Symmetric Uniform Quantization Overview

• Symmetric uniform quantization is defined by reconstruction levels arranged with equal spacing and symmetry constraints across various domains.
• It is applied in neural networks and geometric scenarios to ensure zero-centered representations and uniform error distributions.
• Variants such as dithered, periodic, and binary/ternary schemes balance computational efficiency with quantization accuracy.

Symmetric uniform quantization denotes a family of quantization schemes in which reconstruction levels or decision regions are arranged with a uniform spacing rule and a symmetry constraint. In the literature, this phrase covers several closely related but nonidentical constructions: zero-centered signed scalar quantizers with no zero-point for neural-network weights and activations; optimal quadratic quantizers for uniform measures supported on symmetric geometric sets; randomized or dithered quantizers whose error is uniform over a symmetric set; and periodic quantizers on angular domains. Across these settings, the common structural theme is that symmetry is imposed either on the quantizer grid, on the support of the source distribution, or on the error law itself (Kim et al., 2021, Pham et al., 2021, Ling et al., 2023).

1. Formal scope and core meanings

A general quadratic quantization problem starts from a Borel probability measure $P$ on $\mathbb R^d$ and the distortion functional

$$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$

A minimizer is an optimal set of $n$-means, and its geometry is governed by Voronoi regions and the centroid condition

$a=E(X\mid X\in M(a\mid \alpha)).$

This formulation underlies the studies of uniform distributions on an equilateral triangle, on symmetric curves, on polygonal boundaries, and on the stretched Sierpiński triangle (Dettmann et al., 2015, Rosenblatt et al., 2018, Pena et al., 2019, Comez et al., 2016).

In neural-network quantization, “symmetric” usually means that the quantizer is zero-centered, signed, and uses no zero-point. Q-Rater explicitly assumes “layer-wise and symmetric quantization structure for both weights and activations,” with clipping interval $[-Th_c,Th_c]$ and scale $s=Th_c/(2^{q-1}-1)$ (Kim et al., 2021). LG-LSQ is explicitly a linear symmetric uniform quantizer with learned positive scale $a$, while UniQ defines a strict symmetric weight quantizer whose levels are centered around zero through an offset $\alpha=\Delta(N-1)/2$ (Lin et al., 2022, Pham et al., 2021). By contrast, SYQ uses symmetric binary or ternary weight codebooks with learned subgroup scales, but for weights it is not ordinary multi-level uniform quantization in the usual equal-step sense (Faraone et al., 2018).

The phrase “uniform” also changes meaning across domains. On a line segment or circle it refers to normalized length or arc-length measure; on the equilateral triangle it refers to constant area density; on self-similar fractals it refers to equal mass across congruent similarity branches; in randomized vector quantization it can mean that the error is uniform over a prescribed set such as a ball; and in TurboAngle it refers to equal-width angular bins on $[0,2\pi)$ (Rosenblatt et al., 2018, Dettmann et al., 2015, Comez et al., 2016, Ling et al., 2024, Patel, 29 Mar 2026).

Domain	Meaning of symmetry	Meaning of uniformity
DNN weight/activation quantization	Zero-centered signed grid, usually no zero-point	Equal step size in value space
Geometric optimal quantization	Symmetry of support and Voronoi structure	Uniform source measure on support
Dithered/vector quantization	Symmetric support of error law	Constant density over a target set
Angular quantization	Rotational or periodic symmetry	Equal angular bin width

2. Canonical quantizer forms

In LG-LSQ, the $\mathbb R^d$0-bit quantizer is given by

$\mathbb R^d$1

For ReLU activations, the paper states that activations are clamped to $\mathbb R^d$2, so that

$\mathbb R^d$3

Because code $\mathbb R^d$4 is present, zero is represented exactly. The paper further states that the quantizer is inferably mid-tread because it uses rounding to the nearest integer and includes an exact zero reconstruction level (Lin et al., 2022).

Q-Rater keeps the symmetric uniform grid fixed and changes clipping and rounding. It clips weights as

$\mathbb R^d$5

defines

$\mathbb R^d$6

and uses the conventional rounding-to-nearest baseline

$\mathbb R^d$7

Its 1st-order and 2nd-order schemes modify only the rounding boundary, not the uniformly spaced output grid. The paper explicitly emphasizes that the formulation remains symmetric signed quantization for both weights and activations, with no affine offset or zero-point (Kim et al., 2021).

UniQ gives a different strict symmetric form for weights: $\mathbb R^d$8 with

$\mathbb R^d$9

Its reconstruction levels are

$$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$0

so they are uniformly spaced and exactly symmetric around zero. For even $$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$1, the paper explicitly notes that zero is not a reconstruction level for weights. Activations instead use

$$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$2

which is uniform but nonnegative rather than symmetric (Pham et al., 2021).

SYQ clarifies a recurrent distinction. Its weight quantization is symmetric because every positive codebook value has a negative counterpart of equal magnitude, but for weights it is not standard uniform quantization over many evenly spaced levels. Binary weights use $$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$3, ternary weights use $$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$4, and learned positive scales are shared across structured subgroups (Faraone et al., 2018).

3. Symmetry, Voronoi geometry, and optimal quantization of uniform measures

In the geometric literature, symmetric uniform quantization is studied through exact optimal $$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$5-means for highly symmetric supports. On a line segment $$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$6, the optimal set of $$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$7-means is

$$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$8

with

$$V_n:=V_n(P)=\inf \Big\{\int \min_{a\in\alpha} \|x-a\|^2 dP(x) : \alpha \subset \mathbb R^d,\ \operatorname{card}(\alpha) \leq n \Big\}.$$9

This is the classical case where symmetry of the support and uniformity of the source coincide with equal-length Voronoi cells and equally spaced centroids (Rosenblatt et al., 2018).

On the unit circle, the optimal $n$0-means remain uniformly spaced in angle, but the centroids lie on a smaller concentric circle: $n$1 and

$n$2

Uniformity is therefore intrinsic to arc length rather than Euclidean coordinate spacing (Rosenblatt et al., 2018).

For the uniform distribution on the equilateral triangle with vertices

$n$3

the density is constant over the area, the mean is

$n$4

and

$n$5

The equilateral triangle has full dihedral symmetry $n$6, and the paper repeatedly exploits the three medians and three rotations by multiples of $n$7. For $n$8, symmetry yields three equivalent optimal configurations along the medians; for $n$9, the optimal set forms a smaller equilateral triangle with sides parallel to the original; for $a=E(X\mid X\in M(a\mid \alpha)).$0, the numerically verified arrangement is reflection-symmetric, with three equivalent orientations (Dettmann et al., 2015).

The same structural role of symmetry appears on one-dimensional supports embedded in $a=E(X\mid X\in M(a\mid \alpha)).$1. For the boundary of a regular hexagon, the support and law have full dihedral symmetry $a=E(X\mid X\in M(a\mid \alpha)).$2, and for $a=E(X\mid X\in M(a\mid \alpha)).$3 the optimal codebook is built from one prototype side or corner configuration replicated six times. The paper gives, for example,

$a=E(X\mid X\in M(a\mid \alpha)).$4

and an optimal two-mean set

$a=E(X\mid X\in M(a\mid \alpha)).$5

For the semicircular boundary mixture, symmetry is only with respect to the vertical axis $a=E(X\mid X\in M(a\mid \alpha)).$6, so optimal codebooks are mirror-symmetric rather than fully rotationally symmetric. For the ellipse $a=E(X\mid X\in M(a\mid \alpha)).$7, only the major and minor axes remain as reflection symmetries, and even and odd $a=E(X\mid X\in M(a\mid \alpha)).$8 exhibit different optimality patterns (Pena et al., 2019).

The stretched Sierpiński triangle gives a recursive self-similar counterpart. Its support is generated by three similarities of ratio $a=E(X\mid X\in M(a\mid \alpha)).$9, its measure assigns equal mass $[-Th_c,Th_c]$0 to the three first-level pieces, and the mean is again

$[-Th_c,Th_c]$1

For $[-Th_c,Th_c]$2, the unique optimal set is the set of centroids of all level-$[-Th_c,Th_c]$3 basic triangles: $[-Th_c,Th_c]$4 For intermediate $[-Th_c,Th_c]$5, optimal sets are obtained by replacing selected level-$[-Th_c,Th_c]$6 centroids by scaled copies of optimal 2-point or 3-point local codebooks. The quantization dimension exists and equals $[-Th_c,Th_c]$7, but the $[-Th_c,Th_c]$8-dimensional quantization coefficient does not exist (Comez et al., 2016).

These examples show a consistent pattern. Symmetry reduces the search space, but it does not force globally equal Euclidean spacing. Corner singularities, curvature, or self-similar refinement can preserve symmetry while producing nontrivial local structure (Dettmann et al., 2015, Rosenblatt et al., 2018).

4. Randomized, dithered, and periodic variants

A second major interpretation of symmetric uniform quantization comes from randomized constructions. In universal quantization for neural compression, the central identity is

$[-Th_c,Th_c]$9

so additive uniform noise can be implemented at test time as randomized shifted-lattice quantization. This removes the train/test mismatch between differentiable additive-noise training and discrete inference. The underlying quantizer is uniform with equal-width bins, but for fixed dither realization it is not necessarily centered at zero; it is a randomly shifted uniform lattice (Agustsson et al., 2020).

The deep image compression literature studies the same unit-step rounding quantizer

$s=Th_c/(2^{q-1}-1)$0

together with training-time approximations such as additive uniform noise,

$s=Th_c/(2^{q-1}-1)$1

and universal quantization,

$s=Th_c/(2^{q-1}-1)$2

Here the operative test-time quantizer is the standard integer-lattice uniform quantizer, while symmetry appears through the centered interval $s=Th_c/(2^{q-1}-1)$3 used for noise or dither (Tsubota et al., 2023).

Higher-dimensional generalizations make the error law itself uniform over a symmetric set. One construction introduces shift-periodic vector quantizers whose error is uniformly distributed over an arbitrary bounded measurable set $s=Th_c/(2^{q-1}-1)$4, and in particular over the unit $s=Th_c/(2^{q-1}-1)$5-ball $s=Th_c/(2^{q-1}-1)$6. With subtractive dithering, the error becomes exactly independent of the input and equal to the prescribed law $s=Th_c/(2^{q-1}-1)$7 (Ling et al., 2023). A closely related construction, rejection-sampled universal quantization, starts from subtractive dithered lattice quantization

$s=Th_c/(2^{q-1}-1)$8

and rejects proposals until the error lands in a target set $s=Th_c/(2^{q-1}-1)$9. The accepted error then satisfies

$a$0

For $a$1, the error is uniform over a symmetric ball, and the paper gives

$a$2

This is a randomized vector analogue of symmetric uniform quantization with a ball-shaped error law (Ling et al., 2024).

Blind-Adaptive Quantizers attack a different problem: source–quantizer mismatch. The paper keeps a bounded uniform quantizer over a symmetric interval $a$3, but inserts a preprocessing map

$a$4

The modulo-folded signal always lies in $a$5, and for sufficiently large amplification $a$6 the folded distribution approaches $a$7 for Gaussian, exponential, and uniform source families. This does not redefine the quantizer; it improves the operating conditions of a bounded symmetric uniform quantizer (Chemmala et al., 2024).

TurboAngle extends the notion of uniform quantization to a periodic angular domain. After a random diagonal sign matrix and normalized Fast Walsh–Hadamard transform,

$a$8

each consecutive pair $a$9 is converted to polar coordinates and the angle is quantized by

$\alpha=\Delta(N-1)/2$0

The codebook is

$\alpha=\Delta(N-1)/2$1

with equal angular width $\alpha=\Delta(N-1)/2$2. This is uniform and symmetric on a periodic circle rather than on a zero-centered real interval (Patel, 29 Mar 2026).

5. Neural, communication, and coding applications

In low-bit neural-network inference and training, symmetric uniform quantization is valued because it eliminates zero-point arithmetic and keeps the hardware interface simple. Q-Rater uses a plain symmetric signed layer-wise uniform quantizer and argues that low-bit post-training quantization should optimize clipping and rounding directly against task loss rather than rely on convex surrogates. The reported low-bit gains are large: for ResNet-18 on ImageNet at $\alpha=\Delta(N-1)/2$3 bits, the MSE baseline gives $\alpha=\Delta(N-1)/2$4 whereas Q-Rater with Bayesian optimization gives $\alpha=\Delta(N-1)/2$5; for ResNet-32 on CIFAR-10 at $\alpha=\Delta(N-1)/2$6 bits, the MSE baseline gives $\alpha=\Delta(N-1)/2$7 and Q-Rater full gives $\alpha=\Delta(N-1)/2$8 (Kim et al., 2021).

LG-LSQ addresses quantization-aware training. It introduces SSG for scale learning, ASR as a differentiable soft-round surrogate, and MDE for reducing the mismatch between full-precision and quantized values. The paper reports full-precision baseline accuracy in various 3-bit networks including ResNet18, ResNet34, and ResNet50, and less than $\alpha=\Delta(N-1)/2$9 accuracy drop for $[0,2\pi)$0-bit weights and $[0,2\pi)$1-bit activations in lightweight models such as MobileNetV2 and ShuffleNetV2 (Lin et al., 2022).

UniQ shows that a strict symmetric quantizer can unify multi-bit quantization and 1-bit binarization. For ImageNet, the paper reports, for example, ResNet-34 accuracies of $[0,2\pi)$2 at $[0,2\pi)$3, $[0,2\pi)$4 at $[0,2\pi)$5, and $[0,2\pi)$6 at $[0,2\pi)$7, compared with LSQ values of $[0,2\pi)$8, $[0,2\pi)$9, and $\mathbb R^d$00. It attributes the gains to the combined result of the symmetric quantizer and optimal initialization (Pham et al., 2021). SYQ, in turn, keeps activations uniform and nonnegative but uses subgroup-wise learned symmetric binary or ternary weight codebooks. On ImageNet, it reports, for example, ResNet-50 accuracies of $\mathbb R^d$01 for $\mathbb R^d$02-$\mathbb R^d$03, $\mathbb R^d$04 for $\mathbb R^d$05-$\mathbb R^d$06, and $\mathbb R^d$07 for $\mathbb R^d$08-$\mathbb R^d$09 (Faraone et al., 2018).

Deep image compression uses a standard unit-step uniform quantizer at test time, but the best training surrogate is architecture dependent. Across three architectures and two datasets, the combination using universal quantization for the entropy model and differentiable soft quantization for the decoder is reported as a comparatively good choice, with average BD-rate $\mathbb R^d$10 across the three architectures (Tsubota et al., 2023).

In federated learning, symmetric clipped uniform quantization is used for communication reduction. The clipping interval is explicitly $\mathbb R^d$11, the number of levels is $\mathbb R^d$12, and stochastic quantization adds noise from

$\mathbb R^d$13

The paper reports, for a model with 80,848 weights, about $\mathbb R^d$14 communication saving for $\mathbb R^d$15-$\mathbb R^d$16-$\mathbb R^d$17-$\mathbb R^d$18 bits, about $\mathbb R^d$19 for $\mathbb R^d$20-$\mathbb R^d$21-$\mathbb R^d$22-$\mathbb R^d$23, about $\mathbb R^d$24 for $\mathbb R^d$25-$\mathbb R^d$26-$\mathbb R^d$27-$\mathbb R^d$28, and about $\mathbb R^d$29 for $\mathbb R^d$30-$\mathbb R^d$31-$\mathbb R^d$32-$\mathbb R^d$33 (Bozorgasl et al., 2024).

In channel coding, uniform quantization is used as a hardware simplification inside low-resolution LDPC decoders. The proposed symmetric uniform quantizer is

$\mathbb R^d$34

The paper reports that the uniformly quantized decoder causes only minor performance degradation within $\mathbb R^d$35 dB compared to the non-uniform alternative, while reducing node-operation complexity approximately by half (Mohr et al., 2022).

In information theory, uniform output quantization is studied through saturation and wrapping. For wrapping quantization with levels $\mathbb R^d$36, the capacity-achieving input is an equiprobable $\mathbb R^d$37-point constellation

$\mathbb R^d$38

For arbitrarily many uniform quantization levels, the paper further shows that the gap between Ihara’s upper and lower bounds is only $\mathbb R^d$39 bits (0901.2545).

6. Distinctions, misconceptions, and recurrent trade-offs

A central distinction is that symmetric does not mean the same thing as uniform. A quantizer can be symmetric without being ordinary equal-step multi-level quantization, as in SYQ’s symmetric binary and ternary weight codebooks (Faraone et al., 2018). Conversely, a quantizer can be uniform without being zero-centered, as in universal quantization with random dither or in activations quantized over a nonnegative range (Agustsson et al., 2020, Pham et al., 2021).

Zero handling is another point of divergence. In LG-LSQ and Q-Rater, zero is represented exactly because the integer code $\mathbb R^d$40 is included and the quantizer has no zero-point (Lin et al., 2022, Kim et al., 2021). UniQ explicitly notes the opposite for even-level weight quantization: strict symmetry around zero implies that zero is not itself a weight reconstruction level (Pham et al., 2021). In the LDPC decoder setting, the exchanged message alphabet is symmetric in sign-magnitude form but has no zero output symbol at all (Mohr et al., 2022). The literature therefore does not support a single universal rule that “symmetric quantizers always include zero.”

Uniformity itself is domain dependent. On a line segment it means equal subintervals; on a circle it means equal arc-length sectors; on an equilateral triangle or polygonal boundary it means a uniform source measure whose optimal Voronoi cells are shaped by symmetry but also by corners and curvature; in vector quantization it can mean error uniform over a ball; in TurboAngle it means equal angular bins on a periodic manifold (Rosenblatt et al., 2018, Dettmann et al., 2015, Ling et al., 2023, Patel, 29 Mar 2026). It follows that “uniform” is not restricted to equal-width bins on the real line.

Several papers also identify a low-bit optimization tension. Q-Rater states that for low-bit post-training quantization, non-convex optimization is unavoidable, because task loss is non-convex and weight MSE correlates poorly with task loss (Kim et al., 2021). Blind-Adaptive Quantizers identify a different mismatch: a bounded symmetric uniform quantizer performs best when the input distribution is close to uniform on its support, motivating modulo-based preprocessing (Chemmala et al., 2024). The geometric papers show an analogous phenomenon in a non-neural setting: even under perfect support symmetry, corners or self-similar hierarchy can prevent globally uniform cell shapes and can even destroy the existence of the quantization coefficient, as in the stretched Sierpiński triangle (Comez et al., 2016).

Taken together, these results define symmetric uniform quantization not as a single formula but as a design principle. The principle is to exploit symmetry, regular spacing, or uniform source structure to simplify representation, hardware, or analysis, while accepting that exact implementation details vary sharply between scalar, vector, geometric, periodic, and neural settings.