Professional Documents
Culture Documents
4 (1999) 345-354
The placement accuracy and resolution of direct-write patterning tools, in particular the atomic force
microscope (AFM), is considered for application to fabricating multi-passband integrated optical filters.
Because of its simpler fabrication a grating structure is proposed that consists of identical stripes that are
non-periodically spaced. The recently developed pseudorandom encoding method from the field of
computer generated holography is modified to effectively assign analog reflectances at each point along the
grating by selective withdrawal and offsetting of the stripes from a periodic spacing. An example filter
designed by this method has two 1.5 nm bandwidth passbands and - 23 dB of rejection for lightly coupled
stripes. As with single band filters, the passbands broaden as the coupling increases. A calculation of the
coupling coeflicient of stripes on a fundamental mode, slab waveguide indicate that stripes on the order of
100 nm in depth and width support low insertion loss, multipassband filtering applications at visible
wavelengths. Lines of these dimensions patterned with an AFM on (110) silicon indicates the feasibility of
fabricating these fllters. These conclusions are specific to current AFM's that are limited to writing fields of
100 1lrn. Increased rejection and decreased passband widths will result from incorporating precise field-
stitching into future AFM'S.
Key words: atomic force microscopy, nanolithography, photonic crystals, optical information processing,
waveguide optics, nanometer optics
345
346 OPTICAL REVIEW Vol. 6, No. 4 (1999) R. W. CoHN et al
q -~
~1, I~2, ~3 -V
lllIIll ~ cally illuminated computer generated hologram (CGH).
~2 ~II . ~~ This similarity can be used to directly encode a desired
complex-valued temporal function/impulse response
into a grating. The Fourier transform of this function is
the desired temporal frequency response. These similari-
I~ ~II I I I I I ~
~l'2~3
~1' I~2' ~3
ties are used in Sect. 2 to adapt CGH methods, specifical-
where ai is the (real-valued) reflectance, ti is the time be applied with consequent improvements in the accu-
delay produced by the i 'th stripe of the N stripe grating racy of the encoding to approximate the desired
and v is the temporal frequency of the light. The stripe spectrum.7) Even finer placement resolutions lead to
reflectance can be interpreted as being complex-valued even more accurate encoding methods. The writing of
by rewriting Eq. (1) using the definitions v = vo + 6 v and thinner lines is also desirable in that the SPM writing
ti :~ ito + ~ ti Where 6 ti are the offsets of the stripes ito from speed can be increased. However, thin lines must be etch-
a perfectly periodic grating, and 6v is the frequency ed more deeply to produce reflection strengths equal to
of~set from the center frequency vo. Multiplying out those of thicker lines. The relationship between etch
these terms in the argument of the exponential in Eq. (1) depth and reflectance is considered further in Sect. 3.
yields a product of four complex exponentials. One term For the frequency responses developed here, positive
is exp (j27T6v6ti)-1 for frequency ranges of concern and negative real valued modulation is sufficient and pro-
6 v << vo• This condition is usually easy to meet in current
wavelength division multiplexing systems where laser
vides the least strict requirements on line width and
placement accuracy.
tuning ranges and system bandwidths are usually less Based on the above considerations on the dual pass-
than 100 nm. Ignoring this term leads to Eq. (1) being ap- band filter, each stripe will be limited by the fabrication
proximated as process to be identical. Therefore ai the magnitude of the
reflectance of each stripe is identical. However, there is
N
F(v) * ~ ai exp (j~i) exp (j27ritov), (2) the option to not place a stripe at certain locations on the
i=1 sampling grid. Therefore, either a unity amplitude l ''i "
where ~i ~ 21Tv06 ti is the nominal phase shift produced by or a zero amplitude "O" can be realized at each sample
offsetting the stripe positions from those of a periodic point of the grid. In general, any phase ~pi=27T6ti/ to Can
grating. Equation (2) is the Fourier transform of a period- be realized by offsetting/ delaying the stripes from the ~o
ic grating in which the stripes have fully complex valued sampling points on the grid. We however limit the offsets
reflectances. It is mathematically identic,al to the far-field to O or ~0/2. Therefore, the phases of the reflectances
diffraction pattern of the original Lohmann CGH,3) in can be either O (for 6ti=0) or lr (for 6ti= to /2). The corn-
which case v would represent the spatial coordinate plex reflectances ai~ai exp (j~Pi) in Eq. (2) that can be
across the diffraction plane. For the wide range of optical realized are "I", when a stripe is written at the iAO Sam-
frequencies over which Eq. (2) is valid, nearly arbitrary pling point of the grid, " - 1" when a stripe is written at
frequency responses can be designed based on the values the (i+(1/2))~o Sampling point on the grid, and "O"
selected for the number of stripes, and the magnitude when a stripe is written at neither of the two sampling
and phase of the stripe reflectances. points. [Stripes also could be written simultaneously at
2.2 Selection of Stripe Reflectances for Dual Passband both the i~O and (i + (1 / 2))~o locations, but we do not con-
Fil ters sider this possibility here.]
Following the CGH design philosophy, the first step in 2.3 Pseudoraudom Encoding.' A CGH Algorithm for En-
a design is to identify the available modulation values coding Fully Complex Values
that can be implemented. Then an encoding scheme is de- One recently developed class of CGH methods that can
veloped to represent all the modulation values needed to be adapted to the problem of encoding continuous valued
design a spectrum. In this section we apply this approach reflectances with only the three amplitudes - I , O and 1
to the design of a dual-passband grating filter. As dis- is referred to as pseudorandom encoding.6,8) A specific al-
cussed in Sect. I , it is desirable to achieve the filter func- gorithm already developed for the case of bi-magnitude
tion with the simplest fabrication processes possible. For SLM's will be used.9) Given two available values of mag-
this reason we have specifled a lithography in which each nitude O and 1, bi-magnitude pseudorandom encoding
stripe is identical in geometry and stripes are written on can represent/ encode any desired magnitude aci between
a periodic grid corresponding to a sample spacing of ~o / O and 1. Used together with the additional sign reversal
2 optical path difference where Ao is the wavelength at available by offsetting a stripe a half wavelength, all
center frequency vo• This prescription allows reflectance
desired amplitudes between - I and I can be encoded.
The basic algorithm and the results of a theoretical per-
values of 1, O and - I to be implemented. From these
values an encoding method is developed that effectively formance analysis are given here. References 6-9 may be
realizes a continuum of reflectances from - I to I . This consulted for additional background and theory on pseu-
particular algorithm is by no means the only possible dorandom CGH algorithms.
CGH algorithm that could be employed, but its numeri- In pseudorandom encoding the magnitude ai for the
cally simple implementation makes it especially useful i'th stripe is selected using a random number generator.
for purposes of illustration. Specifically, the random number generator is configured
In passing we note that the particular encoding to produce random numbers from the probability density
method can be generalized from real, to complex-valued function (pdf)
representations if the pitch of the sampling grid on which
p(ai) = a,i ' 6 (ai - 1) + (1 - a,i) ' 6 (ai), (3)
the reflectors are placed is reduced from ~0/2 to ~0/3.
Then the CGH method of ternary-valued encoding can
where 6( • ) is the Dirac delta function,
348 OPTICAL REVIEW Vol. 6, No. 4 (1999) R. W. CoHN et al.
probability of selecting the magnitude to be ai=1 and
1S e "-e '- ll
,
truncated at a maximum magnitude of 0.077. This lated to that for a sinusoidal grating.15) These relation-
sacrifices the sidelobe level somewhat, but for a fixed ships are used to 'estimate the appropriate stripe geo-
field of view grating the passband is narrower than if the metry as a function of the magnitude of the grating reflec-
Dolph weights were allowed to decay to near zero. The tance. The remainder of this section summarizes these
modified Dolph apodization produces a -35 dB sidelobe analysis procedures and uses them to evaluate the perfor-
level and a - 3 dB bandwidth (i.e. full width at half maxi- mance of the dual passband filter design.
mum power) of 1.5 nm. This can be compared with the 3.4 Discrete Layer Analysis
frequency response for 512 periodically spaced, unit The wavelength dependent reflectance of an etalon
strength reflectors. The periodic filter would have a side- stack can be analyzed by cascading the reflection and
10be level of - 13 dB and a passband bandwidth of ap- transmission properties of the individual layers. For this
proximately ~o/N= 1.25 nm, however the actual - 3 dB Discrete Layer (DL) analysis, each layer is modeled using
bandwidth found numerically is 1.1 nm. a 2 x 2 characteristic scattering matrix as described in
The encoded function is represented by the dots of Ref. 14. Each section is designed to introduce a quarter
values 1, O and - I in Fig. 2(a). As prescribed by pseu- wavelength optical delay In = ~o / 4 where n is the refrac-
dorandom encoding, desired values close to I are usually, tive index of a particular layer and I is its physical length.
but not always, represented by ai=1. Likewise values The characteristic matrices for all the layers are multi-
close to O and - I are most frequently represented by plied in sequence and the resulting matrix is evaluated to
those values. The Fourier transform of the encoded give the frequency dependent complex reflectance.
values ai Produces the spectrum (thick line) in Fig. 2(b). For the proposed grating filter the sections that
The bandwidth of each passband is 1.5nm and the represent values of either 1, - I or O are modeled as fol-
highest sidelobe level is -23 dB. The sidelobe level 10ws. Each section consists of two quarter wavelength
reflects the noise introduced by the noise term (i.e. the layers. A O corresponds to two layers of refractive index
summation) in Eq. (6). The average noise level calculated n = I . A I corresponds to a layer with n > I followed by a
from this term is - 24 dB below the peak of the pass- layer n=1. A - I corresponds to a layer with n=1 fol-
band. Thus the sidelobes from the apodization are low 10wed by a layer with le > 1, which is the reverse of the or-
enough that the noise from the encoding procedure is the dering used for the value I . In this way the higher index
principal contributor to the sidelobe level. The sidelobe layers represent stripes located at the desired positions
level can be improved by using more samples in the fil- in the proposed grating filter.
ter. This could be achieved by designing for even shorter This method is adequately general for analyzing
wavelengths or by increasing the field over which the aperiodic structures. However, in order to relate the DL
patterning tool can write. Increasing the writing field analysis to COM analysis it is useful to consider the spe-
would also allow the passbands to be narrowed further. cial case for a periodic structure. Kogelik's analysis of
3.3 Evaluation ofthe GratingFilter in a Slab Waveguide the periodic, sinusoidally perturbed waveguide of length
Configura tion L and coupling coefEicient K gives essentially identical
The frequency response of a periodic corrugated results as the DL analysis of N pairs of quarter-
waveguide has been analyzed using coupling of modes wavelength layers that differ in refractive index by A n if
(COM) analysis by Kogelnik.12) Closed form COM solu-
tions for non-periodic linear and quadratically chirped A n = lCAO / 2 . (7)
gratings were also developed by Kogelnik.13) However, a We have also checked this correspondence through nu-
method of analyzing general nonperiodic structures is merical simulation. We specifically evaluated a 512
needed. The analysis of arbitrary nonperiodic gratings period structure as a function of A n for both analyses.
should be analogous to the Born and Wolf analysis of a The DL geometry consists of 1024 Iayers of alternating
stack of nonidentical etalons.14) Kogelnik has already refractive index n+An and n. The correspondence be-
adapted their method to propagation in a layered or tween the two models is compared in (Fig. 3) in terms of
stratified waveguide.12) Instead of propagation being the - 3 dB bandwidth as a function of filter reflectance at
parallel to the layers we consider the case of plane wave center wavelength ~o = 640 nm. The results are identical
propagation normal to the layers. We only discuss the for the two analyses except for small errors that are due
TE case, in which case the electric field is parallel to the to the small number of sample points used in calculating
stri pes. the spectra. The bandwidth broadening is a direct result
While we use the analysis to evaluate the spectrum of of strong multiple reflections that saturate the frequency
the dual passband filter, it does not by itself provide infor- response around the center frequency. In each case iden-
mation on the dependence of grating reflectance on the tical centerband reflectances are found if the stacked
stripe width and depth. However, it is possible to relate etalon and COM analyses use values of A n and lc that
the coupling coefiicient K used in COM analysis of are related by Eq. (7). These correspondences between
sinusoidally perturbed guides to the refractive index the two analyses indicate that the DL analysis for the aperi-
difference Ale between the two types of layers used in a odic filters will reasonably model our grating structures of
periodic etalon stack. Furthermore lc for a square wave interest.
grating on top of a slab waveguide has been directly re-
350 OPTICAL REVIEW Vol. 6, No. 4 (1999) R. W. CoHN et al.
5 O (a)
,~,
~
OO
1) -1 o
E4 ~8 -20
,~. CC~
1:
~~
~3
.-~
~o
o -30
q=
~) o
O(
-40
~2
1
O (b)
o 0.5 1 .o ~
OO
~ -1 O
Insertion loss ~oo
CU
-20
Fig. 3. Bandwidth at - 3 dB of peak intensity for a single
~~
passband (512 period) grating filter and the dual passband grating o
o -30
filter as a function of filter insertion loss. The correspondence be- ~:
tween the COM (coupling of modes) and DL (discrete layer) o
O(
analyses for the periodic structure indicates the validity of the DL -40
analysis for the analysis of the non-periodic dual passband filter.
, dual band filter (DL); , periodic grating (DL); ' ' ' ' ' ' ,
periodic grating (COM).
o ( c)
~
ao
I~ -1 O
3.5 Grating Depth Analysis
Yariv describes a method of calculating the coupling
~8
~ -20
coeificient for TE waves propagating in a slab waveguide
~o
that is perturbed by a square wave corrugation.15) Follow-
ing this procedure we derived from the unperturbed field
o
~:
-30
a)
distribution a closed form expression for the coupling O(
-4 O
coefiicient. For simplicity of presentation, we only report
our result for symmetric slab waveguides. We find that 633 640 647
the coupling coefElcient can be expressed as wavelength (nm)
y~ y, n. /~ n~ - n~ Fig. 4. Reflectance spectra for the dual passband filter design.
K ~ y~ + y~ 2 + h y, cos2 (h yg 1 2)
The DL analyses (thick lines) are shown for filter insertion loss (i.e.
peak intensity reflectance at centerband) of (a) 0.068, (b) 0.532 and
x d+Sin (hyg)-sin (hyg-2d) (8) (c) 0.917. The spectrum for the dual passband design from Fig. 2 is
2 yg '
where h is the thickness of the guiding layer, d is the cor-
also replotted (thin line) for comparison. The reflectance spectra are
normalized so as to bring their sidelobe structure into
correspondence with the design spectrum.
rugation depth, n, is the refractive index of the cladding,
ng is the refractive index of the guiding region, and
n. =fi~ / (27r) is the effective index for the wav'e having
where the term sin (7cA ) is the ratio of the Fourier
propagation constant fi. The eigenvalues that describe
coefficient of a grating of duty cycle A to the coeificient
the field distribution of the unperturbed field are yg for
for A = 1/2.
the guiding layer and y* for the cladding regions. Yariv
3. 6 DL Analysis of the Dual Passband Filter
has presented an approximate expression for K (Eq.
Figure 4 shows representative spectra resulting from
13.4-17 of Ref. 15). This approximation is valid only for
the DL analysis for values of An of 0.0015, 0.0065, and
h(ng ~ n*)/A >> I . This condition is valid only if the guide
can support multiple modes.12) However, Eq. (8) is valid 0.14 for a-c respectively in Fig. 4. For a centerband
reflectance (i.e. insertion loss) of 0.068 (-11,7 dB) the
even for single mode guides.
DL analysis in Fig. 4(a) is nearly identical to the Fourier
Equation (8) specifically describes K for the fundamen-
transform of the encoded function (from Fig. 2). For an
tal waveguide mode and the fundamental Fourier series
insertion loss of 0.532 (-2.7 dB) the mainlobes in Fig.
harmonic of a 500/0 duty cycle grating. The expression of
4(b) are slightly saturated and the sidelobe are nearly
coupling coefiicient is generalized for a rectangular gr•at-
ing of any duty cycle according to
identical to the designed spectrum. For insertion loss of
0.917 ( - 0.38) dB, Fig. 4(c) shows a strong intensity satu-
K~ = K Sin (7cA), (9) ration and frequency broadening of the passbands.
OPTICAL REVIEW Vol. 6, No. 4 (1999) R. W. COHN et al. 351
/
//
Equation (8), the relationship between stripe depth d
and coupling coefiicient for a periodic grating of 500/0 ~~~
/ /
//
/
/
duty cycle (where duty cycle is the ratio of stripe width to
1~ ,
,, / /
grating period) is evaluated in Fig. 5 for four values of 11~
, ; , • ; --, • ; , 3.0pm.
0.5 ,Im 1 O pm 2 O pm
o 4 -4 O A (a)
?\J~~i~ ~
'~~o
~y2~\;~ -~~ ' __
*:~
70/1/~ ?\vl~J3-_
~6~(~
.2
'~~ ~b um
Fig. 6. A non-periodic silicon dioxide grating that has been direct
written on a silicon surface using an atomic force microscope. The
structure (which also was profiled by an AFM) is annotated with A (b)
wavelengths and stripe reflectances in the same manner as the
proposed filter of Fig. 1(c). The closest spacing of adjacent lines in
this AFM profile is 240 nm for a I followed by a - 1.
silicon because the (110) plane is known to etch as much grating is placed between the two sections. The guided
as 100X faster than the (111) plane for specific etch- wave velocity can be compensated by depth etching, to
ants.21) Thus by writing oxide lines on the surface of (1 10) properly phase the two sections together. A single etch
silicon that are parallel to the wafer flat we anticipate depth (i.e. identical etch conditions) can be used for phas-
that near vertical sidewalls corresponding to the (111) ing any two sections if the length of the guided region
plane will be formed. The AFM written silicon dioxide that is exposed to etchant is varied.
lines are oriented in this way. (2) While commercial AFM's are quoted with around
A single wafer containing lines of various widths is wet 2-3 pm of stage positioning error, it would be possible to
etched at 90'C in a solution containing 83 ml of 250/0Wt. use the AFM head itself to determine exactly where the
tetramethyl-ammonium hydroxide (TMAH) and 17 ml of stage moves to. The previously written pattern (or other
isopropyl alcohol for 15 s. AFM profiles of the etched prewritten fiducials) can be identified through AFM scan-
structures reveals several interesting results. (1) The ning and the measured offset (and possibly tilt) errors can
thicker oxide lines etch to a depth of - 100 nm while the be used to offset (and possibly rotate) the patterning in-
thinner lines etch to depths of - 50 nm. (2) The typical structions. This is a quite reasonable approach if one has
sidewall slope is 30' from vertical. (3) Thicker lines have the ability and adequate time to modify the AFM control
a flat surface between the sidewalls while the thinner sof tware.
lines form a continuously curved hillock. These results There are various applications and configurations of
suggest that the thin edges of the oxide are not protect- multipassband grating filters. A single grating cus-
ing the line during the entire etch. For the very thin lines tomized to the demands of a subscriber can be used to
the oxide is being completely undercut and removed. tap off a number of non-sequential frequency channels
Figure 7 shows one of the lines before (Fig. 7(a)) and af- from a wavelength multiplexed fiber channel. Slanted
ter etch (Fig. 7(b)). The width across the base of the line gratings (Fig. 1(d)) can be used to form wavelength selec-
is - 250 nm both before and after etch. After etch the tive crosspoints. Arranging these crosspoints an x-y
width of the plateau of the line has narrowed to - 100 fashion on a single substrate can be used to realize vari-
nm. Additional studies are needed to find ways to better ous other network topologies. Programmable filters can
protect silicon from etchants, such as modifying oxide also be envisioned in which arrays of stripes can be in-
shape or optimizing the properties of the etchant. Alter- dividually placed in or removed from the beam path. Elec-
natively, different material processes altogether may trostatic attraction could be used to displace the stripes
lead to lines of the desired width and depth. in a manner similar to micromechanical mirror array tech-
nology of Texas Instruments that is currently used in
5. Discussion video projectors and printing engines.24) CGH encoding
We have considered the possibility of generalizing the algorithms, such as those described provide the flexibili-
frequency response of grating reflection filter by em- ty and adaptivity to design and compute desired filter
ploying signal encoding techniques from computer gener- functions instead of storing large tables of anticipated
ated holography. A specific case of a grating on a slab stripe settings.
waveguide has been considered. Fabrication constraints In summary this paper has considered the possibility
set by the limited field of view of current AFM pattern- of nanofabricating multipassband grating filters using
ing systems have been considered. One consideration is current AFM's as direct write patterning tools. We have
that encoded functions generally have lower intensity fre- demonstrated that useful designs are possible even with
quency responses than do periodic structures of the same the limited writing field of current AFM's and that there
length. In order to compensate for these differences the are approaches that can permit precise field stitching.
stripes must be correspondingly deeper than for periodic Writing times are currently quite slow but not critical for
structures in order to obtain equivalent insertion losses. using AFM's to develop single experimental or proto-
Likewise, the use of phase reversed strip placement for type devices. Current writing speed can be increased by
the dual passband filter requires narrower stripes which using a controlled atmosphere for the silicon oxidation
also requires deeper stripes. Even greater depths (or thin- process or using different material systems which are
ner guiding layers) will be needed to extend this ap- known to be faster.25)
proach from bi-phase to polyphase encoding algorithms.
Already for some of the examples considered here, stripe Acknowledgments
aspect ratios (depth over width) in excess I : I have been We thank Frank Luecke of FrankDesign, LLC for discussions
found . and ideas on field stitching methods. This study was supported by
These aspect ratios can be reduced for designs having ONR grant NOO014-96-1-1296, NASA cooperative agreement
NCC5-222, and NSF grant ECS-9724371.
more stripes. Current AFM's (unlike ebeam pattern
generators) do not include high precision stages neces- Ref erences
sary to stitch together multiple fields. At least two reason-
1) H. Asai and Y. Wada: Proc. IEEE 85 (1997) 505.
able extensions are possible: 2) E. S. Snow, P. M. Campbell and F. K. Perkins: Proc. IEEE 85
(1) Two gratings could be placed in close proximity (1997) 601.
to each other. A short region that is unperturbed by a 3) B. R. Brown and A. W. Lohmann: Appl. Opt. 5 (1966) 967.
354 OPTICAL REVIEW Vol. 6, No. 4 (1999) R. W. CoHN et al.
4) W. J. Dallas: The Computer in Optical Research, ed. B. R. 16) J. A. Dagata, J. Schneir, H. H. Harry, C. J. Eaves, M. T.
Frieden (Springer, 1980) Chap. 6, p. 332. Postek and J. Bennet: Appl. Phys. Lett. 56 (1990) 2001.
5) R. W. Cohn and L. G. Hassebrook: Optical Information Proc- 17) E. S. Snow, P. M. Campbell and P. J. McMarr: Appl. Phys.
essing, eds. F. T. S. Yu and S. Jutamulia (Cambridge U. Press, Lett. 63 (1993) 749.
1998) Chap. 15, p. 396. 18) E. S. Snow, W. H. Juan, S. W. Pang and P. M. Campbell:
6) R. W. Cohn: J. Opt. Soc. Am. A 15 (1998) 868. Appl. Phys. Lett. 66 (1995) 1729.
7) R. W. Cohn and M. Duelli: J. Opt. Soc. Am. A 16 (1999) 71. 19) H. Sugimura and N. Nakagiri: Nanotechnology 6 (1995) 29.
8) R. W. Cohn and M. Liang: Appl. Opt. 33 (1994) 4406. 20) H. Sugimura and N. Nakagiri: Nanotechnology 8 (1997) A15.
9) R. W. Cohn and W. Liu: 1996 OSA Technical Digest Series, 5 21) K. E. Petersen: Proc. IEEE 70 (1982) 42.
(Boston, MA, 1996) pp. 237-240. 22) G. T. A. Kovacs, N. I. Maluf and K. E. Petersen: Proc. IEEE
10) C. S. Hartmann: 1973 IEEE Proc. Ultrasonics 1973, 423. 86 (1998) 1536.
11) F. J. Harris: Proc. IEEE 66 (1978) 51. 23) Y. Xia and G. M. Whitesides: Annu. Rev. Mater. Sci. 28
12) H. Kogelnik: Guided-Wave Optoelectronics, 2nd ed, ed. T. (1998) 153.
Tamir (Springer, 1990) p. 7. 24) P. F. Van Kessel, L. J. Hornbeck, R. E. Meier and M. R.
13) H. Kogelnik: Bell Syst. Tech. J. 55 (1976) 109. Douglass: Proc. IEEE 86 (1998) 1687.
14) M. Born and E. Wolf: Prileciples of Optics (Cambridge U. 25) K. Wilder, C. F. Quate, B. Singh and D.F. Kyser: Electron. Ion
Press, 1980) 6th ed. aud Photon Beam Techleology and Nanofabrication, paper NT7
15) A. Yariv: Optical Electrordcs (Oxtord U. Press, 1997) 5th ed, (26-29 May 1998, Chicago, IL).
Chap. 13, p. 491.