Cascading Two-Ports with T-Parameters Instead of S
S-matrices of cascaded two-ports cannot simply be multiplied, but T-matrices can: T_total = T_A·T_B in signal-flow order. Two matched 6.02 dB pads give T11 = 4, so S21 = 0.25, or −12.04 dB.
Engineer’s note: People often try to predict a whole chain by simply multiplying a filter's S2P and an amplifier's S2P in a spreadsheet or script, and it fails. S-parameters can't be multiplied directly when you cascade. Convert to T-parameters, multiply, and convert back to S, and the result includes the reflections between the blocks.
For practitioners
Why S_A·S_B is not the cascade
An S-matrix answers one question: given the waves arriving at each port, which waves leave? For a two-port:
The input vector holds one incident wave from each side of the network. The output vector holds one outgoing wave from each side. Neither vector belongs to one end of the network only.
Now consider what S_A·S_B would mean. S_B turns [a1B; a2B] into [b1B; b2B]. Feeding that result into S_A treats b1B as a1A and b2B as a2A. That connects B's port 1 to A's port 1 and B's port 2 to A's port 2. A real cascade does neither. It joins A's port 2 to B's port 1, and the wave leaving one network enters the other.
The real connection also contains a feedback loop. Part of the wave leaving A at port 2 reflects off B (S11B), travels back into A, reflects again off A's output (S22A), and returns to B. Summing that geometric series gives the textbook cascade result:
Look at the S21 entry of the plain product S_A·S_B. It is S21A·S11B + S22A·S21B. For two well-matched networks, that is close to zero. The product would call a low-loss chain opaque.
Defining T from S
The fix is to reorder the wave variables so that one side of the equation holds only port 1 and the other holds only port 2. One common convention is:
Solving the S equations for a1 and b1 in terms of b2 and a2 gives:
The reverse conversion follows from the same definition. With a2 = 0, S21 = 1/T11 and S11 = T21/T11. With a1 = 0, S22 = −T12/T11 and S12 = det(T)/T11. The determinant of T works out to S12/S21, so a reciprocal network (S12 = S21) has det(T) = 1. That gives you a quick check on any converted data.
The cascade as a matrix product
Connect port 2 of network A to port 1 of network B, with the same reference impedance on both sides of the junction. The wave leaving one network is the wave entering the other:
a1B = b2A and a2A = b1B
So the right-hand vector of A, [b2A; a2A], is exactly the left-hand vector of B, [a1B; b1B]. Substitute:
That gives T_total = T_A·T_B, multiplied in signal-flow order. The same argument extends to any number of stages: T_total = T_1·T_2·…·T_N. Matrix multiplication does not commute, so order matters. Putting a mismatched filter before or after a mismatched amplifier changes the loop term, and the T product reflects that.
The multiple reflections have not gone away. They are built into the product. Expanding the top-left element:
T11,total = T11A·T11B + T12A·T21B = (1 − S22A·S11B) / (S21A·S21B)
Since S21 = 1/T11, this is the signal-flow result from the first section. The matrix product reproduces the reflection series without any series summation.
The same-reference-impedance condition is not a formality. If A's port 2 data is normalized to one impedance and B's port 1 data to another, a1B = b2A no longer holds. Renormalize one data set first.
Worked example: two matched 6.02 dB pads
Take a matched attenuator with S11 = S22 = 0 and S21 = S12 = 0.5, which is −6.02 dB.
- Δ = 0·0 − 0.5·0.5 = −0.25
- T11 = 1/0.5 = 2
- T12 = −0/0.5 = 0
- T21 = 0/0.5 = 0
- T22 = 0.25/0.5 = 0.5
So T = [2, 0; 0, 0.5]. Two pads in cascade give:
T_total = [2, 0; 0, 0.5]·[2, 0; 0, 0.5] = [4, 0; 0, 0.25]
Converting back gives S21 = 1/T11 = 0.25, and 20·log10(0.25) = −12.04 dB. Also S11 = T21/T11 = 0, S22 = −T12/T11 = 0, and S12 = det(T)/T11 = 1/4 = 0.25. The result is a matched, reciprocal 12.04 dB pad, as expected.
Compare this with the naive product. With S = [0, 0.5; 0.5, 0], S·S = [0.25, 0; 0, 0.25]. That matrix describes a network with S11 = 0.25 and S21 = 0, a partial reflector that passes nothing. Two attenuators in series obviously do not behave like that.
Where T-parameters break and where they earn their keep
Every T element is divided by S21, so T is undefined when S21 = 0. An ideal open switch or a perfect reverse-direction element has no T-matrix in this convention. When S21 is small but not zero, as in a deep filter stopband, the T elements become large. Noise in the measured S21 then spreads into all four elements, and the cascade loses accuracy. For stopband work, compute the cascade with the signal-flow formula or keep the S-domain result as a cross-check.
The biggest practical payoff is de-embedding. If a measurement is fixture X, then the device, then fixture Y, then T_meas = T_X·T_DUT·T_Y, and:
This only works if both fixture T-matrices are invertible and use the same convention as the measurement data. The de-embedding article covers how to get the fixture models.
In plain words · for students and beginners
T-parameters in plain words
An everyday analogy
Think of a chain of gears. Each gear changes the speed by a fixed ratio, and to find the ratio of the whole chain you multiply the ratios together. T-parameters work the same way for RF blocks. Each block gets a small table of numbers (a 2×2 matrix), and to get the table for the whole chain you multiply the tables in order.
The analogy breaks down because gears don't push energy backwards, but RF blocks do. A wave hitting a block can partly bounce back, which is called a reflection. That's why the T table has two rows: one tracks the wave going forward and one tracks the wave coming back. When you multiply the tables, the bounces between blocks are counted for you.
The same example, step by step
- One pad (an attenuator, a block that weakens the signal on purpose) lets half the wave through. S21, the "how much gets through" number, is 0.5. In decibels that is −6.02 dB.
- T is written as "left-side waves = T × right-side waves". If only half gets through, the input must be twice the output, so T11 = 1/0.5 = 2.
- Two pads in a row: 2 × 2 = 4. The input must be four times the output.
- Flip it back: 1/4 of the wave gets through, so S21 = 0.25.
- In decibels: 20·log10(0.25) = −12.04 dB, which is the two −6.02 dB losses added together.
Mistakes beginners make
- Multiplying S-matrices. S-parameters (scattering parameters, which describe what each port sends out for what comes in) mix both ends of a block in each column. Multiplying them connects the wrong wires. For the two pads it gives "nothing gets through", which is clearly wrong.
- Multiplying in the wrong order. With matrices, A·B is usually not the same as B·A. Write them in the order the signal passes through.
- Mixing conventions. Books define T in different ways. A formula from one book combined with data from another can give wrong answers.
- Using T when S21 is zero. Every T number divides by S21, and you can't divide by zero.
Remember
- To chain blocks, convert each S-matrix to T, multiply the T's in signal order, then convert back.
- For a single block, S21 = 1/T11, so a −6.02 dB pad has T11 = 2.
- T only works when S21 is not zero and both blocks use the same reference impedance at the joint.
Related pages on RF View
Frequently asked questions
Why can't I multiply S-parameter matrices to cascade two networks?
An S-matrix maps the waves arriving at both ports to the waves leaving both ports, so its input and output vectors each mix the two ends of the network. Multiplying two S-matrices connects the wrong waves and ignores the reflections that bounce between the networks.
How do I convert S-parameters to T-parameters?
With T defined by [a1; b1] = T·[b2; a2], T11 = 1/S21, T12 = −S22/S21, T21 = S11/S21 and T22 = −Δ/S21, where Δ = S11·S22 − S12·S21. Other texts use different conventions, so check which one a formula assumes.
In what order do T-matrices multiply for a cascade?
In signal-flow order: if the signal passes through A first and then B, T_total = T_A·T_B. Matrix multiplication does not commute, so reversing the order gives a different and generally wrong result.
When do T-parameters fail?
They are undefined when S21 = 0, because every element is divided by S21. Very small S21 values also make the T elements large and sensitive to noise in the S21 data.