Wireless Infrared Communications JOSEPH M. KAHN, JOHN R. BARRY MEMBER, IEEE, AND

Wireless Infrared Communications JOSEPH M. KAHN, JOHN R. BARRY MEMBER, IEEE, AND
Wireless Infrared Communications
The use of infrared radiation as a medium for high-speed,
short-range wireless digital communication is discussed. Currently
available infrared links and local-area networks are described.
Advantages and drawbacks of the infrared medium are compared
to those of radio and microwave media. Physical characteristics of
infrared channels using intensity modulation with direct detection
(IM/DD) are presented, including path losses and multipath responses. Natural and artificial ambient infrared noise sources are
characterized. Strategies for designs of transmitters and receivers
that maximize link signal-to-noise ratio (SNR) are described. Several modulation formats are discussed in detail, including on–off
keying (OOK), pulse-position modulation (PPM), and subcarrier
modulation. The performance of these techniques in the presence
of multipath distortion is quantified. Techniques for multiplexing
the transmissions of different users are reviewed. Performance
of an experimental 50-Mb/s on–off-keyed diffuse infrared link is
The emergence of portable information terminals in work
and living environments is accelerating the introduction
of wireless digital links and local area networks (LAN’s).
Portable terminals should have access to all of the services
that are available on high-speed wired networks. Unlike
their wired counterparts, portable devices are subject to
severe limitations on power consumption, size and weight.
The desire for inexpensive, high-speed links satisfying
these requirements has motivated recent interest in infrared
wireless communication [1]–[44].
A. Comparison Between Infrared and Radio Media
As a medium for short-range, indoor communication,
infrared1 radiation offers several significant advantages over
Manuscript received October 10, 1996; revised December 12, 1996.
This work was supported by National Science Foundation Grants ECS9408957, ECS-9632829, and NCR-9308968, Hewlett–Packard, the University of California MICRO Program, IBM, and Siemens.
J. M. Kahn is with the Department of Electrical Engineering and
Computer Sciences, University of California, Berkeley, CA 94720 USA
(e-mail: jmk@eecs.berkeley.edu).
J. R. Barry is with the School of Electrical and Computer Engineering,
Georgia Institute of Technology, Atlanta, GA 30332-0250 USA (e-mail:
Publisher Item Identifier S 0018-9219(97)01832-X.
1 In general, the infrared region includes wavelengths between about
700 nm and 100 m. In this paper, unless otherwise noted, “infrared”
refers to the near-infrared band between about 780 nm and 950 nm.
radio.2 Infrared emitters and detectors capable of highspeed operation are available at low cost. The infrared
spectral region offers a virtually unlimited bandwidth that is
unregulated worldwide. Infrared and visible light are close
together in wavelength, and they exhibit qualitatively similar behavior. Both are absorbed by dark objects, diffusely
reflected by light-colored objects, and directionally reflected
from shiny surfaces. Both types of light penetrate through
glass, but not through walls or other opaque barriers, so
that infrared transmissions are confined to the room in
which they originate. This signal confinement makes it easy
to secure transmissions against casual eavesdropping, and
it prevents interference between links operating in different rooms. Thus, infrared wireless LAN’s can potentially
achieve a very high aggregate capacity, and their design
may be simplified, since transmissions in different rooms
need not be coordinated. When an infrared link employs
intensity modulation with direct detection (IM/DD), the
short carrier wavelength and large-area, square-law detector
lead to efficient spatial diversity that prevents multipath
fading. By contrast, radio links are typically subject to
large fluctuations in received signal magnitude and phase.
Freedom from multipath fading greatly simplifies the design
of infrared links.
The infrared medium is not without drawbacks, however.
Because infrared cannot penetrate walls, communication
from one room to another requires the installation of
infrared access points that are interconnected via a wired
backbone. In many indoor environments there exists intense
ambient infrared noise, arising from sunlight, incandescent
lighting and fluorescent lighting, which induces noise in
an infrared receiver. In virtually all short-range, indoor
applications, IM/DD is the only practical transmission
technique. The signal-to-noise ratio (SNR) of a directdetection receiver is proportional to the square of the
received optical power, implying that IM/DD links can
tolerate only a comparatively limited path loss. Often,
infrared links must employ relatively high transmit power
levels and operate over a relatively limited range. While the
transmitter power level can usually be increased without
fear of interfering with other users, transmitter power may
2 In this paper, the term “radio” is inclusive of the frequency bands that
are often referred to as “radio frequency,” “microwave,” and “millimeter
0018–9219/97$10.00  1997 IEEE
Table 1 Comparison Between Radio and IM/DD Infrared Systems for
Indoor Wireless Communication
Property of Medium
Bandwidth Regulated?
IM/DD Infrared
Passes Through Walls?
Approval not required.
Worldwide compatibility.
Less coverage.
More easily secured.
Independent links in different rooms.
Multipath Fading?
Simple link design.
Multipath Distortion?
Path Loss
Dominant Noise
Other Users
Background Light
X (t) Represents
SNR Proportional to
Average Power Proportional to
X (t) 2 dt
X (t) 2 dt
X (t) 2 dt
X (t)dt
Implication for IR
Limited range.
Difficult to operate outdoors.
Hight transmitter power requirement.
Choose waveform
X (t) with high peak-to-average ratio.
Fig. 1. Classification of simple infrared links according to the degree of directionality of the
transmitter and receiver and whether the link relies upon the existence of a LOS path between them.
be limited by concerns of power consumption and eye
safety, particularly in portable transmitters.
The characteristics of radio and infrared indoor wireless
links are compared in Table 1.
Radio and infrared are complementary transmission media, and different applications favor the use of one medium
or the other. Radio is favored in applications where user
mobility must be maximized or transmission through walls
or over long ranges is required and may be favored when
transmitter power consumption must be minimized. Infrared
is favored for short-range applications in which per-link bit
rate and aggregate system capacity must be maximized,
cost must be minimized, international compatibility is required, or receiver signal-processing complexity must be
B. Infrared Link Designs
Infrared links may employ various designs, and it is
convenient to classify them according to two criteria. This
classification scheme is shown in Fig. 1. The first criterion is the degree of directionality of the transmitter and
receiver. Directed links employ directional transmitters and
receivers, which must be aimed in order to establish a link,
while nondirected links employ wide-angle transmitters and
receivers, alleviating the need for such pointing. Directed
link design maximizes power efficiency, since it minimizes
path loss and reception of ambient light noise. On the other
hand, nondirected links may be more convenient to use,
particularly for mobile terminals, since they do not require
aiming of the transmitter or receiver. It is also possible
to establish hybrid links, which combine transmitters and
receivers having different degrees of directionality.
The second classification criterion relates to whether the
link relies upon the existence of an uninterrupted line-ofsight (LOS) path between the transmitter and receiver. LOS
links rely upon such a path, while non-LOS links generally
rely upon reflection of the light from the ceiling or some
other diffusely reflecting surface. LOS link design maximizes power efficiency and minimizes multipath distortion.
Non-LOS link design increases link robustness and ease of
use, allowing the link to operate even when barriers, such as
people or cubicle partitions, stand between the transmitter
and receiver. The greatest robustness and ease of use are
achieved by the nondirected-non-LOS link design, which is
often referred to as a diffuse link.
C. IM/DD Channels
Modulation techniques for radio wireless systems include
amplitude, phase, and frequency modulation (AM, PM, and
FM), as well as some generalizations of these techniques
[45]. Radio receivers employ one or more antennas, each
followed by a heterodyne or homodyne down-converter,
which is comprised of a local oscillator and a mixer.
Efficient operation of this mixer relies upon the fact that it
receives both the carrier and the local oscillator in a common electromagnetic mode. The down-converter output is
an electrical signal whose voltage is linear in the amplitude
of the received carrier electric field.
In a low-cost wireless infrared system, it is extremely
difficult to collect appreciable signal power in a single
electromagnetic mode. This spatially incoherent reception
makes it difficult to construct an efficient heterodyne or
homodyne downconverter for AM, PM, or FM, or to
detect AM or PM by any other means. For infrared links,
the most viable modulation is intensity modulation (IM),
in which the desired waveform is modulated onto the
instantaneous power of the carrier. The most practical
down-conversion technique is direct detection (DD), in
which a photodetector produces a current proportional to
the received instantaneous power, i.e., proportional to the
square of the received electric field.3
The modeling of infrared channels with IM/DD is illustrated in Fig. 2. The transmitted waveform
the instantaneous optical power of the infrared emitter.
The received waveform
is the instantaneous current
in the receiving photodetector, which is proportional to
the integral over the photodetector surface of the total
instantaneous optical power at each location. As indicated
in Fig. 2(a), the received electric field generally displays
spatial variation of magnitude and phase,4 so that “multipath
fading” would be experienced if the detector were smaller
than a wavelength. Fortunately, typical detector areas are
millions of square wavelengths, leading to spatial diversity
that prevents multipath fading.5 Thus when the detector
is moved by a distance of the order of a wavelength, no
change in the channel is observed. As the transmitted optical power
propagates along various paths of different
lengths, infrared channels are still subject to multipathinduced distortion. As we will see, this distortion is most
pronounced in links utilizing nondirectional transmitters
and receivers, and especially when non-LOS propagation
is employed. The channel can be modeled as a baseband
3 It is possible to employ FM of the infrared source, and to use an optical
filter to convert the received FM signal to an IM signal, which can then
be directly detected. For detection of binary frequency-shift keying, this
scheme can be made more efficient if the receiver employs two optical
filters and two photodetectors. Nonetheless, it does not offer a performance
gain over IM/DD. Moreover, with simple optical transmitters, it is difficult
to obtain a frequency deviation sufficiently large that practical filters could
be used for FM-IM conversion.
4 Under very unusual circumstances, with point or plane-wave source,
LOS propagation and carefully aligned detector, this spatial variation
would exhibit a regular pattern, but in typical cases of LOS or diffuse
propagation, it will appear random.
5 The detector is equivalent to a two-dimensional array of many antennas
whose receptions are squared, lowpass filtered, and summed.
Fig. 2. (a) Transmission and reception in an infrared link with
IM/DD. (b) Modeling link as a baseband linear, time-invariant
system having impulse response ( ), with signal-independent,
additive noise ( ). The photodetector has responsivity .
linear system, with instantaneous input power
, output
, and an impulse response
, as shown in
Fig. 2(b). Alternately, the channel can be described in terms
of the frequency response
which is the Fourier transform of
. It is usually appropriate to model the channel
as fixed, since
it usually changes only when the transmitter, receiver, or
objects in the room are moved by tens of centimeters. The
linear relationship between
is a consequence
of the fact that the received signal consists of many
electromagnetic modes [17]. By contrast, we note that
when IM/DD is employed in dispersive single-mode optical
fibers, the relationship between
is sometimes
nonlinear [46].
In many applications, infrared links are operated in the
presence of intense infrared and visible background light.
While received background light can be minimized by
optical filtering, it still adds shot noise, which is usually
the limiting noise source in a well-designed receiver (see
Section II-F). Due to its high intensity, this shot noise
can be modeled as white, Gaussian [47], and independent6 of
. When little or no ambient light is present,
the dominant noise source is receiver preamplifier noise,
which is also signal-independent and Gaussian (though
often nonwhite). Thus we usually model the noise
Gaussian and signal-independent. This stands in contrast to
the signal-dependent, Poisson noise considered in photoncounting channel models. Fluorescent lamps emit infrared
that is modulated in nearly periodic fashion; when present,
this adds a cyclostationary component to
. Ambient
noise sources are discussed in detail in Section II-D.
6 Because both signal and ambient light are received in many electromagnetic modes, ( ) contains no significant signal-ambient cross
Our baseband channel model is summarized by
where the “ ” symbol denotes convolution and
the detector responsivity (A/W). While (1) is simply a
conventional linear filter channel with additive noise, infrared systems differ from conventional electrical or radio
systems in several respects. Because the channel input
represents instantaneous optical power, the channel input is
and the average transmitted optical power
is given by
, which is
rather than the usual time-average of
represents amplitude. The average
appropriate when
received optical power is given by
where the channel dc gain is
. As we
will see in Sections II and IV, the performance of a digital
link at bit rate
is related to the receiver electrical SNR7
is dominated by a white Gaussian
assuming that
component having double-sided power-spectral density
From (5), we see that the SNR depends on the square of
the received optical average power, implying that IM/DD
infrared links must transmit at a relatively high power and
can tolerate only a limited path loss. This stands in contrast
to the case of conventional channels, where the SNR is
proportional to the first power of the received average
D. Current Infrared Communication Systems
At present, most infrared links are of the directed-LOS
or hybrid-LOS designs. The low path loss of these designs
minimizes the transmitter power requirement and permits
the use of a simple, low-cost receiver. Typically, these
links transmit using a single light-emitting diode (LED),
which emits an average power of several tens of mW
that is concentrated within a semiangle of 15 –30 . The
LED emission wavelength typically lies between 850 and
950 nm. This wavelength matches the responsivity peak
of the silicon positive-intrinsic-negative (p-i-n) photodiode.
In hybrid-LOS link designs, the photodiode is most often
encapsulated in a planocylindrical or hemispherical plastic
lens that serves to concentrate the received light, while
maintaining a relatively wide field of view (FOV), e.g., a
semiangle of the order of 60 . Directed-LOS links employ
an optical concentrator that restricts the FOV, usually with
7 We define SNR in terms of average optical power to facilitate comparison of the average optical power requirements of different modulation
techniques. Our definition of SNR differs from the conventional Eb =N0 .
the goal of providing a higher degree of optical concentration. Directed-LOS and hybrid-LOS links are relatively free
from multipath distortion, sometimes permitting them to
achieve bit rates above 100 Mb/s while maintaining a very
simple design. These link designs are well suited for pointto-point and some point-to-multipoint applications, but are
not suited for multiple-access networks, since it is difficult
to establish full bidirectional connectivity between more
than two transceivers.
Directed-LOS and hybrid-LOS links have been used for
many years in remote-control units and other unidirectional,
low-bit-rate applications. Over the past three years, the
Infrared Data Association (IrDA) has established standards
for short-range, half-duplex LOS links operating at bit
rates up to 4 Mb/s [44]. Two of the key features of IrDA
transceivers are low cost (well under $10) and low power
consumption (under 1 W while transmitting, and under 100
mW when idle or receiving). At present, more than 130
companies worldwide are members of IrDA.
There are several key components of the IrDA standards
[44]. The IrDA Serial Infrared Physical Layer defines standards for half-duplex links at several bit rates up to 4
Mb/s. 4 Mb/s links employ four-pulse-position modulation
(4-PPM), while 1.152 Mb/s links utilize on-off keying
(OOK) with return-to-zero (RZ) pulses having a duty cycle
of 0.25. Links operating at bit rates of 115.2 kb/s and
below employ OOK with RZ pulses having a duty cycle
of 0.1875 (shorter pulses are permitted in some cases).
IrDA-compliant transmitters must emit at a wavelength
between 850 and 900 nm into a semiangle (at half-power) of
15 –30 . Compliant receivers must have a FOV (semiangle
at half-effective light-collection area) of at least 15 . Most
IrDA receivers have a much larger FOV, so that most IrDA
links are of the hybrid-LOS type. IrDA links are required
to achieve a bit error rate (BER) not exceeding 10 (10
for 4 Mb/s links) over a range of at least 1 m, but many
links achieve a range as long as 3 m.
The IrDA Infrared Link Access Protocol (IrLAP) is derived from an existing asynchronous data communication
standard, the high-level data-link control (HDLC) protocol.
IrLAP utilizes most of the standard frame types defined
by HDLC. IrLAP links may be point-to-point or point-tomultipoint. A key feature of IrLAP is that when a link
is established, a negotiation process defines one node as
primary, and all other nodes as secondary. All transmissions
over the link must go to, or from, the primary node.
IrLAP defines procedures for link initialization, device
address discovery, connection start-up (including bit-rate
negotiation), data exchange, disconnection, link shutdown,
and device address conflict resolution.
The IrDA Infrared Link Management Protocol provides
the means for multiple software applications running in
each node to operate independently and concurrently, sharing the single link provided by IrLAP between the primary
node and each secondary node. This involves three processes: discovery of the services that the link currently
has available, multiplexing of the communications of several applications over the single link, and management of
the link, including provision for applications that demand
exclusive use of the link.
IrDA-standard transceivers are now an integral feature
of numerous portable and fixed information appliances, including laptop computers, personal digital assistants, printers, and wireless access points to wired networks. It is also
envisioned that IrDA transceivers will be incorporated into
cellular and desktop telephones, pagers, watches, digital
cameras, automobiles, public telephones, automatic teller
machines, information kiosks, and industrial machinery,
enabling new applications of short-range wireless communication.
Fig. 3(a)–(c) illustrate three general ways that IrDAstandard (or similar) links can be utilized (these three usage
models are not mutually exclusive). In Fig. 3(a), a portable
device (e.g., a laptop computer or personal digital assistant)
establishes an infrared link to another portable device or to
a fixed device (e.g., a desktop computer or printer). Typical
applications include printing, file system synchronization,
and “business card” exchange. In Fig. 3(b), a portable
device establishes an infrared connection to an access point
to a wired network (e.g., a networked desktop computer or a
dedicated infrared access point). At present, both Extended
Systems and Hewlett-Packard offer dedicated access points
that bridge between IrDA links and Ethernet LAN’s. Such
access points make a wide range of networked applications
available to portable devices. In the future, infrared access
points in public telephones may enable wireless access
to the Internet, while infrared links in automatic teller
machines might allow one to download “digital cash.”
As previously mentioned, links using directional, LOS
transmitters, such as current IrDA links, cannot easily
achieve full connectivity between more than two nodes,
making them unsuitable for forming multiple-access networks. However, it might be possible to build a hub
capable of establishing simultaneous point-to-point links
with several portable devices, as illustrated in Fig. 3(c).
Such a hub could be equipped with an internal switching
fabric, buffering and control circuitry to interconnect the
portables in a multiple-access LAN. At the same time,
the hub could serve as a bridge to a wired network.
For example, such a hub might be used for information
exchange among several portables in a conference room.
One technical challenge to building this hub is cochannel
interference between different inbound transmissions. Since
these transmissions will arrive from different directions,
it might be possible to separate them using an angle
diversity receiver. Angle diversity receivers are discussed
in Section V below.
As mentioned previously, among all infrared link designs,
diffuse links (nondirected-non-LOS links) are the most
easy-to-use and robust, since no aiming of the transmitter
or receiver is required, and since no LOS path between the
transmitter and receiver is required. However, diffuse links
have a higher path loss than their LOS counterparts, requiring higher transmit power and a receiver having a larger
light-collection area. Typical diffuse transmitters employ
several 850–900-nm LED’s, which are sometimes oriented
in different directions, to provide a diversity of propagation
paths. When transmitting, they typically emit an average
power in the range of 100–500 mW, making their power
consumption higher than a typical IrDA transmitter. Diffuse
receivers typically employ silicon p-i-n detectors encapsulated in hemispherical or plano-cylindrical lenses, which
provide some light concentration while maintaining a wide
FOV. Often they employ several detectors, in which case
each is oriented in a different direction.
When several diffuse transceivers are located in proximity to each other, they naturally form a shared bus
topology, making diffuse links suitable for multiple-access
LAN’s. However, “hidden nodes” may be present, i.e.,
each receiver cannot receive from, or even detect the presence of, each transmitter. When hidden nodes are present,
random medium-access control (MAC) protocols that rely
upon collision avoidance or detection, such as carriersense multiple access with collision detection (CSMA/CD)
or with collision avoidance (CSMA/CA), do not always
work reliably. Wireless LAN’s using diffuse infrared can
be formed in two different ways, which are illustrated in
Fig. 3(d).
In the first technique, diffuse infrared links are used
to achieve access to resources on a wired LAN. Clearly,
this architecture also permits communication among the
portable terminals via the wired backbone. This wireless
LAN architecture is well suited for wireless data communication in offices, hospitals, schools, factories, restaurants,
financial trading centers, or other heavily used environments, in which the cost of installing a backbone and
wireless access points can be justified. An example of
this type of wireless LAN is SpectrixLiteTM , made by the
Spectrix Corporation. This system utilizes a base station to
connect together up to 16 wireless access points, forming
a LAN having an aggregate capacity of 4 Mb/s. The
base station also bridges to a wired network (Ethernet or
Token Ring). Portable terminals equipped with wireless
LAN interfaces connect to the access points via 4 Mb/s
links employing OOK with RZ pulses. These diffuse links
are intended to achieve a BER of 10
over a range
of 15 m. Transmission over the wireless LAN is controlled by the centralized, deterministic CODIAC protocol.
Uplink bandwidth-reservation requests, uplink data and
downlink data employ a single wavelength, and are timedivision multiplexed together within a superframe interval.
All transmissions occur at times scheduled by the CODIAC
protocol, permitting portable transmitters and receivers to
“sleep” at other times, thus saving power.
In the second technique, diffuse infrared links are employed to achieve direct, peer-to-peer communication between a number of portable and/or fixed terminals. This
type of ad hoc interconnection is well suited to new or
temporary work groups, for collaboration while traveling
or at off-site meetings, or for setting up LAN’s in a home
or office environment in which all nodes are located within
a single room. IBM supplies a diffuse infrared ad hoc LAN
operating at 1 Mb/s using 16-PPM. It is intended to achieve
coverage of a 10 m
10 m region. The transceivers are
Fig. 3. Types of infrared wireless communication systems. (a) LOS, point-to-point (or
point-to-multipoint) link, such as those standardized by the IrDA. (b) Point-to-point infrared
link to a fixed device, which serves as a bridge to a wired network. (c) Hub capable of establishing
simultaneous point-to-point links with several portable devices. The hub interconnects the portables
in a multiple-access LAN, and bridges to the wired network. (d) LAN using diffuse, multiple-access
links to connect portable devices to a wired network. A peer-to-peer connection of two portable
devices is also shown.
not able to receive with high sensitivity while transmitting,
so they cannot perform collision detection, and the LAN
employs a CSMA/CA protocol. In order to perform collision avoidance, prior to initiating transmission, a transceiver
listens to the shared channel. If the channel is free, it
transmits a jam signal to reserve the channel, waits a time
sufficient for all the stations to receive the jam signal, then
transmits the payload packet. If acknowledgments indicate
that a packet has been lost (usually due to a collision),
the lost packet is retransmitted. When hidden nodes are
present, collision avoidance can fail, causing the CSMA/CA
protocol to crash. In this case, the IBM LAN changes the
protocol to a deterministic, bandwidth-reservation scheme,
which is stable, although it achieves lower throughput than
CSMA/CA would in the absence of hidden nodes.
Through judicious use of the technologies employed in
currently available systems, it is possible to enhance the
performance of wireless infrared systems significantly. It
appears likely that 10-Mb/s diffuse links and low-cost LOS
links operating at tens of Mb/s can be achieved. Even higher
bit rates will be desirable in future applications. Recent
research work suggests that using new techniques, lowcost infrared links operating in the 100 Mb/s range may be
achieved. In the remainder of this paper, we describe the
nature of the infrared communication medium, obstacles
to improved system performance and capacity, and the
Table 2 Comparison Between LED’s and LD’s (Shading Denotes an Advantage)
< 1005 to 5 nm
Spectral width
25–100 nm
(10–50 THz)
Modulation Bandwidth
Tens of kiloherz to tens of megahertz
(< 1 MHz to 2 THz)
Tens of kiloherz to tens of gigahertz
E/O Conversion Efficiency
Eye Safety
Generally considered eye-safe
Must be rendered eye-safe, especially for
Moderate to high
some of the means to overcome them. Section II describes
how to achieve a high SNR, which is the single most
difficult problem faced by the designer of an infrared link.
Multipath distortion on infrared channels is characterized
in Section III. Section IV provides a survey of various
modulation techniques for infrared systems, comparing
their power and bandwidth efficiencies, and characterizing
their performance on multipath channels. In Section V,
angle-diversity receivers and quasidiffuse transmitters are
discussed. Multiple-access techniques are the subject of
Section VI. Section VII describes an experimental 50-Mb/s
diffuse infrared link, and Section VIII provides some concluding remarks.
Achieving a high electrical SNR is the single biggest
problem facing the designer of an infrared link. The difficulty arises for two reasons. Firstly, the SNR of an IM/DD
link depends upon the square of the received optical average
power. This implies that one should transmit at relatively
high power, but available transmitter power may be limited
by considerations of eye safety and power consumption.
It also implies that one should design the link so as to
minimize path loss and employ a receiver having a large
light-collection area. Second, in many environments there
exists intense ambient infrared noise, which introduces
white shot noise and low-frequency cyclostationary noise
into the receiver. This noise can be minimized through
optical filtering and by employing a directional receiver,
which can separate the desired signal from the ambient
A. Infrared Transmitters and Eye Safety
The wavelength band between about 780 and 950 nm is
presently the best choice for most applications of infrared
wireless links, due to the availability of low-cost LED’s
and laser diodes (LD’s), and because it coincides with the
peak responsivity of inexpensive, low-capacitance silicon
photodiodes. The primary drawback of radiation in this
band relates to eye safety: it can pass through the human
cornea and be focused by the lens onto the retina, where it
can potentially induce thermal damage [49]. The cornea is
opaque to radiation at wavelengths beyond about 1400 nm,
considerably reducing potential ocular hazards, so that it has
been suggested that the 1550-nm band may be better suited
for infrared links. Unfortunately, the photodiodes presently
available for this band, which are made of germanium or
< 1400 nm
InGaAs, have much higher costs and capacitances per unit
area than their silicon counterparts. To our knowledge, at
present, all commercially available systems operate in the
shorter-wavelength band.
Table 2 presents a comparison between LED’s and LD’s.
LED’s are currently used in all commercial systems, due
to their extremely low cost and because most LED’s emit
light from a sufficiently large surface area that they are
generally considered eye-safe. Typical packaged LED’s
emit light into semiangles (at half power) ranging from
about 10 –30 , making them suitable for directed transmitters. Nondirected transmitters frequently employ multiple
LED’s oriented in different directions. Potential drawbacks
of LED’s include: 1) typically poor electro-optic power conversion efficiencies of 10–20% (though new devices have
efficiencies as high as 40%), 2) modulation bandwidths that
are limited to tens of MHz in typical low-cost devices, 3)
broad spectral widths (typically 25–100 nm), which require
the use of a wide receiver optical passband, leading to
poor rejection of ambient light, and 4) the fact that wide
modulation bandwidth is usually obtained at the expense of
reduced electro-optic conversion efficiency.
LD’s are much more expensive than LED’s, but offer
many nearly ideal characteristics: 1) electro-optic conversion efficiencies of 30–70%, 2) wide modulation bandwidths, which range from hundreds of MHz to more than 10
GHz, and 3) very narrow spectral widths (spectral widths
ranging from several nm to well below 1 nm are available).
To achieve eye safety with an LD requires that one pass
the laser output through some element that destroys its
spatial coherence and spreads the radiation over a sufficiently extended emission aperture and emission angle. For
example, one can employ a transmissive diffuser, such as
a thin plate of translucent plastic. While such diffusers can
achieve efficiencies of about 70%, they typically yield a
Lambertian radiation pattern, offering the designer little
freedom to tailor the source radiation pattern. Computergenerated holograms [29] offer a means to generate customtailored radiation patterns with efficiencies approaching
100%, but must be fabricated with care to insure that any
residual image of the LD emission aperture is tolerably
The eye safety of infrared transmitters is governed by
International Electrotechnical Commission (IEC) standards
[49]. It is desirable for infrared transmitters to conform to
the IEC Class 1 allowable exposure limit (AEL), implying
that they are safe under all foreseen circumstances of use,
and require no warning labels. At pulse repetition rates
higher than about 24 kHz, compliance with this AEL can
be calculated on the basis of average emitted optical power
alone. The AEL depends on the wavelength, diameter, and
emission semiangle of the source. At present, the IEC is
in the midst of revising the standards applying to infrared
transmitters. Based on proposed revisions, at 875 nm, an
IrDA-compliant source having an emission semiangle of
15 and diameter of 1 mm can emit an average power up
to 28 mW. At the same wavelength, a Lambertian source
(60 semiangle) having a diameter of 1 mm can emit up to
280 mW; at larger diameters, the allowable power increases
as the square of the diameter.
B. Optical Filters and Concentrators
Infrared receivers typically employ either longpass or
bandpass optical filters to attenuate ambient light. Longpass
filters can be thought of as essentially passing light at
all wavelengths beyond the cutoff wavelength.8 They are
usually constructed of colored glass or plastic, so that their
transmission characteristics are substantially independent of
the angle of incidence. Fig. 4(a) shows the transmission of
a common longpass filter, superimposed upon the responsivity curve of a typical silicon photodiode. As the silicon
device does not respond to wavelengths beyond about 1100
nm, the filter-photodiode combination effectively exhibits
a bandpass optical response, whose bandwidth is several
times that required to pass typical LED radiation. Longpass
filters are used in almost all present commercial infrared
Bandpass filters are usually constructed of multiple thin
dielectric layers, and rely upon the phenomenon of optical interference [50]. These filters can achieve narrow
bandwidths, leading to superior ambient light rejection
(bandwidths below 1 nm are available commercially). In
order to maximize the SNR, however, the transmitter optical
spectrum must lie within the filter bandwidth, implying that
when the filter bandwidth is made small, LD transmitters
need to be used. The transmission spectrum of a typical
bandpass filter is shown in Fig. 4(b). It is seen that the
bandpass shifts to shorter wavelengths as one increases ,
the angle at which light strikes the filter. Such a filter must
be used carefully if the receiver is intended to achieve a
wide FOV.
An infrared receiver detects an optical power
is proportional to its effective light-collection area. Increasing the photodiode area is expensive, and tends to
decrease receiver bandwidth and increase receiver noise.
Hence, it is desirable to employ an optical concentrator
to increase the effective area. Concentrators may be of
the imaging or nonimaging variety. The telescopes used
in long-range, free-space optical links represent examples
of imaging concentrators. Most short-range infrared links
employ nonimaging concentrators.
8 This is true within the band over which silicon detectors respond,
though the filters do attenuate at still longer wavelengths.
Fig. 4. (a) Responsivity of typical silicon p-i-n photodiode, transmission of typical longpass optical filter (Schott RG-780), and
overall responsivity, which is the product of transmission and
responsivity. (b) Polarization-averaged transmission of an optical
bandpass filter for rays incident at angle . This filter is of a
25-layer, three-cavity design [10], [20].
Ignoring reflection losses, a bare detector achieves an
effective signal-collection area of
is the detector physical area and
is the angle
of incidence with respect to the receiver axis. Adding a
concentrator and filter, the effective signal-collection area
is the signal transmission of the filter,9
is the concentrator gain and
is the concentrator FOV
9 T ( ) may represent an average over the filter transmission at different
wavelengths (if the source spectrum is not narrow) and/or angles of
incidence upon the filter (if different rays strike the filter at different angles
of incidence). All losses arising from reflections, e.g., at the concentratordetector interface) are included in Ts ( ).
(semiangle). Usually,
. Nonimaging concentrators
exhibit a trade-off between gain and FOV. An idealized
nonimaging concentrator [51] having an internal refractive
index achieves a gain:
From (8), we see that as the FOV is reduced, the gain
within the FOV is increased.
The hemispherical lens is an important nonimaging concentrator [4], [16], [20], and is widely employed in commercial infrared systems [see Fig. 5(a) and (b)]. It achieves
a wide FOV and omnidirectional gain, making it especially
suitable for use in nondirected links. A hemisphere can
over its entire FOV.10
While this is sometimes called “omnidirectional gain,” a
hemisphere-based receiver is not truly omnidirectional, but
has an effective area
When longpass
filtering is employed, a planar longpass filter can be placed
between the hemisphere and the detector, as shown in
Fig. 5(a).
When bandpass filtering is utilized, it is not desirable
to employ a planar filter in the configuration shown in
Fig. 5(a). As , the angle from which rays are received,
shifts, so does , the angle at which light strikes the filter.11
This shifts the filter passband, as described above, decreasing the filter transmission
for some . Instead, as
shown in Fig. 5(b), the bandpass filter should be deposited
or bonded onto the outer surface of the hemispherical
concentrator [16], [20]. Regardless of the angle
which the signal is received, rays that reach the detector
are incident upon the filter at small values of the angle ,
minimizing the shift of the filter passband, and maximizing
its transmission. Thus with a hemispherical filter, it is
possible to simultaneously obtain a narrow bandwidth and
wide FOV.12
The compound parabolic concentrator (CPC) [51] is
another nonimaging concentrator that is widely used in
infrared links [12]. It can achieve much higher gain than
the hemisphere, but at the expense of a narrower FOV,
making it especially suitable for directed links. A CPC
having FOV
can achieve a gain close to that
given by (8). As shown in Fig. 5(c), a longpass or bandpass
filter can be placed on the front surface of the CPC. The
restricted FOV of a typical CPC is well matched to the
10 In
order to achieve this, the hemisphere must be sufficiently large in
relation to the detector, i.e., R > n2 r , where r and R are the detector and
hemisphere radii, respectively [10], [20].
11 When light is incident at angle
across a broad area of the receiver,
rays strike the filter over a range of that are close to .
12 Rays are incident upon the filter at angles
01 nr=R , where r and R are the detector and hemispheremaxradii,
respectively [10], [20]. Typically max is of the order of 30 . In order
to choose a narrow filter bandwidth that still achieves high Ts
reasonable procedure is to: 1) choose the filter center wavelength so that
, the short-wavelength edge of the filter passband lies at the
for signal wavelength, and then 2) choose the filter bandwidth so that
for max , the long-wavelength edge of the passband is at the signal
wavelength. For example, the filter shown in Fig. 4(b) would result from
this procedure for a signal wavelength of 840 nm and for max
sin (
( )
= 30
Fig. 5. Nonimaging optical concentrators: (a) hemisphere with
planar optical filter, (b) hemisphere with hemispherical optical
filter, (c) CPC with planar optical filter, and (d) effective
light-collection areas achieved by ideal concentrators with lossless
with respect to the
filters. Light rays are received from angle
receiver axis, and strike the filter at angle . The concentrator
has a FOV (semiangle) c and refractive index n, while the
photodetector has area A.
restricted angular acceptance of a narrow bandpass filter.
For example, a 30-nm-wide bandpass filter can be coupled
to achieve near-ideal
with a CPC having a FOV
performance [18]. Moreover, a second, inverted CPC can be
placed in front of such a bandpass filter-CPC combination
to widen its input FOV to
, while reducing the
gain to . Such a CPC-filter-CPC structure can achieve a
wide FOV and narrow passband, but without requiring a
hemispherical bandpass filter [18]. The principal drawback
of CPC’s is their excessive length, particularly for small
. There exist more compact nonimaging concentrators
that achieve performance near that of CPC’s [52].
Fig. 5(d) compares the effective signal-collection areas
achieved by ideal nonimaging concentrators with lossless
filters, which have been computed using (7) and (8). The
trade-off between FOV and gain is readily apparent.
It is important to note that a large fraction of the
energy exiting a nonimaging concentrator does so at oblique
angles, so that careful attention to antireflection coating and
index matching at the concentrator-photodetector interface
is required in order to achieve near-ideal performance.
C. Channel DC Gains
The frequency responses of infrared channels are relatively flat near dc, so for most purposes, the single most
important quantity characterizing a channel is the dc gain
,13 which relates the transmitted and received average
powers via (4). In this section we compute the dc gains of
common link configurations (see Fig. 6).
In LOS links (either directed, hybrid, or nondirected), the
dc gain can be computed fairly accurately by considering
only the LOS propagation path. This approximation is
particularly accurate in directed-LOS links. We consider the
link geometry shown in Fig. 6(a). Suppose the transmitter
emits an axially symmetric radiation pattern described by
the radiant intensity (W/sr)
.14 At the receiver,
located at distance and angle with respect to the transmitter, the irradiance
The received power is
and using (7),
we obtain the channel dc gain:
. From (9), we
which we observe is proportional to
observe that if and
are fixed, the most effective
means to increase
are to increase the detector area
, and increase the concentrator gain
(by increasing
the refractive index and decreasing the FOV
). Under
some conditions, the power efficiency of a LOS link can
be maximized by optimization of the transmitter radiant
[10], [20]. For example, suppose the transmitter is pointed straight down from the ceiling, and as the
receiver is moved about the room, it is pointed straight up
at the ceiling. Then the condition
is maintained.
In this case,
can be enhanced at some values of
to compensate for changes in
, and as the
receiver is moved.
13 The channel has an optical path loss of 10 log
10 H (0) (measured in
optical decibels), and an equivalent electrical dc gain of 10 log10 H 2 (0)
(measured in electrical decibels).
14 Here, R () is normalized so that 2 R () sin d = 1:
Fig. 6. Geometries used in channel gain calculations: (a) LOS,
(b) directed-non-LOS or hybrid non-LOS, and (c) nondirected
non-LOS (diffuse).
The emission from a variety of practical LOS transmitters
can be modeled reasonably using a generalized Lambertian
radiant intensity
[1]. The
, the transmitter semiangle (at half
order is related to
power), by
. For example,
(Lambertian transmitter) corresponds to
, while
(typical directed transmitter) corresponds to
. The channel dc gain is given by
by narrowing
If is kept very small, we can increase
the transmitter semiangle
, thereby increasing .
Non-LOS infrared links exploit the fact that a wide
variety of common building materials are efficient diffuse
infrared reflectors. In the 800–900-nm range, typical plaster
walls and acoustical ceiling tiles have diffuse reflectivities
in the range of 0.6–0.9, while darker materials often
exhibit lower values of [1], [17]. Most building materials
(with the notable exception of glass) are approximately
Lambertian reflectors, i.e., they scatter light with a power
per unit solid angle proportional to the cosine of the angle
with respect to the surface normal, independent of the angle
of incidence.15
To compute the gain of directed-non-LOS or hybridnon-LOS channels, we refer to the geometry shown in
Fig. 6(b). One of the principal advantages of this link design
is that, assuming that the transmitter illuminates a fairly
small region of the ceiling,16 the channel gain depends only
, the horizontal separation between the illuminated
spot and the receiver, not the transmitter-receiver horizontal
. If the ceiling of diffuse reflectivity
located a distance above the receiver, then the received
signal irradiance is
. The
received power is
, and the channel
dc gain is
In directed-non-LOS or hybrid-non-LOS links, the most
effective means to increase
are to increase the detector
and the concentrator gain
(by increasing the
refractive index and decreasing the FOV
To compute the dc gain of nondirected-non-LOS (diffuse) channels, one must consider the effect of multiple
reflections from surfaces within the room. As a first-order
approximation, we consider only the first bounce off of the
ceiling [1], [3], and consider the configuration shown in
Fig. 6(c). We will assume that the transmitter is pointed
straight upward and emits a Lambertian pattern. We will
also assume that the receiver is pointed straight upward,
and employs a concentrator of FOV
that achieves
omnidirectional gain
) and an
. The transmitter
omnidirectional filter having
and receiver are located, respectively, at coordinates (0, 0)
in the horizontal
plane. We integrate the
power reflected from each ceiling element to obtain (see
(12) at the bottom of the page). The most effective means
to increase
are to increase the detector area
the concentrator gain (this should be done by increasing
the refractive index , but not by reducing the FOV
Equation (12) predicts that at large , the channel gain
is expected to be proportional to
. A diffuse
link need not employ a Lambertian transmitter, and the
15 As shown in [1], for very large incidence angles (of order 70
and larger), even nominally diffuse reflectors exhibit a strong specular
16 The diameter of the illuminated spot should be somewhat smaller than
dsr .
Fig. 7. Optical path loss of a diffuse infrared links employing
a Lambertian transmitter and a detector of area A = 1 cm2 .
(a) Measured in a single room having a ceiling of 80% diffuse
reflectivity, for transmitter and receiver located 1.2 m and 1.6 m
below the ceiling, respectively. (b) Measured in a collection of
rooms, showing effect of shadowing at receiver [17].
increase of path loss with
can be made more gradual by
employing several emitters angled in different directions
[1], [32]. This strategy is employed in some commercial
diffuse links.
Fig. 7(a) shows the path loss of a diffuse link, which
was measured in a typical office [17]. The one-bounce
theory (12) is reasonably accurate, but it underestimates
by a few decibels. Accurate prediction of
requires the inclusion of higher-order reflections (see the
discussion of multipath channel simulation in Section IIIB). For example, to model propagation in a large open
exceeding about 5 m, reflections up to fifth
office, for
order are required [11].
Fig. 7(b) presents the path losses of diffuse links mea-
Fig. 8. (a) Optical power spectra of common ambient infrared sources. Spectra have been scaled
to have the same maximum value. (b) Detected electrical power spectrum of infrared emission from
a fluorescent lamp driven by a 22-kHz electronic ballast [38].
sured in several rooms [17]. In the unshadowed case, the
propagation path was not blocked, while shadowing was
effected by a person standing next to the receiver, so as to
block the strongest propagation path (which was dominated
by the first bounce off of the ceiling). In these diffuse links,
shadowing decreased
by 2–5 dB. For comparison, we
note that in nondirected-LOS links, shadowing was found
to decrease
by 7–10 dB. This comparison illustrates
the robustness of diffuse links against shadowing.
D. Ambient Light Noises
Many environments contain intense ambient infrared
radiation arising from sunlight, skylight, incandescent and
fluorescent lamps, and other sources [1], [41]. The optical
power spectra of some common infrared sources are shown
in Fig. 8(a).17 Sunlight, skylight, and incandescent lamps
17 It should be emphasized that the power spectra in Fig. 8(a) have been
normalized to have equal maximum value. Direct sunlight, when present,
is typically much stronger than the other two sources.
represent essentially unmodulated sources18 that can be
received at an average power much larger than the desired
signal, even when optical filtering is employed. The resulting dc photocurrent causes shot noise, which is a dominant
noise source in typical infrared receivers, as shown below.
Here, we compute the optical power received from steady
ambient light sources, which will be used to compute the
shot noise they induce. We assume that the receiver employs
a bandpass optical filter of noise bandwidth19
and peak
transmission . The ambient light noise is assumed to have
a spectral irradiance
nm)] that is independent
of wavelength within the filter bandwidth. If the ambient
light originates from a localized source at angle
respect to the receiver normal,20 then the received ambient
18 As shown in [41], incandescent lamps are modulated periodically at
the power-line frequency, but because of their slow response time, this
modulation contains few higher harmonics of that frequency.
19 This noise bandwidth is generally close to, but slightly greater than,
the 3 dB bandwidth of the filter.
20 This is similar to the geometry of Fig. 6(a) or (b) but with the
transmitter replaced by the noise source.
optical average power is
If, instead, the ambient light is “isotropic”21 and the receiver
, over
employs an ideal concentrator having a FOV
which it achieves a constant gain given by (8), then the
received ambient optical power is given by [10], [18]
We note that (14) is independent of the concentrator FOV
. This occurs because as
is varied, the gain (8)
varies in such a way that the total received power
remains constant. If the concentrator is nonideal, (14) will
Fluorescent lamps emit strongly at spectral lines of
mercury and argon that lie in the 780–950-nm band of
interest for low-cost infrared systems. Fluorescent-lamp
emission is modulated in a near-periodic fashion at the
lamp drive frequency, and the detected electrical power
spectrum contains discrete components at harmonics of
the drive frequency. Traditionally, such lamps have been
driven at the power-line frequency (50 or 60 Hz), and
their electrical spectrum has contained energy at harmonics up to tens of kilohertz [3], [41]. However, recently
introduced, high-efficiency “electronic ballasts” drive the
lamps at frequencies of tens to hundreds of kilohertz.
Their detected electrical spectrum contains energy up to
hundreds of kilohertz [38], [41], making such lamps a
potentially much more serious impairment to infrared links.
Fig. 8(b) presents the detected electrical power spectrum
from a lamp driven by a 22-kHz ballast. The system penalty
caused by fluorescent-light noise depends strongly on the
modulation scheme employed, so we defer discussion of
this impairment until Section IV-D.
E. Photodetectors and Preamplifiers
As mentioned previously, the availability of low-cost,
low-capacitance, large-area silicon photodiodes strongly
favors choice of the 780–950 nm band over the region
beyond 1400 nm for most infrared link applications. Two
types of medium- and large-area silicon photodiodes are
widely available: ordinary positive-intrinsic-negative (p-in) photodiodes and avalanche photodiodes (APD’s) [53].
APD’s are essentially p-i-n devices that are operated at very
high reverse bias, so that photogenerated carriers create
secondary carriers by impact ionization, resulting in internal
electrical gain. APD’s are favored in DD optical receivers
when there is little ambient-induced shot noise, because
their internal gain helps overcome preamplifier thermal
noise, increasing the receiver SNR. APD-based receivers
can lead to impressive infrared link performance when
ambient light is weak [12]. When ambient-induced shot
noise is dominant, however, use of an APD results in a net
decrease in SNR [10], because the random nature of the
APD’s internal gain increases the variance of the shot noise
21 More
precisely, it is received with equal radiant intensity for 0
by a factor greater than the signal gain. Additional drawbacks of APD’s include their high cost, requirement for
high bias, and their temperature-dependent gain. Ordinary
silicon p-i-n photodiodes are employed in all commercial
infrared links at present. We will restrict the remaining
discussion to ordinary p-i-n photodiodes.
When it receives an instantaneous optical power
p-i-n photodiode produces an instantaneous photocurrent
, where the responsivity is
(A/W). The
responsivity of a typical silicon p-i-n photodiode is shown
in Fig. 4(a), and is seen to peak near 950 nm. We assume
that the desired signal and ambient light are received with
average optical powers and
, respectively. Assuming
, the ambient light induces in the detector
a shot noise current
, which is essentially white,
Gaussian, and independent of the desired signal,22 and
which has a one-sided power-spectral density (PSD):
is the electronic charge.
has units of
Among preamplifier designs, the transimpedance type
is best suited to most infrared link applications, because
its achieves a large dynamic range and a wide bandwidth
without the need for equalization [54]. Under typical conditions, lower noise is achieved if the front-end device is a
field-effect transistor (FET), rather than a bipolar-junction
transistor (BJT) [22], [54]. If power consumption is constrained, however, a BJT may achieve superior results [22].
We will assume the use of a FET-based transimpedance
preamplifier. A simplified schematic of such a circuit is
shown in Fig. 9(a). Assuming that the amplifier of gain
is ideal, the circuit has a single-pole response with
cutoff frequency
In infrared receivers, the total input capacitance
is usually dominated by the detector
capacitance , because of the large detector area required
to achieve a high SNR. The total input-referred noise PSD
of this preamplifier is
where the dominant contributions to the input-referred
thermal noise PSD are given by
Here, is Boltzmann’s constant, is absolute temperature,
is the FET channel noise factor, and are the FET
noise coefficients, and
is the FET drain current. These
PSD’s are plotted in Fig. 9(b), assuming parameters that
might be typical of a receiver operating in a 10-Mb/s diffuse
link. The first term in (17) is a white noise arising from the
22 Although the desired signal does contribute to this shot noise, it is
usually not important to include this contribution, even when Pn = 0,
because it is typically much smaller than the preamplifier thermal noise.
Fig. 9. (a) Simplified schematic of FET-based transimpedance preamplifier. (b) Dominant input-referred noise power spectral densities (one-sided). (c) Dominant input-referred noise variances. Parameters assumed in (b) and (c) include: isotropic bright skylight of spectral irradiance
= 2 nm), optical filter noise bandwidth n
nm, and optical concentrator
2 , responsivity
. The p-illuminated silicon p-i-n photodetector has area A
noise gain g
: A/W, depletion width w
mm, capacitance Cd
: pF, and reverse bias V
V. The ambient-induced dc photocurrent is Idc
: mA. The feedback resistance is RF
k . The FET parameters are gm
mS, Cgs Cgd
: pF, fT
: GHz, ID
mA, K
fA, a
= 6 mW (cm
= 0 53
= 294
= 60
= 40
1 = 30
= 17 5
= 28 6
+ = 10
feedback resistor
and is minimized by choosing
large as possible, while still achieving a sufficiently large
preamplifier cutoff frequency.23 The second term in (17)
arises from the FET white channel noise. It increases as
, so that it tends to become the dominant noise source in
receivers for very high bit rates. This term is minimized by
. Because
minimizing the total input capacitance
is often dominated by the detector capacitance
, this
often amounts to minimizing
. The second term is also
minimized by choosing an FET having a transconductance
as high as possible, subject to constraints on power
consumption. The third term in (17) is proportional to ,
so that it also becomes important at high bit rates. It arises
from the FET
channel noise and is minimized through
choice of small
, high
, and an FET type having
small .
A typical p-i-n photodiode is illuminated through either
the p or n contact and has a capacitance given by
, where
is the device area,
is the depletion23 For OOK with nonreturn-to-zero (NRZ) pulses, a cutoff frequency
equal to the bit rate is typically required.
= 0 1 cm
= 64
= 30
= 10
= 20
layer thickness, and is the semiconductor permittivity.
For high-bit-rate receivers, in which the second two terms
of (17) become important, this seems to imply that
should be chosen as large as possible. It should be noted,
however, that increasing eventually leads to transit-time
limitations in the photodetector frequency response [53].
Simple considerations show that for photodiodes much
thicker than the photon absorption length and at electric
fields well below saturation, the transit-time-limited cutoff
frequency is approximately proportional to
, where
is the mobility of the carrier type that is collected by
the contact opposite the illuminated contact, and
is the
photodiode reverse bias. Fig. 10 shows the theoretically
calculated transit-time-limited frequency response of silicon
p-i-n photodiodes illuminated through the p and n contacts.
The p-illuminated photodiode has a much wider bandwidth
than the n-illuminated device, since electrons have a mobility much higher than holes. It has been shown that when
a constraint is placed on the bias voltage , receiver performance is optimized when is chosen so that the transittime-limited 3-dB cutoff approximately equals the bit rate
(assuming OOK modulation) [22]. Metal-semiconductor
while the thermal-noise variance is given by
Fig. 10. Transit-time-limited frequency responses of silicon p-i-n
photodiodes illuminated through the n- and p-doped contacts.
Responses have been normalized to 0 dB at dc. These responses
assume: depletion width w = 100 mm, reverse bias V = 30 V,
uniform electric field in depletion region, and absorption coefficient
= 103 cm01 ( = 800 nm).
metal (MSM) photodiodes have a different relationship
between capacitance and transit time than conventional pi-n photodiodes, owing to their different geometry. While
MSM detectors can yield improved performance in opticalfiber receivers [56], analysis suggests that an improvement
is not likely in very large-area infrared receivers [22].
F. Receiver SNR and BER
In this section we compute the receiver SNR and BER.
We assume here that the receiver transmits at bit rate
using on–off keying (OOK) with NRZ pulses. The
transmitted average power is , and the received average
power is
, where the channel dc
gain can
be computed as in Section II-C above. We assume that the
channel is distortionless, i.e., it has a gain
all frequencies of interest. Following [55], at the receiver,
the preamplifier is followed by an equalizer that converts
the received pulse to one having a raised-cosine Fourier
transform with 100% excess bandwidth. The equalizer gain
is chosen so that when sampled, its output is either zero
(A), ignoring noise. Each sample of the equalizer
output contains a Gaussian noise having a total variance
that is the sum of contributions from shot and thermal
The SNR is expressed as
and the BER is given by
We have defined the noise-bandwidth factors
, and
. Examining (20) and (21),
we see that the shot noise and
noise, which have
white input-referred PSD’s, lead to variances proportional
to the bit rate
. The FET white channel noise and
channel noise, with input-referred PSD’s proportional to
and , lead to variances proportional to
respectively. The various terms in the noise variances (20)
and (21) are plotted as a function of bit rate Fig. 9(c). For
the parameters considered there, which might be typical of
a 10-Mb/s diffuse receiver, shot noise is dominant up to
a bit rate of tens of Mb/s, above which the white channel
noise becomes the leading term.
Examining the SNR (19), we note that the numerator is
always proportional to the square of the detector area, i.e.,
. We note that the shot noise variance (20) is always
proportional to the detector area, . Hence, if shot noise
is the dominant noise source, then the SNR is proportional
to the detector area, i.e.,
. The thermal noise
variance (21) is a complicated function of , and is only
independent of A if the
term is always dominant.
When the
noise is the dominant noise source, then the
SNR is proportional to the square of the detector area, i.e.,
Fig. 11 illustrates the average transmitter power
required to achieve 10
BER in nondirected-LOS links
[10], [20]. These links operate in the presence of bright
skylight, and shot noise is assumed to be the dominant noise
source. This example considers the use of a hemispherical
concentrator, and compares the performance of bandpass
filters in planar and hemispherical geometries [as shown in
Fig. 5(a) and (b)]. The transmitter is placed at ceiling height
in the center of the cell. For each cell radius, the optical
filter bandwidth and center wavelength are optimized jointly
with the transmitter radiation pattern so as to minimize
the transmitter power required to achieve the desired BER
everywhere within the cell. Use of the hemispherical filter
reduces the transmitter power requirement by about 3
dB over the planar filter. This occurs because only the
hemispherical filter is able to achieve simultaneously a
narrow bandwidth and high signal transmission over a
wide FOV. We note that if the link were to employ PPM
instead of OOK, the transmitter power requirements could
be further reduced by several dB (see Section IV).
, where
For example, to achieve
dB. The shot-noise variance is given by
While nondirected propagation alleviates the need for
physical alignment between the transmitter and the receiver,
a major drawback of this approach is the signal distortion
caused by reflections from ceilings, walls, and other objects.
In this section, we describe experimental characterization of
Fig. 11. Average transmitter power required to achieve 1006
BER using OOK in nondirected-LOS links, comparing planar and
hemispherical bandpass filters. The transmitter is placed at ceiling
height in the center of the cell. Shot noise induced by skylight is
the dominant noise source. The detector area is A = 1 cm2 , and
the hemispherical concentrator has a 2-cm radius and refractive
index n = 1:7: For each cell radius, the bandpass filter bandwidth
and center wavelength are optimized jointly with the transmitter
radiation pattern. For example, at a 5-m cell radius, the optimized
planar and hemispherical filters have bandwidths of 70.6 nm and
10.7 nm, respectively [10], [20].
multipath channels, as well as techniques to simulate and
model them. In Section IV, we will quantify the penalties
that multipath distortion causes in very high-bit-rate links.
A. Experimental Results
Two experimental studies have characterized multipath
distortion in nondirected links, obtaining similar results
[13], [17]. In [17], we employed a vector network analyzer to perform swept-modulation-frequency characterization [48] of the channel frequency response
The 832-nm transmitter emitted a Lambertian radiation
pattern. During all measurements, the receiver was placed
at desk height and pointed upwards. To form nondirectedLOS configurations (referred to in this section simply
as “LOS”), the transmitter was placed near the ceiling
and pointed straight down, while for nondirected-non-LOS
(diffuse) configurations, it was placed at desk height and
pointed straight up. The channel impulse response
obtained by inverse Fourier transformation of
Fig. 12 presents the magnitude and impulse responses
of four different nondirected link configurations, measured
in an empty conference room. While the details of channel
responses depend on the link geometry, responses measured
at all positions in all rooms exhibit qualitative similarity
to Fig. 12. Unshadowed LOS impulse responses are dominated by a short initial pulse, and the strongest distinct
reflections typically arrive 15–20 ns after the initial pulse.
Dominance of the short initial pulse leads to magnitude
responses that are flat at high frequencies. Unshadowed
diffuse impulse responses exhibit a significantly wider
initial pulse, which has a width of about 12 ns at 10%
height, corresponding to the existence of a continuum of
different path lengths between illuminated portions of the
Fig. 12. Responses
nondirected-non-LOS (diffuse) channels. Measurements were performed
in a 5.5 m 7.5 m room having a 3.5-m-high ceiling. Shadowing
was effected by a person standing next to receiver. Detector area
was A = 1 cm2 : (a) Frequency responses. (b) Impulse responses
obtained by inverse Fourier transformation of the frequency
responses using a 300-MHz Hamming window [17].
ceiling and the receiver. This continuous distribution of path
delays leads to a steady decrease in the channel magnitude
response at high frequencies. For all channels, the impulse
response may contain significant energy as long as 70
ns after its initial nonzero excursion. The dc gain of all
channels is enhanced over that at high frequencies, because
it includes the contribution due to the entire duration of the
impulse response.
Shadowed channels exhibit characteristics that are
slightly less predictable than the unshadowed channels.
The shadowed LOS impulse response typically resembles
the corresponding unshadowed response with the dominant
initial pulse removed, since only indirect propagation paths
remain. We observe that in LOS configurations, shadowing
significantly degrades the channel frequency and impulse
responses. Diffuse configurations are far less vulnerable to
shadowing than their LOS counterparts, because in diffuse
configurations there exist many possible propagation paths
between the illuminated ceiling area and the receiver.
In diffuse configurations, shadowing produces a slight
broadening of the impulse response, and a slight increase in
the rate of rolloff of the magnitude response with increasing
A useful measure of the severity of intersymbol interference (ISI) induced by a multipath channel
is the
channel root-mean-square (rms) delay spread . As shown
in Section IV, the delay spread of a channel is a remarkably
accurate predictor of ISI-induced SNR penalties, independent of the particular time dependence of that channel’s
impulse response. The delay spread is computed from the
impulse response using
where the mean delay
is given by
The impulse response
and delay spread
can be
considered to be deterministic quantities, in the sense that
as long as the positions of the transmitter, receiver and
intervening reflectors are fixed,
are fixed. This
stands in contrast to the case of time-varying radio channels,
where the rms delay spread is interpreted as a statistical
expectation of a random process [48].
Fig. 13(a) and (b) present the rms delay spread versus horizontal separation for measured multipath channels
[17]. In the absence of shadowing, LOS channels, whose
impulse response is dominated by a short initial pulse,
generally yield the smallest delay spreads, ranging from
a measurement-limited 1.3 ns to about 12 ns. Unshadowed
diffuse channels exhibit delay spreads that lie in the same
range, but which are systematically slightly larger, due
to the finite temporal spread of the dominant reflection
from the ceiling. Shadowing increases the delay spread of
both LOS and diffuse channels but, as might be expected,
the increase is relatively modest for the latter channels.
Shadowed LOS channels consistently exhibit the largest
delay spreads, typically between 7 and 13 ns.
B. Simulation and Modeling
Simulation techniques for radio multipath channels frequently assume specular reflection and employ “ray tracing.” By contrast, infrared multipath propagation is usually
dominated by diffuse reflection, a process in which the
energy reflected from each surface element follows a Lambertian distribution, independent of the angle of incidence.
We have developed a technique to simulate multipath
infrared propagation, which can account for an arbitrary
number of diffuse reflections from room surfaces [9], [10].
This technique is based on the observation that given a
particular source and receiver in a room with reflectors,
light from the source can reach the receiver after any
number of reflections. Therefore, the impulse response can
Fig. 13. RMS delay spreads of (a) nondirected-LOS and (b)
nondirected-non-LOS (diffuse) channels. Shadowing was effected
by a person standing next to receiver [17].
be written as an infinite sum:
is the response of the light undergoing exactly
reflections. Reflecting surfaces in the room are divided
into small elements,
The LOS response
between and
is computed first. Each higherorder term
, is then calculated recursively
Examining (25), we see that each reflecting element
makes a contribution to the -bounce response
is composed of the LOS response from the source
element , convolved with the
-bounce response
between that reflecting element and the receiver. Fig. 14
illustrates the simulation of a diffuse infrared channel. The
individual contributions from bounces
are shown
response having one free parameter, which controls the
channel delay spread. By considering an impulse response
of this functional form and varying this parameter, one can
reproduce a set of channels that produce ISI equivalent in
effect to an ensemble of real channels. For purposes of
evaluating ISI caused by multipath infrared propagation,
this functional model is analogous to the statistical models
used in analysis of radio systems. A suitable channel model
is obtained by considering a diffuse link consisting of a
Lambertian transmitter that is directed toward a diffuse
reflector of infinite extent, and a receiver co-located with
the transmitter [35]. For such a link, the impulse response
can be obtained in closed form as
is the unit step function and is related to
, the rms delay spread, by
. Although
this model is derived by considering an unshadowed diffuse
link, it is found to be accurate for LOS and diffuse links,
with or without shadowing, as shown in Section IV.
Fig. 14. Simulation of a multipath propagation in a diffuse link.
(a) Contributions to impulse response from light that has undergone
k bounces for k = 1; 2; 3. (b) Comparison of simulation to experiment. In (b), the simulated impulse response has been convolved
with the impulse response of the experimental measurement system
[9], [10].
in Fig. 14(a). The term
is absent, as there is no
direct LOS path between the transmitter and receiver. We
successively become
see that as increases, the
broader in time, while their integrated areas decrease.
Fig. 14(b) compares the simulated impulse response to an
experimental measurement, showing that the two are in
fairly good agreement. In order to facilitate comparison,
the simulated impulse response has been convolved with the
impulse response of the experimental measurement system.
Simulations of LOS channels are also in good agreement
with experimental measurements, leading us to believe that
multipath infrared propagation is well described by the
model based on multiple diffuse reflections.
As we will see in Section IV, the delay spread of a
channel is a remarkably accurate predictor of ISI-induced
SNR penalties, independent of the particular time dependence of that channel’s impulse response. This suggests that
there should exist some functional model for the impulse
Nondirected infrared channels can be modeled as fixed,
linear systems with additive white Gaussian noise (AWGN),
as summarized in (1). The transmitted waveform
be nonnegative, and it is related to the average transmitted
optical power
by (3). The channel, described by impulse
, exhibits multipath distortion that can cause
ISI in high-bit-rate links. When the ISI is relatively mild,
it leads to an optical power penalty, but if it is severe, it
may lead to a BER floor.
The average transmitted optical power governs the eye
safety and electrical power consumption of the transmitter.
Hence, the most important criterion for evaluating modulation techniques is the average received optical power
required to achieve a desired BER. We will normalize
these received average optical power requirements to those
required to achieve the desired BER when transmitting
OOK over an ideal channel having impulse response
. That is, on an ideal channel, OOK has a normalized
power requirement of 0 dB. We have defined the receiver
electrical SNR (5) to be proportional to , so an equivalent
criterion is the value of the SNR required to achieve
a desired BER. We note that because of this squarelaw relationship, a 1-dB change in the average power
corresponds to a 2-dB change in the SNR. In evaluating
the power or SNR requirements of a modulation technique,
at high bit rates one must consider the impact of multipath
ISI, as well as any reduction of the ISI that can be achieved
through equalization techniques.
It can be difficult to achieve flat frequency response
and low noise over a wide bandwidth using large-area
photodiodes. Thus the second most important attribute of
a modulation technique is the receiver electrical bandwidth
Fig. 15. Common binary modulation schemes. Transmitted waveforms of (a) OOK with NRZ
pulse, (b) OOK with RZ pulse of duty cycle = 0:5, (c) 2-PPM, and (d) BPSK subcarrier with
subcarrier frequency f0 = 1=T . Corresponding power spectra are shown in (e). In all cases, Pt is
the average transmitted optical power and T is the symbol interval (equal to the bit interval).
it requires.24 For all baseband and single-subcarrier modulation schemes, we define this bandwidth requirement
as the span from dc to the first null in
, the PSD
of the transmitted waveform
. For multiple-subcarrier
schemes, the bandwidth requirement is the span from dc
to the first null in the highest-frequency subcarrier. We will
present bandwidth requirements
, i.e., normalized to
the bit rate . In Fig. 15(a)–(d), we present the waveforms
of four common binary modulation techniques: (a) OOK
with NRZ pulses, (b) OOK with RZ pulses of duty cycle
, (c) 2-PPM, and (d) a BPSK subcarrier at a
frequency equal to the reciprocal of the symbol duration.
24 We note that the electrical bandwidth requirement of a modulation
technique has little bearing on the optical bandwidth occupied by an IM
signal. This optical bandwidth is dominated by the large spectral spread
of practical infrared sources. For example, a 1-nm width corresponds to
469 GHz, assuming a wavelength of 800 nm.
A transmission is of the form:
encodes the th information bit and
. The PSD’s of these four modulations are shown
in Fig. 15(e). Based on our definition, OOK with NRZ has a
normalized bandwidth requirement of unity, while the other
three modulations have twice this normalized bandwidth
Table 3 and Fig. 16 present a comparison of the power
and bandwidth requirements of a number of binary and
nonbinary modulation schemes, assuming an ideal channel
and AWGN [10], [43]. These modulations include OOK,
PPM, differential PPM (DPPM) and multiple-subcarrier
Fig. 16. Theoretical comparison of average power efficiency and bandwidth efficiency of several
modulation schemes on nondistorting channels with IM/DD and AWGN. Techniques include OOK
with NRZ or RZ pulses, PPM with soft and hard decisions, differential PPM, BPSK subcarriers,
and QPSK subcarriers [10], [25], [43].
Table 3 Comparison of Normalized Average Power and
Normalized Bandwidth Requirements of Several Modulation
Schemes on Ideal Distortionless Channels with AWGN. Some of
the Expressions Here Represent Approximations That Are Valid
at High SNR. OOK with NRZ Pulses Has Power and Bandwidth
Requirements of 0 dB and Unity, Respectively [10], [43]
OOK, RZ, duty
cycle Normalized Average Power
Requirement (Optical dB)
5 log10 1
1:5 + 5 log
10 N
(soft decisions)
(hard decisions)
(hard decisions)
1:5 + 5 log
10 N
05 log10( L log2
L log L
05 log10 ( 4
05 log10[ L log8
log L
2 log L
modulation (MSM) using either binary phase-shift keying
(BPSK) or quaternary phase-shift keying (QPSK). In what
follows, we discuss the performance of these modulation
techniques, including their performance in the presence of
multipath ISI and fluorescent-light noise.
A. On–Off Keying
Among all modulation techniques suitable for wireless
infrared links, OOK is the simplest to implement. The
waveforms of OOK using NRZ pulses and using RZ
pulses of duty cycle
are shown in Fig. 15(a) and
(b), respectively. If the channel is distortionless, the ideal
maximum-likelihood (ML) receiver for OOK in AWGN
consists of a continuous-time filter matched to the transmitted pulse shape, followed by a sampler and threshold
detector set midway between the “low” and “high” pulse
amplitudes. Referring to Fig. 16, we see that OOK with
NRZ pulses represents a good compromise between the
power requirement and bandwidth requirement. The use
of RZ pulses having a duty cycle
. However,
the bandwidth requirement by a factor of
it decreases the average power requirement, because the
increased noise associated with this expanded bandwidth is
increase in peak optical power. For
outweighed by the
this reason, OOK with RZ pulses is utilized in a number of
current infrared systems (see Section I-D). However, when
the pulse duty cycle is made sufficiently small, we will see
that it becomes more efficient to encode information in its
position, i.e., to employ PPM.
The performance of OOK on multipath channels without equalization can be computed using the enumeration
technique developed in [9]. Fig. 17(a) shows the theoretical
power requirements of unequalized OOK (and PPM) links
at 10 and 30 Mb/s, computed on an ensemble of about 100
measured multipath channels, which include nondirectedLOS and diffuse configurations, without and with shadowing [17]. We see that the optical power requirement
essentially depends only on the normalized delay spread
(the delay spread normalized to the bit duration), regardless
of the link configuration or the particular time dependence
. On the worst chanof the channel impulse response
Fig. 17. Theoretical performance of OOK and 2-, 4-, 8-, and 16-PPM at 10 and 30 Mb/s on
measured multipath channels using: (a) unequalized receivers, which are optimal only on the ideal
channel and (b) optimal MLSD. The dashed lines indicate performance on the ideal channel, and
the solid lines indicate performance on model channels having an impulse response of the form
(t + a)07 u(t), where a is a parameter governing the delay spread [34], [35].
nels at 30 Mb/s (normalized delay spread of about 0.35),
the normalized power requirement is about 6 dB. While
relatively high, the power requirement may be acceptable
for some applications, particularly when transmitting from
a base station. Results for 100 Mb/s are presented in
[17], where it is found that unequalized OOK incurs very
large power penalties, and develops an irreducible BER
for normalized delay spreads above about 0.60, implying
that on multipath channels, unequalized OOK reception is
not feasible at this bit rate. The solid lines in Fig. 17(a)
represent the power requirement computed using the model
channel impulse response (26), for different values of the
parameter , which governs the channel delay spread. This
model channel reproduces the ISI penalties of real channels
remarkably well for all of the modulation/demodulation
techniques considered in Fig. 17(a) and (b).
In the presence of multipath ISI, the optimum receiver for
OOK employs a whitened matched filter (WMF)25 frontend and performs maximum-likelihood sequence detection
(MLSD), which can be implemented efficiently using the
25 The WMF is the optimum front-end for all pulse-amplitude modulation schemes in the presence of ISI and AWGN. It consists of a
continuous-time filter matched to the received pulse shape, followed by a
sampler and a discrete-time noise-whitening filter [47].
Viterbi algorithm (VA) [45], [47]. Fig. 17(b) shows the
theoretical power requirements of 10- and 30-Mb/s OOK
(and PPM) links using optimum MLSD, computed on an
ensemble of measured channels. We see that with OOK,
MLSD is extremely effective in mitigating ISI, reducing
the normalized power requirements at 30 Mb/s to less than
2 dB. A practical, though suboptimal, means to mitigate
the multipath ISI penalty is by using a decision-feedback
equalizer (DFE) [45], [47], which can adapt automatically
to the channel impulse response.26 The power requirements
of OOK links using DFE at 10, 30, and 100 Mb/s have been
evaluated in [17]. At 10 and 30 Mb/s, power requirements
with DFE are very close to those with MLSD, which
are shown in Fig. 17(b). At 100 Mb/s, even on the worst
channels, no irreducible BER is observed with a DFE, in
contrast to the unequalized case, though the normalized
power requirements can reach 7.1 dB for shadowed diffuse
channels and 9.1 dB for shadowed LOS channels. These
power requirements, while high, may be small enough to
make OOK with a DFE practical at 100 Mb/s, particularly
for transmission from a base station.
26 Implementation of the WMF and MLSD require knowledge of the
multipath channel impulse response, which is typically learned using an
adaptive equalizer, such as a DFE.
Fig. 18. Comparison of transmitted waveforms for (a) 4-PPM, and (b) 4-differential pulse-position
modulation (4-DPPM). Pt is the average transmitted power. In the case of 4-PPM, T represents
the symbol duration, while for 4-DPPM, T represents the average symbol duration. In this figure,
the two sets of waveforms have been adjusted to correspond to the same peak power and the
same average bit rate.
Practical DFE’s are implemented using digital or
discrete-time analog signal-processing techniques, and
usually employ finite-impulse-response filters. As described
in Section VII, an experimental 50-Mb/s diffuse link
employing OOK with DFE achieved performance close to
that predicted by theory [15], [40]. In numerical simulations
of a 100-Mb/s OOK link [19], a DFE was found to adapt
to a training sequence within about 200 b, i.e., about 2
s. Most indoor applications of nondirected infrared links
will not involve rapidly moving receivers, so that channel
impulse responses will change significantly only on time
scales of tens to hundreds of ms. Thus we expect that once
it is initially adapted to the channel response, a DFE should
be able to track easily any changes in that response.
B. Pulse-Position Modulation
PPM is an orthogonal modulation scheme that offers a
decrease in average-power requirement compared to OOK,
at the expense of an increased bandwidth requirement. PPM utilizes symbols consisting of time slots, which we
is transmitted
will refer to as chips. A constant power
during one of these chips and zero power is transmitted
during the remaining
chips, thereby encoding
bits in the position of the “high” chip. Waveforms of
2- and 4-PPM are shown in Fig. 15(c) and Fig. 18(a),
respectively. For a given bit rate, -PPM requires more
bandwidth than OOK by a factor
, e.g., 16-PPM
requires four times more bandwidth than OOK. In the
absence of multipath distortion, -PPM yields an averagepower requirement that decreases steadily with increasing
; the increased noise associated with a
wider receiver noise bandwidth is outweighed by the fold increase in peak power. This decreased average-power
requirement makes PPM especially suitable for portable
infrared transmitters. As shown in Section IV-D, PPM
can achieve much greater immunity than OOK to near-dc
noise from fluorescent lamps. For these reasons, PPM is
employed in several commercial wireless infrared systems
(see Section I-D). In addition to the increased bandwidth
requirement, two drawbacks of PPM, as compared to OOK,
are an increased transmitter peak-power requirement, and
the need for both chip- and symbol-level synchronization.
In the absence of multipath distortion, an optimum
ML receiver for -PPM employs a continuous-time filter
matched to one chip, whose output is sampled at the
chip rate. Each block of
samples is passed to a block
decoder, which makes a symbol decision, yielding
information bits. In soft-decision decoding, the samples
are unquantized,27 and the block decoder chooses the
largest of the samples. In hard-decision decoding, each
sample is quantized to “low” or “high” using a simple
threshold detector, and the block decoder makes a symbol
decision based on which sample is “high,” mediating
in cases where no sample, or more than one sample, is
“high.” While hard decoding is easier to implement, and
is thus used in most commercial implementations, it incurs
approximately a 1.5-dB optical power penalty with respect
to soft decoding [25], [43].28 Table 3 and Fig. 16 present
the power and bandwidth requirements of PPM with soft
and hard decoding. For example, 16-PPM has normalized
power requirements of 7.5 dB and 6.0 dB with soft and
hard decoding, respectively.
When -PPM is transmitted over multipath channels, the
nonzero transmitted chips can induce interference in chips
both within the same symbol (intrasymbol interference) and
in adjacent transmitted symbols (intersymbol interference);
we will refer to these effects collectively as ISI. Fig. 17(a)
presents the power requirements of systems employing no
equalization [34], i.e., using the same receiver filter and
soft-decision decoder that is optimal on a distortionless
channel. For small values of the normalized delay spread
(i.e., small delay spreads and/or low bit rates), the power
requirement is still reduced by employing larger , and 1627 In a digital implementation of the detector, these samples would be
quantized. A resolution of the order of 4–6 b is typically sufficient to
achieve near-ideal performance.
28 This corresponds to a 3-dB SNR penalty, which is fairly typical of
binary block codes [45].
PPM yields the best power efficiency among techniques
studied. As the normalized delay spread increases, the
power requirements increase more rapidly for PPM than
for OOK, and increase most rapidly for large , because
of the short chip duration. For the worst channels shown in
Fig. 17(a), 4-, 8-, and 16-PPM incur an irreducible BER,
unlike OOK and 2-PPM.
When multipath ISI is present, the optimum PPM receiver
employs a chip-rate WMF, followed by a MLSD, which can
be implemented using a symbol-rate VA [10]. The power
requirements of PPM systems using MLSD on multipath
channels [34] are shown in Fig. 17(b). The use of MLSD
significantly improves the performance of PPM, preventing
irreducible BER’s and restoring the average-power gain
obtained by increasing the value of . However, as the
normalized delay spread increases, even when this optimum
detection technique is employed, the power requirements of
PPM increase much more rapidly than those of OOK, and
on the worst channels, 16-PPM offers only about a 1-dB
power advantage over OOK.
There exist several reduced-complexity, suboptimal adaptive equalization techniques for PPM [10], which include
linear and DFE’s operating at either the chip or symbol
rates, as well as hybrid DFE’s that make use of tentative chip decisions to cancel intrasymbol interference, but
use more reliable symbol decisions to cancel intersymbol
interference. At 10 and 30 Mb/s on measured multipath
channels, their performance is close to the optimum MLSD
The performance of PPM on multipath channels can be
improved [36] by using trellis-coded modulation (TCM)
[57], which is designed to maximize the minimum Euclidean distance between allowed signal sequences. A key
concept of TCM is the partitioning of the signal set into subsets having unequal minimum Euclidean distances. Since
PPM is an orthogonal modulation scheme, on a distortionless channel, the Euclidean distance between any two
symbols is the same, and there is no advantage in performing set partitioning. However, multipath distortion
causes the Euclidean distances between PPM symbols to
become unequal. Partitioning of the 8-PPM symbol set
is illustrated in Fig. 19(a), and Fig. 19(b) shows how the
minimum Euclidean distances within the subsets29 depend
on normalized channel delay spread. As the delay spread
increases, the distances
between symbols with adjacent
chips falls quite rapidly, whereas the distances,
between symbols with chips separated by two or four chip
periods, respectively, decrease less rapidly. As the delay
spread increases, trellis codes designed with proper symbol
mapping and set partitioning exploit these differences in
distance to improve the system performance as compared
to that of uncoded PPM. Reference [36] describes a search
for optimal rate-2/3 codes for 8-PPM and rate-3/4 codes
for 16-PPM. In Fig. 20, the performance of these codes
29 More precisely, we compute the minimum distance between all
possible symbol sequences ending with a symbol drawn from the subset.
is compared to that of uncoded 16- and 32-PPM.30 The
parameter is the constraint length of the code, and the
receiver is assumed to employ optimal MSLD on a trellis
of combined code and ISI states. We note that as the delay
spread increases, trellis-coded PPM exhibits a much more
gradual increase in power requirement than uncoded PPM.
Trellis-coded 16-PPM is the most power-efficient technique
known for use on multipath channels.
Differential PPM (DPPM) [43] is a simple modification
of PPM that can achieve improved power and/or bandwidth
efficiency in applications where low cost dictates the use of
hard-decision detection and multipath ISI is minimal (e.g.,
at low bit rates or in directed-LOS links). The 4-PPM and
4-DPPM signal sets are compared in Fig. 18(a) and (b),
respectively. Each 4-PPM symbol consist of four chips, of
which one is “high” and three are “low.” The 4-DPPM
symbols omit the “low” chips that follow the “high” chip,
and hence have unequal durations. Because of this variable
symbol duration, symbol boundaries are not known prior
to detection, and optimal soft decoding of DPPM requires
use of MLSD, even in the absence of coding or ISI. If hard
decoding is used, DPPM is easier to decode than PPM, since
the former requires no symbol-level timing recovery. The
power and bandwidth requirements of DPPM are shown in
Fig. 16. Using hard decoding, for a fixed , DPPM requires
much less bandwidth and slightly more power than PPM,
mainly because the former has a higher duty cycle. If we
compare 16-DPPM to 4-PPM, however, we see that 16DPPM requires 6% more bandwidth, but requires 3.1 dB
less optical average power. Because of the variable symbol
duration, DPPM offers a nonuniform throughput,31 which
may be problematic for some applications. Reference [43]
describes several techniques for mitigating this throughput
variation. The performance of DPPM in the presence of
multipath ISI is currently under investigation.
C. Subcarrier Modulation
In single-subcarrier modulation (SSM) [10], [33], a bit
stream is modulated onto a radio-frequency subcarrier, and
this modulated subcarrier is modulated onto
, the
instantaneous power of the infrared transmitter. Because
the subcarrier is typically a sinusoid taking on negative
or positive values, a dc bias must be added to it to
satisfy the requirement that
be nonnegative. The
transmitted waveforms and PSD of a BPSK subcarrier
are shown in Fig. 15(d) and (e), respectively, assuming a
subcarrier frequency equal to the bit rate. Such a BPSK
subcarrier requires twice the bandwidth of an OOK signal,
while a QPSK subcarrier requires the same bandwidth as
OOK. After optical-to-electrical conversion at the receiver,
the subcarrier can be demodulated and detected using a
30 Note that rate-2/3 coded 8-PPM has the same bandwidth requirement
as uncoded 16-PPM, while rate-3/4-coded 16-PPM requires slightly less
bandwidth than uncoded 32-PPM.
31 To provide an example of nonuniform throughput, assume that the
4-DPPM symbols s0 (t), s1 (t), s2 (t), s3 (t) are used to encode the bit
pairs {0, 0}, {0, 1}, {1, 0}, {1, 1}, respectively. Referring to Fig. 18(b),
we see that transmission of the bit sequence {0, 0, 0, 0, 0, 0} will require
a time 1.2 T , while {1, 1, 1, 1, 1, 1} will require 4.8 T .
Fig. 19. (a) Partitioning the 8-PPM signal set. (b) Minimum Euclidean distance between symbols
drawn from subsets. In the absence of multipath distortion (zero delay spread), the minimum intraset
distances are the same in all three sets. As the channel delay spread increases, the minimum intraset
distance decreases most rapidly in the set 0 , and least rapidly in the subsets 2 [36].
standard BPSK or QPSK receiver [45]. On distortionless
channels, a single BPSK or QPSK subcarrier requires 1.5
dB more optical power than OOK. This can be explained
by observing that BPSK and OOK transmissions of average
power are equivalent to binary antipodal signals plus a dc
carrying no information, but the BPSK waveform
uses sinusoidal pulses having 3-dB less electrical power,
thus requiring 1.5 dB more optical power for achievement
of the same receiver SNR.
Multiple-subcarrier modulation (MSM) [10], [33] makes
it possible to perform frequency-division multiplexing,
while maintaining the simplicity of IM/DD. In MSM,
several independent bit streams are modulated onto
subcarriers at different frequencies, and their frequencydivision-multiplexed sum is used to modulate the intensity
of an optical transmitter. At the receiver, the individual
bit streams can be recovered using multiple bandpass
-subcarrier transmission requires
demodulators. An
dB more optical power than the corresponding
single-subcarrier scheme, because the amplitude of each
, to insure that
subcarrier must not exceed
is nonnegative. The power and bandwidth requirements
of SSM and MSM systems using BPSK and QPSK on
distortionless channels are summarized in Table 3 and
Fig. 16.
When MSM is transmitted over a multipath channel,
several effects further degrade its power efficiency, as
compared to transmission of OOK over an ideal channel
[33]. As multipath channels are generally lowpass in nature,
subcarriers are subject to an attenuation that generally
increases with increasing subcarrier frequency. In addition,
subcarriers may be subject to ISI, interference between
in-phase and quadrature phases of one subcarrier, and
interference between adjacent subcarriers that may overlap
partially in frequency. To reduce these three interferences,
it is desirable to use a large number of subcarriers, but this
leads to an excessive penalty for large , as described
above. We have evaluated the performance of a large
number of different MSM schemes, for transmission at
total bit rates of 30 and 100 Mb/s over measured multipath
channels [33], and have found that the best MSM schemes
require several decibels more optical power than OOK with
While less power-efficient than OOK or -PPM, MSM
may be well suited for transmission of multiplexed bit
streams from a base station to a collection of several
portable receivers. Through simultaneous transmission of
several narrowband subcarriers, MSM can enable very high
aggregate bit rates without requiring adaptive equalization
to overcome ISI, and may allow individual receivers to
Fig. 20. Theoretical performance of rate-2/3-coded 8-PPM, rate-3/4-coded 16-PPM, and uncoded
16- and 32-PPM. represents the constraint length of the code. Receivers employ optimal
maximum-likelihood sequence detection on a trellis of combined code and intersymbol-interference
states. Multipath distortion is modeled by considering an impulse response of the form (t+a)07 u(t),
where a is a parameter governing the delay spread [36].
process only a subset of the total transmission. Furthermore,
SSM and MSM can achieve much greater immunity than
OOK to near-dc noise from fluorescent lamps, as shown in
Section IV-D.
D. Effect of Fluorescent-Light Noise
As mentioned in Section II-D, electronic-ballast fluorescent lamps are potentially much more detrimental to
infrared links than their conventional-ballast counterparts.
Reference [38] studies the effects of the infrared emission
from 22- and 45-kHz electronic-ballast lamps upon various
modulation techniques, by computing the BER and optical
power requirements in the presence of the nearly periodic fluorescent-light waveform. As an example, Fig. 21
shows the normalized average optical power requirements
of 10-Mb/s links in the presence of 22-kHz emissions,
as a function of
, the optical power received from the
fluorescent lamp after optical filtering. This figure considers
a single BPSK subcarrier, OOK, and PPM with softdecision decoding. For the latter two modulation schemes,
we consider both receivers employing no highpass filtering
and those employing first-order highpass filters that induce
sufficient ISI to cause a 2.0-dB SNR penalty, corresponding
approximately to the highest cut-on frequency one might
employ in the absence of line coding or active baseline
In principle, subcarrier modulation can be made immune
to fluorescent-light noise, by choosing the subcarrier frequency high enough. Hence, a BPSK subcarrier has an
optical power requirement that is independent of . When
is sufficiently large, OOK suffers a significant penalty
from fluorescent-light noise. First-order highpass filtering
is not able to reduce the penalty, because the filter cuton frequency cannot exceed about 0.004 , which is too
low to block most of the fluorescent-light noise.32 For a
PPM suffers much smaller penalties than
given value of
OOK. This can be understood by observing that the OOK
slicer is sensitive to the actual values of the fluorescent-light
samples, whereas the PPM “choose largest” soft decoder is
affected only by the differences among blocks of samples.
For this reason, PPM is less susceptible to fluorescent-light
noise at high bit rates, where the block duration is shorter.
However, if a PPM receiver employs hard decoding, it
loses this “natural” immunity to fluorescent-light noise. We
note that simple highpass filtering is effective in countering
fluorescent-light noise with PPM, because the continuous
component of the -PPM PSD goes to zero at dc [see, e.g.,
Fig. 15(e)], so the filter cut-on frequency can be as high as
without inducing excessive ISI.
An angle-diversity receiver utilizes multiple receiving
elements that are oriented in different directions. It can be
used in place of a single-element receiver in either LOS
32 As shown in Section VII, the filter cut-on frequency can be raised
significantly if active baseline restoration is employed.
Fig. 21. Theoretical performance comparison of various modulation schemes at 10 Mb/s in the
presence of noise from a fluorescent lamp driven by a 22-kHz electronic ballast. Calculations
consider two alternate receiver designs: 1) no highpass filter; 2) first-order highpass filter whose
cutoff frequency is chosen so that in the absence of fluorescent lighting, the filter induces a 2-dB
electrical SNR penalty due to intersymbol interference. The ratio Pf =P0 represents the maximum
absolute excursion (with respect to the mean) of the received fluorescent optical power waveform,
divided by the average optical power required to achieve 1009 BER with OOK in the absence
of fluorescent lighting [38].
or non-LOS links. In a conventional angle-diversity receiver, each receiving element utilizes its own nonimaging
concentrator, such as a CPC or a hemisphere. A CPCbased realization is illustrated in Fig. 22(a), and the angledependent effective area of its various elements is shown
in Fig. 22(c). A principal advantage of angle-diversity
reception is that it allows the receiver to achieve high
optical gain and a wide FOV simultaneously, circumventing
the gain-FOV trade-off implied in (8). Moreover, an anglediversity receiver can reduce the impact of ambient light
noise, co-channel interference and multipath distortion, in
part by exploiting the fact that they are often received from
different directions than the desired signal. The advantages
achieved by angle-diversity reception depend on how the
signals received in the different elements are processed and
detected, as summarized in Table 4.
When multipath distortion is significant, the optimum
reception technique is maximum-likelihood combining
(MLC). Assuming that no cochannel interference or
fluorescent-light noise is present, a -element anglediversity receiver yields receptions of the form:
is the channel between the transmitter and
the th receiver and the
are mutually independent
white, Gaussian noises. In MLC, each
is processed
by a separate continuous-time matched filter
. The
matched-filter outputs are sampled and combined in
memoryless fashion, with the th sample weighted in
. This
inverse proportion to the PSD of the noise
sample sequences is a sufficient
weighted sum of the
statistic. Using this sufficient statistic, the receiver performs
MLSD, which can be implemented using the VA. We note
that implementation of MRC requires separate estimation of
each of the channel impulse responses and noise PSD’s.
The complexity of MLC is likely to be too high for
many applications, and a number of simpler approaches
are possible. An easily implemented class of techniques
in a memoryless, linear cominvolves combining the
biner, filtering the weighted sum using a single continuoustime filter, and sampling the filter output. The resulting
sample sequence can be processed using MLSD, DFE,
or a simple slicer to yield the detected data. This class
of techniques includes maximal-ratio combining (MRC),
selection diversity (SD), and equal-gain combining (EGC),
which differ according to how the combining weights are
, are summed together
In MRC, the
with weights proportional to the signal current to noisePSD ratios, thereby maximizing the SNR of the weighted
sum. When multipath distortion is not significant, the
optimum MLC reduces to MRC followed by a simple slicer.
Fig. 22. Angle-diversity receivers using (a) CPC’s and (b) an imaging lens and a segmented
photodetector. Effective light-collection areas (schematic) of (c) CPC-based receiver and (d) imaging
Table 4
Combining Techniques for Angle-Diversity Receivers
Achieves high optical gain and wide field of view simultaneously.
Selection Diversity
Mitigates noise and cochannel interference.
Mitigates multipath distortion.
Can employ a dingle preamplifier
Under some circumstances.
Avoids need for channel and noise estimation.
Simulations have shown that in diffuse links at low bit
rates (where multipath distortion is not a concern), anglediversity detection with MRC can reduce transmitter optical
power requirements by 4–6 dB [28]. MRC can result in
a net decrease in multipath distortion, as compared to a
single, wide-FOV receiver, as long as both the ambient light
noise and multipath reflections both arrive from directions
sufficiently far away the strong signal components. If this
is not the case, however, an increase in multipath distortion
could result. MRC requires estimation of the SNR’s in the
each of the receiving elements, representing an increase in
complexity over nondiversity reception.
In SD, only the signal having the best SNR is utilized.
This technique can often separate the signal from ambientlight noise, resulting in a SNR improvement, but the gains
are not as large as those achieved using MRC. For example,
[14] found that in diffuse links employing angle-diversity
receivers, SD requires 1–2 dB more optical power than
MRC. SD can yield a significant reduction in multipath
distortion, provided that directional receiving elements are
employed, making it a promising technique for high-bit-rate
systems. Simulations have shown that substantial reductions
in multipath delay spread occur when each element has a
FOV (semi-angle) smaller than 50 [26]. SD is probably
not much simpler to implement than MRC, since it still
requires SNR estimation.
In EGC, the multiple signals are summed together with
equal weights. This technique increases the receiver FOV,
but is unable to separate signal from noise or cochannel
interference. Moreover, it can result in an increase of
multipath distortion, making it unsuitable for very highbit-rate links. It is attractive in its simplicity, as it avoids
the need for SNR estimation, and the signals from several
photodetectors can be processed by a single preamplifier,
provided that the total receiver input capacitance is acceptably small. EGC is used in some diffuse systems at bit rates
up to 4 Mb/s, helping them achieve very robust operation,
even in the face of shadowing [32].
A promising means to implement angle-diversity reception is by using an imaging diversity receiver [8], [31].
This consists of an imaging optical concentrator that has a
segmented photodetector array placed at its focal plane, as
shown in Fig. 22(b). Each detector output is preamplified
separately, and the resulting signals can be processed using either SD, MRC, or MLC.33 Fig. 22(d) schematically
illustrates the angle-dependent effective area of an imaging
diversity receiver.34 An imaging diversity receivers has two
major advantages over a nonimaging angle-diversity receiver. First, an imaging receiver employs a planar detector
array, which can be monolithically fabricated, enabling the
use of a large number of detector segments. Second, the
imaging receiver employs one concentrator, regardless of
how many detector segments are used. Thus the imaging
receiver can use more receiving elements, leading to more
effective separation of the signal from unwanted noise and
interference. The main disadvantage of the imaging receiver
is that it cannot collect light over as wide a FOV as a
nonimaging angle-diversity receiver, since the elements of
the latter can be oriented in any direction.
The power efficiency of nondirected-non-LOS links can
be improved significantly if the diffuse transmitter is replaced by one that emits a number of relatively narrow
beams that illuminate a lattice of spots on the ceiling [8],
[31]. Each beam should have a divergence angle large
enough to make it eye-safe, but small enough that it
does not spread excessively when traversing a room. A
divergence semiangle of about 2 might be typical. This
transmitter design, which we call quasi-diffuse, is illustrated
in Fig. 23(a). If the signal spots form a regular lattice on
the ceiling such that at least one lies within the receiver
FOV, then the number of beams required to cover a circle
of radius
is approximately proportional to
. The
required transmitter power required is proportional to the
same factor, i.e.,
. Assuming that the beams do not
spread much, the propagation loss from the transmitter to
the receiver is independent of their horizontal separation ,
and depends only on the spot-receiver horizontal separation
, i.e.,
. Hence the channel gain is proportional to
the inverse square of distance,
, as opposed to
the fourth-power dependence characteristic of diffuse links
[see (12)].35
The quasi-diffuse transmitter design can be combined
with either a single-element or angle-diversity receiver to
33 There is probably no advantage to using an imaging diversity receiver
with EGC, because it is unlikely that an imaging concentrator can achieve
simultaneously a higher gain and wider overall FOV than a single
nonimaging concentrator.
34 It should be noted that the “dips” in the angle-dependent effective area
shown in Fig. 22(d) do not correspond to angles at which the receiver has
a significantly smaller overall effective area; rather, they occur when the
reception is divided between two or more photodetector segments.
35 In order to achieve protection against shadowing, it may be desirable
that at least two spots lie within the receiver FOV, which would require
at least a doubling of the transmitter power Pt .
Fig. 23. (a) Nondirected-nonline-of-sight link employing a
quasi-diffuse transmitter. (b) Comparison of three nondirected-nonline-of-sight link designs: average transmitter power
09 BER using OOK at 100 Mb/s. Shot
required to achieve
noise induced by ambient light is the dominant noise source. In
all cases, the transmitter and receiver are located 3 m below the
ceiling, which has 100% diffuse reflectivity, and the receiver has a
2 and
, an entrance area A
FOV (semiangle) s
: , and it employs a 50-nm-wide optical
a refractive index n
bandpass filter. The quasi-diffuse transmitter employs a number
of beams that ranges between 1 and 11 as the cell radius varies
between 2 and 10 m [31].
9 = 45
= 9 4 cm
significantly improve the power efficiency of nondirectednon-LOS links. Fig. 23(b) illustrates the transmitter power
requirements of 100-Mb/s OOK links [31]. Shot noise
induced by ambient light sources is assumed to be the
dominant noise source. When a single-element receiver
is used, the quasi-diffuse transmitter requires as much as
about 10 dB less power than the diffuse transmitter, simply
because the former achieves a much higher channel dc gain.
If the quasi-diffuse transmitter is employed, exchanging the
single-element receiver for a 100-element imaging diversity
receiver yields another 10-dB power reduction, because
the imaging receiver receives far less ambient light noise
than its single-element counterpart.36 We note that if these
links employed PPM instead of OOK, all the transmitter
power requirements shown in Fig. 23(b) could be reduced
by several dB (see Section IV).
36 The power requirements for the 100-element imaging receiver shown
in Fig. 23(b) assume that the signal spot falls entirely upon a single
photodetector segment, in which case MRC and SD are equivalent. In
the worst case, when the spot is incident equally upon three detectors
(assuming hexagonal detector segments), and assuming the use of MRC,
power requirements for the 100-element imaging receiver are increased
from the values shown in Fig. 23(b) by 2.4 dB.
Table 5
Techniques for Multiplexing Transmissions (Shading Denotes an Advantage)
Wavelength Division
Necessary Loss of
Per-User Capacity
Space Division Multiplexing with
Angle Diversity Receiver
Time Division
Code Division
Subcarrier Frequency Division
An angle-diversity receiver (of either nonimaging or
imaging design) might be used to minimize cochannel
interference at a LOS hub, an application that was mentioned in Section I-D. This application is discussed further
in Section VI-A, where it is presented as an example of
space-division multiple access.
In this section, we discuss how multiple users can share
the infrared medium, placing emphasis on physical-layer
multiplexing techniques, as opposed to medium-access protocols or higher network layers. The medium-access protocols used in some current infrared systems were discussed
in Section I-D.
Infrared and radio physical media differ in a number of
ways that have significant implications for multiple-access
1) Infrared cannot pass through walls or other opaque
barriers, making it possible to reuse the same infrared
bandwidth in each room of a building.
2) As explained in Section I-C, typically it is not practical to perform homodyne or heterodyne detection of
the infrared carrier, and DD must be employed. If a
DD system employs multiple infrared carriers, these
must be separated by optical filtering prior to DD.
3) The short wavelength of infrared makes it possible to
achieve high angular resolution in an angle-diversity
receiver, e.g., by using imaging optics, as discussed
in Section V. This may make it possible to perform
space-division multiplexing.
4) Achieving a high SNR is usually the greatest challenge in infrared system design. Achievement of
high average-power efficiency with IM favors the
use of waveforms having a short duty cycle (see
Section IV). This tends to favor the use of timedivision multiple access (TDMA) over other electrical
multiplexing techniques.
5) The SNR of a DD receiver is proportional to the
square of the received optical power [e.g., see (5)].
Because of this, even moderately strong co-channel
interference may be “buried” in the thermal and
shot noises. It is well known that the capacity of
a multiuser system employing bandwidth reuse is
maximized when cochannel interference dominates
over noise. With infrared systems, however, it is
usually not possible to increase the power of all
Optical Average
Power Efficiency
Permits Simultaneous
transmitters or decrease the noises until the system
becomes interference-limited.
Interference is modeled differently in infrared systems
than in radio systems. Consider
simultaneous IM transmissions
, which are incident upon a
DD receiver. Let
denote the impulse response of the
channel between transmitter and the receiver. Then the
total received photocurrent
is given by
We emphasize that
is linear in each of the IM
, and that there is no need to consider
the relative phases of the underlying optical carriers. The
derivation of (29) is a simple generalization of the derivation of (1), which is provided in [17].
We classify techniques for multiplexing transmissions
into two categories: optical and electrical. Some of their
characteristics are summarized in Table 5.
A. Optical Multiplexing Techniques
Optical multiplexing techniques may permit different
users to transmit simultaneously37 within the same space,
without requiring a loss of per-user capacity (at least in
In wavelength-division multiple access (WDMA),38 different users transmit at different infrared wavelengths using
narrow-spectrum sources (e.g., LD’s), and a receiver employs a bandpass optical filter to select the desired channel prior to DD. Wavelength-tunable optical transmitters
(e.g., tunable LD’s) are currently expensive, and fairly
sophisticated techniques are required to tune them to a
precisely defined wavelength. Large-area, tunable bandpass
filters (e.g., single- or multiple-stage Fabry–Perot filters)
are costly and, while they could be made in planar form,
would be difficult to fabricate in hemispherical form for
wide-FOV receivers. Therefore, if a transceiver is to have
multiwavelength capability, it must probably include multiple transmitters and receivers, which are likely to be too
costly for many applications.
There exist some viable applications of WDMA in which
each transceiver needs to transmit at only one wavelength,
37 That
is, using the same time slot, code, or subcarrier frequency.
order to simplify terminology, multiplexing techniques will include
the suffix “multiple access” even when they are not used to provide users
with random access to the medium.
38 In
and receive at only one wavelength (these two wavelengths may or may not be the same). When used in this
manner, WDMA can enable simultaneous, uncoordinated
operation of heterogeneous links within the same space.
For example, WDMA could enable simultaneous operation
of a remote-control device and a data-communication link
without interference. As another example, consider a system
employing infrared links between portable terminals and a
wired backbone network, as shown in Fig. 3(d). WDMA
could be used for duplexing the uplinks and downlinks,
i.e., all uplinks would employ wavelength
, while all
downlinks would use a second wavelength, . The main
drawback of this approach is that it would not permit direct
communication between the portables, unless each one were
equipped with a second transmitter and/or receiver.
Space-division multiple access (SDMA) involves the use
of an angle-diversity receiver (see Section V) to separate
signals that are received from different directions. The
angle-diversity receiver can be of an imaging or nonimaging design and may employ either SD or MRC. As an
example of SDMA, consider a hub capable of establishing
simultaneous, independent LOS links with several portable
transceivers, as shown in Fig. 3(c). The hub can employ
an angle-diversity receiver to minimize co-channel interference between different inbound receptions. In another
potential application of SDMA, one might construct a
multiple-access LAN using quasi-diffuse transmitters and
an imaging angle-diversity receivers, like those shown in
Fig. 23(a). Suppose that two quasi-diffuse transmitters are
located in close proximity and transmit simultaneously,
each illuminating a lattice of spots on the ceiling. If an
imaging receiver “sees” a desired signal spot that is not
overlapped by a spot from another transmitter, it may be
able to detect the desired signal with acceptably small cochannel interference. Assuming that some or all of the
transmitters are mobile, as the number of users increases,
spot overlap becomes more probable. Hence, this form
of SDMA might need to be combined with an electrical
multiplexing technique to achieve robust operation.39
B. Electrical Multiplexing Techniques
When different users share the same optical channel,
electrical multiplexing techniques can enable reliable transmission, but they necessarily entail a loss of per-user
capacity (just as in radio systems). The first of these techniques is TDMA, whereby users transmit in nonoverlapping
time slots. TDMA has high average-power efficiency, since
the transmissions are of low duty cycle, but requires some
form of coordination to insure that transmissions do not
coincide in time. Since achieving high power efficiency
is of paramount importance for portable transmitters, it
is not surprising that all current infrared multiple-access
LAN’s employ some variant of TDMA (see Section I39 It is interesting to note the analogy between this form of SDMA and
CDMA. In this context, the various ceiling-spot arrays can be viewed
as quasiorthogonal spatial signals, and an imaging receiver configured to
receive one spot array corresponds to a spatial matched filter.
Fig. 24. (a) Fixed assignment of channels in a cellular system
having a reuse factor of three. (b) Theoretical performance comparison of six fixed reuse schemes, assuming hexagonal cells and
a reuse factor of three. Diffuse transmission is employed, and the
throughput in each cell is 10 Mb/s. A BER of 1009 is achieved
when the receiver is placed at the worst-case location within the cell
(the cell corner). The factor is equal to the SNR for unit optical
path gain, and is proportional to the square of the transmitted
optical power. A 10-dB change in corresponds to a 5-dB change
in the transmitter optical power [39].
D).40 In code-division multiple access (CDMA), different
users employ different orthogonal or quasiorthogonal code
sequences, permitting them to transmit simultaneously.
The power efficiency achieved by CDMA varies, depending upon the duty cycle of the transmitted waveforms.
In subcarrier frequency-division multiple access (FDMA),
different users can transmit simultaneously at different
subcarrier frequencies. The power efficiency achieved by
FDMA is poor and worsens as the number of subcarriers
increases (see Section IV-C)
We will compare the performance of TDMA, CDMA, and
FDMA for infrared within the context of a cellular system
with spatial reuse [39]. Specifically, we consider a system
that uses diffuse infrared links to access a wired backbone
network, like that shown in Fig. 3(d). We will consider
the downlinks only, assuming that the uplinks employ a
different wavelength or otherwise avoid interfering with
the downlinks. Suppose that we wish to provide seamless
downlink coverage within a very large room, which requires
a very large number of diffuse transmitters. One option is
40 Recall that the SpectrixLite LAN employs CODIAC, a deterministic
TDMA protocol, while the IBM diffuse LAN employs CSMA/CA, which
can be considered a form of TDMA in which time blocks are randomly
utilized by various users.
Fig. 25.
Design of 50-Mb/s diffuse infrared link employing OOK with DFE [15], [40].
to employ unison broadcast, i.e., to have all transmitters
emit identical signals. This technique will lead to increased
multipath distortion [10] and requires the entire available
bandwidth to be shared by all users within a room.
Here, we consider the use of TDMA, CDMA, or FDMA
to divide the downlink bandwidth into several equal partitions; the number of partitions is called the reuse factor. The
room is divided into cells covered by different downlink
transmitters, and equal bandwidth is allocated to each transmitter.41 The same bandwidth can be reused by nonadjacent
transmitters. As an example, Fig. 24(a) shows the layout of
a system using a reuse factor of three. Receptions within
the central cell numbered “0” are subject to little or no
interference from adjacent nearby cells labeled “1” and “2,”
but are subject to cochannel interference from the six nearby
cells labeled “0,” which are separated from central cell by
the reuse distance.42 The signal-to-cochannel interference
ratio (SIR) depends upon the ratio of the reuse distance
to the cell radius, and also depends upon how the channel
dc gain
varies43 with the horizontal range . The
41 Higher capacity can be achieved by dynamically allocating bandwidth
to transmitters depending upon the number of portables each is presently
serving, though at the price of increased complexity.
42 In TDMA and FDMA systems, no interference is received, assuming
perfect synchronization and the absence of multipath distortion. In CDMA
systems, the adjacent cells “1” and “2” do interfere with cell “0.”
04 (see Section II43 Recall that in diffuse links at large d ; H (0)
reuse distance is proportional to the cell radius and to the
square root of the reuse factor. In order to increase the SIR,
one can either increase the cell radius or increase the reuse
factor. As the latter entails a loss of capacity, it is desirable
to use the smallest possible reuse factor.
Fig. 24 presents a comparison of the optical averagepower efficiency of several fixed cellular reuse schemes
[39], including: TDMA with 4-PPM, 2-PPM, and OOK;
FDMA with BPSK; and CDMA using -sequences [45],
and optical orthogonal codes (OOC’s) [58]. A reuse factor
of three is assumed. These calculations consider the gainversus-range relation of experimentally measured diffuse
channels. The throughput in each cell is 10 Mb/s, and a
10 BER is achieved when the receiver is placed at the
worst-case location within the cell (i.e., the cell corner). In
order to minimize the transmit power requirement, we wish
to minimize the required value of , which is equal to the
SNR for unit optical path gain, and is proportional to the
square of the transmitted optical power. A 10-dB change in
corresponds to a 5-dB change in the transmitter optical
In Fig. 24, we see that for cell radii smaller than 2
increases, because of
to 2.5 m, the required value of
cochannel interference, and at very small cell radii, it
BER (this does not
becomes impossible to achieve 10
occur with CDMA using -sequences, making this the only
viable scheme for these small cell radii). Of greater practical
interest, at cell radii larger than about 3 m, where the curves
for all modulation schemes become parallel, all modulation
schemes become noise-limited, i.e., the performance is
virtually the same as if there were no cochannel interference
present. In this regime, the lowest power requirement is
achieved by TDMA with 4-PPM, the lowest-duty-cycle
modulation considered, while FDMA with BPSK has the
highest power requirement. The comparison made here does
not consider the bandwidth requirement, nor the effect of
multipath distortion [39].
We have built and tested an experimental 50-Mb/s diffuse
infrared link using OOK [15], [40]. The main goals of
this project were 1) to explore the performance limits of
very high-bit-rate diffuse infrared links, and 2) to test the
performance of a wide-FOV, narrowband receiver based
on a hemispherical concentrator and hemispherical bandpass filter. The link design is depicted in Fig. 25. The
transmitter uses a cluster of eight LD’s whose output is
passed through a translucent plastic diffuser to create an
approximately Lambertian radiation pattern having 475mW average power at a wavelength of 806 nm. In typical
operation, the transmitter emission is directed upward toward the ceiling, creating a diffuse link configuration. In
order to create a receiver having large collection area, wide
FOV and narrow passband, a 1-cm silicon p-i-n detector
is index-matched to a hemispherical concentrator of 2cm radius, having a refractive index of 1.78. An optical
bandpass filter having a 30-nm bandwidth centered at 815
nm is bonded to the hemisphere’s outer surface. This filterconcentrator combination achieves a bandwidth of 30 nm,
a net gain
ranging from 0.5 to 1.75 dB, and
. The photodiode capacitance of 35 pF,
in conjunction with the preamplifier load resistance of 10
, leads to a 455-kHz pole that is compensated by a
passive R-C equalizer. The equalized receiver achieves
a 3-dB cutoff frequency of 25 MHz, which is limited
by the transit time of holes across the depletion region
of the photodiode, which is illuminated through the
contact. The preamplifier has an input-referred thermal
noise variance of
when a fivepole, 25-MHz Bessel lowpass receive filter is employed.
Residual interference from fluorescent lighting is removed
using a 1.6-MHz, single-pole highpass filter, and quantized
feedback through a 1.6-MHz, single-pole lowpass filter is
used to prevent baseline wander. In order to reduce the
impact of multipath ISI, the receiver employs a DFE, with
the forward and reverse filters realized using cable delays
and manually adjusted, variable-gain amplifiers. Both filters
have four taps; those of the forward filter are half-baudspaced, while those of the reverse filter are baud-spaced.
Fig. 26(a) presents the BER performance of the system
in the presence of various ambient lighting conditions.
The emission from fluorescent lamps driven by 22-kHz
electronic ballasts induces only about a 0.1-dB power
penalty, while bright skylight causes a penalty of about
Fig. 26. Performance of 50-Mb/s diffuse infrared link. (a) BER
versus received signal irradiance under various types of ambient
lighting. (b) Comparison between theoretical and measured optical
power gains achieved by DFE on various multipath channels. Solid
symbols denote shadowed link configurations [15], [40].
1.6 dB. The simple DFE is extremely effective in countering multipath ISI, yielding performance gains as high
as about 7 dB (optical power), corresponding to a 14-dB
SNR gain. Measured performance gains are in excellent
agreement with the predictions of theory, which can be
seen in Fig. 26(b). In the absence of shadowing and ambient
lighting, the link achieves a horizontal transmission range of
about 4.4 m (at
BER), while bright skylight reduces
this range to about 2.9 m.
As portable computers and communication terminals
become more powerful and are more widely deployed, the
demand for high-speed wireless communication is increasing. Infrared represents an attractive choice for many shortrange applications. Its advantages include the availability
of a wide bandwidth that is unregulated worldwide and
that can be reused in a very dense fashion, immunity to
eavesdropping, the ability to achieve very high bit rates,
low signal-processing complexity, and potentially very low
cost. The most difficult challenge in infrared link design
is achieving a high SNR, but major improvements in
link efficiency are possible through careful transmitter and
receiver design. The unique nature of the IM/DD infrared
channels necessitates reexamination of what are appropriate
modulation and multiple-access techniques. In most applications, average-power efficiency is of paramount importance, favoring the use of PPM-based modulation schemes
and TDMA-based multiple-access techniques. Multipath
propagation causes significant ISI in nondirected links at
bit rates above 10 Mb/s, but can be mitigated through
proper modulation and detection techniques. Advanced
components, such as quasi-diffuse transmitters and anglediversity receivers, promise significant increases in link
efficiency and bit rate, at the price of increased complexity.
The authors are grateful for the valuable contributions of
M. D. Audeh, J. B. Carruthers, H. T. Chee, K. P. Ho, W. J.
Krause, C. S. Lee, D. C. Lee, E. A. Lee, G. W. Marsh, D.
G. Messerschmitt, R. K. Narasimhan, T. D. Nguyen, D. S.
Shiu, and A. P. Tang.
[1] F. R. Gfeller and U. H. Bapst, “Wireless in-house data communication via diffuse infrared radiation,” Proc. IEEE, vol. 67,
pp. 1474–1486, Nov. 1979.
[2] Y. Nakata, J. Kashio, T. Kojima, and T. Noguchi, “In-house
wireless communication system using infrared radiation,” Proc.
7th Int. Conf. on Computer Commun., Sydney, Australia, Oct.
30–Nov. 2, 1984, pp. 333–337.
[3] M. D. Kotzin, “Short-range communications using diffusely
scattered infrared radiation,” Ph.D. dissertation, Northwestern
Univ., Evanston, IL, June 1981.
[4] M. E. Marhic, M. D. Kotzin, and A. P. van den Heuvel, “Reflectors and immersion lenses for detectors of diffuse radiation,”
J. Optical Soc. Amer., vol. 72, no. 3, pp. 352–355, Mar. 1982.
[5] M. D. Kotzin and A. P. van den Heuvel, “A duplex infrared system for in-building communications,” in IEEE Vehic.
Technol. Conf. Proc., pp. 179–185, 1986.
[6] T. S. Chu and M. J. Gans, “High speed infrared local wireless
communication,” IEEE Commun. Mag., vol. 25, no. 8, pp. 4–10,
Aug. 1987.
[7] A. Lessard and M. Gerla, “Wireless communication in the
automated factory environment,” IEEE Network Mag., vol. 2,
no. 3, pp. 64–69, May 1988.
[8] G. Yun and M. Kavehrad, “Spot diffusing and fly-eye receivers
for indoor infrared wireless communications,” in Proc. 1992
IEEE Conf. on Sel. Topics in Wireless Commun., Vancouver,
B.C., Canada, June 25–26, 1992, pp. 286–292.
[9] J. R. Barry, J. M. Kahn, W. J. Krause, E. A. Lee, and D. G.
Messerschmitt, “Simulation of multipath impulse response for
wireless optical channels,” IEEE J. Select. Areas in Commun.,
vol. 11, no. 3, pp. 367–379, Apr. 1993.
[10] J. R. Barry, Wireless Infrared Communications. Boston:
Kluwer, 1994.
[11] F. R. Gfeller, P. Bernasconi, W. Hirt, C. Elisii, and B. Weiss,
“Dynamic cell planning for wireless infrared in-house data
transmission,” in Mobile Commun.: Advanced Syst. and Components, Proc. 1994 Int. Zurich Seminar on Digital Commun.,
Zurich, Switzerland, Mar. 8–11, 1994, pp. 261–272.
[12] M. J. McCullagh and D. R. Wisely, “155 Mb/s optical wireless
link using a bootstrapped silicon APD receiver,” Electron. Lett.,
vol. 30, no. 5, pp. 430–432, 1994.
[13] H. Hashemi, G. Yung, M. Kavehrad, R. Behbahani, and P. A.
Galko, “Indoor propagation measurements at infrared frequencies for wireless local area networks applications,” IEEE Trans.
Vehic. Technol., vol. 43, pp. 562–576, Aug. 1994.
[14] R. T. Valadas and A. M. de Oliveira Duarte, “Sectored receivers
for indoor wireless optical communication systems,” in Proc.
5th IEEE Int. Symp. on Personal, Indoor, and Mobile Radio
Commun., The Hague, The Netherlands, Sept. 21–23, 1994, pp.
[15] G. W. Marsh and J. M. Kahn, “50-Mb/s diffuse infrared
free-space link using on-off keying with decision feedback
equalization,” IEEE Photon. Technol. Lett., vol. 6, no. 10, pp.
1268–1270, Oct. 1994.
[16] J. P. Savicki and S. P. Morgan, “Hemispherical concentrators
and spectral filters for planar sensors in diffuse radiation fields,”
Appl. Optics, vol. 33, no. 34, pp. 8057–8061, Dec. 1994.
[17] J. M. Kahn, W. J. Krause, and J. B. Carruthers, “Experimental characterization of non-directed indoor infrared channels,” IEEE Trans. Commun., vol. 43, pp. 1613–1623, Apr.
[18] K. P. Ho and J. M. Kahn, “Compound parabolic concentrators
for narrow-band wireless infrared receivers,” Opt. Engineering,
vol. 34, no. 5, pp. 1385–1395, May 1995.
[19] M. D. Audeh and J. M. Kahn, “Performance evaluation of
baseband OOK for wireless indoor infrared LAN’s operating
at 100 Mb/s”, IEEE Trans. Commun., vol. 43, no. 6, pp.
2085–2094, June 1995.
[20] J. R. Barry and J. M. Kahn, “Link design for non-directed
wireless infrared communications,” Appl. Optics, vol. 34, no.
19, pp. 3764–3776, July 1995.
[21] H. Park and J. R. Barry, “Modulation analysis for wireless infrared communications,” in Proc. IEEE Int. Conf. on Commun.,
Seattle, WA, June 18–22, 1995, pp. 1182–1186.
[22] T. D. Nguyen, M.S. report, Univ. Calif., Berkeley, July 1995.
[23] D. B. Medved and R. Halpern, “UWIN: A universal wireless
infrared network system,” in Proc. SPIE Conf. on Wireless
Commun., vol. 2556, San Diego, CA, July 12–13, 1995, pp.
[24] D. B. Medved and Y. Azancot, “Wireless optical communication for FDDI, fast Ethernet and ATM connectivity,” in Proc.
SPIE Conf. on Wireless Commun., vol. 2556, San Diego, CA,
July 12–13, 1995, pp. 294–304.
[25] M. D. Audeh, “Power-efficient modulation for high-speed nondirected wireless infrared communication,” Ph.D. dissertation,
Univ. Calif., Berkeley, 1995.
[26] C. R. A. T. Lomba, R. T. Valadas, and A. M. de Olveira
Duarte, “Sectored receivers to combat the multipath dispersion
of the indoor optical channel,” in Proc. 6th IEEE Int. Symp. on
Personal, Indoor and Mobile Radio Commun., Toronto, Canada,
Sept. 27–29, 1995, pp. 321–325.
[27] H. Zeino and M. Misson, “Adaptive infrared time-division
multiple access (AIR TDMA) for an in-house wireless hybrid
LAN,” in Proc. 6th IEEE Int. Symp. on Personal, Indoor, and
Mobile Radio Commun., pp. 332–337, Toronto, Canada, Sept.
27–29, 1995.
[28] A. M. R. Tavares, R. J. M. T. Valadas, and A. M. de Olveira
Duarte, “Performance of an optical sectored receiver for indoor
wireless communication systems in presence of artificial and
natural noise sources,” in Proc. SPIE Conf. on Wireless Data
Transmission, vol. 2601, Philadelphia, PA, Oct. 23–25, 1995,
pp. 264–273.
[29] P. P. Smyth, P. L. Eardley, K. T. Dalton, D. R. Wisely, P.
McKee, and D. Wood, “Optical wireless: A prognosis,” in SPIE
Proc. on Wireless Data Transmission, vol. 2601, Philadelphia,
PA, Oct. 23–25, 1995, pp. 212–225.
[30] P. Nicholls, R.T. Unwin, and K. T. Dalton, “A 10 Mb/s optical
wireless Ethernet: Practical results,” in Proc. SPIE Conf. on
Wireless Data Transmission, vol. 2601, Philadelphia, PA, Oct.
23–25, 1995, pp. 294–304.
[31] A. P. Tang, J. M. Kahn, and K. P. Ho, “Wireless infrared communication links using multi-beam transmitters and imaging
receivers,” in Proc. IEEE Int. Conf. on Commun., Dallas, TX,
June 23–27, 1996, pp. 180–186.
[32] F. Gfeller, W. Hirt, M. de Lange, and B. Weiss, “Wireless infrared transmission: How to reach all office space,” in
Proc. IEEE Vehic. Technol. Conf., Atlanta, GA, Apr. 1996, pp.
J. B. Carruthers and J. M. Kahn, “Multiple-subcarrier modulation for non-directed wireless infrared communication,” IEEE
J. Select. Areas in Commun., vol. 14, pp. 538–546, Apr. 1996.
M. D. Audeh, J. M. Kahn, and J. R. Barry, “Performance
of pulse-position modulation on measured non-directed indoor
infrared channels,” IEEE Trans. Commun., vol. 44, pp. 654–659,
June 1996.
J. B. Carruthers and J. M. Kahn, “Modeling of non-directed
wireless infrared channels,” Proc. IEEE Int. Conf. on Commun.,
Dallas, TX, June 23–27, 1996, pp. 1227–1231.
D. C. Lee, J. M. Kahn, and M. D. Audeh, “Trellis-coded
pulse-position modulation for indoor wireless infrared communications,” in Proc. 7th IEEE Int. Symp. on Personal, Indoor,
and Mobile Radio Commun., Taipei, Taiwan, Oct. 15–18, 1996,
pp. 349–353.
R. Valadas et al., “Experimental results of a pulse-position modulation infrared transceiver,” in Proc. 7th IEEE Int. Symp. on
Personal, Indoor, and Mobile Radio Commun., Taipei, Taiwan,
Oct. 15–18, 1996, pp. 252–256.
R. Narasimhan, M. D. Audeh, and J. M. Kahn, “Effect of
electronic-ballast fluorescent lighting on wireless infrared
links,” IEE Proc.-Optoelectron., Dec. 1996.
G. W. Marsh and J. M. Kahn, “Channel reuse strategies for
indoor infrared wireless communications,” to be published in
IEEE Trans. Commun.
, “Performance evaluation of experimental 50-Mb/s diffuse infrared wireless link using on-off keying with decisionfeedback equalization,” IEEE Trans. Commun., vol. 44, pp.
1496–1504, Nov. 1996.
A. J. C. Moreira, R. T. Valadas, and A. M. de Olveira Duarte,
“Optical interference produced by artificial light,” to be published in Wireless Networks.
M. D. Audeh, J. M. Kahn, and J. R. Barry, “Decision-feedback
equalization of pulse-position modulation on measured nondirected indoor infrared channels,” IEEE Trans. Commun., June
D. S. Shiu and J. M. Kahn, “Differential pulse-position modulation for power-efficient optical communication,” to be published
in IEEE Int. Conf. on Commun., Montreal, Quebec, Canada,
June 8–12, 1997.
IrDA Standards can be obtained at http://irda.org.
J. G. Proakis, Digital Communications, 3rd ed. New York:
McGraw-Hill, 1995.
R. D. Gitlin, J. F. Hayes, and S. B. Weinstein, Data Communications Principles. New York: Plenum, 1992.
E. A. Lee and D. G. Messerschmitt, Digital Communication,
2nd ed. Boston: Kluwer, 1994.
J. D. Parsons, The Mobile Radio Propagation Channel. New
York: Halsted, 1992.
Int. Electrotech. Commission, CEI/IEC825-1: Safety of Laser
Products, 1993.J. D. Rancourt, Optical Thin Films. New York:
Macmillan, 1987.
W. T. Welford and R. Winston, High Collection Nonimaging
Optics. San Diego: Academic, 1989.
X. Ning, R. Winston, and J. O’Gallagher, “Dielectric totally
internally reflecting concentrators,” Appl. Optics, vol. 26, no. 2,
pp. 300–305, Jan. 1987.
[52] B. E. A. Saleh and M. C. Teich, Fundamentals of Photonics.
New York: Wiley, 1991.
[53] B. L. Kasper, “Receiver design,” in Optical Fiber Telecommunications II, S. E. Miller and I. P. Kaminow, Eds. New York:
Academic, 1988, pp. 689–723.
[54] S. D. Personick, “Receiver design for digital fiber optic communications systems, I and II,” Bell Syst. Techn. J., vol 52, no.
6, pp 843–886, July–Aug. 1973.
[55] D. L. Rogers, “Integrated optical receivers using MSM
detectors,” IEEE J. Lightwave Technol., vol. 9, no. 12, pp.
1635–1638, Dec. 1991.
[56] G. Ungerboeck, “Trellis-coded modulation with redundant signal sets—Part I: Introduction,” IEEE Commun. Mag., vol. 25,
no. 2, pp. 5–11, Feb. 1987.
[57] J. A. Salehi, “Code division multiple access techniques in
optical fiber networks—Part I: Fundamental principles,” IEEE
Trans. Commun., vol. 37, pp. 824–833, Aug. 1989.
Joseph M. Kahn (Member, IEEE) received the
A.B., M.A., and Ph.D. degrees in physics from
the University of California at Berkeley in 1981,
1983, and 1986, respectively.
From 1987 to 1990 he was a Technical
Staff Member of AT&T Bell Laboratories’
Lightwave Communications Research Department, Holmdel, NJ. His research focused on
multigigabit-per-second coherent optical fiber
transmission systems and related device and
subsystems technologies. In 1990 he became a
Faculty Member at the University of California at Berkeley, where his
research interests include infrared and radio wireless communications
and optical fiber communications. He is also a Technical Editor of IEEE
Dr. Kahn received the National Science Foundation Presidential Young
Investigator Award. He is a member of the IEEE Communications Society
and the IEEE Lasers and Electro-Optics Society.
John R. Barry received the B.S. degree from
the State University of New York at Buffalo
in 1986, and the M.S. and Ph.D. degrees from
the University of California at Berkeley in 1987
and 1992, respectively, all in electrical engineering.
Since 1992 he has been an Assistant Professor
with the School of Electrical and Computer
Engineering at the Georgia Institute of Technology, Atlanta, where his research interests
include wireless communications, blind equalization, and multiuser communications. He authored Wireless Infrared
Communications (Kluwer, 1994).
Dr. Barry received the 1992 David J. Griep Memorial Prize, the 1993
Eliahu Jury Award from University of California at Berkeley, and a 1993
IBM Faculty Development Award.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF