however, and you could choose a wider RBW
with a corresponding narrower gate length. The minimum usable RBW in gated FFT
applications is always lower than the minimum usable RBW in other gating techniques,
because the IF must fully settle during the burst in other techniques, which takes longer
than 1.83 divided by RBW.

Gated LO
LO gating, sometimes referred to as gated sweep, is another technique for performing
time gating. With this method, we control the voltage ramp produced by the scan
generator to sweep the LO, as shown in Figure 2-40. When the gate is active, the LO
ramps up in frequency like any spectrum analyzer. When the gate is blocked, the voltage
out of the scan generator is frozen, and the LO stops rising in frequency. This technique
can be much faster than gated video because multiple buckets can be measured
during each burst. As an example, let’s use the same GSM signal described earlier in
this chapter.

RF IF resolution Envelope Video

step bandwidth IF log detector bandwidth Peak/sample Analog-digital
attenuator Mixer filter amplifier (IF to video) filter detector converter
RF
input

Display logic

Local Scan generator

oscillator
Gate
control
Display

Figure 2-40. In gated LO mode, the LO sweeps only during gate interval

Find us at www.keysight.com Page 49

Using an X-Series signal analyzer, a standard, non-gated, spectrum sweep over a
1-MHz span takes 14.6 ms, as shown in Figure 2-41. With a gate length of 0.3 ms, the
spectrum analyzer sweep must be built up in 49 gate intervals (14.6 divided by 0.3).
Or, if the full frame of the GSM signal is 4.615 ms, the total measurement time is 49
intervals times 4.615 ms = 226 ms. This represents a significant improvement in speed
compared to the gated video technique, which will be described in the following section.
LO gating is available on X-Series signal analyzers and PSA Series spectrum analyzers.

Figure 2-41. Spectrum of the GSM signal

Find us at www.keysight.com Page 50

Gated video
Gated video is the analysis technique used in a number of spectrum analyzers, including
the Keysight 8560, 8590 and ESA Series. In this case, the video voltage is switched
off, or to “negative infinity decibels,” during the time the gate is supposed to be in its
“blocked” mode. The detector is set to peak detection. The sweep time must be set so
that the gates occur at least once per display point, or bucket, so the peak detector
is able to see real data during that time interval. Otherwise, there will be trace points
with no data, resulting in an incomplete spectrum. Therefore, the minimum sweep time
is N display buckets times burst cycle time. For example, in GSM measurements, the
full frame lasts 4.615 ms. For an ESA spectrum analyzer set to its default value of 401
display points, the minimum sweep time for GSM gated video measurements would be
401 times 4.615 ms or 1.85 s. Some TDMA formats have cycle times as large as 90 ms,
resulting in long sweep times using the gated video technique.

Now that you’ve seen how a classic analog spectrum analyzer works and how to use
some of the important features and capabilities, let’s take a look at how replacing some
analog circuits with digital technology improves spectrum analyzer performance.

RF IF resolution Envelope Video

step bandwidth IF log detector bandwidth Peak/sample Analog-digital
attenuator Mixer filter amplifier (IF to video) filter detector converter

RF
input
– ∞ dB
Reset
Gate control

Display logic

Local Scan generator

oscillator

Display

Figure 2-42. Block diagram of a spectrum analyzer with gated video

Find us at www.keysight.com Page 51

Chapter 3. Digital IF Overview
Since the 1980s, one of the most profound changes in spectrum analysis has been the
application of digital technology to replace portions of spectrum analyzers that had been
implemented previously as analog circuits. With the availability of high-performance
analog-to-digital converters, the latest spectrum analyzers digitize incoming signals
much earlier in the signal path compared to spectrum analyzer designs of just a few
years ago. The change has been most dramatic in the IF section of the spectrum
analyzer. Digital IFs1 have had a great impact on spectrum analyzer performance, with
significant improvements in speed, accuracy and the ability to measure complex signals
using advanced DSP techniques.

Digital filters
You will find a partial implementation of digital IF circuitry in the Keysight ESA-E Series
spectrum analyzers. While the 1-kHz and wider RBWs are implemented with traditional
analog LC and crystal filters, the narrowest bandwidths (1 Hz to 300 Hz) are realized
using digital techniques.

As shown in Figure 3-1, the linear analog signal is mixed down to an 8.5-kHz IF and
passed through a bandpass filter only 1 kHz wide. This IF signal is amplified, then
sampled at an 11.3-kHz rate and digitized.

Log
21.4 MHz

Video ADC µC

Linear

3rd LO Sample and hold

at 11.3 kHz

8.5 kHz CF
1 kHz BW

Figure 3-1. Digital implementation of 1-, 3-, 10-, 30-, 100- and 300-Hz resolution filters in
ESA-E Series spectrum analyzers

1. Strictly speaking, once a signal has been digitized, it is no longer at an intermediate frequency, or
IF. At that point, the signal is represented by digital data values. However, we use the term “digital
IF” to describe the digital processing that replaces the analog IF processing found in traditional
spectrum analyzers.

Find us at www.keysight.com Page 52

Once in digital form, the signal is put through a fast Fourier transform algorithm. To
transform the appropriate signal, the analyzer must be fixed-tuned (not sweeping).
That is, the transform must be done on a time-domain signal. Thus the ESA-E Series
analyzers step in 900-Hz increments, instead of sweeping continuously, when we select
one of the digital resolution bandwidths. This stepped tuning can be seen on the display,
which is updated in 900-Hz increments as the digital processing is completed.

As you will see in a moment, other spectrum and signal analyzers, such as the
Keysight X-Series analyzers, use an all-digital IF, implementing all resolution bandwidth
filters digitally.

A key benefit of the digital processing done in these analyzers is a bandwidth selectivity
of about 4:1. This selectivity is available on the narrowest filters, the ones we would
choose to separate the most closely spaced signals.

In Chapter 2, we did a filter skirt selectivity calculation for two signals spaced 4 kHz
apart, using a 3-kHz analog filter. Let’s repeat that calculation using digital filters. A
good model of the selectivity of digital filters is a near-Gaussian model:

H(∆f) = –3.01 dB x [ RBW/2

∆f
]α

where H(∆f) is the filter skirt rejection in dB.

∆f is the frequency offset from the center in Hz, and α is a parameter that controls
selectivity. α = 2 for an ideal Gaussian filter. The swept RBW filters used in Keysight
spectrum analyzers are based on a near-Gaussian model with an α value equal to 2.12,
resulting in a selectivity ratio of 4.1:1.

Entering the values from our example into the equation, we get:

[ ]
4000 2.12
H(4 kHz) = –3.01 dB x = –24.1 dB
3000/2

At an offset of 4 kHz, the 3-kHz digital filter is down –24.1 dB compared to the analog
filter which was only down –14.8 dB. Because of its superior selectivity, the digital filter
can resolve more closely spaced signals.

Find us at www.keysight.com Page 53

All-digital IF
Analyzers such as the Keysight X-Series combine several digital techniques to achieve
the all-digital IF. The all-digital IF offers users a wealth of advantages. The combination
of FFT analysis for narrow spans and swept analysis for wider spans optimizes sweeps
for the fastest possible measurements. Architecturally, the ADC is moved closer to the
input port, a move made possible by improvements to the A-to-D converters and other
digital hardware.

Let’s begin by taking a look at the block diagram of the all-digital IF in the X-Series
signal analyzer, as shown in Figure 3-2.

In this case, all 160 resolution bandwidths are digitally implemented. However, there is
some analog circuitry prior to the ADC, starting with several stages of down conversion,
followed by a pair of single-pole prefilters (one an LC filter, the other crystal-based).
A prefilter helps prevent succeeding stages from contributing third-order distortion in
the same way a prefilter would in an analog IF. In addition, it enables dynamic range
extension via autoranging. The output of the single-pole prefilter is routed to the
autorange detector and the anti-alias filter.

As with any FFT-based IF architecture, the anti-alias filter is required to prevent aliasing
(the folding of out-of-band signals into the ADC sampled data). This filter has many
poles and thus has substantial group delay.

Custom IC

Analog Digital Counter

Anti-alias gain gain I
filter
I, Q Display
ADC VBW det
Q r, log (r)

log pwr pwr log

Hilbert log v v log
Prefilter Ranging –1 transform
rules log log log log

Autoranging ADC system

Processor

Display
FFT Processing
log/lin dB/div Display

Figure 3-2. Block diagram of the all-digital IF in the Keysight X-Series signal analyzers

Find us at www.keysight.com Page 54

Even a very fast-rising RF burst, downconverted to the IF frequency, will experience a
delay of more than three cycles of the ADC clock (30 MHz) through the anti-alias filter.
The delay allows time for an impending large signal to be recognized before it overloads
the ADC. The logic circuitry controlling the autorange detector will decrease the gain in
front of the ADC before a large signal reaches it, thus preventing clipping. If the signal
envelope remains small for a long time, the autoranging circuit increases the gain,
reducing the effective noise at the input. The digital gain after the ADC is also changed
to compensate for the analog gain in front of it. The result is a “floating point” ADC with
very wide dynamic range when autoranging is enabled in swept mode.

Figure 3-3 illustrates the sweeping behavior of the X-Series analyzers. The single-pole
prefilter allows the gain to be turned up high when the analyzer is tuned far from the
carrier. As the carrier gets closer, the gain falls and the ADC quantization noise rises.
The noise level will depend on the signal level frequency separation from the carrier,
so it looks like a step-shaped phase noise. However, phase noise is different from this
autoranging noise. Phase noise cannot be avoided in a spectrum analyzer. However,
reducing the prefilter width can reduce autoranging noise at most frequency offsets from
the carrier. Since the prefilter width is approximately 2.5 times the RBW, reducing the
RBW reduces the autoranging noise.

Custom digital signal processing

Turning back to the block diagram of the digital IF (Figure 3-2), after the ADC gain
has been set with analog gain and corrected with digital gain, a custom IC begins
processing the samples. First, it splits the 30-MHz IF samples into I and Q pairs at half

Amplitude
(log)

ADC
clipping threshold
Prefilter gain
Typical
analog IF Digital IF RBW response
response
Noise floor after autoranging
Typical LO phase noise

Frequency or time

Figure 3-3. Autoranging keeps ADC noise close to the carrier and lower than LO noise
or RBW filter response

Find us at www.keysight.com Page 55

the rate (15 Mpairs/s). The I and Q pairs are given a high-frequency boost with a single-
stage digital filter that has gain and phase approximately opposite to that of the single-
pole analog prefilter. Next, I and Q signals are low-pass filtered with a linear-phase filter
with nearly ideal Gaussian response. Gaussian filters have always been used for swept
spectrum analysis, because of their optimum compromise between frequency domain
performance (shape factor) and time-domain performance (response to rapid sweeps).
With the signal bandwidth now reduced, the I and Q pairs may be decimated and sent
to the processor for FFT processing or demodulation. Although FFTs can be performed
to cover a segment of frequency span up to the 10-MHz bandwidth of the anti-alias
filter, even a narrower FFT span, such as 1 kHz, with a narrow RBW, such as 1 Hz,
would require FFTs with 20 million data points. Using decimation for narrower spans,
the number of data points needed to compute the FFT is greatly reduced, speeding
up computations.

For swept analysis, the filtered I and Q pairs are converted to magnitude and phase
pairs. For traditional swept analysis, the magnitude signal is video-bandwidth (VBW)
filtered and samples are taken through the display detector circuit. The log/linear display
selection and dB/division scaling occur in the processor, so a trace can be displayed on
any scale without remeasuring.

Additional video processing features

The VBW filter normally smoothes the log of the magnitude of the signal, but it has
many additional features. It can convert the log magnitude to a voltage envelope before
filtering and convert it back for consistent behavior before display detection.

Filtering the magnitude on a linear voltage scale is desirable for observing pulsed-RF
envelope shapes in zero span. The log-magnitude signal also can be converted to a
power (magnitude squared) signal before filtering, and then it can be converted back.
Filtering the power allows the analyzer to give the same average response to signals
with noise-like characteristics, such as digital communications signals, as to CW signals
with the same rms voltage. An increasingly common measurement need is total power
in a channel or across a frequency range.

In a measurement such as this, the display points might represent the average power
during the time the LO sweeps through that point. The VBW filter can be reconfigured
into an accumulator to perform averaging on either a log, voltage or power scale.

Find us at www.keysight.com Page 56

Frequency counting
Swept spectrum analyzers usually have a frequency counter. This counter counts the
zero crossings in the IF signal and offsets that count by the known frequency offsets
from LOs in the rest of the conversion chain. If the count is allowed to run for a second,
you can achieve a resolution of 1 Hz.

Because of its digitally synthesized LOs and all-digital RBWs, the native frequency
accuracy of the X-Series signal analyzer is very good (0.1% of span). In addition, the
X-Series signal analyzer includes a frequency counter that observes not just zero
crossings, but also the change in phase. Thus, it can resolve frequency to the tens-of-
millihertz level in 0.1 second. With this design, the ability to resolve frequency changes
is not limited by the spectrum analyzer, but rather is determined by the noisiness of the
signal being counted.

More advantages of all-digital IF

We have already discussed a number of advantages of signal analyzers with all-digital
IF: power/voltage/log video filtering, high-resolution frequency counting, log/linear
switching of stored traces, excellent shape factors, an average-across-the display-
point detector mode, 160 RBWs, and of course, FFT or swept processing. In spectrum
analysis, the filtering action of RBW filters causes errors in frequency and amplitude
measurements that are a function of the sweep rate. For a fixed level of these errors,
the all-digital IF’s linear phase RBW filters allow faster sweep rates than analog filters
permit. The digital implementation also allows well-known compensations to frequency
and amplitude readout, permitting sweep rates typically twice as fast as older analyzers
and excellent performance at even four times the sweep speed. Keysight X-Series
signal analyzers can achieve over 50 times faster sweep speeds (see Chapter 2 - Digital
resolution filters).

Digitally implemented logarithmic amplification is very accurate. Typical errors of the

entire analyzer are much smaller than the measurement uncertainty with which the
manufacturer proves the log fidelity. The log fidelity on all digital IF implementations is
specified at ± 0.07 dB for any level up to –20 dBm at the input mixer of the analyzer.
The range of the log amp does not limit the log fidelity at low levels, as it would be in
an analog IF; the range is only limited by noise around –155 dBm at the input mixer.
Because of single-tone compression in upstream circuits at higher powers, the fidelity
specification degrades to ± 0.13 dB for signal levels down to –10 dBm at the input
mixer. By comparison, analog log amps are usually specified with tolerances in the
± 1 dB region.

Find us at www.keysight.com Page 57

Other IF-related accuracies are improved as well. The IF prefilter is analog and must be
aligned like an analog filter, so it is subject to alignment errors, but it is much better than
most analog filters. With only one stage to manufacture, that stage can be made much
more stable than the 4- and 5-stage filters of analog IF-based spectrum analyzers. As a
result, the gain variations between RBW filters is held to a specification of ± 0.03 dB for
general digital IF implementations, which is ten times better than all-analog designs.

The accuracy of the IF bandwidth is determined by settability limitations in the digital

part of the filtering and calibration uncertainties in the analog prefilter. Again, the prefilter
is highly stable and contributes only 20 percent of the error that would exist with an
RBW made of five such stages. As a result, most RBWs are within 2 percent of their
stated bandwidth, compared to 10 to 20 percent specifications in analog-IF analyzers.

Bandwidth accuracy is important for minimizing the inaccuracy of channel power

measurements and similar measurements. The noise bandwidth of the RBW filters
is known to much better specifications than the 2 percent setting tolerance, and
noise markers and channel-power measurements are corrected to a tolerance of
± 0.5 percent. Therefore, bandwidth uncertainties contribute only ± 0.022 dB to the
amplitude error of noise density and channel-power measurements.

Finally, with no analog reference-level-dependent gain stages, there is no “IF gain” error
at all. The sum of all these improvements means that the all-digital IF makes a quantum
improvement in spectrum analyzer accuracy. It also allows you to change analyzer
settings without significantly impacting measurement uncertainty. We will cover this
topic in more detail in the next chapter.

Find us at www.keysight.com Page 58