# Using the Three-Dimensional Smith Chart

The Smith Chart is a standard tool in RF design. Developed by Phillip Smith in 1939, it has become the most popular graphic method for representing impedances and solving operations with complex numbers. Traditionally, the Smith Chart has been used as 2-D polar form, centered at an unit radius circle. However, the 2D format has some restrictions when the active impedances (oscillators) or stability circles (amplifiers) are represented, since these ones usually leave the polar chart. Last years, three-dimensional Smith Chart has become popular. Advances in 3D rendering software make it easy to use for design. In this post, I will try to show the handling of the three-dimensional Smith Chart and its application for a low-noise amplifier design.

When Phillip Smith was working at Bell Labs, he have to match one antenna and he was looked for a way to solve the design graphically. By means of the mathematical expressions that define the impedances in the transmission lines, he got to represent the impedance complex plane by circles with constant resistances and reactances. These circles made it easier for him to represent any impedance in a polar space, with the maximum matching placed in the center of the chart and the outer circle representing the pure reactance. Traditionally, Smith’s Chart has been represented in polar form as shown below

Fig. 1 – Traditional Smith’s Chart

The impedance is normalized calculating the ratio between the impedance and the generator impedance. The center of the chart is pure unit resistance (maximum matching) while the peripheral circle that limits the chart is the pure reactance. The left end of the chart represents the pure short circuit and the right end, the pure open circuit. The chart was then very popular to be able to perform calculations for matching networks with transmission lines using a graphical method. However, the design difficulties with the chart happened when active impedances were analyzed, studying amplifiers stability and designing oscillators.

By its design, the chart is limited to the impedances with positive real part, but it could represent, extending the complex plane through the Möbius transformation, impedances with negative real part [1]. This expanded chart, to the negative real part plane, can be seen in the following figure

Fig. 2- Smith’s Chart expanded to active impedances

However,this chart shows two issues: 1) although it allows to represent all the impedances, there is a problem with the complex infinity, so it remains limited and 2) the chart has large dimensions that make it difficult to us in a graphic environment, even in a computer-aided environment. However, the extension is needed when the amplifier stability circles are analyzing, since in most of cases the centers of these circles are located outside the passive impedance chart.

In a graphical computer environment, representing the circles is already performed by the software itself through the calculations, being able to limit the chart to the passive region and drawing only a part of the circle of stability. But with oscillators still have the problem of complex infinity, which could be solved through a representation in a Riemann’s sphere.

RIEMANN’S SPHERE

The Riemann’s sphere is a mathematical solution for representing the complete complex plane, including infinity. The entire complex surface is represented on a spherical surface by a stereographic projection of this plane.

Fig. 3 – Projection of the complex plane on a sphere

In this graphic form the southern hemisphere represents the origin, the northern hemisphere represents infinity and the equator the circle of unitary radius. The distribution of complex values in the sphere can be seen in the following figure

Fig. 4 – Distribution of complex values in the sphere

So, it is possible to represent any complex number on a surface easy to handle.

SMITH’S CHART ON A RIEMANN’S SPHERE

Since Smith’s Chart is a complex representation, it can be projected in the same way to a Riemann’s sphere [2], as shown in the following figure

Fig. 5 – Projection of the Smith’s Chart on a Riemann’s sphere

In this case, the northern hemisphere shows the impedances with positive resistance (passive impedances), in the southern hemisphere, the impedances with negative resistance (active impedances), in the eastern hemisphere, the inductive impedances, and in the western one the capacitive impedances. The main meridian shows the pure resistive impedance.

Thus, when we wish to represent any impedance, either active or passive, it can be represented at any point in the sphere, greatly facilitating its drawing. In the same way, we can represent the stability circles of any amplifier without having to expand the chart. For example, if we want to represent the stability circles for one transistor, which parameters S at 3GHz are the next

S11=0,82/-69,5   S21=5,66/113,8   S12=0,03/48,8  S22=0,72/-37,6

its representation in the conventional Smith’s Chart is

Fig. 6 – Traditional representation for stability circles

while in the three-dimensional chart it is

Fig. 7 – Stability circles on the 3D chart

where both circles can be seen, a fraction in the northern hemisphere and the other one in the south. Thus, its representation has been greatly facilitated.

A PRACTICAL APPLICATION: LOW NOISE AMPLIFIER

Let’s see a practical application of the 3D chart matching the previous amplifier with the maximum stable gain and minimum figure of noise, at 3GHz. Using traditional methods, and knowing the transistor parameters which are the next

S11=0,82/-69,5   S21=5,66/113,8   S12=0,03/48,8  S22=0,72/-37,6

NFmin=0,62  Γopt=0,5/67,5 Rn=0,2

S-parameters are represented in the3D Smith’s chart and the stability circles are drawn. For a better representation 3 frequencies are used, with a 500MHz bandwidth.

Fig. 8 – S-parameters and stability circles for the transistor (S11 S21 S12 S22 Input Stability Circle Output Stability Circle)

It can be seen that S-parameters as well as the stability circles in both the conventional Smith’s chart and 3D one. In the conventional Smith’s chart, the stability circles leave the chart.

One amplifier is unconditionally stable when the stability circles are placed in the active impedance area of the chart, in the southern hemisphere, under two conditions: if the circles are placed in the active region and do not surround the passive one, the unstable impedances are located inside the circle. If the circles surround the passive region, the unstable impedances are located outside the circle.

.

Fig. 9 – Possible cases for stability circles in the active region

In this case, since part of the circles enters on the passive impedances region, the amplifier is conditionally stable.Then the impedances that could unstabilize the amplifier are placed inside the circles. This is something that cannot be seen clearly in the three-dimensional chart yet, the app does not seem to calculate it and would be interesting to include in later versions, because it would greatly facilitate the design.

Let’s match now the input for the minimum noise. For this, it is needed to design a matching network to transform from 50Ω to reflection coefficient Γopt, being its normalized impedance Zopt=0,86+j⋅1,07. In the app, opening the design window and writing this impedance

Fig. 10 – Representation of Γopt

Using now the admittance, we translate in the circle of constant conductance until the real part of the impedance is 1. This is down by estimation and a 0,5 subsceptance is got. It should be increased 0,5 – (- 0,57) = 1.07 and this is a shunt capacitor, 1,14pF.

Fig. 11 – Translating to circle with real part 1.

Now it is only needed to put a component that makes zero the reactance, when the resistance is constant. As the reactance is -1.09, the added value should be 1.09, so that the reactance is zero. This is equivalent to a series inductor, 2,9nH.

Fig. 12 – Source impedance matched to Γopt

Once calculated the input matching network for the lower noise figure, we recalculate the S-parameters. Being an active device, the matching network transforms the S parameters, which are:

S11=0,54/-177   S21=8,3/61,1   S12=0,04/-3,9  S22=0,72/-48,6

and which are represented in the Smith’s chart to get the stability circles.

Fig. 13 – Transistor with matching network to Γopt and stability circles.

The unstable regions are the internal regions, so the amplifier remains stable.

Now the output matching network is got for maximum stable gain, and the ouput reflection coefficient S22=0,72/-48,6 should be loaded by ΓL (S22  conjugate), translating from 50Ω to ΓL=0,72/48,6. This operation is performed in the same way that input matching network. By doing the complete matching , S parameters are recalculated, with input and oputput matching networks. These are

S11=0,83/145   S21=12/-7.5   S12=0,06/-72,5  S22=0,005/162

The gain is 20·log(S21)=21,6dB, and the noise figure, 0,62dB (NFmin). Now it is only represented these parameters in the three-dimensional chart to get the stability circles.

Fig. 14 – Low noise amplifier and stability circles

In this case, the stable region in the input stability circle is inside and in the otuput stabiliy circle is outside. Due to both reflection coefficients, S11 y S22 are into the stable regions, then the amplifier is stable.

CONCLUSIONS

In this entry I had the first contact with the three-dimensional Smith’s chart. The object was to study its potential with respect the traditional chart in microwave engineering. New advantages are observed in this respect in that it is possible to represent the infinite values ​​from the Möbius transform to a Riemann’s sphere and thus having a three-dimensional graphical tool where practically all passive and active impedances and parameters which can be difficult to draw in the traditional chart as stability circles.

In its version 1, the app, which can be found on the website 3D Smith Chart / A New Vision in Microwave Analysis and Design, shows some design options and configurations, although some applications should be undoubtedly added In future versions. In this case, one of the most advantageous applications for the chart, having studied the stability circles of an amplifier, is the location of the stability regions graphically. Although this can be solved by calculation, the visual image is always more advantageous.

The app has a user manual with examples explained in a simple way, so that the designer becomes familiar with it immediately. In my professional opinion, it is an ideal tool for those of us who are used to using Smith’s chart to perform our matching network calculations.

REFERENCES

1. Müller, Andrei; Dascalu, Dan C; Soto, Pablo; Boria, Vicente E.; ” The 3D Smith Chart and Its Practical Applications”; Microwave Journal, vol. 5, no. 7, pp. 64–74, Jul. 2012
2. Andrei A. Muller, P. Soto, D. Dascalu, D. Neculoiu and V. E. Boria, “A 3D Smith Chart based on the Riemann Sphere for Active and Passive Microwave Circuits,” IEEE Microwave and Wireless Components. Letters, vol 21, issue 6, pp 286-288, june, 2011
3. Zelley, Chris; “A spherical representation of the Smith Chart”; IEEE Microwave, vol. 8, pp. 60–66, July 2007
4. Grebennikov, Andrei; Kumar, Narendra; Yarman, Binboga S.; “Broadband RF and Microwave Amplifiers”; Boca Raton: CRC Press, 2016; ISBN 978-1-1388-0020-5

# Ajustando filtros mediante el método de Dishal

En Telecomunicaciones es usual tener que usar filtros para poder eliminar frecuencias indeseadas. Estos filtros suelen ser de bandas muy estrechas y se suelen utilizar técnicas de líneas acopladas, por lo que en la mayor parte de los diseños se debe recurrir a la simulación electromagnética para verificar el diseño. La simulación electromagnética, aunque es una potente herramienta, suele ser lenta si se desea optimizar mediante algoritmos convencionales. Aunque estos algoritmos están incluidos en la mayor parte de los simuladores electromagnéticos, ya sea en 2D o en 3D, si la respuesta del filtro está muy alejada de la deseada, la optimización suele ser muy lenta, por lo que se requieren otros métodos que permitan ajustar previamente antes de realizar una optimización final. Uno de los métodos es el de Dishal, en el que se puede sintonizar un filtro de varias secciones a base de sintonizar cada una de ellas. En esta entrada, sintonizaremos un filtro microstrip de tipo HAIRPIN, de resonadores λ/2 acoplados, usando un simulador electromagnético como HPMomentum.

Los filtros son los dispositivos más comunes que se usan en Telecomunicaciones. Eliminan las frecuencias interferentes y el ruido, pudiendo procesar la señal recibida o transmitida de una forma más eficiente. Tienen bastante literatura para su diseño, y existen muchas combinaciones para obtener su respuesta. Sin embargo, es uno de los dispositivos en los que es más difícil obtener un óptimo resultado. Su sintonía física requiere habilidad y entrenamiento, y su sintonía en simulación paciencia y tiempo. Sin embargo, existen técnicas que permiten la optimización de un filtro a base de usar metodologías de ajuste que permita acercarse a los parámetros ideales del filtros. Una de metodología que permite sintonizar un filtro de forma sencilla es el método de Dishal y es el que vamos a usar para sintonizar un filtro paso banda HAIRPIN para la banda de subida de LTE-UHF.

Esta metodología permite realizar el ajuste de un filtro paso banda acoplado sintonizando tanto de los factores de calidad Qi y Qo que necesita el filtro para ser cargado, como de los factores de acoplamiento Mi,i+1 que acoplarán las diferentes etapas, de forma independiente. Estos parámetros son calculados a través de los parámetros del filtro prototipo, que se pueden obtener ya sea a través de las tablas presentes en cualquier libro de diseño de filtros como en programas de cálculo como MatLab. Las expresiones para calcular los parámetros fundamentales de un filtro paso banda acoplado son

$Q_i=\dfrac {g_0g_1}{FBW}$

$Q_o=\dfrac {g_ng_{n+1}}{FBW}$

$M_{i,i+1}=\dfrac {FBW}{\sqrt{g_ig_{i+1}}}$

$FBW=\dfrac {f_h-f_l}{f_0}$

$f_0^2=f_hf_l$

donde fh y fl son las frecuencias de corte de la banda pasante, f0 es la frecuencia central y FBW el ancho de banda fraccional. Los valores g0..gn son los coeficientes del filtro prototipo normalizado. Con estos valores obtendremos los parámetros de acoplamiento de nuestro filtro.

FILTRO PASO BANDA HAIRPIN DE 5 SECCIONES

Vamos a desarrollar un filtro paso banda en tecnología microstrip, usando una configuración HAIRPIN de resonadores λ/2 acoplados. En este filtro, la línea resonante es una línea λ/2, que se acopla al siguiente resonador mediante la sección λ/4. O más concretamente, entre un 85 y un 95% de λ/4. Su denominación HAIRPIN es debida a que tiene forma física de peine. Nuestro filtro va a tener las siguientes características fundamentales:

• Banda pasante : 791÷821MHz (banda de UHF para LTE de subida)
• Número de secciones: 5
• Tipo de filtro: Chebychev 1
• Impedancias de generador y carga: 50Ω

Con estos valores acudimos a las tablas para obtener los coeficientes g0..g6 del filtro prototipo y aplicando las expresiones anteriores obtenemos que

• Qi=Qo=30,81
• M12=M45=0,0297
• M23=M34=0,0226

Con estos coeficientes se pueden calcular las impedancias Zoe y Zoo que definirán las líneas acopladas, así como la posición de los feeds de entrada y salida. En este último caso, esta posición se puede obtener a partir de

$t=\dfrac {\lambda}{4\pi} \sin^{-1} \sqrt {\dfrac {\pi}{2Q_{i/o}}\dfrac {Z_G}{Z_L}}$

Como soporte vamos a usar un substrato Rogers, el RO3006, que tiene una εr=6,15, usando un espesor de 0,76mm y 1oz de cobre (35μm). Con este substrato, el filtro obtenido es:

y con estos valores, pasaremos a la simulación.

SIMULACIÓN DEL FILTRO PASO BANDA

Usando HPMomentum, el simulador electromagnético de ADS, vamos a poder simular la respuesta de este filtro, que se puede ver en la siguiente gráfica

Resultado de la simulación del filtro

que, la verdad sea dicha, no se nos parece ni por asomo a lo que pretendíamos realizar. El filtro está cerca de la frecuencia f0, tiene un ancho de banda de 30MHz, pero ni está centrado ni el rizado es, ni de lejos, 0,1dB. Por tanto, habrá que recurrir a una sintonía usando el método de Dishal y así llevar el filtro a la frecuencia deseada, con el acoplamiento deseado.

AJUSTANDO EL Q EXTERNO

En primer lugar vamos a ajustar los factores de calidad de los resonadores de generador y de carga, que tienen que ser de 30,81. Como ambos son iguales, la sintonía obtenida servirá para los dos. Para ajustar los Qi y Qo, tendremos que buscar la posición adecuada de la alimentación para que el valor sea el deseado.

Para calcular el Qext, se evalúa el coeficiente de reflexión del resonador y se obtiene su retardo de grupo. El factor de calidad será

$Q_{ext}=- \dfrac {d(phase(S_{11}))}{df}f_0$

Cuando hacemos la primera simulación y representamos Qext, obtenemos

donde se puede comprobar que ni el filtro está centrado ni su factor de calidad es el deseado. Para centrar el filtro, aumentamos la distancia entre las líneas en 1,1mm y recortamos las líneas resonantes en 0,34mm. De este modo, obtenemos

en el que ya están centradas las líneas, siendo el Qext de 37,28. Ahora aumentamos la distancia del feed al extremo de la pista en 0,54mm y obtenemos el Qext deseado.

Ya tenemos centrado el filtro y con el Qext requerido. Ahora tocaría ajustar los acoplamientos.

AJUSTE DE LOS ACOPLAMIENTOS

Para ajustar los acoplamientos, primero separamos el feed unos 0,2mm de la línea, y hacemos un espejo de la misma para que quede como sigue

En este caso, para medir el acoplamiento usamos los picos que salen en la transmisión (S21), y aplicamos la expresión

$M=\dfrac {f_h^2-f_l^2}{f_h^2+f_l^2}$

El resultado de la simulación, para el primer acoplo, es

que como podemos comprobar está en el valor requerido.

En el caso del segundo acoplo

que también está cerca de su valor requerido. Por tanto, con los cambios obtenidos, simulamos el filtro total y obtenemos

Filtro después de la primera sintonía

que ya se acerca al filtro deseado.

REITERANDO LA SINTONÍA

Si reiteramos sobre la sintonía, podremos llegar a mejorar el filtro hasta los valores que deseemos. Así, disminuyendo el Qext obtenemos

Disminución del Qext

que supone ya una mejora importante. Jugando ahora con los acoplamientos, disminuyéndolos, llegamos a obtener

Ajuste de los acoplamientos

que podemos dar por válido. Por tanto, el método de Dishal nos ha permitido, a partir de los parámetros calculados, ajustar el filtro hasta obtener las características deseadas.

CONCLUSIONES

Hemos analizado el método de Dishal como herramienta para el ajuste y sintonización de un filtro paso banda de 5 secciones, con óptimos resultados. La sencillez del método permite ajustar los principales parámetros de forma independiente, de manera que el ajuste final u optimización sean más sencillas, cosa de agradecer en simuladores electromagnéticos, que requieren de potencia de cálculo y tiempo de simulación. Vemos que el método, realizado paso a paso, nos permite ir ajustando las características hasta obtener el resultado deseado, por lo que podemos concluir que es un método muy útil en sintonización de filtros, tanto en discretos como en distribuidos, y que bien usado permite acercarse lo suficientemente al resultado final como para que la optimización electromagnética sea innecesaria.

REFERENCIAS

1. Zverev, Anatol I., “Handbook of Filter Synthesys”, Hoboken, New Jersey : John Wiley & Sons Inc., 1967. ISBN 978-0-471-74942-4.

# Statistical analysis using Monte Carlo method (II)

In the previous post, some single examples of the Monte Carlo method were shown. In this post it will be deeply analyzed, making a statistical analysis on a more complex system, analyzing its output variables and studying the results so that they will be quite useful. The advantage of simulation is that it is possible to get a random generation of variables, and also a correlation between variables can be set, achieving different effects in the analysis of the system performance. Thus, any system not only can be analyzed statistically using a random generation of variables, but also this random generation can be linked in a batch analysis or failures in production and in a post-production recovery.

The circuits studied in the previous post were very simple circuits, allowing to see the allocation of random variables and their results when these random variables are integrated a more complex system. With this analysis, it is possible to check the performance and propose corrections which would limit statistically the variations in the final system.

In this case, the dispersive effect of the tolerances will be studied on one of the circuits where it is very difficult to achieve an stability in its features: an electronic filter. An electronic filter, passband type, will be designed and tuned to a fixed frequency, with a certain bandwidth in passband and stopband, and several statistical analysis will be done on it, to check its response with the device tolerances.

DESIGN OF THE BANDPASS FILTER

A bandpass filter design is done, with a 37,5MHz center frequency, 7MHz pass bandwidth (return losses ≥14dB) and a 19MHz stopband bandwidth (stopband attenuation >20dB). When the filter is calculating, three sections are got, and its schematic is

3-sections bandpass filter

With the calculated values of the components, standard values which can make the filter transfer function are found, and its frequency response is

Bandpass filter frequency response

where it is possible to check that the center frequency is 37.5 MHz, the return losses are lower than 14dB at ± 3.5Mhz of the center frequency, and the stopband width is 18,8MHz, with 8,5MHz from the left of the center frequency and 10,3MHz to the right of the center frequency.
Then, once the filter is designed, a first statistical analysis is done, considering that the capacitor tolerance is ± 5% and the inductors are adjustable. In addition, there is not any correlation between the random variables, being able to take an random value independently.

STATISTICAL ANALYSIS OF THE FILTER WITHOUT CORRELATION BETWEEN VARIABLES

As it could be seen in the previous post, when there are random variables there is an output dispersion, so limits to consider a valid filter must be defined, from these limits, to analyze its valid frequency response. Yield analysis is used. This is an analysis using the Monte Carlo algorithm that it allows  to check the performance or effectiveness of the design. To perform this analysis, the limits-for-validation specifications must be defined. The chosen specifications are return losses >13,5dB at 35÷40MHz, with a 2 MHzreduction in the passband width and an attenuation >20dB at frequencies ≤29MHz and ≥48MHz. By statistical analysis, it is got

Statistical analysis of the filter . Variables without correlation.

whose response is bad: only 60% of possible filters generated by variables with a ±5% tolerance could be considered valid. The rest would not be considered valid by a quality control, which would mean that 40% defective material should be returned to the production, to be reprocessed.

It can be checked in the graph that the return loss are the primarily responsible for this bad performance. What could it be done to improve it? In this case, there are 4 random variables. However, two capacitors have of the same value (15pF), and when they are assembled in a production process, usually belong to the same manufacturing batch. If these variables show no correlation, variables can take completely different values. When they are not correlated, the following chart is got

C1, C3 without correlation

However, when these assembled components belong to the same manufacturing batch, their tolerances vary always to the same direction, therefore there is correlation between these variables.

STATISTICAL ANALYSIS OF THE FILTER WITH CORRELATION BETWEEN VARIABLES

When the correlation is used, the influence of tolerances is decreased. In this case, it is not a totally random process, but manufacturing batches in which the variations happen. In this case, it is possible to put a correlation between the variables C1 and C3, which have the same nominal value and belong the same manufacturing batch, so now the correlation graph is

C1, C3 with correlation

where the variation trend in each batch is the same. Then, putting a correlation between the two variables allows studying the effective performance of the filter and get

Statistical analysis with correlation in C1, C3

that it seems even worse. But what happens really? It must be taken into account that the variable correlation has allowed analyzing complete batches, while in the previous analysis was not possible to discern the batches. Therefore, 26 successful complete manufacturing processes have been got, compared to the previous case that it was not possible to discern anything. Then, this shows that from 50 complete manufacturing processes, 26 processes would be successful.

However, 24 complete processes would have to be returned to production with the whole lot. And it remains really a bad result. But there is a solution: the post-production adjustment.

As it was said, at this point the response seems very bad, but remembering that the inductors had set adjustable. What happens now? Doing a new analysis, allowing at these variable to take values in ±10% over the nominal value, and setting the post-production optimization in the Monte Carlo analysis and voilà! Even with a very high defective value, it is possible to recover 96% of the filters within the valid values.

Statistical analysis with post-production optimization

So an improvement is got, because the analysis is showing that it is possible to recover almost all of the batches with the post-production adjustment, so this analysis allows showing not only the defective value but also the recovery posibilities.
It is possible to represent the variations of the inductors (in this case corresponding to the serial resonances) to analyze what is the sensitivity of the circuit to the more critical changes. This analysis allows to set an adjustment pattern to reduce the adjustment time that it should have the filter.

Analysis of the adjustment patterns of the serial resonance inductors

So, with this analysis, done at the same time design, it is possible to take decisions which set the patterns of manufacturing of the products and setting the adjustment patterns for the post-production, knowing previously the statistic response of the designed filter. This analysis is a very important resource before to validate any design.

CONCLUSIONS

In this post, a more grade in the possibilities of using Monte Carlo statistical analysis is shown, using statistical studies. The algorithm provides optimal results and allows setting conditions for various analysis and optimizing more the design. Doing a post-production adjustment, it is possible to get the recovery grade of the proposed design. In the next post, another example of the Monte Carlo method will be done that allows seeing more possibilities over the algorithm.

REFERENCES

1. Castillo Ron, Enrique, “Introducción a la Estadística Aplicada”, Santander, NORAY, 1978, ISBN 84-300-0021-6.
2. Peña Sánchez de Rivera, Daniel, “Fundamentos de Estadística”, Madrid,  Alianza Editorial, 2001, ISBN 84-206-8696-4.
3. Kroese, Dirk P., y otros, “Why the Monte Carlo method is so important today”, 2014, WIREs Comp Stat, Vol. 6, págs. 386-392, DOI: 10.1002/wics.1314.

# Statistical analysis using Monte Carlo method (I)

When any electronic device is designed, we can use several deterministic methods for calculating its main parameters. So, we can get the parameters that we measure physically in any device or system. These preliminary calculations allow the development and their results are usually agreed with the prediction. However, we know that everything we manufacture is always subject to tolerances. And these tolerances cause variations in the results that often can not be analyzed easily, without a powerful calculation application. In 1944, Newmann and Ulam developed a non-deterministic, statistical method called Monte Carlo. In the following blog post.  we are going to analyze the use of this powerful method for predicting possible tolerances in circuits, especially when they are manufactured industrially.

In any process, the output result is a function of the input variables. These variables generate a response which can be determined, both if the system is linear and if it is not linear. The relationship between the response and the input variables is called transfer function, and its knowledge allows us to get any result concerning the input excitation.

However, it must be taken in account that the input variables are random variables, with their own distribution function, and are subject to stochastic processes, although their behavior is predictable through the Theory of Probability. For example, when we make any measure, we get its average value and the error in which can be measured that magnitude. This allows to limit the environment in which it is correct and decide when the magnitude behaves incorrectly.

For many years, I have learned to successfully transform the results obtained by simulations in real physical results, with predictable behavior and I got valid conclusions, and I have noticed that in most cases the use of the simulation is reduced to get the desired result without studying the dependence of the variables in that result. However, most simulators have very useful statistical algorithms that, properly used, allow to get a series of data that the designer can use in the future, predicting any system behavior, or at least analyzing what it can happen.

However, these methods are not usually used. Either for knowledge lack of statistical patterns, or for ignorance of how these patterns can be used. Therefore, in these posts we shall analyze the Monte Carlo method on circuit simulations and we shall discover an important tool which is unknown to many simulator users.

DEVICES LIKE RANDOM VARIABLES

Electronic circuits are made by simple electronic devices, but they have a statistical behavior due to manufacturing. Device manufacturers usually show their nominal values and tolerances. Thus, a resistance manufacturer not only publishes its rating values and its dimensions. Tolerances, stress, temperature dependance, etc., are also published. These parameters provide important information, and propertly analyzed with a powerful calculation tool (such as a simulator), we can predict the behavior of any complex circuit.

In this post, we are going to analyze exclusively the error environment around the nominal value, in one resistor. In any resistor, the manufacturer defines its nominal value and its tolerance. We asume these values 1kΩ for the nominal value and ± 5% for its tolerance. It means the resistance value can be found between 950Ω and 1,05kΩ. In the case of a bipolar transistor, the current gain β could take a value between 100 and 600 (i.e. NXP BC817), which may be an important and uncontrollable variation of current collector. Therefore, knowing these data, we can analyze the statistical behavior of an electronic circuit through the Monte Carlo method.

First, let us look resistance: we have said that the resistance has a ± 5% tolerance. Then, we will analyze the resistor behavior with the Monte Carlo method, using a circuit simulator. A priori, we do not know the probability function, although most common is a Gaussian function, whose expression is well known

$f_{\mu,\sigma^2}(x)=\dfrac {1}{\sigma \sqrt {2 \pi}}e^{\dfrac {(x-\mu)^2}{\sigma^2}}$

being μ the mean and σ² the variance. Analyzing by the simulator, through Monte Carlo method and with 2000 samples, we can get a histogram of resistance value, like it is shown in the next figure

Histogram of the resistor

Monte Carlo algorithm introduces a variable whose value corresponds to a Gaussian distribution, but the values it takes are random. If these 2000 samples were taken in five different processes with 400 samples each one, we would still find a Gaussian tendency, but their distribution would be different

Gaussian distributions with different processes

Therefore, working properly with the random variables, we can get a complete study of the feasibility of any design and the sensitivity that each variable shows. In the next example, we are going to analyze the bias point of a bipolar transistor, whose β variation is between 100 and 600, being the average value 350 (β is considered a Gaussian distribution), feeding it with resistors with a nominal tolerance of ± 5% and studying the collector current variation using 100 samples.

STATISTICAL ANALYSIS OF A BJT BEHAVIOR IN DC

Now, we are going to study the behavior of a bias circuit, with a bipolar transistor, like the next figure

Bias point circuit of a BJT

where the resistors have a ±5% tolerance and the transistor has a β variation between 100 and 600, with a nominal value of 350. Its bias point is  Ic=1,8mA, Vce=3,2V. Making a Monte Carlo analysis, with 100 samples, we can get the next result

BJT current distribution respect to the random variables

Seeing the graph form, we can check that the result converges to a Gaussian distribution, being the average value Ic=1,8mA and its tolerance, ±28%. Suppose now that we do the same sweep before processing, in several batches of 100 samples each one. The obtained result is

BJT current distribution respect several batches

where we can see that in each batch we get a graph which converges to a Gaussian distribution. In this case, the Gaussian distribution has an average value μ=1,8mA and a variance σ²=7%. Thus, we have been able to analyze each process not only like a global statistical analysis but also like a batch. Suppose now that β is a random variable with an uniform distribution function, between 100 and 600. By analyzing only 100 samples, the next graphic is got

Results with a BETA uniform distribution

and it can be seen that the current converges to an uniform distribution, increasing the current tolerance range and the probability at the ends. Therefore, we can also study the circuit behaviour when it shows different distribution functions for each variable.

Seeing that, with the Monte Carlo method, we are able to analyze any complex circuit behavior in terms of tolerances, in the same way it will help us to study how we could correct those results. Therefore, in the next posts we shall analyzed deeply this method, increasing the study of its potential and what we can be achieved with it.

CORRECTING THE TOLERANCES

In the simulated circuit, when we have characterized the transistor β like an uniform random variable, we have increased the probability into unwanted current values (at the ends). This is one of the most problematic features, not only on bipolar transistors but also on field effect transistor: the variations of their current ratios. This simple example let see what happens when we use a typical correction circuit for the β variation, like the classic polarization by emitter resistance.

Bias circuit by emitter resistance

Using this circuit and analyzing by Monte Carlo, we can compare its results with the analysis obtained in the previous case, but using 1000 samples. The result is

Results with both circuits

where we can check that the probability values have increased around 2mA, reducing the probability density at the low values of current and narrowing the distribution function. Therefore, the Monte Carlo method is a method that not only enables us to analyze the behavior of a circuit when subjected to a statistical, but also allow us to optimize our circuit and adjust it to the desired limit values. Used properly, it is a powerful calculation tool that will improve the knowledge of our circuits.

CONCLUSIONS

In this first post, we wish to begin a serie dedicated to Monte Carlo method. In it, we wanted to show the method and its usefulness. As we have seen in the examples, the use of Monte Carlo method provides very useful data, especially with the limitations and variations of the circuit we are analyzing if we know how they are characterized. On the other hand, it allows us to improve it using statistical studies, in addition to setting the standards for the verification of in any production process.

In the next posts we shall go more in depth on the method, by performing a more comprehensive method through the study of a specific circuit of one of my most recent projects, analyzing what the expected results and the different simulations that can be performed using the method of Monte Carlo, like the worst case, the sensitivity, and the post-production optimization.

REFERENCES

1. Castillo Ron, Enrique, “Introducción a la Estadística Aplicada”, Santander, NORAY, 1978, ISBN 84-300-0021-6.
2. Peña Sánchez de Rivera, Daniel, “Fundamentos de Estadística”, Madrid,  Alianza Editorial, 2001, ISBN 84-206-8696-4.
3. Kroese, Dirk P., y otros, “Why the Monte Carlo method is so important today”, 2014, WIREs Comp Stat, Vol. 6, págs. 386-392, DOI: 10.1002/wics.1314.

# A 900Mhz Feedfordward Amplifier with MOSFET

In this article, we are going to demonstrate a 900Mhz feedforward amplifier design. Feedforward is a linearization technique for the IM distortion, caused by the nonlinear feature of the active device. Lateral interference, generated by the IM distortion on both sides of the main frequency, affecting the Adjacent Channel. Decreasing this interference is the purpose of this entry.

We are going start with a LDMOS amplifier, tuned to 900Mhz. The active device is a STMicroelectronics’ MOSFET, PD84001, It
operates at 8 V in common source mode at frequencies of up to 1 GHz. POUT is 31dBm (IDQ=50mA) and Drain Efficiency, 60%. Once we have designed the amplifier, we’ll make the linearization of the IM products, using the feedforward technique.

THE MOSFET AMPLIFIER

The amplifier is designed in common source mode. The Operating Point is chosen as the optimal features of the manufacturer: VDS=8V, IDQ=50mA . At this OP, and at 900MHz, ZIN=3,6+j·4,3Ω and ZOUT= 3,9 + j·5,5Ω. Maximum power transfer is obtained with a conjugate matching network at the generator and load impedances, which are Z0=50. Once the matching networks are calculated, the purposed schema for the amplifier is

PD84001 Two-Stage Amplifier

Amplifier’s gain is 34,3dB, and its phase is 76,4deg. Input and Output Return Losses are respectively 30,7 and 39,8dB. Then, the amplifier is matched and the maximum power at 900MHz is 27dBm, for 1-Tone.

2 Tone Output Power and TOD vs. Input Power, at 900Mhz

For a 2-Tone input, the IM distortion generates a power drop, caused by the Third Order Distortion. TOI (Third Order Intercept) is 31,7dBm, near of the maximum output power of the datasheet, and it causes the power drop.

Intermodulation products are 12dB below the carrier, and this value may cause interference on the Adjacent Channel. Therefore, we must reduce this value as much as possible, using a linearization technique.

There are many linearization techniques, but we are going to use the feedforward technique, because it is a technique that requires only the use of RF networks.

The amplifier gain, including the second and third order distortions, could be expressed by

$P_o= \hat{g} (P_i)=\hat{g}_1 P_i+\hat {g}_2 P_i^2+\hat {g}_3 P_i^3$

with

$\hat{g}_1=g_1e^{j \theta_1}$

$\hat{g}_2=g_2e^{j \theta_2}$

$\hat{g}_3=g_3e^{j \theta_3}$

Where the input signal Pi is a 2-Tone signal. In this case, we will not take into consideration the second order distortion, since the Pfrequencies will be very close together. A bandpass filter could remove the second order spurious.

The amplifier gain is complex, The coefficients g1 and g3 could be expressed using the polar notation (in mag/phase). Then, these are

$\hat{g}_1=2961,54 \cdot e^{j 76,4}$

$\hat{g}_3=11,43 \cdot e^{-j 95,3}$

These coefficients are going to use to calculate the phase shifter of the first stage. Now, we shall describe shortly the feedforward technique.

FEEDFORWARD PRINCIPLE

The Feedfordward Principle is based on reducing the distortion by mixing in phase opposition with the same distortion. In a RF amplifier, an output distorted signal is generated due to the active device’s nonlinearity. It could be mixed with the input signal in phase opposition, adjusting the levels of both signals.  So, we get the distorted signal on one port, and on the other port, only the distortion spurious.

Cancellation of the main signals on the second port is achieved by placing a delay line (τ1), in the secondary network of the first stage. One sample of the signal output of the amplifier (G1) is derived to combine with the secondary network, with a combiner. The levels of both signals are equalized by an inter-stage attenuator (β). Then, both signals are combined. Then, the ouput signal of the amplifier is called MAIN, and the combined signal, AUX.

Feedforward Principle

AUX is now used as an error signal in the second stage, which is amplified by an error amplifier (G2), while the MAIN is delayed with another delay line (τ2). In this second stage, we want to get the same effect than the first stage: put both signals in phase opposition, and combine them. Then, the distortion is cancelled and reduced the interference on the Adjacent Channel.

The level at the output of the amplifier could be written as

$P_{MAIN}= \hat{g} \left( \dfrac {P_i}{2} \right)=g_1e^{j \theta_1} \dfrac {P_i}{2}+g_2e^{j \theta_2} \dfrac {P_i^2}{4}+g_3e^{j \theta_3} \dfrac {P_i^3}{8}$

and PAUX1 (the sample level before the error combiner) could be expressed by

$P_{AUX_1}=\beta g_1e^{j \theta_{A1}} \dfrac {P_i}{2}+\beta g_2e^{j \theta_{A2}} \dfrac {P_i^2}{4}+\beta g_3e^{j \theta_{A3}} \dfrac {P_i^3}{8}$

and $\theta_{Ai}=\theta_i + \theta_{\beta}$

The level PAUX2 at the secondary network is

$P_{AUX_1}=\dfrac {P_i}{2} e^{j{\theta}_{A2}}$

In these expressions, β is the magnitude of the losses of the inter-stage attenuator and θβ is its phase; and θA2 is the phase of the delay line τ1. It must be satisfied

$\theta_1+\theta_{\beta}=\theta_{A2} \pm 180$

$\dfrac {d \theta_1}{d \omega}+\dfrac {d \theta_{\beta}}{d \omega}=\dfrac {d \theta_{A2}}{d \omega}$

θ1 is the phase of the linear gain of the amplifier. Then, not only the phases must be in phase opposition, but also the delay time must be the same in every subnetworks. In magnitude, it must be satisfied |β·g1|=1.

In the second stage, the gain of the amplifier must equalize the g2 and g3 levels, and their phases must satisfy the same equations (absolute phase and delay time) of the first stage, to combine and cancel the distortions.

$\left| \beta G_2 \right|=1$

$\theta_{\beta}+\theta_{G_2}=\theta_{\tau 2} \pm 180$

$\dfrac {d \theta_{\beta}}{d \omega}+\dfrac {d \theta_{G_2}}{d \omega}=\dfrac {d \theta_{\tau 2}}{d \omega}$

In RF designs, the adders must be replace by hybrid couplers or directional couplers, which have insertion and coupler losses. Using two hybrid couplers (3dB for insertion losses) to split Pi and combine PAUX1 and PAUX2, and two directional couplers (with C for coupler losses) to take the sample in the first stage and combine the error sample in the second stage, the expressions are now

$\left| \dfrac {\beta}{IL_{hyb} C_{coup}}g_1 \right|=1$

$\theta_1+\theta_{\beta}+\theta_{coup}=\theta_{A2} \pm 180$

$\dfrac {d \theta_1}{d \omega}+\dfrac {d \theta_{\beta}}{d \omega}+\dfrac {d \theta_{coup}}{d \omega}=\dfrac {d \theta_{A2}}{d \omega}$

at the first stage and

$\left| \dfrac {\beta}{IL_{hyb}^2 C_{coup}^2}G_2 \right|=1$

$\theta_{\beta}+2 \theta_{coup}+\theta_{G2}=\theta_{\tau 2} \pm 180$

$\dfrac {d \theta_{\beta}}{d \omega}+2\dfrac {d \theta_{coup}}{d \omega}+\dfrac {d \theta_{G2}}{d \omega}=\dfrac {d \theta_{\tau 2}}{d \omega}$

at the second stage.

In a narrowband amplifier, delay time could not be considered, because its phase slope will be smallest than the phase slope of a broadband amplifier.

900MHz FEEDFORDWARD AMPLIFIER

Now, we are going to design our feedforward amplifier, based in our two-stage LDMOS amplifier. In first, we must split the input signal in two outputs, one to the amplifier and the other to the phase shifter. We are going to use an 180-deg hybrid coupler, with 3dB of insertion losses. At this frequencies, hybrid couplers could be easily found in the market, as a Surface Mounting Device (SMD). The designed amplifier is a narrowband amplifier.

The output levels of the hybrid coupler are the same, in magnitude and phase. The phase of the amplifier gain was 76,4deg in linear mode, but in nonlinear mode, we have got a phase of 69,4deg, with 0dBm of input power. Taking a sample of the output level of the amplifier with a directional coupler, which introduces a 90deg coupling phase, with 10dB of coupling level, we have got a sample level of 12dBm, with a phase of 159,4deg.

Then, we are going to combine with another hybrid coupler, and as in the secondary network the level is -6dBm, we have to equalize both levels with the attenuator, whose attenuation must be ≈20dB. The phase shifter should be adjusted to a phase of ≈-12 deg.

Adjusting the phase and the level with the phase shifter and the attenuator, we are able to optimize the response for several input levels.

AUX Main an Intermodulation Frequencies, vs Input Power

We are going to complete now the second stage amplifier, where an error amplifier increases the level of AUX spurious intermodulation to combine in phase opposition with the MAIN line. The error amplifier G2 should not be a power amplifier, at this stage. A linear, general-purpose amplifier maybe used. The gain is calculated by the difference between the MAIN and AUX IM spurious. This value is 45,5dB, because we are combining with a directional coupler, to reduce the insertion losses in the MAIN line. Using an amplifier with a magnitude of 45,5dB and a phase of -145deg, we have got a phase shifter with the same value, and after the coupler, the IM distortion decreases around 65dB. The output level is now 31dBm, and the TOI increases to 75dBm.

Output level and TOD, after the feedforward correction

The definitive amplifier is

Definitive Feedforward Amplifier

CONCLUSIONS

With the amplifier designed we have achieved a significant improvement: increasing efficiency around 40dB for the same output level, on adjacent channel. Furthermore, the amplifier is very simple to realize with a few RF devices. The design is very easy and intuitive.

However, the Feedforward has two serious disadvantages: on the PCB, it needs a lot of surface; and the input level cannot be increased above the input level that provides maximum output level of the MOSFET, because the distortion can be increased above the value we have corrected.

In broadband we must take into consideration not only the phase of the amplifiers but also the group delay, because the phase slope of the amplifiers has to be compensated by the phase shifter. Then, the phase shifter could have a larger surface dimensions, because it must be a delay line, too.

References

1. R. Cordell, “A MOSFET Power Amplifier with Error Correction”; JAES, vol. 32, nr. 1/2, 1984 Jan/Feb
2. J. Vanderkooy, S.P. Lipshitz, “Feed-Forward Error Correction in Power Amplifiers”, JAES, Vol. 28, Nr. 1/2, 1980 Feb
3. A.M. Sandman, “Reducing Distortion by ‘Error add-on‘”, Wireless World, vol.79, p.32, 1974 Oct