Electronics Guide

Bain's Synchronized Pendulums

Alexander Bain, a Scottish clockmaker, received British patent number 9745 on 27 May 1843 for an apparatus that is now recognized as the first facsimile machine. Bain's insight was that a picture can be reduced to a sequence of contacts in time if a stylus sweeps it in a regular raster and if the receiving stylus sweeps an identical raster in step. He achieved that synchronism with the tool he understood best: a pair of pendulums, one at each end of the line, each released by an electromagnet and each regulated by a clock.

The transmitting stylus swept across a bed of raised metal type. Where the stylus touched metal, the circuit closed; where it fell into the gaps, the circuit opened. Between sweeps, a ratchet advanced the type bed by one line. At the receiving end, an identical pendulum dragged a stylus across paper soaked in a solution of ammonium nitrate and potassium ferrocyanide. Current passing through the stylus decomposed the salt and deposited Prussian blue, so the mark appeared without ink, impact, or heat. This electrochemical marking principle outlived Bain's mechanism by more than a century and reappeared in the electrolytic recorders of mid-twentieth-century weather-chart facsimile.

Bain's machine established the three functional blocks that every facsimile system since has contained: a scanner that converts spatial variation into a time-varying signal, a channel that carries the signal, and a marker that converts the signal back into spatial variation. It also established the central engineering difficulty, which is synchronization. If the receiver sweeps even slightly faster or slower than the transmitter, the image skews into a diagonal smear, and if the two lose line synchronism entirely, the page becomes noise.

Bakewell's Drum and Caselli's Pantelegraph

Frederick Bakewell replaced the pendulum with the arrangement that dominated facsimile for the next hundred and thirty years. In an 1848 British patent he described a "copying telegraph" in which the document was wrapped around a rotating cylinder while a stylus advanced slowly along the cylinder's axis on a lead screw. The stylus therefore traced a helix over the document surface, covering the whole page in one continuous motion. The drum is mechanically simpler than a reciprocating pendulum, it runs at constant velocity rather than the sinusoidal velocity of a swinging arm, and its rotation rate is easy to regulate. Bakewell demonstrated the apparatus at the Great Exhibition in London in 1851, but it was never reliable enough for service.

The first commercially operated facsimile system was the pantelegraph of Giovanni Caselli, an Italian physicist and priest working in France. Caselli combined Bain's electrochemical receiver with a regulating clock and a large pendulum frame, and solved the synchronization problem well enough for revenue service. The sender wrote the message in insulating ink on a tin sheet; the scanning stylus made contact with the bare metal between the strokes, and the receiving stylus deposited Prussian blue on chemically treated paper. French law authorized the service in 1864, and the Paris-to-Lyon line opened in 1865, later extending to Marseille. Traffic was modest, the tariff was high, and the service closed around 1870 amid the Franco-Prussian War, but it proved that facsimile could work over an ordinary telegraph line as a public offering.

The Photoelectric Turn

Every system to this point required a specially prepared original—raised type or insulating ink on metal—because the scanner sensed electrical contact. Facsimile became general only when the scanner learned to sense light. Shelford Bidwell demonstrated a scanning phototelegraph using a selenium cell in 1881, transmitting silhouettes. Arthur Korn in Germany made the technique practical for photographs in the first decade of the twentieth century, and by 1907 his equipment was carrying press pictures between Munich, Berlin, and Paris. Édouard Belin's portable Belinograph of 1913 worked over an ordinary telephone line and became the basis of the newspaper wirephoto services that AT&T and RCA launched in the 1920s.

Photoelectric scanning also changed the signal. Contact scanning produces a two-level signal by its nature, but a photocell produces a continuously variable output, so early phototelegraphy transmitted a genuinely analog gray scale, usually by modulating the amplitude or frequency of a tone. The bilevel, black-or-white character of modern office fax is not a limitation of scanning but a deliberate later choice, made because two-level images compress enormously well and text loses nothing by being rendered in two levels.

Office facsimile arrived slowly. Western Union's Deskfax, introduced in 1948, put a small drum machine on a subscriber's desk connected to a Western Union circuit. The decisive product was the Xerox Magnafax Telecopier of 1966, which coupled acoustically to an ordinary telephone handset and sent a page in about six minutes. It required no special line and no operator, and it made the fax machine a piece of office equipment rather than a carrier installation.

Drum Scanning

In a drum scanner the document wraps around a cylinder that turns at constant speed while the optical head advances along the axis on a lead screw. The scan is therefore a helix, and the image geometry follows directly from two mechanical constants. Vertical resolution—the line pitch—equals the lead screw advance per drum revolution. Horizontal resolution along the scan line depends on the spot size, the drum circumference, and the sampling rate of the electronics. The "index of cooperation," a number defined as the product of drum diameter and line density, had to match at both ends for the received page to have the same proportions as the original, and international facsimile services standardized on specific values so that machines from different manufacturers would interoperate.

Drum scanning gives excellent optical quality because a single detector and a single small illuminated spot are used for the entire page, so there is no sensor-to-sensor variation to correct. Its drawbacks are mechanical: the document must be flexible enough to wrap, it must be loaded and clamped by hand, bound volumes cannot be scanned at all, and the total scan time is fixed by the drum speed. Drum machines dominated wirephoto and weather facsimile through the 1970s and survive today only in high-end graphic-arts scanners built around photomultiplier tubes.

Photoelectric Line Scanning

Modern fax machines and multifunction devices move the paper past a stationary linear sensor, or move the sensor past stationary paper on a flatbed, and capture a whole scan line at once. Two optical architectures are in use. A charge-coupled device or CMOS linear array behind a reduction lens images the full page width onto a sensor a few centimeters long; this permits a large depth of field and tolerates books and thick originals. A contact image sensor instead places a full-width array of photosites directly beneath the document, illuminated by a light-emitting-diode bar and imaged one-to-one through a rod lens array. The contact image sensor is thinner, cheaper, and lower in power, and its very shallow depth of field is unimportant for the loose sheets that fax machines actually handle. The underlying detectors are treated in Image Sensors and the electrical interface in Image Sensor Interfaces.

Whatever the optics, the analog output must be conditioned before coding. Shading correction removes the cosine falloff of the illumination and the response variation between photosites, using a white reference strip scanned at the start of every page. A background-suppression stage then sets the threshold that separates black from white. Because fax is bilevel, this threshold decision is the single most consequential image-processing step in the machine: set too high, and faint pencil or thermal-paper originals disappear; set too low, and the gray of a photocopied background codes as black and inflates the transmission time.

Standard Resolutions

ITU-T Recommendation T.4 fixes the sampling grid. The standard scan line is 1728 picture elements, called pels, across a nominal 215 millimeters of paper width, which works out to 8.04 pels per millimeter, or approximately 204 pels per inch. Wider paper is supported at 2048 pels for 255 millimeters and 2432 pels for 303 millimeters, so that B4 and A3 originals keep the same horizontal density.

Vertical resolution is selectable and is the parameter a user actually notices. Standard vertical resolution is 3.85 lines per millimeter, approximately 98 lines per inch, which produces the familiar squat, slightly coarse fax page. Fine resolution doubles this to 7.7 lines per millimeter, approximately 196 lines per inch, at roughly double the transmission time. Superfine resolution of 15.4 lines per millimeter is optional, as is a doubled horizontal density of approximately 16 pels per millimeter. Later revisions of T.4 added true inch-based square grids of 200 by 200, 300 by 300, and 400 by 400 pels per inch so that Group 3 machines could exchange pages with the Group 4 and computer imaging equipment that had standardized on those numbers.

The asymmetry of the standard grid—roughly 204 by 98—is a deliberate trade-off. Horizontal resolution costs nothing in transmission time under run-length coding, because a wider line of the same content produces longer runs and therefore similar code lengths, whereas each additional scan line is an additional coded line to send. Spending resolution horizontally rather than vertically therefore buys legibility at low cost, which is why the standard fax page looks vertically stretched.

Groups 1 and 2: Analog Facsimile

Group 1, standardized in Recommendation T.2, transmitted an analog signal by frequency modulation of a voiceband subcarrier and required approximately six minutes for an A4 page at 3.85 lines per millimeter. Group 2, standardized in Recommendation T.3, halved that to approximately three minutes by using amplitude modulation with phase modulation and vestigial-sideband transmission around a 2100 hertz carrier, a considerably more bandwidth-efficient scheme. Neither group compressed the image, because neither digitized it; both simply modulated a continuous scan signal onto the telephone line. Both are long obsolete, though Group 2 reception lingered as a backward-compatibility mode in early Group 3 machines.

Group 3: The Standard That Endured

Group 3, approved in 1980, is the standard that made fax ubiquitous, and it is defined by a pair of recommendations that are best understood as separate layers. ITU-T Recommendation T.4 specifies the image: the scanning grid, the bilevel representation, and the source coding that compresses it. ITU-T Recommendation T.30 specifies the session: how two terminals find each other on a telephone call, declare their capabilities, agree on a common set, train the modem, transfer pages, acknowledge them, and disconnect. Between them they reduced an A4 page from six minutes to under a minute, and in the common case of an ordinary typed letter at standard resolution over a 14,400 bit-per-second modem, to a few seconds.

Group 3 succeeded for reasons that had little to do with image quality. It ran over the ordinary switched telephone network described in Public Switched Telephone Networks, so it needed no new subscriber line, no new tariff, and no new numbering plan; a fax number is simply a telephone number. It negotiated capabilities rather than mandating them, so a machine could implement only the mandatory minimum and still interoperate with anything. And it degraded gracefully: a poor line produced a slower transmission rather than a failed one.

Group 4 over ISDN

Group 4 was designed for digital networks rather than analog lines, principally the Integrated Services Digital Network at 64,000 bits per second. Because the underlying bearer is error-controlled and inherently digital, Group 4 needed no modem, no training sequence, and no facsimile-specific error correction; the image coding is ITU-T Recommendation T.6, and the rest of the stack is assembled from separate recommendations: T.62 for the facsimile control procedures, T.70 for network-independent transport, T.503 and T.521 for the document and communication application profiles, and T.90 for operation on the ISDN. Recommendation T.563, which defines the terminal itself, sets the basic pel density at 200 per inch and offers 240, 300, and 400 per inch as options. A page transfers in a handful of seconds.

Group 4 was the better design by almost every measure—an error-controlled bearer, no modem training, a square basic grid, and far shorter page times—and it failed commercially anyway. It required an ISDN line at both ends, and ISDN penetration was never high enough outside a few national markets to give a Group 4 terminal anyone to call. The single lasting contribution of Group 4 is its coding: T.6 was retrofitted into Group 3 as an optional negotiated capability and remains a standard compression option in the TIFF fax file profiles that fax servers store to this day.

Modified Huffman One-Dimensional Coding

The mandatory coding scheme of T.4, universally called Modified Huffman, codes each scan line independently as an alternating series of white and black run lengths. Every line is defined to begin with a white run, which may have length zero if the line starts in black, and thereafter the colors alternate so that no color flag needs to be transmitted.

Run lengths are represented by a fixed code table, not by a Huffman code computed per document. The CCITT derived the table from the measured run-length statistics of eight standard test documents, and it is built into every machine, which removes any need to transmit a code book. Runs of 0 through 63 pels have single terminating codes. Longer runs are coded as a makeup code representing a multiple of 64 followed by a terminating code for the remainder, with makeup codes defined up to 1728 and an extended set from 1792 to 2560 that is shared between the two colors. Separate tables are used for white and black runs because their statistics differ sharply: white runs are long and black runs are short in ordinary text, so the shortest codes are assigned accordingly.

Each coded line is terminated by an end-of-line code, twelve bits consisting of eleven zeros followed by a one. This pattern cannot occur inside any valid run-length code, so it serves as a resynchronization marker: a receiver that loses the bit stream discards data until it sees the next end-of-line code and resumes at the start of a fresh line, corrupting only one line rather than the remainder of the page. A page ends with a return-to-control sequence of six consecutive end-of-line codes. Fill bits of zero may be inserted before an end-of-line code to guarantee a minimum transmission time per line, which prevents the coded data from arriving faster than a mechanical printer can mark it.

Modified READ Two-Dimensional Coding

Modified Huffman ignores the strongest redundancy in a document image, which is vertical. In a page of text, a given scan line usually differs from the line above it by a pel or two at the edges of each character stroke. The optional two-dimensional scheme of T.4, called Modified READ for Relative Element Address Designate, codes those differences instead of the runs themselves.

The coder works with changing elements, meaning the pels at which the color changes. On the line being coded it tracks the reference element a0 and the next two changing elements a1 and a2; on the reference line immediately above it tracks b1, the first changing element to the right of a0 with the opposite color of a0, and b2, the next changing element after b1. Three coding modes follow from the relationship between these positions. Vertical mode applies when a1 lies within three pels of b1 and codes the signed offset in one to seven bits, with the common case of exact alignment costing a single bit. Pass mode applies when b2 lies to the left of a1, meaning a run on the reference line has no counterpart below, and costs four bits. Horizontal mode is the fallback for genuinely new content and codes the two runs a0a1 and a1a2 with the Modified Huffman tables after a three-bit escape.

Because a two-dimensional line is coded relative to the line above it, a single bit error corrupts not only its own line but every line that references it. T.4 bounds this propagation with the parameter K: at most K−1 two-dimensionally coded lines may follow a one-dimensionally coded line, so the coder resynchronizes to absolute coding at regular intervals. K is 2 at standard vertical resolution and 4 at fine resolution. A tag bit immediately after each end-of-line code declares whether the line that follows is one- or two-dimensional. Modified READ improves compression appreciably over Modified Huffman on ordinary text, and the K limit is exactly the trade-off between that gain and error containment.

Modified Modified READ and T.6

Recommendation T.6, the Group 4 coding scheme, removes the K limit entirely. Every line after a notional all-white reference line is coded two-dimensionally, no end-of-line codes are sent, and the page terminates with an end-of-facsimile-block pattern. This scheme, universally called Modified Modified READ or MMR, is the most efficient of the three on ordinary documents, because it spends nothing on end-of-line codes and never resynchronizes to absolute coding.

MMR is only safe on a channel that delivers the data intact, which is why it belongs natively to Group 4 over ISDN. Group 3 may negotiate MMR as an option, but responsible implementations do so only in conjunction with the error correction mode described below, because a single uncorrected bit error in an MMR stream destroys the remainder of the page. MMR is also available in the TIFF fax file profiles and is what the Group 4 variant of the CCITTFaxDecode filter in PDF expects, so a page that arrived under MH or MR is often recoded to MMR when it is filed.

Two further schemes appear as negotiated options in later revisions. Recommendations T.82 and T.85 bring JBIG arithmetic coding to Group 3, which compresses halftoned images far better than run-length methods. Recommendation T.81, the JPEG standard, together with the mixed raster content model of Recommendation T.44, supports color and continuous-tone facsimile. Neither option is widely deployed; the overwhelming majority of fax traffic remains bilevel and MH, MR, or MMR coded.

Halftones and the Limits of Bilevel Coding

Run-length coding assumes long runs, and a photograph has none. To send a photograph, the scanner must convert continuous gray into a bilevel pattern by dithering or error diffusion, and the resulting field of scattered isolated pels is close to the worst case for the T.4 tables: runs are one or two pels long, so nearly every run costs a full terminating code, and adjacent lines do not correlate, so two-dimensional coding gains nothing. A halftoned photograph can code larger than the uncompressed bitmap.

This is the reason a photograph faxes slowly and arrives looking poor, and the reason machines offer a separate "photo" mode that sacrifices resolution for tolerable transmission time. It is also the practical justification for the JBIG option, whose arithmetic coder with a context model handles dithered images gracefully where run-length coding cannot. The general treatment of these techniques appears in Coding and Error Control.

Phase A: Call Establishment

The calling terminal dials and then emits the calling tone CNG, an 1100 hertz tone in a repeating pattern of 0.5 seconds on and 3 seconds off, which identifies the call as a fax call to any equipment listening. The answering terminal responds with the called station identification tone CED, 2100 hertz held for between 2.6 and 4 seconds. Modern implementations usually send the CED as the ANSam variant defined in Recommendation V.8, a 2100 hertz tone carrying a low-frequency amplitude modulation that advertises V.8 capability to the far end. When the answer tone additionally carries periodic phase reversals, it is those reversals that command any network echo canceller in the path to disable itself. That instruction matters: an echo canceller left active will treat the far end's modem signal as an echo of the near end's and suppress it.

Phase B: Capability Negotiation and Training

All control signaling is carried in HDLC frames sent at 300 bits per second using Recommendation V.21 channel 2, a frequency-shift-keyed modem chosen for robustness rather than speed. Each frame carries a flag, address and control octets, a facsimile control field identifying the message, optional parameters, and a frame check sequence.

The answering terminal opens the exchange with DIS, the digital identification signal, a bit field declaring everything it can do: supported modulations and speeds, vertical resolutions, page widths and lengths, coding schemes, whether it supports error correction mode, and its minimum scan line time. It may precede DIS with CSI, the called subscriber identification, carrying its telephone number, and with NSF, a non-standard facilities frame used by manufacturers for proprietary features between their own machines.

The calling terminal replies with DCS, the digital command signal, which is not a negotiation but a decision: it selects one setting for each parameter from the intersection of its own capabilities with those the DIS advertised. It may precede DCS with TSI, the transmitting subscriber identification. The calling terminal then switches to the selected high-speed modulation and sends TCF, the training check field, consisting of 1.5 seconds of continuous zero bits. The receiver examines TCF and answers on the V.21 channel with CFR, confirmation to receive, if the pattern arrived clean, or FTT, failure to train, if it did not. FTT drives the sender down to the next lower speed, and the training repeats. This fallback ladder is why a fax to a poor line completes slowly instead of failing.

Phases C, D, and E: Message, Post-Message, and Release

Phase C is the image transfer itself at the negotiated modulation and coding. Phase D follows each page with a post-message command on the V.21 channel: MPS, the multipage signal, means another page follows at the same parameters; EOM, end of message, means another page follows but the parameters must be renegotiated from Phase B; EOP, end of procedure, means the document is complete. The receiver answers MCF, message confirmation, if the page was acceptable, or RTN, retrain negative, if it was not, which forces retraining before the page is resent. Phase E is the release, in which the sender transmits DCN, disconnect, and both terminals go on hook.

The strict half-duplex turn-taking of this protocol has an important consequence for networks: T.30 is timing-sensitive. Each phase has timers, typically on the order of three to six seconds for a response, and a delay that exceeds a timer causes the call to fail even though no data was actually lost. This is the root of most fax-over-packet trouble.

Error Correction Mode

The original Group 3 design had no retransmission. A burst of noise produced a black smear or a torn line on the page, which was visible and tolerable for text but fatal for MMR coding and unacceptable for documents of record. Annex A of Recommendation T.30 added the error correction mode, usually abbreviated ECM, which is negotiated in DIS and DCS like any other capability.

Under ECM the image data of Phase C is no longer a raw modem stream but is carried in HDLC frames, normally 256 octets of payload each, grouped into partial pages of up to 256 frames. At the end of each partial page the sender transmits PPS, the partial page signal. The receiver answers MCF if every frame passed its frame check sequence, or PPR, the partial page request, which carries a bitmap identifying exactly which frames failed. The sender then retransmits only those frames. After four unsuccessful rounds the sender may issue CTC to continue correcting at a lower speed, or EOR to abandon the partial page.

ECM changes the failure mode rather than eliminating failure. A page under ECM either arrives bit-exact or does not arrive, which is what a document of record requires, and it makes MMR coding safe over the switched network. The trade-off appears on marginal lines: a connection with a steady low error rate that would once have produced a slightly speckled but readable page now consumes the retransmission budget and disconnects with nothing printed. Disabling ECM is therefore a standard, if inelegant, remedy for a line that will not complete a fax, precisely because it restores the older behavior of accepting a damaged page.

V.27ter, V.29, and V.17

Group 3 uses a small family of voiceband modems, all of them half-duplex, all specified by ITU-T V-series recommendations, and all selected by the DIS and DCS exchange rather than by an independent modem handshake. Recommendation V.27ter is mandatory and provides 4800 bits per second using eight-phase differential phase-shift keying at 1600 symbols per second, falling back to 2400 bits per second using four-phase differential phase-shift keying at 1200 symbols per second. Recommendation V.29 adds 9600 and 7200 bits per second using quadrature amplitude modulation at 2400 symbols per second on a 1700 hertz carrier. Recommendation V.17 adds 14,400, 12,000, 9600, and 7200 bits per second using trellis-coded modulation at 2400 symbols per second, whose coding gain is what makes 14,400 bits per second practical in a 3.1 kilohertz channel. The general principles behind these constellations are covered in Digital Modulation Techniques.

Because the DIS and DCS frames name the modulation and speed explicitly, and because the TCF training check verifies the choice before any image is sent, the fax rate ladder is deterministic and observable. A fax that repeatedly falls back to 4800 bits per second on V.27ter is reporting a real channel impairment—typically noise, group-delay distortion, or excessive loss—and the fallback is diagnostic information rather than a defect.

V.34 and Super G3

Annex F of Recommendation T.30 defines the highest-speed Group 3 mode, marketed as Super G3, which uses the half-duplex mode of Recommendation V.34 to reach 33,600 bits per second. The arrangement differs structurally from the older modes. Rather than negotiating capabilities on the V.21 channel and then training, the terminals first run the V.8 handshake to establish that both support V.34, then open a V.34 control channel at 1200 bits per second that carries the T.30 control frames, thereafter switching between that control channel and the high-speed primary channel without retraining. V.34 also performs line probing, sending a sequence of tones to measure the channel response and select the symbol rate, carrier frequency, and precoder settings that suit it.

Super G3 roughly halves page time relative to V.17 and effectively eliminates the training and turnaround delays that dominate multipage calls. Its weakness is that the V.34 probing and adaptive equalization assume a stable end-to-end analog channel, an assumption that packet networks violate more thoroughly than they violate the simpler assumptions of V.27ter. It is common practice, when a fax will not complete over an IP path, to disable V.34 and force the call down to V.29 or V.27ter.

Direct Thermal

The machine that put fax on every desk in the 1980s printed by direct thermal marking on a roll of coated paper. The coating carries a colorless leuco dye and an acidic developer held apart in a solid binder; a thermal print head containing a linear array of small resistive elements heats selected points above the binder's melting temperature, the two components mix, and the dye turns black. There is no ink, no toner, and no ribbon, so the mechanism is cheap, quiet, and nearly maintenance-free, and the print head can be driven line by line in step with the incoming data.

The defects are equally well known. The paper arrives curled from the roll, cuts awkwardly, and feeds poorly through copiers. More seriously, the image is impermanent: the same chemistry that responds to the print head also responds to sunlight, heat, and common solvents including the plasticizers in vinyl folders, so a thermal fax left in a warm car or a plastic sleeve fades to illegibility. Offices that needed to keep faxes routinely photocopied them onto plain paper on arrival, which is a fair measure of how poor the archival properties were. The related technology is treated in Label Makers and Thermal Printers.

Thermal Transfer

Thermal transfer marking keeps the same print head but replaces the coated paper with a donor ribbon carrying a wax or resin pigment, which the head melts onto ordinary plain paper. The output is permanent, handles and files like any other document, and does not fade. The consumable cost is higher, and the ribbon advances by one page length whether the page is dense or nearly blank.

Thermal transfer carries a security consequence that surprised many organizations: the spent ribbon retains a legible negative image of every character it printed. A discarded ribbon is a complete record of the faxes a machine received, readable by anyone who unspools it, and the disposal of used fax ribbons became a recognized item in document-handling policy. The broader subject appears in Document and Data Security.

Electrophotographic and Inkjet Marking

Laser and light-emitting-diode electrophotography, the mechanism of the office copier, became the standard fax marking engine once the fax function migrated into multifunction devices. A charged photoconductive drum is discharged where the light source writes, toner adheres to the pattern, the toner is transferred to plain paper, and heat and pressure fuse it. The output is permanent, sharp, and indistinguishable from ordinary printing, and the same engine serves printing, copying, and fax reception, which is why the standalone fax machine effectively disappeared. Inkjet marking occupies the low-cost end of the same role. These engines are described in Printers and Printing Technology and Multifunction Devices.

Marking on demand also changed the buffering architecture. A direct thermal machine could print each scan line as it arrived, which is why T.4 defines a minimum scan line time to keep the data from outrunning the head. An electrophotographic engine must receive a complete page before it starts, so the machine holds the whole coded page in memory. That memory turned out to be the enabling condition for several familiar features: reception while out of paper, delayed transmission at off-peak tariffs, broadcast to multiple destinations from one scan, and confidential mailbox reception that prints only when a code is entered.

Why the Handshake Fails on Compressed Codecs

Fax is a modem call, and a modem call is not speech. When a fax call is carried as ordinary voice over an IP network, as described in Voice over IP Systems, several mechanisms designed to improve voice quality actively destroy it.

Low bit-rate codecs are the first and worst. G.729, G.723.1, AMR, and similar coders are built on a linear-prediction model of the human vocal tract; they transmit model parameters rather than a waveform, and they reproduce nothing that the model cannot represent. A V.17 constellation carries information in precise amplitude and phase, none of which survives such a coder. A fax will not complete over G.729 at any speed, and often the CED tone itself is distorted enough that the call never leaves Phase A.

Even with the uncompressed G.711 codec, four further mechanisms cause failures. Packet loss produces bit errors that ECM must repair, and a loss rate that is barely audible in speech—a fraction of a percent—can exhaust the retransmission budget. Adaptive jitter buffers change their depth mid-call by inserting or deleting samples, which is inaudible in speech and catastrophic to an equalized modem. Voice activity detection interprets the steady tones of the handshake as silence and replaces them with comfort noise. Clock slip between gateways with unsynchronized timing inserts or drops a sample periodically, and each such event is a symbol error. Echo cancellers that fail to recognize the ANSam disable command suppress the far end's carrier outright.

The mitigations for a G.711 pass-through path follow directly: force G.711 with no transcoding anywhere in the path, disable voice activity detection and comfort noise generation, disable or correctly arm echo cancellation, use a fixed rather than adaptive jitter buffer, ensure synchronized clocking on the gateways as discussed in Synchronization in Telecommunications, and cap the negotiated rate at V.29 or V.27ter. Pass-through remains fragile even when configured correctly, which is why the standards body defined a purpose-built alternative.

T.38 Real-Time Relay

ITU-T Recommendation T.38 defines procedures for real-time Group 3 facsimile over IP networks, and it works by refusing to carry the modem signal at all. A T.38 gateway demodulates the incoming fax, recovers the T.30 control frames and the T.4 image data as data, and carries them across the IP network as Internet Facsimile Protocol packets. The gateway at the far end remodulates them onto the terminating analog line. No modem waveform ever traverses the packet network, so codec distortion, jitter buffer adaptation, and clock slip cease to matter.

T.38 defines two transports. Carriage over TCP with TPKT framing gives reliable delivery but adds retransmission latency that the T.30 timers may not tolerate. Carriage over UDP using the UDPTL framing is the near-universal choice in practice, and because UDP does not retransmit, UDPTL provides its own protection by repeating a configurable number of previous packets inside each new one, or optionally by forward error correction. The redundancy is what makes T.38 tolerant of loss; a T.38 deployment with redundancy disabled is no more robust than pass-through.

A T.38 call normally begins as a voice call. The gateway detects the CNG or CED tone or the V.21 preamble, and the session is switched to the image/t38 media type by a SIP re-INVITE. Because the two ends of the call are now separated by a packet network with variable delay, the gateway must also manage the T.30 timers on behalf of the terminals it is proxying: while waiting for data from the network, it transmits HDLC flags to the local machine to hold the session open, a technique commonly called spoofing. Misconfiguration of the re-INVITE, a firewall that blocks the renegotiated media port, or a session border controller that declines to pass the image/t38 media type are the ordinary causes of a T.38 deployment that never engages and silently falls back to a pass-through call that then fails.

T.37 Store and Forward

ITU-T Recommendation T.37 takes the opposite approach and abandons real-time operation altogether. The page is scanned, coded as a TIFF file conforming to the fax profile, attached to an ordinary electronic mail message, and delivered through the mail infrastructure. Because there is no live session, there are no timers to violate and no modem to impair, and the transmission tolerates any amount of network delay.

The trade-off is the confirmation. A real-time T.30 call ends with an MCF from the terminating machine, which is positive evidence that the receiving equipment accepted the page; a store-and-forward message returns only mail-level delivery status, which says nothing about whether a document was printed or read. Since the evidentiary value of the transmission report is a principal reason organizations still use fax, this is not a minor loss. In practice most modern fax traffic runs on hybrid server platforms that accept documents by mail, web upload, or a virtual printer driver and then originate an ordinary T.30 call, by T.38 or over a remaining analog line, so that the terminating confirmation is still obtained.

Law, Evidence, and Regulation

Fax survives where a transmission must be attributable and provable rather than merely delivered. A T.30 call yields a confirmation that names the receiving station and records the time, and the connection is a point-to-point circuit between two identified telephone numbers rather than a message passing through intermediate stores. Many courts and administrative agencies accordingly continued to accept filings by fax long after electronic mail was universal, and a number of contract and notice provisions still name fax explicitly.

Regulation reinforced the evidentiary character. In the United States the Telephone Consumer Protection Act of 1991 and the Federal Communications Commission rules implementing it require a fax sent from a computer or similar device to be marked, in a margin at the top or bottom of each transmitted page or on the first page of the transmission, with the date and time of sending, an identification of the sender, and the sending telephone number—the source of the header line that fax machines print. The same statute, extended by the Junk Fax Prevention Act of 2005, restricts unsolicited advertising by fax and requires an opt-out mechanism, a body of law that continues to generate litigation.

Medicine

Health care is among the largest remaining users of facsimile in several countries. The United States Health Insurance Portability and Accountability Act names no transmission technology. It requires reasonable administrative, physical, and technical safeguards, and the Department of Health and Human Services has treated fax as a permissible channel for protected health information where such safeguards—confirming the destination number, controlling access to the receiving machine, using a cover sheet—are in place. Faxing therefore carried no compliance penalty during the long period in which genuinely interoperable electronic record exchange arrived slowly and unevenly. A fax also crosses organizational boundaries without any shared system: a clinic, a pharmacy, a laboratory, and an insurer running four incompatible record systems can all send and receive a page. That property—interoperability by lowest common denominator—is the actual reason fax persisted in medicine, and it is why replacing it required national programs rather than better products. NHS England barred its organizations from buying new fax machines from January 2019 and set 31 March 2020 as the date by which machines already in service were to be retired, with progress tracked by quarterly self-declaration; comparable interoperability programs are under way elsewhere. Related material appears in Health Information Systems.

The Withdrawal of the Analog Line

The constraint that will finally settle the question is not preference but bearer availability. Incumbent operators in many countries are retiring circuit-switched telephony and migrating subscribers to all-IP access, which removes the analog local loop that Group 3 was designed around. A fax machine on such a service reaches the network through an analog terminal adapter, so every packet-network impairment described earlier applies from the first meter of the connection rather than only at some distant gateway. Organizations that must keep faxing therefore end up on T.38 gateways or hosted fax platforms whether they planned to or not, and the decision narrows to which packet transport to use rather than whether to use one.

Retirement schedules differ by country and have repeatedly moved, so the defensible planning assumption is that the analog bearer is going away without a reliable date attached to it. The practical consequence for anyone maintaining fax service is to treat the analog line as a temporary arrangement, to validate the T.38 path before the analog one is withdrawn rather than after, and to record which counterparties still require fax at all, since that list is usually shorter than institutional habit suggests.

What Fax Actually Provides

It is tempting to treat the persistence of fax as institutional inertia, and inertia is certainly part of it. But fax offers a specific combination that its replacements did not immediately match: it requires no account, no directory, and no shared platform, only a telephone number; it delivers to a device rather than to a mailbox, so arrival is physical and visible; it produces a positive confirmation from the terminating equipment; and it carries a signature as an image without any question of format compatibility. Electronic mail with attachments matches none of those properly, and secure messaging platforms match them only within a closed membership. The lesson generalizes: a technology is displaced not when a better one exists but when a replacement covers every property that users depended on, including the ones nobody articulated.