Designing a spectrum analyzer into your own product

1The problem an integrator actually has

A company that sells spectrum monitoring systems does not want to be in the receiver business. It wants to sell nodes, a detection engine, a database and a map, and underneath all of that it needs a calibrated receiver. Design one, buy one as a module, or buy an uncalibrated board and finish it yourself.

The instinct is to compare receivers, so datasheets go side by side and the argument is about noise floor and phase noise. Three years later, when the program is in trouble, the cause is almost never noise floor. It is a supplier who changed a front-end filter without calling the change calibration-affecting, so a hundred fielded nodes now disagree with the fifty built last quarter, and nobody kept as-found data to say by how much.

Take the escape route seriously, because most integrators try it. Buy a low-cost wideband board, characterize it against a reference source, correct it in software and keep the margin. It works on the bench and fails at the second unit. The radio design is not the expensive part; the calibration fixture, the traceability chain, the qualification and ten years of obsolescence engineering are, and none of those scale to fifteen units a year.

2Where the boundary falls

An embedded receive module is a supply boundary before it is a component. What matters is what crosses the line: calibrated baseband samples, an absolute timestamp with a stated uncertainty, a calibration table with its traceability statement, a versioned interface, and a device identity carrying no vendor branding. Read Figure 1 as a list of deliverables. The timestamp uncertainty, the traceability statement and the interface deprecation period are the three most often left unspecified.

3What sets accuracy in a distributed system

Two measurements dominate these products: locating an emitter from the difference in arrival time between nodes, and taking a bearing on it from one node. Neither is limited by a receiver specification.

Time difference of arrival: a nanosecond is thirty centimeters

The conversion from timing to distance is the whole physics of the method:

One nanosecond is thirty centimeters. That conversion is worth more than any comparison of noise floors, because it says the answer is set by whatever assigns time to a sample. Geometry then multiplies the error:

Note what G is made of. It contains no receiver parameter at all. It is a site survey.

Figure 2 plots both effects on one axis. The logarithmic time axis is what bends the lines; the underlying relations are straight.

Direction finding: a decibel is three and a third degrees

Amplitude-comparison direction finding takes the arctangent of the ratio of two channel outputs, so any gain error between them rotates the answer. For a source at 45 degrees, where the channels should read equal, a gain mismatch ΔG gives:

Work it for one decibel. 10^1/20 is 1.1220, whose arctangent is 48.29 degrees, so the error is 3.29 degrees, larger than most published direction-finding accuracy figures and produced by a mismatch smaller than most published amplitude accuracy figures.

Figure 3 carries the commercial point. For two tolerances uniform on ±A, the difference has a 97.5th percentile of 1.553A, so a stack of perfect absolute certificates does not make a matched set.

4Why the specification page is the wrong comparison

Arrival-time uncertainty at a node is the root sum of squares of independent terms:

Table 1 evaluates it for one realistic node, and the variance column is where the argument of this paper is won or lost.

Table 1. The synchronization term is 98.5 percent of the variance, and it is set by the GNSS option on the order form rather than by anything inside the receiver.

Term	1σ	Variance (ns²)	Share
Correlator, 5 MHz signal, 10 dB SNR, 10 ms integration [illustrative]	0.23 ns	0.05	0.008 %
Time transfer, from a ±50 ns 1PPS bound taken as 2σ	25 ns	625.0	98.5 %
Antenna and feeder delay residual [illustrative]	2 ns	4.00	0.6 %
Timestamp quantization at the 8 ns engine clock	2.31 ns	5.33	0.8 %
Root sum of squares	25.19 ns	634.38	100 %
Range difference, Eq. 1	7.55 m
Position error at G = 3, Eq. 2	22.7 m
The 1PPS figure is published, quoted with its option; the other terms are illustrative.

Check it. 0.05 plus 625.0 plus 4.00 plus 5.33 is 634.38 ns², whose square root is 25.19 ns; times c that is 7.55 m, and at a dilution of precision of 3 it is 22.7 m. Make the correlator ten times better and the first term falls to 0.0005 ns², the total to 634.33 ns², and the position answer improves by 0.9 mm.

That is the buying criterion, and it applies to any supplier. Score a candidate on the uncertainty of the timestamp it hands you, the pairwise matching of a set of units, the temperature coefficient of amplitude, and the terms under which those may change. Sensitivity still has to be adequate. It is not the discriminator.

5Derived requirements, and an embedded module against them

The ICX-FieldHawk family includes two variants for this role: the ICX-090U, 9 kHz to 9.5 GHz with 50 MHz of analysis bandwidth standard and 100 MHz optional, and the ICX-400U, 9 kHz to 40 GHz with 100 MHz. Both are USB or LAN devices with no display, and published values hold after 10 minutes of warm-up, at 25 °C ambient, with spur reject standard on. Table 2 sets them against requirements derived for a 25 m position and 2 degree bearing specification. It is the hinge: before it is physics, after it is instrumentation and contract.

Table 2. Six of the seven requirements are properties of timing, matching or thermal behavior. Note the third row: amplitude accuracy is absolute and says nothing about how two units agree, so pairwise matching is a separate deliverable that has to be asked for.

Requirement	Derived value	Published module value
Node time-transfer uncertainty	≤ 28 ns, 1σ, from Eq. 2 at G = 3	GNSS 1PPS ±100 ns standard; ±75 or ±50 ns with options
Timestamp granularity	≤ 10 ns, under the sync term	8 ns time resolution
Pairwise gain match across a set	≤ 0.6 dB, giving 1.97° from Eq. 3	Not specified; amplitude accuracy is ±2.0 dB to 9.5 GHz, ±3.0 dB above, and absolute
Reference stability through an outage	≤ 168 ns/hour of holdover drift, to hold 28 ns through a 10 minute outage, from Eq. 2	TCXO < 1 ppm; OCXO option 01 < 0.15 ppm
Continuous IQ to host	≥ 20 MHz sustained	25 MHz continuous; 100 MHz burst into 128 Mbyte
Dissipation and envelope	≤ 20 W	9 to 16 W; < 305 g (090U USB), < 420 g (400U USB)
Operating temperature	site ambient plus internal rise	0 to 50 °C; −20 to +65 °C option 40; −40 to +65 °C option 41
Published values from the ICX-FieldHawk USB and LAN module datasheet, conditions as stated above.

Read the first row again, because it is the one the paper turns on. The standard 1PPS does not meet the derived requirement. Taken as a 2σ figure, ±100 ns is a 50 ns synchronization term, which puts the node at 45.1 meters rather than 22.7; the ±75 ns option reaches 33.8 meters, and only the ±50 ns option reaches the 22.7 meters Table 1 reports. That option is the line most often left off a quotation, and it is the only one that meets the specification.

The fourth row cannot be scored from a datasheet at all. TCXO and OCXO figures are frequency accuracy, not holdover drift, and a part specified at 0.15 ppm reaches the 28 ns budget in 0.19 seconds of free run if that accuracy is read as drift. Ask for drift after the last correction, over the outage you intend to survive, and treat any answer quoted in ppm as an answer to a different question.

The fifth row is the one an integrator underestimates: burst depth and sustained rate read alike on a datasheet and behave nothing alike in a product.

Figure 4 is what the fifth row buys: every panel is a projection of one array of complex samples, so a host that cannot sustain the stream does not lose a panel, it loses the record.

The host software, SpectraCore, is a full desktop application, which for an embedded product is the wrong deliverable. What matters is that it can be bypassed: the modules are driven from C, C++, C#, Python, MATLAB, Qt and LabVIEW over SCPI, on Windows and Linux, x64 and AArch64. Bring a set up under SpectraCore on a bench, then ship an ARM node with none of it.

The thermal term, decided by somebody who was not at the RF review

A sealed enclosure reaches a steady rise above ambient set by its thermal resistance, on a time constant set by its mass:

Evaluate it. 15 W times 1.67 K/W is 25.1 K, and 1.2 times 900 times 1.67 is 1,804 s, so the enclosure is within 5 percent of final temperature after 90 minutes and not before, and at 0.02 dB/K that rise is 0.50 dB of amplitude walk. At 45 °C ambient it also puts internal air at 70 °C, outside even the −40 to +65 °C class of option 41. The fix is area, a conduction path or a lighter finish, and it is not a receiver purchase.

Figure 5 is worth reading against Table 1. The frequency domain is where a reference looks good and the time domain is where it has to perform, and the variance column shows which of the two the position answer depends on.

6Build, buy, and the volume where the answer reverses

The crossover is where a fixed cost spread over a lifetime volume meets a flat unit price:

Table 3. At fifteen units a year the built front end costs about $33,000 per unit against $9,000 bought, and arrives two years later. Eq. 6 puts the crossover at $4,125,000 over ten times $3,500: 118 units a year.

Line item [all illustrative]	Build	Buy
Design and layout: RF, digital, firmware, mechanical, three spins	$1,610,000	n/a
Calibration fixture development and infrastructure	$500,000	n/a
Qualification, documentation, first article	$265,000	n/a
Integration: enclosure, thermal path, host software, product qualification	included above	$450,000
Sustaining, metrology and obsolescence, 10 years	$2,200,000	included
Fixed total	$4,575,000	$450,000
Marginal cost per unit; time to first delivery	$2,500; 24 to 30 months	$6,000; lead time
Per unit at 15 units/year	$33,000	$9,000
Per unit at 500 units/year	$3,415	$6,090
Illustrative structure to substitute into, not a quotation.

Table 3 is the honest version, and the Buy column is still generous to itself. An integrator pays for the enclosure, the thermal path, the host software and the qualification of the finished product whichever way the decision goes; put $450,000 against those and the crossover falls from 131 units a year to 118. Buying still wins at fifteen units by nearly four to one and still loses at five hundred.

Lifecycle risk belongs in the same arithmetic. For M parts with no second source and end-of-life notices at rate h, expected events over a life t are M(1 − e^−ht): at 25 parts and 0.08 per year over ten years, 13.8 events, with a two-in-a-billion chance of none. Obsolescence is a workload, and the contract decides whose engineers absorb it.

Terms that make a module a component rather than a dependency

• Two-tier change control. A short notice period for changes touching nothing calibrated, a longer one for anything touching amplitude, phase, timing or the interface, and the classification fixed in the agreement rather than at the time of the change.
• Continuity and calibration. An end-of-life notice period, a last-time-buy window, a named migration target, and a serialized calibration certificate with its traceability chain, plus a pairwise matching record where bearings matter.
• Interface stability and neutral identity. A versioning policy, a deprecation notice period and firmware pinning. Neutral enclosure, model number and documentation, and also device descriptors, service names, server headers and log strings, any of which will show a supplier name to a customer with a packet capture.

7Measurement practice

1. Measure your own arrival-time uncertainty. Feed two nodes from one source through a known splitter and cable set, log for an hour, and take the standard deviation of the difference. That is what goes into Eq. 4, not a datasheet value.
2. Compute the dilution of precision before siting. Evaluate Eq. 2 over the coverage area for each candidate layout. Moving one node a few hundred meters routinely buys more than a tenfold receiver improvement.
3. Wait out the thermal transient, and watch the feeder. Amplitude means nothing until three time constants after power-on, roughly 90 minutes. Three meters of ordinary PTFE coax at 1 GHz, at 50 ppm/K and a velocity factor of 0.7, drifts 0.26 degrees per kelvin, so two nodes twenty kelvin apart are five degrees of phase apart before the receiver does anything. Phase-stable assembly at 10 ppm/K takes that to 0.05 degrees per kelvin, and it is the cheapest line in the node.
4. Test bearing accuracy the way ITU-R SM.2060-0 prescribes. A full azimuth traverse in 5 or 10 degree steps, across the band, reporting RMS error over all azimuths with the signal-to-noise ratio stated. Plot error against azimuth too: gain mismatch peaks at 45 degrees and nulls at 0 and 90, phase mismatch grows off broadside, and the shape says which defect you have.
5. Keep as-found calibration data from the first unit. A 200-unit fleet recalibrated annually at three weeks turnaround holds 11.5 units in transit and costs about $226,000 a year at $1,130 per unit [illustrative]. A 24-month interval halves both, and the only thing in the way is drift data nobody kept.

Field evidence

The pattern here is long-running rather than a single case. Several independent system houses have built their own monitoring and direction-finding products on receive modules from this family, write their own software and sell the result under their own name, and none is in the receiver business. Each made the section 6 judgement the same way, at volumes well below the crossover, where the arithmetic says buy and the instinct says build.

8Where an embedded module is the wrong answer

• Below about one nanosecond of time transfer, no receiver helps. Meter-class geolocation needs length-calibrated fibre, GNSS common view with long averaging, or a physical timing network, and that infrastructure becomes the product.
• For narrowband emitters, time difference of arrival is the wrong method. A 12.5 kHz channel at 10 dB signal-to-noise ratio with a full second of integration reaches 181 ns of estimation error, 54 m before geometry or synchronization. Bandwidth decides the method; receiver quality does not.
• A monitoring receiver is not a compliance measuring receiver. Conformity measurement needs the CISPR detectors and bandwidths, a defined input impedance and overload factor, and an accredited uncertainty budget, and no CISPR function appears in the module specification. Nor is the module the limiting element for field strength: a ±0.3 dB receiver inside a ±2.5 dB antenna factor gives ±2.5 dB.
• Above roughly 120 units a year, build. Table 3 says so plainly. The argument for buying is an argument about this volume, not a universal claim.

9Summary, selection and definitions

Table 4 collects the governing relations and published values with their conditions.

Table 4. Governing relations and published values with their conditions, in a form that lifts into an evaluation matrix.

Quantity	Relation or value	Condition
Range from time	Δr = cΔτ; 1 ns = 0.2998 m	vacuum
Position error	G cστ	G = 1/(2 sin(θ/2)), two nodes
Bearing error from mismatch	arctan(10ΔG/20) − 45°; pairwise spread 1.553A	source at 45°; two uniform ±A
Build/buy crossover	118 units/year	illustrative Table 3 values, including buy-side integration cost
Module GNSS 1PPS; bandwidth; power	±100 ns std, ±75 or ±50 ns optional, and only the ±50 ns option meets the 25 m case; 50/100 MHz (090U), 100 MHz (400U); 9 to 16 W	10 min warm-up, 25 °C, spur reject standard on

Selecting a variant

Take the ICX-090U for nodes below 9.5 GHz, which covers most civil monitoring and direction finding, and the ICX-400U where coverage reaches microwave backhaul or radar bands, accepting ±3.0 dB above 9.5 GHz. Take the ±50 ns GNSS option on every node in a locating network, because section 5 shows nothing below it meets a 25 m specification, option 01 where time must also survive an outage, and option 40 or 41 wherever site ambient plus the rise from Eq. 5 exceeds 0 to 50 °C.

Geometric dilution of precision

The factor by which node geometry multiplies measurement error into position error. A property of the layout, with no instrument parameter in it.

Pairwise matching

How closely two units of a set agree with each other, as distinct from how closely each agrees with an absolute standard.

Traceability

An unbroken chain of comparisons linking a calibration to a national standard, each with a stated uncertainty.

Symbols used in this paper.

Symbol	Meaning	Units
c	Speed of light	m/s
Δr, Δτ	Range and arrival-time difference	m, s
G	Geometric dilution of precision	dimensionless
στ, σp	Time and position uncertainty	s, m
ΔG, εθ	Gain mismatch, bearing error	dB, degrees
σest, σsync, σcal, σq	Estimation, time-transfer, residual delay and quantization uncertainty	s
τth, m cp	Thermal time constant, thermal mass	s, J/K
θ, L, A	Subtended angle; program life; tolerance half-width	degrees, years, dB
P, Rth, kT	Dissipation, thermal resistance, amplitude coefficient	W, K/W, dB/K
N*, Cbuild, Cbuy, M, h	Crossover volume, the two fixed costs, single-sourced parts, end-of-life rate	units/yr, currency, count, 1/yr

Further reading

• IEEE Std 1588-2019, Precision Clock Synchronization Protocol, and IEC 62402, Obsolescence management.
• Recommendation ITU-R SM.2060-0, Test procedure for measuring direction finder accuracy, and ITU-R SM.854-3 on direction finding at monitoring stations.
• Berkeley Nucleonics BNC-AN-102, Wide-area spectrum monitoring with automated classification, and the ICX-FieldHawk USB and LAN module datasheet.

If you are sizing an embedded receiver into a product of your own, our application engineers would be glad to work the timing budget and the build-against-buy arithmetic against your volume and your emitters.