Baseline drift is a change with time in the non-signal level of a measurement readout. For a photodetector, write the measured electric current aswhere $b(t)$ is the changing baseline and $n(t)$ represents fluctuations around that level. Detector baseline subtraction using one fixed value can leave a time-dependent residual even when it correctly removes the initial offset.

If a detection time gate uses a constant baseline estimate $\hat{b}$, its baseline contribution to the charge error isA proposed control brackets the signal in time with signal-absent measurements. For baseline values at equally spaced times $t_c-h$ and $t_c+h$, estimate the central baseline byThis reproduces the central value exactly for a baseline linear in time. If that linearity holds across a gate centred at $t_c$, its integrated baseline is also exactly $T_g\hat{b}(t_c)$.

For an original example, choose baseline readings $2$ and $6\mathrm{nA}$ before and after the signal, with linear drift between them. The central baseline is $4\mathrm{nA}$. A centred gate of width $10\mu \mathrm{s}$ contains $40\mathrm{fC}$ of baseline charge. Subtracting only the earlier reading removes $20\mathrm{fC}$ and leaves a false positive contribution of $20\mathrm{fC}$.

Bracketing is not an unconditional drift correction. For $b(t_c+s)=b_0+vs+c{s}^2/2$, the two-point estimate of the central level is $b_0+c{h}^2/2$, so curvature introduces an error. Random baseline-readout errors also contribute measurement uncertainty. Taking reference readings close to the signal limits some drift effects, but does not remove a source-correlated artifact or compensate for all changes in the measurement conditions.

A real gated-integrator implementation alternates signal-present and signal-absent triggers and reverses the latter contribution before averaging. The manufacturer's stated purpose is active baseline subtraction.^[ref] Its effectiveness depends on whether the signal-absent readings represent the baseline during the signal. Plotting those readings against time, as well as reporting their average, can distinguish a changing zero from uncorrelated measurement noise.

Gate timing jitter is shot-to-shot variation in the timing of a detection time gate relative to the signal being measured. It differs from a fixed delay error: repeated gates may have the correct mean delay while admitting different portions of a rapidly varying detector signal. The SR250 catalogue describes a trigger-controlled gate, finite delay jitter and a gate-monitor output that can be overlaid with the signal on an oscilloscope.^[ref]

For an original sensitivity derivation, let a reproducible baseline-corrected electric current waveform be $i(t)$ and the intended integrated electric charge beSmall edge errors giveIf both edges shift by the same amount, the gate width is unchanged butThus a gate translating within a flat plateau has little first-order sensitivity, while a gate whose edges intersect different signal levels can be sensitive even at fixed width.

For stipulated edge currents $i(t_0)=100\mathrm{nA}$ and $i(t_1)=20\mathrm{nA}$, a common timing error with standard deviation $2\mathrm{ns}$ contributes approximately $0.16\mathrm{fC}$ charge standard deviation. This calculation assumes smooth waveform values near both edges and timing excursions small enough for linearization. A gate crossing an abrupt onset requires the actual waveform integral.

Gate-width jitter is a separate effect. For a constant plateau current $I$, $Q=I(t_1-t_0)$, so $\delta Q=I\delta T_g$ even though a common gate translation would cancel. More generally, the edge-error contribution has varianceThe covariance determines whether two individually noisy edges largely preserve the width.

An apparatus check records both the gate monitor and signal waveform over many triggers, using the same time reference, to estimate relative delay and width distributions. A single aligned trace establishes neither distribution. Timing motion shared by the pulse and gate can cancel in their relative delay; the uncertainty relevant to the measured charge is their relative timing, not solely a gate generator's standalone specification.

A gate-normalized detector output reports an integrated detector signal divided by the duration of the detection time gate. For an input electric potential difference waveform $v(t)$ and a gate from $t_0$ to $t_0+T_g$, the input-referred average isIts units are those of electric potential difference, whereas the integral has units of electric potential difference times time. The SR250 gated integrator implements integration followed by gate-width normalization and an adjustable amplification stage.^[ref]

For an original idealized detector pulse, take $v(t)=2\mathrm{mV}$ from0 to4nanoseconds and zero outside that interval. A gate covering0–4nanoseconds integrates $8\mathrm{mV}\mathrm{ns}$ and reports an average of2millivolts. A gate covering0–8nanoseconds integrates the same area but reports1millivolt. Thus unchanged admitted pulse area need not give an unchanged normalized output.

The gate average also differs from the average over a full repetition period. For the same pulse repeated once per microsecond, the full-period mean is only8microvolts. Comparing instruments therefore requires knowing whether the reported quantity is pulse area, within-gate average or full-period average.

A constant baseline $b$ contributesGate-width normalization does not remove that baseline. A separate reference or subtraction measurement is needed to distinguish it from the desired signal.

An apparatus check overlays the gate monitor and detector waveform on an oscilloscope, then varies the gate width while keeping its start fixed. A flat plateau should retain its input-referred average while the gate remains inside it; a short pulse diluted by additional quiet time behaves like the example above. Gate timing jitter and gate-width jitter can alter which signal portion is admitted and which duration is used for normalization.^[ref]

The major index of a permutation $\pi ={\pi }_1\cdots {\pi }_n$ is the sum of its permutation descent positions:It uses positions, rather than the values being compared. For $3142$, the permutation descents occur at positions one and three, giving $\mathrm{maj}=4$, whereas the descent number is two.

The major index has the same distribution over all permutations of $n$ elements as the inversion number:For $n=3$, the six values give $1+2q+2q^2+q^3$. Individual statistics need not agree: $312$ has major index one and two inversions. The shared generating polynomial is a statement about their distribution across the whole symmetric group.^[ref]

The Bourgain-Demeter decoupling theorem controls a function whose Fourier support lies near a positively curved hypersurface by the contributions from small frequency caps. For a compact $C^2$ hypersurface in ${\mathbb{R}}^n$, $n\ge 2$, with positive-definite second fundamental form, decompose its $\delta$-neighborhood into caps of tangential width ${\delta }^{1/2}$ and normal thickness $\delta$. If $f_{\theta }$ denotes the corresponding Fourier pieces, the critical estimate isThe constant is independent of $\delta \in (0,1)$ and $f$, for the fixed surface and chosen cap decomposition. The same arbitrarily small loss holds for $2\le p\le p_c$; supercritical exponents require an additional scale power.^[ref]

For a parabola in ${\mathbb{R}}^2$, $p_c=6$. A naive triangle inequality sums all cap norms, whereas decoupling sums their squares before taking a square root. Curvature makes this much stronger control possible. The theorem links Fourier restriction theory with estimates for oscillatory sums and dispersive equations.

A decomposition has a lossless join when projecting any permitted Database relation onto the component schemas and taking their natural join reconstructs exactly the original relation. Every original tuple reappears; the substantive requirement is that no spurious tuples are introduced. The dependencies defining permitted instances are essential to the claim.

For example, decompose $R(A,B,C)$ into $R_1(A,B)$ and $R_2(B,C)$. If $B\to C$ is a functional dependency, this is lossless. Any joined tuple combines $(a,b)$ and $(b,c)$ from original rows; the dependency forces every row with that $b$ to have the same $c$, so the combined tuple was already present. Without that dependency, original rows $(a_1,b,c_1)$ and $(a_2,b,c_2)$ yield two additional crossed combinations on rejoining. Removing redundancy is therefore insufficient: the decomposition must preserve which facts belonged together. A binary decomposition is lossless under functional dependencies when the common attributes determine one of its component schemas.