Integrate-and-Fire TEM (IF-TEM)

• IF-TEM is an event-driven analog-to-digital converter that encodes signals as time-stamped spikes when integrated voltages cross a preset threshold.
• It leverages analog integration and nonuniform sampling to enable asynchronous, energy-efficient, and sparse signal processing ideal for low-power applications.
• Enhanced CIF-TEM variants implement dynamic-range windowing for quantization, significantly reducing reconstruction MSE and bit rate while preserving signal fidelity.

An Integrate-and-Fire Time-Encoding Machine (IF-TEM) is a nonuniform, event-driven analog-to-digital encoding architecture that produces a time-sequence of events ("spikes") driven by input signal integrals crossing a threshold. Unlike traditional amplitude-sampling ADCs operating under synchronous clock domains, the IF-TEM exploits analog integration to realize an asynchronous, energy-efficient, and sparse encoding, making it highly suitable for low-power and sub-Nyquist signal processing. The Compressed Integrate-and-Fire Time-Encoding Machine (CIF-TEM) is an enhancement of the canonical IF-TEM that exploits the statistical stationarity of the inter-spike intervals for significant analog-to-digital compression prior to quantization, yielding substantial reductions in reconstruction mean squared error (MSE) and bit rate for a fixed reconstruction fidelity (Tarnopolsky et al., 2022).

1. Canonical IF-TEM Architecture and Sampling Principle

The standard IF-TEM consists of an adder, an integrator (with gain parameter $1/\kappa$), a threshold detector set at $\delta$, a time-stamp generator, and a reset mechanism. The continuous-time model can be expressed as: $$\dot{y}(t) = \frac{1}{\kappa}(b + x(t)), \quad y(t_n) = 0$$ where $x(t)$ is the analog input, $b$ is a bias ensuring positivity, and $\kappa > 0$.

A spike is emitted at the next time $t_{n+1} > t_n$ such that: $y(t_{n+1}) = \delta$ which is equivalent to the integral condition: $$\frac{1}{\kappa} \int_{t_n}^{t_{n+1}} [b + x(\tau)]\,d\tau = \delta$$

Each inter-spike interval $T_n = t_{n+1} - t_n$ naturally encodes a local signal amplitude via: $$x_n = \int_{t_n}^{t_{n+1}} x(\tau)\,d\tau = \kappa \delta - b T_n$$ For signals bounded as $|x(t)| \leq c < b$, the firing interval is bounded: $$\Delta t_{\min} = \frac{\kappa\delta}{b + c} \leq T_n \leq \frac{\kappa\delta}{b - c} = \Delta t_{\max}$$

Perfect recovery of a $2\Omega$-bandlimited input is guaranteed if $\Delta t_{\max} < \pi/\Omega$; i.e., firing rate exceeds the Nyquist rate (Tarnopolsky et al., 2022).

2. Stationarity of Inter-Spike Intervals and Motivations for Compression

The output of the integrator inherently acts as a low-pass filter, rendering the distribution of inter-spike intervals $\{T_n\}$ sharply concentrated within a much narrower sub-range of the full dynamic range $[\Delta t_{\min}, \Delta t_{\max}]$: $\sigma \ll (\Delta t_{\max} - \Delta t_{\min})^2$ where $\sigma$ denotes the variance of $\{T_n\}$. This stationarity invites analog compression techniques: most $T_n$ lie close to a mean value, with only rare excursions. Uniform quantization across the entire dynamic range is thus highly inefficient; more efficient representation is possible by first localizing $T_n$ into adaptively or statically defined windows, followed by fine quantization of the small residuals (Tarnopolsky et al., 2022).

3. CIF-TEM Analog Compression: Dynamic-Range Windowing and Encoding

The CIF-TEM algorithm subdivides the dynamic range $\Delta T = \Delta t_{\max} - \Delta t_{\min}$ into $L$ windows $\{W_i\}$, each of size $\Delta T / L$. For each $T_n$:

• Window index: $$i_n = \left\lfloor \frac{T_n - \Delta t_{\min}}{\Delta T / L} \right\rfloor$$
• Residual: $$r_n = T_n - [\Delta t_{\min} + i_n \cdot (\Delta T / L)],\quad r_n \in [0, \Delta T / L)$$
• Residual is quantized with $K$ uniform levels over $[0, \Delta T / L)$ (step size $\Delta_C = (\Delta T/L)/K$).

This produces a codeword $(i_n, q_n)$ per event, where $q_n$ is the quantization index.

CIF-TEM provides two instantiations:

• CCIF-TEM (Constant Compression IF-TEM): $L$ is fixed based on prior variance estimate, with $L =\lceil \Delta T/(2\sqrt{\sigma}) \rceil$.
• DCIF-TEM (Dynamic Compression IF-TEM): $L_n$ is updated online with a sliding window variance estimator and is adjusted every $\ell$ events.

For both, Popoviciu’s inequality $\sigma < (\Delta T)^2/4$ ensures $L > 1$, and the quantization step is always strictly finer than in classical uniform quantization.

4. Decoding and Signal Reconstruction Methods

Given the compressed codewords $(i_n, q_n)$, each inter-spike interval is reconstructed as: $$\hat{T}_n = \Delta t_{\min} + \left(i_n + \frac{q_n + 1/2}{K}\right) \frac{\Delta T}{L}$$ And the corresponding amplitude estimate is: $\hat{x}_n = \kappa\delta - b\hat{T}_n$

For recovery of the original signal, standard irregular sampling techniques for bandlimited functions are deployed, e.g., frame-based reconstruction or irregular sinc interpolation, exploiting pairs $(\hat{x}_n, t_n)$: $$\hat{x}(t) = \sum_n \hat{x}_n \cdot \mathrm{sinc}(\Omega(t - t_n))$$ Alternatively, one may solve for Fourier coefficients $\{d_k\}$ via frame equations if required.

For the quantization error, since the windowed quantization step $\Delta_{CCIF} = \Delta_{IF}/L$, the MSE is reduced by approximately $20\log_{10}(L)$ dB (Tarnopolsky et al., 2022).

5. Performance Metrics and Empirical Results

Empirical evaluation over 100 random $2\Omega$-bandlimited signals ($|x(t)| \leq c$) with oversampling factor $\approx 3.5$, and $K$-level quantization (bits per interval: $\log_2 K$), yields the following improvement in mean-square error (MSE), given equal sample count:

$K$	IF-TEM MSE (dB)	CCIF-TEM MSE (dB)
8	–28	–42
10	–32	–46
12	–35	–50

The MSE improvement is $5$–$20$ dB for the same number of samples and up to $7\%$ additional encoding bits for the window indexes. Conversely, CIF-TEM can save $1$–$2$ quantizer bits ($10$–$20\%$ bit-rate reduction) for the same reconstruction MSE (Tarnopolsky et al., 2022).

6. Advantages, Limitations, and Algorithmic Implications

Advantages:

• CIF-TEM directly leverages IF-TEM's stationarity to achieve efficient analog compression prior to quantization.
• Substantial MSE reductions for fixed bit-rate and sample count, or conversely, significant bit-rate reduction for a fixed distortion level.
• Preserves the asynchronous, event-based, and energy-efficient character of IF-TEMs.
• Provides both fixed (CCIF) and adaptive (DCIF) schemes, covering diverse application requirements.

Limitations:

• Requires knowledge of, or the capability to estimate, the variance $\sigma$ of the inter-spike intervals either a priori (CCIF) or online (DCIF).
• Introduces coding overhead to track and transmit the window index ($i_n$), although this is typically sparse in time.
• Signal reconstruction from irregular samples remains an off-line process in the framework described, requiring additional processing.

7. Context within IF-TEM and Event-Based ADC Research

CIF-TEM is a significant advance in event-driven ADC architectures, enhancing the bitrate/distortion efficiency by integrating analog compression within IF-TEM encoding. This approach reflects a broader paradigm shift in ADC design, focusing on asynchronous, clockless, and signal-adaptive representations in analog-to-digital conversion. The method is compatible with further innovations such as adaptive IF-TEMs and hybrid estimators, provided that the key assumption of stationarity (or near-stationarity) of inter-spike intervals is maintained. CIF-TEM thus offers a flexible, low-power, and highly compressive alternative for ADC applications where both energy and bandwidth cost are paramount (Tarnopolsky et al., 2022).