PHYSICAL REVIEW E 70, 016608 (2004)

Robust method to retrieve the constitutive effective parameters of metamaterials

Xudong Chen, Tomasz M. Grzegorczyk, Bae-Ian Wu, Joe Pacheco, Jr., and Jin Au Kong
Research Laboratory of Electronics, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA
(Received 25 February 2004; published 26 July 2004)

We propose an improved method to retrieve the effective constitutive parameters (permittivity and perme-
ability) of a slab of metamaterial from the measurement of S parameters. Improvements over existing methods
include the determination of the first boundary and the thickness of the effective slab, the selection of the
correct sign of effective impedance, and a mathematical method to choose the correct branch of the real part of
the refractive index. The sensitivity of the effective constitutive parameters to the accuracy of the S parameters
is also discussed. The method has been applied to various metamaterials and the successful retrieval results
prove its effectiveness and robustness.

DOI: 10.1103/PhysRevE.70.016608 PACS number(s): 41.20.Jb, 42.25.Bs, 78.20.Ci

I. INTRODUCTION ber of the incident wave in free space. The S parameters are
related to refractive index n and impedance z by [7,9,10]:
Left-handed (LH) structures have been realized so far as
metamaterials [1–3] and very quickly, researchers have been R01s1 − ei2nk0dd
working on retrieving their effective permittivity and perme- S11 = , s1ad
1 − R201ei2nk0d
ability to better characterize them [4–6]. Known methods to
date [7,8] use S parameters calculated from a wave incident s1 − R201deink0d
normally on a slab of metamaterial, from which the effective S21 = 2 i2nk0d , s1bd
refractive index n and impedance z are first obtained. The 1 − R01 e
permittivity e and permeability m are then directly calculated
from m = nz and e = n / z.
/
where R01 = z − 1 z + 1 .
As it has been pointed out in [4,5], the refractive index n
It is also known that this retrieval process may fail in and the impedance z are obtained by inverting Eqs. (1a) and
some instances, such as when the thickness of the effective (1b), yielding
slab (exhibiting bulk properties) is not accurately estimated
[4] or when reflection sS11d and transmission sS21d data are
very small in magnitude [6]. Although these issues have been
z= ± Î s1 + S11d2 − S221
s1 − S11d2 − S221
, s2ad
addressed to some extent in previous works, we have found
that the retrieved results in some cases are still unsatisfac- eink0d = X ± iÎ1 − X2 , s2bd
tory. In particular, the determination of the first boundary of
the effective homogeneous slab, the selection of the sign of where X = 1 / 2
2S21 s1 − S11 2
+ S21 d.
Since the metamaterial un-
the roots when solving for the impedance z, the determina- der consideration is a passive medium, the signs in Eqs. (2a)
tion of the branch of the real part of refractive index n, and and (2b) are determined by the requirement
the origin of the spikes appearing in the retrieved impedance
z8 ù 0, s3ad
z, are many problems that deserve further investigation. The
aforementioned issues are addressed in the next sections of
this paper and some typical retrieval results are presented to n9 ù 0, s3bd
show the robustness and effectiveness of the method. Note where s·d and s·d denote the real part and imaginary part
that the values of e, m, and z are relative to those in free 8 9
operators, respectively. The value of refractive index n can
space, thus dimensionless. be determined from Eq. (2b) as
1
II. RETRIEVAL METHOD n= hfflnseink0ddg9 + 2mpg − iflnseink0ddg8j, s4d
k 0d
A. Theoretical retrieval equations
where m is an integer related to the branch index of n8. As
In order to retrieve the effective permittivity and perme- mentioned above, the imaginary part of n is uniquely deter-
ability of a slab of metamaterial, we need to characterize it as mined, but the real part is complicated by the branches of the
an effective homogeneous slab. In this case, we can retrieve logarithm function.
the permittivity and permeability from the reflection sS11d Equations (2a) and (2b) can be directly applied in the case
and transmission sS21d data. For a plane wave incident nor- of a homogeneous slab for which the boundaries of the slab
mally on a homogeneous slab of thickness d with the origin are well defined and the S parameters are accurately known.
coinciding with the first face of the slab, S11 is equal to the However, since a metamaterial itself is not homogeneous, the
reflection coefficient, and S21 is related to the transmission two apparently straightforward issues mentioned above need
coefficient T by S21 = Teik0d, where k0 denotes the wave num- to be carefully addressed. First, the location of the two

1539-3755/2004/70(1)/016608(7)/$22.50 70 016608-1 ©2004 The American Physical Society

CHEN et al. PHYSICAL REVIEW E 70, 016608 (2004)

metamaterial of different thicknesses. The reflection S11 de-

pends on the position of the first boundary and the transmis-
sion S21 depends on the thickness of the slab. In addition,
since the impedance z is also a function of S11 and S21, z
depends on the first boundary and the thickness of the slab as
well. Taking into account the above-mentioned properties,
we propose a method whereby the first boundary and the
thickness of the sample are determined by optimization so
that z1 matches z2 at all frequencies. Figure 1 illustrates the
configuration of the problem for a metamaterial made of two
cells in the propagation direction (x direction). The geometry
of the metamaterial has been taken from [11,12], in which
the dimensions have been slightly modified for ease of dis-
cretization in finite-difference time-domain (FDTD) simula-
tions. With the split-ring resonator (SRR) and rod in the cen-
FIG. 1. Illustration of the effective boundaries of a two-cell
ter of the unit cell, the periodicity along the propagation
metamaterial. The SRRs and rods are periodic along ŷ and ẑ direc-
direction is d0. The first boundary of the effective homoge-
tions with periodicity ay = 4 mm, az = 3 mm. The unit-cell thickness
neous medium is located at x1 below sx1 ù 0d or above
sd0d in the direction of wave incidence is 4 mm. We choose the first
and the last unit-cell boundary as the reference plane for the param- sx1 , 0d the first unit-cell boundary, and the thickness of the
eters x1 and x2, respectively. The value of x1 and x2 are positive effective medium is 2d0 + x2 − x1. The optimization model is
(negative) if the dashed lines are below (above) their reference set up to minimize the mismatch of impedances of different
planes. The thickness of the effective homogeneous medium is numbers of cells of metamaterial:
2d0 + x2 − x1smmd. N
1 f uz1sf i,x̄d − z2sf i,x̄du
min fsx̄d = o
N f i=1 maxhuz1sf i,x̄du,uz2sf i,x̄duj
,
boundaries of the effective slab need to be determined, which
we do here by ensuring a constant impedance for various
slab thicknesses. Second, the S parameters obtained from s. t. :− 0.5d0 ø x1, x2 ø 0.5d0, x̄ = sx1,x2d, s5d
numerical computation or measurements are noisy which can
cause the retrieval method to fail, especially at those frequen- where N f is the total number of sample frequencies and z jsf id
cies where z and n are sensitive to small variations of S11 and represents the impedance of slab jsj = 1 , 2d at frequency f i.
S21. These two problems are examined in detail in the fol- In the ideal case, z1 matches z2 for all frequencies with the
lowing sections. objective function value equal to zero. We use the differential
evolution algorithm [13] to optimize the objective function,
and the optimized solution is x̄opt = s3.85653 10−4d0 , 1.0479
B. Determination of the first boundary and the thickness 3 10−4d0d. The corresponding values of impedance are
of the effective slab
shown in Fig. 2. It can be seen that the retrieved impedances
A homogeneous slab of material can be characterized by for one, two, and three cells of SRR-rod metamaterial match
the fact that its impedance does not depend on its thickness. well for most frequencies, while matching was not as satis-
Our understanding of the physical meaning of the first effec- factory when the method in Ref. [4] was used (which corre-
tive boundary is a plane beyond which the reflected wave sponds to x1 = 0.5d0 in our formulation). We also calculated
behaves like a plane wave for a plane wave incidence. When the impedance z for the case of x̄ = s0 , 0d and found that the
a plane wave is incident on a metamaterial, currents will be results are almost the same as the optimized ones. We have
induced on the metals creating a scattered field. The field corroborated this result with many other metamaterial thick-
produced by the induced currents is not uniform: It is stron- nesses and geometries to eventually conclude empirically
gest around the metal and decay at a certain distance. The that the first effective boundary of a symmetric one-
first effective boundary is located where the reflected wave dimensional (1D) metamaterial [1,4,14] coincides with the
behaves like a plane wave, and it has to be determined. We first unit-cell boundary and the second effective boundary
use z1 and z2 to represent the impedances of two slabs of coincides with the last unit-cell boundary. For two-

FIG. 2. Optimized impedance z for one, two,

and three cells of SRR-rod metamaterial of Fig. 1
in the direction of propagation.

016608-2
ROBUST METHOD TO RETRIEVE THE CONSTITUTIVE… PHYSICAL REVIEW E 70, 016608 (2004)

FIG. 3. Comparison of the retrieved imped-

ance z (real and imaginary parts) for one cell of
metamaterial shown in Fig. 1 by the method pre-
sented in this paper and a traditional method us-
ing only the requirement z8 ù 0.

dimensional (2D) [11,14] and asymmetric 1D metamaterials, the branches of logarithm function as seen in Eq. (4). In
no such rule could be found and the effective boundaries of order to address this problem, it has been suggested to use a
the slab need to be determined from optimization. slab of small thickness and applying the requirement that
esfd and msfd are continuous functions of frequency [4,5].
C. Determination of n and z from S11 and S21 However, no further details on the continuity argument were
provided. In our method, we determine the proper branch by
It is a common method to determine z and n from Eqs. using the mathematical continuity of the parameters, with
(2a) and (2b) with the requirement of Eqs. (3a) and (3b), special attention to possible discontinuities due to reso-
where z and n are determined independently. However this nances. The method is an iterative one: Assuming we have
method may fail in practice when z8 and n9 are close to zero: obtained the value of the refractive index nsf 0d at frequency
A little perturbation of S11 and S21, easily produced in experi- f 0, we obtain nsf 1d at the next frequency sample f 1 by ex-
mental measurements or numerical simulations, may change panding the function einsf 1dk0sf 1dd in a Taylor series:
the sign of z8 and n9, making it unreliable to apply the re-
quirement of Eqs. (3a) and (3b), as discussed in Ref. [6]. In
fact, z and n are related and we should use their relationship
to determine the signs in Eqs. (2a) and (2b). In order to
S 1
D
einsf 1dk0sf 1dd < einsf 0dk0sf 0dd 1 + D + D2 ,
2
s7d

determine the correct sign of z, we distinguish two cases.

where D = insf 1dk0sf 1dd − insf 0dk0sf 0dd and k0sf 0d denotes the
The first is for uz8 u ù dz, where dz is a positive number, for
wave number in free space at frequency f 0.
which we apply Eq. (3a). In the second case, for uz8 u , dz, the
In Eq. (7), the branch index [m in Eq. (4)] of the real part
sign of z is determined so that the corresponding refractive
of nsf 1d is the only unknown. Since the left-hand side of Eq.
index n has a non-negative imaginary part, or equivalently
ueink0d u ø 1, where n is derived from Eqs. (1a) and (1b): (7) is obtained from Eq. (6), Eq. (7) is a binomial function of
the unknown nsf 1d. Out of the two roots, one of them is an
S21 approximation of the true solution. Since we have obtained
eink0d = . s6d n9sf 1d, we choose the correct root among the two by com-
z−1
1 − S11 paring their imaginary parts with n9sf 1d. The root whose
z+1
imaginary part is closest to n9sf 1d is the correct one, and we
Note that once we obtain the value of z, the value of eink0d denote it as n0. Since n0 is a good approximation of nsf 1d, we
is obtained from Eq. (6), so that we avoid the sign ambiguity choose the integer m in Eq. (4) so that n8sf 1d is as close to n80
in Eq. (2b) [it can be proven that only one sign of the imagi- as possible.
nary part in Eq. (2b) makes it equivalent to Eq. (6)]. Figure 3 The branch of n8 at the initial frequency is determined as
shows the retrieved impedance using the method presented follows: From m = nz and e = n / z, we have
in this paper and using only the condition of Eq. (3a). It is
noted that the discontinuities obtained when only applying m 9 = n 8z 9 + n 9z 8 , s8ad
the criterion z8 ù 0 are removed.
1
D. Determination of the branch of n8 e9 = s− n8z9 + n9z8d. s8bd
uzu2
We have presented in the previous sections a method of
solving for z and n9, but n8 remains ambiguous because of The requirements m9 ù 0 and e9 ù 0 lead to

016608-3
CHEN et al. PHYSICAL REVIEW E 70, 016608 (2004)

FIG. 5. Range of z (real and imaginary parts) for tolerance dr

= 0.015 and dt = 0.0 in Eqs. (11a) and (11b). The impedance is for a
three-cell metamaterial shown in Fig. 1.
FIG. 4. Retrieved refractive index n (real and imaginary parts)
for one, two, and three cells of the metamaterial structure shown in
values of n for different cell numbers are identical above the
Fig. 1.
resonant region. Below the resonant band, however, the re-
trieved n for one and two cells match well, but the result for
un8z9u ø n9z8 . s9d three cells differs significantly from the previous two. This
discrepancy is due to the small magnitude of S21 in this fre-
In particular, when n9z8 is close to zero but z9 is not, n8 quency band, as we shall discuss in the next section.
should be close to zero. At the initial frequency, we solve for
the branch integer m satisfying Eq. (9). If there is only one E. Sensitivity analysis
solution, it is the correct branch. In case of multiple solu-
A close examination of the retrieved z and n for one, two,
tions, for each of the candidate branch index m, we deter-
and three cells of metamaterial presented so far shows that
mine the value of n8 at all subsequent frequencies using the
the three results do not always match. There are two cases of
above-mentioned iterative method. Because the requirement
discrepancies. The first is that the retrieved refractive index n
of Eq. (9) applies to n8 at all frequencies, we use it to check
for three cells of metamaterial does not match the value for
the validity of n8 at all frequencies produced by the candi-
one and two cells at low frequencies (5 GHz– 11 GHz in
date initial branch. Note in the special case when n9z8 is
Fig. 4). The second is that the impedance z appears to spike
close to zero but z9 is not, the checking process can easily be
at some frequencies (around 12 GHz, 17 GHz, and 19.5 GHz
carried out. Therefore, the initial branch that satisfies Eq. (9)
in Fig. 2). We shall show here that these discrepancies are
at both the initial frequency and the subsequent frequencies
due to the sensitivity of z and n to the accuracy of S11 and
is the correct one.
S21.
For the SRR-rod structure, we found that there is a fre-
The first case appears when uS21u is close to zero. In the
quency region at which there is no branch index m satisfying
region below the resonance band, the transmission is usually
Eq. (9). We call this frequency region the resonance band.
small, especially for thicker metamaterials. From Eq. (2b),
The properties of the resonance band are still disputed by
we see that S21 appears in the denominator, so that the values
researchers. Some papers [15–17] mention the existence of
of n are very sensitive to small perturbations of S21. Yet, a
multiple modes in this region since the real part of n is large,
small transmission has little influence on the retrieval of z,
yielding a wavelength comparable to or smaller than the unit
which can be seen by computing:
size of the metamaterial thereby rendering the retrieval of the
effective parameters of the metamaterials impossible. Other ] z2 8S21S11
papers [5,18] state that retrieval is possible and the retrieved = 2 2, s10d
] S21 fs1 − S11d2 − S21 g
permittivity e has a negative imaginary part in the resonance
band. In this paper, we do not address this issue and for this from which it is clear that a small uS21u makes ]z2 ]S21 /
reason the retrieved results presented here are interrupted in small (approximately zero). In addition, we can see from Eq.
frequency by the resonance region (see, for example, Fig. 4). (1b) that if n9 is not small, S21 will decrease exponentially
In this case, since nsfd is not continuous through all frequen- with d, i.e., with an increasing number of cells in the propa-
cies, we have to determine the initial branches for two fre- gation direction. Therefore, the smaller S21, the higher the
quency regions: Below and above the resonance band. Note computation and measurement relative errors, which leads to
that below the resonant band, the retrieved branch index is less accurate retrieval results.
2
zero, which confirms the validity of the traditional method The second case appears when S21 is close to unity while
used for low-frequency retrieval. The retrieved refractive in- S11 is close to zero. Similar to the first case, the denominator
dexes n for one, two, and three cells in the propagation di- in the expression of z [see Eq. (2a)] approaches zero, thus
rection are shown in Fig. 4, where the resonance band is seen making it difficult to retrieve z. However, in this case, the
to extend between 11 GHz and 12 GHz. We observe that the value of n is stable. In this situation, instead of solving for n

016608-4
ROBUST METHOD TO RETRIEVE THE CONSTITUTIVE… PHYSICAL REVIEW E 70, 016608 (2004)

FIG. 6. S11 and S21 (real and imaginary parts)

for three cells: Comparison between FDTD re-
sults (dot line with *) and calculated S parameters
based on the retrieved e and m (solid line with h)
for a one-cell metamaterial shown in Fig. 1.

and z which exactly satisfy Eqs. (1a) and (1b), we solve for that although the results satisfy the condition e9 ù 0 and m9
the following inequalities: ù 0, the positive energy requirement ]sevd / ]v . 0 [19,20] is

U S11 −
R01s1 − ei2nk0dd
2 i2nk0d
1 − R01 e
U ø dr , s11ad
violated in the frequency band 12 GHz– 12.2 GHz. As a re-
sult, the resonance band is extended to 11 GHz– 12.2 GHz,
as shown by the vertical dashed lines in Fig. 7(a). The value
of e and m are both negative in the frequency range
U S21 −
2
s1 − R01 deink0d
2 i2nk0d
1 − R01 e
U ø dt , s11bd
12.2 GHz– 12.8 GHz, showing that in this band, the
metamaterial exhibits a LH behavior. We also retrieved the
effective parameters of four and five cells of metamaterial
where dr and dt are small positive numbers. Figure 5 shows shown in Fig. 1, and the retrieval results are close to those
the range of z satisfying Eqs. (11a) and (11b) for dr = 0.015 for one, two, and three cells.
and dt = 0.0. At each frequency, all z having a real and imagi-
In addition, we also applied our method to retrieve the
nary parts between the bounds shown in Fig. 5 satisfy Eqs.
effective parameters of the structure taken from [14,21], as
(11a) and (11b). It can be seen that the magnitude of the
shown in the inset of Fig. 8(a). For a 1D structure, by match-
spikes is within the tolerance error, which implies that they
are due to the noise in the S11 and S21 data. ing the impedance z for one and two cells of the metamate-
Finally, note that although the retrieved n and z for mul- rial using the previously described method, we obtain x̄opt
tiple cells may be different from that for one cell at some = s2.20533 10−3d0 , 1.13563 10−3d0d, where d0 is the length
specific frequencies, the calculated S11 and S21 for multiple of unit cell. Again, we find that the two boundaries of the
cells using the retrieved e and m from one cell data match effective homogeneous medium are close to the outer unit-
well with the S11 and S21 data computed for multiple cells cell boundaries of the 1D metamaterial. Figure 8 shows the
directly from numerical simulation, as is illustrated in Fig. 6. retrieved z, n, e, and m for one cell of this metamaterial. It
can be seen that the frequency range of 13.8 GHz– 14.5 GHz
is a LH band, which agrees with the conclusion in Ref. [14].
F. Results
It should be noted, however, that for a 2D version of this
The retrieved permittivity e and permeability m of a one metamaterial, the effective boundaries should be obtained
cell of SRR-rod structure of Fig. 1 are shown in Fig. 7. Note from the optimization process, as they do not necessarily

FIG. 7. Retrieved e and m (real and imaginary

parts) for a one-cell metamaterial shown in Fig.
1. The vertical dashed lines denote the limits of
the resonance band.

016608-5
CHEN et al. PHYSICAL REVIEW E 70, 016608 (2004)

FIG. 8. Retrieved z, n, e, and m (real and

imaginary parts) for a one-cell metamaterial
structure taken from Refs. [14,21] and shown in
the inset (a). The vertical dashed lines denote the
limits of the resonance band.

match the unit-cell boundaries of the metamaterial. Indeed, the sign of z by the value of its corresponding n so that n9
in this specific case, we obtain x̄opt = s0.33110d0 , 0.30342d0d. ù 0.
(3) There is a resonance band characterized by the fact
that the requirement m9 ù 0 and e9 ù 0 cannot be satisfied at
III. CONCLUSION those frequencies. On each side of the resonance, the branch
We have proposed an improved method to retrieve the of n8 can be obtained by a Taylor expansion approach con-
effective parameters (index of refraction, impedance, permit- sidering the fact that the refractive index n is a continuous
tivity, and permeability) of metamaterials from transmission function of frequency. Since the refractive index n at the
and reflection data. The successful retrieval results for vari- initial frequency determines the values of n8 at the subse-
ous metamaterial structures show the effectiveness of the quent frequencies, we determine the branch of the real part
method. Our main conclusions are as follows: of n at the initial frequency by requiring that m9 and e9 are
(1) The first boundary and the thickness of the effective non-negative across all the frequency band.
media can be determined by matching z through all sample (4) Due to the noise contained in the S parameters, the
frequencies for different lengths of the slabs in the propaga- retrieved n and z at some specific frequencies are not reli-
tion direction. For symmetric 1D metamaterials, we have able, especially for thicker metamaterials at lower frequen-
drawn the empirical conclusion that the first boundary coin- cies. In spite of this, the fact that S11 and S21 for multiple
cides with the first boundary of the unit cell facing the inci- cells of metamaterial calculated from the retrieved e and m
dent wave, and the thickness of the effective medium is ap- for a unit-cell metamaterial match the S11 and S21 computed
proximately equal to the number of unit cells multiplied by directly from numerical simulation confirms that the
the length of a unit cell. For 2D and asymmetric 1D metama- metamaterials can be treated as an effective homogeneous
terials, the effective boundaries have to be determined by material.
optimization.
ACKNOWLEDGMENTS
(2) The requirement z8 ù 0 cannot be used directly for
practical retrievals when z8 is close to zero because the nu- This work was supported by DARPA (Contract No.
merical or measurement errors may flip the sign of z8, mak- N00014-03-1-0716) and ONR (Contact No. N00014-01-1-
ing the result unreliable. In this case, we have to determine 0713).

[1] D. R. Smith, W. J. Padilla, D. C. Vier, S. C. Nemat-Nasser, and Kong, Chin. Sci. Bull. 48, 1325 (2003).
S. Schultz, Phys. Rev. Lett. 84, 4184 (2000). [4] D. R. Smith, S. Schultz, P. Markoš, and C. M. Soukoulis, Phys.
[2] R. A. Shelby, D. R. Smith, and S. Schultz, Science 292, 77 Rev. B 65, 195104 (2002).
(2001). [5] P. Markoš and C. M. Soukoulis, Opt. Express 11, 649 (2003).
[3] L. Ran, X. Zhang, K. Chen, T. M. Grzegorczyk, and J. A. [6] R. W. Ziolkowshi, IEEE Trans. Antennas Propag. 51, 1516

016608-6
ROBUST METHOD TO RETRIEVE THE CONSTITUTIVE… PHYSICAL REVIEW E 70, 016608 (2004)

(2003). magnetics Research Symposium, Honolulu, Hawaii, 2003

[7] A. M. Nicolson and G. F. Ross, IEEE Trans. Instrum. Meas. (EMW, Cambridge, MA, 2003), p. 286.
19, 377 (1970). [15] C. R. Simovski, P. A. Belov, and S. He, IEEE Trans. Antennas
[8] J. Baker-Jarvis, E. J. Vanzura, and W. Kissick, IEEE Trans. Propag. 51, 2582 (2003).
Microwave Theory Tech. 38, 1096 (1990). [16] J. E. Sipe and J. V. Lranendonk, Phys. Rev. A 9, 1806 (1974).
[9] J. A. Kong, Electromagnetic Wave Theory (EMW, Cambridge,
[17] P. A. Belov, S. A. Tretyakov, and A. J. Viitanen, Phys. Rev. E
MA, 2000).
66, 016608 (2002).
[10] J. A. Kong, Prog. Electromagn. Res. 35, 1 (2002).
[18] T. Koschny, P. Markoš, D. R. Smith, and C. M. Soukoulis,
[11] D. A. Shelby, D. R. Smith, S. C. Nemat-Nasser, and S.
Schultz, Appl. Phys. Lett. 78, 489 (2001). Phys. Rev. E 68, 065602 (2003).
[12] C. D. Moss, T. M. Grzegorczyk, Y. Zhang, and J. A. Kong, [19] L. D. Landau, E. M. Lifshitz, and L. P. Pitaevskii, Electrody-
Prog. Electromagn. Res. 35, 315 (2002). namics of Continuous Media, 2nd ed. (Pergamon, Oxford,
[13] R. Storn and K. Price, J. Global Optim. 11, 341 (1997). 1984).
[14] T. M. Grzegorczyk, C. D. Moss, J. Lu, and J. A. Kong, New [20] V. Veselago, Sov. Phys. Usp. 10, 509 (1968).
Ring Resonator for the Design of Left-Handed Materials at [21] C. D. Moss, Ph.D. thesis, Massachusetts Institute of Technol-
Microwave Frequencies, Proceedings of Progress in Electro- ogy, 2004.

016608-7