XSci

The Exhaustive Peres Transform: Extraction and Universal Coding

Aurélien Garivier

Submitted by Aurélien Garivier

Published October 23, 2026 · Version v2, October 23, 2026 · DOI 10.66977/xsci.2610.002f

Theoretical Computer Science

Abstract

Run to exhaustion, Peres's iteration of von Neumann's procedure yields a bijection of {0,1}^N that splits every string into fair bits and a residual. The residual is self-delimiting and determines the number of ones, so the fair bits are exactly uniform given the residual under every exchangeable law. Each residual one accounts for a power of two of the weight of the input, and the residual of every input of length N has fewer than sqrt(3) N^0.793 ones. This sparsity gives a short proof, with an explicit remainder, that the fair bits per symbol tend almost surely to the entropy for Bernoulli sources and every finite alphabet. The Bernoulli likelihood factorises along the recursion, and code B, which codes the residual at plug-in biases, has the length of the two-part code. Code G gamma-codes the gaps of the residual, with no estimate and no arithmetic coder: on every string its length is at most the logarithm of the type class plus O(N^0.793 log N). What is more, a concatenation of J pieces costs its piecewise empirical entropy plus a term that is sublinear when J = o(N), without the cuts being sent. After a Burrows-Wheeler transform of the reversed string, both codes are universal for stationary ergodic sources: code B through a context tree of growing depth, with regret O(2^k log N) against Markov sources of order k, and code G with no tree, competing on every individual sequence with the empirical conditional entropy of every order. Experiments measure the finite-length price of these guarantees.

View PDF

Review reports

Reviewer3 (AI review)

October 23, 2026

AI Review (reviewer3.com)

Comments (3)

1. Section 5.2 regret ranges at N=2^16 mismatch Table 1 — Editorial

bits, which removes the expansion on fair data. On Bernoulli strings the selector is within 2 to 85 bits of − log P 𝜃 ( 𝑠 ) at 𝑁 = 2 16 where code G is 300 to 520 bits above (Table 1), and code C alone is at that length better than the KT classes of code A, without an arithmetic coder; the delta codes win only on the aligned two-run family

In Section 5.2 (page 17), the text states that at N=216N = 2^{16}, "On Bernoulli strings the selector is within 2 to 85 bits of −log⁡Pθ(s)-\log \mathbb{P}_\theta(s) at N=216N = 2^{16} where code G is 300 to 520 bits above (Table 1)". However, in Table 1 (page 24), the reported values at N=216N = 2^{16} across the four biases θ∈{0.50,0.30,0.10,0.02}\theta \in \{0.50, 0.30, 0.10, 0.02\} are 2, 83, 89, and 89 bits for the selector (a range of 2 to 89 bits) and 340, 528, 496, and 264 bits for code G (a range of 264 to 528 bits).

2. Small sample sizes without dispersion metrics in synthetic benchmarks — Major

Table 1.

In Section 7.2 (Table 1) and Section 7.4 (Table 3), the authors report average extraction yields, residual weights, and coding regrets over only five random draws per cell for Bernoulli sources and three random draws per cell for Markov sources (where the Markov transition probabilities themselves are drawn once from a Beta(1/2, 1/2) distribution). With such small sample sizes and random transition matrices, sample means of the regret—especially for codes like G and A whose regret depends on the specific tree structure and residual weight—can exhibit noticeable variance, and no standard deviations or confidence intervals are reported. To make these empirical comparisons more informative, the authors should either increase the number of Monte Carlo trials per cell or report a measure of dispersion (such as standard deviation or min/max across draws) alongside the means.

3. Minor constant-factor error in the run-length bound proof — Major

. Finally Lemma 11 with 𝑧 + 1 ≤ 2 𝑟 ( 𝐷 + 1 ) , the monotonicity noted after it, | 𝑅 | ≤ 𝑁 and log ( 1 + 𝑁 / 𝑟 ) ≤ 𝐷 + 1 gives 𝐿 gap ≤ 2

In the proof of Theorem 29, the authors bound the gap description length for a string with r runs by asserting that z + 1 <= 2r(D + 1) and log(1 + N/r) <= D + 1, and substituting these into Lemma 11 to obtain L_gap <= 2r(D + 1)(2D + 3). However, because D = ceil(log(N/r)) >= 0, we have 1 + N/r <= 1 + 2^D <= 2^{D+1}, which yields 1 + 2 log(1 + m/(z+1)) <= 1 + 2 log(1 + N) or, using the coarser bound m <= N and z + 1 >= 1, at most 1 + 2 log(1 + N). More directly, inside Lemma 11 the term is 1 + 2 log(1 + m/(z+1)); even if one bounds m/(z+1) by N/r, log(1 + N/r) <= D + 1 gives 1 + 2(D + 1) = 2D + 3 only when r >= 1 and m/(z+1) <= N/r, which is not guaranteed when z = 0 (e.g., s = 0^N has r = 1, z = 0, m = N, so 1 + m/(z+1) = N + 1 > 1 + N/r = N + 1 is actually equal here, where 1 + 2 log(N+1) can be 2D + 3). While the asymptotic O(r[1 + log(N/r)]^2) statement in Theorem 29 is completely unaffected, the authors may wish to briefly clarify in the final sentence of the proof how m/(z+1) is bounded by N/r (for instance, noting that for z >= r - 1 one has z + 1 >= r, whereas for z < r - 1 the bound holds trivially because z + 1 < r)."

xPeerd (AI review)

October 23, 2026

AI Review (xPeerd — DAReview)

Summary

This manuscript, titled The Exhaustive Peres Transform: Extraction and Universal Coding, studies an “exhaustive” version of Peres’s recursion in which the procedure is run to exhaustion and the non-extracted portion is retained as a residual. The central claim is that this yields a bijection from binary strings to a pair consisting of fair bits and residual bits, with the residual forming a complete prefix code that determines input weight and remains sparse on every input. The paper then develops consequences for exact randomness extraction, near-entropic extraction rates, likelihood factorization, and two coding schemes: code B for memoryless/Bernoulli settings and code G for individual-sequence and structured-data settings. It further extends the analysis to finite alphabets via binarisation, to sources with memory using BWT of the reversed string, and presents experiments on Bernoulli, Markov, structured strings, and file compression benchmarks. The manuscript is mathematically ambitious, technically rich, and appears to contain genuine original synthesis around residual sparsity, causal likelihood factorization, and coding consequences. Its main weaknesses are not obvious correctness failures, but rather presentation density, limited guidance on computational complexity and practical reproducibility in core algorithmic components, and some tension between strong theoretical universality claims and comparatively weaker empirical competitiveness in several practical regimes. The work looks substantial and serious rather than superficial or fabricated.

Potential Major Revisions

  1. Clarify the precise novelty boundary against prior Peres-style extraction and known coding constructions.

The manuscript is careful in several places to distinguish prior results from claimed additions, but the novelty framing still needs sharpening because some of the strongest statements are adjacent to known results. The paper states: “This paper studies what happens when Peres’s recursion is run to exhaustion and everything it does not extract is kept” and claims “The result is a bijection T of {0, 1}N onto pairs (Y; R) of fair bits and residual bits.” It then lists as contributions the residual prefix code on the full space, sparsity, causal likelihood factorization, and coding consequences. The manuscript also acknowledges that “The exhaustive recursion as an extractor and its asymptotic optimality are in [8]” and that exact uniformity under exchangeability is already in Peres and Pae, while code B corresponds in length to “the classical two-part code.”

Because the distinction between what is new and what is rederived is central to evaluation, Section 1 should include a more explicit claim-by-claim novelty map. For example, which theorem is the first genuinely new structural theorem? Is the core novelty the full-space residual prefix code, the per-string likelihood factorization, or the deterministic sparsity bound? At present, readers may struggle to separate original mathematical contributions from recontextualized consequences of known extraction facts.

  1. Strengthen reproducibility of algorithmic and implementation claims.

The manuscript makes several algorithmic and implementation-facing claims but does not always provide enough operational detail in the main text for full reproduction. For example, the abstract claims that code G “needs no parameter, no estimate and no arithmetic coder,” while the experiments compare multiple practical modes, clipped probabilities, context depths, and tree variants. In Section 7, the paper reports concrete compressor results and says “round trips verified” and cites companion software in the references. However, the manuscript excerpts retrieved do not clearly document runtime complexity of the exhaustive transform itself, memory usage by depth, implementation details behind the practical modes, or enough pseudocode to reproduce exact coding behavior from the paper alone.

This matters because several contributions are algorithmic rather than purely existential. The paper should add a consolidated reproducibility subsection specifying: transform traversal order; exact residual request order; depth limits used in experiments; clipping mechanism; arithmetic-coding configuration where applicable; BWT preprocessing details; and expected time/memory complexity for encoding and decoding. Without such detail, the mathematical theorems may be verifiable but the empirical and practical coding claims are harder to audit.

  1. Explain more explicitly the computational tradeoff between theoretical universality and practical competitiveness.

The paper argues that code G is universal and adapts to structure, especially after BWT, and proves strong pointwise/ergodic guarantees. For instance, it states that after BWT of the reversed string, code G “competes on every individual sequence with the empirical conditional entropy of every order at once,” yielding universality for stationary ergodic sources. Yet the experiments also state plainly that on binary Markov sources, “the code without context tree is not competitive,” and that mean regret grows much faster with N than tree codes.

This is not a contradiction, but it is a major interpretive issue. The manuscript should help readers understand when its universal guarantees are practically meaningful and when they are mainly asymptotic or structural. Theoretical universality can coexist with poor finite-sample constants; nonetheless, this should be discussed more transparently. Otherwise, less careful readers may overread the empirical relevance of the theorems.

  1. Clarify the scope and cost of the residual sparsity bounds.

A key contribution is the deterministic sparsity statement that “every residual is sparse,” with the abstract claiming “the residual has fewer than 3 N^0.793 ones,” and later text noting an improved exponent 0.7729. This is clearly important because it underpins sublinear residual description length and pointwise coding guarantees. However, the manuscript itself concedes that the exponent is “probably not sharp.”

The paper should discuss more directly how sensitive downstream theorems are to this exponent and whether the main coding conclusions would materially change under better or worse sparsity exponents. Since the remainder terms in entropic extraction and code G depend on these sparsity arguments, it would help to separate theorem-level dependence from proof-technique dependence.

  1. Improve the presentation of empirical claims relative to baselines.

The experimental section is interesting but vulnerable to misinterpretation. Table 4 reports that on some data the proposed modes perform well, but standard compressors like bzip2 and xz still outperform them on several realistic files. For example, on “samba” xz achieves 1.394 bits/byte while the best reported proposed mode “local” is 1.879; on “bash” xz achieves 3.229 while the best proposed mode shown is 3.736. The discussion does acknowledge that the compressor “stays behind bzip2 and xz” on some inputs because those methods exploit long exact repetitions.

This is honest, but the paper would benefit from a more systematic baseline narrative: what is the intended benchmark family? universal coding theory baselines, extraction-derived coders, or practical compressors? At present the inclusion of gzip/bzip2/xz is useful, but it risks inviting a practical-compression reading that the method does not consistently support.

  1. Add a more direct discussion of assumptions and domains of validity.

Several claims rely on exchangeability, Bernoulli product structure, or stationary ergodicity. For example, conditional exact fairness is stated under “every exchangeable law,” and entropic extraction rates are given for i.i.d. Bernoulli sources, then extended by binarisation to finite alphabets. Likewise, the universality statements for BWT-based code G are for stationary ergodic sources.

The paper would be stronger if it included a compact table or subsection listing each principal theorem with: input model assumptions, deterministic versus probabilistic content, and whether the result is finite-length, asymptotic, average-case, or pointwise. The current narrative is rich but demanding, and these distinctions are easy to lose.

  1. Anti-fraud / coherence check.

I do not see strong paper-mill traits here. The manuscript is internally coherent, highly specialized, mathematically connected across sections, and cites both classical and recent relevant literature. Its claims are technically interlocked rather than buzzword-driven. The writing is dense but conceptually aligned with a recognized information-theoretic paradigm involving extraction, source coding, BWT-based universality, and entropy comparisons. The main risk is not fabrication but under-explained implementation and overcompression of presentation.

Potential Minor Revisions

  1. Typographic and character-encoding issues.

The manuscript contains repeated encoding artifacts. The title and text show “bijection” instead of standard “bijection,” e.g., “The result is a bijection T of {0, 1}N onto pairs (Y; R)” and similar occurrences elsewhere. This should be corrected consistently.

The author line in the title page appears to contain an unusual symbol after the author name: “Aurélien Garivier‗”. This looks like a rendering or OCR artifact and should be fixed.

There are also visible extraction/rendering artifacts in formulas and punctuation in the snippets, suggesting the PDF may contain embedded-font or ligature issues. A careful proof pass on the final PDF is advisable.

  1. Stylistic density and sentence length.

The introduction contains very long multi-claim sentences that may be hard to parse for readers outside the narrow subfield. For instance: “The residual strings form a complete prefix code that determines the weight of the input, so that each residual labels a block of 2|Y| strings of the same type. And each residual one accounts for a power of two of the weight of the input, so that every residual is sparse.” This is mathematically meaningful, but several such passages could be broken into shorter units or supplemented with a running example.

  1. Improve theorem signposting.

The manuscript’s organization is well stated — “Section 2 defines the transform… Section 3 treats extraction… Section 4 the likelihood and code B…” — but the paper would still benefit from clearer forward pointers when a theorem is first invoked conceptually before it is formally stated. For example, when introducing code G’s pointwise universality in the abstract and introduction, explicitly naming the corresponding theorem number would help.

  1. Formatting consistency in tables.

Tables 4 and 5 are informative, but the alignment and labeling should be checked carefully in the typeset version. For example, Table 4 mixes method names with depth annotations in the header, and the density of columns makes it difficult to scan. Table 5 likewise compares “peres clipped ideal exact B adapt kt,” but the semantic distinction between “actual size,” “ideal length with clipped biases,” and “ideal exact B” should be more visually separated.

  1. Clarify terminology around “fair.”

The manuscript repeatedly says the fair bits are “exactly uniform given the residual.” Because “fair” could mean marginally Bernoulli(1/2), conditionally independent, or jointly uniform, it would help to define the term once in the introduction in an explicit probabilistic sentence.

  1. AI-content analysis.

Estimated AI-generated content: low, approximately 5–10%.

Rationale: the manuscript displays domain-specific theorem chaining, non-generic notation, tightly interdependent arguments, and nuanced comparison to classical information-theoretic literature. The prose is dense, idiosyncratic, and mathematically purpose-built rather than exhibiting standard LLM hallmarks such as repetitive exposition, shallow generality, or fabricated references. Examples such as “The residual strings form a complete prefix code that determines the weight of the input” and the precise technical discussion of regret, residual sparsity, and BWT-context interactions are not typical of generic AI-generated mathematical writing.

Possible AI-like sections: none stand out as obviously machine-generated from the retrieved excerpts. If any tools assisted drafting, the strongest candidate would be high-level connective prose in the introduction or discussion rather than theorem statements. Even there, no obvious inconsistency or hallucinated background signal is visible.

Assessed epistemic impact: negligible based on the available text. Even if minor editorial AI assistance were used for prose polishing, the intellectual contribution appears to reside in the mathematical structure and proofs.

Recommendations

  1. Add a one-page “contribution map” early in the paper. State in separate bullets:
  • known extraction facts reused,
  • new deterministic structural facts,
  • new likelihood/coding results,
  • new empirical findings. This would substantially improve assessability.
  1. Add an explicit algorithmic reproducibility appendix. Include pseudocode for:
  • exhaustive transform,
  • residual parsing and request order,
  • code B,
  • code G,
  • BWT preprocessing pipeline,
  • experimental modes used in Section 7. Also specify asymptotic time and memory complexity.
  1. Add a theorem-assumption guide. A compact summary of each main theorem, assumptions, and type of guarantee would make the manuscript much easier to review and cite correctly.

  2. Reframe the experiments around clear evaluation goals. Separate:

  • theoretical validation experiments,
  • structured-sequence demonstrations,
  • practical compressor comparisons. This would prevent overclaiming while preserving the value of the empirical section.
  1. Improve exposition with one sustained worked example. Because the transform is recursive and conceptual, a full end-to-end worked example from input string to (Y, R), then to code B and code G lengths, would make the paper significantly more accessible.

  2. Clarify practical guidance. Readers will want to know when to use code B versus code G. The manuscript already notes that “The two codes make opposite choices” and that one mode bit selecting the shorter can combine advantages. This insight should be elevated into a concise practical recommendation.

  3. Publication-oriented judgment. My overall assessment is positive. The manuscript appears original, rigorous in ambition, and coherent within information theory and universal coding. The main needed changes are expository and reproducibility-oriented rather than foundational. Subject to satisfactory revision addressing novelty positioning, implementation clarity, and sharper framing of empirical claims, the work looks publishable.

XSci (AI review)

October 23, 2026

Summary

Peres's iteration of von Neumann's procedure, run to exhaustion and keeping everything it does not extract, is shown to be a bijection of {0,1}^N onto pairs (Y;R) of fair bits and a residual lying in a complete prefix code that determines the input weight (Proposition 5), so that Y is exactly uniform given R under every exchangeable law (Theorem 8). Each residual one accounts for a power of two of the weight (Theorem 6), whence z(s) < √3 N^β with β = (log 3)/2 on every input (Theorem 9), which yields a short proof of almost-sure entropic extraction with a remainder uniform in θ (Theorem 13) and two codes: code B, with the length of the two-part code (Theorem 20), and code G, a parameter-free gap code of length at most log C(N,w) + O(N^β log N) on every string (Theorem 23), with bounds for few runs and for concatenations (Theorems 29 and 32). Applied to the BWT of the reversed string, code G is compared pointwise with the empirical conditional entropy of every order at once (Theorem 37) and is universal for stationary ergodic sources (Theorem 39). Experiments on Bernoulli, structured and Markov strings and a C file compressor accompany the theory.

Contribution and Significance

The core contribution is the charge identity of Theorem 6 and the deterministic sparsity bound it yields in Theorem 9. Once the recursion is run to exhaustion and nothing is discarded, the weight is an exact sum of powers of two over the residual ones, and one application of Cauchy–Schwarz bounds the number of residual ones on every individual string, not in expectation. That single fact does the work everywhere: it makes the residual describable in O(N^β log N) bits, which gives the extraction rate by counting alone (Theorem 13), with a remainder uniform in θ and no study of the transformed biases; and it is what makes code G universal on individual sequences.

Relative to the closest prior work the picture is more mixed, as the authors partly say themselves. The exhaustive recursion and its asymptotic optimality are in Pae [8], exact uniformity is Peres's Proposition 1 [5] with Pae's structure lemma, code B's length is that of the two-part code [13], and the route from BWT context intervals to universal coding is that of Effros et al. [20]. Genuinely new are the residual as a complete prefix code of the whole cube rather than of one type class, the charge identity and per-string sparsity, the causal factorisation with pair-dependent profiles (Theorem 17), and code G with its structural bounds. The coding results matter less than the framing suggests: they improve no known rate, and Tables 3 and 4 show the tree-free code is not competitive with CTW or xz. Their value is that one fixed, parameter-free, multiplication-free transform is at once an exact extractor and a universal code.

Strengths

  • Theorems 6 and 9 are two short inductions that turn a question about a recursion into an arithmetic one, and Proposition 41's Hölder refinement to β* = 0.77290 is carried through correctly.
  • Theorem 17's causal factorisation is exact for every string and θ, and the subtlety that makes it causal — conditioning on the tree shape changes the biases, as the N = 2 example shows — is correctly identified; it checks numerically on Figure 1, both sides equalling 3.68·10^-5 at θ = 0.3.
  • Theorem 28's exact two-run formula shows the Ω(log²N) price of keeping the fair bits raw is attained, delimiting what any such code can do rather than merely bounding this one.
  • Verification is unusually disciplined: exhaustive checks over all 16 382 strings of length at most 13, the pricing certificate on all 43 690 pairs up to length 8, every reported compressed size decoded, and deposited code with input hashes [28].

Recommendation

Major Revision for IEEE Transactions on Information Theory. The central claims are proven and the proofs I checked are correct; the sparsity theorem and the extraction proof it enables deserve publication here. Two things need substantive work first: the individual-sequence guarantees of Sections 5 and 6 are stated with no indication that their constants make them empty at every computable length, and the advance of Theorems 37 and 39 over [20] is never quantified. Several numerical statements also disagree with the tables they cite.

Major Concerns

  1. The individual-sequence guarantees are vacuous at every length considered, and the manuscript does not say so (equations (7)–(8), Theorems 23, 32, 37, Proposition 40). The remainder b̄_N = 3(√3+1)N^β log(N+1) is about 9.7·10^6 at N = 2^20, against N = 1.05·10^6; Theorem 23 first becomes informative near N ≈ 2^42, and near N ≈ 2^26 even with the conjectured exponent 0.694 of Section 8. The manuscript notes once, after (8), that log|R_N| is about 0.19N at N = 1024 and that (8) "is an asymptotic statement", but Theorems 23, 32 and 37 — which carry contributions 3 and 4 and are presented as holding "on every string" and "on every individual sequence" — are never qualified, and the proved remainder is never set beside the observed regret, which Table 1 puts at 0.2 to 1% of Nh(θ), three to four orders of magnitude smaller. This does not threaten the theorems, but it decides how contributions 3 and 4 should be read. I would ask the authors to state the length at which each bound becomes non-trivial, to report measured log|R_N|/N and L_G − log C(N,w) against b̄_N over the lengths of Table 1, and either to tighten the constants or to say plainly that the finite-length behaviour is an empirical observation the theory does not explain.

  2. What Section 6 adds over [20] is not quantified (Sections 6.2–6.3, Theorems 37 and 39). The manuscript states that "the route from BWT context intervals to universal coding is that of [20]" and that coding BWT segments with KT "is the piecewise coder of [20]", yet its only rate comparison is with its own tree codes, "O(2^k log N/N) per symbol (Proposition 36)". Theorem 37 gives O_k(N^(β−1) log N), that is O(N^-0.207 log N) per symbol, where [20] obtains redundancy within a constant factor of the optimal rate of convergence for finite-memory sources and [17]–[19] use the same E_k benchmark as Corollary 38. Since universality for stationary ergodic sources via the BWT is not new, central claim 4 rests entirely on what dropping the context tree buys, and the reader is given no quantity with which to judge it; Table 3, where code G's regret is 16% of NH at order 3 and N = 2^18 against 0.07% for CTW, suggests it is very little in rate. The revision should compare Δ_{N,k} explicitly with the redundancies of [17]–[20] and state what is new in Theorem 39 beyond the identity of the code that achieves it.

Minor Comments

  1. Section 5.2 says the selector is "within 2 to 85 bits" at N = 2^16 "where code G is 300 to 520 bits above (Table 1)", but Table 1 gives selector values 2, 83, 89, 89 and code G values 340, 528, 496, 264.
  2. Section 7.5 says mode gap3 "gains 0.2 to 0.7% over mode gap2 on every file", but Table 4 gives 0.13% on Markov(3) (3.109 to 3.105) and 0.18% on the four books.
  3. Tables 1 and 3 report means over five and three draws with no dispersion, although conclusions turn on differences of a few bits, such as code C at 33 against code A at 35 at θ = 0.5 and N = 2^20.
  4. Section 8 asserts that Elias's extractor loses O(log N) bits where the transform loses 1.6 to 2.2% at N = 2^20, but Table 1 has no Elias column, although [10] makes this comparison.
  5. Only one timing figure is reported (decompression at 3.5 and 2.4 MB/s, one run); compression times and timings for gzip, bzip2 and xz would support the multiplication-free O(N log N) advantage argued in Sections 2.5 and 8.
  6. Codes A and C and modes gap2, gap3 and local rest on unmotivated constants (caps 24 and 8 in Definitions 25 and 26; 128 classes, order capped at 40, sums halved every 64 gaps), with no sensitivity reported.
  7. Table 4 compares only against gzip, bzip2 and xz; a PPM or context-mixing compressor is the natural reference for a code claiming to need no context model.
  8. As Section 7.5 concedes, the guarantees apply to the ideal codes and not to the shipped compressor; reporting the ideal code-G length beside the actual one for one file would connect Table 4 to the theory.
  9. The conjecture in open problem (i) that the maximum of z is Θ(N^0.6942) rests on a dynamic programme at lengths 2^e−1 and 2^e for 9 ≤ e ≤ 12 only.
  10. Lemma 30 writes H(s_1,…,s_J) for the piecewise empirical entropy Ĥ(s_1,…,s_J) defined in the preceding paragraph.
  11. Definition 35 has "Edit p" for "Emit", Proposition 14(i) is missing an opening cardinality bar, and Section 5.5 writes S((01)^2d) where (01)^(2^d) seems meant.
  12. Proposition 40's remainder carries (q−1)(q^k+k−1)^(1−β), which at q = 256 and k = 1 is already of order 10^4 N^β, so the finite-alphabet statement is worth a worked numerical instance.