Threshold Current Extraction Methods Compared

Scope

This article compares the four standard methods for extracting threshold current $I_\text{th}$ from a semiconductor laser diode LIV curve: the two-segment linear fit, the second-derivative peak, the first-derivative-of-log inflection point, and the fixed-power threshold. All four methods are applied to a single representative LIV dataset to demonstrate the numerical differences they produce. The conditions under which each method is preferred are discussed.

Why the choice matters

Threshold current values reported in the literature for nominally similar devices vary by 10–30% across publications, and the dominant cause of this spread is inconsistent extraction methodology rather than device-to-device variation. A semiconductor laser does not have a single, mathematically unambiguous threshold current — it has a gradual transition region in which spontaneous emission rolls off and stimulated emission turns on. Any extraction method places the threshold somewhere within this transition region, and different methods place it at different points.

The implications:

• $I_\text{th}$ values from different methods are not interchangeable
• Comparisons across publications require matched methodology
• Extracted parameters that depend on $I_\text{th}$ (notably $T_0$, slope efficiency, wall-plug efficiency at fixed power above threshold) inherit the methodological choice
• Method selection should be reported alongside the value

The four methods reviewed here cover the substantial majority of published practice. The relations between them are predictable: for a clean LIV curve, the four methods agree to within $\sim 1$ mA for a typical telecom DFB, but the rank order is consistent across devices.

The transition region

Above threshold, the optical output power $P$ versus drive current $I$ is well-approximated as linear:

$$P(I) \;=\; \eta_s (I - I_\text{th}), \qquad I \gg I_\text{th},$$

with slope efficiency $\eta_s$ and an extrapolated intercept $I_\text{th}$. Below threshold, $P$ is approximately linear in $I$ as well, with a much smaller slope set by spontaneous emission outcoupling.

In the transition region near $I_\text{th}$, $P(I)$ smoothly curves between the two linear regimes. The width of this transition is determined by the spontaneous emission coupling factor $\beta$ — the fraction of spontaneously emitted photons that couple into the lasing mode. For a typical Fabry–Pérot semiconductor laser, $\beta \sim 10^{-5}$ to $10^{-4}$, producing a sharp transition of width $\Delta I \sim 0.1$ mA. For high-$\beta$ devices (microcavity lasers, single-photon-emitter cavities), the transition can extend over decades of current.

The four extraction methods locate the threshold at different points within this transition.

Method 1: Two-segment linear fit

Two independent linear fits are performed: one to the sub-threshold region (gentle slope, near-zero output), and one to the above-threshold region (steep slope, linear output). The intersection of the two fit lines is taken as $I_\text{th}$.

Procedure:

1. Identify the lasing region — typically $I > 1.2 I_\text{th}^\text{est}$, where $I_\text{th}^\text{est}$ is a rough visual estimate. Fit a linear function $P_\text{above}(I) = \eta_s (I - I_a)$ to this region.
2. Identify the sub-threshold region — typically $I < 0.7 I_\text{th}^\text{est}$. Fit a linear function $P_\text{below}(I) = a I + b$.
3. Solve for the intersection: $I_\text{th} = (a I_a)/(a - \eta_s) - b/(a - \eta_s)$. In the limit $a \to 0$, this simplifies to $I_\text{th} = -b/\eta_s + I_a$.

For typical devices where the sub-threshold slope is much smaller than the above-threshold slope ($a \ll \eta_s$), the two-segment intersection is well-approximated as $I_\text{th} \approx I_a$ — that is, the linear extrapolation of the lasing region back to zero output power.

This method is the most common in practice and is the implicit assumption in many published laser characterization tools.

Advantages. Simple to compute, robust to measurement noise, intuitive interpretation, requires only modest LIV resolution near threshold.

Disadvantages. Sensitive to the choice of fitting ranges; substantial dependence on where the lasing region is taken to begin. For devices with curvature extending well above the kink (high-$\beta$ lasers, thermally compromised devices), the extracted $I_\text{th}$ depends strongly on the lower bound of the fitting range.

Method 2: Second-derivative peak

The optical output power is differentiated twice with respect to current. The threshold is taken at the current where $d^2 P / d I^2$ reaches its maximum value.

Mathematically this corresponds to the inflection point of the slope itself — the point of maximum curvature in the transition. For an ideal step-function transition between the two linear regimes, $d^2 P / d I^2$ would be a delta function at the kink; for a real device with finite transition width, it is a smooth peak.

Procedure:

1. Smooth the LIV data, typically with a Savitzky–Golay filter or a moving-average window.
2. Numerically differentiate twice with respect to $I$ (or fit a spline and take the analytical second derivative).
3. Identify the peak of $d^2 P / d I^2$ and locate its position $I_\text{th}$.

Advantages. No fitting range to choose — the method operates directly on the data. Mathematically well-defined and unambiguous. Particularly useful for devices with significant sub-threshold curvature where the two-segment method's fitting range becomes critical.

Disadvantages. Highly sensitive to measurement noise — taking two numerical derivatives amplifies noise by orders of magnitude. Requires fine current step size near threshold (typically $\leq 0.1 \cdot I_\text{th}$) for stable identification of the peak. The smoothing filter parameters influence the extracted threshold value.

Method 3: First-derivative-of-log inflection

The natural logarithm of optical output power is differentiated once with respect to current. The threshold is taken at the inflection point of $d (\ln P) / d I$, equivalently where the second derivative $d^2 (\ln P) / d I^2$ changes sign.

This method exploits the fact that $\ln P$ versus $I$ has a characteristic S-shape near threshold, with the inflection point identifying the steepest portion of the log-power transition. The method is particularly useful for low-power LED-like operation below threshold transitioning to lasing operation above.

Procedure:

1. Compute $\ln P_i$ for each measured point.
2. Smooth and numerically differentiate to obtain $d (\ln P) / d I$.
3. Locate the maximum of $d (\ln P) / d I$, which corresponds to the inflection point of $\ln P$.

Advantages. Robust for measurements where the sub-threshold output is non-negligible. Less sensitive to noise than the second-derivative method (one fewer differentiation). Well-defined for high-$\beta$ devices where the transition is broad.

Disadvantages. Requires LIV data with positive output power throughout the measurement range — measurements that drop to noise floor or detector dark in the sub-threshold region cannot be log-transformed. Sensitive to background offset; baseline subtraction is required before log transformation.

Method 4: Fixed-power threshold

The threshold is taken at the current that produces a fixed reference output power, typically 0.1 mW or 1 mW depending on device class.

Procedure:

1. Choose a reference power $P_\text{ref}$ appropriate to the device class.
2. Find the current $I_\text{th} = I(P = P_\text{ref})$, by interpolation if necessary.

Advantages. Trivial to compute. Reproducible — does not depend on fitting choices or differentiation parameters. Useful in production-test environments where consistency across thousands of devices is more important than absolute accuracy.

Disadvantages. Not a real threshold by any physical definition; the extracted $I_\text{th}$ depends on the choice of reference power. Cannot be compared across publications unless $P_\text{ref}$ is matched. Strongly device-dependent — a 0.1 mW reference is meaningful for a low-power telecom DFB but produces poorly-defined results for a 1 W pump diode.

Worked example

A clean LIV curve from a 1310 nm Fabry–Pérot InP laser at 25 °C is used to compare the four methods. Drive current is swept from 0 to 30 mA in 0.5 mA steps with optical power measured to $\pm 0.5\%$ relative accuracy.

Applying each method:

Two-segment linear fit. Linear fit on $I \in [12, 26]$ mA: $P = 0.525 (I - 7.33)$ in mW with $R^2 = 0.9999$. Sub-threshold fit on $I \in [0, 7]$ mA: $P = 0.000210 I - 0.00002$ in mW. Intersection: $I_\text{th} = 7.34$ mA.

Second-derivative peak. Numerical differentiation with a 5-point Savitzky–Golay smoothing filter. The $d^2 P / d I^2$ trace peaks at $I = 8.5$ mA, with peak height 0.45 mW/mA². $I_\text{th} = 8.5$ mA.

First-derivative-of-log inflection. $d (\ln P) / d I$ peaks at $I = 8.0$ mA. $I_\text{th} = 8.0$ mA.

Fixed-power threshold at $P_\text{ref} = 0.1$ mW. Interpolating, $I(P = 0.1 \text{ mW}) = 7.6$ mA.

Summary:

The four methods span 1.16 mA — roughly 15% of the value. The rank order (two-segment lowest, second-derivative highest) is consistent across devices for clean LIV curves.

For this dataset, the two-segment value coincides nearly exactly with the linear-extrapolation intercept $I_\text{th}^*$ from the slope efficiency extraction. The second-derivative value coincides nearly exactly with the kink in the LIV — where the curve visibly transitions from sub-threshold to lasing.

Choosing among methods

The four methods locate the threshold at slightly different physical features of the transition. For a given application:

For new measurements, the two-segment linear fit is the most defensible default. It is the implicit assumption in much of the published literature on slope efficiency and characteristic temperature, and the extracted $I_\text{th}$ value enters cleanly into the corresponding formulas. The second-derivative method is the appropriate companion measurement when the visible kink location is the property of interest.

The fixed-power method should be reported as a fixed-power threshold and not labeled simply as "$I_\text{th}$" — the value depends on $P_\text{ref}$, and the convention is publication-dependent.

Sources of error common to all methods

Independent of the extraction method, several measurement issues bias all four methods in the same direction.

Self-heating in CW measurements. At currents approaching $I_\text{th}$, self-heating elevates the active region temperature, increasing $I_\text{th}$. For CW measurements, the apparent threshold is somewhat higher than the true low-temperature value. Pulsed measurement at $\leq 0.1\%$ duty cycle eliminates this contribution.

Optical detector noise floor. Sub-threshold output power is often below the detector noise floor, which puts $P = 0$ in the sub-threshold region of the LIV data. The two-segment fit handles this correctly (the sub-threshold fit slope approaches zero). The first-derivative-of-log method requires the noise floor to be subtracted before log transformation.

Background or stray light reaching the detector. Stray light produces a power-independent baseline offset in the LIV. The two-segment fit handles this via the sub-threshold intercept; the other methods require explicit baseline subtraction.

Insufficient current resolution near threshold. Methods 2 and 3 require fine current steps near threshold (typically $\leq 0.5\%$ of $I_\text{th}$) to resolve the second-derivative peak or inflection point. Coarse-step measurements ($> 5\%$ steps near threshold) produce poor extraction by these methods.

Validation

For all methods, extracted $I_\text{th}$ should lie within the apparent transition region of the LIV — between where the sub-threshold and above-threshold linear approximations clearly hold. A threshold value outside the visible transition indicates extraction failure.

The four methods should be in approximate agreement (within $\sim 20\%$) for a clean LIV. Large disagreement among methods indicates either a non-standard device transition, measurement noise contamination, or a non-physical artifact in the dataset.

References

For the original characterization of the LIV transition region and the role of spontaneous emission coupling factor $\beta$, see Coldren, Corzine, and Mašanović (2012), chapter 2. For comparative studies of threshold extraction methods including statistical analysis across many devices, see Anderson et al. (2003) on semiconductor laser characterization standards. For the high-$\beta$ regime where standard methods break down, see Björk et al. (1994) on definition of laser threshold.