How to Check Passivity of Measured S-Parameter Data
A network is passive at a frequency only if every eigenvalue of I − S^H S is ≥ 0, which is the same as the largest singular value of S being ≤ 1. Entries of 0.5 and 0.9 each pass |Sij| ≤ 1 but give a column power sum of 1.06, so the data is not passive.
Engineer’s note: Measured data for a passive part sometimes shows S21 above 0 dB, apparent gain. The usual cause is a poor contact when connectors were swapped, or phase error from a cable that moved after calibration. Push data like that into a circuit simulator and a transient analysis can diverge and stop, so screen files before you use them.
For practitioners
A passive network cannot deliver more power than it receives. That sounds like an easy property to check in a Touchstone file. People often scan the magnitudes, see that nothing goes above 0 dB, and decide the data is fine. That check is incomplete. You need a matrix test, and you need to run it at every frequency point.
Why checking each |Sij| ≤ 1 is not the test
Every S-parameter on its own describes one path: power going into port j and coming out of port i. Passivity is about the total. Power that enters one port can leave through every port at once, and all of those outgoing powers have to fit inside what went in.
When you check each entry against 1, you only confirm that no single path exceeds the input. You don't confirm that the paths together stay within it. A network can pass the entry check on every element and still create power.
The passivity condition
Write the incident power waves as a vector a and the reflected waves as b = S·a. With power waves normalized to real, positive reference impedances (for example 50 Ω at every port, as in standard VNA data), the net power absorbed by the network is:
Here S^H is the conjugate transpose of S. The network is passive at a given frequency if P_net ≥ 0 for every possible excitation a. By definition, that is the statement that I − S^H S is positive semidefinite:
The eigenvalues of S^H S are the squares of the singular values of S. So an equivalent test is:
Both forms give the same answer. The eigenvalue form shows how much margin you have. The singular-value form is often a single library call.
Some special cases follow directly:
- One-port: S is a 1 × 1 matrix, so the test becomes |S11| ≤ 1.
- Column check: the diagonal entries of S^H S are the column power sums Σk |Ski|². A positive semidefinite matrix can't have a negative diagonal entry, so each column must satisfy Σk |Ski|² ≤ 1. This condition is necessary but not sufficient, because the off-diagonal cross terms can still push the largest eigenvalue of S^H S above 1.
- Lossless: S^H S = I, so every eigenvalue of I − S^H S is exactly zero. A lossless network sits right on the boundary.
A worked two-port example
Take a reciprocal, symmetric two-port with real entries:
Every entry has a magnitude below 1, so the entry check passes. Now drive port 1 alone and add up the power in column 1: |S11|² + |S21|² = 0.25 + 0.81 = 1.06. That's more than 1, so the column condition already fails.
The full test confirms it. S is real and symmetric, so S^H S = S², and the eigenvalues of S^H S are the squares of the eigenvalues of S. The eigenvalues of S are 0.5 ± 0.9, which are 1.4 and −0.4. Squaring gives 1.96 and 0.16. The eigenvalues of M = I − S^H S are therefore:
- 1 − 1.96 = −0.96
- 1 − 0.16 = 0.84
One eigenvalue is negative, so the network is non-passive at this frequency. The largest singular value of S is 1.4, which also breaks the σmax ≤ 1 limit.
The procedure for one frequency point is:
- Read the S-matrix at frequency f with a 50 Ω reference.
- Compute M = I − S^H S.
- Find the eigenvalues of M. Because M is Hermitian, they are real.
- If all eigenvalues are ≥ 0, the data is passive at f and you move to the next frequency. If any eigenvalue is negative, flag f as non-passive.
Running the check across a sweep
Passivity is a property of each frequency point, so the test runs once per point. For each frequency, keep the minimum eigenvalue of I − S^H S and plot that value against frequency. Zero is the passivity limit. A passive point sits on or above zero, and a non-passive point falls below it.
This plot shows you more than a pass/fail flag does. You can see where the violations are, how deep they go, and whether they are isolated points or a continuous band. A few shallow dips usually point to measurement noise. A wide, deep region usually points to a systematic problem.
Some practical points:
- Convert first. Touchstone files can store data as dB/angle, magnitude/angle, or real/imaginary. Convert everything to complex linear values before you build S. The line-by-line S2P walkthrough covers the formats.
- Check the reference. The test as stated assumes real, positive reference impedances. If the file uses a different reference at some port, make sure you understand what normalization the data actually has before you trust the result.
- Use a Hermitian solver. M is Hermitian, so a Hermitian eigenvalue routine returns real eigenvalues. A general solver can return small imaginary parts that come only from rounding.
- Allow for rounding. A lossless or nearly lossless structure has eigenvalues close to zero. Floating-point rounding alone can push those slightly negative, so judge small violations against the precision of your data and arithmetic.
Why measured data goes non-passive
A real passive device can't create power, so a violation in measured data comes from the measurement or the processing. The usual suspects are calibration residuals, noise on paths where the true value is close to lossless, fixture de-embedding that over-corrects, and interpolation between frequency points. Paths near 0 dB transmission are the most exposed, because they have almost no room before the power sum crosses 1.
If you find violations, look at the measurement first. Check the calibration, the connectors, and the de-embedding steps. Only then consider a passivity-enforcement step, which adjusts the data slightly so it meets the condition. Treat enforcement as a correction you document, not a way to hide a bad measurement. To view the raw traces before and after, load the file into the S2P Viewer.
In plain words · for students and beginners
S-parameter passivity in plain words
An everyday analogy
Think of a passive network as a shop that can't print money. Each customer (a signal) brings in some cash (power). The shop can hand some back at the door (reflection), pass some to a branch across town (transmission), and keep the rest as heat (loss). What it can't do is pay out more than it took in.
The analogy breaks down in one place. Signals are waves, so when signals arrive at more than one port together, they can add up or cancel each other depending on their timing (phase). That's why the real test uses the whole matrix at once and not one number at a time.
The same example in simple steps
Our two-port has these values:
- S11 = 0.5: the share of the wave at port 1 that bounces back.
- S21 = 0.9: the share of the wave at port 1 that comes out at port 2.
- S22 = 0.5 and S12 = 0.9 are the same values seen from port 2.
All four numbers are below 1, so the data looks fine at first glance. Now count power. Power goes with the square of the wave size.
- Put 1 W into port 1.
- Reflected power is 0.5² = 0.25 W.
- Transmitted power is 0.9² = 0.81 W.
- Total power out is 0.25 + 0.81 = 1.06 W.
That's 0.06 W more out than in. A passive part can't do that, so this data is not passive, even though each entry is below 1.
The full check builds a matrix called I − S^H S. Here "I" is the identity matrix and "S^H" means S flipped across its diagonal with each complex number conjugated. You then find its eigenvalues, the special numbers that describe how the matrix scales its inputs. For this example they are −0.96 and 0.84. A negative eigenvalue means that some way of driving the ports gets out more power than you put in.
Mistakes beginners make
- Checking only single entries. "Nothing above 0 dB" isn't a passivity test. You have to add up the powers.
- Adding magnitudes instead of squares. Power goes with |S|², not |S|. Here 0.5 + 0.9 = 1.4 would give the wrong picture, while 0.25 + 0.81 = 1.06 is the correct power sum.
- Testing one frequency. A file can pass at most points and fail at a few. Check every point.
- Mixing up formats. A value of −3 dB isn't 0.5 in linear terms. Convert dB to linear values first.
- Panicking over tiny negatives. An eigenvalue just below zero on a nearly lossless part can come from rounding or noise and doesn't always mean the measurement is wrong.
Remember
- Passive means power out is never more than power in, for every way you drive the ports.
- The test: all eigenvalues of I − S^H S are ≥ 0 at every frequency. Equivalently, the largest singular value of S is ≤ 1.
- Every |Sij| < 1 isn't enough. S11 = 0.5 with S21 = 0.9 already sums to 1.06.
Related pages on RF View
Frequently asked questions
Is it enough to check that every |Sij| is below 1 for passivity?
No. Every entry can be below 1 while the network still puts out more power than it takes in. For example, S11 = S22 = 0.5 and S21 = S12 = 0.9 give a column power sum of 1.06, so the network is not passive.
What is the passivity condition for an S-parameter matrix?
At each frequency, I − S^H S must be positive semidefinite, which means all of its eigenvalues are ≥ 0. The same condition says the largest singular value of S must not be greater than 1.
How do I check passivity of a one-port S1P file?
For a one-port the matrix test becomes |S11| ≤ 1 at every frequency. Any point where the reflection magnitude goes above 1 is non-passive.
Does the passivity test depend on the reference impedance?
The eigenvalue test as stated applies to power-wave S-parameters normalized to real, positive reference impedances, such as 50 Ω at every port. Standard VNA data meets this condition.