Efficient Mitigation of Error Floors in Quantum Error Correction using Non-Binary Low-Density Parity-Check Codes

Kenta Kasai

REVIEW 4 major objections 6 minor 16 references

Efficient Mitigation of Error Floors in Quantum Error Correction using Non-Binary Low-Density Parity-Check Codes

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a post-processing step that identifies the length-2L cycle trapping the sum-product decoder can correct all Type-I and Type-III estimation errors in non-binary quantum LDPC codes, leaving only uncorrectable Type-II…

desk verdict Plausible post-processing for L=6 non-binary QLDPC error floors, but the load-bearing assumption that the SP estimate is exact outside the trapped cycle is unchecked, so the 'all Type-I/III corrected' claim outruns the evidence. read the letter →

arxiv 2501.13923 v2 pith:67RERGVU submitted 2025-01-23 quant-ph cs.ITmath.IT

Kenta Kasai This is my paper · ORCID

classification quant-phcs.ITmath.IT MSC 81P70 PACS 03.67.Pp

keywords quantumerrorcorrectionnon-binarylow-densityparity-checkcodesfloorsum-productdecodingcycleclassificationCSSdegenerateerrors

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes a post-processing step for non-binary quantum LDPC codes that targets the error floor caused by the sum-product decoder getting trapped in short cycles of the parity-check matrix. It classifies the length-2L trapping cycles into three types and gives a correction for each: Type-I errors are rendered degenerate, Type-II errors are shown to be undetectable and uncorrectable, and Type-III errors are solved exactly because the corresponding cycle matrix has full rank. Numerical experiments for the L=6 construction show that the method corrects all Type-I and Type-III failures, so the remaining error floor consists entirely of Type-II errors. This matters because error floors are the main obstacle to using these high-rate non-binary quantum LDPC codes at low noise rates, and the added correction has complexity O(L), independent of the code length.

What carries the argument

The central object is the length-2L cycle \hat C_\$\Delta$ in the Tanner graph of the F_q-valued matrix H_\$\Delta$, together with its classification by rank and by membership in \tilde C_\$\Delta$^{(2L)}, the cycles coming from rows of H_\Gamma. The post-processor uses the fact that for column weight J=2, a cycle is identified from the set J(\ell) of oscillating variable indices in O(L) time, and then reduces recovery to solving the linear equation \hat C_\$\Delta$ \hat\xi_J = (\sigma_0)_I. The rank of \hat C_\$\Delta$ dictates the outcome: rank L-1 gives a one-dimensional solution space whose offset from the true noise is a row of H_\Gamma, making the difference degenerate; rank L gives a unique solution equal to the true noise. This rank-based dichotomy is what carries the argument.

What would settle it

Run the proposed decoder on the L=6 code at flip probabilities inside the former error floor, collect every failed frame, and classify the trapped cycle by its type; any residual failure whose trapped cycle is not Type-II, or any failure in which the estimate is wrong at a position outside the identified cycle, would refute the claim that the remaining error floor consists entirely of Type-II errors.

Watch

Extended reading notes

Core claim

For the J=2, L=6 non-binary CSS LDPC code construction, a decoding failure in the error-floor region is caused by the estimated X-noise differing from the true noise only on a set of indices J contained in the columns of a length-2L cycle C_\$\Delta$ in the Tanner graph of H_\$\Delta$. The paper classifies such cycles as Type-I, II, or III according to whether the cycle belongs to the special set \tilde C_\$\Delta$^{(2L)} and whether its F_q-valued matrix has rank L-1 or L. The proposed post-processor identifies C_\$\Delta$ from the past L iterations of the SP decoder, then solves the restricted linear system \hat C_\$\Delta$ \hat\xi_J = (\sigma_0)_I. For Type-I, every solution differs from the true noise by a scalar multiple of a row of H_\Gamma, so the difference is a degenerate error and the codeword is recovered; for Type-III, the system has a unique solution equal to the true noise. The paper therefore claims that all Type-I and Type-III estimation errors are corrected by post-processing, and the remaining error floor consists entirely of Type-II errors, which produce the same syndrome as the true noise and cannot even be detected.

Load-bearing premise

The whole correction assumes the sum-product decoder's estimates are already correct at every position outside the identified cycle, so the syndrome computed from those outside estimates matches the true noise; if an estimate is wrong there too, the linear solve produces a wrong correction.

Editorial extensions

If this is right

For the L=6 code, every Type-I and Type-III estimation error in the error-floor region is corrected by the post-processing step, so the residual floor is exactly the Type-II contribution.
Because correction is obtained by solving a linear system on a single cycle, the added complexity is O(L), independent of the code length n = ePL.
Type-II failures are undetectable: the wrong estimate has the same syndrome as the true noise, so a decoder cannot tell that it failed; eliminating Type-II cycles in the code construction is the stated route to a much deeper error floor.
The method preserves the near-hashing-bound threshold of the original SP decoder while lowering the floor, extending the usable operating range of the code toward smaller flip probabilities.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: if a code construction can provably avoid Type-II cycles, this post-processor would convert the L=6 quantum LDPC family from floor-limited to threshold-limited, likely matching the sharp-threshold behavior already seen for L=8, 10, and 16.
Editorial inference: the cycle-identification scheme, which watches only which variable-node estimates flip during the last L SP iterations, is not tied to quantum specifics and could be tested as a general error-floor diagnostic for classical non-binary LDPC decoders.
Editorial inference: the argument's load-bearing assumption is that the SP estimate is correct everywhere outside the identified cycle; measuring how often this assumption fails at low flip probabilities would show whether the residual floor is really Type-II-only in practice.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a real attempt at a known problem—error floors in the non-binary QLDPC codes from [3,4]—and the cycle classification (Type I/II/III) is a useful way to organize failure modes. The O(L) post-processing for Type-I and Type-III cycles is neat, and the observation that non-binary cycles need not correspond to codewords is correctly exploited for Type-III. The paper also honestly states that Type-II errors are left uncorrectable and even undetectable, which is an important negative result if it holds.

The soft spots are mostly in the gap between what is simulated and what is claimed. Section VII assumes \hat xi_{bar J} = xi_{bar J} (noise correctly estimated outside the identified cycle J). That is load-bearing: if any symbol outside J is wrong, the syndrome σ0 is wrong and the linear solve doesn't give a valid correction. The paper gives no bound, no failure-rate data, and no theoretical argument for why this should hold. Section V only says that in 'most cases' the error set is contained in some cycle, not that the SP estimate is exact outside the identified cycle. So the closing claim that 'all Type-I and Type-III estimation errors were corrected' is not established by the evidence. It may be true—the FER plot does show a big improvement—but the mechanism the paper proposes does not prove it.

Also, Theorem 1 is imported from [3] with a one-line proof; that's acceptable if [3] is solid, but it makes the paper dependent on prior work that is itself not yet peer-reviewed. Numerical support is one FER curve, with no code, data, trial counts, or error bars. For an error-floor study, that is thin.

Where I land: the idea is worth taking seriously and the paper is readable, but the central claim needs either a proof that the outside-J estimate is reliable with high probability, or a simulation that counts how often the assumption fails and what happens then. Type-II uncorrectability should also be stated as an experimental observation, not a proven fact, unless there's a proof I'm missing.

I'd send this to review—it's a plausible subfield contribution and the classification may stick—but I'd ask for the missing validation before publication.

Referee Report

4 major / 6 minor

Summary. The paper proposes a post-processing method to lower the error floor of non-binary quantum LDPC codes with parameters J=2, L=6. The decoder first identifies the length-2L cycle in the Tanner graph of H_Δ in which the sum-product estimate is trapped, classifies the cycle as Type I, II, or III, and then applies a tailored correction: Type-I cycles are corrected by exploiting degeneracy, Type-II cycles are declared uncorrectable, and Type-III cycles are corrected by solving a full-rank linear system. The central claim is that after this post-processing, all remaining error-floor frames are Type-II estimation errors. Numerical results are presented as a single frame-error-rate plot comparing the proposed method with conventional SP decoding and the hashing bound.

Significance. If the proposed method works as claimed, it would offer a low-complexity (O(L), independent of code length) way to mitigate error floors in a family of non-binary quantum LDPC codes, which is of practical interest for scalable quantum error correction. The paper also gives a useful classification of trapping cycles and explicitly identifies an uncorrectable type, which is informative for code design. However, the significance is not yet established: the correction algorithm contains a load-bearing algebraic step that appears incorrect as written, the key assumption that the SP estimate is exact outside the identified cycle is neither proved nor quantified, and the numerical evidence is too sparse to verify the central claim that all Type-I and Type-III errors are corrected.

major comments (4)

[Section VII] The linear solve used for both Type-I and Type-III correction is not the equation that follows from the syndrome constraint. With J := J(Ĉ_Δ) and I := I(Ĉ_Δ), the true noise satisfies C ξ_J + (H_Δ)_{I, J̄} ξ_{J̄} = σ_I, where C = Ĉ_Δ. Since the assumption gives (H_Δ)_{J̄} ξ_{J̄} = σ_0, this implies C ξ_J = σ_I - (σ_0)_I, not C ξ_J = (σ_0)_I as stated. Consequently, the true noise is not generally a particular solution of the equation being solved, and the subsequent claims that ξ_J + ĉ_ξ_J equals a multiple of a row of H_Γ (Type-I) or that ĉ_ξ_J = ξ_J (Type-III) do not follow. This is a central error: the Type-I and Type-III correction derivations in Sections VII-A and VII-C, and the Section VIII assertion that all such errors were corrected, all depend on this step. Please correct the equation and re-derive the claimed properties.
[Section VII] The decoder design assumes ĉ_ξ^{(ℓ)}_{J̄} = ξ_{J̄} (the estimate is exact outside the identified cycle). This assumption is load-bearing: if any symbol outside J is misestimated, then σ_0 does not equal the true syndrome contribution from J̄, and the solved correction can fail to satisfy G_Γ(ξ + ĉξ) = 0. Section V only reports that in 'most cases' the error set is contained in some cycle, and that statement is about existence of a trapping cycle, not about absence of errors outside the specific cycle identified by the J(ℓ)/I(ℓ) heuristic. No proof, bound, or empirical failure-rate data is provided for the assumption. The paper should either prove the assumption for the identified cycle, or report how often it is violated and how violations are handled.
[Section VIII] The numerical evidence is insufficient to support the central claim that 'all Type-I and Type-III estimation errors were corrected' and that 'the remaining errors ... consisted entirely of Type-II estimation errors.' The paper reports a single FER plot with no trial counts, no error bars, no description of how the decoder classified frames into Type-I/II/III, and no data release. To make this claim verifiable, the authors should specify the number of simulated frames per point, report the breakdown of failure causes (including any frames where the J̄ assumption failed), and ideally release the code and data.
[Theorem 1] Theorem 1 is the formal bridge between the binary syndrome condition and the finite-field representation used throughout the correction algorithm, but its proof is deferred entirely to the author's preprint [3]. Since the current paper's central derivation depends on this theorem, the proof should be included or summarized in a self-contained appendix, rather than only citing an unreviewed arXiv manuscript.

minor comments (6)

[Section V] There are typos in this section: 'esitimation' should be 'estimation', and the phrase 'all the esitimation errors were observed' should be 'all the estimation errors were observed'.
[Section VII] There are typos: 'concatenatin' should be 'concatenating', and 'wherer' should be 'where'.
[Section II] The notation for the error operator is inconsistent: the text writes 'Applying ˆE† to this state yields E†E|ψ⟩' but the preceding and following expressions concern ˆE†E, not E†E. Please use a consistent notation such as $\hat{E}^\dagger E|\psi\rangle$.
[Section VI] The uniqueness of the cycle identification relies on the claim that no two cycles in C_Δ^{(2L)} share the same pair of columns. This is plausible given column weight J=2, but it would help to state explicitly that the pair (j,j') used for identification is chosen from J(ℓ) and that the O(L) identification assumes both indices belong to the same cycle.
[Section VII-B] The statement that Type-II cycles 'always' cause the SP decoder to converge to a wrong syndrome-matching estimate is stated as an experimental observation without supporting data. Since Type-II errors are declared uncorrectable, this claim should be either proved or accompanied by a quantitative statement about how often it occurs in the simulation.
[Figure 1] The figure caption does not explain what the solid and dashed curves represent beyond 'proposed method (dashed)'. It would help to include the code parameters (e, J, L, P, R) in the legend or caption, and to report the number of Monte Carlo trials used for each point.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the post-processing step is a conditional algebraic derivation, and the unproven 'correct outside J' assumption is an evidence gap, not a circular step.

full rationale

The paper's derivation chain is not circular. Section II reduces the degenerate-error condition to finite-field equations (8)-(10); these are algebraic equivalences, not the target result. The proposed decoder in Section VII explicitly states its conditional assumption: "The decoder design assumes that the noise is correctly estimated outside the indices in J := J(\hat C_\Delta), i.e., \hat\xi^{(\ell)}_{\bar J} = \xi_{\bar J}." This is a transparent hypothesis, not a fitted parameter disguised as a prediction, and not a quantity defined in terms of the claimed outcome. Given the assumption, the Type-I/Type-III corrections solve a linear system over the identified cycle; the fact that any solution differs from the true noise by a multiple of a row of H_\Gamma makes the residual syndrome-valid by Theorem 1. That theorem is imported from the author's earlier [3], but it is a parameter-free algebraic statement (finite-field image of (4)-(7)) that does not include the empirical error-floor claim as an input, so the self-citation is not load-bearing in a circular sense. The statement that "all Type-I and Type-III estimation errors were corrected" is an experimental observation over the simulated frames, not an equation forced by construction. The unproven character of the 'correct outside J' assumption is a genuine correctness/evidence risk: if the SP estimate errs outside J, the linear solve can output a wrong correction and the residual can remain non-degenerate. But that is a missing bound, not circularity. Similarly, the empirical classification into Type-I/II/III rests on the observation that floor errors lie on 2L cycles; this is a stated restriction, not a renaming of the conclusion. No fitted parameter is relabeled as a prediction, and no uniqueness theorem is invoked to forbid alternative decoders. The paper is therefore self-contained in the sense relevant to circularity.

Assumptions & free parameters 0 free parameters · 6 assumptions · 0 invented entities

The code construction, the finite-field maps, and the error-correcting condition are imported from the author's prior work [3,4]; the paper adds empirical assumptions about the error-floor mechanism and the performance of the SP decoder outside the trapped cycle.

assumptions (6)

domain assumption The stabilizer code is a CSS code with check matrix H having HX HZ^T = O, and decoding succeeds iff \hat E^\dagger E is in the stabilizer group.
Section II, equations (1)-(8); standard CSS stabilizer formalism.
standard math The code construction from [4] and [3] yields girth-12 protographs with rank L-1 for row-induced cycles C_Delta(i).
Section III and Section V; the rank statement is cited to [4,3] rather than proved here.
standard math The finite-field representation maps v, w and Theorem 1 are correct.
Theorem 1's proof is 'evident from the discussion in [3, Appendix B]', a self-cited reference.
ad hoc to paper Error-floor failures are exclusively caused by SP estimates trapped in length-2L cycles with Hamming distance to the true noise at most L.
Section V states this as an experimental observation, not a proven characterization.
ad hoc to paper Outside the trapped cycle, the SP estimate matches the true noise: \hat xi_{\bar J} = xi_{\bar J}.
Section VII explicitly assumes this; it is load-bearing for both Type-I and Type-III corrections but not validated.
ad hoc to paper Type-II cycles always cause the SP decoder to converge to a wrong syndrome-matching estimate, making failure undetectable.
Section VII-B reports this from experiments; no proof or data is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Efficient Mitigation of Error Floors in Quantum Error Correction using Non-Binary Low-Density Parity-Check Codes." pith.science (2026). https://pith.science/paper/67RERGVU

@misc{pith2026250113923,
  author       = {Pith},
  title        = {Pith review of: Efficient Mitigation of Error Floors in Quantum Error Correction using Non-Binary Low-Density Parity-Check Codes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/67RERGVU}},
  note         = {Machine review of arXiv:2501.13923}
}

read the original abstract

In this paper, we propose an efficient method to reduce error floors in quantum error correction using non-binary low-density parity-check (LDPC) codes. We identify and classify cycle structures in the parity-check matrix where estimated noise becomes trapped, and develop tailored decoding methods for each cycle type. For Type-I cycles, we propose a method to make the difference between estimated and true noise degenerate. Type-II cycles are shown to be uncorrectable, while for Type-III cycles, we utilize the fact that cycles in non-binary LDPC codes do not necessarily correspond to codewords, allowing us to estimate the true noise. Our method significantly improves decoding performance and reduces error floors.

Figures

Figures reproduced from arXiv: 2501.13923 by the authors.

**Figure 1.** Decoding performance (fm,FER) of the proposed method (dashed). for fm smaller than these threshold points. The threshold points where the curves intersect are marked with circles. On the other hand, for L = 6, a high error floor was observed. The error floor has been largely improved by the proposed method, but an error floor is still observed. A closer look at the experimental results shows that all Type-I and Type… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 13 canonical work pages

[3]

Quantum error correction near th e coding theoretical bound,

D. Komoto and K. Kasai, “Quantum error correction near th e coding theoretical bound,” arXiv:2412.21171, 2024

arXiv 2024
[1]

Logical quantum processor based on reconﬁgurable atom arrays,

D. Bluvstein, S. J. Evered, A. A. Geim, S. H. Li, H. Zhou, T. Manovitz, S. Ebadi, M. Cain, M. Kalinowski, D. Hangleiter , et al., “Logical quantum processor based on reconﬁgurable atom arrays,” Nature, vol. 626, no. 7997, pp. 58–65, 2024

work page 2024
[2]

Quantum computing in the NISQ era and beyon d,

J. Preskill, “Quantum computing in the NISQ era and beyon d,” Quantum, vol. 2, p. 79, 2018

work page 2018
[4]

Quantum error correction beyond the bounded distance decoding limi t,

K. Kasai, M. Hagiwara, H. Imai, and K. Sakaniwa, “Quantum error correction beyond the bounded distance decoding limi t,” IEEE Trans. Inf. Theory , vol. 58, no. 2, pp. 1223–1230, 2012

work page 2012
[5]

Explicit construction of classi cal and quantum quasi-cyclic low-density parity-check codes w ith column weight 2 and girth 12,

D. Komoto and K. Kasai, “Explicit construction of classi cal and quantum quasi-cyclic low-density parity-check codes w ith column weight 2 and girth 12,” arXiv:2501.13444, 2025

arXiv 2025
[6]

Spati ally coupled quasi-cyclic quantum LDPC codes,

M. Hagiwara, K. Kasai, I. Hideki, and K. Sakaniwa, “Spati ally coupled quasi-cyclic quantum LDPC codes,” in Proc. 2011 IEEE Int. Symp. Inf. Theory (ISIT) , Aug. 2011, pp. 543–547

work page 2011
[7]

Quantum error correction with girth-16 non- binary LDPC codes via aﬃne permutation construction,

K. Kasai, “Quantum error correction with girth-16 non- binary LDPC codes via aﬃne permutation construction,” arXiv:2504.17790, 2025

arXiv 2025
[8]

Stabilizer codes and quantum error corre ction,

D. Gottesman, “Stabilizer codes and quantum error corre ction,” Ph.D. dissertation, California Institute of Technology, M ay 1997

work page 1997

Show all 16 references

[9]

Good quantum error- correcting codes exist,

A. R. Calderbank and P. W. Shor, “Good quantum error- correcting codes exist,” Phys. Rev. A , vol. 54, no. 2, pp. 1098– 1105, Aug. 1996

1996
[10]

Multiple particle interference and quan tum error correction,

A. M. Steane, “Multiple particle interference and quan tum error correction,” vol. 452, no. 1954, pp. 2551–2577, 1996

1954
[11]

Error ﬂoors of non - binary LDPC codes,

T. Nozaki, K. Kasai, and K. Sakaniwa, “Error ﬂoors of non - binary LDPC codes,” in 2010 IEEE International Symposium on Information Theory , 2010, pp. 729–733

2010
[12]

Analysis of error ﬂoors of non-binary LDPC codes ov er MBIOS channel,

——, “Analysis of error ﬂoors of non-binary LDPC codes ov er MBIOS channel,” in 2011 IEEE International Conference on Communications (ICC), 2011, pp. 1–5

2011
[13]

Analysis of error ﬂoors of generalized non-binary LDPC codes over q-ary memoryless symmetric channels,

——, “Analysis of error ﬂoors of generalized non-binary LDPC codes over q-ary memoryless symmetric channels,” in Proc. 2012 IEEE Int. Symp. Inf. Theory (ISIT) , 2012, pp. 2341–2345

2012
[14]

Analysis of error ﬂoors of non-binary LDPC codes ov er MBIOS channel,

——, “Analysis of error ﬂoors of non-binary LDPC codes ov er MBIOS channel,” IEICE Trans. Fundamentals , vol. E94-A, no. 11, pp. 2144–2152, November 2011

2011
[15]

Analysis of error ﬂoors of non-binary LDPC codes ov er BEC,

——, “Analysis of error ﬂoors of non-binary LDPC codes ov er BEC,” IEICE Trans. Fundamentals, vol. E95-A, no. 1, pp. 381– 390, January 2012

2012
[16]

Analysis of error ﬂoors for non-binary LDPC codes over general linear group through q-ary memoryless symmetric channels,

——, “Analysis of error ﬂoors for non-binary LDPC codes over general linear group through q-ary memoryless symmetric channels,” IEICE Trans. Fundamentals, vol. E95-A, no. 12, pp. 2113–2121, December 2012

2012

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [3]

Quantum error correction near th e coding theoretical bound,

D. Komoto and K. Kasai, “Quantum error correction near th e coding theoretical bound,” arXiv:2412.21171, 2024

arXiv 2024

[2] [1]

Logical quantum processor based on reconﬁgurable atom arrays,

D. Bluvstein, S. J. Evered, A. A. Geim, S. H. Li, H. Zhou, T. Manovitz, S. Ebadi, M. Cain, M. Kalinowski, D. Hangleiter , et al., “Logical quantum processor based on reconﬁgurable atom arrays,” Nature, vol. 626, no. 7997, pp. 58–65, 2024

work page 2024

[3] [2]

Quantum computing in the NISQ era and beyon d,

J. Preskill, “Quantum computing in the NISQ era and beyon d,” Quantum, vol. 2, p. 79, 2018

work page 2018

[4] [4]

Quantum error correction beyond the bounded distance decoding limi t,

K. Kasai, M. Hagiwara, H. Imai, and K. Sakaniwa, “Quantum error correction beyond the bounded distance decoding limi t,” IEEE Trans. Inf. Theory , vol. 58, no. 2, pp. 1223–1230, 2012

work page 2012

[5] [5]

Explicit construction of classi cal and quantum quasi-cyclic low-density parity-check codes w ith column weight 2 and girth 12,

D. Komoto and K. Kasai, “Explicit construction of classi cal and quantum quasi-cyclic low-density parity-check codes w ith column weight 2 and girth 12,” arXiv:2501.13444, 2025

arXiv 2025

[6] [6]

Spati ally coupled quasi-cyclic quantum LDPC codes,

M. Hagiwara, K. Kasai, I. Hideki, and K. Sakaniwa, “Spati ally coupled quasi-cyclic quantum LDPC codes,” in Proc. 2011 IEEE Int. Symp. Inf. Theory (ISIT) , Aug. 2011, pp. 543–547

work page 2011

[7] [7]

Quantum error correction with girth-16 non- binary LDPC codes via aﬃne permutation construction,

K. Kasai, “Quantum error correction with girth-16 non- binary LDPC codes via aﬃne permutation construction,” arXiv:2504.17790, 2025

arXiv 2025

[8] [8]

Stabilizer codes and quantum error corre ction,

D. Gottesman, “Stabilizer codes and quantum error corre ction,” Ph.D. dissertation, California Institute of Technology, M ay 1997

work page 1997