De-Embedding a Fixture with T-Parameter Matrix Inversion
T_DUT = T_A^-1 · T_meas · T_B^-1 removes both fixtures at every frequency point. With two matched fixture halves of S21 = 0.9, measured S21 is 0.81 × S21_DUT, so de-embedding adds 1.83 dB.
Engineer’s note: De-embedding removes a fixture mathematically by inverting its T-matrix. If the fixture model is off by even 0.1 mm, the impedance after de-embedding jumps somewhere it shouldn't on the Smith chart. A careful TRL calibration built on measured standards usually serves fixture correction better.
For practitioners
The problem: the VNA measures the fixture too
A VNA calibration moves the reference planes to the ends of the test cables. Most DUTs don't connect straight to a cable. They sit on a test board or in a fixture, with a launch and some trace on each side. The S2P file you save describes the whole chain: fixture A, then the DUT, then fixture B.
If you know the S-parameters of each fixture half, you can remove them with arithmetic. Simulation or a separate fixture measurement can give you those values. You can't do this directly in S-parameters, because cascading S-matrices is messy. T-parameters (transfer parameters) turn the cascade into plain matrix multiplication. Removing a block then means multiplying by its inverse.
The calibration planes sit at the outer ends of the fixtures, where the cables connect. The DUT planes sit at the inner ends, where the fixture meets the DUT. De-embedding moves the reference from the first pair of planes to the second.
The T-matrix convention and the conversion from S
This article uses the convention [a1; b1] = T·[b2; a2]. The waves at port 1 depend on the waves at port 2. From the S-parameter definitions b1 = S11·a1 + S12·a2 and b2 = S21·a1 + S22·a2, solve for a1 and b1:
The conversion back to S is:
Two properties matter for de-embedding. First, every entry contains 1/S21, so the T-matrix only exists if S21 ≠ 0. Second, the determinant works out to:
So the matrix can only be inverted if S12 ≠ 0. For a reciprocal fixture, S12 = S21 and det T = 1. Any real fixture made of traces, connectors and launches is reciprocal, and both conditions hold unless the fixture is broken or far outside its usable band.
Other textbooks define T with the waves in a different order. The algebra changes, but the cascade logic doesn't. Pick one convention and use it for every matrix.
Cascade and inversion
At the junction between fixture A and the DUT, the wave leaving A at its port 2 is the wave entering the DUT at its port 1. So [b2; a2] of A equals [a1; b1] of the DUT. The same holds at the next junction. Chaining the definitions gives:
Matrix multiplication isn't commutative, so the order matters. Multiply on the left by T_A^-1 and on the right by T_B^-1:
For a 2×2 matrix the inverse is short. Swap the diagonal, negate the off-diagonal, and divide by the determinant. For a reciprocal fixture the determinant is 1, so you only swap and negate.
The full procedure, at each frequency point:
- Convert measured S to T_meas.
- Convert fixture A and fixture B S-parameters to T_A and T_B.
- Compute T_DUT = T_A^-1 · T_meas · T_B^-1.
- Convert T_DUT back to S.
Everything is complex. Magnitude and phase are both carried through. The fixture data must have the same frequency points as the measurement. It must also use the same reference impedance, usually 50 Ω. If it doesn't, renormalize it first.
Orientation of fixture B
This is the most common error. Fixture B sits between the DUT and VNA port 2. In the cascade, its port 1 must face the DUT and its port 2 must face the VNA. A fixture file is often simulated or measured with port 1 at the connector, because that's the natural way to set it up. If so, swap S11 with S22 and S12 with S21 before converting to T.
For a symmetric, matched fixture the swap changes nothing, so the error stays hidden. For a real launch with a different match at each end, the wrong orientation puts a reflection in the wrong place. The de-embedded S11 and S22 are then wrong even though S21 may look fine.
Worked example: a matched fixture with S21 = 0.9
Take two identical fixture halves. Each is matched (S11 = S22 = 0) and reciprocal, with S21 = S12 = 0.9. That is a loss of 20·log10(0.9) = −0.915 dB per half. Using the conversion above, ΔS = −0.81, so:
The determinant is (1/0.9)·0.9 = 1, as it should be for a reciprocal fixture. Now cascade with a general DUT. Because the fixture matrices are diagonal, the products are easy:
- T_meas,11 = (1/0.9)·T_DUT,11·(1/0.9) = T_DUT,11 / 0.81
- T_meas,21 = 0.9·T_DUT,21·(1/0.9) = T_DUT,21
So S21_meas = 1/T_meas,11 = 0.81 × S21_DUT. In decibels that's 2 × 0.915 = 1.83 dB of loss. De-embedding adds 1.83 dB back to the measured S21.
The same algebra gives S11_meas = T_meas,21 / T_meas,11 = 0.81 × S11_DUT. A reflection from the DUT passes through fixture A twice, so the measured return loss looks 1.83 dB better than the DUT's real return loss. That flattering error matters for match work. De-embedding removes it too.
A real fixture also has phase delay and some mismatch. Then T has non-zero off-diagonal terms and complex entries, and the result is no longer a simple scale factor. The procedure stays the same.
In plain words · for students and beginners
De-embedding with T-parameters in plain words
An everyday analogy
Picture looking at a painting through two sheets of tinted glass, one in front and one behind a backlight. You know exactly how much each sheet darkens and tints the light. To work out the painting's real colors, you undo the front sheet and the back sheet. De-embedding does the same for a test fixture. The fixture is the glass. The DUT (device under test, the part you actually care about) is the painting.
The analogy breaks down in one place. Glass only dims light. A fixture can also reflect signal back and delay it, and these effects mix together. That's why we need matrices, and not just a single number per fixture.
The same example in simple steps
- S-parameters describe how much signal goes through or bounces back at each port. S21 is "how much gets through."
- T-parameters are the same information, rearranged so that blocks in a row simply multiply.
- Each fixture half passes 0.9 of the signal amplitude. That's a loss of 0.915 dB.
- Two halves in a row pass 0.9 × 0.9 = 0.81 of the signal. So the VNA (vector network analyzer, the instrument that measures S-parameters) shows S21 = 0.81 × the DUT's real S21.
- Multiplying by the inverse of each fixture's T-matrix undoes that. In dB, you add 0.915 + 0.915 = 1.83 dB back.
- You repeat this at every frequency in the measurement file, then convert the result back to S-parameters.
Mistakes beginners make
- Wrong order. Matrix multiplication depends on order. Fixture A's inverse goes on the left and fixture B's inverse goes on the right.
- Fixture B turned around. Fixture B's port 1 must face the DUT. If the file has port 1 at the connector, swap its ports first.
- Different reference impedance. The fixture data and the measurement must both be referenced to the same impedance, usually 50 Ω.
- Subtracting dB from S11 by hand. This only works for a perfectly matched fixture. Real fixtures reflect a bit, and only the matrix method handles that.
Remember
- T_DUT = T_A^-1 · T_meas · T_B^-1, at every frequency point.
- Each fixture needs S21 ≠ 0 and S12 ≠ 0, or the math has no answer.
- Two matched halves with S21 = 0.9 hide 1.83 dB of the DUT's S21.
Related pages on RF View
Frequently asked questions
How do you de-embed a fixture using T-parameters?
Convert the measured S-parameters and both fixture S-parameters to T-matrices. Compute T_DUT = T_A^-1 · T_meas · T_B^-1, then convert T_DUT back to S. Do this at every frequency point.
Why does de-embedding fail when a fixture has S21 = 0?
The T-matrix has 1/S21 in its entries, so it does not exist when S21 = 0. Its determinant is S12/S21, so it also cannot be inverted when S12 = 0.
Which way round should the output fixture be described?
Describe fixture B with its port 1 facing the DUT. Then its wave order matches the cascade and T_meas = T_A · T_DUT · T_B holds.
How much does a matched fixture with S21 = 0.9 change the measured S21?
Each half has a loss of 0.915 dB. With one half on each side, measured S21 = 0.81 × S21_DUT, so de-embedding adds 1.83 dB to the measured S21.