GATE Electrical: Probability, Statistics and Complex Variables (Engineering Mathematics), previous-year questions solved
The 12 questions from the last five GATE Electrical papers that test probability, statistics and complex variables: normalising a density, conditional expectation, functions of a random variable, unbiased estimators, least squares, analytic functions, Taylor coefficients and contour integrals by residues. Each solution matches the official answer key, was checked by enumeration or numerical integration, names the trap that costs marks, and shows where the idea is used in electrical and communication engineering.
Questions from official GATE EE papers (IIT Kharagpur 2022, IIT Kanpur 2023, IISc Bengaluru 2024, IIT Roorkee 2025, IIT Guwahati 2026); solutions by Skill-Lync. · Updated 16 Nov 2026 · 12 questions
Probability density f(x) = a e^(−2|x|): what is a?
Answer: 1 Official key: 0.99 to 1.01
- Any probability density must integrate to 1 over all : .
- The function is even, so integrate over and double it: .
- So .
Trap: forgetting the absolute value and integrating only from 0 to gives and . The density is defined for negative too. Check: numerical integration of from to gives .
This double-sided exponential is the Laplacian distribution. Speech amplitudes and the prediction errors inside audio and image codecs are modelled with it, and quantiser step sizes are designed from it. Whatever density you fit to measured noise or error data, the first check is the same as here: it must integrate to 1 before any probability you read off it means anything.
Expected number of trials to the first head is 4: probability that the first head comes on trial 2?
Answer: 0.1875 Official key: 0.187 to 0.188
- The number of tosses up to the first head is geometric: , with mean .
- , so the head probability is .
- First head on the second toss means tail, then head: , which is to three decimals.
Traps: answering itself, or (two heads in a row). Check: summed to 400 terms is 4.
The same distribution counts retransmissions on a radio link. If each packet gets through with probability , the number of attempts until the first success is geometric with mean : a link with needs 4 attempts on average. Wi-Fi and LTE retransmission (ARQ/HARQ) limits and latency budgets are sized from exactly this calculation.
X is uniform on {−10, …, 10}: which of X², X³, (X−5)², (X+10)² are uniform?
- (A)
- (B)
- (C)
- (D)
Answer: (B) and (D) Official key: B; D
- Each of the 21 values of has probability . A function of stays uniform if it never sends two values of to the same output (it is one-to-one on the set): then each output value still has probability .
- is one-to-one (it keeps the sign): (B) is uniform over 21 values.
- runs over , all non-negative, so squaring it is one-to-one: (D) is uniform too.
- folds and together: but . Not uniform. runs over , so collide after squaring: has probability , only . Answer (B) and (D).
We tabulated all 21 outcomes for each option: only the cube and give 21 distinct values with equal counts.
A square-law (power) detector does what option (A) does: it maps and to the same reading, so a signal whose voltage is spread evenly around zero gives a power reading that is not evenly spread. In test equipment, a uniformly distributed code (from an LFSR, say) stays uniform under any one-to-one mapping such as an offset or a bit reordering, which is why those are safe ways to shape dither and test patterns.
Joint probability table of X and Y: what is E(Y | X = 1)?
- (A)
- (B)
- (C)
- (D)
Answer: (B) Official key: B
- Condition on : .
- and .
- : option (B).
Traps: taking the joint probability without dividing by gives option (C); is option (A). The unconditional mean is not an option.
A receiver on a noisy binary channel reasons exactly like this. From the joint table of sent and received bits (measured in a bit-error-rate test), the probability that a 1 was sent given that a 1 was received is the joint probability divided by the probability of receiving a 1. Maximum a posteriori detectors and error-correcting decoders are built on these conditional probabilities.
X is uniform on (0, 1] and Y = X²: which is the density of Y?
- (A)
- (B)
- (C)
- (D)
Answer: (A) Official key: A
- Work with the CDF. For : , because is uniform on and positive.
- Differentiate: on : option (A).
- Shortcut: with , so .
Trap: all four options integrate to 1 on , so checking the normalisation does not decide. Squaring numbers below 1 makes them smaller, so piles up near 0; only (A) grows towards . We checked the CDF of every option against a direct count of values with at four values of : only (A) matches.
Power goes as the square of voltage, so this is how a voltage distribution turns into a power distribution. If a sensor voltage is spread evenly between 0 and 1 V, the power in a 1 Ω load is not spread evenly: it has density and half of the readings fall below 0.25 W. The same change-of-variable rule converts Rayleigh-distributed signal amplitudes on a fading radio link into the exponential distribution of received power.
Zero-mean Gaussian samples: is dividing by 10000 or by 9999 the unbiased variance estimate?
- (A)
- (B)
- (C)
- (D)
Answer: (C) Official key: C
- The mean is known to be zero, so for every sample. Then : option (C).
- , slightly too big, so (A) is false.
- is a fixed multiple of , so their means differ: (D) is false.
- Square roots are biased even when the square is not: and (from the chi distribution with 10000 degrees of freedom), so (B) is false too.
Trap: the rule (Bessel's correction) is for when the mean is estimated from the same samples, which uses up one degree of freedom. Here the mean is known to be zero and nothing is estimated, so dividing by is the unbiased choice. A Monte Carlo run with 200,000 sets of 5 samples confirms both effects: the variance is unbiased and its square root comes out low.
This is how input-referred noise of an ADC or amplifier is measured: short the input, record N samples and compute the RMS. If you subtract the sample mean to remove the offset, divide by N − 1; if the offset was removed separately (AC coupling, a known zero), divide by N. With 10,000 samples the difference is 0.01 %; the bigger lesson is that an RMS (a square root) is biased low for small N, which matters when a datasheet quotes noise from a few dozen samples.
Least-squares straight line through (−1, 0.8), (0, 2.2), (1, 2.8): what is the slope?
Answer: 1.00 Official key: 0.9 to 1.1
- For the normal equations give .
- Here , , , , .
- , and .
Shortcut: with three equally spaced values the middle point only moves the intercept, so the slope is . We checked that changing the middle leaves the fitted slope unchanged. The residuals are , and , adding to zero as a least-squares line's must. Trap: the slope of one pair of neighbouring points, 1.4 or 0.6.
Sensor calibration is this calculation. Apply known currents to a Hall-effect current sensor (or known temperatures to a thermocouple), record the outputs and fit a least-squares line: the slope is the gain, the intercept the offset. Energy meters and battery management systems store those two numbers per channel and correct every reading with them.
Given the moments of X and Y, which α minimises E((X − αY)²)?
Answer: 0.1 Official key: 0.1 to 0.1
- Expand: .
- Set the derivative to zero: , so . The minimum value is .
Trap: and are not needed. Removing the means and using answers a different question (the best fit with an intercept); is also wrong. We scanned from to in steps of : the minimum is at .
This is the one-tap Wiener filter: the best scale factor for predicting one signal from another, in the mean-square sense, is the cross-correlation over the power, . Echo cancellers in phones and adaptive noise cancellers use LMS filters that converge to exactly this solution, tap by tap.
Which of j Re(z), Im(z), e^|z| and z² − z are analytic on the whole complex plane?
- (A)
- (B)
- (C)
- (D)
Answer: (D) Official key: D
- Write and . Analytic means the Cauchy-Riemann equations hold: and .
- (A) : , . but . Not analytic.
- (B) : , . but . Not analytic.
- (C) is real-valued and not constant. A real-valued function has , so Cauchy-Riemann would force , a constant. Not analytic.
- (D) A polynomial in is analytic everywhere (entire): . Answer (D).
Shortcut: anything built from , , or is not analytic (except where it reduces to a constant). We checked each function at 50 random points: only (D) has the same difference quotient along the real and the imaginary direction.
A circuit's transfer function is a ratio of polynomials, so it is analytic everywhere except at its poles. That property is what makes the Nyquist criterion work: by the argument principle, the encirclements of by count the right-half-plane poles of the closed loop. A magnitude such as is real-valued and not analytic, so it carries no such information on its own.
Coefficient of z⁵ in the Taylor series of cos z + e^(z²)
Answer: 0 Official key: 0.0 to 0.0
- has only even powers.
- also has only even powers.
- So the sum has no term: the coefficient is .
Shortcut: both terms are even functions, , so every odd coefficient is zero without expanding anything. Trap: computing the coefficient instead, . Check: the Cauchy integral formula on , evaluated numerically, gives .
Symmetry decides which harmonics a nonlinearity makes. A push-pull amplifier has a nearly odd transfer curve, so its power series has mainly odd powers and its distortion is mostly 3rd and 5th harmonic; a full-wave rectifier, , is even and puts out only DC and even harmonics of the input frequency. Power-quality engineers read the harmonic spectrum to tell which kind of load is on the line.
Integral of f(z) = jz clockwise around |z| = 1
Answer: 0 Official key: 0 to 0
- is a polynomial, analytic everywhere inside and on .
- Cauchy's theorem: the integral of an analytic function around a closed curve is . The direction (clockwise) does not matter when the answer is zero.
Direct check: on , . Numerically, the clockwise integral comes out below . Trap: mixing this up with , which is (here clockwise) because has a pole inside.
Cauchy's theorem is the zero case of the Nyquist test: if a contour encloses no poles or zeros of , the plot of around it does not encircle the origin. Control-design tools such as MATLAB's nyquist count encirclements on exactly this basis to decide closed-loop stability.
Magnitude of the contour integral of (z+1)²/((z−i)(z−2)) around |z − 2 − i| = 3/2
Note: is a complex variable and
Answer: 25.29 Official key: 25.00 to 25.60
- The contour is a circle of radius centred at . Pole is at distance : inside. Pole is at distance : outside.
- Residue at : .
- , so the magnitude is .
Trap: adding the residue at as well gives , outside the key range. Always check which poles the contour encloses before summing residues. Check: integrating numerically around the circle (4000 points) matches the residue to 9 decimal places.
Inverse Laplace transforms of circuit responses are sums of residues: the time response is over the poles of , which is what partial fractions compute. Each pole gives one term of the transient, and the residue sets its size and phase. Simulation tools that print a 'pole-residue' model of a filter or interconnect are listing exactly these numbers.