Recommended Free Tools
Parity does not need a wavelet transform: an integer is even when its least significant bit (LSB) is 0 and odd when it is 1. That makes parity a useful diagnostic task. In a 2026 experiment, a wavelet-and-clustering pipeline classified a particular binary encoding above chance—but its accuracy changed sharply when the bit layout, boundary handling, or numeric range changed. The more revealing question is not whether wavelets can solve parity, but what the representation makes accessible to a simple model.
Why build a complicated classifier for a one-bit answer?
For an integer in binary, parity is already explicit: the LSB is 0 for an even number and 1 for an odd one. A direct parity test can read that bit. A wavelet pipeline is therefore unnecessary for solving the task; its value here is as a controlled probe of representation. It lets us ask whether a simple downstream method can recover information after a signal has been transformed and summarized.
As an Amazon Associate I earn from qualifying purchases.
That distinction matters. A model can perform well because its input representation preserves a useful cue, without discovering a general arithmetic principle. The experiment tests accessibility of information in one encoding and pipeline, not whether wavelets independently infer the rule of parity.
Free tools Windows power users keep installed
One-click scans. No signup required.
What the experiment actually did
Ertuğrul Mutlu’s revised 2026 study encoded every integer from 0 through 10,000 as a fixed-width 32-bit binary signal with zeros padded on the left. It applied a level-3 Daubechies-2 (db2) discrete wavelet transform using symmetric boundary extension, then summarized coefficients by mean absolute value (MAV). Separate k-means clustering with k = 2 was run for each wavelet subband.
#1 Best Overall
- Use scikit-learn to track an example ML project end to end
- Explore several models, including support vector machines, decision trees, random forests, and ensemble methods
- Exploit unsupervised learning techniques such as dimensionality reduction, clustering, and anomaly detection
- Dive into neural net architectures, including convolutional nets, recurrent nets, generative adversarial networks, autoencoders, diffusion models, and transformers
- Use TensorFlow and Keras to build and train neural nets for computer vision, natural language processing, generative models, and deep reinforcement learning
The dataset was divided into 6,000 training, 2,000 validation, and 2,001 held-out test examples. Clustering itself did not use labels, but the assignment of clusters to even or odd used training labels. So the clustering step was unsupervised; the full classifier was not fully unsupervised.
In this frozen setup, reported held-out test accuracy was 84.26%, with a 95% Wilson confidence interval of 82.60%–85.79%. Across 20 stratified random 80/20 resplits, reported accuracy was 84.20% ± 0.57%. These figures describe this specific protocol and range; they do not establish a useful general-purpose parity classifier. The revised arXiv paper and author repository document the experiment and its revised setup.
Rank #2
Which parts of the representation mattered?
Removing the LSB removed the useful signal
When the natural LSB was masked while the rest of the pipeline remained unchanged, validation accuracy fell to 48.15%, close to chance for a balanced binary task. The level-3 approximation band A3 alone reached 83.20%, while detail bands were near chance. This is evidence that the pipeline’s performance depended on information carried through the encoded signal, rather than a result showing it had found parity independently of the input cue.
Moving the bit changed the result
Mutlu reports that moving the parity-carrying bit to different signal positions altered accuracy, with the best tested position reaching 98.60%. That is not a directly comparable replacement for the 84.26% held-out result: it is an ablation result under a different bit layout. The large change nonetheless shows that the same underlying label can be easier or harder for this pipeline to access depending on where its informative bit appears.
Boundary handling also changed validation accuracy
Changing the wavelet boundary-extension mode produced validation results from 54.45% to 83.20%. Boundary handling determines how a transform treats signal edges; with a finite, fixed-width binary word, that choice can affect coefficient summaries. The experiment therefore ties performance not just to the abstract transform family, but to implementation details in the representation and processing pipeline.
Can wavelets classify parity?
Under the reported encoding and procedure, yes: the pipeline classified the selected examples above chance. But that answer needs two limits. First, parity is already directly encoded in the LSB. Second, the outcome is sensitive to bit position, subband, boundary mode, and the numeric distribution. As Mutlu states in the revised paper’s abstract, “These results do not show that wavelets discover the arithmetic rule of parity.”
Rank #4
The supported claim is narrower and more useful: a transform plus coefficient summaries and clustering can make an existing cue available to a simple classifier in some configurations. Accuracy alone does not tell us that the classifier learned a representation-independent rule.
What happened outside the training range?
The frozen model’s accuracy fell to 79.98% on integers 10,001–20,000 and to 59.69% on 100,001–1,000,000. In contrast, models trained and tested separately within fixed bit-length bands reportedly stayed around 78%–88%. Mutlu interprets this pattern as evidence that representation or distribution shift is a major factor in the tested setup. It does not establish a universal claim about all parity classifiers; it shows that this frozen pipeline did not maintain its in-range result on more distant numeric ranges.
Best Value
The author also reports that increasing training data from 500 to 80,000 examples on a wider 0–100,000 distribution barely moved the performance ceiling. That figure is reported in Mutlu’s DEV article, rather than in the detailed values available in the paper record and repository, so it should be treated as the author’s account rather than an independently cross-checked measurement.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Why the revision matters for interpreting the result
Mutlu says an earlier version of the experiment had label leakage during cluster-to-label calibration and described the approach too broadly as unsupervised. The revised account separates training, validation, and test data and uses training labels to calibrate clusters. This correction is important: calibration using labels means the total method is not fully unsupervised, and a held-out test result is meaningful only when test examples do not influence choices or label mapping.
The revised arXiv record identifies version 2 as last revised on 26 September 2026. The repository identifies the September 2026 revision and points to the Git tag paper-v2 as the exact manuscript snapshot; its README also records dependencies and runtime details. Anyone reproducing the reported setup should use that tag rather than the moving main branch: the public repository.
How to read results like these
When comparing variants, first check that the evaluation protocol is comparable. A validation score from a changed bit layout should not be placed on the same leaderboard as held-out test accuracy from the primary setup. Then check the representation and data conditions that can alter what the model sees:
- Bit layout and padding: whether the signal is fixed-width, and where the informative bit sits.
- Transform and subband: which wavelet family and decomposition band produce the features.
- Boundary extension: how edge values are handled during the transform.
- Information availability: whether the LSB is retained or masked.
- Numeric range: whether training and evaluation cover similar magnitudes and bit lengths.
- Choice timing: whether configuration decisions were made using training or validation data rather than the held-out test set.
These checks do not turn the experiment into a general parity benchmark. They help explain why one representation may make a cue easy for a particular model to recover and another may not.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
