Dynamic Range Expansion (DRE)

• Dynamic Range Expansion (DRE) is a set of techniques that extend the range between the smallest and largest distinguishable signals, essential for applications like HDR imaging and ADCs.
• It employs methodologies such as nonlinear signal folding, inverse tone mapping with deep networks, and adaptive quantization to recover lost information and enhance signal fidelity.
• Practical implementations demonstrate significant improvements in dynamic range through hardware-based ADCs and data-driven methods, balancing trade-offs in latency, noise, and reconstruction accuracy.

Dynamic Range Expansion (DRE) refers to a family of signal processing, sensing, and computational techniques that extend the measurable or representable range between the smallest and largest reliably distinguishable signal levels—often far beyond native electronic or numerical constraints. DRE is critical across diverse domains: high dynamic range imaging, high-fidelity data acquisition, quantum metrology, low-precision neural network training, and quantitative phase imaging. Research on arXiv demonstrates a suite of system-level and algorithmic strategies—ranging from nonlinear signal folding, sequential measurement, and data-driven mapping to adaptive quantization—that realize practical and substantial dynamic range expansion.

1. Fundamental Concepts and Definitions

Dynamic range (DR) is formally defined as the ratio of the largest to smallest nonzero value faithfully captured or reconstructed by a system. For a generic signal $L$, this is $DR = L_{\max} / L_{\min}$, with $L_{\max}$ and $L_{\min}$ denoting extremal measurable or representable values. In digital imaging, audio, and analog-to-digital conversion, DR is limited by quantizer resolution and physical front-end constraints.

Classic DR manipulation involves either compressing (via nonlinear mappings such as logarithm or gamma correction) or expanding (reconstructing or encoding information lost to quantization or saturation) the signal’s range. Two crucial ill-posed operations are tone mapping (HDR $\rightarrow$ LDR) and its inverse, expansion/inverse tone mapping (LDR $\rightarrow$ HDR), the latter attempting to restore content lost to clipping and quantization (Marnerides et al., 2018).

2. Algorithmic DRE Techniques in Imaging

In image processing, DRE enables reconstruction of scene radiance exceeding the input encoding’s capacity. Key approaches include:

• Inverse Tone Mapping Networks: For single-image HDR expansion, the ExpandNet model fuses local high-frequency (detail), medium-range (dilation), and global (context) features using a three-branch, fully-convolutional SELU architecture. The output HDR is mapped in an end-to-end, data-driven manner, eschewing upsampling to avoid artifacts. ExpandNet is trained with multi-TMO augmentation and a compound loss that includes both L1 distance and channel-wise cosine similarity, operating quantitatively superior to both classical and deep-learning-based expansion operators across metrics such as SSIM, PSNR, and HDR-VDP-2.2 (Marnerides et al., 2018).
• Logarithmic Model DRE: In the theoretical context of logarithmic gray-level image models, DRE can be framed as finding optimal pointwise homothetic transforms that maximize either (i) the algebraic dynamic range or (ii) the mean dynamic range, based on closed-form solutions for the optimal exponent. These transforms generalize classical gamma correction, yielding a one-parameter family that achieves mathematically optimal expansion under the assumed statistical properties of the image (Patrascu et al., 2014).

Approach	Domain	Key Mechanism
ExpandNet	LDR $\rightarrow$ HDR	Multiscale CNN, no upsampling, multi-TMO loss
Logarithmic	Gray-level images	Optimal exponential homothety

3. Hardware and Signal Acquisition: Modulo Folding and Self-Reset ADCs

State-of-the-art data converters achieve DRE by leveraging modulo (folding) operations:

• Modulo and Self-Reset ADCs: Instead of clipping when the input exceeds the ADC’s reference range, a modulo ADC wraps the input into the allowed interval via $y(t) = (y(t) + \lambda) \bmod 2\lambda - \lambda$. The original, potentially unbounded dynamic range is recovered by algorithmically reconstructing integer fold-counts post-conversion, provided sampling constraints on input slew rate are met. This enables “unlimited” dynamic range in principle, with practical folding factors reported at up to 8$\times$ (prototype board), and more than 100$\times$ (FPGA-based system), with minimal impact on quantization SNR or system area/power (Mulleti et al., 2023, Krishna et al., 2019, Li et al., 27 Nov 2025).
• Recovery Algorithms: Unfolding from modulo samples uses difference-based residue tests (e.g., B$DR = L_{\max} / L_{\min}$0R$DR = L_{\max} / L_{\min}$1, RSoD), iterative sieving, or sparse recovery. Architecture advances (e.g., multi-bit folding, under-compensation digital calibration on an FPGA) achieve both high folding factors (over 100) and high-fidelity operation (SINAD $DR = L_{\max} / L_{\min}$244 dB, ENOB $DR = L_{\max} / L_{\min}$37 bits at 92$DR = L_{\max} / L_{\min}$4 expansion). Trade-offs center around folding latency, loop bandwidth, and the robustness of digital post-processing (Li et al., 27 Nov 2025).

Implementation	Achievable DRE Factor	Bandwidth / Speed	SNR / ENOB
Prototype Board	8$DR = L_{\max} / L_{\min}$5	up to 10 kHz	$DR = L_{\max} / L_{\min}$6
FPGA Platform	$DR = L_{\max} / L_{\min}$7100$DR = L_{\max} / L_{\min}$8	up to 400 kHz	44 dB, 7 bits
UDR-ADC (65 nm CMOS)	%%%%19$L_{\min}$20%%%% per-samp	$L_{\max}$150 kHz	63-75 dB

4. Adaptive and Hybrid DRE in Experimental Measurement

DRE also addresses the sensitivity-limit trade-offs in measurement science:

• Quantum Deamplification: For entanglement-enhanced quantum metrology, extending DR while retaining sensitivity is realized using sequential spin-squeezing (TACT)—“squeeze–encode–squeeze” cycles. A first squeezing pulse prepares an entangled probe, phase encoding rotates collective spin, and a second squeezing (“deamplification”) narrows the effective encoding window, multiplying the range by $L_{\max}$2, where $L_{\max}$3 is a gain parameter set by squeezing duration. Sequential iterations arbitrarily extend DR; hybrid protocols that combine deamplification with quantum amplification yield resilience to detection noise while spanning multi-$L_{\max}$4 phase intervals, near the optimal quantum interferometer bound (Liu et al., 2024).
• Adaptive QPI via ADRIFT: Adaptive dynamic range shift in quantitative phase imaging is accomplished by measuring large optical phase delay (OPD) components conventionally, canceling them optically (spatial light modulator), and boosting sensitivity for residual measurement with dark-field QPI. This two-stage ADRIFT procedure multiplies DR by up to $L_{\max}$56.6$L_{\max}$6 experimentally, with theoretical limits set by SLM quantization and detector capacity. Extensions apply to wide-field, nanoscale imaging (Toda et al., 2020).

5. DRE in Computational and Learning Systems

Dynamic range constraints impact memory and numeric error in modern distributed computation:

• FP8 Neural Training with DRE: In low-precision neural training, such as FP8 (E4M3) optimizers, DRE is applied to optimizer states. Per-group exponentiation (learned $L_{\max}$7) expands the group dynamic range, aligning with FP8 encoding bounds and minimizing quantization error. The data-driven expansion uses $L_{\max}$8 before quantization, and contracts on dequantize via $L_{\max}$9. This technique in the COAT framework reduces memory by 1.54$L_{\min}$0 and quantization MSE by 1.63$L_{\min}$1, while preserving downstream performance across LLM and VLM benchmarks (Xi et al., 2024).

Application	DRE Mechanism	Key Outcome
Quantum metrology	Squeeze–encode–squeeze, hybrid QA+QD	$L_{\min}$22–$L_{\min}$3 DR, sub-SQL
FP8 Neural Train	Per-group exponent DRE, adaptive $L_{\min}$4	$L_{\min}$51.5$L_{\min}$6 memory reduction

6. Multi-Exposure and Sensor Fusion for Extended Dynamic Range

Photographic sensing and 3D vision exploit DRE by aggregating information from multiple exposures:

• Dual-Exposure Stereo: Stereo 3D imaging under large scene dynamic range employs automatic dual-exposure control (ADEC) that dynamically steers each camera of a stereo pair to divergent exposures when scene DR exceeds sensor capability. Features from both exposures are fused using exposure-aware weights, and depth is estimated via a motion-aware stereo network. This pipeline yields up to 160% DR expansion, reduces holes in dark/bright regions, and achieves depth MAE (real robot-vision scenario) of 1.91 m compared to 2.58–2.77 m for other exposure control methods (Choi et al., 2024).

7. Implementation Trade-offs and Limitations

DRE approaches are inherently constrained by hardware (speed, noise, resolution, overshoot in folding circuits), algorithmic complexity (reconstruction from residues, fusion accuracy), and context-specific properties (phase ambiguities in quantum and phase imaging, or representable range in fixed-point formats). In modulo ADCs, the maximal expansion is set by recovery algorithm robustness, fold-event tracking speed, and hardware settling time. In neural quantization, group size trade-offs yield diminishing returns past a threshold; in ADRIFT-QPI, SLM quantization and alignment precision bound feasible DRE. Adaptive and hybrid schemes (e.g., QD+QA in quantum metrology) aim to circumvent regime-specific vulnerabilities while balancing efficiency and error resilience.

• "ExpandNet: A Deep Convolutional Neural Network for High Dynamic Range Expansion from Low Dynamic Range Content" (Marnerides et al., 2018)
• "Unlimited Dynamic Range Analog-to-Digital Conversion" (Krishna et al., 2019)
• "A Hardware Prototype of Wideband High-Dynamic Range ADC" (Mulleti et al., 2023)
• "FPGA-Enabled Modulo ADC with x100 Dynamic-Range Expansion: Hardware Design and Performance Evaluation" (Li et al., 27 Nov 2025)
• "Enhancing Dynamic Range of Sub-Quantum-Limit Measurements via Quantum Deamplification" (Liu et al., 2024)
• "Adaptive dynamic range shift (ADRIFT) quantitative phase imaging" (Toda et al., 2020)
• "COAT: Compressing Optimizer states and Activation for Memory-Efficient FP8 Training" (Xi et al., 2024)
• "Image Dynamic Range Enhancement in the Context of Logarithmic Models" (Patrascu et al., 2014)
• "Dual Exposure Stereo for Extended Dynamic Range 3D Imaging" (Choi et al., 2024)