These semantic aspects of communication are irrelevant to the engineering problem.
7.1Message alternatives and a channel
Begin with a deliberately thin model. A sender holds a variable X ranging over a finite alphabet of alternatives and selects exactly one of them. An encoding maps that selection to a code C; a channel carries C; and a receiver faces a specified recovery task: reconstruct X in full, or recover some declared feature of X. Only once the task is stated does "successful communication" have determinate content. Before that, an engineering analogy risks adding formal appearance without explanatory value, and it cannot yet be compared fairly with careful ordinary editing or source pointers.
Take the stipulated case of four equally likely alternatives, written 00, 01, 10, 11, with the receiver given the code exactly. With the identity encoding C = X, both coordinates of the selected alternative are exactly recoverable: the receiver recovers the whole symbol, and the conditional entropy of X given C is zero. With C = a, where a is the first coordinate, the receiver recovers that coordinate and no more. Each arrangement is lossless with respect to its declared task. Losslessness, in other words, is a relation among an alphabet, an encoding, and a recovery demand, not a property of a message floating free of them.
The boundary is finite and strictly formal. This model says nothing about numerical human meaning or political legitimacy, and its quantities describe finite encodings rather than measuring either. Shannon's engineering formulation addresses selecting and transmitting messages, a task distinct from interpreting their meaning (Shannon, 1948). That boundary does not declare meaning unimportant; it marks what the present apparatus cannot settle.
Once the alphabet is fixed, the model has already decided what counts as distinct. Within that declared model, lossless recovery has a precise criterion: the output matches a declared sequence of symbols or it does not. It does not decide what those symbols signify, who composed them, or whether a reader ought to comply. A symbol error rate alone does not evaluate civic interpretation or political legitimacy. Coding theory therefore transfers as a way of separating recovery of a chosen distinction from recovery of meaning, not as a substitute account of meaning itself. The one qualification to carry forward is that every distinction in the model is model-relative; the model's alphabet fixes what counts as distinct, but it does not automatically encompass what readers could make of those symbols.
7.2Exact recovery of symbols
Lossless recovery is a task with a defined answer: given an alphabet and a specified sequence, the receiver's output either matches that sequence exactly or it does not. The criterion is binary and internal to the symbol system. Neither what the sequence signifies, nor who composed it, nor whether a reader ought to comply enters the specification. That narrowness is the point. When a question is stored, its bytes may be recovered perfectly while its meaning and its authority remain unsettled.
Shannon's introduction frames the engineering problem as the selection and transmission of messages from a defined set of alternatives (Shannon, 1948). That boundary does specific work: it identifies what a channel analysis can measure, and so helps separate a communication criterion from an account of meaning. It does not assert that semantic or political meaning is unimportant, and reading such a conclusion into the source would exceed what the source states.
Consider a hypothetical institutional archive. A submitted question is stored losslessly and later retrieved byte for byte, yet the summary displayed to the intake committee drops a negation, names the wrong requester, and sets a suggested answer in the question's voice. The archive succeeds under its specification; the display misleads. Because the two tasks have separate success criteria, neither result corrects the other, and perfect fidelity in storage does not repair a distorted rendering.
What transfers from communication theory is therefore modest but real: a recovery criterion can be stated exactly, and quality claims can then be tied to it. What does not transfer is any presumption that exact symbol recovery settles interpretation, provenance, or decision authority. Abstraction still earns its keep; a compact rendering can preserve the distinction on which a decision turns and spare readers irrelevant detail. But that is a claim about a deliberately chosen distinction, not about symbols as such. The next section asks what happens when a map deliberately preserves only part of the symbol.
The stored archive and the displayed rendering are different objects. A byte archive may preserve an original exactly while the display derived from it drops a negation or assigns a statement to the wrong contributor. Storage recovery has succeeded; the display is misleading. No error in the stored symbols can be inferred from the rendering alone, and conversely a byte-perfect archive does not correct what the rendering says.
An inspectable archive provides a comparison point. If it remains inspectable, the misleading display can be checked against it and corrected; the correction is a further act performed by someone with access to both. Inspection is the condition, not an automatic consequence of storage. Without it, the archive merely preserves the possibility of correction and may sit unexamined beneath a confidently wrong screen. Its value is therefore conditional: correction by comparison with this archive requires the stored sequence to be retrieved and compared. Other surviving sources might support a different route.
7.3Lossy maps with explicit alphabets
Take the alphabet stipulated in 07.1 as the set X = {00, 01, 10, 11}, with the four states equally likely. Under the identity encoding C = X, both coordinates of each symbol are exactly recoverable, and the conditional entropy H(X|C) is zero. That is lossless recovery: the code determines the symbol. A display map need not work this way. Define D from X to the two labels 0 and 1 by D(00) = D(01) = 0 and D(10) = D(11) = 1; in words, D keeps the first coordinate and drops the second.
This map is many-to-one. The distinct inputs 00 and 01 both produce the label 0, and 10 and 11 both produce 1. The map therefore preserves exactly one distinction — whether the first bit is 0 or 1 — while collapsing a second. That is the sense in which a map can be lossy with respect to symbols yet adequate for a chosen distinction: the two aims are not the same target. Suppose a limited lookup asks only for the retained coordinate; then the two-bit input carries more than the task requires.
Take the comparison left open by the many-to-one map. Under the projection just defined, 00 and 01 both display 0 while 10 and 11 both display 1. In each comparison the decoder receives only the declared code. Without additional information, then, it cannot uniquely recover both inputs from a shared output: 0 could have come from either 00 or 01. Knowing the map and the equiprobability of the states does not remove either collision. What is lost is the second coordinate, not the first.
The inference is a finite one. If two admissible inputs map to the same output, the output alone does not determine which occurred, and exact recovery of the full input is impossible. This is not a claim about every interpretive situation. A person may reconstruct a likely reading from context, memory, or expectations, but that reconstruction uses information outside the declared map. Such a decoder may be right or wrong; the map argument only records that the code does not settle the question.
7.4An example with two preserved bits
Against that finite-map result, the opposite case is worth stating exactly. Let X range over four equally likely states, written as the pairs 00, 01, 10, 11, and stipulate the identity encoding C = X: the code sent is the pair itself. A decoder receives only this declared code. Since each code corresponds to exactly one state, the decoder recovers both coordinates exactly. In the associated bookkeeping, H(X) = 2 bits, H(C) = 2 bits, and H(X|C) = 0 bits: seeing the code leaves no uncertainty about which state occurred.
The inferential step is deliberately narrow. Unlike a map that merges distinct inputs into a shared output, the identity encoding is one-to-one over the stipulated alphabet. Nothing is discarded, so the preservation of both bits follows from the structure of the code rather than from any observed regularity in the states. Lossless means precisely this here: an exact relabelling of a finite set of symbols. The two coordinates are recoverable, but the example says nothing yet about what either coordinate is for.
This is where the model boundary has to be declared before the coding analogy goes further. The example concerns finite symbol alphabets and stipulated encodings; its bit counts carry no numerical human meaning and no measure of political legitimacy. The formalism does limited but real work: here it establishes that each code identifies one allowed state. It does not supply an equivalence test for interpretations of those states. Beyond that identification, further formal restatement adds nothing to the example.
The identity code supplies a baseline for the next comparison: a deliberately reduced code may keep a chosen coordinate while no longer identifying the whole state. The changed encoding alters what can be recovered; the chosen requirement determines whether that change is acceptable.
Keep the same alphabet and the same four equally likely states, but change the encoding: let the code \(C=a\) retain only the first coordinate. The code now takes two values, so it carries one bit rather than two. A decoder receiving only the declared code can recover that coordinate exactly, but it cannot tell 00 from 01 or 10 from 11. The conditional entropy is one bit: the unresolved quantity is precisely the second coordinate. The identity code and the projection therefore support different recoveries, and their different recoverability follows from the encoding. The task determines which recovery is adequate.
This is where the coding analogy earns its keep. Entropy bookkeeping locates the distinction that has been dropped: one bit of the original description remains undetermined after the code is observed. If the task is only to retrieve the first coordinate, the projection is adequate and the dropped coordinate is outside that requirement. If the task is exact recovery of the pair, the same projection fails. What matters is the gap between the two recovery requirements. In this uniform finite example, the entropy and conditional entropy quantify the preserved and unresolved uncertainty. Neither quantity measures the importance, civic or otherwise, of any coordinate, and no political significance follows from retaining one bit rather than two.
7.5Why semantic equivalence needs a task
A declared comparison is also needed when assessing differences between strings. Consider a limited lookup task in which the answer is a requested fact about a room. The criteria are these: preserve the negation, so that the room is not open remains the stipulated assertion, and treat the spelling difference between colour and color as irrelevant to the requested fact. Under those criteria the two spellings are equivalent for this lookup, while deleting not yields a different assertion and so a failure of adequacy. Adequacy for this task is therefore not a property of the strings alone; it depends on the declared comparison.
For question compression, then, the transferable part of coding is exactness relative to a specified comparison. What coding does not supply is the choice of which distinction is worth preserving, and that choice must come from the task before any recovery claim is meaningful.
Suppose the task is no longer the limited lookup for which "colour" and "color" agree, but reproduction of an original inscription. Then the spelling difference becomes part of the requested output, while the truth value of the asserted statement may be irrelevant. Nothing in the text has changed; the declared comparison has. Adequacy for this task is therefore a relation among an output, a target, and a criterion, not a property that a compressed form carries by itself.
The finite encoding example shows the same structure without importing any claim about human meaning. Let X range over 00, 01, 10, and 11. With the identity encoding C = X, a decoder receiving C recovers both coordinates, so H(X|C) = 0 bits. With C = a, only the first coordinate is retained and the second is lost. Which encoding is adequate depends on which distinction the task requires. The inscription and coordinate cases have the same pattern of changed requirements. No encoding from that inscription to these two bits has been defined, so the comparison is an analogy, not a literal bit count for the text.
7.6Errors of importing an analogy
The identity code uses two bits and the projection uses one. The latter is more economical in transmitted length, but that observation alone does not establish better performance against the required recovery task.
The unsupported step is therefore the inference from "this code is shorter" to "this code represents better." Shortness is measured relative to what the code records; better representation is measured relative to what the comparison requires. If the declared task is to recover both coordinates, the one-bit code is strictly worse despite its economy. For a first-coordinate task, however, both codes succeed: a receiver can read that coordinate from either. Under this same task the one-bit code uses less transmission. A fair comparison holds the input domain and recovery requirement fixed while allowing the encoding to vary. Holding the encoding itself fixed would prevent precisely the comparison being proposed.
The one qualification is that a lower message length is evidence of economy under a stated recovery test. It is not evidence, by itself, that the shorter representation preserves the distinction that matters.
The useful contribution is an explicit map that can be checked rather than admired, and a recovery test run on declared inputs. This study tells a reader which coordinates survive a specified encoding and which do not. An analogy without a declared map and recovery question may still suggest an inquiry, but it has not yet supplied a tested account of preserved distinctions.
7.7A modest role for information measures
For the four equally likely pairs, H(X) = 2 bits. Under the stipulated identity encoding C = X, the code carries H(C) = 2 bits and leaves H(X|C) = 0 bits: a decoder given only the code recovers both coordinates exactly. Under the encoding C = a, which retains only the first coordinate, H(C) = 1 bit and H(X|C) = 1 bit. The comparison is tractable precisely because the alphabet, the map and the recovery test are all declared. If the task is to restore the whole pair, the identity encoding passes and the reduced one fails; if the task is only to return the first coordinate, the reduced encoding passes just as well. The measures register what each code preserves and discards. They do not register whether that preservation is worth wanting.
So the contribution is modest but real. Information quantities support narrow comparisons among encodings of fixed variables, and, in this example, distinguish full recovery from a one-coordinate projection. Which distinctions ought to be encoded, and who is entitled to settle that for a group, remain interpretive and normative questions that sit alongside the quantities rather than inside them.
These quantities leave open which coordinate matters, or which distinctions the inquiry is meant to preserve. As the preceding section put it, these are interpretive and normative judgments, not readings taken off the quantities. A number characterising recoverability under a declared task appoints no editor to speak for a group. If members disagree about which feature a summary should retain or who may narrow a question, the computed entropies do not resolve their dispute: the dispute concerns the relevance target and the authority to set it, not the size of the encoding.
The role is therefore bounded. Declare an input alphabet, a recovery task and each candidate encoding, and the quantities can inform a comparison under that task; conditional on choosing a different task, the same raw states may yield a different ordering. For evaluating a proposed summary, the arithmetic needs an accompanying account of what is to be preserved, for whom, and for what purpose. That statement remains an argument to be made, and it cannot be delegated to the computation that follows it.