WINT: A Novel Weighted Integer Representation with Improved Error Characteristics

Cheng-Yen Lee; Gautham Nemani; Sunil P. Khatri; Zach Assad

REVIEW 2 major objections 5 minor 12 references

A simple mantissa-exponent integer format can beat plain integers on both range and mean relative error for the same bit width.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 15:55 UTC pith:JS2T5NHO

load-bearing objection Clean, usable format plus careful O(1) MRE analysis that actually shows simultaneous range and error wins over plain integers for e=2–3. the 2 major comments →

arxiv 2607.09775 v1 pith:JS2T5NHO submitted 2026-07-07 cs.AR cs.NAmath.NA

WINT: A Novel Weighted Integer Representation with Improved Error Characteristics

Cheng-Yen Lee , Zach Assad , Gautham Nemani , Sunil P. Khatri This is my paper

classification cs.AR cs.NAmath.NA

keywords weighted integer representationconfigurable number formatmean relative errorharmonic numberprecision-range tradeoffmantissa-exponent encodingembedded systems

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Many memory-tight systems still need large dynamic range. This paper introduces Weighted Integer (WINT): store an m-bit mantissa and an e-bit exponent and decode the value simply as M times 2 to the E, with no bias, no implicit leading one, and no special encodings. The user chooses the split {m,e} at design time. The authors measure error by Mean Relative Error (MRE) over the continuous intervals between consecutive representable values and give exact integration plus O(1) harmonic and Taylor approximations that stay within 0.2 percent of the exact figure. Across 8- to 32-bit widths they show that spending two exponent bits both doubles the range and cuts MRE by 12-33 percent relative to a pure integer of the same width (for widths of 12 and above); three exponent bits multiplies range by 16 while cutting MRE by 15-50 percent (for widths of 16 and above). The practical claim is that a designer can obtain simultaneously larger range and lower average relative error than a standard integer without the complexity of floating-point hardware.

Core claim

For fixed total bit width K = m + e, the WINT representation V = M * 2^E with modest e (especially e = 2 or e = 3) simultaneously increases maximum representable range and reduces Mean Relative Error relative to a pure K-bit unsigned integer, while remaining far simpler to encode and decode than IEEE-style floating point.

What carries the argument

Mean Relative Error (MRE) obtained by integrating relative error over every interval between consecutive unique WINT values, together with closed-form harmonic-number (Euler-Maclaurin) and Taylor-series approximations that evaluate those sums in constant time independent of mantissa width.

Load-bearing premise

True values are treated as continuous and uniformly distributed so that average error can be obtained by integration over each interval; a strongly non-uniform or purely discrete distribution could change the ranking of configurations.

What would settle it

For any concrete bit width K >= 12, recompute MRE under a discrete uniform distribution over the integers up to the integer baseline's maximum and check whether the e = 2 (or e = 3) WINT configuration still shows both larger range and lower average relative error than the pure integer.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Clean, usable format plus careful O(1) MRE analysis that actually shows simultaneous range and error wins over plain integers for e=2–3.

read the letter

The one thing worth knowing is that WINT is a stripped-down mantissa-times-power-of-two format with fully free (m,e) allocation, and the authors give you both an exact MRE integral and two O(1) closed-form approximations (harmonic via Euler-Maclaurin, Taylor) that stay inside 0.2 % of the exact numbers while turning 40-hour runs into milliseconds. The headline empirical claim holds under their metric: for total widths ≥12, e=2 doubles range and cuts MRE 12–33 % versus the pure integer; e=3 gives 16× range and 15–50 % lower MRE for widths ≥16. Those numbers come straight out of the tables, not from tuning.

What is actually new is the combination of unrestricted bit allocation, the complete unique-value accounting (they correctly handle the non-canonical encodings), and the conversion of the mantissa sum into harmonic numbers so the whole MRE becomes constant-time. Related-work section is honest: they locate themselves relative to BFP, MX, Flexpoint, Posit, LNS and RNS without over-claiming. Derivations in §IV are careful; the continuous-uniform model is stated up front and applied uniformly to the integer baseline, so the relative rankings are internally consistent. Approximation error tables and the Bernoulli-term study are thorough.

Soft spots are real but modest. Continuous true values and uniform measure are modeling choices; a highly skewed or purely discrete distribution could reorder things, though the paper never claims otherwise. No silicon or application-level numbers, just the metric. Free parameters (Bernoulli terms, Taylor segments) exist but are validated rather than fitted to force the result. Novelty is incremental—the encoding itself is a natural generalization of integer-mantissa BFP—but the analysis package is the real contribution and is fully specified.

Math, tables and citation pattern look solid. This is for people who design number formats for memory-tight embedded or edge-ML work and need a simple knob between range and relative error. It deserves a serious referee; I would engage with it and expect it to survive review with only the usual polishing.

Referee Report

2 major / 5 minor

Summary. The paper introduces Weighted Integer (WINT), a configurable unsigned mantissa-exponent format V = M × 2^E with user-selectable bit allocations {m, e} and total width K = m + e. It develops an analytical framework for Mean Relative Error (MRE) under a continuous uniform model of true values, deriving an exact Integration Method (IM) for cumulative relative error (CRE) over intervals between consecutive representable values (Eqs. 7–9), a Triangle Approximation (TA) that reduces CRE to a simple geometric form (Eqs. 10–13), and two O(1) closed-form approximations—Taylor Series (TSA) via segmented expansions and Harmonic Approximation (HA) via conversion of the mantissa sum into harmonic numbers plus Euler-Maclaurin truncation (Eqs. 20–23). Tables III–VII compare all four methods for 8–32-bit configurations against the pure-integer baseline (e = 0), showing that HA/TSA stay within 0.2 % of IM while delivering enormous speed-ups, and that e = 2 yields both ~2 imes range and 12–33 % lower MRE for K ≥ 12, while e = 3 yields 16 imes range and 15–50 % lower MRE for K ≥ 16.

Significance. If the continuous-uniform MRE ranking holds for the intended applications, WINT supplies a simple, design-time-configurable alternative that simultaneously improves range and average relative error over plain integers for the same bit budget—an attractive property for memory-constrained embedded and edge-ML systems. The closed-form HA derivation (identity to harmonic numbers + controlled Euler-Maclaurin truncation) is a clean technical contribution that makes exhaustive configuration search practical even at 128 bits. The paper is fully self-contained: every approximation is validated against the exact integral, free parameters (Bernoulli terms, Taylor order) are fixed by explicit accuracy tables, and the reported percentage improvements emerge without post-hoc fitting. These strengths make the work a useful reference for both number-format designers and practitioners who need a lightweight integer-like format with tunable dynamic range.

major comments (2)

The continuous-uniform model of true values is stated explicitly in §IV-A and is applied uniformly to every configuration, including the integer baseline; the relative rankings are therefore internally consistent. However, the abstract and §V-D present the 12–33 % / 15–50 % MRE reductions as application-relevant without any discrete or non-uniform sensitivity check. A short experiment (or analytic bound) showing how the ranking changes under a discrete uniform measure or a power-law density would strengthen the claim that the reported improvements survive the modeling assumptions most likely to be violated in practice.
Section III-A notes that WINT encodings are non-canonical and that the number of unique values is 2^{m-1}(2^e + 1). The MRE derivation correctly sums only over unique intervals (Table I and the split in Eq. 13), yet the paper never quantifies the density of redundant encodings or their effect on encoding/decoding hardware. Because the central claim is that WINT is simpler than floating-point while offering better error/range than integers, a brief discussion of the hardware cost of handling (or ignoring) non-uniqueness would make the practicality argument complete.

minor comments (5)

Abstract and §I: the phrase “2X greater range” is slightly ambiguous; “doubles the range” or “2 imes the range” would be clearer.
Table I caption and surrounding text: the bold/italic distinction that identifies unique versus duplicate values is helpful but should be stated once in the caption itself so the table is self-contained.
Equation (5) and the subsequent range-increase examples: a short closed-form expression for the range-increase factor relative to a K-bit integer would make the trade-off discussion more precise.
Section V-B / Table II: the asymptotic divergence for m = 4 is correctly noted; a one-sentence recommendation that IM be used for m ≤ 4 would prevent readers from applying HA outside its validated regime.
References: the Posit and Microscaling citations are appropriate; a pointer to the classic block-floating-point literature (e.g., Oppenheim 1970 or more recent DSP surveys) would better situate the BFP discussion in §II-A.

Circularity Check

0 steps flagged

No significant circularity: MRE is derived from first principles and approximations are validated against the exact integral, not fitted to force the claimed improvements.

full rationale

The paper defines WINT as V = M imes 2^E with user-chosen {m, e}, then defines relative error, CRE via integration over consecutive representable intervals, and MRE as the average of those CREs (Eqs. 6–9). The triangle, Taylor, and harmonic approximations (Eqs. 10–13, 14–16, 20–23) are algebraic reductions of that same integral expression; they introduce no free parameters that are later tuned to data. Tables III–VII report both the exact IM values and the approximation errors (<0.2 %), so the numerical claims (12–33 % MRE reduction and 2 imes range for e = 2; 15–50 % MRE reduction and 16 imes range for e = 3) emerge directly from evaluating the formulas on every configuration, including the pure-integer baseline e = 0. There are no self-citations that carry the central argument, no uniqueness theorems imported from prior author work, and no renaming of a known empirical pattern. The continuous-uniform modeling choice is stated explicitly and applied uniformly, so the relative rankings are internally consistent by construction of the metric, not by circular fitting. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 1 invented entities

The central numerical claims rest on a small set of modeling choices (continuous true values, MRE definition, unique-value enumeration) plus two free truncation parameters for the asymptotic series. No new physical entities are postulated; WINT is an encoding, not a postulated object of nature. All other mathematics is standard.

free parameters (2)

number of Bernoulli correction terms in HA = 10
Chosen empirically as 10 after inspecting Table II; the asymptotic series diverges for small m if too many terms are kept.
number of Taylor terms / expansion segments in TSA = 10 terms, mid-point a
Ten terms and midpoint expansion points a = 1.5 × 2^k are selected by hand; accuracy is reported but the choice is free.

axioms (4)

domain assumption True values inside each representable interval are treated as continuous so that CRE can be obtained by definite integration.
Stated explicitly in §IV-A; required for the closed-form CRE expression (Eq. 8).
ad hoc to paper Mean Relative Error is defined as the sum of interval CREs divided by the maximum representable value V_range.
Eq. 9; this particular normalization is a modeling choice that weights larger absolute intervals more heavily.
domain assumption Only unique (m_i, e_j) pairs contribute to the MRE sum; duplicates are excluded by the bold/italic partitioning of Table I.
Necessary because WINT is non-canonical; derived in §IV-B.
standard math Euler-Maclaurin expansion of the harmonic numbers is a valid asymptotic approximation for the mantissa sums when m ≥ 5.
Invoked in §IV-D with reference to Lagarias; validated numerically in Table II.

invented entities (1)

Weighted Integer (WINT) format no independent evidence
purpose: Provide a simple, fully configurable mantissa-exponent integer encoding that improves both range and MRE relative to plain integers.
Defined by V = M × 2^E with free (m,e) allocation; the paper’s entire contribution is the definition plus its error analysis.

pith-pipeline@v1.1.0-grok45 · 19738 in / 2701 out tokens · 30245 ms · 2026-07-14T15:55:51.868630+00:00 · methodology

0 comments

read the original abstract

In computing, there is a need for number representation schemes that provide large dynamic range with low error. Many applications, including embedded systems and edge machine learning, have stringent memory constraints yet require large dynamic range for data representation. We present Weighted Integer (WINT), a simple and configurable mantissa exponent number format with user-selectable mantissa (m) and exponent (e) bit allocations (also referred to as configurations) that enables application-specific precision versus range tradeoffs at design time. We develop a complete analytical framework for computing Mean Relative Error (MRE), the primary metric for characterizing WINT's error. Since exact MRE calculations grow exponentially with mantissa size, we introduce harmonic and Taylor series approximation methods that achieve O(1) time complexity regardless of configuration. The Taylor series and harmonic approximations demonstrate significant speedups over the exact method while maintaining accuracy within 0.2% for the configurations presented. Our experiments across 8 to 32-bit configurations show that allocating 2 exponent bits consistently yields both lower MRE by 12-33% and 2X greater range than the integer baseline for bit widths of 12 and above. Allocating 3 exponent bits extends range by 16X while reducing MRE by 15-50% for bit widths of 16 and above

Figures

Figures reproduced from arXiv: 2607.09775 by Cheng-Yen Lee, Gautham Nemani, Sunil P. Khatri, Zach Assad.

**Figure 2.** Figure 2: illustrates the relative error profile within a single interval [Vk, Vk+1). For any true value V in the interval [Vk, Vk+1), it rounds to the nearest representable WINT value. The decision boundary is the midpoint Vhalf = (Vk + Vk+1)/2. Values below Vhalf round to Vk and values above round to Vk+1. V RE Vk Vhalf Vk+1 REmax CREIM,k CRETA,k [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

12 extracted references · 5 linked inside Pith

[1]

IEEE Standard for Floating-Point Arithmetic,

“IEEE Standard for Floating-Point Arithmetic,”IEEE Std 754-2019 (Re- vision of IEEE 754-2008), pp. 1–84, 2019

2019
[2]

Quantization and Deployment of Deep Neural Networks on Microcontrollers,

P.-E. Novac, G. Boukli Hacene, A. Pegatoquet, B. Miramond, and V . Gripon, “Quantization and Deployment of Deep Neural Networks on Microcontrollers,”Sensors, vol. 21, no. 9, 2021. [Online]. Available: https://www.mdpi.com/1424-8220/21/9/2984

2021
[3]

FP8 Formats for Deep Learning,

P. Micikevicius, D. Stosic, N. Burgess, M. Cornea, P. Dubey, R. Grisenthwaite, S. Ha, A. Heinecke, P. Judd, J. Kamalu, N. Mellempudi, S. Oberman, M. Shoeybi, M. Siu, and H. Wu, “FP8 Formats for Deep Learning,” 2022. [Online]. Available: https://arxiv.org/abs/2209.05433

Pith/arXiv arXiv 2022
[4]

Training DNNs with Hybrid Block Floating Point,

M. Drumond, T. Lin, M. Jaggi, and B. Falsafi, “Training DNNs with Hybrid Block Floating Point,” 2018. [Online]. Available: https: //arxiv.org/abs/1804.01526

Pith/arXiv arXiv 2018
[5]

Microscaling Data Formats for Deep Learning,

B. D. Rouhani, R. Zhao, A. More, M. Hall, A. Khodamoradi, S. Deng, D. Choudhary, M. Cornea, E. Dellinger, K. Denolf, S. Dusan, V . Elango, M. Golub, A. Heinecke, P. James-Roxby, D. Jani, G. Kolhe, M. Langhammer, A. Li, L. Melnick, M. Mesmakhosroshahi, A. Rodriguez, M. Schulte, R. Shafipour, L. Shao, M. Siu, P. Dubey, P. Micikevicius, M. Naumov, C. Verrill...

Pith/arXiv arXiv 2023
[6]

Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks,

U. K ¨oster, T. J. Webb, X. Wang, M. Nassar, A. K. Bansal, W. H. Constable, O. H. Elibol, S. Gray, S. Hall, L. Hornof, A. Khosrowshahi, C. Kloss, R. J. Pai, and N. Rao, “Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks,” 2017. [Online]. Available: https://arxiv.org/abs/1711.02213

Pith/arXiv arXiv 2017
[7]

OCP Microscaling Formats (MX) Specification, Version 1.0,

Open Compute Project, “OCP Microscaling Formats (MX) Specification, Version 1.0,” https://www.opencompute.org/documents/ ocp-microscaling-formats-mx-v1-0-spec-final-pdf, September 2023

2023
[8]

Beating Floating Point at its Own Game: Posit Arithmetic,

J. L. Gustafson and I. Yonemoto, “Beating Floating Point at its Own Game: Posit Arithmetic,”Supercomputing Frontiers and Innovations, vol. 4, no. 2, pp. 71–86, Jun. 2017. [Online]. Available: https://doi.org/10.14529/jsfi170206

doi:10.14529/jsfi170206 2017
[9]

Number Systems for Deep Neural Network Architectures: A Survey,

G. Alsuhli, V . Sakellariou, H. Saleh, M. Al-Qutayri, B. Mohammad, and T. Stouraitis, “Number Systems for Deep Neural Network Architectures: A Survey,” 2023. [Online]. Available: https://arxiv.org/abs/2307.05035

Pith/arXiv arXiv 2023
[10]

High-Performance Computation in Residue Number System Using Floating-Point Arithmetic,

K. Isupov, “High-Performance Computation in Residue Number System Using Floating-Point Arithmetic,”Computation, vol. 9, no. 2, 2021. [Online]. Available: https://www.mdpi.com/2079-3197/9/2/9

2021
[11]

Euler’s constant: Euler’s work and modern developments,

J. C. Lagarias, “Euler’s constant: Euler’s work and modern developments,” Bulletin of the American Mathematical Society, vol. 50, no. 4, pp. 527–628, Jul. 2013. [Online]. Available: http://dx.doi.org/10.1090/ S0273-0979-2013-01423-X

2013
[12]

Johanssonet al.,mpmath: a Python library for arbitrary-precision floating-point arithmetic (version 1.3.0), 2023, https://mpmath.org/

F. Johanssonet al.,mpmath: a Python library for arbitrary-precision floating-point arithmetic (version 1.3.0), 2023, https://mpmath.org/

2023

[1] [1]

IEEE Standard for Floating-Point Arithmetic,

“IEEE Standard for Floating-Point Arithmetic,”IEEE Std 754-2019 (Re- vision of IEEE 754-2008), pp. 1–84, 2019

2019

[2] [2]

Quantization and Deployment of Deep Neural Networks on Microcontrollers,

P.-E. Novac, G. Boukli Hacene, A. Pegatoquet, B. Miramond, and V . Gripon, “Quantization and Deployment of Deep Neural Networks on Microcontrollers,”Sensors, vol. 21, no. 9, 2021. [Online]. Available: https://www.mdpi.com/1424-8220/21/9/2984

2021

[3] [3]

FP8 Formats for Deep Learning,

P. Micikevicius, D. Stosic, N. Burgess, M. Cornea, P. Dubey, R. Grisenthwaite, S. Ha, A. Heinecke, P. Judd, J. Kamalu, N. Mellempudi, S. Oberman, M. Shoeybi, M. Siu, and H. Wu, “FP8 Formats for Deep Learning,” 2022. [Online]. Available: https://arxiv.org/abs/2209.05433

Pith/arXiv arXiv 2022

[4] [4]

Training DNNs with Hybrid Block Floating Point,

M. Drumond, T. Lin, M. Jaggi, and B. Falsafi, “Training DNNs with Hybrid Block Floating Point,” 2018. [Online]. Available: https: //arxiv.org/abs/1804.01526

Pith/arXiv arXiv 2018

[5] [5]

Microscaling Data Formats for Deep Learning,

B. D. Rouhani, R. Zhao, A. More, M. Hall, A. Khodamoradi, S. Deng, D. Choudhary, M. Cornea, E. Dellinger, K. Denolf, S. Dusan, V . Elango, M. Golub, A. Heinecke, P. James-Roxby, D. Jani, G. Kolhe, M. Langhammer, A. Li, L. Melnick, M. Mesmakhosroshahi, A. Rodriguez, M. Schulte, R. Shafipour, L. Shao, M. Siu, P. Dubey, P. Micikevicius, M. Naumov, C. Verrill...

Pith/arXiv arXiv 2023

[6] [6]

Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks,

U. K ¨oster, T. J. Webb, X. Wang, M. Nassar, A. K. Bansal, W. H. Constable, O. H. Elibol, S. Gray, S. Hall, L. Hornof, A. Khosrowshahi, C. Kloss, R. J. Pai, and N. Rao, “Flexpoint: An Adaptive Numerical Format for Efficient Training of Deep Neural Networks,” 2017. [Online]. Available: https://arxiv.org/abs/1711.02213

Pith/arXiv arXiv 2017

[7] [7]

OCP Microscaling Formats (MX) Specification, Version 1.0,

Open Compute Project, “OCP Microscaling Formats (MX) Specification, Version 1.0,” https://www.opencompute.org/documents/ ocp-microscaling-formats-mx-v1-0-spec-final-pdf, September 2023

2023

[8] [8]

Beating Floating Point at its Own Game: Posit Arithmetic,

J. L. Gustafson and I. Yonemoto, “Beating Floating Point at its Own Game: Posit Arithmetic,”Supercomputing Frontiers and Innovations, vol. 4, no. 2, pp. 71–86, Jun. 2017. [Online]. Available: https://doi.org/10.14529/jsfi170206

doi:10.14529/jsfi170206 2017

[9] [9]

Number Systems for Deep Neural Network Architectures: A Survey,

G. Alsuhli, V . Sakellariou, H. Saleh, M. Al-Qutayri, B. Mohammad, and T. Stouraitis, “Number Systems for Deep Neural Network Architectures: A Survey,” 2023. [Online]. Available: https://arxiv.org/abs/2307.05035

Pith/arXiv arXiv 2023

[10] [10]

High-Performance Computation in Residue Number System Using Floating-Point Arithmetic,

K. Isupov, “High-Performance Computation in Residue Number System Using Floating-Point Arithmetic,”Computation, vol. 9, no. 2, 2021. [Online]. Available: https://www.mdpi.com/2079-3197/9/2/9

2021

[11] [11]

Euler’s constant: Euler’s work and modern developments,

J. C. Lagarias, “Euler’s constant: Euler’s work and modern developments,” Bulletin of the American Mathematical Society, vol. 50, no. 4, pp. 527–628, Jul. 2013. [Online]. Available: http://dx.doi.org/10.1090/ S0273-0979-2013-01423-X

2013

[12] [12]

Johanssonet al.,mpmath: a Python library for arbitrary-precision floating-point arithmetic (version 1.3.0), 2023, https://mpmath.org/

F. Johanssonet al.,mpmath: a Python library for arbitrary-precision floating-point arithmetic (version 1.3.0), 2023, https://mpmath.org/

2023