Fractional-order hardware has closed the wrong gap, and the open one sits inside your electrode

What the work claims

This is a single-author topical review with original analysis, and its central claim is deliberately contrarian: the fractional-order neuromorphic literature treats the fractional order (the exponent alpha of the power-law memory kernel) as the hard design variable, and that worry is largely obsolete. Fabricated constant-phase elements, the passive devices whose impedance follows a fractional power law, already span orders from 0.33 to 0.94, covering the task-optimal orders that two independent research groups report on shared benchmarks. What remains open is not an order gap but a frequency-band gap: no thin-film or solid-state device integrable in a neuromorphic process has been characterized below roughly 10 Hz, while the adaptation neuromorphics wants to emulate runs from about a kilohertz down toward hundredths of a hertz¹. The only devices that work at that corner are discrete electrochemical double-layer components measured in millifarads, and the field's single end-to-end fractional neuron demonstration was built from supercapacitor stacks, not from anything a fab can make.

How it works

A fractional derivative of order alpha acts on a signal through a power-law memory kernel: the present output depends on the entire past, with weights decaying as a power of elapsed time, tuned by one continuous parameter. That is exactly what a neuron with power-law adaptation needs, and exactly what is expensive: an integer-order circuit buys each additional timescale with one more state variable, while the fractional operator's cost is the retained history itself. The review's key derivation quantifies it. Holding the truncation error of a discretized fractional derivative below a tolerance epsilon requires a retained history of N = (1 / (epsilon |Gamma(1 - alpha)|))^(1/alpha) samples. The exponent 1/alpha is the result the rest of the review turns on, and the author verified its asymptotic form against exact partial sums at 100,000 samples to five significant figures for alpha of 0.15, 0.5 and 0.9¹. Read the table and the cost changes character, not just size: at one percent truncation error, alpha = 0.9 needs a 14-sample window (a shift register), alpha = 0.5 needs about 3,200 samples (a design decision), alpha = 0.3 needs of order 10^6, and alpha = 0.15 needs 10^13, which the author bluntly calls non-viable at any sample rate one would build.

That matters because the fractional orders reported for cortical neurons sit precisely in the forbidden zone: about 0.15 for cortical pyramidal cells, with companion models of biological adaptation calling for 0.2 to 0.3. A second, independent limit closes the door further. In fixed point, the kernel weights decay as j^-(1+alpha), so beyond some index every weight rounds to zero; a longer buffer multiplies stored samples by zero. The crossing of the truncation bound and this underflow bound decides whether any practical word length can serve a given order at all, and for the direct digital form the answer fails below roughly alpha = 0.27. The review sorts the published hardware into three strategies around these two limits and notes a fourth, diffusive representations (recasting the fractional operator as a bank of ordinary first-order relaxations), that the numerical literature has developed and the hardware literature has largely ignored¹.

On the device side, the survey's conclusion is the reverse of the field's assumption. Constant-phase thin-film elements hold their fractional impedance from about 100 Hz upward; electrochemical double-layer electrodes, whose porous ion transport is a physical power-law memory, work down into the millihertz range, but only as discrete components; and nothing integrable has been characterized in the roughly three decades in between. One caveat the author states carefully: four of the six tabulated thin-film papers report a band whose lower endpoint coincides with their instrument's sweep window, so the gap may partly reflect a measurement nobody has bothered to make rather than a physical impossibility¹.

Where a skeptic should push

The most load-bearing assumption is that task-optimal orders, established almost entirely in simulation, translate into requirements on physical devices. Simulations run in discrete time, where the operative parameter is retained history in samples, not a bandwidth; mapping a simulated order onto a fabricated element's frequency response requires an assumed timestep, and the review says so in its own figure caption. The strongest quantitative claim in the review, the order-0.5 optimum on the cart-pole benchmark, comes from the author's own group, and while an independent group reproduces the same optimum on a harder variant of the same task, both efforts draw on the same NSF grant, a dependence the author discloses rather than hides¹.

Second, the worst-case bounds are ceilings, not specifications, and the author is admirably clear about this: on the spoken-digit task at alpha = 0.5, classification accuracy saturates at a retained history of about 150 samples against a bound of 3,200, more than an order of magnitude cheaper. But the counter-case is in the same review: in a fractional integrate-and-fire neuron at alpha = 0.2, deleting the accumulated history even after 100 seconds still changed the firing rate substantially. Classification forgives a truncated tail; dynamics does not. Whether your application is a classifier or a behaving model changes the affordable history by more than an order of magnitude, and the field habitually borrows conclusions from the forgiving case. Finally, the supercapacitor-based end-to-end demonstration, validated against recordings from weakly electric fish, is an existence proof built from bench parts, not a path to silicon.

Fractional memory at the tissue-silicon junction

For microelectrode array instrumentation the review reads backwards, and that is its value here. The device the review treats as a problem, the electrochemical double-layer electrode, is the front end of every extracellular recording ever made. An electrode in saline is a constant-phase element by construction: a broad distribution of interfacial RC time constants, set by porous ion transport, which is why electrode impedance has a fractional frequency response and why low-frequency recording drifts the way it does. The three-decade gap the review maps, roughly 10 Hz down to hundredths of a hertz, is not an exotic neuromorphic specification. It is the band where organoid and slice preparations put their slow envelopes: network-burst amplitude waxing and waning over seconds, infra-slow oscillations, baseline wander from electrode drift and from the tissue itself. The instrumentation connecting living tissue to silicon already contains the only kind of fractional-order device physics that works there.

The opportunity is to stop fighting that and use it. A front end whose adaptation or baseline normalization is implemented as a physical power-law relaxation, rather than as a bank of high-pass filters with hand-set time constants, would track slow signal statistics with one parameter instead of many, and the review's analysis of why you cannot simply clock such a circuit a thousand times faster, the timescale is set by the physical input, and a tissue input cannot be replayed faster without reinstating the storage the device was meant to remove, applies verbatim to tissue. The threat is the same physics wearing its other face. Adaptation computed through the electrode is indistinguishable from artifact generated at the electrode: if your fractional-order element and your recording electrode share a mechanism, drift compensation and drift corruption are the same process, and validating which one you got becomes a calibration problem most labs are not equipped for. There is also an obsolescence warning for the digital route: at the cortical descriptive order near 0.15, the review's arithmetic says faithful digital emulation at biological wall-clock timescales costs 10^13 samples of history, so any near-term analog-adaptive front end will be built from discrete electrochemistry or not at all, and the vendors who understand the double-layer physics as a compute resource rather than a parasitic will own that corner.

The most useful single habit the review hands to array designers is its insistence on separating the descriptive order, what a biological neuron actually does, near 0.15, from the task-optimal order, what a given workload wants, which three groups place in a band of roughly 0.3 to 0.9, with at least one benchmark where the tuned optimum is the integer-order limit, fractional dynamics switched off. A front-end adaptation circuit should be specified in seconds and hertz against the statistics of the preparation it will see, not in alpha against a target model. Teams that specify in alpha, the review's evidence says, are quoting a number the devices can already hit and the band cannot.

The bottom line

Established, with unusually candid caveats: the digital cost of fractional-order dynamics scales as epsilon^(-1/alpha), making the direct form unaffordable at the cortical order near 0.15; fixed-point underflow independently excludes the direct form below roughly alpha = 0.27; fabricated constant-phase devices cover orders 0.33 to 0.94 but are characterized only from about 100 Hz up; and no integrable device has been characterized in the three low-frequency decades where biological adaptation lives. Hypothesis: that the order gap is closed and the band gap is the real frontier. What would confirm or break it is almost embarrassingly cheap: point a standard low-frequency impedance spectrometer at the existing thin-film fractional capacitors and publish below 10 Hz. If the bands extend, the review's central claim softens; if they do not, the gap is physical. For MEA instrumentation, the actionable reading stands either way: the sub-10 Hz band is an unsolved component problem, the electrode interface is already a fractional-order device, and adaptation requirements for the acquisition chain should be written in time, not in Greek letters.