circuit_modes_v5

1

Abstract — We provide a local linear relation between the
circuit parameters and the poles and response of an RLC
network. The sensitivity matrix which defines this relationship is
an important tool in causal analyses used in “six-sigma” methods.
It also enables fast yield predictions and co-variance estimations
or, more generally, can be used as a basis for DOE studies. The
sensitivity matrix is obtained from first-order derivatives of the
poles and transfer function. These derivatives are calculated
from the eigenvalues and eigenvectors of the network. The
method is illustrated on a simple example.
Index Terms—RLC circuits, Computer aided engineering,
Sensitivity analysis, Yield estimation, DOE
I. INTRODUCTION
ENSITIVITY analysis is an integral part of computer-
aided circuit design. Efficient statistical analysis
algorithms are available to simulate the circuit response at
fixed frequencies [1], from which the sensitivity to
components variation may be obtained by regression. In the
context of resonant circuits, however, the designer is primarily
interested in the poles and the circuit response on-resonance.
Tracking the resonances at each trial requires an extra
computational step that undermines the efficiency of existing
methods.
In this paper, we follow the reverse procedure: we first
determine the local linear relationship between the circuit
parameters and the poles and response of the system, then
carry out “primitive” Monte Carlo simulations. For yield
predictions, which require a large number of trials, the fixed
cost associated with the determination of the sensitivity matrix
is lower than the recurring costs in existing approaches.
We define the sensitivity matrix as the Jacobian of the
transformation between the circuit parameters and the poles
and system response [2]. We present a general method to
calculate the sensitivity matrix based on the solution of the
eigenvalue problem associated with the circuit. This method is
illustrated with a simple example.
II. BACKGROUND
Consider the transfer function of an arbitrary network of
resistances, inductors, and capacitors. The observable response
of the circuit is entirely characterized by the poles
and the values of the transfer function on-resonance
Olivier Buu is with the Agilent Technologies, Santa Clara, CA 95051 USA
(e-mail: olivier.buu@agilent.com).
1. Formally, there exists a complicated relationship
between the circuit parameters [ ] and the
column vector formed by the real and imaginary parts of the
poles and each independent component of the transfer function
at each resonance frequency. To study the sensitivity to small
component variations , we follow [2] and linearize this
relation:
[ ] (1)
where [ ] denotes the matrix transpose. In practice, only a
subset of the observable parameters may be under
specification, and the size of the sensitivity matrix is reduced
accordingly.
Although the calculation of the Jacobian may be
computationally costly for large systems, the simple linear
relation (1) allows for efficient statistical analysis of the
circuit. In some cases, the Jacobian gives direct access to the
multivariate distribution function. In particular, if the relation
between and is bijective, the probability density function
associated with is known locally from the relation:
| | (2)
where is the probability density function associated with the
random vector . More generally, when the second moment of
the distribution of the circuit parameters exists, the co-
variance matrix for the random vector is given by:
[ ] [ ] (3)
where is the co-variance matrix of the vector .
III. SENSITIVITY MATRIX CALCULATION
The sensitivity matrix is assembled from the derivatives of
the poles and the transfer function with respect to the circuit
parameters. In this section, we review the perturbation method
used to calculate these derivative terms based on the solution
of the circuit eigenvalue problem. The eigenvalue problem is
formulated from the circuit state equation.
A. State Equation
The circuit equation for a Linear Time-Invariant network is
assumed to take the standard form:
̇ (4)
1
This list may include a pole at infinity.
Sensitivity Analysis of Resonant Circuits
Olivier Buu
S

2
where the vector includes the node voltages at the capacitors
terminals and the currents flowing through the inductors. is
the input matrix and is the excitation vector. is a matrix
composed of capacitance and inductance values. The matrix
includes the resistance values and the node-branch incidence
matrix describing the network under study. Both and
may be written as sums of sparse matrices corresponding to
the individual component contributions. These sparse
matrices, sometimes called “component stamps” in the
literature [3], are useful for the sensitivity analysis presented
below. Provided does not contain any linearly-dependent
variables, then (4) is a state equation and the matrix is full-
rank. We will assume this condition to be fulfilled in the rest
of this paper.
The transfer function is determined by the output equation:
(5)
where is the output vector, is the output matrix, and is
the transmission matrix. By taking the Laplace transform of
(3) and (4) and applying the definition of the transfer function
to the zero-state output vector, we obtain:
(6)
where is the Laplace variable.
B. Generalized eigenvalue problem
The circuit is entirely characterized by the eigenvalues and
eigenvectors of the state equation. The square matrices of right
eigenvectors [ ] and left eigenvectors
[ ] are solutions of the generalized eigenvalue
problem:
and (7)
where [ ] is a diagonal matrix of eigenvalues.
Since and are real matrices, the eigenvalues are either
real or complex conjugate pairs. For passive networks, all the
eigenvalues are located in the left half of the complex plane.
From (7), it can be shown that the eigenvectors are bi-
orthogonal. Since is assumed to be non-singular, we can
always find a normalization such that:
and (8)
where is the identity matrix. Using (5) and (7) we obtain the
following expression for the transfer function:
(9)
Together with the eigenvalues, this last expression forms
the basis of the sensitivity analysis described in the next
sections.
C. Derivative of Eigenvalues
For simple eigenvalues, the derivatives with respect to a
circuit parameter is [4]:
[ ] [ ] (10)
As noted above, the matrix derivatives and are sparse
and closely related to the “stamp” for the circuit component
parameterized by .
The case of multiple eigenvalues is addressed in [5]: for an
eigenvalue of multiplicity with associated
eigenvectors [ ] and [ ],
there are derivatives which are the eigenvalues of the matrix
[ ] .
D. Derivative of the Transfer Function
The derivative of the on-resonance transfer function
includes two terms:
[ ] [ ] [ ] (11)
where is the imaginary part of the nth
pole. To obtain these
two terms we introduce some intermediate calculation steps:
(12)
(13)
(14)
In the previous expressions, we have assumed that the
resonance of interest is damped, so the matrix is
non-singular. Note that (12) does not require a full matrix
inversion. The right-hand-side terms of (11) follows from
derivatives of (6):
[ ] (15)
[ ]
[ ] (16)
IV. COMPUTATIONAL COST
The most costly step of the sensitivity matrix calculation is
the solution of the eigenvalue problem, which scales as
operations. If the probability density function can be
obtained, from equ. (2) or otherwise, the cost of a Monte Carlo
trial is the cost of sampling the distribution. In the worst case,
an additional matrix multiplication (equ. (1)) is required,
.
By contrast, the cost of setting up a quadratic approximation
of the response function at a given frequency is . Since
the number of degrees of freedom is equal the number of
independent active elements, this cost is equivalent to .
Each simulation involves a matrix multiplication, with a

3
cost. However, the response function has to be
evaluated at different frequencies to characterize the circuit,
so the fixed and recurring costs are, respectively and
in this method. Moreover, there is a cost associated
with tracking the resonance frequencies at each trial.
The standard deviation of yield predictions scales as the
inverse of the square root of the number of trials. For a ~10-2
accuracy on simulation results, we assume a 104
trials run.
With a circuit with 50 eigenmodes, the sensitivity matrix
method would require ~107
operations. The same simulation
would cost ~109
operations to achieve the same accuracy with
the quadratic approximation.
V. EXAMPLE
Fig. 1 shows a circuit example used in Magnetic Resonance
instruments [6]. The inductive transducer Lcoil is embedded in
a two-port matching network, where port 1 is tuned to 200
MHz, and port 2 to 50 MHz.
In this RF application, the S-parameters are the preferred
signal representation. However, it is convenient to first
obtain the 2x2 port-admittance matrix and component
stamps by Modified Nodal Analysis. In this case, the
matrices and are:
[ ] and [ ] (17)
The sub-matrices are obtained by inspection:
[ ] (18)
[ ] (19)
[ ] (20)
[ ] (21)
We used Matlab™ to generate the eigenvalues and
eigenvectors, which are reported in table 1. In addition to the
nominal resonances at 200 MHz and 50 MHz, there is a
spurious resonance at 179 MHz and a pole at DC.
TABLE I
EIGENVALUES AND EIGENVECTORS
Channel 1a
Channel 2a
Spuriousa
DC
Frequencyb
199.9 MHz 50.0 MHz 179.0 MHz 0
c
-1.0 0.0 -1.0 1
c
0.0 -0.5 0.0 1
c
0.7 -0.4 -0.5 1
c
0.1 1.0 0.1 1
d
-0.1-j10.8 -j8.2 j10.4 0
d
j2.1 j7.9 j2.4 0
d
j8.9 -j8.4 j2.7 0
a
Right eigenvector from the eigenvalue of positive natural frequency.
b
Each frequency corresponds to a complex-conjugate pair of poles.
c
Voltage nodes in V.
d
Inductor currents in mA.
Because we eliminated the redundant variables
from the state vector, the driving term depends on the time
derivative on the input vector [7] and takes the form ̇ with:
[ ] and [ ] (22)
Similarly, the transmission term in the output equation takes
the form ̇ because the DC mode is not observable [7]. The
coefficients of the output equation are:
and [ ] (23)
[ ] (24)
These modifications to the standard circuit equation add
terms to (15) but do not drastically alter the sensitivity
analysis. The scattering matrix follows from the well-
known relation:
(25)
where 50 Ohm is the characteristic port impedance.
Differentiating (25) yields:
(26)
In this example the specifications are:
| | (27)
| | (28)
| | (29)
| | (30)

4
Correspondingly, we characterize the circuit response with
the vector:
[ ] (31)
where denotes the real part and the imaginary part. The
corresponding sensitivity matrix is compiled in table 2.
TABLE II
SENSITIVITY MATRIX
Component
| | | |
-14.4a
1.78d
-0.1a
0.06d
-5.1a
3.85d
0.0a
0.00d
-5.1a
1.87d
0.0a
0.00d
-2.4a
0.55d
-0.1a
0.06d
0.0a
0.00d
-0.2a
0.08d
0.0a
0.00d
-0.2a
0.71d
-2.4a
0.55d
0.0a
0.00d
0.0a
0.01d
-0.6a
0.00d
-0.6b
0.14e
0.0b
0.00e
0.0b
0.01e
0.0b
0.00e
-0.4b
0.02e
0.0b
0.00e
0.0c
4.14f
0.0c
4.04f
a
In units of MHz/pF.
b
In units of MHz/nH.
c
In units of MHz/Ohm.
d
In units of 1/pF.
e
In units of 1/nH.
f
In units of 1/Ohm.
Assuming a 5% variance and no correlation between the
component values, we can calculate the second moment of the
chosen engineering parameters from the diagonal elements of
the co-variance matrix obtained from (3). The values are
listed in table 3.
TABLE III
PREDICTED YIELD
a
| |b
| |
Variance 4.54 MHz 0.69 1.25 MHz 0.25
Partial Yields 73% 35% 79% 79%
Combined Yields 26% 62%
Total Yield 19%
a
√ √
b
√ | | √
To calculate the yield, we further assume the circuit
parameters to be normally distributed. The distribution of the
random vector is then multivariate normal. Instead of
integrating the 6-dimensional probability density function over
the specification domain, we found it more accurate and faster
to use a Matlab routine to sample the distribution. The various
conditional probabilities estimated by averaging 109
trials are
reported in table 3. A convenient representation of the results
is the area-proportional Venn diagram, which was created with
the VennEuler algorithm [10] and is shown on Fig. 2. The
standard deviation on these figures is ~10-3
.
In this example, the detailed yield analysis points to the
return loss on channel 1 as the largest risk of failure. The
sensitivity matrix suggests that reducing the variance of
, , and would improve the yield. Replacing by
a trimmer capacitor and adding a tuning process is another
solution to the yield issue. These different assumptions and
their economic implications may be tested by re-calculating
the yield with the Monte Carlo method. The experimental
validation of the model may be done by Design of Experiment
(DOE) based on the sensitivity matrix calculated above.
VI. CONCLUSION
In this paper we have shown that the sensitivity matrix
method is an efficient way to carry out the statistical analysis
of resonant circuits. Despite the computational cost of
determining the sensitivity matrix, its value may be realized in
the direct calculation of the co-variance matrix, DOE studies,
or substantiating causal analyses used in “six-sigma” quality
control frameworks.
Our approach may be improved with more efficient
eigenvectors calculation algorithms [8] or faster Monte Carlo
methods [9]. A low-cost approximation of the sensitivity
matrix may also be obtained by the quadratic approximation,
if the eigen-frequencies are known in advance. Finally, this
approach is applicable to any linear time-invariant system, and
may be expanded to other network characterizations, as
described in the appendix.

5
APPENDIX
ALTERNATE CIRCUIT CHARACTERIZATIONS
Engineering specifications may include circuit
characterizations other than the quantities we considered in the
analysis presented above. In this appendix, we provide the
elements of perturbation theory that may be used with other
well-known characterizations. For simplicity, we restrict this
section to the case of circuits with single eigenvalues. The
case of multiple eigenvalues has been worked out [10] but is
outside the scope of this paper.
A. Time Domain Analysis
The circuit natural response may be calculated in terms of
eigenvectors from (4) and (7):
(32)
where is the vector of initial conditions. Since is
diagonal, the calculation of the exponential term presents no
numerical difficulty. The differentiation of (30) with respect to
a circuit parameter involves the derivative of eigenvectors.
Ref. [11] gives their expression as linear combinations of the
un-perturbed eigenvectors in the case of distinct eigenvalues:
∑
[ ]
(33)
An alternate expression involving only one un-perturbed
eigenvector is given by [12].
B. Residues
The matrix may be expanded as a sum of rational
functions:
∑ (34)
The poles and residue matrices provide a complete
characterization of the observable response. By expanding (8)
we can express the residue matrices in terms of eigenvectors:
(35)
Similarly to the time-domain analysis, the perturbation of the
residue matrices involves the derivative of eigenvectors.
C. Zeros
Single Input Single Output systems are often analyzed in
terms of pole-zero loci. The first-order derivative of a zero
may be obtained by differentiating the implicit relation
( )=0 with respect to the circuit parameter :
[ ]
[ ]
(36)
The numerator and denominator are obtained similarly to
(15) and (16). In this case, the vectors and may be
interpreted as the respective solutions of the direct and adjoint
systems at .
REFERENCES
[1] R. Biernacki, J.W. Bandler, J. Song, Q. Zhang, “Efficient quadratic
approximation for statistical design,” IEEE Trans. Circuits and Systems,
vol. 36, no. 11, pp, 1449-1454, Nov. 1989.
[2] A.S. Cook, T. Downs, “Estimating manufacturing yield by means of
Jacobian of transformation, ” IEE Proc. G, vol. 129, issue 4, pp. 127-
133, Aug. 1982.
[3] F. N. Najm, “Circuit Simulation,” Hoboken, NJ, Wiley, 2010, p 35.
[4] S. B. Haley, “The Generalized Eigenproblem: Pole-Zero Computation,”
IEEE Proc., vol. 76, no. 2, pp. 103-120, Feb. 1988.
[5] A. Papoulis, “Perturbations of the Natural Frequencies and Eigenvectors
of a Network,” IEEE Trans. Circuit Theory, vol. 13, Issue 2, pp. 188-
195, Jun. 1966.
[6] M. Lupu, A. Briguet, J. Mispelter, “NMR Probeheads: For Biophysical
and Biomedical Experiments: Theoretical Principles & Practical
Guidelines,” 1st
ed., Imperial College Press, 2006, pp. 209-210.
[7] S. Natarajan, “A systematic method for obtaining state equations using
MNA,” IEE Proc. G, vol. 138, no. 3, pp. 341-346, Jun. 1991.
[8] A. Expósito, A. B. Soler, J. A. R. Macías, “Application of Generalized
Phasors to Eigenvector and Natural Response Computation of LTI
Circuits,” IEEE Trans. Circuits and Systems, vol. 53, no. 7, pp, 1533-
1543, Jul. 2006.
[9] M. Keramat and R. Kielbasa, “Modified latin hypercube sampling
Monte Carlo (MLHSC) estimation for average quality index,” Int. Jour.
Analog Integrated Circuits Signal Process., vol. 19, no. 1, pp. 87–98,
Apr. 1999.
[10] L. Wilkinson, "Exact and Approximate Area-Proportional Circular Venn
and Euler Diagrams," Visualization and Computer Graphics, IEEE
Transactions on , vol.18, no.2, pp.321,331, Feb. 2012.
[11] M. I. Friswell, “The Derivatives of Repeated Eigenvalues and Their
Associated Eigenvectors,” ASME J. Vibration and Acoustics, vol. 118,
Issue 7, pp. 943-944, Jul. 1970.
[12] R. H. Plaut, “Derivatives of eigenvalues and eigenvectors in non-self-
adjoint systems,” AIAA J., vol. 11, Issue 2, pp. 250-251, Feb. 1973.
[13] R. B. Nelson, “Simplified Calculation of Eigenvector Derivatives,”
AIAA J., vol. 14, Issue 9, pp. 1201-1205, Sep. 1976.

circuit_modes_v5

More Related Content

What's hot (18)

Viewers also liked (16)

Similar to circuit_modes_v5 (20)

circuit_modes_v5