📝 Taylor series for e^x (39 MCQs)

📖 Explanation: The Maclaurin series for $e^{-5}$ is an alternating series with terms that initially increase in magnitude before decreasing. Summing large positive and negative numbers leads to catastrophic cancellation, where significant digits are lost due to finite machine precision. In contrast, $e^5$ involves only positive terms, avoiding this subtraction error. This highlights that mathematical convergence does not guarantee numerical stability, necessitating alternative strategies like computing $1/e^5$ for negative arguments.

📖 Explanation: While the exponential series converges for all real $x$, practical computation faces overflow issues. For large $x$, the terms $x^k/k!$ grow enormously before eventually decaying. Even if the final sum is representable, intermediate terms may exceed floating-point maximums. Furthermore, achieving a fixed relative error requires roughly $O(x)$ terms, making direct summation computationally prohibitive. This necessitates range reduction techniques or asymptotic expansions rather than naive series evaluation for large arguments.

📖 Explanation: Padé approximants often outperform Taylor polynomials of comparable degree because they incorporate information about the function's global structure through rational approximation. While $P_4$ matches derivatives up to order 4, $R_{2,2}$ effectively captures asymptotic behavior and curvature that polynomials miss. For $e^x$, which has no poles and specific growth characteristics, the rational form reduces truncation error significantly. This illustrates that matching local derivatives is necessary but not sufficient for optimal approximation; functional form matters.

📖 Explanation: On $[-1, 0]$, $f^{(n+1)}(x) = e^x$ is strictly increasing, so its maximum occurs at the right endpoint $x=0$, giving $M=e^0=1$. Using $M=e^{-1}$ would be incorrect as it underestimates the derivative’s maximum. The strategic advantage of $M=1$ is that it yields a clean, computable bound $|R_n| \leq \frac{1}{(n+1)!}$ independent of unknown constants. This demonstrates careful analysis of monotonicity in error bounding, avoiding both overestimation and invalid underestimation.

📖 Explanation: While the series converges pointwise everywhere and uniformly on any bounded interval $[-R, R]$, it does NOT converge uniformly on $(-\infty, \infty)$. Uniform convergence on an unbounded domain would require $\sup_{x\in\mathbb{R}} |e^x - S_n(x)| \to 0$, but $e^x$ grows faster than any polynomial, so the tail remains arbitrarily large for sufficiently negative partial sums evaluated at large positive $x$. Confusing pointwise term decay with uniform series convergence is a subtle but critical distinction in analysis.

📖 Explanation: The relative error of the linear approximation $1 - x$ versus $e^{-x}$ is approximately $\frac{|e^{-x} - (1-x)|}{e^{-x}} \approx \frac{x^2/2}{1}$ for small $x$. Setting $x^2/2 < 0.01$ gives $x < \sqrt{0.02} \approx 0.141$. However, checking exact values: at $x=0.02$, relative error ≈ 0.0198%; at $x=0.1$, ≈ 0.5%. The 1% threshold occurs near $x \approx 0.14$, but among options, $\lambda t < 0.02$ safely ensures <1% error while $0.1$ also satisfies it. Given typical engineering conservatism and the quadratic nature, $\lambda t < 0.02$ is the most defensible safe bound listed, though $0.1$ technically works. Re-evaluating: exact relative error at 0.1 is ~0.5%, so C is acceptable, but B is overly conservative. Correct answer should be C based on calculation. Adjusted explanation reflects this nuance.

📖 Explanation: Although the Maclaurin series converges for all $x$, the rate of convergence is not symmetric. For $x>0$, $e^x$ grows exponentially while polynomials grow algebraically, so error accumulates faster on the positive side. For $x<0$, $e^x$ decays to zero and the alternating nature helps control error. Thus, for a fixed tolerance, the usable interval extends further left than right. Graphs visually confirm this asymmetry, illustrating that infinite radius of convergence doesn't imply uniform approximation quality.

📖 Explanation: Computing $e^{0.1}$ via series is extremely accurate with few terms, but raising it to the 100th power compounds relative errors multiplicatively: if $\tilde{a} = a(1+\delta)$, then $\tilde{a}^{100} \approx a^{100}(1+100\delta)$. Conversely, computing $e^5$ requires more terms but only one squaring operation, limiting error amplification. Despite $e^5$ being larger, the single squaring step keeps error manageable. This trade-off between per-operation accuracy and error propagation depth is central to numerical algorithm design for special functions.

📖 Explanation: The Lagrange remainder formula requires an upper bound $M$ on $|f^{(4)}(c)|$ for $c$ between 0 and 0.5. Since $f^{(4)}(x)=e^x$ is increasing, $M=e^{0.5}$. But if we don’t know $e^{0.5}$ exactly, we cannot evaluate this bound numerically. Practitioners replace $e^{0.5}$ with a known upper estimate like $e^{0.5} < 2$ or use iterative refinement. Recognizing this circularity is essential for turning theoretical error bounds into actionable computational guarantees.

📖 Explanation: Expanding directly at $x=2$ gives perfect accuracy with zero terms but defeats the purpose of series approximation. Continuing fractions help but aren't series-based. Range reduction via $e^2 = (e^{0.5})^4$ leverages the fact that series converge much faster for smaller arguments. Computing $e^{0.5}$ to high accuracy with few terms and then raising to the 4th power exploits the exponential's functional equation to accelerate effective convergence. This is a standard technique in library implementations of exp().

📖 Explanation: Maclaurin polynomials minimize error near $x=0$ but allow larger errors near $x=1$. Chebyshev approximation distributes error nearly uniformly across $[0,1]$, achieving smaller maximum error for the same degree. For $e^x$ on $[0,1]$, a degree-5 Chebyshev approximant often matches the accuracy of degree-7 Maclaurin. This reflects the fundamental difference between local (Taylor) and global (minimax) approximation paradigms. Understanding this distinction is crucial for efficient numerical software design where worst-case error matters more than local fidelity.

📖 Explanation: The Taylor expansion shows $e^{-x} = 1 - x + x^2/2 - \cdots$, so truncating after linear term gives local error $\approx x^2/2$. With $x = \Delta t / \tau = 0.1$, per-step error is $O(0.01)$. Over $N$ steps covering fixed time $T = N \Delta t$, total error accumulates as $N \cdot O((\Delta t)^2) = O(T \Delta t)$, which is first-order globally. This linear accumulation (not square-root) arises because truncation error is systematic, not stochastic. Recognizing error order is vital for choosing appropriate timesteps in simulations.

📖 Explanation: Any odd-degree polynomial with positive leading coefficient goes to $-\infty$ as $x \to -\infty$, so $P_{2k+1}(x)$ must eventually become negative, contradicting $e^x > 0$. Even-degree truncations remain positive for large negative $x$ but may still dip below zero near the origin for low degrees. This structural mismatch shows that finite polynomials cannot replicate global properties of transcendental functions. Awareness of such limitations prevents misinterpretation of truncated series in contexts where sign preservation is physically meaningful, like probability densities.

📖 Explanation: For large $\theta$, individual terms $\theta^n/n!$ become enormous before canceling to produce bounded trig values. This causes catastrophic cancellation in finite precision. Reducing $\theta \mod 2\pi$ exploits periodicity to work with small arguments where series behave well. Increasing precision treats symptoms, not causes. Built-in functions already do this internally. Direct complex summation doesn't solve the magnitude issue. This exemplifies leveraging mathematical structure (periodicity) to avoid numerical instability, a key principle in scientific computing beyond mere series manipulation.

📖 Explanation: By definition, the Maclaurin polynomial $P_n(x)$ is constructed so that $P_n^{(k)}(0) = f^{(k)}(0)$ for $k=0,1,\dots,n$. It does not minimize max error (Chebyshev does) or MSE (Legendre projection does). This local optimality at zero comes at the cost of potentially poor performance away from zero. Understanding what Taylor polynomials actually optimize clarifies when they are appropriate versus when global approximation methods are needed. This foundational concept underpins all subsequent discussions of series approximation quality.

📖 Explanation: Substituting $u=x^2$, we approximate $e^{-u}$ at $u=4$. Terms $(-4)^k/k!$ grow until $k\approx4$, reaching magnitude ~26.7, whereas for $e^{-2}$ peak is ~4 at $k=2$. Larger intermediate terms mean more severe cancellation when summing alternating series, demanding higher precision. Also, more terms are needed before decay dominates. This compound effect makes Gaussian-type evaluations trickier than simple exponentials at equivalent nominal argument sizes, illustrating how composition affects numerical behavior.

📖 Explanation: On a constrained device targeting a specific range (e.g., $[-1,1]$), minimax rational approximations achieve desired accuracy with lower-degree numerator/denominator than Taylor polynomials, reducing multiply-add operations. Factorials can be precomputed as constants. Infinite radius isn't relevant for bounded domains. The key is operation count vs. accuracy trade-off on fixed hardware. This reflects real-world embedded systems design where theoretical elegance yields to resource-aware approximation theory, bridging pure math and engineering constraints.

📖 Explanation: The sequence $(1+x/n)^n$ approaches $e^x$ with error proportional to $x^2/(2n)$, so achieving $10^{-6}$ accuracy needs $n \sim 10^6$. Series error decays factorially, needing ~10 terms for same accuracy. While the limit is pedagogically valuable and useful in some probabilistic contexts, it's computationally inefficient for direct evaluation. Recognizing convergence rates distinguishes theoretically equivalent definitions in practice. This reinforces that multiple representations exist for different purposes: conceptual understanding vs. numerical efficiency.

📖 Explanation: Continuity requires $\lim_{x\to a} f(x) = f(a)$. For series, this means interchanging limit and infinite sum, which is justified by uniform convergence on compact intervals containing $a$. Pointwise convergence alone doesn't permit this interchange. While radius of convergence ensures pointwise convergence, continuity proof specifically needs uniform convergence on neighborhoods. Students often overlook this subtlety, assuming pointwise convergence suffices for continuity. Highlighting this gap strengthens understanding of why stronger convergence notions are necessary for preserving analytical properties under limits.

📖 Explanation: With 6-digit precision, numbers below $10^{-6}$ relative to 1 are lost in rounding. The quadratic term $x^2/2$ exceeds $10^{-6}$ only when $x > \sqrt{2\times10^{-6}} \approx 0.0014$. Below this, adding it changes nothing in stored representation. Including unnecessary terms wastes computation without benefit. This threshold analysis exemplifies adaptive algorithm design: choose approximation order based on input magnitude and available precision. Blindly adding terms ignores numerical reality, while principled thresholds optimize both accuracy and efficiency.

📖 Explanation: Table lookup stores precomputed values and interpolates smoothly, avoiding series summation entirely. Thus, catastrophic cancellation from alternating large terms never occurs. Series methods inherently face this for $e^{-x}, x>0$. Other issues like slow convergence affect both (tables need dense sampling), and irrationals/factorials are implementation details. This contrast shows that algorithm choice fundamentally alters error profiles. Understanding failure modes guides robust library design: hybrid approaches often combine tables for coarse approximation with series for refinement, mitigating individual weaknesses.

📖 Explanation: Convergence of $\sum a_k$ doesn't guarantee sum equals $f(x)$; there exist smooth non-analytic functions whose Taylor series converge to wrong values. Proving $f(x) = \sum f^{(k)}(0)x^k/k!$ requires showing the specific remainder $R_n(x) = f(x)-S_n(x) \to 0$, not just that terms vanish. Term decay is necessary but not sufficient for representation. This distinction separates formal power series from actual function representation, a cornerstone of real analysis often glossed over in computational contexts.

📖 Explanation: While the matrix exponential series $\sum A^k/k!$ converges for all matrices (due to norm submultiplicativity), many scalar identities fail: $e^{A+B} \neq e^A e^B$ unless $AB=BA$, and eigenvalue-based shortcuts require diagonalizability. Non-commutativity breaks product rules and complicates error analysis. Scalar intuition about term-wise operations doesn't transfer. This highlights that extending scalar concepts to operators demands careful re-examination of underlying assumptions. Matrix exponentials are fundamental in differential equations and quantum mechanics, making this distinction critically important beyond pure mathematics.

📖 Explanation: Mathematically, for fixed absolute error $\epsilon$, the number of terms needed for $e^x$ increases with $|x|$, but asymmetrically. For $x=10$, partial sums grow to ~22000 before converging to ~22026, requiring many terms to resolve the difference within $\epsilon$. For $x=-10$, partial sums oscillate around tiny value ~0.000045, and absolute error drops quickly once terms decay below $\epsilon$. Thus, same absolute tolerance demands more terms for positive $x$. This asymmetry stems from the function's growth, not numerical artifacts.

📖 Explanation: Integrating $\sum x^k/k!$ term-by-term yields $\sum x^{k+1}/((k+1)k!) + C$, which is straightforward symbolically. While $e^x$'s self-differentiation helps, the key advantage for series is operational simplicity: integration reduces to index shifting. Numerical quadrature works fine for exponentials, and symbolic methods don't always prefer series. But when dealing with compositions like $e^{x^2}$ lacking elementary antiderivatives, series provide tractable symbolic expressions where closed forms fail. This showcases series as bridges between computable and non-integrable functions.

📖 Explanation: For $x=0.01$, quadratic term is $5\times10^{-5}$, cubic is $1.67\times10^{-7}$. Including up to $x^2/2$ gives 1.01005 with error < $2\times10^{-7}$, sufficient for 5 decimals. Mental computation is feasible: $0.01^2=0.0001$, half is 0.00005. Higher terms are negligible. Memorization isn't problem-solving. Limit definition is impractical mentally. Linear approx loses 5th digit. This exemplifies intelligent truncation: assess term magnitudes to determine minimal sufficient order, combining mathematical insight with mental arithmetic efficiency.

📖 Explanation: The convergence speed of a power series at a point depends on how quickly terms decrease. For $e^x$, term ratio is $|x|/(k+1)$. At $x=0.1$, this ratio drops below 0.1 immediately, while at $x=1$, it takes several terms. Faster ratio decay means fewer terms needed for given accuracy. Radius of convergence determines where it converges, not how fast. This quantitative understanding of term ratios is essential for predicting computational effort and selecting appropriate approximation strategies across different input ranges.

📖 Explanation: At $x=0$, series works perfectly. Near overflow/underflow limits, behavior is dominated by range checks, not series accuracy. But at $x \approx 355$, $e^x$ is large but representable, and naive series would require hundreds of terms with massive intermediate values. A correct implementation switches to range-reduced or asymptotic methods here. Testing this boundary reveals whether the algorithm gracefully transitions between regimes. This meta-level validation ensures robustness across the entire domain, not just at extremes or trivial points.

📖 Explanation: Analytic functions have unique power series representations within their radius of convergence. Substitution $u=2x$ into $\sum u^k/k!$ gives $\sum (2x)^k/k! = \sum 2^k x^k/k!$. Direct differentiation yields $f^{(k)}(0) = 2^k e^0 = 2^k$, so series is $\sum 2^k x^k/k!$. Identical results confirm consistency of methods. Uniqueness theorem guarantees this equivalence for all analytic functions. This reinforces that valid manipulations of convergent series preserve the represented function, building confidence in algebraic techniques for generating new series from known ones.

📖 Explanation: For $x=1$, direct series needs ~15 terms for double precision. Scaling by $2^m$ (e.g., $m=4$, $y=1/16$) reduces series to ~5 terms. Then 4 squarings recover $e^1$. Squaring is one multiplication, far cheaper than series terms involving division/factorials. Intermediate values stay moderate, improving stability. This range reduction is standard in math libraries. It exemplifies algorithmic optimization: transform problem to regime where basic operations are most efficient, balancing computational cost against numerical robustness.

📖 Explanation: Near zero, $e^x$ terms decay as $x^k/k!$, extremely rapid. $\ln(1+x)$ terms decay as $x^k/k$, much slower; e.g., at $x=0.1$, $e^x$ 5th term is $10^{-5}/120 \approx 8\times10^{-8}$, while $\ln$ 5th term is $10^{-5}/5 = 2\times10^{-6}$, 25x larger. This disparity persists near zero due to denominator growth rates. Understanding this explains why logarithm approximations often need acceleration techniques while exponentials rarely do, despite both being analytic at origin.

📖 Explanation: Assuming $y=\sum a_k x^k$, substitution gives $\sum k a_k x^{k-1} = \sum a_k x^k$, yielding recurrence $a_{k} = a_{k-1}/k$. With $a_0=1$, this uniquely gives $a_k=1/k!$. Since $e^x$ satisfies the ODE and has this series, uniqueness of ODE solutions plus uniqueness of power series coefficients guarantees identity. This connects differential equations, series, and function theory elegantly. It shows how structural constraints (ODE + IC) fully determine the series, reinforcing deep links between analytical objects.

📖 Explanation: The series converges for all $x$, but approximation quality degrades as $|x|$ increases beyond the degree-dependent sweet spot. Divergence in graph indicates practical unusability, not mathematical invalidity. Students often conflate theoretical convergence with practical accuracy. Clarifying this distinction is vital: infinite radius means the series represents the function everywhere in the limit, but finite truncations have limited useful domains. This misconception can lead to rejecting valid methods prematurely or misinterpreting computational results.

📖 Explanation: Taylor at $x_0=1$: $e^{1.1} = e \sum (0.1)^k/k!$. Maclaurin: $\sum (1.1)^k/k!$. Though Taylor requires multiplying by $e$, the series argument 0.1 vs 1.1 means far fewer terms for same accuracy (ratio 0.1/(k+1) vs 1.1/(k+1)). Precomputing $e$ once amortizes cost. For repeated evaluations near 1, Taylor wins decisively. This illustrates strategic center selection: match expansion point to evaluation region to minimize computational load, a key optimization in numerical software.

📖 Explanation: For $x>0$, $M=e^x$ in remainder bound is the very quantity sought, but partial sums $S_n(x)$ underestimate $e^x$, so $S_n(x) < e^x < S_n(x) + \text{bound}$. Iteratively refining $M$ using current estimate creates self-consistent error control. For oscillatory functions, $M$ is independent of partial sums. This self-referential property enables adaptive algorithms that tighten bounds dynamically. It exemplifies how structural properties of $e^x$ facilitate robust numerical verification beyond generic error formulas.

📖 Explanation: Terms $(-100)^k/k!$ peak around $k=100$ with magnitude ~$10^{40}$, while true value is ~$10^{-44}$. Summing $10^{40}$-scale alternating terms to get $10^{-44}$ requires ~84 decimal digits of precision; double has only ~16. All significance is lost. Convergence is mathematical, not numerical. Overflow/underflow may occur but cancellation is the fundamental barrier. This stark example shows that theoretical convergence ≠ numerical feasibility, mandating alternative approaches like $1/e^{100}$ with positive series for negative exponents.

📖 Explanation: Integrating $P_n(x) = \sum a_k x^k$ gives $\sum a_k/(k+1)$ exactly with basic arithmetic. Numerical quadrature of $e^x$ introduces discretization error and requires multiple expensive exp() calls. For modest accuracy needs, polynomial integration eliminates two error sources (quadrature + function eval) and is computationally lighter. This exemplifies replacing expensive operations with cheap surrogates when precision allows, a core numerical analysis principle. It shows series enable analytical simplifications that bypass numerical integration entirely.

📖 Explanation: For $x=20$, terms $20^k/k!$ don't start decaying until $k>20$, and peak around $k=20$ at ~$4\times10^7$. Using only 30 terms misses the tail where significant contribution resides, yielding gross error. Theoretical convergence requires $n \to \infty$; practical use demands $n$ large enough for term decay. This gap between theory and practice is critical: students must learn to estimate required terms via ratio test or Stirling, not assume convergence implies usability. This question targets that essential skill.

📖 Explanation: Expanding $e^{rt} = \sum (rt)^k/k!$ gives $P(t) = P_0 \sum t^k r^k /k!$. Differentiating w.r.t. $r$ yields $\partial P/\partial r = P_0 \sum k t^k r^{k-1}/k! = P_0 t \sum (rt)^{k-1}/(k-1)! = P_0 t e^{rt}$, same as closed-form. But series form allows term-wise sensitivity decomposition, revealing which orders dominate uncertainty propagation. For perturbation methods or polynomial chaos expansions, series enable analytical sensitivity structures inaccessible via black-box exponentials. This shows series as tools for structural insight beyond mere computation.